JP7469828B2 - Structure diagnosis system, structure diagnosis method, and structure diagnosis program - Google Patents

Description

The present invention relates to a structure diagnosis system, a structure diagnosis method, and a structure diagnosis program.

Conventionally, there has been technology that estimates characteristic quantities (vibration characteristics) such as the natural frequency, damping coefficient, and mode shape of vibration based on the acceleration of structures such as bridges, and diagnoses abnormalities in the structure based on changes in the estimated vibration characteristics.

For example, Patent Document 1 discloses the following technology. In particular, independent component analysis (ICA) is performed on sensor signals indicating vibration obtained from acceleration sensors installed at multiple points on the bridge, and the natural frequency of the bridge is determined by performing spectral analysis on the independent vibration components obtained by the independent component analysis. Then, based on a comparison between the natural frequency of the bridge in a healthy state obtained in advance and the natural frequency of the bridge in its current state, an abnormality diagnosis is performed on the entire bridge and abnormal locations are identified.

JP 2008-255571 A

However, the sensitivity to changes in response to damage varies depending on the location and vibration mode where the vibration is measured. Furthermore, with previous technologies, estimations vary depending on the environmental conditions (external force, temperature, etc.) at the time of vibration measurement. However, even an experienced person has difficulty in pre-setting the installation location and vibration mode of the acceleration sensor appropriately. For this reason, in the case of abnormality diagnosis of a bridge, for example, based on the natural frequency, if the vibration mode is set incorrectly, it is not possible to distinguish between changes in vibration characteristics due to abnormalities in the bridge and accidental changes in vibration characteristics due to reasons other than abnormalities in the bridge, and there is a problem that quantitative abnormality diagnosis cannot be performed with high accuracy.

Taking the above points into consideration, the present invention aims to enable quantitative abnormality diagnosis of structures with high accuracy.

In order to solve the above-mentioned problems, the present invention provides a structure diagnosis system, which is characterized in that it has a diagnosis unit that calculates an evaluation index, which is the ratio between a first marginal likelihood representing the probability distribution of a feature quantity indicating the state of a structure generated from a time series of information about the structure, when it is assumed that the structure is in a healthy state, and a second marginal likelihood representing the probability distribution of the feature quantity when it is assumed that the structure is not in the healthy state, and diagnoses the state of the structure based on the evaluation index.

According to the present invention, quantitative abnormality diagnosis of structures can be performed with high accuracy.

FIG. 1 is a diagram showing a schematic configuration of a diagnostic system according to a first embodiment. FIG. 2 is a block diagram showing the configuration of a sensor node according to the first embodiment. FIG. 1 is a block diagram showing the configuration of a diagnostic device according to a first embodiment. FIG. 4 is a sequence diagram showing a reference time process in the diagnostic system of the first embodiment. FIG. 4 is a sequence diagram showing a diagnosis process in the diagnosis system of the first embodiment. FIG. 4 is an explanatory diagram of a diagnosis process in the diagnosis system of the first embodiment. 11A and 11B are diagrams showing a comparison of the data amounts of coefficient matrices and principal component matrices for each number of sensors in the first embodiment and the prior art; 10 is a flowchart showing a diagnostic process in a diagnostic system according to a second embodiment. 13A and 13B are diagrams for explaining a method of calculating a threshold value for determining an abnormality according to the second embodiment. 13A to 13C are diagrams for explaining another example of the abnormality determination method using a plurality of threshold values according to the second embodiment. 13A to 13C are diagrams for explaining the results of an experiment on abnormality determination performed using the diagnosis system of the second embodiment. FIG. 11 is a diagram showing the execution conditions of an experiment performed using the diagnostic system of the second embodiment. FIG. 11 is a diagram showing the results of an experiment performed using the diagnostic system of the second embodiment. 13 is a diagram showing the change in the natural frequency of a bridge that differs depending on the vibration mode measured under the same conditions as in the experiment conducted using the diagnostic system of embodiment 2. FIG. A graph showing the time-dependent changes in the outside temperature and Bayes factor values for a certain girder of a bridge during a certain period of time. A graph showing the change in deflection and Bayes factor values over time for a bridge over a certain period of time. A comparison of the Bayes Factor values over time for two spans of a bridge. 13 is an explanatory diagram of sequential updating of a coefficient matrix in the diagnostic system according to the third embodiment. 13 is a diagram showing the convergence of the probability distribution of the coefficient matrix by sequentially updating the coefficient matrix in the diagnostic system of the embodiment 3. FIG. FIG. 2 is a block diagram showing the configuration of a sensor node terminal used as a sensor node according to the embodiment.

Below, an embodiment of the present invention will be described with reference to the drawings. The following embodiment is an example to fully explain the present invention, and appropriate omissions and simplifications have been made. Not all of the configurations and processes described in the following embodiment are essential to the implementation of the present invention, and may be omitted as appropriate. In addition to the following embodiment, the present invention can be implemented in various other forms that can achieve the object of the present invention. Furthermore, forms that combine some or all of the embodiments and variations are also included in the embodiments of the present invention, so long as they are consistent within the scope of the technical concept of the present invention.

In the following embodiments, later configurations and processes that are similar to previously described configurations and processes may be given the same reference symbols and explanations may be omitted, with only the differences being explained.

In the following embodiment, a bridge is used as an example of a structure to be diagnosed for the presence or absence of abnormalities such as cracks and other damage and deterioration. However, the present invention is not limited to bridges, and can be applied to any civil engineering or architectural structure that serves as social infrastructure, such as tunnels, road structures, river structures, port structures, water supply and sewerage systems, and buildings, to diagnose the presence or absence of abnormalities.

In the following description of the embodiments, an expression in which a first symbol is followed by a second symbol, such as "A^", is the same as an expression in which the second symbol is written directly above the first symbol.

<Mathematical Background of the Invention>
First, before describing the embodiments, the mathematical background of the present invention will be described. The present invention utilizes the fact that the system matrix of the state equation in the state space model of a structure can be approximated by a matrix of regression coefficients of an autoregressive model that expresses an observation signal at a certain time as a linear combination of a time series of observation signals before the certain time. The present invention focuses on the fact that the matrix of regression coefficients contains various information related to the state of the structure contained in the system matrix, and expresses the state of the structure using the matrix of regression coefficients, and evaluates the state of the structure by analyzing the matrix of regression coefficients.

In the following, the sensor information acquired by the sensor will be described as acceleration, but it is not limited to acceleration and may be other physical quantities.

Generally, the equation of motion of a structure is expressed as the following equation (1). Assume that n sensors (n is a natural number equal to or greater than 1) are installed at various positions on the structure, and a time series of n-th order observation vectors is obtained with the installation positions as observation points.

The equation of motion (1) above can be transformed into an equation of state as shown in the following equations (2-1) and (2-2).

Here, y in the above formula (2-2) represents the observation vector, z in the above formula (2-3) represents the state variable vector, and A _s in the above formula (2-5) represents the system matrix that represents the state of the system (structure to be diagnosed) in state space. C in the above formula (2-2) is a matrix that associates the observation vector y with the state of the system, and becomes a unit matrix when the measurement information at the observation point is regarded as the state of the system as it is.

Now, it is known that the state equations of the above equations (2-1) and (2-2) can be expressed as an autoregressive model, which is a linear combination of the time series of discretized observation signals, as shown in the following equation (3).

In the above formula (3), k is the time index, y(k) is the sensor information vector at time k, p is the order of the autoregressive model, _Ai is the matrix of regression coefficients, and e(k) is the error term of the autoregressive model at time k. The matrix _Ai is the regression coefficient multiplied by each of the time series of the linearly combined sensor information vector y(ki) in the autoregressive model.

Each element of the matrix A _i , which is a regression coefficient in the autoregressive model of the above formula (3), is associated with a physical quantity of the structure included in the system matrix A _s of the above formula (2-1), and includes at least one of information on the displacement, velocity, acceleration, mass matrix m, damping coefficient matrix c, and stiffness matrix k at the observation point where the sensor is installed. Therefore, by using the matrix A _i , it is possible to express the state of the structure so as to include various information related to vibration. In other words, it is possible to capture changes in the state of the structure by capturing changes in the matrix A _i .

However, when there is one sensor (n=1), the matrix A _i in the autoregressive model is a scalar.

Here, the matrix _Ai is an n-th order square matrix as shown in the following formula (4): n in the following formula (4) is the number of sensors installed at each position of the structure.

And when the pth-order autoregressive model shown in the above equation (3) generated from the time series of the sensor information vector y(k) measured by n sensors is expressed as a determinant, it becomes as shown in the following equation (5).

Here, _Yf on the left side of the above formula (5) represents a predicted state, _Yp in the first term on the right side represents the p-th past acceleration, and E in the second term on the right side represents an error caused by uncertainty in observing the state of the structure. Furthermore, the coefficient matrix A in the first term on the right side of the above formula (5) is an n×(n×p) matrix combining matrices A _i for i from 1 to p, and is expressed as in the following formula (6).

The coefficient matrix A inherits the information of the system matrix A _s . For example, the poles z _k , z _k ^* of the system at time k expressed using the natural frequency ω _k and damping coefficient h _k of the system at time k can be obtained from the above equation (5) as shown in the following equation (7).

The natural frequency ω _k and the damping coefficient h _k obtained by converting the posterior distribution of the coefficient matrix A into the posterior distribution of the poles and using the above formula (7) can also be used for diagnosing a structure. The natural frequency ω _k and the damping coefficient h _k obtained in this way have an advantage over the conventional technology in that vibration characteristics such as the natural frequency and the damping coefficient can be identified while taking into account variations. Furthermore, by converting the posterior distribution of the coefficient matrix A into the posterior distribution of the poles and obtaining vibration characteristics from poles with small variance, it is possible to automate the selection of physically meaningful vibration characteristics. Furthermore, it is possible to automate the selection of the model order p based on BIC (Bayesian information criterion).

By transmitting the coefficient matrix A obtained as described above to a diagnostic device as a feature representing the state of the structure, a time series having the vibration characteristics of the structure can be reproduced in the diagnostic device, making it possible to diagnose the state of the structure.

However, the coefficient matrix A contains components of poor quality information that are not related to the diagnosis result. By removing these components of poor quality information and sending the reconstructed matrix A^ to the diagnosis device, the amount of data sent and received can be reduced. This matrix A^ is called the principal component matrix because it is found by principal component analysis of the coefficient matrix A.

For example, the coefficient matrix A estimated at the reference time (e.g., when the structure is healthy) is subjected to singular value decomposition as shown in the following equation (8).

In the above formula (8), the matrices U and P ^T (where P ^T is the transpose of matrix P) are orthogonal matrices, and the matrix Λ is an eigenvalue of the coefficient matrix A. Then, the right-hand side of the above formula (8) is decomposed into an n×N matrix U ₁ (N<n) and an n×(np−N) matrix U ₂ by principal component analysis, as shown in the following formula (9).

This matrix _U1 at the reference time (e.g., healthy time) is a matrix that projects the coefficient matrix A in the probability space onto the principal component space. As shown in the following formula (10), ^the transposed matrix _U1T of the matrix _U1 is applied to the coefficient matrix A estimated at each time after the reference time to generate the principal component matrix A^, which is an N×np matrix from which components of poor quality information have been removed.

As described above, by estimating the coefficient matrix A that inherits information from the system matrix A _s , generating the principal component matrix A^ from the coefficient matrix A, and capturing changes in the principal component matrix A^, it is possible to detect abnormalities in the structure. In the present invention, as shown in the following formula (11), the probability distribution of the elements of the principal component matrix A^ is estimated by Bayesian estimation. Then, by using the estimated probability distribution to diagnose the condition of the structure, it is possible to detect abnormalities in the structure while taking into account the uncertainty included in the above formula (5).

In this invention, a probability distribution with a mean and variance is used as the evaluation index for evaluating the condition of a structure, which not only makes it possible to capture minute changes in the mean value of the evaluation index when the structure is damaged compared to when it is healthy, but also makes it possible to indicate the reliability of the evaluation index using the variance.

In addition, the condition of a structure can also be diagnosed by using the coefficient matrix A as a feature instead of the principal component matrix A^. In that case, the above equation (11) becomes an equation in which A^ is replaced with A.

<Embodiment 1>
Hereinafter, the first embodiment will be described with reference to FIGS.

(Configuration of diagnostic system S of embodiment 1)
1 is a diagram showing a schematic configuration of a diagnostic system S of the first embodiment. The diagnostic system S includes a sensor node 1 and a diagnostic device 2. The sensor node 1 generates a feature quantity indicating the state of the bridge 4 based on sensor information acquired via wireless communication (or wired communication) from a sensor 3 installed on the bridge 4. The diagnostic device 2 acquires the feature quantity indicating the state of the bridge 4 generated by the sensor node 1 via wireless communication (or wired communication), and diagnoses the presence or absence of an abnormality in the bridge 4 using the feature quantity.

Sensor 3 is, for example, an acceleration sensor, and is installed at each location of bridge 4. In this embodiment, sensor 3 is a plurality of sensors installed at multiple locations of bridge 4, but is not limited to this and may be a single sensor installed at one location. Also, in this embodiment, sensor 3 is an acceleration sensor that detects acceleration, but is not limited to this and may be a sensor that can measure other physical quantities (for example, strain, displacement, speed, etc.).

In this embodiment, for ease of understanding, as shown in FIG. 1, a configuration is described as an example in which the sensor node 1 processes sensor information acquired from the sensor 3 and transmits the results to the diagnostic device 2. However, the present invention is not limited to this, and may be configured as a distributed cooperative system in which the sensors 3 form a sensor network and perform distributed cooperative processing to achieve functions equivalent to those of the sensor node 1. When there is only one sensor 3, the single sensor 3 achieves functions equivalent to those of the sensor node 1.

(Configuration of the sensor node 1 of the first embodiment)
2 is a block diagram showing the configuration of the sensor node 1 of the embodiment 1. The sensor node 1 performs edge processing to generate a coefficient matrix A (or a principal component matrix A^) as a feature quantity representing the state of the bridge 4 from sensor information acquired from the sensor 3. The sensor node 1 has a processor 11, a memory 12, a storage 13, and a communication I/F (Inter/Face) unit 14.

The processor 11 is an arithmetic processing device such as a CPU (Central Processing Unit), a PLD (Programmable Logic Device), or a microprocessor. The memory 12 is a main storage device. The storage 13 is an auxiliary storage device. The communication I/F unit 14 is a communication interface that enables the sensor node 1 to perform wireless communication (or wired communication) with the sensor 3 and the diagnostic device 2.

The processor 11 executes a program in cooperation with the memory 12 to realize the functional units of a sensor information acquisition unit 111, an autoregressive model generation unit 112, a feature generation unit 113, and a feature transmission unit 114.

The sensor information acquisition unit 111 acquires, in time series, observation signals observed by the sensors 3 at each installation position on the bridge 4 via the communication I/F unit 14, and stores the signals in the storage 13 as sensor information 131. In this embodiment, the sensor information 131 acquired by the sensor information acquisition unit 111 is time-series information at continuous time, but may be time-series information at discrete time. The sensor information 131 may be acquired continuously, or may be acquired for a fixed period of time in response to an acquisition instruction.

The autoregressive model generation unit 112 discretizes the sensor information 131 stored in the storage 13 by time. Then, as shown in the above formula (3), the autoregressive model generation unit 112 generates an autoregressive model that expresses the sensor information 131 at a certain time k as a linear combination of the time series of the sensor information 131 at p times k-i (i = 1, ..., p) prior to the certain time k. Details of the processing by the autoregressive model generation unit 112 will be described later with reference to a flowchart.

The feature generation unit 113 generates a feature representing the state of the bridge 4 at a certain time k from the coefficient matrix A (formula (6) above) which combines the regression coefficients of the linear combination of the autoregressive models generated by the autoregressive model generation unit 112. The feature may be the coefficient matrix A itself, or a principal component matrix A^ (formula (10) above) which projects the coefficient matrix A onto a principal component space based on principal component analysis. As described below, based on such feature, the surface and internal state of the bridge 4 can be represented as a probability distribution. Details of the processing by the feature generation unit 113 will be described later with reference to a flowchart.

The feature transmission unit 114 transmits the features generated by the feature generation unit 113 to the diagnostic device 2 via the communication I/F unit 14.

(Configuration of diagnosis device 2 of embodiment 1)
3 is a block diagram showing the configuration of the diagnosis device 2 of embodiment 1. The diagnosis device 2 includes a processor 21, a memory 22, a storage 23, a communication I/F unit 24, and an output unit 25. The processor 21 cooperates with the memory 22 to execute a program, thereby realizing each of the functional units, that is, a feature receiving unit 211 and a diagnosis unit 212.

The processor 21 is an arithmetic processing device such as a CPU, PLD, or microprocessor. The memory 22 is a main storage device. The storage 23 is an auxiliary storage device. The communication I/F unit 24 is a communication interface that enables the diagnostic device 2 to perform wireless communication (or wired communication) with the sensor node 1. The output unit 25 is a monitor, display, etc., and outputs various information.

The feature receiving unit 211 receives a feature 231 indicating the state of the bridge 4 at each time k from the sensor node 1 via the communication I/F unit 24. The feature receiving unit 211 stores the received feature 231 in the storage 23.

The diagnosis unit 212 uses Bayesian estimation to obtain a probability distribution (the above formula (11)) of the feature at each time k from the feature 231 stored in the storage 23 as an evaluation index. The diagnosis unit 212 then calculates the Mahalanobis distance MD of the probability distribution obtained as the evaluation index at each time k, as shown in the following formula (12).

In the above formula (12), matrix X is coefficient matrix A (or principal component matrix A^), and matrix S is the covariance matrix of matrix X. The diagnosis unit 212 obtains in advance the Mahalanobis distance MD of coefficient matrix A (or principal component matrix A^) of bridge 4 at a reference time (e.g., when healthy) as a reference evaluation index 232 and stores it in storage 23. The diagnosis unit 212 also calculates the Mahalanobis distance MD of coefficient matrix A (or principal component matrix A^) of bridge 4 at the time of diagnosis when the presence or absence and degree of damage are unknown.

The diagnosis unit 212 then compares the Mahalanobis distance MD at the time of diagnosis with the reference evaluation index 232, and if a statistical distance such as the difference or ratio between them is equal to or greater than a certain value, diagnoses that an abnormality such as damage has occurred or progressed in the bridge 4 compared to the reference time. The diagnosis unit 212 outputs the diagnosis result to the output unit 25 such as a display.

In addition, the diagnosis unit 212 may diagnose abnormalities or the progression of deterioration of the bridge 4 by using Z values or other statistical indices instead of the Mahalanobis distance MD and determining the threshold value of the statistical distance between each statistical indices at the reference time and the diagnosis time.

(Reference time process in diagnostic system S of embodiment 1)
4 is a sequence diagram showing the reference time processing in the diagnosis system S of embodiment 1. In the reference time processing in the diagnosis system S, a coefficient matrix A and a principal component matrix A^ are generated from the time series of sensor information acquired from the sensor 3 at a reference time (for example, when the bridge 4 is healthy), and a process of calculating a reference evaluation index 232 based on the principal component matrix A^ is performed.

First, in step S101, the sensor information acquisition unit 111 of the sensor node 1 acquires the sensor information 131 from the sensor 3 and stores it in the storage 13. Next, in step S102, the autoregressive model generation unit 112 generates a p-th order autoregressive model of the sensor information vector y(k) at a certain time k obtained by discretizing the sensor information 131 acquired in step S101 based on the above formula (3). The order p of the autoregressive model is determined appropriately, but may be automatically determined from the above formula (5) based on the BIC, for example, as described above.

Next, in step S103, the feature generator 113 generates a coefficient matrix A by combining matrices _Ai representing regression coefficients of the p-th order autoregressive model of the sensor information vector y(k) as in the above formula (6). Next, in step S104, the feature generator 113 generates a principal component space matrix U 1 T that projects the coefficient matrix A in the probability space representing the state of the bridge 4 at the reference time onto the principal component space, based on the above formulas (8) to ( ₉ ⁾ .

Next, in step S106, the feature generating unit 113 applies the principal component space matrix U ₁ ^T generated in step S105 to the coefficient matrix A to generate a principal component matrix A^ as a feature, as shown in the above formula (10). Next, in step S106, the feature transmitting unit 114 transmits the feature generated in step S105 to the diagnostic device 2.

Next, in step S201, the feature receiving unit 211 of the diagnostic device 2 receives the feature 231 from the sensor node 1 and stores it in the storage 23. Next, in step S202, the diagnostic unit 212 obtains the probability distribution p(A^|Σ, Y) of the feature 231 as in the above formula (11), and calculates the Mahalanobis distance MD of the probability distribution p(A^|Σ, Y) as the reference evaluation index 232. Next, in step S203, the diagnostic unit 212 stores the reference evaluation index 232 calculated in step S202 in the storage 23.

Note that, not limited to the illustration in Figure 4, in the reference time processing, the steps executed by the sensor node 1 may be executed by the diagnostic device 2, and the steps executed by the diagnostic device 2 may be executed by the sensor node 1. In other words, there is no limitation as to whether the sensor node 1 or the diagnostic device 2 is the executing entity of each step shown in Figure 4.

For example, the processing from any of steps S102 to S105 onwards may be performed by the diagnostic device 2 rather than the sensor node 1. Also, for example, the processing of step S202 may be performed by the sensor node 1 rather than the diagnostic device 2, and the calculated reference evaluation index is transmitted to the diagnostic device 2 and used by the diagnostic device 2 to diagnose the condition of the bridge 4.

(Diagnosis process in the diagnosis system S of the first embodiment)
Fig. 5 is a sequence diagram showing a diagnosis process in the diagnosis system S of the embodiment 1. The diagnosis process in Fig. 5 differs from the reference time process in Fig. 4 only in that it handles sensor information and feature amounts at the time of diagnosis rather than at the reference time, and steps S111 to S116 are the same as steps S101 to S106 in Fig. 4.

If the principal component space matrix U ₁ ^T is generated in step S104 of the reference-time processing in Fig. 4, the step of generating the principal component space matrix U ₁ ^T in step S114 can be omitted in the diagnosis-time processing in Fig. 5. If step S114 is omitted, in step S115 following step S113, the feature generator 113 applies the principal component space matrix U ₁ ^T generated in step S104 of the reference-time processing in Fig. 4 to the coefficient matrix A generated in step S113 to generate the principal component matrix A^ as a feature. Next, in step S116, the feature transmitter 114 transmits the feature generated in step S115 to the diagnosis device 2.

Next, in step S211, the feature receiving unit 211 of the diagnostic device 2 receives the feature from the sensor node 1. Next, in step S212, the diagnostic unit 212 calculates the probability distribution p(A^|Σ, Y) of the feature based on the feature received in step S211 using Bayesian estimation as in the above formula (11). The diagnostic unit 212 then calculates the Mahalanobis distance MD at the time of diagnosis as an evaluation index based on the above formula (12), compares the evaluation index with the reference evaluation index 232, and diagnoses that an abnormality such as damage has occurred or progressed in the bridge 4 compared to the reference time if the difference or ratio between them is equal to or greater than a certain value.

As with FIG. 4, FIG. 5 does not limit whether each step is executed by the sensor node 1 or the diagnostic device 2. For example, the evaluation index may be calculated by the sensor node 1 instead of the diagnostic device 2. The evaluation index calculated in this manner is transmitted to the diagnostic device 2, where it is used to diagnose the condition of the bridge.

4 and 5, the principal component matrix A^ is used as the feature, but the coefficient matrix A may be used as the feature. In this case, A^ is replaced with A in the above formula (11) to calculate the probability distribution p(A|Σ, Y).

Figure 6 is an explanatory diagram of the diagnosis processing in the diagnosis system S of embodiment 1. The illustrated portion 601 in Figure 6 shows an overview of the reference time processing (Figure 4) in which feature amounts are extracted from observation data representing the vibration of the bridge 4 at the reference time, and a statistical index based on this feature amount is calculated as a reference evaluation index. The illustrated portion 602 shows an overview of the diagnosis time processing (Figure 5) in which feature amounts are extracted from observation data representing the vibration of the bridge 4 at the diagnosis time, and a statistical index based on this feature amount is calculated as a diagnosis time evaluation index.

Then, in the diagnosis process (Fig. 5), the presence or absence of an abnormality in the bridge 4 and its progression are judged based on the comparison result between the reference evaluation index and the diagnosis evaluation index. That is, as shown in the illustrated portion 603 of Fig. 6, in a probability space, an abnormality in the bridge 4 is diagnosed by comparing a reference evaluation index based on the probability distribution of the reference feature amount with a diagnosis evaluation index based on the probability distribution of the diagnosis feature amount. In the comparison between the reference evaluation index and the diagnosis evaluation index, the degree to which the diagnosis evaluation index deviates from the reference evaluation index is evaluated using a statistical index such as the Mahalanobis distance. If the statistical distance between the diagnosis evaluation index and the reference evaluation index is equal to or greater than a certain value, the presence or absence of an abnormality in the bridge 4 can be judged to be that the state of the bridge 4 at the time of diagnosis has changed from the reference time.

In this embodiment, the presence or absence of an abnormality in bridge 4 is determined based on the statistical distance between the evaluation indexes at the reference time and the diagnosis time, but this is not limited to the above. For example, the coefficient matrix A (or the principal component matrix A^) at each time is clustered, and if the coefficient matrix A based on new observation data is not classified into an existing cluster but into a new cluster, it is determined that the pattern of the features of coefficient matrix A has changed, and an abnormality has occurred in bridge 4.

(Effects of the First Embodiment)
In the above-mentioned first embodiment, the condition evaluation based on the feature quantity representing the state of the structure is performed using an evaluation index based on a probability distribution that includes various physical quantities of the structure and that can take into account the estimation error of the measurement value while absorbing the difference in sensitivity of the change in the measurement value to damage that differs depending on the measurement location. Therefore, according to the first embodiment, it is not necessary to set a vibration mode that is sensitive to damage, which requires a high level of specialized knowledge, and the state of the structure can be evaluated at low cost and with high accuracy without performing numerical analysis by using a probability distribution having a mean and variance based on the feature quantity that is sensitive to damage of the structure. In addition, the validity of the evaluation can be expressed by the variance.

In addition, the probability distribution evaluation index used in embodiment 1 captures minute changes that progress slowly as a result of damage to a structure and that are easily obscured by noise during measurement and forced vibrations such as live loads, making it possible to perform highly accurate condition diagnosis of the structure.

In the first embodiment, the coefficient matrix A representing the state of the structure is projected into the principal component space to remove information of low quality, generating a principal component matrix A^ of lower order. Therefore, the feature quantities representing the state of the structure are compressed into feature quantities capable of reproducing the state of the structure without reducing the quality of the information, so that the feature quantities can be transferred by edge processing using low-cost advanced technology such as low-power wide area radio, and the sensors and transfer system can be constructed inexpensively, which is expected to reduce management costs significantly.

More specifically, in the past, it was necessary to use acceleration data for several minutes to evaluate such vibration conditions, which required a large amount of communication. However, in this method, the only information required is the value of the principal component matrix A^, so the principal component matrix A^ is calculated by edge processing in the measuring device, and by making the data small, it is possible to use inexpensive advanced technologies such as specific low-power radio, and it is considered to have an advantage in terms of management costs as shown in Figure 7. Figure 7 is a diagram showing a comparison of the data amount of the coefficient matrix A and the principal component matrix A^ in the first embodiment and the conventional technology for each number of sensors. Figure 7 shows the case of transmitting data measured for 60 seconds with a period of 5 ms. According to Figure 7, it can be seen that the amount of data transferred from the sensor node 1 can be significantly reduced by the principal component matrix A^ in any number of sensors, whether 8, 4, or 1.

<Embodiment 2>
In the first embodiment, abnormality diagnosis of the bridge 4 is performed using a statistical index such as the Mahalanobis distance of a probability distribution based on the feature amounts of the bridge 4 at the reference time and the diagnosis time, whereas in the second embodiment, abnormality diagnosis of the bridge 4 is performed using, as an evaluation index, a Bayes factor based on the feature amounts of the bridge 4. Note that the configuration of the diagnosis system S in the second embodiment is similar to the configuration of the diagnosis system in the first embodiment, except that a part of the processing of the diagnosis unit 212 of the diagnosis device 2 is different.

(Diagnosis process in diagnosis system S of embodiment 2)
Fig. 8 is a flowchart showing diagnostic processing in the diagnostic system S of the second embodiment. The flowchart shown in Fig. 8 is another example of the processing of step S212 of the diagnostic processing of the first embodiment shown in Fig. 5. That is, in the second embodiment, the diagnostic device 2 executes the diagnostic processing shown in Fig. 8 following step S211. Note that in the second embodiment, the diagnostic device 2 can execute the diagnostic processing of Fig. 5 without executing the reference time processing of Fig. 4.

First, in step S2121, the diagnosis unit 212 of the diagnosis device 2 calculates a Bayes factor B defined by the following formula (13) as an evaluation index based on the coefficient matrix A (or the principal component matrix A^), which is the feature amount received in step S211. When distinguishing it from a local Bayes factor ^Bj described later, the Bayes factor B is called a global Bayes factor.

In the above formula (13), H ₀ is the null hypothesis (in this embodiment, the sound state of the bridge 4), H ₁ is the alternative hypothesis (in this embodiment, the abnormal/damaged state of the bridge 4), Y _t is the feature to be tested for the hypothesis (in this embodiment, the coefficient matrix A or the principal component matrix A^), and Φ _t is a parameter under the hypothesis of H ₀ or H ₁ (for example, the average value or standard deviation of the probability distribution of the feature).

In the Bayes factor B shown in the above formula (13), the denominator p( _Yt | _Φt , _H0 ) indicates the marginal likelihood that the feature _Yt at the time of diagnosis is the same feature as the reference state (healthy state) of the bridge 4, and the numerator p( _Yt | _Φt , _H1 ) indicates the marginal likelihood that the feature _Yt at the time of diagnosis is different from the reference state of the bridge 4. Therefore, when the Bayes factor B, which is the ratio of p( _Yt | _Φt , _H1 ) to p( _Yt | _Φt , _H0 ), exceeds a threshold value, it can be determined that an abnormality or damage has occurred in the bridge 4 based on a certain feature.

Next, in step S2122, the diagnosis unit 212 determines whether the Bayes factor B calculated in step S2121 is greater than a threshold value. This threshold value is set based on past inspection data, etc., and is the category that an inspection engineer for bridge 4 has determined to be abnormal. If the Bayes factor B is greater than the threshold value (step S2122 Yes), the diagnosis unit 212 determines that there is an abnormality in bridge 4 (step S2123), and if the Bayes factor B is equal to or less than the threshold value (step S2122 No), the diagnosis unit 212 determines that bridge 4 is normal (step S2124).

The diagnostic process shown in Figure 8 may be performed by the sensor node 1 rather than by the diagnostic device 2.

(Method of calculating threshold for abnormality determination)
Here, a method for calculating the threshold value for abnormality determination used in step S2122 of the diagnosis process (FIG. 8) will be described. FIG. 9A is a diagram for explaining a method for calculating the threshold value for abnormality determination in the second embodiment. In the graph (a) on the left side of FIG. 9A, the horizontal axis represents time, the vertical axis represents the logarithm of the Bayes Factor B, and the values of the Bayes Factor B at each time are plotted. In the graph (b) on the right side of FIG. 9A, the horizontal axis represents frequency, the vertical axis represents the logarithm of the Bayes Factor B, and the frequency distribution of each value of the Bayes Factor B is shown. The calculation of the threshold value for abnormality determination may be performed by the diagnosis device 2 or another information processing device.

On actual bridges, various situations occur that affect anomaly detection, such as the passing of a truck carrying a heavy load or braking on a bridge. Bayes Factor B is a value with variation that includes specific data under various situations. When calculating the threshold value for Bayes Factor B, by including specific data from multiple measurement data over a specified period of time in the past as the basis for calculating the threshold value rather than eliminating such data, it is possible to calculate a threshold value that takes into account the various situations that occur on actual bridges.

Among the past measurement data of the bridge 4, the measurement data for a predetermined period (for example, one year) as a reference (see graph (a) in FIG. 9A) is applied to the above formula (13) to create a distribution of the values of the Bayes factor B (see graph (b) in FIG. 9A). Based on the average value μ and variance σ of the distribution of the values of the Bayes factor B, for example, any value such as μ, μ + σ, μ + 2σ, etc. can be determined as the threshold. In graph (b) in FIG. 9A, the values of the Bayes factor B are normalized, so that the distribution is a normal distribution with an average μ = 0, and the threshold value is set to the average μ. The threshold value of the Bayes factor B can be quantitatively determined using the average μ and variance σ based on index values such as the characteristics of the bridge 4, the conditions of use, the geographical conditions, and the weather conditions. In this way, the accuracy of abnormality determination can be improved by using a threshold value that takes into account various situations that occur in an actual bridge.

(Another example of the threshold for abnormality diagnosis)
Another example of the threshold value for abnormality diagnosis will be described. Fig. 9B is a diagram for describing another example of the abnormality determination method using a plurality of threshold values according to the second embodiment. In Fig. 9B (graph a), the horizontal axis represents frequency distribution, and the vertical axis represents the logarithm of Bayes Factor B, and the frequency distribution of each value of Bayes Factor B is shown. In Fig. 9B (graph b), the horizontal axis represents days, the vertical axis represents the logarithm of Bayes Factor B, and each value of Bayes Factor B for each day is plotted.

The value of Bayes factor B is determined to be abnormal when it exceeds a threshold value. For this reason, in the distribution of values of Bayes factor B based on the measurement data described above with reference to FIG. 9A, multiple threshold values can be defined using the mean μ and standard deviation σ for values greater than the mean μ, and abnormality can be determined in stages. Based on multiple threshold values, ranges representing diagnostic levels, such as a "normal" range, a "caution required" range, and an "abnormal" range, may be defined.

For example, multiple values based on the mean μ and variance σ of the distribution of the Bayes Factor B values (such as μ, μ+σ, μ+2σ in (graph a) of FIG. 9B) are determined as multiple stepped thresholds for abnormality diagnosis of the bridge 4 based on the Bayes Factor B. As shown in (graph a) of FIG. 14, if the Bayes Factor B is B≦μ+σ, it is diagnosed as "normal", if μ+σ<B≦μ+2σ, it is diagnosed as "needs attention", and if μ+2σ<B, it is diagnosed as "abnormal". In the example of FIG. 9B, the number of stepped thresholds for abnormality diagnosis is "2", but is not limited to this. The stepped thresholds and the number of the Bayes Factor B can be quantitatively determined using the mean μ and variance σ based on index values such as the characteristics of the bridge 4, the conditions of use, the geographical conditions, and the weather conditions.

In this way, the accuracy of anomaly detection can be improved by using multiple thresholds that take into account various situations that occur on actual bridges. Furthermore, by gradually determining the thresholds for diagnosing anomalies in the bridge 4 based on the Bayes Factor B, the manager of the bridge 4 can gradually detect anomalies in the bridge 4 that progress gradually over time, and can perform maintenance on the bridge 4 in a planned manner.

(Local Bayes Factor B)
Now, the Bayes factor B based on the above formula (13) is an evaluation index for detecting abnormalities when viewing the bridge 4 as a whole. On the other hand, it is also possible to calculate individual Bayes factors based on feature quantities obtained from the time series of sensor information from each sensor 3 installed on the bridge 4. An individual Bayes factor is called a local Bayes factor. For example, a local Bayes factor B j based on the feature quantities of the time series of sensor information from the j-th sensor 3 (j=1, ..., n) out of the multiple sensors ³ installed on the bridge 4 is defined by the following formula (14).

By performing the diagnostic process shown in FIG. 8 with the local Bayes factor ^Bj replaced with the Bayes factor B, it is possible to identify, from the index j of the local Bayes factor ^Bj indicating an abnormality, that an abnormality or damage has occurred in the bridge 4 at the location where the j-th sensor 3 is installed.

The local Bayes factor ^Bj can also be used as follows. The type of sensor information that can sensitively detect anomalies using the Bayes factor may differ depending on the structure such as the bridge 4 and the type of anomaly or damage that occurs in the structure. Therefore, sensors 3 of multiple sensor types are installed on the structure such as the bridge 4, and diagnosis is performed based on the local Bayes factor ^Bj based on the feature amount generated for each sensor type.

That is, sensors 3 that detect different physical quantities (different sensor types) are installed on the bridge 4, and a local Bayes factor ^Bj is calculated based on the time-series feature amount of the sensor information for each sensor type. When an abnormality is determined based on the local Bayes factor ^Bj of the jth sensor type, it can be identified from the index j of the local Bayes factor ^Bj that an abnormality or damage has occurred based on the sensor information of the sensor 3 of the jth sensor type. By using the local Bayes factor ^Bj in this way, it is possible to detect an abnormality using any one of multiple physical quantities, thereby improving the sensitivity of anomaly detection.

(Abnormality determination in bridge diagnosis system S of embodiment 2)
Fig. 10 is a diagram for explaining the results of an experiment on abnormality determination performed using the diagnosis system S of the second embodiment. Fig. 10 shows the time progression of the binary logarithm of ten types of local Bayes factors ^Bj , A1 to A10, with the horizontal axis representing time and the vertical axis representing the binary logarithm axis of the local Bayes factors ^Bj . In Fig. 10, A1 to A10 represent installation positions of the sensor 3 as an example. INT (initial stage), DMG1, and DMG2 in Fig. 10 represent each period.

In the INT, since no damage is done to the bridge 4, the local Bayes factor ^Bj based on the sensor information of the sensor 3 at any installation position is approximately 0 and does not exceed the threshold. Conversely, if the local Bayes factor ^Bj is a value like the INT shown in Fig. 10, it can be estimated that the bridge 4 is in a sound state with no abnormalities or damage occurring.

In DMG1, which caused damage to a portion of bridge 4 near A6 following INT, the value of the local Bayes factor based on the sensor information of sensor 3, particularly A6 among A1 to A10, exceeds the threshold and is larger than INT. Conversely, if the local Bayes factor ^Bj is a value like DMG1 compared to INT, it can be estimated that damage has occurred in the portion of bridge 4 near A6.

Furthermore, in DMG2, which follows DMG1 and further inflicts damage on a portion of bridge 4 near A6, the local Bayes factor ^Bj based on the sensor information of sensor 3 at A6 is even larger than that of DMG1. Conversely, if the local Bayes factor ^Bj has a value similar to that of DMG2 compared to DMG1, it can be estimated that the damage that occurred in the portion of bridge 4 near A6 has progressed further. Also, by comparing the sum or average of the local Bayes factors ^Bj of DMG1 with the sum or average of the local Bayes factors ^Bj of DMG2, the degree of progression of damage from DMG1 to DMG2 can be evaluated.

An evaluation similar to that shown in Figure 10 can also be performed using the global Bayes factor.

By using the Bayes factor B or the local Bayes factor ^Bj in this manner, it is possible to detect when the bridge 4 has changed from a healthy state to an abnormal state, and to evaluate the progression of damage that has already been identified through inspection or the like.

(Experimental Results of the Second Embodiment)
Fig. 11 is a diagram showing the execution conditions of an experiment performed using the diagnostic system S of the embodiment 2. Fig. 12 is a diagram showing the results of an experiment performed using the diagnostic system S of the embodiment 2.

Figure 11 shows the load pattern applied to the bridge over time, with the horizontal axis representing time and the vertical axis representing load. As shown in Figure 11, vibration was applied to the bridge at times t0-t1 (Stage 1), t2-t3 (Stage 2), t4-t5 (Stage 3), t6-t7 (Stage 4), t7-t8 (Stage 5), t9-t10 (Stage 6), t10-t11 (Stage 7), and t12- (Stage 8). A crack load was applied to the bridge at times t1-t2 (Loading 1). A yield load was applied to the bridge at times t3-t4 (Loading 2). Times t5-t6 (Loading 3) are the duration of the load on the bridge. At times t8 to t9 (Loading 4), the design load was applied to the bridge. At times t11 to t12 (Loading 5), the maximum load was applied to the bridge.

As a result of conducting experiments under these execution conditions, it can be said that the Bayes factor generally tends to increase as the loading stage progresses, as shown in Figure 12. Therefore, a clear interpretation can be given to the observed results of changes in the Bayes factor, making it easy to detect abnormalities or damage that have occurred in the bridge.

The advantage of this embodiment, that abnormalities or damage occurring in a bridge can be easily detected regardless of an increase or decrease in the bridge's natural frequency, can be seen from Figure 13. Figure 13 shows, as a comparative example, the change in the natural frequency of a bridge that differs depending on the vibration mode measured under the same conditions as in the experiment conducted using the diagnosis system S of the second embodiment shown in Figure 12.

Graph 1101 in Figure 13 shows the change in the natural frequency of the primary bending mode from Stage 1 to Stage 8. Graph 1102 in Figure 13 shows the change in the natural frequency of the secondary bending mode from Stage 1 to Stage 8. As shown in graphs 1101 and 1102, the natural frequency increases or decreases depending on the change in the condition of the bridge. Theoretically, the natural frequency decreases due to a decrease in rigidity caused by damage, but the natural frequency may increase if, for example, there is a change in the condition of the bridge's support parts. Also, as shown in graphs 1101 and 1102, the change in the natural frequency differs depending on the mode.

In this way, abnormality diagnosis based on natural frequency makes it difficult to set an appropriate vibration mode and determine the presence or absence of damage and its location, because the manner in which the natural frequency changes varies depending on the location of the damage and the vibration mode. In this regard, abnormality diagnosis using the Bayes factor according to this embodiment does not require the setting of a vibration mode, and can quantitatively detect abnormalities and damage that have occurred in the bridge.

(Temperature and Bayes factor values over time)
Fig. 14 is a diagram showing the change over time in the outside air temperature and the Bayes factor value for a certain period of time for a certain girder of bridge 4. Graph a in Fig. 14 is a graph showing the change over time in the outside air temperature for a certain period of time, and graph b is a graph showing the change over time in the Bayes factor B of the outside air temperature. The solid line in graph b in Fig. 14 indicates the threshold for determining an anomaly in the Bayes factor B, and if this threshold is exceeded, it is determined to be an anomaly.

Figure 14 shows an example of data on changes in outside air temperature of the girders of bridge 4, where it has been confirmed that there are no abnormalities in the structure itself. The outside air temperature fluctuates on an annual cycle, but when this embodiment is applied to the outside air temperature, the abnormality determination determines that the change in outside air temperature during the period for which an abnormality is to be determined is not abnormal. In other words, Figure 14 shows that this embodiment makes it possible to determine an abnormality in a structure by taking into account the presence or absence of the influence of periodic fluctuations that occur in nature, such as temperature changes.

(Changes in the deflection and Bayes factor values of Bridge 4 over time)
Figure 15 shows the change over time in the deflection and Bayes Factor values of a bridge over a certain period of time. Deflection is detected using a connecting pipe. In Figure 15, one year from April to March of the following year is taken as one fiscal year, and (graph a) shows the change over time in deflection over a certain period, while (graph b) shows the change over time in the Bayes Factor B of deflection. The solid line in (graph b) of Figure 15 indicates the threshold for determining an abnormality in Bayes Factor B, and if this threshold is exceeded, it is determined to be abnormal. The threshold for determining an abnormality is determined based on measurement data for one year in the base fiscal year.

Although the temperature fluctuates throughout the year, as shown in (graph a) of Figure 15, the results of the judgment made by this embodiment show that an abnormality is judged to exist during periods when the amount of deflection is large (for example, from September of the fourth year to January of the following year). It can also be seen that an abnormality is clearly judged to exist from the fifth year onwards.

(Comparison of changes in Bayes factor values over time for two spans of a bridge)
Figure 16 is a diagram showing a comparison of the change over time in the Bayes factor values of two spans of a bridge. In Figure 16, the fiscal year is one year from May to April of the following year. Bridge 4 whose data is shown in Figure 16 is the same as Bridge 4 whose data is shown in Figure 15.

In graphs a and b of Figure 16, the results of abnormality determination are calculated based on acceleration measurement results at two different points on the same bridge 4 (the damaged first span and the relatively less damaged second span). In Figure 16, thresholds are set based on one year's worth of measurement data in the base year. However, the base year, first year, second year, and third year in Figure 16 are different from those in Figure 15, so they are distinguished by notating them as, for example, the first year in Figure 15 and the first year in Figure 16.

In the abnormality judgment according to this embodiment, an abnormality in the damaged first span was detected in September of the first year in Figure 16, and deterioration over time was detected thereafter. On the other hand, the second span, which was relatively less damaged, was judged to be normal except for temporary abnormal values, even in the year in which an abnormality was detected in the first span in Figure 16. This result for the first span is the same as the period when the fluctuation in the amount of deflection became large as shown in Figure 15 (from the fourth and fifth years in Figure 15), so it is believed that appropriate abnormal values can be judged.

(Modification of the second embodiment)
In this embodiment, the diagnostic device 2 uses the coefficient matrix A or the principal component matrix A^ of the bridge 4 as a feature representing the state of the bridge 4, and performs abnormality diagnosis of the bridge using a Bayes factor B or a local Bayes factor ^Bj based on the feature. However, without being limited to this, the diagnostic device 2 may extract a feature from time-series data obtained by continuously measuring one or more physical quantities of each part of the structure, and perform abnormality diagnosis of the structure using a Bayes factor based on this feature. The physical quantity may be displacement, speed, acceleration, external force, strain, temperature, etc. For example, abnormality diagnosis of the structure may be performed using a Bayes factor based on a feature extracted from data for a predetermined period of time-series data obtained by fixed-point observation of the temperature of the bridge 4.

The diagnosis unit may also correct the Bayes factor using the results of an on-site inspection of the bridge 4 by an inspection worker, and diagnose the condition of the bridge 4 based on the corrected Bayes factor. For example, if an abnormality is determined based on the Bayes factor of a certain feature, but no abnormality is found as a result of an on-site inspection, the Bayes factor is corrected so that the same feature is not subsequently determined to be abnormal. Alternatively, if an abnormality is found as a result of an on-site inspection, the threshold may be set or corrected by referring to the Bayes factor based on a sensor installed near the location where the abnormality was detected.

(Effects of the Second Embodiment)
In the second embodiment, the condition of the structure is diagnosed based on a Bayes factor, which is the ratio of the probability of a healthy state to an abnormal state calculated from the time series of observation data of the structure in each time domain, and a threshold value set based on damage states that inspection engineers with advanced expertise have previously determined to be abnormal. Therefore, even a person without advanced expertise can perform a highly accurate evaluation of the health of the structure using quantitative evaluation indices without the need for costly processing such as numerical analysis, in the same way as an inspection engineer with advanced expertise.

<Embodiment 3>
In the third embodiment, in the first or second embodiment, the coefficient matrix A, which is a feature representing the state of the bridge 4, is sequentially learned using Bayesian estimation, thereby improving the accuracy of detecting an anomaly in the bridge 4. The sequential learning of the coefficient matrix A may be performed by the sensor node 1 (e.g., the feature generating unit 113) or the diagnosis device 2 (e.g., the diagnosis unit 212). Note that the same effect can be obtained by performing the same processing using the principal component matrix A^ instead of the coefficient matrix A.

FIG. 17 is an explanatory diagram of sequential updating of the coefficient matrix A in the diagnostic system S of the third embodiment. In FIG. 17, the acceleration detected by the sensor 3 is shown on the vertical axis, and the time is shown on the horizontal axis. The posterior probability p(A,Σ|Y) _i of the coefficient matrix A at time i shown in FIG. 17 is obtained using Bayesian estimation as shown in the following formula (15-1). However, the prior probability p(A,Σ) _i at time i is the posterior probability p(A,Σ|Y) _i-1 at time (i-1) as shown in the following formula (15-2). By replacing i with i+1 in the following formulas (15-1) and (15-2), the posterior probability p(A,Σ|Y) _i+1 of the coefficient matrix A at time (i+1) can be obtained in the same manner.

(Effects of the Third Embodiment)
FIG. 18 is a diagram showing the convergence of the probability distribution of the coefficient matrix by sequentially updating the coefficient matrix in the diagnosis system of the third embodiment. In FIG. 18, the horizontal axis represents the values that the elements of the feature (for example, the coefficient matrix A) can take, and the vertical axis represents the posterior probability at each value of the element. In the third embodiment, by sequentially learning the probability distribution of the coefficient matrix A, as shown in FIG. 18, the average value of the probability distribution approaches the true value and converges in a direction in which the variance becomes smaller (the probability distribution shown in FIG. 18 converges from the thin line to the dashed line and then to the thick line), so that the coefficient matrix A represents the state of the structure with higher accuracy. Then, the diagnosis unit 212 diagnoses the state of the structure using the coefficient matrix A or the principal component matrix A^ after the sequential learning as a feature. Therefore, it is possible to improve the accuracy of the abnormality estimation of the state of the structure. Since the coefficient matrix A is a feature including various features of the vibration data of the structure, it is possible to capture a more minute change in the state of the structure and perform an abnormality determination than the conventional technology that performs an abnormality determination using the natural frequency or the like.

Other Embodiments
In the above-described embodiment, the diagnostic system S has been described as having the sensor node 1 and the diagnostic device 2, but the embodiment is not limited to this. For example, the diagnostic system S may be configured without the diagnostic device 2. In such a case, the diagnostic system S is configured to include the sensor node 1 having an acquisition unit, an autoregressive model generation unit, a feature generation unit, and a diagnosis unit. That is, the sensor node 1 has an acquisition unit that acquires sensor information in a time series from a sensor attached to the structure, an autoregressive model generation unit that generates an autoregressive model that expresses the sensor information acquired by the acquisition unit at a certain time as a linear combination of the time series of sensor information acquired before the certain time, a feature generation unit that generates features indicative of the state of the structure at a certain time based on the regression coefficients of the autoregressive model, and a diagnosis unit that calculates an evaluation index for the features indicative of the state of the structure generated from the time series of information about the structure, which is the ratio between a first marginal likelihood that represents the probability distribution of the feature when it is assumed that the structure is in a healthy state and a second marginal likelihood that represents the probability distribution of the feature when it is assumed that the structure is in an unhealthy state, and diagnoses the state of the structure based on the evaluation index.

In the above embodiment, a case has been described in which a feature indicating the state of a structure at a certain time is generated based on the regression coefficients of an autoregressive model, but the embodiment is not limited to this. For example, a feature indicating the state of a structure may be generated from something other than the regression coefficients of an autoregressive model.

(Example of sensor node 1 implementation)
19 is a block diagram showing the configuration of a sensor node terminal 1B used as the sensor node 1 of the embodiment. The sensor node terminal 1B implemented as the sensor node 1 further includes an acceleration sensor 15, in comparison with the configuration of the sensor node 1 (FIG. 2). Such a sensor node terminal 1B exerts the same functions as the sensor node 1 and the sensor 3 by executing a predetermined application. That is, the sensor node terminal 1B is installed on the bridge 4 in the same way as the sensor 3, generates feature quantities of the bridge 4 from the acquired acceleration data, and transmits them to the diagnostic device 2.

The diagnostic device 2 is constructed, for example, on a cloud server, and provides diagnostic results based on the features as a cloud service. The diagnostic device 2 transmits the diagnostic results to the sensor node terminal 1B or other terminal device. The user checks the diagnostic results by viewing the screen output of the sensor node terminal 1B or other terminal device.

A plurality of sensor node terminals 1B may also be implemented as a plurality of sensor nodes 1 and installed at a plurality of locations on the bridge 4 in the same manner as the sensors 3. In this case, a feature of the bridge 4 is generated from the acceleration data acquired by all of the plurality of sensor node terminals 1B by independent processing of one sensor node terminal 1B representing the plurality of sensor node terminals 1B or by cooperative processing of a predetermined number of sensor nodes 1, and transmitted to the diagnostic device 2.

The present invention is not limited to the above-described embodiments, and the configuration of each embodiment can be added to, deleted, replaced, integrated, or distributed. Furthermore, the configurations and processes shown in the embodiments can be appropriately distributed, integrated, or replaced based on the efficiency of processing or implementation. A program for executing each process of the diagnostic system described in the above-described embodiments is installed in one or more computers via a recording medium or transmission medium, or is provided as an embedded program.

S: diagnostic system, 1: sensor node, 2: diagnostic device, 3: sensor, 4: bridge, 111: sensor information acquisition unit, 112: autoregressive model generation unit, 113: feature generation unit, 114: feature transmission unit, 131: sensor information, 211: feature reception unit, 212: diagnostic unit, 231: feature, 232: reference evaluation index

Claims (12)

a diagnosis unit that calculates an evaluation index, which is a ratio between a first marginal likelihood representing the likelihood of a feature quantity indicating a state of a structure generated from a time series of information about the structure, when it is assumed that the structure is in a healthy state, and a second marginal likelihood representing the likelihood of the feature quantity when it is assumed that the structure is not in the healthy state, and diagnoses the state of the structure based on the evaluation index;
The diagnosis unit includes:
Estimating the first marginal likelihood and the second marginal likelihood from the feature quantity using Bayesian estimation;
a Bayes factor calculated as the evaluation index by dividing the second marginal likelihood by the first marginal likelihood. The time series of information is obtained for each position in the structure,
The structure diagnosis system according to claim 1, characterized in that the diagnosis unit calculates the Bayes factor for each position based on the time series for each position in the structure, and diagnoses the condition of each position of the structure based on the Bayes factor for each position. The time series of the information is obtained for each of a plurality of types of the information,
The structure diagnosis system according to claim 1 , wherein the diagnosis unit calculates the Bayes factor for each type based on the time series for each type, and diagnoses the condition of the structure based on the Bayes factor for each type. an acquisition unit that acquires sensor information in time series from a sensor attached to the structure;
an autoregressive model generation unit that generates an autoregressive model that expresses the sensor information acquired by the acquisition unit at a certain time as a linear combination of a time series of the sensor information acquired before the certain time;
The structure diagnosis system according to any one of claims 1 to 3, further comprising: a feature generation unit that generates the feature indicating the state of the structure at the certain time based on the regression coefficients of the autoregressive model. The structural diagnostic system includes:
5. The structure diagnosis system according to claim 4, further comprising a sensor node including the acquisition unit, the autoregressive model generation unit, the feature generation unit, and a transmission unit that transmits the feature generated by the feature generation unit to the diagnosis unit. The feature generating unit is
The structural diagnosis system according to claim 4 or 5, characterized in that the features are generated by sequential learning, which repeats a process of setting a posterior distribution obtained by Bayesian estimation, in which a probability distribution of the previously generated features is used as a prior distribution, as the probability distribution of the currently generated features. The structure diagnosis system according to any one of claims 1 to 6 , characterized in that the feature includes at least one of information on the displacement, velocity, acceleration, mass, stiffness, and damping coefficient of the structure at the observation point of the feature. The structure diagnosis system of any one of claims 1 to 7, characterized in that the diagnosis unit corrects the evaluation index using inspection results from a field inspection of the structure, and diagnoses the condition of the structure based on the corrected evaluation index. The diagnosis unit diagnoses a state of the structure based on a comparison between the evaluation index and a threshold value;
The structure diagnosis system according to any one of claims 1 to 8, characterized in that the threshold value is determined based on the average and variance of multiple measurement data of the structure over a specified period of time in the past. A plurality of the thresholds are determined based on the average and the variance;
The structure diagnosis system according to claim 9 , wherein the diagnosis unit diagnoses the state of the structure based on a comparison between the evaluation index and a range based on a plurality of the threshold values. A structure diagnosis method performed by a structure diagnosis system that diagnoses a state of a structure, comprising:
calculating an evaluation index which is a ratio between a first marginal likelihood representing a likelihood of the feature quantity in a case where it is assumed that the structure is in a healthy state and a second marginal likelihood representing a likelihood of the feature quantity in a case where it is assumed that the structure is not in the healthy state, for the feature quantity which is generated from a time series of information about the structure and which indicates the state of the structure;
diagnosing a state of the structure based on the evaluation index;
The structural diagnostic system includes:
Estimating the first marginal likelihood and the second marginal likelihood from the feature quantity using Bayesian estimation;
a Bayes factor obtained by dividing the second marginal likelihood by the first marginal likelihood is calculated as the evaluation index. Computer,
1. A structure diagnosis program for functioning as a diagnosis unit that calculates an evaluation index, which is a ratio between a first marginal likelihood representing a likelihood of a feature quantity indicating a state of a structure generated from a time series of information about the structure, when it is assumed that the structure is in a healthy state, and a second marginal likelihood representing a likelihood of the feature quantity when it is assumed that the structure is not in the healthy state, and diagnoses the state of the structure based on the evaluation index,
The diagnosis unit includes:
Estimating the first marginal likelihood and the second marginal likelihood from the feature quantity using Bayesian estimation;
a Bayes factor obtained by dividing the second marginal likelihood by the first marginal likelihood, as the evaluation index.

Images