An oil and gas tank sound recognition method based on subdivision abstract features

Abstract

The application discloses a kind of based on subdivision abstract feature's press oil gas tank sound identification method.The method includes the following steps: sound data acquisition and utilize high sampling rate of sound sensor subdivision reprocessing;Utilize matrix self-encoder to carry out equipment sound abstract feature extraction;Build convolution-cyclic parallel neural network, and train by subdivision abstract coding feature;According to the convolution-cyclic parallel neural network after training, obtain press oil gas tank sound identification result.The application effectively solves the current power plant auxiliary equipment state detection technology and identifies slowly, and the problems such as low reliability, in addition, by the way of extracting subdivision abstract feature, reduce the redundancy of data, while effectively avoid the possibility that gas leakage feature is hidden, the application combines self-feature extraction and neural network mode reaches better gas leakage identification effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electrical equipment defect identification, and particularly relates to an oil and gas tank sound identification method based on subdivided abstract features. BACKGROUND

[0002] Although auxiliary equipment in power plants is a non-electric main device, it also plays an important role in the safe operation of the whole power plant. There are many auxiliary devices in power plants, such as air valves, air tanks and air pumps, which are driven by gas. Power plant inspectors also need to detect air leakage of these auxiliary devices.

[0003] The traditional method for monitoring whether the equipment leaks includes the soap bubble leak detection method, the vacuum monitoring leak detection method and the gas leak detection method. The traditional method cannot achieve real-time monitoring, and it depends on the operation of the staff and does not have the function of automatic identification.

[0004] The current device sound identification method mainly studies transformers, cables, circuit breakers and other electrical main devices, and lacks sound identification research on auxiliary equipment in power plants. The commonly used classification identification method uses neural network identification. Wang Ronghao et al. used incremental learning to improve the training time of the fast incremental SVDD and the gated recurrent unit network, and successfully realized the classification and identification of three states of the transformer, namely normal working condition, loose core and loose coil (Wang Ronghao, Li Zhe, Sun Zheng, Hu Zhaoyu, Sun Hanwen, Jiang Xiucheng. Transformer voiceprint recognition technology based on FISVDD and GRU [J / OL]. High Voltage Technology: 1-12 [2022-09-25]. DOI: 10.13336 / j.1003-6520.hve.20211923.). Zhou Mengxie et al. studied the relationship between the discharge ultrasonic signal and the corresponding defect, and used the Mel frequency cepstrum coefficient and the gamma pass frequency cepstrum coefficient as the feature vector to establish a mixed Gaussian model for defect classification (Zhou Mengxie, Tang Zhiguo, Wang Zerui, Cao Zhi, He Ninghui, Liu Bo. Recognition of partial discharge ultrasonic signals based on voiceprint recognition system [J]. High Voltage Apparatus, 2022, 58(09): 127-133. DOI: 10.13296 / j.1001-1609.hva.2022.09.016.).

[0005] According to the investigation, Zhang Xinmin et al. of Zhejiang University realized fault classification through a context attention dynamic feature extractor, obtained dynamic features through a sliding window and a recurrent neural network encoder-decoder using historical data of an industrial process, and then realized fault classification through a classifier (Zhang Xinmin, He Boqun, Song Zhiluan, Zhu Zheren. Fault classification method based on context attention dynamic feature extractor[P]. Zhejiang province: CN114298220B, 2022-09-16.); Qiu Tingting et al. of Anhui University proposed a wavelet LSTM autoencoder reconstruction method for detecting arc fault, first decomposed the feature vectors of the data before and after arc of a photovoltaic inverter using discrete wavelet transform, then obtained the reconstructed feature vectors through four LSTM autoencoders, and finally determined whether there was an arc fault through a threshold (Qiu Tingting, Qi Xing, Cao Wenping. Arc fault detection method and system based on wavelet LSTM autoencoder single-class reconstruction[P]. Anhui province: CN115078934A, 2022-09-20.); Yuan Li et al. of Beijing Control Research Engineering Institute proposed a space control system fault feature acquisition method based on automatic coding, extracted the preprocessed control system data through an automatic encoder to obtain fault features (Yuan Li, Wei Chunling, Liu Chengrui, Wang Shuyi, Li Wenbo. Space control system fault feature acquisition method based on automatic encoder[P]. Beijing: CN115017607A, 2022-09-06.); Niu Sijie et al. of Jinan University added attention correction based on a context variational autoencoder, obtained spatial and context features through two encoders, and effectively detected image anomalies (Niu Sijie, Zhou Xueying, Li Xiaohui, Gao Xizhan, Dong Zihao, Liu Bowen, Zhang Fenghang, Dong Jiweng. Image anomaly detection method based on spatial context variational autoencoder[P]. Shandong province: CN114913377A, 2022-08-16.); Zhang Hui et al. of Hunan University realized latent feature extraction of product data based on a variational autoencoder, and completed defect detection using a puzzle solver on the decoder (Zhang Hui, Hu Feiyi, Chen Yurong, Liu Jiaxuan, Jiubing Xu, Zhu Qing, Yuan Xiaofang, Wang Yaonan. Defect detection method based on variational autoencoder[P]. Hunan province: CN114862811A, 2022-08-05.); Therefore, there are problems that current equipment air leakage diagnosis technologies judge whether air leakage occurs through traditional detection methods, air leakage occurs due to the uncertainty of the equipment, which leads to untimely warning, and the accuracy of the conventional recognition method is not high. SUMMARY

[0006] The application is to solve the problems of current equipment gas leakage diagnosis technology, which judges whether gas leakage occurs through traditional detection methods, the uncertainty of equipment gas leakage, resulting in untimely early warning, and the low accuracy of conventional identification methods. The current sound recognition algorithm is based on prior knowledge after sound preprocessing, such as frequency domain, wavelet decomposition, and principal component analysis. The method of the application obtains the global sound signal based on a high sampling rate sound sensor, and uses an encoder to extract the deep abstract features of the sound. On the one hand, it avoids the preprocessing process of hiding the gas leakage fault features, and on the other hand, it reduces the dimension of the training data, improves the speed and accuracy. The traditional method relies on technical personnel in the detection process and is restricted by equipment and environmental conditions. The application uses parallel network recognition technology to identify abnormal sound, which opens up the time and space dimensions, facilitates the processing of massive equipment data in the big data era, and builds a cloud-side collaborative power plant sealing equipment state monitoring system, which is not achieved by traditional detection methods.

[0007] The application provides a pressure oil and gas tank sound recognition method based on subdivided abstract features, which can identify normal working conditions, normal gas leakage conditions and abnormal gas leakage conditions. On the one hand, a high sampling rate sensor is used to collect global signals, and an encoder is used to extract subdivided abstract features, so as to complete data dimension reduction and feature extraction. On the other hand, a deep learning algorithm is used to build a convolution-cyclic double parallel network, which retains the identification of time sequence information and spatial information, and further improves the accuracy of identification.

[0008] The application has the advantages that the subdivided abstract features of the sound are extracted by using an improved encoder, and a convolution-cyclic double parallel network is built to identify the sound, so that a more efficient and better gas leakage monitoring method is obtained by combining the two.

[0009] The application is achieved by at least one of the following technical solutions.

[0010] A pressure oil and gas tank sound recognition method based on subdivided abstract features comprises the following steps:

[0011] S1, sound data is acquired and subdivided and reprocessed by using a high sampling rate sound sensor;

[0012] S2, device sound abstract feature extraction is performed by using a matrix autoencoder;

[0013] S3, a convolution-cyclic parallel neural network is built, and training is performed by using subdivided abstract coding features;

[0014] S4, a pressure oil and gas tank sound recognition result is obtained according to the trained convolution-cyclic parallel neural network.

[0015] Further, in step S1, the sound data is a sound signal of the pressurized oil tank obtained by a sound sensor arranged on the pressurized oil tank of the auxiliary equipment of the power plant.

[0016] Further, the subdivision reprocessing is a subdivision processing and a repeated processing on the obtained sound data, including the following steps:

[0017] S1.1, subdivision processing:

[0018] l c = Sr x s c

[0019]

[0020] wherein, l c is the length of the sound signal of each segment after the subdivision processing of the sound data, Sr is the sampling rate of the sound sensor, f j min is the minimum frequency under the jth working state of the pressurized oil tank, S c is the minimum time length for ensuring the sound full frequency domain information;

[0021] S1.2, repeated processing:

[0022] Sa i = S[l c x (i-1) / 2, l c x (i+1) / 2] i = 1, 2, ···, W

[0023] wherein, S is the sound sample data, [] represents the value in the interval, Sa i is the ith sound sample after the subdivision reprocessing, and W is the number of sound samples after the subdivision reprocessing.

[0024] Further, in step S2, the sound signal after the subdivision reprocessing is subjected to feature extraction by the stack autoencoder, including the following steps:

[0025] S2.1, matrix autoencoder construction;

[0026] S2.2, optimization and improvement of the stack autoencoder;

[0027] S2.3, feature extraction by the constructed stack autoencoder.

[0028] Further, in step S2.1, the matrix autoencoder is constructed as follows:

[0029] The encoding connection layer and the decoding connection layer are constructed, and the weight is set to a symmetrical structure, and the weight connection of the encoder layer is:

[0030] vb = g bd x v d + u bd

[0031] wherein v d is a node d located in an upper layer, v b is a node b located in a lower layer, g bd , u bd are parameters to be optimized from the node d to the node b;

[0032] The weight connection at the decoder layer is:

[0033] v p = 1 / g bd x v q - 1 / g bd x u bd

[0034] wherein v q is a node q located in an upper layer, v p is a node p located in a lower layer.

[0035] Further, in step S2.2, the optimization of the stack autoencoder is improved, and the details are as follows:

[0036] A regularization term and 2-norm are introduced to improve the error optimization function of the stack autoencoder:

[0037]

[0038] wherein Sa is the encoded input, is the decoded output, and λ is the regularization term.

[0039] Further, in step S3, the convolutional neural network is built, and the details are as follows:

[0040] According to the subdivided autoencoder features, the two-dimensional data is reconstructed for input:

[0041] x1 = x input reshape(a, a)

[0042] wherein x input is the sound feature extracted by the encoder, x1 is the input of the convolutional neural network, reshape(a, a) is the data reorganization operation, and a is the dimension of the input data.

[0043] The convolutional neural network is built, including sequentially connected convolutional layers, pooling layers, flattening layers, a first full connection layer, a second full connection layer, and a third full connection layer, wherein the convolution operation is as follows:

[0044]

[0045] in, For convolution operations, w conv Let Z be the training parameters of a×a. conv This is the output of the convolutional layer;

[0046] Pooling operations are as follows:

[0047] g = pooling(Z) conv )

[0048] y1 = wg + b

[0049] Where y1 is the output of the convolutional neural network, g is the output of the pooling layer, pooling is the pooling operation, b is the bias term, and w is the weight.

[0050] Furthermore, in step S3, a recurrent neural network is constructed, as follows:

[0051] Based on the detailed autoencoder features, reconstruct the two-dimensional data for input:

[0052] x2=x input reshape(step, dim)

[0053] Where step is the time step, dim is the time sequence length, and x2 is the input of the recurrent neural network;

[0054] Construct a recurrent neural network, including recurrent units and fully connected layers, where the recurrent units are configured as follows:

[0055] R (t) =k r ·h(s (t-1) x2 (t) ,w r )+c r

[0056] s = O(R, x2, m)

[0057] Where t is the time step, s (t-1) Let x2 be the internal state of the system at time t-1. (t) R is the input received by the loop unit at time t. (t) k is the output of the loop unit at time t. r w r c r Let O be the training parameters of the recurrent unit, and let h be the first activation function and h be the second activation function, respectively. Let m be the output of the first activation function h, R be the output of the recurrent unit, and s be an internal state calculated by the recurrent unit.

[0058] The output of the fully connected layer is as follows:

[0059] y2=wR+b

[0060] wherein, y2 is the output of the recurrent neural network, R is the output of the recurrent neural unit after time steps, b is the bias term, and w is the weight;

[0061] Further, in step S3, a convolution-recurrent parallel neural network is built to output results in parallel:

[0062] The results are integrated and output using a connection layer:

[0063] y al = ave (y1, y2, η)

[0064] wherein, y1 is the output of the convolutional neural network, y2 is the output of the recurrent neural network, and η is the average weight, which is a value set by a person, ave is the connection operation, and y al is the parallel output result, i.e., the output of the convolution-recurrent parallel neural network.

[0065] Further, in step S4, the device state is obtained according to the parallel output result of the trained convolution-recurrent parallel neural network:

[0066] name 状态 = reonehot (y al )

[0067] wherein, name is the state name, and reonehot is the reverse one-hot encoding.

[0068] Compared with the prior art, the present application has the following advantages:

[0069] The present application proposes a method for identifying the sound of a pressurized oil and gas tank based on subdivided abstract features, effectively solving the problems of insufficient practicability and poor reliability of the current gas leakage detection technology through gas monitoring and physical detection. In addition, the method of identifying the sound of a device through a convolution-recurrent parallel neural network effectively utilizes the information of sound features in two-dimensional arrangement and time sequence arrangement, improves the accuracy of identification, and achieves better gas leakage identification effect. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 is a step flowchart of a method for identifying the sound of a pressurized oil and gas tank based on subdivided abstract features in an embodiment of the present application;

[0071] Figure 2 is a step flowchart of a method for extracting abstract features of device sound using a matrix autoencoder in an embodiment of the present application;

[0072] Figure 3 is a structural diagram of a convolution-recurrent parallel neural network in an embodiment of the present application. DETAILED DESCRIPTION

[0073] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0074] Embodiment 1

[0075] A method for identifying the sound of a pressurized oil tank based on subdivided abstract features, as shown in the figure, comprising the following steps: Figure 1

[0076] S1, sound data acquisition and subdivided reprocessing using a high sampling rate of a sound sensor;

[0077] The sound data is the sound signal of the pressurized oil tank obtained by the sound sensor installed on the auxiliary equipment of the power plant;

[0078] The subdivided reprocessing is a fine subdivision and repeated processing of the obtained sound data, comprising the following steps:

[0079] S1.1, fine subdivision processing:

[0080] l c = Sr x s c

[0081]

[0082] Wherein, l c is the length of the sound signal of each segment after the fine subdivision processing of the sound data, Sr is the sampling rate of the sound sensor, f j min is the minimum frequency under the working state of the jth pressurized oil tank, S c is the minimum time length to ensure the full frequency domain information of the sound;

[0083] S1.2, repeated processing:

[0084] Sa i = S[l c x (i-1) / 2, l c x (i+1) / 2] i = 1, 2, ···, W

[0085] Wherein, S is the sound sample data, [] represents the value in the interval, Sa i is the ith sound sample after subdivided reprocessing, and W is the number of sound samples after subdivided reprocessing.

[0086] S2, using a matrix autoencoder to extract device sound abstract features, as shown in the figure:​Figure 2 As shown, comprising the following steps:

[0087] S2.1, matrix auto-encoder construction:

[0088] The encoding connection layer and the decoding connection layer are constructed, and the weight is set to a symmetric structure. The weight connection of the encoder layer is:

[0089] v b = g bd x v d + u bd

[0090] Where v d is the node d located in the upper layer, v b is the node b located in the lower layer, g bd , u bd are the to-be-optimized parameters of the node d to the node b;

[0091] The weight connection of the decoder layer is:

[0092] v p = 1 / g bd x v q - 1 / g bd x u bd

[0093] Where v q is the node q located in the upper layer, and v p is the node p located in the lower layer.

[0094] S2.2, optimization and improvement of the stack auto-encoder:

[0095] The regularization term and 2-norm are introduced to improve the error optimization function of the stack auto-encoder:

[0096]

[0097] Where Sa is the encoding input, is the decoding output, and λ is the regularization term.

[0098] S2.3, feature extraction using the constructed stack auto-encoder.

[0099] S3, build a convolution-cyclic parallel neural network, train by subdividing abstract coding features, comprising the following steps:

[0100] S3.1, build a convolutional neural network:

[0101] According to the subdivided auto-encoding features, reconstruct the two-dimensional data for input:

[0102] x1=x inputreshape(a,a)

[0103] where x input is the sound feature extracted by the encoder, x1 is the input of the convolutional neural network, reshape(a,a) is a data reorganization operation, and a is the dimension of the input data;

[0104] A convolutional neural network is built, including sequentially connected convolutional layers, pooling layers, a first fully connected layer, a second fully connected layer, and a third fully connected layer, wherein the convolutional operation is as follows:

[0105]

[0106] wherein, is the convolutional operation, w conv is a training parameter of a x a, and Z conv is the output of the convolutional layer;

[0107] The pooling operation is as follows:

[0108] g = pooling(Z conv )

[0109] y1 = wg + b

[0110] wherein, y1 is the output of the convolutional neural network, g is the output of the pooling layer, pooling is the pooling operation, b is the bias term, and w is the weight;

[0111] S3.2 Build a recurrent neural network:

[0112] According to the subdivided auto-encoding feature, reconstruct the two-dimensional data for input:

[0113] x2 = x input reshape(step, dim)

[0114] wherein, step is the time step, dim is the time length, and x2 is the input of the recurrent neural network;

[0115] A recurrent neural network is built, including a recurrent unit and a fully connected layer, wherein the recurrent unit is set as follows:

[0116] R (t) = k r ·h(s (t-1) ,x2 (t) ,w r )+c r

[0117] s = O(R, x2, m)

[0118] wherein, t is the time step, s (t-1) is the internal state of the system at t-1(t) R is the input received by the recurrent unit at time t (t) R is the output of the recurrent unit at time t r w r c r O and h are the first and second activation functions, respectively, m is the output of the first activation function h, R is the output of the recurrent unit, and s is an internal state computed by the recurrent unit

[0119] The output of the fully connected layer is as follows:

[0120] y2 = wR + b

[0121] where y2 is the output of the recurrent neural network, R is the output of the recurrent neural unit after a time step, b is the bias term, and w is the weight

[0122] S3.3 Convolution-Recurrence Parallel Neural Network Parallel Output Results:

[0123] The results are integrated and output using the connection layer:

[0124] y al = ave(y1, y2, η)

[0125] where y1 is the output of the convolutional neural network, y2 is the output of the recurrent neural network, η is the average weight, which is a manually set value, ave is the connection operation, and y al is the parallel output result, i.e., the output of the convolution-recurrence parallel neural network.

[0126] S4. Obtain the oil gas tank sound recognition result according to the network output

[0127] Obtain the device state according to the parallel output result of the trained convolution-recurrence parallel neural network

[0128] name 状态 = reonehot(y al )

[0129] where name is the state name, and reonehot is the reverse one-hot encoding

[0130] In this embodiment, the sound data of three states of the pressure oil tank is collected, which are normal working.wav, normal leakage.wav and abnormal leakage.wav. After subdivision processing and reprocessing, the sound original sample is obtained, the data dimension is [5180, 1600], the abstract feature extraction is carried out by using the self-encoder, the 5-layer stack type encoder and stack type decoder are built, the input and output are 1600, 800, 784, 800, 1600 respectively, the improved loss function is set for feature extraction, and the coding error of three kinds of sound is 0.2%, 2.5% and 0.3%.

[0131] The time step of the recurrent neural network is set to 4, the output is 20, the predicted label value is output by using the full connection layer, the convolution layer of the convolution network is two-dimensional convolution, the convolution kernel is 16, the number is 3, the step of the pooling layer is 2, the full connection layer is used for dimension reduction output respectively, three full connection layers are set, the first full connection layer is set to output 784 nodes, the second full connection layer is set to output 20 nodes, and the third full connection layer is set to output 3 nodes. Finally, the combined output result is obtained by using the connection layer, and the structure is as shown in Figure 3 .

[0132] The training result is iterated 120 times, and the result is shown in Table 1.

[0133] Table 1

[0134] Category Normal operation Abnormal air leakage Normal air leakage Accuracy of the method of the present invention 98.7％ 97.3％ 98.6％

[0135] Embodiment 2:

[0136] The sound data of three states of the pressure oil tank is collected, which are normal working.MP3, normal leakage.MP3 and abnormal leakage.MP3. After subdivision processing and reprocessing, the sound original sample is obtained, the data dimension is [5180, 1600], the abstract feature extraction is carried out by using the self-encoder, the 5-layer stack type encoder and stack type decoder are built, the input and output are 1600, 800, 784, 800, 1600 respectively, the improved loss function is set for feature extraction, and the coding error of three kinds of sound is 0.2%, 2.2% and 0.3%.

[0137] The time step of the recurrent neural network is set to 4, the output is 20, the predicted label value is output by using the full connection layer, the convolution layer of the convolution network is two-dimensional convolution, the convolution kernel is 16, the number is 3, the step of the pooling layer is 2, the full connection layer is used for dimension reduction output respectively, three full connection layers are set, the first full connection layer is set to output 784 nodes, the second full connection layer is set to output 20 nodes, and the third full connection layer is set to output 3 nodes. Finally, the combined output result is obtained by using the connection layer, and the structure is as shown in

[0138] The training result is iterated 80 times, and the result is shown in Table 2.

[0139] Table 2

[0140]

[0141] Example 3:

[0142] The sound data of three states of the oil gas tank is collected, which are normal working.WAV, normal gas leakage.WAV and abnormal gas leakage.WAV. After subdivision processing and reprocessing, the sound original sample is obtained, the data dimension is [5180, 1600], the abstract feature extraction is performed by using the self-encoder, the 5-layer stack type encoder and stack type decoder are built, the input and output are 1600, 1000, 784, 1000, 1600 respectively, the improved loss function is set for feature extraction, and the coding error of the three sounds is 0.2%, 2.5% and 0.3%.

[0143] The time step of the recurrent neural network is set to 4, the output is 20, the predicted label value is output by using the full connection layer, the convolution layer of the convolution network is two-dimensional convolution, the convolution kernel is 16, the number is 3, the step of the pooling layer is 2, the full connection layer is used for dimension reduction output, the output of the full connection layer 1 is set to 784 nodes, the control group 2 is set to a single cycle unit layer and a full connection layer, the input of the cycle unit layer is set to [4, 196], the output is set to 20 nodes as the input of the full connection layer, and the full connection layer is set to 4 nodes.

[0144] The training result is iterated 80 times, and the result is shown in Table 3.

[0145] Table 3

[0146]

[0147] The above identification and tracking method is combined as the preferred embodiment of the present application, but the embodiment of the present application is not limited by the above examples, and any modification, modification, substitution, combination, simplification made without departing from the spirit and principles of the present application should be equivalent to the replacement method, and should be included in the protection scope of the present application.

Claims

1. A method for sound recognition of pressurized oil and gas tanks based on subdivided abstract features, characterized in that, Includes the following steps: S1. Acquire sound data and perform detailed reprocessing using a high sampling rate sound sensor; the detailed reprocessing involves finely segmenting and repeating the acquired sound data, including the following steps: S1.1 Fine cutting process: in, c The length of each audio signal segment after fine segmentation of the audio data. Sr It is the sampling rate of the sound sensor. f j min It is the first j The minimum frequency of the pressurized oil gas tank under operating conditions. S c It is the minimum time length to ensure the information of the entire frequency domain of sound; S1.2, Repeated processing: in, S For sound sample data, [] indicates that the values ​​are taken within a range. Sa i It is the i-th sound sample after subdivision and reprocessing. W To further refine the number of reprocessed audio samples; S2. Extract abstract features of device sound using a matrix autoencoder; S3. Construct a convolutional-recurrent parallel neural network and train it by subdividing and abstracting the encoded features; S4. Obtain the sound recognition result of the oil tank based on the trained convolutional-recurrent parallel neural network.

2. The method for sound recognition of pressurized oil and gas tanks based on subdivided abstract features according to claim 1, characterized in that, In step S1, the sound data is the sound signal of the pressure oil tank obtained from the sound sensor installed in the pressure oil tank of the power plant auxiliary equipment.

3. The method for sound recognition of pressurized oil and gas tanks based on subdivided abstract features according to claim 1, characterized in that, In step S2, the subdivided and reprocessed audio signal is subjected to feature extraction using a stacked autoencoder, including the following steps: S2.1 Construction of matrix autoencoder; S2.2 Optimization and improvement of stack-based autoencoders; S2.

3. Feature extraction is performed using the constructed stacked autoencoder.

4. The method for sound recognition of pressurized oil and gas tanks based on subdivided abstract features according to claim 3, characterized in that, In step S2.1, the matrix autoencoder is constructed as follows: Construct an encoder connection layer and a decoder connection layer, with weights set to a symmetric structure. The weight connections in the encoder layer are as follows: in, v d It is a node located at the upper level. d , v b It is a node located at the lower level. b , g bd , u bd It is a node d To the node b The parameters to be optimized; The weighted connections in the decoder layer are as follows: in, v q It is a node located at the upper level. q , v p It is a node located at the lower level. p .

5. The method for sound recognition of pressurized oil and gas tanks based on subdivided abstract features according to claim 4, characterized in that, In step S2.2, the optimizations and improvements to the stack-based autoencoder are as follows: The error optimization function of the stacked autoencoder is improved by introducing a regularization term and a 2-norm: in, Sa For encoded input, For decoding output, λ This is a regularization term.

6. The method for sound recognition of pressurized oil and gas tanks based on subdivided abstract features according to claim 1, characterized in that, In step S3, the convolutional neural network is constructed, as follows: Based on the detailed autoencoder features, reconstruct the two-dimensional data for input: in, x input These are the sound features extracted by the encoder. x 1 represents the input to the convolutional neural network. reshape ( a , a This refers to a data reorganization operation. a Dimensions of the input data; Construct a convolutional neural network, consisting of sequentially connected convolutional layers, pooling layers, flattening layers, a first fully connected layer, a second fully connected layer, and a third fully connected layer. The convolutional operation is as follows: in, For convolution operations, w conv for a × a The training parameters, Z conv This is the output of the convolutional layer; Pooling operations are as follows: in, y 1 represents the output of the convolutional neural network. g For pooling layer output, pooling For pooling operations, b For deviation terms, w As weight.

7. The method for sound recognition of pressurized oil and gas tanks based on subdivided abstract features according to claim 6, characterized in that, In step S3, a recurrent neural network is constructed, as follows: Based on the detailed autoencoder features, reconstruct the two-dimensional data for input: in, step For time step, dim For timing length, x 2 represents the input to the recurrent neural network; Construct a recurrent neural network, including recurrent units and fully connected layers, where the recurrent units are configured as follows: in, t For time steps, s (t-1) In order to be in t- Internal state of the system at time 1 The input received by the loop unit at time t. This represents the output of the loop unit at time t. k r , w r , c r These are the training parameters for the recurrent unit. O and h These are the first activation function and the second activation function, respectively. m The first excitation function h The output, R For output of the loop unit, s An internal state calculated for a loop unit; The output of the fully connected layer is as follows: in, y 2 represents the output of the recurrent neural network. R The output of the recurrent neural unit after a time step. b For deviation terms, w As weight.

8. The method for sound recognition of pressurized oil and gas tanks based on subdivided abstract features according to claim 7, characterized in that, In step S3, a convolutional-recurrent parallel neural network is built to output the results in parallel: Use a connection layer for result integration and output: in, y 1 represents the output of the convolutional neural network. y 2 represents the output of the recurrent neural network; η The average weight is a value set by the individual. ave For join operations, y al This is the output of the parallel output, i.e., the output of the convolutional-recurrent parallel neural network.

9. A method for sound recognition of pressurized oil and gas tanks based on subdivided abstract features according to any one of claims 1 to 8, characterized in that, In step S4, the device status is obtained based on the parallel output of the trained convolutional-recurrent parallel neural network: Where name is the state name and reonehot is the reverse one-hot encoding.

Images