An acoustic identification method for gas leakage in pressure vessels with automatic feature extraction

Abstract

The present invention discloses an acoustic recognition method for gas leakage of pressure vessels that automatically extracts features in the field of intelligent recognition of gas leakage, and the recognition method comprises the following steps: Step 1: Collecting leakage acoustic signals through a collection device; Step 2: Preprocessing each signal collected in step 1; Step 3: Dividing the preprocessed data set into a training set and a test set, and reducing the dimension of the collected signal; Step 4: Using the training set data to train the designed collaborative bidirectional long short-term memory neural network, and using the test set data to test the classification accuracy of the trained network. The present invention is a pressure vessel gas leakage detection method that is fully functional, has high recognition accuracy, and is convenient for providing fault information to maintenance personnel, providing a strong guarantee for industrial production safety; with high recognition accuracy, it can not only detect whether there is a leak, but also determine the size of the leak, and provide assistance for subsequent leak location research, and has great practical application value.

Description

Technical Field

[0001] The invention belongs to the field of intelligent identification of gas leakage, and in particular relates to an acoustic identification method for gas leakage of a pressure vessel for automatically extracting features. Background Art

[0002] Gas production is an important branch of modern chemical production. Pressure pipes and containers are usually used in the process of gas generation, collection, storage and transportation. Once pressure pipes and containers have material manufacturing defects, corrosion and damage, gas leakage will occur. If the leaked gas is flammable, explosive or toxic and harmful hazardous gas, it will have catastrophic consequences on human life, property and the ecological environment. It is extremely important to use electronic information technology to monitor gas leaks in real time, so as to timely discover, alarm and respond, and eliminate dangers at an early stage.

[0003] Therefore, researchers have proposed various gas leakage detection technologies, including resistive catalytic combustion, semiconductor, electrochemical, infrared and acoustic detection methods, among which the acoustic detection method has high sensitivity, efficiency and accuracy. In the acoustic detection method, the sound signal of gas escaping from the pressure vessel is usually non-stationary, and the time-frequency analysis method can effectively analyze the non-stationary signal. In recent years, the time-frequency analysis method represented by wavelet transform has good time resolution and frequency resolution, and has been proved to be used to describe non-stationary signals. In view of the defect of wavelet transform that changes with time shift, researchers have improved a wavelet scattering transform, which has the characteristics of translation invariance, local deformation stability and information integrity, and has been proved to be superior through rigorous mathematical derivation and is widely used in the feature extraction of non-stationary signals such as audio signals, heart rate signals and bearing faults. In addition, in the actual engineering environment, it is not enough to judge the degree of danger of the leakage only to determine whether there is a leakage, and it is impossible to provide maintenance personnel with sufficient fault repair information. Therefore, it is necessary to propose a set of intelligent recognition systems that can determine the leakage level.

[0004] To this end, we propose an acoustic identification method for pressure vessel gas leakage with automatic feature extraction to solve the above problems. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention aims to provide a method for acoustic identification of gas leakage in pressure vessels by automatically extracting features, so as to solve the problems raised in the above-mentioned background technology.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] A method for acoustically identifying gas leakage in a pressure vessel by automatically extracting features, the method comprising the following steps:

[0008] Step 1: Collect leakage sound signals under different working conditions through the acquisition device, and establish a signal data set at different distances, different pressures, and different leakage apertures to provide a prerequisite for subsequent training steps;

[0009] Step 2: Preprocess each signal collected in step 1, decompose the signal into 8 sub-bands using wavelet packet transform, select the first three sub-bands with the highest wavelet entropy to reconstruct the signal, and remove the interference of other irrelevant frequency bands;

[0010] Step 3: The preprocessed data set in step 2 is divided into a training set and a test set in a ratio of 8:2. The wavelet scattering network is used to automatically extract features from the training set and the test set, thereby reducing the dimension of the collected signal while maximally retaining the data features.

[0011] Step 4: Use the training set data divided in step 3 to train the designed collaborative bidirectional long short-term memory neural network, and use the test set data to test the classification accuracy of the trained network.

[0012] Preferably, the collection device in step 1 includes an air compressor, a high-pressure gas cylinder, a pressure reducing valve, a pressure gauge, a low-pressure gas cylinder, a fine-adjustment valve, a detachably mounted circular leakage hole with different sizes, an acoustic sensor array, a collection device, and a host computer.

[0013] Preferably, the sampling rate set when collecting signals in step 1 is 500kHz, the sampling interval is 0.5s, the sampling distance is 30cm and 100cm, the sampling pressure is 0.2MPa and 0.5MPa, and the leakage aperture is 0.2mm and 0.5mm.

[0014] Preferably, during the data preprocessing in step 2, it is assumed that the signal at time ω is decomposed into l frequency bands after the j-th wavelet packet decomposition. The frequency band is filtered by high-pass and low-pass filters. Decompose and get the i+1th layer sub-band and expression:

[0015]

[0016]

[0017] The signals of the first two sub-bands are reconstructed, and the wavelet coefficients of the remaining six sub-bands are set to 0. 4096 data points are taken from the preprocessed signal and input into the wavelet scattering network of step 3.

[0018] Preferably, each layer in the wavelet scattering network in step 3 performs two independent operations including scattering propagation and scattering output;

[0019] The first step in the first layer of the wavelet scattering network is to combine the signal x with the complex wavelet ψ λ Convolution and modular operation to obtain the first layer of wavelet modulus coefficient U1x=|x*ψ λ | and used as the output value of the second layer, where λ represents the scattering propagation path;

[0020] The second step of the first layer combines the input signal x with a scale of 2 j The Gaussian low-pass filter φ j (u) Convolution obtains wavelet coefficient S0x = x*φ j (u);

[0021] For the second layer output of the scattering network, the hollow nodes represent the first layer wavelet modulus operator U1x, which is then combined with the complex wavelet ψ λ The convolution modulus is used to obtain the third layer wavelet modulus operator, namely U2x;

[0022] The second step of the second layer is to convert the scattering operator S j Applied to the second layer wavelet modulus coefficient U2x and scale 2 j The Gaussian low-pass filter φ j (u) Convolution obtains wavelet coefficients S2x = U2x * φ j (u).

[0023] Preferably, the wavelet scattering network operation steps are repeated to the third layer, as shown below:

[0024] U3x(u,λ1,λ2,λ3)=|||x*ψ λ1 |*ψ λ2 |*ψ λ3 |

[0025] S3x(u,λ1,λ2,λ3)=|||x*ψ λ1 |*ψ λ2 |*ψ λ3 |*φ j (u)=U3x(u,λ1,λ2,λ3)*φ j (u) (2)

[0026] After the data passes through the wavelet scattering network, the feature dimension of each data is reduced to 349*4.

[0027] Preferably, the framework diagram expression of the collaborative bidirectional long short-term memory neural network is:

[0028] h t =f(W1x t +W2h t-1 )

[0029] h′ t =f(W3xt +W4h′ t+1 )

[0030] o t =g(W5h t +W6h′ t ) (3).

[0031] Preferably, the collaborative bidirectional long short-term memory neural network uses a Softmax layer to convert the score output by the hidden layer into a vector output, where each element value of the vector is between 0 and 1, and the sum of all elements is 1, and the result of network training is converted into a probability representation;

[0032] Use the sigmoid function to map all score values ​​to 0 to positive infinity. The sigmoid function formula is as follows:

[0033]

[0034] Then use the Softmax function to obtain the probability value of the correct classification. The Softmax function is as follows:

[0035]

[0036] Preferably, the parameters of the collaborative bidirectional long short-term memory neural network are set as follows:

[0037] The input layer Inputsize is set to 349, the hidden unit is set to 500, the output mode is "last", and the last element of the data is used as the output. The fully connected layer corresponds to the number of types recognized by this method, which is 9. The training parameters use the Adam optimizer, the learning rate is 0.01, the gradient threshold is set to 1, and the loss function selects the cross entropy loss. The number of small samples selected in each round is 1000, the maximum number of iterations is 80, and the recognition accuracy is used as the evaluation indicator of classification, and the accuracy acc = N acc / N, where N acc is the number of samples whose predicted labels are the same as the actual labels of the test samples, and N is the total number of test samples.

[0038] A pressure vessel gas leakage acoustic identification controller for automatically extracting features stores a pressure vessel gas leakage acoustic identification method program for running the automatic feature extraction.

[0039] Beneficial effects of the present invention:

[0040] 1. The present invention can quickly detect the state of the pressure vessel by observing the leakage sound signal, which is of great help to the factory safety monitoring. The pressure vessel gas leakage detection method has complete functions and high recognition accuracy, and is convenient for providing fault information to maintenance personnel, which provides a strong guarantee for industrial production safety.

[0041] 2. The present invention has a high recognition accuracy rate for leak detection. It can not only detect whether there is a leak, but also determine the size of the leak and provide assistance for subsequent leak location research. It has great practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1 It is the overall flow chart of the present invention;

[0044] Figure 2 It is a schematic diagram of the collection scheme in the present invention;

[0045] Figure 3 It is the principle diagram of the three-layer wavelet scattering network in the present invention;

[0046] Figure 4 It is a schematic diagram of the structure of the collaborative bidirectional long short-term memory neural network in the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] See also Figure 1 As shown, a method for acoustically identifying gas leakage in a pressure vessel by automatically extracting features, the identification method comprises the following steps:

[0049] Step 1: Collect leakage sound signals under different working conditions, including different distances of 30cm and 100cm, different pressures of 0.2MPa and 0.5MPa, and different leakage apertures of 0.2mm and 0.5mm, to establish a signal data set to provide a prerequisite for subsequent training;

[0050] Step 2: Preprocess each collected signal, decompose the signal into 8 sub-bands using wavelet packet transform, select the first three sub-bands with the highest wavelet entropy to reconstruct the signal, and remove the interference of other irrelevant frequency bands;

[0051] Step 3: Divide the preprocessed data set into a training set and a test set in a ratio of 8:2, and use the wavelet scattering network to automatically extract features from the training set and the test set, reducing the dimension of the collected signal while maximizing the retention of data features;

[0052] Step 4: Use the training set data to train the designed collaborative bidirectional long short-term memory neural network, and use the test set data to test the classification accuracy of the trained network.

[0053] The wavelet scattering network model draws on the hierarchical structure of deep learning, and uses the overlapping operation of complex wavelet operators and modulo operators to obtain the scattering coefficient. The complex wavelet operation is a linear process, so a nonlinear operator is required to make the operation integral non-zero. The complex modulo operator is a typical nonlinear operator. Its operation process can obtain the envelope of the signal, preparing for the subsequent acquisition and output of low-frequency energy. In the scattering transformation process, the effects of the wavelet operator and the modulo operator will output low-frequency signals, while causing the loss of high-frequency components, and the main features of the signal cannot be effectively extracted. In order to further analyze the lost high-frequency part, it is necessary to continue to iterate the wavelet modulus coefficients at a higher scale, perform complex wavelet operations, modulo operations and low-pass filtering to obtain more low-frequency components. During the iteration process, the signal is scattered to different scales and angles according to the path. These coefficients are different from the convolutional network, which adjusts the parameters by back-propagating through learning sample data, so the scattering network greatly reduces the budget and computational complexity. Data classification first requires data preprocessing and the establishment of a scattering network, followed by classification analysis. The first step is to divide the data set to be classified into training sets and test sets according to different proportions according to the experimental settings. In order to establish a scattering network, the sample data of the training data is input into the scattering network. After all the training data have passed through the scattering network, the feature matrix Q0 can be obtained. All samples of the test data are input into the scattering network respectively. After all the test data have passed through the scattering network, the feature matrix Q1 can be obtained. Finally, the training set feature matrix Q0 and the test set feature matrix Q1 are combined to obtain the total feature matrix Q = [Q0, Q1]. The final scattering feature matrix Q is input into the collaborative bidirectional long short-term memory neural network for classification and discrimination.

[0054] BiLSTM is proposed on the basis of LSTM network. It is a neural network that can recurse and feedback past and future hidden layer states, which can better mine the hidden connections between time series data, thereby improving the prediction accuracy of the model and data utilization. The BiLSTM network is based on a data set collected in the laboratory and is implemented using Matlab language and DeepLearning Toolbox. Specifically, the BiLSTM network training requires a large amount of data for network training. The training data is passed through the network to obtain the predicted category label, and the loss value is obtained by comparing it with the real category label of the sample. The gradient is calculated to update the network parameters. The network parameters include weight matrix and bias variable. After the parameters are updated, the data is re-entered and the above process is repeated to obtain a trained BiLSTM model. Finally, a probability value is assigned to each output classification result through the Softmax layer, indicating the possibility of belonging to each category. This method divides the leakage sound signal data under different working conditions of no leakage and by changing the distance, pressure size, and leakage aperture into 9 categories, and experimental demonstration is carried out by simulating the leakage environment.

[0055] The simulation experiment plan is as follows Figure 2 As shown in the figure, 1 is a 13kW air compressor for generating gas; 2 is a high-pressure gas cylinder for storing the high-pressure gas generated by the air compressor; 3 is a pressure reducing valve with a pressure scale; 4 is a pressure gauge; 5 is a low-pressure gas cylinder for receiving the gas after decompression from the upstream; 6 is a fine-adjusting valve for simulating the precise regulation of the leakage pressure level; 7 is a circular leakage hole with different sizes designed to be detachable and installed, which is used to control gas leakage faults of different severity when simulating gas leakage; 8 is an acoustic sensor array, whose performance indicators are shown in Table 1. In view of the ultrasonic characteristics of gas leakage, the frequency response range and sensitivity parameters of the sensor are required to be high; 9 is the PXI5172 acquisition device of NI Company, which can realize 8-channel real-time synchronous acquisition, meet the picosecond mixed signal application, and is used to convert the analog signal collected in the laboratory environment into a digital signal; 10 is a host computer interface written by LabVIEW software, which is used to store and record the collected data.

[0056] In the data collection stage, 9 different types of data will be collected, as shown in Table 1. The sampling rate set for the collection is 500kHz and the sampling interval is 0.5s.

[0057]

[0058] Table 1

[0059] In the data preprocessing stage, wavelet packet decomposition is a technology that decomposes the original signal into multiple frequency bands. It can decompose the low-frequency and high-frequency parts. By using the good processing ability of wavelet packet decomposition for non-stationary random signals and eliminating random fluctuation data, it can reduce the adverse effect of noise signals on recognition accuracy. Assume that the signal at time ω is decomposed into l frequency bands after the j-th layer of wavelet packet decomposition. The frequency band is filtered by high-pass and low-pass filters. Decompose and get the i+1th layer sub-band and expression:

[0060]

[0061]

[0062] The db6 wavelet is used to decompose the collected data into wavelet packets, and the signal is divided into 8 sub-bands according to the frequency band. By calculating the wavelet entropy value of each sub-band, it is found that the entropy value of the first two sub-bands is the largest, indicating that the first two sub-bands contain the most information, so the signals of the first two sub-bands are reconstructed, and the wavelet coefficients of the remaining 6 sub-bands are set to 0. The high-frequency interference in the acquisition signal conditioning circuit is removed. 4096 data points are taken from the pre-processed signal and input into the wavelet scattering network.

[0063] The wavelet scattering network used in the present invention is a special convolutional network structure, which overlaps the negative wavelet operator (using Gabor wavelet) and the modulus operator, such as Figure 3 As shown in the figure, the hollow nodes in the network structure represent the intermediate results of the iterative operation, namely the wavelet modulus operator, and the solid nodes represent the layer-by-layer output of the complex scattering network, namely the wavelet scattering operator. In each layer of the scattering network, two independent operations including scattering propagation and scattering output are performed.

[0064] The first step of the first layer is to transform the signal x into a complex wavelet ψ λ Convolution and modular operation to obtain the first layer of wavelet modulus coefficient U1x=|x*ψ λ | and is used as the output value of the second layer, where λ represents the scattering propagation path. The second step of the first layer is to convert the input signal x to a scale of 2 j The Gaussian low-pass filter φ j (u) Convolution obtains wavelet coefficient S0x = x*φ j (u). For the second layer output of the scattering network, the hollow nodes represent the first layer wavelet modulus operator U1x, which is then combined with the complex wavelet ψ λ The convolution modulus is used to obtain the third layer wavelet modulus operator, namely U2x. The second operation of the second layer is to convert the scattering operator S j Applied to the second layer wavelet modulus coefficient U2x and scale 2 jThe Gaussian low-pass filter φ j (u) Convolution obtains wavelet coefficients S2x = U2x * φ j (u). Repeat the above steps to the third layer, and we get formula (2):

[0065] U3x(u,λ1,λ2,λ3)=|||x*ψ λ1 |*ψ λ2 |*ψ λ3 |

[0066] S3x(u,λ1,λ2,λ3)=|||x*ψ λ1 |*ψ λ2 |*ψ λ3 |*φ j (u)=U3x(u,λ1,λ2,λ3)*φ j (u) (2)

[0067] The output of the scattering network is obtained by cascading the wavelet scattering coefficients of each layer, where the scattering transform is gradually calculated along a path of decreasing frequency, which includes most of the scattering energy of the signal. The wavelet scattering network structure used in this method adopts a third-order scattering transform, and the corresponding quality factors are 6, 1, 1, that is, the number of wavelets in each wavelet filter group. The sampling frequency of the wavelet scattering decomposition is selected to be 1024Hz, and the time invariant scale is 3s. After the data passes through the wavelet scattering network, the feature dimension of each data is reduced to 349*4, which is 65.92% less than the original data.

[0068] The framework diagram of BiLSTM is as follows Figure 4 As shown in the figure, the input wavelet scattering characteristic signal is from the t0th scattering coefficient point to the t n The forward hidden layer output is calculated from the tth scattering coefficient point. n The reverse hidden layer output is obtained by multiplying the respective weights from the scattering coefficient point to the t0th scattering coefficient point, and then combining them to obtain the final output, which is expressed as:

[0069] h t =f(W1x t +W2h t-1 )

[0070] h′ t =f(W3x t +W4h′ t+1 )

[0071] o t =g(W5h t +W6h′ t ) (3)

[0072] Where ht is the output of the forward LSTM layer; h′ t is the output of the reverse LSTM layer. The output state of BiLSTM will be based on h t and h′ t Merge and output matrices; o t It is the output of the hidden layer after BiLSTM stacking.

[0073] Finally, the Softmax layer is used to convert the scores output by the hidden layer into a vector output. Each element of this vector has a value between 0 and 1, and the sum of all elements is 1, which converts the results of network training into probability representation. First, the sigmoid function is used to map all score values ​​to 0 to positive infinity. The sigmoid function formula is shown below.

[0074]

[0075] Then use the Softmax function to obtain the probability value of the correct classification. The Softmax function is as follows:

[0076]

[0077] In this method, the network parameters of BiLSTM are set as follows. Since each wavelet scattering feature consists of 4 columns of wavelet scattering coefficients, and the size of each column of wavelet scattering coefficients is (349*1), the input layer Inputsize is set to 349, the hidden unit is set to 500, and the output mode is "last", that is, the last element of the data is used as the output. The fully connected layer corresponds to the number of types recognized by this method, which is 9. In terms of training parameters, the Adam optimizer is used, the learning rate is 0.01, the gradient threshold is set to 1, the loss function selects the cross entropy loss, the number of small samples selected in each round is 1000, and the maximum number of iterations is 80. This method uses the recognition accuracy as the evaluation indicator of classification, and the accuracy acc=N acc / N, where N acc is the number of samples whose predicted labels are the same as the actual labels of the test samples, and N is the total number of test samples.

[0078] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0079] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0080] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0081] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0082] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for acoustic identification of gas leakage in pressure vessels by automatically extracting features, characterized in that: The identification method comprises: Step 1: Collect leakage sound signals under different working conditions through the acquisition device, and establish a signal data set at different distances, different pressures, and different leakage apertures to provide a prerequisite for subsequent training steps; Step 2: Preprocess each signal collected in step 1, decompose the signal into 8 sub-bands using wavelet packet transform, select the first three sub-bands with the highest wavelet entropy to reconstruct the signal, and remove the interference of other irrelevant frequency bands; Step 3: The preprocessed data set in step 2 is divided into a training set and a test set in a ratio of 8:

2. The wavelet scattering network is used to automatically extract features from the training set and the test set, thereby reducing the dimension of the collected signal while maximally retaining the data features. Step 4: Use the training set data divided in step 3 to train the designed collaborative bidirectional long short-term memory neural network, and use the test set data to test the classification accuracy of the trained network; Each layer in the wavelet scattering network in step 3 performs two independent operations including scattering propagation and scattering output; The first step in the first layer of the wavelet scattering network is to convert the signal With complex wavelet Convolution and modulus operation to obtain the first layer of wavelet modulus coefficients And as the output value of the second layer, represents the scattered propagation path; The second step of the first layer takes the input signal With the scale Gaussian low-pass filter Convolution to get wavelet coefficients ; For the second layer output of the scattering network, the hollow nodes represent the first layer wavelet modulus operators. ,Will Again with complex wavelets The convolution modulus is used to obtain the wavelet modulus operator of the third layer, namely ; The second step of the second layer is to use the scattering operator Applied to the second layer wavelet mode coefficients With the scale Gaussian low-pass filter Convolution to get wavelet coefficients ; Repeat the wavelet scattering network operation steps to the third layer, as shown below: （2） After the data passes through the wavelet scattering network, the feature dimension of each data is reduced to 349*4; The framework diagram expression of the collaborative bidirectional long short-term memory neural network is: （3） In the formula, is the output of the forward LSTM layer, is the output of the reverse LSTM layer, It is the output of the hidden layer after BiLSTM stacking.

2. The method for acoustic identification of gas leakage in pressure vessels by automatically extracting features according to claim 1, characterized in that: The collection device in step 1 includes an air compressor, a high-pressure gas cylinder, a pressure reducing valve, a pressure gauge, a low-pressure gas cylinder, a fine-adjustment valve, a detachable circular leakage hole with different sizes, an acoustic sensor array, a collection device, and a host computer.

3. The method for acoustic identification of gas leakage in pressure vessels by automatically extracting features according to claim 1, characterized in that: The sampling rate set when collecting signals in step 1 is 500kHz, the sampling interval is 0.5s, the sampling distance is 30cm and 100cm, the sampling pressure is 0.2MPa and 0.5MPa, and the leakage aperture is 0.2mm and 0.5mm.

4. The method for acoustic identification of gas leakage in pressure vessels by automatically extracting features according to claim 1, characterized in that: When preprocessing the data in step 2, set The time signal is After layer wavelet packet decomposition, we get The frequency band is , through high-pass filter and low-pass filter to the frequency band Decompose it and get Layer sub-band and expression: （1） The signals of the first two sub-bands are reconstructed, and the wavelet coefficients of the remaining six sub-bands are set to 0. 4096 data points are taken from the preprocessed signal and input into the wavelet scattering network of step 3.

5. The method for acoustic identification of gas leakage in pressure vessels by automatically extracting features according to claim 1, characterized in that: In the collaborative bidirectional long short-term memory neural network, a Softmax layer is used to convert the score output by the hidden layer into a vector output, where each element value of the vector is between 0 and 1, and the sum of all elements is 1, so as to convert the result of network training into a probability representation; Use the sigmoid function to map all score values ​​to 0 to positive infinity. The sigmoid function formula is as follows: （4） Then use the Softmax function to obtain the probability value of the correct classification. The Softmax function is as follows: （5）。 6. The method for acoustic identification of gas leakage in pressure vessels by automatically extracting features according to claim 1, characterized in that: The parameters of the collaborative bidirectional long short-term memory neural network are set as follows: The input layer Inputsize is set to 349, the hidden unit is set to 500, the output mode is "last", and the last element of the data is used as the output. The fully connected layer corresponds to the number of types recognized by this method, which is 9. The training parameters use the Adam optimizer, the learning rate is 0.01, the gradient threshold is set to 1, and the loss function selects the cross entropy loss. The number of small samples selected in each round is 1000, the maximum number of iterations is 80, and the recognition accuracy is used as the evaluation indicator of classification. ,in is the number of samples whose predicted labels are the same as the actual labels of the test samples, and N is the total number of test samples.

7. A pressure vessel gas leakage acoustic identification controller with automatic feature extraction, storing a program for running the pressure vessel gas leakage acoustic identification method with automatic feature extraction according to any one of claims 1 to 6.

Images