CN112903090A

CN112903090A - Neural network-based langevin transducer resonant frequency detection method

Info

Publication number: CN112903090A
Application number: CN202110300513.4A
Authority: CN
Inventors: 甘屹; 孟斌; 孙福佳; 马洪运; 何伟铭; 高丽
Original assignee: University of Shanghai for Science and Technology
Current assignee: University of Shanghai for Science and Technology
Priority date: 2021-03-22
Filing date: 2021-03-22
Publication date: 2021-06-04

Abstract

The invention relates to a resonance frequency detection method of a langevin transducer based on a neural network, which comprises the steps of firstly extracting main influence factors influencing the resonance frequency of the transducer by a transducer characteristic measuring module, then inputting the main influence factors into a neural network detection module, outputting corresponding resonance frequency by the neural network detection module through calculation of the neural network, then combining a small number of transducers with different specifications to serve as sample data for training the neural network, and simultaneously optimizing parameters of the neural network by adopting a particle swarm optimization algorithm to further improve the precision of the output resonance frequency of the network. The method has the characteristics of low cost, high efficiency and high precision, and can detect the resonant frequency of the transducer in time when the resonant frequency changes.

Description

Neural network-based langevin transducer resonant frequency detection method

Technical Field

The invention relates to a transducer resonance frequency detection method, in particular to a langevin transducer resonance frequency detection method

Background

The ultrasonic transducer is an energy conversion device which converts alternating electric signals into acoustic signals or converts the acoustic signals into the electric signals in an ultrasonic frequency range, and the power supply is controlled to apply proper excitation frequency to enable the whole transducer to resonate, so that the amplitude of the transducer is remarkably improved, the vibration is relatively stable, and the energy conversion efficiency is highest. The frequency at which the transducer resonates is referred to as the resonant frequency of the transducer, and ideally the transducer is caused to resonate if the externally applied excitation frequency is equal to the natural frequency of the transducer.

In actual use, the resonant frequency at which the transducer is brought into resonance varies as a result of variations in a number of characteristic quantities (material, dimensions, voltage, etc.). First, in the production process of the transducer, the characteristics (characteristics of itself) such as the material and the size of the transducer of different specifications are different, and the resonance frequency of the transducer is different. Secondly, in the use scenario of the transducer, the resonance frequency of the transducer is also affected by the change of external characteristics such as gear (voltage), temperature, etc. Generally, in order to obtain the resonant frequency of the transducer, the manufactured transducer is obtained by a laboratory measurement method with high cost, low precision and low efficiency, and expensive experimental equipment and experimental conditions are required. The structure of the ultrasonic transducer is shown in fig. 1, and comprises a pre-tightening bolt, a rear cover plate L1, a piezoelectric ceramic crystal stack L2, a front cover plate L3, a horn transition section L4, a front cover plate L5, the diameter D1 of the pre-tightening bolt and the diameter D2 of a front cover plate L5.

At present, ultrasonic power supplies and transducers are basically configured one to one in the market, transducers with different resonant frequencies are replaced, and the ultrasonic power supplies cannot identify the series resonant frequency of a new transducer and achieve dynamic matching; the theoretical research on the series resonance frequency of the transducer is only limited to static parameter research and measurement, and the research is not suitable for the adaptive adjustment of the ultrasonic power supply, so that the application range of the ultrasonic power supply is greatly limited. Therefore, it is important to measure the resonant frequency of the transducer.

Disclosure of Invention

In order to solve the problems in the prior art, the invention provides a resonance frequency detection method of a langevin transducer based on a neural network.

The technical scheme of the invention is as follows: a resonance frequency detection method of a langevin transducer based on a neural network comprises the steps of firstly extracting main influence factors influencing the resonance frequency of the transducer by a transducer characteristic measuring module, then inputting the main influence factors into a neural network detection module, outputting corresponding resonance frequency by the neural network detection module through calculation of the neural network, then combining a small number of transducers with different specifications to serve as sample data for neural network training, simultaneously optimizing parameters of the neural network by adopting a particle swarm optimization algorithm, and further improving the precision of the output resonance frequency of the network.

Further, the transducer characteristic measuring module comprises a transducer appearance measuring module and a working environment measuring module, and is used for analyzing and extracting main influencing factors of the transducer and measuring the transducer characteristics.

Further, the analysis and extraction of the main influence factors of the transducer comprise the extraction of 6 main influence characteristics of the transducer, which are respectively: the lengths of the front cover plate L3, the front cover plate L5 and the rear cover plate L1, the installation pretightening force F1, the assembly pretightening force F2 and the excitation voltage U, and the 6 main influence characteristics are used as the input of the neural network.

Further, the specific method for measuring the characteristics of the transducer comprises the following steps: processing the extracted main influence factors of the transducer as variables, respectively processing 3 groups of transducers according to 6 main influence factors of the transducer to obtain 3 6-power sample data, wherein the sample data is used for training a neural network model, the transducer characteristic measuring module can be used for detecting the similar transducers after the neural network model is trained, and the corresponding resonant frequency is rapidly output by measuring the characteristics of the transducers.

Further, the neural network computing module adopts a basic structure based on a BP neural network resonant frequency detection model, and the basic structure comprises an input layer, a hidden layer and an output layer, wherein the input layer receives external information of main influencing factors of the transducer, transmits the external information to a middle internal information processing layer, and transmits the external information to the output layer to complete a forward propagation process.

Further, if the result obtained is different from the desired result, the error is gradually increased in the direction opposite to the forward learning, and the weight is corrected in a gradient decreasing manner.

Further, the particle swarm optimization algorithm is used for optimizing the neural network, a particle swarm optimization algorithm is adopted, a PSO-BPNN-based resonant frequency detection model is established, main influence factors of transducer characteristics extracted by the transducer characteristic measurement module are used as input of the neural network, 3 transducers with different sizes are respectively processed according to each characteristic, three groups of different sizes are respectively set according to three characteristics of installation pretightening force F1, assembly pretightening force F2 and excitation voltage U, 100 groups of sample data are randomly extracted to serve as a training set of the neural network, and 50 groups of samples are extracted to serve as a test set.

The invention has the beneficial effects that:

1. the invention determines the main influencing factors influencing the resonance frequency of the transducer through earlier research and analysis, and deeply analyzes the relation between the influencing factors and the resonance frequency of the transducer.

2. The invention indirectly detects the resonance frequency of the transducer by a novel method, and outputs the corresponding resonance frequency value by measuring the structural size of the Langmuim transducer and external applied conditions (voltage, pretightening force and the like) and inputting the measured values into a trained neural network model.

3. In the invention, 729 sample data can be obtained as training and testing data of the neural network by only processing three transducer parts with different specifications aiming at main influence factors and then assembling. The cost is low. The transducers can be made universal for the same material without reworking the sample.

4. The invention can pre-judge the resonant frequency of the transducer in the using process of the transducer, for example, when the voltage at two ends of the transducer is changed, the neural network can quickly output the resonant frequency at the moment, thereby adjusting the excitation frequency in time and keeping the transducer to be in stable resonance. The method has the characteristics of high efficiency and high precision.

In conclusion, the method has the characteristics of low cost, high efficiency and high precision, and can detect the resonant frequency of the transducer in time when the resonant frequency changes.

Drawings

FIG. 1 is a schematic diagram of a transducer configuration;

FIG. 2 is a schematic diagram of a transducer characteristic measurement module;

FIG. 3 is a neural network flow diagram;

fig. 4 is a principle and flow chart of a neural network-based langevin type transducer resonance frequency detection method.

Detailed Description

The invention is further described with reference to the following figures and examples.

As shown in fig. 4, in the resonance frequency detection method of langevin transducer based on neural network of the present invention, the transducer characteristic measurement module is first adopted to extract the main influencing factors influencing the resonance frequency of the transducer, then the main influencing factors are input into the neural network detection module, the neural network detection module outputs the corresponding resonance frequency through the calculation of the neural network, then a small number of transducers with different specifications are combined to be used as the sample data for neural network training, and simultaneously the particle swarm optimization algorithm is adopted to optimize the parameters of the neural network, so as to further improve the precision of the output resonance frequency of the network.

The invention discloses a resonance frequency detection method of a langevin transducer based on a neural network, which mainly adopts two modules:

1. transducer characteristic (influencing factor) measuring module

The transducer characteristic (influencing factor) measuring module is constructed as shown in fig. 2, and mainly realizes two functions:

(1) transducer primary influencing factor analysis and extraction

Theoretical analysis shows that the resonance frequency of the transducer is related to the characteristics of each segment length (L1, L2, L3, L4, L5) of the transducer, the diameters of the large end and the small end (D1, D2), the density (rho 1, rho 2) of materials and the like. The above characteristics are the characteristics of the transducer itself, which affect the resonant frequency of the transducer in the unused state. In the practical use scenario of the transducer, the applied voltage U across the transducer and the difference between the installation preload F2 and the assembly preload F1 also affect the resonant frequency of the transducer.

Because the influence factors are many, the influence factors are respectively subjected to simulation analysis, the influence degree of each influence factor on the resonant frequency can be obtained, main influence factors are extracted from the influence factors to be used as the input of the neural network, the dimensionality of the input layer of the neural network can be reduced, the output precision of the neural network is improved, and meanwhile, the network convergence speed is accelerated.

Simulation analysis shows that 9 transducer characteristic quantity changes can change the resonant frequency, but the resonant frequency changes in different trends and degrees, the amplitude of the change of the resonant frequency can reach delta 322Hz in a given variable interval, and the changes are all nonlinear. The condition for the transducer to resonate is that the excitation frequency is equal to the resonance frequency, and the larger the change of the resonance frequency is, the larger the difference from the excitation frequency is, the worse the resonance effect is.

Of the 6 own characteristics (L1, L3, L4, L5, D1, D2) of the transducer, the effect of the change in the length of the transducer (L1, L3, L4, L5) on the resonance frequency is large, while the effect of the change in the radial diameter (D1, D2) on the resonance frequency is low, and the change in the resonance frequency is Δ 19Hz and Δ 12Hz, respectively, in the change interval of 4 mm.

The 3 external characteristic quantities (F1, F2, U) of the transducer have a large influence on the resonant frequency of the transducer, wherein the influence of the assembly pretension F1 and the installation pretension F2 on the resonant frequency is weaker and weaker. The influence trend of the excitation voltage U on the resonant frequency is close to a straight line, and the larger the excitation voltage is, the lower the resonant frequency is.

All 9 characteristics have strong dependence on the resonant frequency, but the three characteristics of L4, D1 and D2 have no obvious influence on the resonant frequency, so that the influence is eliminated. So 6 main features are extracted, which are respectively: the lengths of the front cover plate L3, the front cover plate L5 and the rear cover plate L1, the installation pretightening force F1, the assembly pretightening force F2 and the excitation voltage U. The above 6 features are used as inputs to the neural network.

(2) Transducer characteristic (influencing factor) measurement

And processing the extracted main influence factors of the transducer as variables, and respectively processing 3 groups of the extracted main influence factors to obtain 3 times of sample data. The sample data may be used to train the neural network model. After the model training is finished, the transducer characteristic (influencing factor) measuring module can be used for detecting the transducers of the same type, and the corresponding resonant frequency can be quickly output by measuring the transducer characteristic (influencing factor).

2. Neural network computing module

The basic structure of the BP neural network resonant frequency detection model comprises an input layer, a hidden layer and an output layer. The input layer receives external information (main influence factors of the transducer), transmits the external information to the middle internal information processing layer and transmits the external information to the output layer (resonant frequency), and the forward transmission process is completed. If the result is different from the desired result, the error is made to progress in the opposite direction of the forward learning, and the weight is corrected in a gradient descending manner.

However, conventional neural networks have some problems. In network training and prediction, the weights and thresholds are randomly generated. As the system order or unknown order increases, the fast growing network structure slows down the convergence speed and may fall into local minimum convergence, leading to an unsatisfactory prediction result. Particle Swarm Optimization (PSO) appears in 1990, is known for simplicity in operation, high precision and fast convergence and is mainly used for solving the optimization problem in industrial design, so that the PSO algorithm is introduced to optimize neural network parameters and establish a resonant frequency detection model based on PSO-BPNN. The flow is shown in fig. 3.

The main influence factors extracted by the transducer characteristic (influence factor) measuring module are used as the input of a neural network, 3 parts with different sizes are respectively processed according to each characteristic, and three groups of different sizes are respectively set according to three characteristics of installation pretightening force F1, assembly pretightening force F2 and excitation voltage U. It will be possible to compose transducers of thousands of different resonance frequencies. 100 groups of sample data are randomly extracted to be used as a training set of the neural network, and 50 groups of sample data are extracted to be used as a testing set. The other partial parameters of the PSO-BPNN take the values as shown in the table I:

table one PSO-BPNN parameter value

The transducer resonance frequency detection method based on the neural network is specific to the Langmuim transducer, if more types of transducers need to be detected, more types of transducer influence factors need to be considered, more variables are added, and the application range of the detection method is wider.

The PSO-BPNN algorithm established by the invention has higher detection precision, and the precision of the algorithm output value can be the common BP neural network, RBF neural network and the like by using a small amount of training data. In the large framework of the invention, the algorithm part can be continuously optimized, or a better algorithm is established for frequency detection, so that the output precision of the frequency detection is further improved.

Claims

1. a Langevin type transducer resonant frequency detection method based on neural network, it is characterized in that: at first adopt transducer characteristic measurement module to extract the main influence factor that affects transducer resonant frequency, then the main influence factor is input Neural network detection module, the neural network detection module outputs the corresponding resonance frequency through the calculation of the neural network, and then combines a small number of transducers of different specifications as the sample data for neural network training, and uses particle swarm optimization algorithm to optimize The parameters of the neural network further improve the accuracy of the output resonance frequency of the network.

2. the Langevin type transducer resonance frequency detection method based on neural network according to claim 1, is characterized in that: described transducer characteristic measurement module comprises transducer shape measurement module and working environment measurement module, It is used for the analysis and extraction of the main influencing factors of the transducer and the measurement of the transducer characteristics.

3. the Langevin type transducer resonance frequency detection method based on neural network according to claim 2, is characterized in that: described transducer main influence factor analysis and extraction comprise extracting transducer 6 main influence characteristics , respectively: the lengths of the front cover plate L3, the front cover plate L5 and the rear cover plate L1, the installation preload force F1, the assembly preload force F2 and the excitation voltage U, and the six main influence characteristics are used as the input of the neural network .

4. the Langevin type transducer resonance frequency detection method based on neural network according to claim 1, is characterized in that: The concrete method of described transducer characteristic measurement is: for the transducer main influence of extraction The factors are processed as variables. According to the 6 main influencing factors of the transducer, 3 groups of transducers are processed respectively, and 3 to the 6th power sample data is obtained. The sample data is used to train the neural network model. After the neural network model is trained, it will be The transducer characteristic measurement module can be used to detect similar transducers, and the corresponding resonant frequency can be quickly output by measuring the transducer characteristics.

5. the Langevin type transducer resonance frequency detection method based on neural network according to claim 1, is characterized in that: described neural network calculation module adopts the basic structure based on BP neural network resonance frequency detection model, this basic The structure includes an input layer, a hidden layer and an output layer, wherein the input layer receives the external information of the main influencing factors of the transducer, and transmits it to the intermediate internal information processing layer, and then transmits it to the output layer to complete the forward propagation process.

6. the Langevin type transducer resonance frequency detection method based on neural network according to claim 5 is characterized in that: if the result obtained is different from the desired result, the error is forward with the front side at this time. Learning proceeds in the opposite direction, and the weights are corrected by gradient descent.

7. the Langevin type transducer resonance frequency detection method based on neural network according to claim 1, it is characterized in that: described particle swarm optimization algorithm optimization neural network adopts particle swarm optimization algorithm, establishes the resonance based on PSO-BPNN The frequency detection model uses the main influencing factors of the transducer features extracted by the transducer feature measurement module as the input of the neural network, and respectively processes 3 transducers of different sizes for each feature. The three characteristics of assembly preload force F2 and excitation voltage U are set in three groups of different sizes respectively. 100 groups of sample data are randomly selected as the training set of the neural network, and 50 groups of samples are selected as the test set.