A soft fault detection method for the core module of underwater acoustic perception system

Abstract

The present invention provides a soft fault detection method for the core module of the underwater acoustic perception system, performs functional modeling on the core module of the underwater acoustic perception system, uses simulink to simulate the model, obtains simulation data under different fault modes, establishes a deep learning model, establishes a pre-training data set, trains the deep learning model through a back propagation algorithm, establishes a fine-tuning data set, fine-tunes the model through real data, and the model output is the fault detection result. The present invention adopts a method that combines physical modeling and data-driven to achieve end-to-end soft fault detection. The method can effectively solve the problem of soft fault detection of analog circuits and the problem of soft fault detection of the core module of the underwater acoustic perception system, and effectively reduce the dependence of deep learning on a large amount of real data. We tested it with actual data, and the test accuracy was 95.17%.

Description

Technical Field

[0001] The invention relates to the field of analog circuit fault detection, and in particular to a soft fault detection method for a core module. Background Art

[0002] With the continuous development of electronic circuits, the proportion of digital circuits is getting higher and higher, but analog circuits are still irreplaceable, especially analog circuit soft failures, parameter drift, module-level working status evaluation and other problems. The underwater acoustic perception system is a system that perceives and processes underwater information through underwater acoustic information. The underwater acoustic perception information system is mainly composed of an underwater acoustic transducer, a signal conditioning circuit and various subsequent information processing modules. The underwater acoustic transducer and the information conditioning circuit are the core parts of the underwater acoustic perception system. At present, the modularization and integration of the underwater acoustic perception system are getting higher and higher. If the soft fault detection at the module level of the underwater acoustic perception system can be realized, the maintenance and repair efficiency of the underwater acoustic perception system can be effectively improved. The present invention proposes a deep learning method for working status evaluation that can be applied to the core modules of a typical underwater acoustic perception system to judge the working status of the system and identify if an abnormal state occurs. Summary of the invention

[0003] In order to overcome the deficiencies of the prior art, the present invention provides a soft fault detection method for a core module of an underwater acoustic sensing system.

[0004] The technical solution adopted by the present invention to solve its technical problem is:

[0005] Step 1: Functional modeling of the core modules of the underwater acoustic perception system. The core modules of the underwater acoustic perception system include the underwater acoustic transducer and the signal conditioning circuit. The underwater acoustic transducer is modeled using an electromechanical equivalent circuit. The modeling of the underwater acoustic transducer and the signal conditioning circuit is performed at the functional level, and its main purpose is to simulate the circuit functional level.

[0006] Step 2: Use Simulink to simulate the model in step 1, set multiple soft fault modes, and then perform Monte Carlo simulation to obtain simulation data under different fault modes;

[0007] Step 3: Establish a deep learning model, which includes a pre-convolutional layer, a channel-space attention mechanism, a ReLu non-linear activation layer, a global average pooling layer, and a two-dimensional convolutional layer;

[0008] Step 4: Convert the one-dimensional signal into a two-dimensional matrix through short-time Fourier transform of the simulation data, establish a pre-training data set, and divide the pre-training data set into a training set and a test set;

[0009] Step 5: Use the pre-training data set established in step 4 to train the deep learning model through the back propagation algorithm;

[0010] Step 6: Fault data of the real circuit is collected to obtain real data. The real data is converted into a two-dimensional matrix through short-time Fourier transform, a fine-tuning data set is established, and the fine-tuning data set is divided into a training set and a test set;

[0011] Step 7: Fine-tune the model trained in the simulation data set in step 5 using the real data established in step 6. The model output is the fault detection result.

[0012] The electromechanical equivalent circuit adopts an LC circuit, and its resonant frequency is the resonant frequency of the transducer. The static equivalent circuit of a piezoelectric transducer is a pure capacitor, and when the transducer vibrates and radiates energy, there is also a dynamic impedance.

[0013] The deep learning model includes a pre-convolutional layer, a bottleneck layer and a post-convolutional layer, which are connected in a cascade form, wherein the pre-convolutional layer consists of a convolutional neural network layer, a batch normalization layer and a ReLu non-linear activation layer, the bottleneck layer includes a convolutional neural network layer, a batch normalization layer, a ReLu non-linear activation layer and a channel-space attention mechanism layer, and the post-convolutional layer includes a convolutional neural network layer, a batch normalization layer, a ReLu non-linear activation layer and a global average pooling layer.

[0014] In step 4, the short-time Fourier transform is:

[0015]

[0016] Among them, x[n] is the discrete signal sequence, m is the window function moving step size, w is the Hamming window function, and after taking the log and normalizing the short-time Fourier transform result, the one-dimensional signal is converted into a two-dimensional matrix to establish the pre-training data set.

[0017] When using the back-propagation algorithm, the optimizer is Adam, the learning rate is 0.0002, the momentum parameter is 0.9, the batch size is 20, and the training objective is the cross entropy loss.

[0018] In the pre-training data set, 70% is a training set and 30% is a test set.

[0019] Among the fine-tuning dataset, 30% is a training set and 70% is a test set.

[0020] In step 7, the model trained on the simulation data set in step 5 is fine-tuned by using the fine-tuning data set established in step 6, the optimizer is Adam, the learning rate is 0.0002, and the cycle is repeated 10-20 times. If the accuracy of the test set does not improve after 3 consecutive cycles, the fine-tuning is stopped.

[0021] The beneficial effect of the present invention is to provide a soft fault detection method for the core module of the underwater acoustic perception system, which adopts a combination of physical modeling and data-driven methods to achieve end-to-end soft fault detection. This method can effectively solve the problem of soft fault detection in analog circuits and the problem of soft fault detection in the core module of the underwater acoustic perception system, and effectively reduce the dependence of deep learning on a large amount of real data. After testing with actual data, the test accuracy is 95.17%. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a schematic diagram of the framework of the method of the present invention.

[0023] Figure 2 It is the electromechanical equivalent circuit diagram of the present invention.

[0024] Figure 3 It is a schematic diagram of the simulink simulation model of the present invention.

[0025] Figure 4 It is a schematic diagram of the convolutional neural network structure of the present invention.

[0026] Figure 5 It is an effect diagram of the actual data test results of the present invention. DETAILED DESCRIPTION

[0027] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0028] Step 1: Functionally model the core modules of a typical underwater acoustic sensing system. The core modules include an underwater acoustic transducer and a signal conditioning circuit. The underwater acoustic transducer is modeled through an electromechanical equivalent circuit. Here we take the static equivalent circuit of a piezoelectric transducer as an example: the lumped circuit model of a piezoelectric transducer can be equivalent to an LC circuit, and its resonant frequency is the resonant frequency of the transducer. The static equivalent circuit of a piezoelectric transducer is a pure capacitor. When the transducer vibrates and radiates energy, there is also dynamic impedance. This is caused by the effective impedance of the vibrating part of the transducer and the reaction of the medium to the vibrating part. Dynamic impedance can be represented by resistance, capacitance and inductance. For example Figure 2 As shown. In the figure: C0 is the static capacitance, R1, C1, L1 are the capacitance, capacitance, and inductance in the dynamic impedance respectively, and Rn is the transducer loss resistance, which represents the electrical loss of the transducer itself. Through the admittance original graph method, the period parameters of the underwater acoustic transducer can be measured after the environmental conditions such as temperature and static pressure reach a stable state. The signal conditioning circuit is modeled through the circuit model. We take the example that the signal conditioning circuit function and circuit module should include amplification, filtering, and automatic gain of the transducer signal.

[0029] Step 2: Use simulink to simulate the electromechanical equivalent circuit model and signal conditioning circuit model in step 1. A variety of soft fault modes can be set according to the parameters of the circuit itself. For example: the first is the normal mode, and all related parameters are set with normal tolerance values. The second is to simulate the drift of the resonance frequency of the underwater acoustic transducer. Assuming that the working resonance frequency of the underwater acoustic transducer is 21.8khz, the soft fault tolerance is [23.98khz, 30.52khz][13.08khz, 19.62Khz]. The third is the drift of the filter parameters of the information conditioning circuit. Assuming that the bandpass characteristic of the signal conditioning circuit parameters is 4khz to 16khz, the soft fault tolerance is [4.4khz, 5.6khz]. The fourth is the clipping distortion caused by saturation of the automatic gain circuit of the signal conditioning circuit. Perform 1000 Monte Carlo simulations for each mode and save them. Figure 3 Signal data of different nodes in the.

[0030] Step 3: Establish a deep learning model, which includes a pre-convolutional layer, a bottleneck layer, and a post-convolutional layer. The three modules are connected in a cascade form. The pre-convolutional layer mainly consists of a convolutional neural network layer and a batch normalization layer ReLu non-linear activation layer. The bottleneck layer mainly includes a convolutional neural network layer, a batch normalization layer ReLu non-linear activation layer, and a channel-space attention mechanism layer. The post-convolutional layer includes a convolutional neural network layer, a batch normalization layer ReLu non-linear activation layer, and a global average pooling layer. The structure of the deep learning model is as follows: Figure 3 shown.

[0031] Step 4: Perform short-time Fourier transform on the simulation data in step 2

[0032]

[0033] Where x[n] is a discrete signal sequence, m is the window function moving step, w is the Hamming window function, and the Fourier transform result is log and normalized to convert the one-dimensional signal into a two-dimensional matrix to establish a pre-training data set, of which 70% is a training set and 30% is a test set;

[0034] Step 5: Use the pre-training data set established in step 4 to train the deep learning model, using the back-propagation algorithm, the optimizer is Adam, the learning rate is 0.0002, the momentum parameter is 0.9, the batch size is 20, and the training target is the cross entropy loss;

[0035] Step 6: Perform the same transformation as step 4 on the real circuit data to establish a fine-tuning dataset, of which 30% is the training set and 70% is the test set.

[0036] Step 7: Fine-tune the model trained on the simulation data set in step 5 using the fine-tuning data set established in step 6. The optimizer is Adam, the learning rate is 0.0002, and the cycle is repeated 10-20 times. If the accuracy of the test set does not improve after three consecutive cycles, stop fine-tuning. The model output is the fault detection result.

[0037] The test accuracy of this embodiment on real data is 95.17%, and the confusion matrix is ​​as follows: Figure 5 shown.

Claims

1. A soft fault detection method for the core module of the underwater acoustic perception system, characterized in that The steps include: Step 1: Functional modeling of the core modules of the underwater acoustic perception system. The core modules of the underwater acoustic perception system include the underwater acoustic transducer and the signal conditioning circuit. The underwater acoustic transducer is modeled using an electromechanical equivalent circuit. The modeling of the underwater acoustic transducer and the signal conditioning circuit is performed at the functional level, and its main purpose is to simulate the circuit functional level. Step 2: Use Simulink to simulate the model in step 1, set multiple soft fault modes, and then perform Monte Carlo simulation to obtain simulation data under different fault modes; Step 3: Establish a deep learning model, which includes a pre-convolutional layer, a channel-spatial attention mechanism layer, a ReLu non-linear activation layer, a global average pooling layer, and a two-dimensional convolutional layer; Step 4: Convert the one-dimensional signal into a two-dimensional matrix through short-time Fourier transform of the simulation data, establish a pre-training data set, and divide the pre-training data set into a training set and a test set; Step 5: Use the pre-training data set established in step 4 to train the deep learning model through the back propagation algorithm; Step 6: Fault data of the real circuit is collected to obtain real data. The real data is converted into a two-dimensional matrix through short-time Fourier transform, a fine-tuning data set is established, and the fine-tuning data set is divided into a training set and a test set; Step 7: Fine-tune the model trained in the simulation data set in step 5 using the real data established in step 6. The model output is the fault detection result.

2. The soft fault detection method for the core module of the underwater acoustic perception system according to claim 1 is characterized in that: The electromechanical equivalent circuit adopts an LC circuit, whose resonant frequency is the resonant frequency of the transducer. The static equivalent circuit of a piezoelectric transducer is a pure capacitor. When the transducer vibrates and radiates energy, dynamic impedance also exists.

3. The soft fault detection method for the core module of the underwater acoustic perception system according to claim 1 is characterized in that: The deep learning model includes a pre-convolutional layer, a bottleneck layer and a post-convolutional layer, which are connected in a cascade form, wherein the pre-convolutional layer consists of a convolutional neural network layer, a batch normalization layer and a ReLu non-linear activation layer, the bottleneck layer includes a convolutional neural network layer, a batch normalization layer, a ReLu non-linear activation layer and a channel-space attention mechanism layer, and the post-convolutional layer includes a convolutional neural network layer, a batch normalization layer, a ReLu non-linear activation layer and a global average pooling layer.

4. The soft fault detection method for the core module of the underwater acoustic perception system according to claim 1 is characterized in that: In step 4, the short-time Fourier transform is: Among them, x[n] is the discrete signal sequence, m is the window function moving step size, w is the Hamming window function, and after taking the log and normalizing the short-time Fourier transform result, the one-dimensional signal is converted into a two-dimensional matrix to establish the pre-training data set.

5. The soft fault detection method for the core module of the underwater acoustic perception system according to claim 1 is characterized in that: When using the back-propagation algorithm, the optimizer is Adam, the learning rate is 0.0002, the momentum parameter is 0.9, the batch size is 20, and the training objective is the cross entropy loss.

6. The soft fault detection method for the core module of the underwater acoustic perception system according to claim 1 is characterized in that: In the pre-training data set, 70% is a training set and 30% is a test set.

7. The soft fault detection method for the core module of the underwater acoustic sensing system according to claim 1 is characterized in that: Among the fine-tuning dataset, 30% is a training set and 70% is a test set.

8. The soft fault detection method for the core module of the underwater acoustic sensing system according to claim 1 is characterized in that: In step 7, the model trained on the simulation data set in step 5 is fine-tuned by using the fine-tuning data set established in step 6, the optimizer is Adam, the learning rate is 0.0002, and the cycle is repeated 10-20 times. If the accuracy of the test set does not improve after 3 consecutive cycles, the fine-tuning is stopped.

Images