CN108519149B - A tunnel accident monitoring and alarm system and method based on sound time-frequency domain analysis - Google Patents

Abstract

The invention discloses a tunnel accident monitoring and alarm system and method based on sound time-frequency domain analysis, comprising a sound collection and processing module, a DSP storage analysis module and a control module, wherein: the sound collection and processing module includes a numerical converter and a sound sensor; The storage analysis module includes a ROM flash memory module, a SRAM data storage module and a DSP core processing module; the present invention senses the accident state in the tunnel through sound signals, better adapts to the tunnel operation mode and improves the operation efficiency. The network analyzes the sound signal generated by the accident in the time-frequency domain, which greatly improves the identification accuracy, coverage, anti-interference and signal-to-noise ratio of the accident signal. The overall monitoring of the tunnel, timely early warning, reduce casualties and property losses, can meet the needs of rapid rescue and reduce the scope of the accident.

Description

A tunnel accident monitoring and alarm system and method based on sound time-frequency domain analysis

technical field

The invention belongs to the field of accident monitoring and wireless communication in a tunnel, and relates to a tunnel accident monitoring and alarm system and method based on sound time-frequency domain analysis.

Background technique

In recent years, a large number of super-long highway tunnels have been built and put into operation one after another. Highway tunnels have gradually shifted from the construction peak period to the operation peak period. Tunnels are important national infrastructure, and it is very important and necessary to maintain their safety. Monitoring monitoring management. However, due to the difficulty of manual detection and technical operation, it has become the primary problem during operation, which brings difficulties to the operation and management of tunnels.

At present, there are the following problems in the accident monitoring and alarming in the tunnel: a. The monitoring system network is relatively low in intelligence, and most of them use fixed-point or manual inspections. It is unrealistic to obtain real-time status information of a certain point in the tunnel; b. The rescue arrangements and efficiency after the accident have been more tested. The tunnel is too long, and it is difficult to communicate with the outside world at the accident site, making it difficult to carry out emergency rescue activities; c. The existing tunnel accident monitoring methods are mainly smoke and video monitoring, but the distance of video monitoring is limited And the scope coverage is limited, it is greatly affected by light and internal traffic factors, and the operating cost is high, while the smoke monitoring distance and scope coverage are lower, the timeliness is poor, and it is easy to miss the golden time for accident rescue.

并进行特征值计算，该算法精准度较差，对撞击信号的识别度较低且误报率较高；又如发明CN 106887105 A与CN 103077609 A，分别提出一种基于受灾人特征和多传感器感知的隧道监控系统，其可行性、造价以及施工技术要求较高，对于传感器和硬件设施依赖性较强，将实际施工环境与人机协调理想化，脱离于现实。For example, the invention CN 104880245 A proposes a localization alarm system based on vehicle impact noise, and obtains

And perform eigenvalue calculation, the algorithm has poor accuracy, low recognition of impact signals and high false alarm rate; another example is the inventions CN 106887105 A and CN 103077609 A, which respectively propose a method based on the characteristics of victims and multi-sensor The perceptual tunnel monitoring system has high requirements for feasibility, cost and construction technology, and is highly dependent on sensors and hardware facilities. It idealizes the actual construction environment and human-machine coordination, and is separated from reality.

SUMMARY OF THE INVENTION

In order to overcome the problems in the prior art, the purpose of the present invention is to provide a tunnel accident monitoring and alarm system and method based on sound time-frequency domain analysis.

For solving the problems existing in the prior art, the technical scheme of the present invention is:

The invention provides a tunnel accident monitoring and alarm method based on sound time-frequency domain analysis, comprising the following steps:

Step 1): collect real-time sound signals in the tunnel, and filter out valid sound signals as valid frames;

Step 2): Fourier transform and filter the valid frame and template library digital signal;

Step 3): the effective frame and the template library digital signal are subjected to power spectrum conversion and screening;

Step 4): the valid frame is decomposed by wavelet and screened as a feature signal;

Step 5): Bring the characteristic signal in step 4 into the neural network for final judgment.

Specific steps are as follows:

Step 1): Filter sound valid frames:

The frequency of the sound signal used is set to 8000HZ, and the data frame whose threshold is 600 after each frame passes 50 points is taken as the valid frame, and the non-valid frame is discarded;

Step 2): Fourier transform and filter the valid frame taken in step 1) and the feature signal in the template library:

Convert the sound characteristic signal corresponding to the tunnel collision accident, the car whistle sound, and the car engine sound characteristic signal into digital characteristic signals and save them in the template library, and perform Fourier transform on the digital characteristic signals in the effective frame and template library, Convert the characteristic signal in the time domain to the characteristic signal in the frequency domain, calculate the correlation coefficient of the above two Fourier integral transformed functions, and keep the data whose correlation coefficient is greater than the threshold value, otherwise discard it, and solve the correlation coefficient Calculate according to the following formula:

In formula (1): Cov(X, Y) represents the covariance formula, D(x)D(y) represents the variance of X and Y respectively;

Step 3): Power spectrum conversion and screening are performed between the valid frame taken in step 1) and the feature signal in the template library:

The discrete Fourier integral transform with a length of 1024 bytes is processed, and the frequency is 8000 Hz. The digital feature signal in the effective frame and the template library is converted into a power spectrum. The valid frame signal whose coefficient exceeds the threshold is further calculated, and the power spectrum is calculated according to the following formula:

In formula (2): S(ω) represents the power spectrum of the effective frame, X(T) represents the time domain signal, and P represents the power spectral density;

Step 4): Wavelet decomposition of the valid frame signal, retaining the characteristic signal:

All the valid frames exceeding the threshold in step 3) are correspondingly converted into a fixed reasonable interval, that is, normalized, and the valid frame signals exceeding the threshold are subjected to wavelet decomposition, and the numerical signals in different time domains and frequency domains are decomposed and eliminated. For high-frequency signals, the decomposed low-frequency signals are retained. Through the Mallat algorithm, the retained low-frequency signals are gradually decomposed and used as characteristic signals. The wavelet decomposition formula is expressed as follows:

In formula (4): h represents the filter coefficient, C _jn represents the scale coefficient of the length space;

Step 5): train a three-layer neural network with the data for feature extraction based on wavelet decomposition, and put the feature signal retained in step 4) into the neural network for final judgment; the specific training process of the neural network: car whistle and car engine sound Training together with the tunnel impact sound, the tunnel impact sound, car whistle and car engine sound are used as training samples after wavelet decomposition; 75% of the samples are extracted by cross-validation as training data, and the rest are test data; A three-layer neural network is used, in which the input layer is the sample feature value, the number of hidden layers is greater than the number of input layers, and the output layer is the recognition result; the weights of the neural network and the threshold θ(W, b) are initialized, and the The training samples are put into the neural network for iterative calculation, the weights W and the thresholds b are adjusted, and the test data is brought into the neural network for classification in the trained neural network, and the weights and thresholds of the best performing neural network are used as The final neural network classifier;

Step 6): input the real-time characteristic signal to judge the output state result of the neural network:

Run the real-time sound signals collected by the sound sensor 2 in the tunnel in the order of steps 1 to 4, and substitute the characteristic signals obtained after running into the neural network model constructed in step 5 for calculation to determine the real-time tunnel status classification.

In step 5, the fine-tuning of the weight W and the threshold b is carried out with reference to the following formula:

表示误差函数对权值求偏导数；In formula (10): W _ij represents the corresponding weight, α represents the convergence rate,

Represents the partial derivative of the error function with respect to the weight;

表示对应阈值，

表示误差函数对阈值求偏导数；In formula (11):

represents the corresponding threshold,

Represents the partial derivative of the error function with respect to the threshold;

In step 5, the fine-tuning of the weight W and the threshold b is carried out with reference to the following formula:

In formula (13): x _i represents the number of neurons in this layer of the neural network.

In the three-layer neural network, the input layer has 11 neurons, the hidden layer has 15 neurons, and the output layer has 3 neurons.

The present invention also provides a tunnel accident monitoring and alarm system based on sound time-frequency domain analysis, comprising a sound acquisition and processing module, a DSP storage analysis module and a control module, wherein:

The sound acquisition and processing module includes a numerical converter and a sound sensor;

DSP storage analysis module includes ROM flash memory module, SRAM data storage module and DSP core processing module;

The control module includes an alarm module, a communication module and a positioning module, wherein the communication module includes a communication control device and a communication transmission device, the positioning module is used to determine the position information of the accident vehicle, and the alarm module includes an alarm signal device, an alarm control system and an alarm communication device. ;

The numerical converter is connected with the sound sensor and the DSP storage analysis module; the DSP storage analysis module is connected with the numerical converter and the control module; the control module is connected with the DSP core processing module and the tunnel monitoring center, and the communication module is connected with the communication system of the tunnel monitoring center. The alarm module is connected with the emergency fire alarm system through the relay interface, and the positioning module is connected with the sound sensor, and is arranged inside the sound sensor;

The sound sensor is used to collect the sound in the tunnel. The collected sound is converted into a digital signal by a numerical converter and then transmitted to the SRAM data storage module; the DSP core processing module is used to load the code in the ROM flash memory for execution and read the SRAM data storage module and send the command obtained after the operation to the communication transmission device.

The DSP storage and analysis module and the control module are placed in the tunnel monitoring center. In the sound acquisition and processing module, the numerical converter is placed in the tunnel monitoring center, and the sound sensor is placed on the inner wall of the tunnel.

The sound sensors are arranged on the side walls on both sides of the tunnel at a height of 2 to 2.5 m, and multiple groups are set at predetermined distances along the extending direction.

The sound sensor adopts ARM9 sound sensor, the numerical converter adopts ads5422 conversion chip, the DSP core processing module adopts TMS320C54 DSP board, and the ROM flash memory module adopts SST39LF/VF160, which is a 1M16bit CMOS multi-function FlashMPF device.

Compared with the prior art, the advantages of the present invention are as follows: the present invention senses the accident state in the tunnel through the sound signal, better adapts to the tunnel operation mode and improves the operation efficiency, and also provides a new mode for the direction of tunnel accident monitoring and alarm. The present invention performs time-frequency domain analysis on the sound signal generated by the accident through wavelet analysis and improved neural network, which greatly improves the identification accuracy, coverage, anti-interference and signal-to-noise ratio of the accident signal; the present invention is more comprehensive And directly obtain tunnel accident information to achieve overall monitoring of the tunnel, timely early warning, reduce casualties and property losses, meet rapid rescue and reduce the scope of accident impact; use the data based on wavelet decomposition for feature extraction to train three-layer neural network, The characteristic signal is put into the neural network for final judgment, and the car whistle and the car engine sound are trained together with the tunnel impact sound. The tunnel impact sound, the car whistle and the car engine sound are used as training samples after wavelet decomposition. The wavelet analysis is Localized analysis of time (space) frequency, time domain analysis refers to the method that the control system directly analyzes the system in the time domain according to the time domain expression of the output quantity, and frequency domain analysis is to decompose the time history waveform through Fourier transform It is a number of single harmonic components to obtain the frequency structure of the signal and the information of each harmonic and phase to improve the accuracy and recognition speed of the algorithm;

In addition, the present invention adds a risk coefficient Rc to the right side of formulas (10) and (11) through a reasonable formula design, thereby reducing the probability of neural network overfitting. The sound sensor is arranged in a reasonable position, while effectively collecting signals. Avoid the risk of being vulnerable to damage.

Description of drawings

Fig. 1 is the system structure schematic diagram of the present invention;

2 is a schematic diagram of an accident monitoring and alarm hardware connection according to an embodiment of the present invention;

FIG. 3 is a flowchart of an improved algorithm according to an embodiment of the present invention.

In the figure, 1-tunnel, 2-sound sensor, 3-value converter, 4-ROM flash memory module, 5-SRAM data storage module, 6-alarm signal device, 7-alarm control system, 8-alarm communication device, 9- -Communication control device, 10-communication transmission device, 11-DSP core processing module.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

1 and 2, a tunnel accident alarm system based on sound time-frequency domain analysis of the present invention includes a sound acquisition and processing module, a DSP storage analysis module, and a control module; the sound acquisition and processing module includes a sound sensor 2 and a numerical value The converter 3, the sound sensor 2 is used to collect the sound signal in the tunnel 1, and the numerical converter 3 amplifies and converts the electrical signal into a digital signal through digital-to-analog conversion; the DSP storage analysis module includes a ROM flash memory module 4, an SRAM data storage module 5. The DSP core processing module 11 and the SRAM data storage module 5 store all digital signals and wait for the DSP core processing module 11 to process them. The ROM flash memory module 4 stores the identification library and executable files. The DSP core processing module 11 passes through the SRAM data storage module 5. Identify, compare and analyze with the files stored in the ROM flash memory module 4; the control module includes an alarm module, a communication module, and a positioning module, and the communication module includes a communication control device 9 and a communication transmission device 10, which are used for the control center to receive accident information and To monitor the coordination between alarms, the positioning module is used to determine the location information of the accident vehicle. The alarm module includes an alarm signal device 6, an alarm control system 7 and an alarm communication device 8, which are used to issue accident warning information to notify the monitoring center and the disaster prevention and rescue department. linkage.

Further, the sound sensors 2 are distributed in the overall road section or important road section of the tunnel 1, arranged on the side walls on both sides of the tunnel 1 and at positions approximately equal to the height of the vehicle, and set multiple groups at a predetermined distance along the extending direction, The reference foldable coverage overlaps the monitoring range between adjacent sound sensors 2, so as to achieve barrier-free, no blind spot collection, and can also be adjusted on the spot according to the sensing properties of the sound sensor 2, which uses an ARM9 sound sensor.

Further, the numerical converter 3 is placed in the tunnel monitoring center, is connected with the sound sensor 2 and the DSP storage analysis module, converts the electrical signal received from the sound sensor 2 to the data signal and transmits it to the DSP storage analysis module, and the numerical conversion Device 3 adopts ads5422 conversion chip.

Further, the DSP storage analysis module is connected with the numerical converter 3 and the control module, is placed in the tunnel monitoring center, receives the digital signal transmitted by the numerical converter 3 and makes a comparative analysis, and the processing result is sent to the corresponding control module; The DSP core processing module 11 adopts TMS320C54 DSP board, and the ROM flash memory module 4 adopts SST39LF/VF160, which is a 1M16bit CMOS multifunctional FlashMPF device.

Further, the control module is placed in the tunnel monitoring center, connected with the DSP core processing module 11 and the relevant traffic supervisor and rescue department, and carries out emergency measures through the conveyed accident status information. The communication module is connected with the communication system of the tunnel monitoring center, the alarm module is connected with the relevant emergency fire alarm system through the relay interface, and the positioning module is connected with the sound sensor 2, placed inside the sound sensor 2, and identifies the sound sensor corresponding to the accident location through audio digital information 2, and then determine the accident location information through the positioning module, and the positioning module adopts the Mtk or Mstar GPS chip.

Further, when the DSP core processing module 11 receives the sound signal transmitted by the sound acquisition module, it first conducts a preliminary analysis of the time domain characteristics, and performs denoising processing on the sound signal through an algorithm to make it reach a recognition rate of more than 60%; After wavelet analysis and improvement, the neural network extracts features in the time domain and frequency domain, so that it can continue to improve the recognition of the impact signal on the basis of being able to be recognized in the time domain, so that the recognition rate can reach more than 90%; when the noise signal Above the threshold, the DSP core processing module 11 immediately links with the control module to alarm and rescue.

The DSP core processing module 11 is the core of the entire tunnel accident monitoring and alarm system, and it completes the functions of audio signal acquisition, control, storage, processing, and communication with the outside world. improvement of:

Step 1): Set the frequency of the sound signal used in this improved algorithm to 8000HZ. Due to the huge number of sound signals collected in the tunnel, in order to reduce the amount of calculation and improve the accuracy, the improved algorithm performs a time domain on the sound signal. For the screening, the data frame with a threshold of 600 after each frame passes 50 points is taken as the valid frame, and the non-valid frame is discarded.

Step 2): Convert the sound characteristic signal, the car whistle sound characteristic signal and the car engine sound characteristic signal corresponding to the occurrence of the tunnel collision accident into digital characteristic signals and save them in the template library. The characteristic signal is Fourier transformed, and the characteristic signal in the time domain is converted into the characteristic signal in the frequency domain. give up. The solution of the correlation coefficient is calculated according to the following formula:

In formula (1): Cov(X, Y) represents the covariance formula, D(x) D(y) represents the variance of X and Y respectively.

Step 3): process the discrete Fourier integral transform with a length of 1024 bytes, the frequency adopts 8000Hz, convert the digital feature signal in the effective frame and the template library into a power spectrum, and calculate the correlation of the above-mentioned two power spectrum converted functions. coefficient, further calculation is performed for the effective frame signal whose cross-correlation coefficient exceeds the threshold, and the power spectrum is calculated according to the following formula:

In formula (2): S(ω) represents the power spectrum of the effective frame, X(T) represents the time domain signal, and P represents the power spectral density.

Step 4): Convert all valid frames that exceed the threshold into a fixed and reasonable interval, that is, normalize, perform wavelet decomposition on the valid frame signals that exceed the threshold, decompose numerical signals in different time domains and frequency domains, and remove them. For high-frequency signals, the decomposed low-frequency signals are retained. Through the Mallat algorithm, the retained low-frequency signals are gradually decomposed and used as characteristic signals. The wavelet decomposition formula is expressed as follows:

In formula (4): h represents the filter coefficient, and C _jn represents the scale coefficient of the length space.

Step 5): Perform wavelet decomposition on the characteristic signals of tunnel impact, car whistle, and car engine sound in the template library according to step 4, take the digital characteristic signals decomposed into different frequency spaces as samples, and take 75% of the samples as training samples A neural network is constructed, and the remaining 25% are used as test samples to detect the error of the neural network. This improved algorithm uses a three-layer neural network, the input layer is the sample feature value, the number of hidden layers is greater than the number of input layers, and the output layer is the recognition result of W and b. Initialize the weights and thresholds θ(W, b) of the neural network, put the training samples into the neural network for iterative calculation, and adjust the weights W and thresholds b, as shown in the following formula:

z(2)=W(1)x+b(1) (5)

a(2)=f(z(2)) (6)

z(3)=W(2)a(2)+b(2) (7)

D=f(z(3)) (8)

In formulas (5) and (7): W(1) and W(2) represent weights, and b(1) and b(2) represent thresholds.

In formulas (6) and (8): a(2) represents the value calculated by the training sample through the neural network threshold, and D represents the stage value of one neural network iteration.

Further, the improved algorithm is based on the traditional back-propagation error neural network analysis, and fine-tunes the weight W and the threshold b, the formula is as follows:

表示误差函数对权值求偏导。In formula (10): W _ij represents the corresponding weight, α represents the convergence rate,

Represents the partial derivative of the error function with respect to the weights.

表示对应阈值，

表示误差函数对阈值求偏导。In formula (11):

represents the corresponding threshold,

Represents the partial derivative of the error function with respect to the threshold.

Using the Widrow-Hoff learning rule, the error function formula is expressed as follows:

In formula (12): d _j represents the output value of one iteration, y _i represents the initial real value, this improved algorithm analyzes the adjustment range of the weight W and the threshold b through the error function.

Further, this improved algorithm adds a risk coefficient Rc to the right side of formulas (10) and (11), thereby reducing the probability of neural network overfitting. Rc is expressed as follows:

In formula (13): x _i represents the number of features.

Through the wavelet decomposition and optimization of the weights W and the threshold b of the tunnel impact, car whistle and car engine sound characteristic signals in the template library, a neural network that can judge the sound signal and classify the state is constructed.

Step 6): Run the real-time sound signal collected by the sound sensor 2 in the tunnel according to steps 1-4, and substitute it into the neural network model constructed in step 5 for calculation to determine the real-time tunnel state classification.

Referring to Fig. 3, the algorithm flow chart describes the method for the tunnel accident monitoring and alarm system to identify whether a collision accident occurs in a tunnel through a sound signal, and the steps are as follows:

1. Collect real-time sound signals in the tunnel;

2. Screen out valid sound signals as valid frames;

3. Fourier transform and filter the valid frame and template library digital signal;

4. Power spectrum conversion and screening of valid frame and template library digital signals;

5. The valid frames are decomposed by wavelet and screened as feature signals;

6. The feature signal is used to judge the output state result by neural network.

In the whole processing process of the DSP core processing module, the improved algorithm analyzes and recognizes multiple domains and features of digital signals, analyzes the time-frequency domain through wavelet analysis and the improved neural network, and extracts through the simulation of Matlab. It greatly improves the recognition efficiency of accident sound signals, and improves the recognition accuracy, anti-interference and signal-to-noise ratio of accident sound signals.

The above content is a further detailed description of the method of the present invention in conjunction with specific embodiments, and it cannot be considered that the specific implementation of the method of the present invention is limited to this. For those of ordinary skill in the technical field to which the present invention pertains, any equivalent substitutions or obvious modifications made without departing from the concept of the present invention, with the same performance or use, shall be regarded as belonging to the claims of the present invention. The scope of patent protection determined by the book.

Claims (8)

1. a tunnel accident monitoring and alarm method based on sound time-frequency domain analysis, is characterized in that, comprises the following steps: Step 1): collect real-time sound signals in the tunnel, and filter out valid sound signals as valid frames; Step 2): Fourier transform and filter the valid frame and template library digital signals; Step 3): perform power spectrum conversion and screening on valid frame and template library digital signals; Step 4): carry out wavelet decomposition on the valid frame and filter it as a characteristic signal; Step 5): bring the characteristic signal in step 4) into the neural network for final judgment; Specific steps are as follows: Step 1): Filter sound valid frames: The frequency of the sound signal used is set to 8000HZ, and the data frame whose threshold is 600 after each frame passes 50 points is taken as the valid frame, and the non-valid frame is discarded; Step 2): Fourier transform and filter the valid frame and template library digital signal taken in step 1): Convert the sound characteristic signal, car whistle sound and car engine sound signal corresponding to the tunnel collision accident into digital signals and save them in the template library. Fourier transform is performed on the valid frames and the template library digital signals, and the time domain is converted into digital signals. The characteristic signal of is converted into a characteristic signal in the frequency domain, and the correlation coefficients are calculated for the functions obtained by the above two Fourier transforms, and the data whose correlation coefficients are greater than the threshold are retained, otherwise they are discarded. The solution of the correlation coefficient is calculated according to the following formula:

In formula (1): Cov(X, Y) represents the covariance formula, D(x) and D(y) represent the variance of X and Y respectively; Step 3): Perform power spectrum conversion and screening of the valid frame taken in step 1) and the digital signal of the template library: The discrete Fourier integral transform with a length of 1024 bytes is processed, and the frequency is 8000 Hz. The digital signal of the effective frame and the template library is converted into a power spectrum. The effective frame signal of the threshold is further calculated, and the power spectrum is calculated according to the following formula:

In formula (2): S(ω) represents the power spectrum of the effective frame, x(t) represents the time domain signal, and P represents the power spectral density; Step 4): Wavelet decomposition of the valid frame signal, retaining the characteristic signal: All the valid frames exceeding the threshold in step 3) are correspondingly converted into a fixed reasonable interval, that is, normalized, and the valid frame signals exceeding the threshold are subjected to wavelet decomposition, and the numerical signals in different time domains and frequency domains are decomposed and eliminated. For high-frequency signals, the decomposed low-frequency signals are retained. Through the Mallat algorithm, the retained low-frequency signals are gradually decomposed and used as characteristic signals. The wavelet decomposition formula is expressed as follows:

In formula (4): h represents the filter coefficient, C _jn represents the scale coefficient of the length space; Step 5): train a three-layer neural network with the data for feature extraction based on wavelet decomposition, and put the feature signal retained in step 4) into the neural network for final judgment; the specific training process of the neural network: car whistle and car engine sound Training together with the tunnel impact sound, the tunnel impact sound, car whistle and car engine sound are used as training samples after wavelet decomposition; 75% of the samples are extracted by cross-validation as training data, and the rest are test data; A three-layer neural network is used, in which the input layer is the sample feature value, the number of hidden layers is greater than the number of input layers, and the output layer is the recognition result; the weights of the neural network and the threshold θ(W, b) are initialized, and the The training samples are put into the neural network for iterative calculation, the weights W and the thresholds b are adjusted, and the test data is brought into the neural network for classification in the trained neural network, and the weights and thresholds of the best performing neural network are used as The final neural network classifier; Step 6): input the real-time characteristic signal to judge the output state result of the neural network: The real-time sound signals collected by the sound sensor (2) in the tunnel are run in the order of steps 1) to 4), and the characteristic signals obtained after the operation are substituted into the neural network model constructed in step 5) for calculation, and the real-time tunnel state classification is judged . 2. a kind of tunnel accident monitoring and alarm method based on sound time-frequency domain analysis according to claim 1, is characterized in that, in step 5), weight W and threshold value b are fine-tuned and carried out with reference to following formula:

In formula (10): W _ij represents the corresponding weight, α represents the convergence rate,

Represents the partial derivative of the error function with respect to the weight; In formula (11):

represents the corresponding threshold,

Represents the partial derivative of the error function with respect to the threshold. 3. a kind of tunnel accident monitoring and alarm method based on sound time-frequency domain analysis according to claim 1, is characterized in that, in step 5), weight W and threshold value b are fine-tuned and carried out with reference to following formula:

In formula (13): x _i represents the number of neurons in this layer of the neural network. 4. a kind of tunnel accident monitoring and alarm method based on sound time-frequency domain analysis according to claim 1, is characterized in that, in described three-layer neural network, input layer is 11 neurons, and hidden layer is 15 neurons The output layer consists of 3 neurons. 5. the system of a kind of tunnel accident monitoring and alarm method based on sound time-frequency domain analysis according to any one of claims 1-4, is characterized in that, comprises sound acquisition processing module, DSP storage analysis module and control module, wherein : The sound acquisition and processing module includes a numerical converter (3) and a sound sensor (2); The DSP storage analysis module includes a ROM flash memory module (4), a SRAM data storage module (5) and a DSP core processing module (11); The control module includes an alarm module, a communication module and a positioning module, wherein the communication module includes a communication control device (9) and a communication transmission device (10), the positioning module is used to determine the position information of the accident vehicle, and the alarm module includes an alarm signal device (6). ), an alarm control system (7) and an alarm communication device (8); The numerical converter (3) is connected with the sound sensor (2) and the DSP storage analysis module; the DSP storage analysis module is connected with the numerical converter (3) and the control module; the control module is connected with the DSP core processing module (11) and the tunnel monitoring center , the communication module is connected with the communication system of the tunnel monitoring center, the alarm module is connected with the emergency fire alarm system through the relay interface, the positioning module is connected with the sound sensor (2), and is arranged inside the sound sensor; The sound sensor (2) is used to collect the sound in the tunnel, and the collected sound is converted into a digital signal by a numerical converter (3) and then transmitted to the SRAM data storage module (5); the DSP core processing module is used to load the data in the ROM flash memory. The code executes and reads the data in the SRAM data storage module (5), and sends the instruction obtained after the operation to the communication transmission device (10). 6. the system of a kind of tunnel accident monitoring and alarming method based on sound time-frequency domain analysis according to claim 5, is characterized in that, DSP storage analysis module and control module are all placed in tunnel monitoring center, in sound acquisition processing module, The numerical converter (3) is arranged in the tunnel monitoring center, and the sound sensor (2) is arranged on the inner wall of the tunnel. 7 . The system of a tunnel accident monitoring and alarm method based on sound time-frequency domain analysis according to claim 6 , wherein the sound sensor (2) is arranged at a height of 2 to 2.5 m at the side walls on both sides of the tunnel. 8 . , and set multiple groups at a predetermined distance along its extending direction. 8. the system of a kind of tunnel accident monitoring and alarm method based on sound time-frequency domain analysis according to claim 7, is characterized in that, sound sensor (2) adopts ARM9 sound sensor, and numerical converter (3) adopts ads5422 to convert Chip, DSP core processing module (11) adopts TMS320C54 DSP board, and ROM flash memory module (4) adopts SST39LF/VF160, which is a 1M16bit CMOS multifunctional FlashMPF device.

Images