Power grid acoustic sensor array self anomaly detection method

Abstract

The application provides a power grid sound sensor array self-abnormality detection method, which utilizes cross-validation, a long short-term memory network and a self-encoder to initially determine through cross-validation, further determine through the long short-term memory network processing time series data and the self-encoder feature learning capability, timely find the abnormality of the sound sensor, improve the accuracy and reliability of the sound sensor self-abnormality detection, and reduce the monitoring errors and potential risks caused by the abnormality of the sound sensor; the method comprises the following steps: step 1, sound sensor array data acquisition; step 2, sound signal cross-correlation coefficient method initial determination; step 3, long short-term memory network feature extraction; step 4, self-encoder abnormality determination model; and step 5, sound sensor abnormality determination.

Description

Technical Field

[0001] This invention relates to the field of acoustic sensor detection technology, and specifically to a method for detecting anomalies in an acoustic sensor array used in power grids. Background Technology

[0002] In power systems, the safe and stable operation of power equipment is crucial. As a core component of DC transmission systems, the operating status of power converter valves directly impacts the reliability and stability of the entire power system. To monitor the operating status of power equipment in real time and promptly detect potential faults, acoustic sensors and their arrays are widely used in partial discharge detection and location of power equipment. Acoustic sensors can capture acoustic signals generated during equipment operation. By analyzing and processing these signals, operating status information can be obtained, providing important data for equipment maintenance and fault diagnosis.

[0003] Although acoustic sensors play a vital role in power equipment monitoring, they may malfunction over long-term use due to various factors such as environmental interference, equipment aging, and component failure, leading to inaccurate or unreliable data. Traditional methods for detecting inherent malfunctions in acoustic sensors have the following limitations:

[0004] Threshold judgment method: This method sets fixed upper and lower thresholds, and judges an anomaly when sensor data exceeds these limits. However, the actual acoustic environment is complex and variable, and the acoustic signal characteristics of different devices and operating conditions vary greatly. A fixed threshold is difficult to adapt to all situations and is prone to misjudgment or missed judgment. For example, under certain special operating conditions, normal acoustic signals may exceed the set threshold, leading to a false judgment of an anomaly; while under some slowly changing fault conditions, the signal may gradually deviate from the normal range but not exceed the threshold, resulting in a missed judgment.

[0005] Statistical analysis methods, while considering the statistical characteristics of the data, are insufficient for uncovering the long-term dependencies inherent in acoustic data with temporal characteristics. Acoustic signals are dynamic processes that change over time, containing complex temporal correlations. Traditional statistical analysis methods often focus only on statistical quantities such as the mean and variance of the data, neglecting the evolution of the data over time. This makes it difficult to accurately capture the dynamic changes in the data, thus affecting the accuracy and timeliness of anomaly detection. For example, when equipment experiences early-stage failures, the acoustic signal may exhibit weak, gradually increasing abnormal changes. Traditional statistical analysis methods may fail to detect these changes in a timely manner, leading to delays in fault diagnosis.

[0006] The actual power equipment operating environment is usually complex, and the complex acoustic environment puts higher requirements on the detection method, and various noise interferences such as mechanical vibration noise, electromagnetic interference noise and the like exist. These noises and normal acoustic signals of the equipment are superimposed on each other, so that the characteristics of the acoustic signals are more complex and difficult to distinguish. In this case, the traditional acoustic sensor itself anomaly detection method is difficult to effectively distinguish the normal signal and the abnormal signal, and cannot meet the detection requirements in the complex acoustic environment.

[0007] The power converter valve equipment is a key equipment of the DC power transmission system, and accurate monitoring of the operating state thereof is crucial for ensuring the safe and stable operation of the power system. Once the acoustic sensor fails, resulting in inaccurate or unreliable data collected, it will directly affect the judgment of the operating state of the power converter valve equipment, may cause incorrect maintenance decisions, and even lead to missed judgment and delayed processing of equipment failure, causing serious safety hazards to the power system. Therefore, developing an effective acoustic sensor array self-detection method that can detect acoustic sensor faults in real time and accurately in a complex acoustic environment is of great practical significance for ensuring the safe and reliable monitoring of the power converter valve equipment.

[0008] In view of the limitations of the traditional acoustic sensor itself anomaly detection method and the special requirements of the complex acoustic environment and power equipment monitoring, developing a more advanced and effective acoustic sensor array self-detection method has become a problem to be solved at present. SUMMARY

[0009] The present application aims to overcome the technical defects of the prior art and provide an acoustic sensor array self-anomaly detection method for power grids. By using the cross-validation method combined with the long short-term memory network and the autoencoder, the cross-validation preliminary judgment, the long short-term memory network processing time series data and the feature learning ability of the autoencoder are used for further judgment, the abnormal condition of the acoustic sensor is discovered in time, the accuracy and reliability of the acoustic sensor itself anomaly detection are improved, and the monitoring errors and potential risks caused by the acoustic sensor anomaly are reduced.

[0010] The present application provides the following technical scheme: an acoustic sensor array self-anomaly detection method for power grids, comprising the following steps:

[0011] Step 1: acoustic sensor array data acquisition; a group of M acoustic arrays are set, the acoustic signals are continuously collected at a fixed sampling frequency, the collected analog signals are converted into digital signals, and through the sound signal data in a period of time, M discrete time sequences can be obtained, the length of each sequence is N;

[0012] Step 2: Initial judgment using the acoustic signal cross-correlation coefficient method; For the k-th acoustic sensor to be verified, calculate its cross-correlation coefficient with the signals of the remaining M-1 acoustic sensors. This yields M-1 cross-correlation coefficients. When the coefficient exceeds a set correlation coefficient threshold... hour, The overall fault judgment coefficient is obtained by summing the counts of the M-1 cross-correlation coefficients, with a count of 1. Set preliminary fault judgment thresholds When the comprehensive fault judgment coefficient Greater than the preliminary fault judgment threshold When determining the fault of the k-th acoustic sensor, the following formula is used:

[0013]

[0014]

[0015] ;

[0016] Step 3: Feature extraction from Long Short-Term Memory (LSTM) network; The LSM network model is used to capture dependency features and complex patterns in the data from the acoustic sensor. A gating mechanism is used to extract features from the input data, obtaining the feature set H of any discrete-time series sample value in the acoustic array. s ;

[0017] Step 4: Autoencoder anomaly detection model; The feature set H of any discrete-time sequence sample value in the acoustic array is obtained using a Long Short-Term Memory (LSTM) network. s Input the autoencoder anomaly detection model to obtain the feature set H of the acoustic sensor. s Reconstruction error ;

[0018] Step 5: Determine if the acoustic sensor is malfunctioning; set the threshold for reconstruction error. Define the anomaly decision function Anomaly(x):

[0019] When Anomaly(x) is 1, the sound sensor is considered abnormal; when Anomaly(x) is 0, the sound sensor is considered normal.

[0020] Furthermore, the feature set H of the sampled values ​​in step 3 s Obtain it through the following function:

[0021]

[0022] Where Sig(x) = (1 + e -x ) -1, tanh(x) = (e x - e -x ) / (e x + e -x ) -1 , and is an element-wise multiplication, x n is defined as the sampling value of the nth discrete point of any one discrete time sequence in the M-numbered acoustic array, h n−1 is the hidden state of the (n-1)th discrete point, W ni , W hi , W nc , W hc , W nf , W hf , W no , W ho are weights, b i , b c , b f , b o are bias vectors, and i n is an input gate function. n is a candidate cell state function; f n is a forget gate function; c n−1 is the cell state of the (n-1)th discrete point; c n is the cell state of the nth discrete point; o n is an output gate function; h n is the hidden state of the nth discrete point, so that the feature set of the sampling value of any one discrete time sequence in the acoustic array can be obtained .

[0023] Further, the long short-term memory network model in step 3 can adjust the number of layers and the number of hidden units of the long short-term memory network according to application scenarios and data characteristics.

[0024] Further, the self-encoding anomaly judgment model training step in step 4 is as follows:

[0025] Step 4-1, acoustic sensor array data acquisition;

[0026] Suppose a group of M-numbered acoustic arrays, continuously collect acoustic signals at a fixed sampling frequency, convert the collected analog signals into digital signals, and through the sound signal data within a period of time, M discrete time sequences can be obtained, each sequence has a length of N;

[0027] Step 4-2, training data preparation;

[0028] Training data preparation is carried out on the normal acoustic array acoustic sensor data. The training data includes non-fault scenario data groups and fault scenario data groups of normal acoustic sensor objects.

[0029] Step 4-3, Feature extraction of Long Short-Term Memory Network;

[0030] The training data set prepared in step 4-2 is used to capture the dependency features and complex patterns in the data from the acoustic sensor using a long short-term memory network model. Feature extraction is performed on the input data through a gating mechanism to obtain the feature set H of any discrete time series sample value in the acoustic array. s ;

[0031] Step 4-4, Autoencoder model training;

[0032] Input 75% of the normal samples obtained from the data preparation of the acoustic sensor into the features extracted by the Long Short-Term Memory network to obtain the feature set H for each acoustic sensor. s The feature set H of the acoustic sensor is composed of Ts Mapping to the low-dimensional latent space z, the expression is:

[0033]

[0034] Where ELU is the activation function. ,

[0035] Reconstruct the latent space z into the original feature space. W d As the weight, b d For the bias vector:

[0036] Define reconstruction error for:

[0037] in For the i-th value in the original feature set, To reconstruct the i-th feature value in the original feature space, when the reconstruction error ζ converges, the training model is output. The training model is then validated using 25% of the training preparation data, and the autoencoded anomaly detection model is output.

[0038] Furthermore, the correlation coefficient threshold The value ranges from 0.4 to 0.6.

[0039] Furthermore, step 3 also incorporates logarithmic power spectrum for feature extraction, employing a discrete Fourier transform on x[n] to obtain the frequency domain representation X(k), which is expressed as follows, where k represents the frequency and j is the imaginary unit:

[0040] , the power spectrum P(k) is calculated:

[0041] , so as to obtain the logarithmic power spectrum L(k):

[0042] L(k)=log(P(k)), M sets of logarithmic power spectrum features of the M discrete time sequences are obtained.

[0043] Further, the reconstruction error of the reconstructed signal is calculated as The reconstruction errors of multiple time windows can also be comprehensively judged, the original discrete signal is divided according to the selected time window to obtain a series of overlapping or non-overlapping data segments, the short time window T1 starts from the marked starting moment of the signal, and the data segment is intercepted with T1 as the length, the medium time window T2 and the long time window T3 are also sequentially intercepted.

[0044] The data of the intercepted short time window T1, medium time window T2 and long time window T3 are respectively subjected to feature extraction of long short memory network and then subjected to autoencoder operation, the reconstruction errors of the discrete sequences of different time windows are calculated, the reconstruction errors of different time windows are set with weights, the weight of the short time window T1 is w1, the weight of the medium time window T2 is w2, and the weight of the long time window T3 is w3, w1+ w2+ w3=1, wherein the weight w1 is 0.1, the weight w2 is 0.2, and the weight w3 is 0.7.

[0045] The reconstruction error of the short time window is The reconstruction error of the medium time window is The reconstruction error of the long time window is The comprehensive reconstruction error is calculated as , The convergence of the comprehensive reconstruction error is used to judge whether the acoustic sensor array has an abnormality.

[0046] The application discloses an acoustic sensor array self-abnormality detection method for power grid, which combines cross-validation preliminary judgment and deep learning model analysis to solve the problems of high misjudgment rate and insufficient time sequence feature mining of traditional methods in complex acoustic environment.

[0047] The specific steps include: data acquisition: acquiring time sequence signals of the acoustic sensor array; cross-validation preliminary judgment: quickly positioning sensors with fixed bias or trend abnormality; LSTM feature extraction: capturing time sequence dependence of the signals; autoencoder reconstruction error analysis: quantifying the abnormality degree of the sensor signals; comprehensive determination: combining cross-validation and reconstruction error to improve detection accuracy; the above method produces the following beneficial effects:

[0048] Improving judgment efficiency: the initial judgment of the cross-correlation coefficient method of acoustic signals combined with the abnormal judgment method of long short memory network and autoencoder can quickly determine the abnormal situation of part of the sensors through statistical methods;

[0049] Improving the accuracy of anomaly detection: combining the advantages of long short memory network and autoencoder, long short memory network can effectively capture the time sequence characteristics and long-term dependencies of acoustic data, and autoencoder can learn the normal mode of data and reconstruct it. Through comparing the reconstruction error, the accuracy of anomaly detection is greatly improved;

[0050] Adapting to complex acoustic environment: this method learns the complex patterns of acoustic data of normal and fault conditions of field power equipment, reducing the misjudgment caused by environmental interference and data fluctuations;

[0051] Strong real-time performance: real-time data processing and abnormal judgment mechanism are adopted, which can timely discover the abnormal situation of acoustic sensors, provide support for timely measures, and avoid monitoring errors and potential risks caused by sensor abnormalities;

[0052] Good scalability: the model structure and parameters of this method can be flexibly adjusted and optimized according to different application scenarios and data characteristics, and have good scalability. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The overall flowchart of the self-abnormal detection method of the present application;

[0054] Figure 2 The normal 8-sensor array data sample;

[0055] Figure 3 The reconstruction error training result;

[0056] Figure 4 The sensor contact failure fault signal;

[0057] Figure 5 The comprehensive judgment coefficient of each sensor;

[0058] Figure 6 The reconstruction error of sensor 1 abnormality;

[0059] Figure 7 The reconstruction error of sensor 1 abnormality; DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0061] The purpose of the present application is to provide a cross-validation combined with long short-term memory network and autoencoder acoustic sensor array self-abnormality detection method. The cross-validation preliminary judgment can quickly determine the fixed deviation and trend abnormality of part of the sensors in the array. For the precision decline and other problems that may exist in the long-term use of acoustic sensors, the long short-term memory network is used to process the time series data, and the self-encoder is used to learn the characteristics. The advantages of the self-encoder improve the accuracy and reliability of the acoustic sensor self-abnormality detection, timely discover the abnormality of the acoustic sensor, reduce the monitoring error and potential risk caused by the abnormality of the sensor, and the essence is a method of acoustic sensor array abnormality determination module. It is a self-abnormality monitoring of acoustic sensor, as shown in Figure 1 The acoustic sensor array self-abnormality detection method is shown in the whole flowchart. The self-encoder abnormality judgment model is a judgment model trained by data, which can accurately detect the acoustic sensor self-failure in a complex acoustic environment in real time. It has important practical significance for ensuring the safety and reliability of the power converter valve equipment monitoring.

[0062] The specific implementation of the method is as follows:

[0063] (1) Data acquisition and preprocessing

[0064] Taking the acoustic sensor in the industrial equipment operation monitoring as an example, a high-sensitivity microphone acoustic sensor is used, the sampling frequency is set to 51200Hz, the sound signals in the equipment operation process are continuously collected, 8 sensors are set as a group, and the normal acoustic signals of the substation converter valve are collected as shown in Figure 2 The data collected is stored in the local solid state disk.

[0065] (2) Long short-term memory network (LSTM) feature extraction

[0066] The feature set H of the sampling value s is obtained by the following function:

[0067]

[0068] Wherein, Sig(x)=(1+e -x ) -1 , tanh(x)=(e x - e -x ) (e x + e -x ) -1 , and is the element-wise multiplication. x n is defined as the sampling value of the nth discrete point of any one discrete time series in the acoustic array with M number of elements, h n−1 is the hidden state of the n-1th discrete point, Wni W hi W nc W hc W nf W hf W no W ho These are the weights, b i b c b f b o It is the bias vector, i n For input gate functions; n f is the candidate cell state function; n Forget gate function; c n−1 The cell state at the (n-1)th discrete point; c n The cell state at the nth discrete point; o n h is the output gate function. n The hidden state of the nth discrete point allows us to obtain the feature set of any discrete-time sequence sample value in the acoustic array. .

[0069] A three-layer LSTM network was constructed, with 20,000 neurons in the input layer, 128 units each in the first and second hidden layers, and 64 neurons in the output layer. The LSTM model was trained using the training set data, with mean squared error (MSE) as the loss function, the Adam optimizer as the optimizer, a learning rate of 0.001, and 50 training epochs. During training, the model was evaluated and parameters were tuned using the validation set data. After training, all data sequences were input into the LSTM model, and the outputs of the hidden layers were extracted as temporal features of the acoustic data.

[0070] (3) Training of autoencoder anomaly detection model

[0071] Input 75% of the normal samples obtained from the data preparation of the acoustic sensor into the features extracted by the Long Short-Term Memory network to obtain the feature set H for each acoustic sensor. s The feature set H of the acoustic sensor is composed of Ts Mapping to a low-dimensional latent space The expression is:

[0072] Where ELU is the activation function. ,

[0073] Reconstruct the latent space z into the original feature space. W d As the weight, b d For the bias vector:

[0074] ; define reconstruction error as:

[0075] wherein is the i-th value of the original feature set, is the i-th feature value of the reconstructed original feature space when the reconstruction error converges, output the training model, and use 25% of the training preparation data to validate the training model to output the auto-encoding anomaly judgment model.

[0076] An auto-encoder model is constructed, the encoder consists of two fully connected layers, which respectively map the 64-dimensional LSTM features to 32-dimensional and 16-dimensional latent space; the decoder consists of two fully connected layers to reconstruct the 64-dimensional original features from the 16-dimensional latent space features. The long short-term memory network features of the training set are used to train the auto-encoder, and the mean square error (MSE) is used as the loss function, and the Adam optimizer is used for parameter update, the learning rate is set to 0.001, and the training rounds are 300 rounds. The reconstruction error is calculated using the validation set data, and the training set and validation set reconstruction error graph obtained is as shown in Figure 3 . The mean value of the reconstruction error plus 3 times the standard deviation is used as the anomaly threshold.

[0077] (4) Real-time anomaly detection

[0078] Set the threshold of the reconstruction error , set the anomaly decision function Anomaly(x):

[0079] When Anomaly(x) is 1, it is determined that the acoustic sensor is abnormal, and when Anomaly(x) is 0, it is determined that the acoustic sensor is normal.

[0080] Real-time data acquisition for the sound sensor of the power grid converter station converter valve, first cross-validation, when the first acoustic sensor connection line contact is set to be poor, the detection data is as shown in Figure 4 . Cross-validation is performed, when is 0.5, is 0.7, the comprehensive judgment coefficient graph as shown in Figure 5 can be obtained, and it can be determined that sensor 1 is abnormal.

[0081] When the sensor 1 precision degradation problem occurs, the comprehensive judgment coefficient method cannot be solved, and the method of using long short memory network and encoder is used for inheritance judgment. The data sequence is input into the trained long short memory network to extract the time sequence feature, and then the feature is input into the trained autoencoder for reconstruction. The reconstruction error of the real-time data is calculated, compared with the pre-determined abnormal threshold value, the threshold value is 0.05, the sensor 1 fault is judged, and the reconstruction errors of the 8 sensors are as shown in Figure 6 When the 8 sensors are all normal, the reconstruction errors of the 8 sensors are as shown in Figure 7 .

[0082] The above method has the following beneficial effects:

[0083] Using the initial judgment of cross-validation, the abnormal situation of individual sensor can be quickly judged by using the simplest mathematical model, and the efficiency of sensor self-checking is improved.

[0084] Combining long short memory network and autoencoder: the application innovatively combines cross-validation, long short memory network (LSTM) and autoencoder, fully utilizes the advantages of LSTM in processing time series data and the feature learning and reconstruction ability of autoencoder, and can more accurately capture the time sequence features and normal patterns of acoustic data, and improve the accuracy of anomaly detection.

[0085] Meanwhile, it is not limited to the above embodiments, and the model structure can be adaptively adjusted. According to specific application scenarios and data characteristics, the number of layers, the number of hidden units of the long short memory network and the structure and the number of layers of the autoencoder can be adjusted to optimize the performance of the model.

[0086] In addition to the time sequence features extracted by the long short memory network, other features such as frequency domain features and time-frequency domain features can also be fused to further improve the accuracy of anomaly detection. In the frequency domain feature, the log power spectrum can be preferably used for feature extraction.

[0087] The acoustic sensor array data acquisition part is used as above, a group of acoustic arrays with M number of elements is set, and the acoustic signals are continuously collected at a fixed sampling frequency. The analog signals collected are converted into digital signals, and through the sound signal data in a period of time, M discrete time sequences x1[n], x2[n], ···, x M [n] can be obtained, and the length of each sequence is N.

[0088] Discrete Fourier transform is performed on x [n] to obtain the frequency domain representation X(k) represented by the following formula, wherein k represents frequency, and j is a virtual unit:

[0089] The power spectrum P(k) is calculated as follows:

[0090] Thus, the logarithmic power spectrum L(k) is obtained:

[0091]

[0092] Thus, M sets of logarithmic power spectrum features of the M discrete time sequences are obtained.

[0093] In terms of optimization of the anomaly determination strategy, a more complex anomaly determination strategy can be adopted, such as comprehensive judgment in combination with reconstruction errors of multiple time windows, or introduction of a dynamic threshold adjustment mechanism to dynamically adjust the anomaly threshold according to real-time changes of data.

[0094] Comprehensive judgment of reconstruction errors of multiple time windows is to select three time windows of different lengths, a short time window T1, a medium time window T2 and a long time window T3. T1 is 10 sampling periods in the present detection method, T2 is 100 sampling periods, and T3 is 1000 sampling periods.

[0095] The original discrete signal is divided according to the selected time windows to obtain a series of overlapping or non-overlapping data segments. For the short time window T1, a data segment is cut off from the marked starting time of the signal with T1 as the length, and the medium time window T2 and the long time window T3 are also sequentially cut off.

[0096] The data of the cut-off short time window, medium time window and long time window are respectively subjected to feature extraction of the long short-term memory network and then subjected to the self-encoding operation module to calculate the reconstruction errors of the discrete sequences of different time windows. The reconstruction errors of different time windows are set with weights to reflect their importance in anomaly determination. Unlike sound signal fault diagnosis, self-anomaly detection of the acoustic sensor can reflect long-term trends and overall changes of the sensor, and the weight can be relatively high, while the medium and short time windows are low. Let the weight of the short time window T1 be w1, the weight of the medium time window T2 be w2, and the weight of the long time window T3 be w3. The recommended values are

[0097] w1+ w2+ w3=1, the recommended value of w1 is 0.1, w2 is 0.2, and w3 is 0.7.

[0098] The reconstruction error of the short time window is , the reconstruction error of the medium time window is , and the reconstruction error of the long time window is . Thus, the comprehensive reconstruction error is calculated as , , and the reconstruction error is used to judge whether the acoustic sensor array has self-anomaly.

Claims

1. A method for power grid acoustic sensor array self anomaly detection, characterized by, The method comprises the following steps: Step 1: Acoustic sensor array data acquisition; a set of M acoustic arrays is set, and the acoustic signals are continuously collected at a fixed sampling frequency. The collected analog signals are converted into digital signals. Through the sound signal data within a period of time, M discrete time sequences x1[n], x2[n], ···, x M [n] can be obtained, and the length of each sequence is N; Step 2: initial judgment by sound signal cross-correlation coefficient method; for the kth sound sensor to be verified, the cross-correlation coefficient of the kth sound sensor with the signals of the remaining M-1 sound sensors is calculated , i≠k, i=1, 2…M, M-1 cross-correlation coefficients can be obtained, when greater than the set correlation coefficient threshold , , the count is 1, and the comprehensive fault judgment coefficient is obtained after the count sum of the M-1 cross-correlation coefficients is calculated , the preliminary fault judgment threshold is set , when the comprehensive fault judgment coefficient is greater than the preliminary fault judgment threshold , the kth sound sensor is determined to be faulty, and the following formula is used, ； Step 3: Feature extraction of long short-term memory network; the long short-term memory network model is used to capture the dependent features and complex patterns in the data in the acoustic sensor, and the input data is extracted through the gating mechanism to obtain the feature set H of any one discrete time sequence sampling value in the acoustic array s ; Step 4: Auto-encoding anomaly detection model; obtain the feature set H of any one discrete time sequence sampling value in the acoustic array by using a long short-term memory network s Input the auto-encoding anomaly detection model to obtain the reconstruction error of the feature set H of the acoustic sensor s ;​ Step 5: Sound sensor abnormality determination; threshold value of reconstruction error is set , set the anomaly decision function Anomaly(x): , When Anomaly(x) is 1, it is determined that the acoustic sensor is abnormal, and when Anomaly(x) is 0, it is determined that the acoustic sensor is normal.

2. The method of claim 1, wherein The feature set H of the sampled value in step 3 s Obtained by the following function: wherein Sig(x) = (1 + e -x ) -1 , tanh(x) = (e x - e -x ) (e x + e -x ) -1 , and is an element-wise multiplication, x n is a sampling value of an n-th discrete point of any one discrete time sequence in the acoustic array with M elements, h n−1 is a hidden state of an (n-1)-th discrete point, W ni , W hi , W nc , W hc , W nf , W hf , W no , W ho are weights, b i , b c , b f , b o are bias vectors, and i n is an input gate function; is a candidate cell state function; f n is a forget gate function; c n−1 is a cell state of the (n-1)-th discrete point; c n is a cell state of the n-th discrete point; o n is an output gate function; h n is a hidden state of the n-th discrete point, and thus a feature set of the sampling value of any one discrete time sequence in the acoustic array can be obtained.

3. The method according to claim 2, wherein The long short-term memory network model in the step 3 can adjust the number of layers and the number of hidden units of the long short-term memory network according to application scenarios and data characteristics.

4. The method according to any one of claims 1 to 3, wherein The self-encoding anomaly judgment model training step in the step 4 is as follows: Step 4-1, acoustic sensor array data acquisition; Let a set of M acoustic arrays be set, continuously collect acoustic signals at a fixed sampling frequency, convert the collected analog signals into digital signals, and through the sound signal data within a period of time, M discrete time sequences, x1[n], x2[n], ···, x M [n] can be obtained, and the length of each sequence is N. Step 4-2, training data preparation; The normal acoustic array acoustic sensor data is prepared for training, and the training data comprises normal acoustic sensor object non-fault scene data groups and normal acoustic sensor object fault scene data groups; Step 4-3, long short-term memory network feature extraction; The long short-term memory network model is used for capturing dependent features and complex patterns in the data in the sound sensor, feature extraction is performed on the input data through a gating mechanism, and a feature set H of any one discrete time sequence sampling value in the sound array is obtained s ; Step 4-4, self-encoder model training; 75% of the normal sample number of the sound sensor obtained by data preparation is input into the feature extracted by the long short memory network to obtain the feature set H of each sound sensor s , and the sound sensor feature set H is composed Ts and mapped to a low-dimensional latent space , and the expression is: wherein ELU is an activation function , and the latent space z is reconstructed into the original feature space s wherein W d is a weight, and b d is a bias vector , and the reconstruction error is defined as follows: wherein is the i th value of the original feature set, is the i th feature value of the reconstructed original feature space, and when the reconstruction error converges, the trained model is output, 25% of the training preparation data is verified on the trained model, and the self-encoding anomaly judgment model is output.

5. The method of claim 4, wherein the correlation coefficient threshold is 0.4~0.

6.

6. The method of claim 5, wherein In step 3, the log power spectrum is also fused for feature extraction. The discrete Fourier transform is performed on x [n] to obtain the frequency domain representation X(k) as follows, where k represents frequency and j is the imaginary unit: The power spectrum P(k) is calculated as follows: The log power spectrum L(k) is obtained as follows: M sets of log power spectrum features of the M discrete time sequences are obtained.

7. The method of claim 6, wherein The reconstruction error The original discrete signal is divided according to the selected time window to obtain a series of overlapping or non-overlapping data segments, a short time window T1 starts from the marked starting moment of the signal, a data segment is cut off with T1 as the length, and a medium time window T2 and a long time window T3 are also sequentially cut off. The data of the short time window T1, the medium time window T2 and the long time window T3 are respectively subjected to long short-term memory network feature extraction and then subjected to self-encoder operation, the reconstruction error of different time window discrete sequences is calculated, the weights of different time window reconstruction errors are set, the weight of the short time window T1 is w1, the weight of the medium time window T2 is w2, and the weight of the long time window T3 is w3, w1+ w2+ w3=1, wherein the weight w1 is 0.1, the weight w2 is 0.2, and the weight w3 is 0.

7. The short-time window reconstruction error is obtained as The medium-time window reconstruction error is obtained as The long-time window reconstruction error is obtained as The comprehensive reconstruction error is calculated as , The self-abnormal situation of the acoustic sensor array is judged by the convergence of the comprehensive reconstruction error .

Images