Self anomaly detection method for acoustic sensor array for power grid

Abstract

The invention provides a self-anomaly detection method of an acoustic sensor array for a power grid, which comprises the following steps of: through a method of combining cross validation with a long-short term memory network and an auto-encoder, performing primary judgment by using the cross validation, processing time sequence data by using the long-short term memory network and further judging by using the feature learning capability of the auto-encoder, and timely finding the abnormal condition of an acoustic sensor; the accuracy and reliability of anomaly detection of the acoustic sensor are improved, and monitoring errors and potential risks caused by anomaly of the acoustic sensor are reduced; the method comprises the following steps: step 1, acoustic sensor array data acquisition; step 2, carrying out primary judgment by a sound signal cross correlation coefficient method; step 3, extracting features of the long and short memory network; 4, a self-coding abnormity judgment model is established; and 5, judging whether the acoustic sensor is abnormal or not.

Description

Technical Field

[0001] The present invention relates to the technical field of acoustic sensor detection, and particularly relates to a method for detecting self-abnormalities of an acoustic sensor array for power grids. Background Art

[0002] In a power system, the safe and stable operation of power equipment is of crucial importance. As a core component of a DC transmission system, the operating state of a power converter valve device directly affects the reliability and stability of the entire power system. In order to monitor the operating state of power equipment in real time and detect potential faults in a timely manner, acoustic sensors and their arrays are widely used in aspects such as partial discharge detection and positioning of power equipment. An acoustic sensor can capture the acoustic signals generated during the operation of the equipment. By analyzing and processing these signals, the operating state information of the equipment can be obtained, providing an important basis for equipment maintenance and fault diagnosis.

[0003] Although acoustic sensors play an important role in power equipment monitoring, during long-term use, due to the influence of various factors such as environmental interference, equipment aging, and component failures, the acoustic sensor may exhibit abnormal conditions, resulting in inaccurate or unreliable collected data. The traditional methods for detecting self-abnormalities of acoustic sensors have the following limitations:

[0004] Threshold judgment method: This method sets fixed upper and lower threshold values, and when the sensor data exceeds this range, it is determined to be abnormal. However, the actual acoustic environment is complex and variable, and the acoustic signal characteristics vary greatly under different equipment and operating conditions. Fixed threshold values are difficult to adapt to various situations, and false positives or false negatives are likely to occur. For example, under certain special operating conditions, normal acoustic signals may exceed the set threshold values, resulting in false positives; while in some cases of slowly changing faults, the signal may gradually deviate from the normal range but not exceed the threshold, resulting in false negatives.

[0005] Statistical analysis method: Although it considers the statistical characteristics of the data, for acoustic data with temporal characteristics, it insufficiently mines the long-term dependence relationships contained in the data. An acoustic signal is a dynamic process that changes over time, and there are complex temporal correlations within it. Traditional statistical analysis methods often only focus on statistical quantities such as the mean and variance of the data, while ignoring the evolution law of the data in the time dimension, making it difficult to accurately capture the dynamic changes of the data, thereby affecting the accuracy and timeliness of anomaly detection. For example, when an early fault occurs in the equipment, the acoustic signal may exhibit weak and gradually increasing abnormal changes, and traditional statistical analysis methods may not be able to detect this change in a timely manner, resulting in delays in fault diagnosis.

[0006] The actual operating environment of power equipment is usually complex. The complex acoustic environment poses higher requirements for detection methods, and there are various noise interferences, such as mechanical vibration noise, electromagnetic interference noise, etc. These noises are superimposed on the normal acoustic signals of the equipment, making the characteristics of the acoustic signals more complex and difficult to distinguish. In this case, it is very difficult for traditional self-abnormality detection methods of acoustic sensors to effectively distinguish normal signals from abnormal signals and cannot meet the detection requirements in complex acoustic environments.

[0007] The power converter valve equipment is a key equipment in the DC transmission system. The accurate monitoring of its operating state is crucial for ensuring the safe and stable operation of the power system. Once the acoustic sensor fails, resulting in inaccurate or unreliable collected data, it will directly affect the judgment of the operating state of the power converter valve equipment, may lead to incorrect maintenance decisions, and even cause missed judgments and delayed handling of equipment failures, posing serious safety hazards to the power system. Therefore, developing an effective self-detection method for acoustic sensor arrays, which can detect the faults of acoustic sensors in complex acoustic environments in real time and accurately, has important practical significance for ensuring the safety and reliability of the monitoring of power converter valve equipment.

[0008] In summary, in view of the limitations of traditional self-abnormality detection methods of acoustic sensors and the special requirements of complex acoustic environments and power equipment monitoring, developing a more advanced and effective self-detection method for acoustic sensor arrays has become an urgent problem to be solved. Summary of the Invention

[0009] The present invention aims at the technical defects of the existing technology and provides a method for self-abnormality detection of an acoustic sensor array for a power grid. By combining cross-validation with a long short-term memory network and an autoencoder, using cross-validation for initial judgment, the long short-term memory network for processing time series data and the feature learning ability of the autoencoder for further judgment, the abnormal conditions of the acoustic sensor can be detected in time, the accuracy and reliability of self-abnormality detection of the acoustic sensor can be improved, and the monitoring errors and potential risks caused by the abnormality of the acoustic sensor can be reduced.

[0010] The present invention provides the following technical solution: A method for self-abnormality detection of an acoustic sensor array for a power grid, the method comprising the following steps:

[0011] Step 1: Data acquisition of the acoustic sensor array; Suppose a set of acoustic arrays with a number of M, 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 series can be obtained. , and the length of each sequence is N;

[0012] Step 2: Preliminary judgment by cross-correlation coefficient method of acoustic signals; For the k-th acoustic sensor to be verified, calculate the cross-correlation coefficients between its signal and the signals of the remaining M - 1 acoustic sensors , and M - 1 cross-correlation coefficients can be obtained. When it is greater than the set correlation coefficient threshold , the count is 1, and after summing up the counts of the M - 1 cross-correlation coefficients, the comprehensive fault judgment coefficient is obtained. Set the preliminary fault judgment threshold . When the comprehensive fault judgment coefficient is greater than the preliminary fault judgment threshold , it is determined that the k-th acoustic sensor is faulty. The formula is as follows,

[0013]

[0014]

[0015] ;

[0016] Step 3: Feature extraction using long short-term memory network; Use the long short-term memory network model to capture the dependency features and complex patterns in the data of the acoustic sensor, and perform feature extraction on the input data through the gating mechanism to obtain the feature set H of any discrete-time series sampling value in the acoustic array s ;

[0017] Step 4: Autoencoder anomaly judgment model; Input the feature set H of any discrete-time series sampling value in the acoustic array obtained by the long short-term memory network s into the autoencoder anomaly judgment model to obtain the reconstruction error s of the feature set H of this acoustic sensor ;

[0018] Step 5: Judgment on whether the acoustic sensor is abnormal; Set the threshold of the reconstruction error , and set the anomaly decision function Anomaly(x):

[0019] . When Anomaly(x) is 1, it is determined that the acoustic sensor is abnormal. When Anomaly(x) is 0, it is determined that the acoustic sensor is normal.

[0020] Furthermore, the feature set H of the sampling value in Step 3 s is obtained through the following function:

[0021]

[0022] where Sig(x)=(1 + e -x ) -1, tanh(x)=(e x - e -x ) (e x + e -x ) -1 , ⊙ is element-wise multiplication, x n is defined as the sampling value of the n-th discrete point of any one discrete-time sequence in a sound array with M elements, 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 respectively, b i , b c , b f , b o are bias vectors, i n is the input gate function; n is the candidate cell state function; f n is the 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 n-th discrete point; o n is the output gate function; h n is the hidden state of the n-th discrete point, from which the feature set of the sampling values of any one discrete-time sequence in the sound array can be obtained .

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

[0024] Further, the training steps of the autoencoder anomaly judgment model in step 4 are as follows:

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

[0026] Suppose a set of sound arrays with M elements, continuously acquire acoustic signals at a fixed sampling frequency, convert the acquired analog signals into digital signals, and through the sound signal data within a period of time, M discrete-time sequences can be obtained, , and the length of each sequence is N;

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

[0028] Prepare training data for the normal sound array acoustic sensor data. The training data includes the non-fault scenario data group of the normal acoustic sensor object and the fault scenario data group of the normal acoustic sensor object;

[0029] Step 4-3, feature extraction of the long short-term memory network;

[0030] Using the long short-term memory network model to capture the dependent features and complex patterns in the data of the acoustic sensor for the training data group prepared in Step 4-2, and extracting features from the input data through the gating mechanism to obtain the feature set H of any discrete time series sampling value in the acoustic array s ;

[0031] Step 4-4, training of the autoencoder model;

[0032] Input 75% of the data of the number of normal samples of the acoustic sensor obtained from data preparation and the features extracted by the long short-term memory network to obtain the feature set H of each acoustic sensor s , and form the acoustic sensor feature set H Ts Map it to the low-dimensional latent space z, and the expression is:

[0033]

[0034] where ELU is the activation function ,

[0035] Reconstruct the latent space z into the original feature space , where W d is the weight, and b d is the bias vector:

[0036] ; Define the reconstruction error as:

[0037] where 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 ζ converges, output the training model, and use 25% of the training preparation data to verify the training model, and output the autoencoding anomaly judgment model.

[0038] Furthermore, the correlation coefficient threshold takes values in the range of 0.4 to 0.6.

[0039] Furthermore, in Step 3, logarithmic power spectrum is also used for feature extraction in combination. Perform discrete Fourier transform on x[n] to obtain the frequency domain representation X(k) as shown in the following formula, where k represents frequency and j is the imaginary unit:

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

[0041] , thus obtaining the logarithmic power spectrum L(k):

[0042] L(k) = log(P(k)), obtaining M sets of logarithmic power spectrum features of M discrete time series.

[0043] Furthermore, the reconstruction error can also be comprehensively judged by combining the reconstruction errors of multiple time windows. 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 T 1 starting from the marked starting moment of the signal, intercept the data segment with length T 1 The medium - time window T 2 and the long - time window T 3 also perform sequential interception operations;

[0044] For the intercepted short - time window T 1 , medium - time window T 2 , long - time window T 3 After extracting the features of the long short - term memory network from the data respectively, perform auto - encoder operations, calculate the reconstruction errors of discrete sequences in different time windows, set weights for the reconstruction errors of different time windows. Let the weight of the short - time window T 1 be w 1 , the weight of the medium - time window T 2 be w 2 , the weight of the long - time window T 3 be w 3 , w 1 + w 2 + w 3 = 1, where the weight w 1 is 0.1, w 2 is 0.2, w 3 is 0.7;

[0045] Obtain the reconstruction error of the short - time window as , the reconstruction error of the medium - time window as , the reconstruction error of the long - time window as , calculate the comprehensive reconstruction error , , judge whether the acoustic sensor array has its own abnormal situation through the convergence of the comprehensive reconstruction error .

[0046] The present invention discloses a method for detecting self - anomalies of an acoustic sensor array for power grids. This method combines cross - validation preliminary judgment and deep - learning model analysis to solve the problems of high false - positive rate and insufficient mining of time - series features in traditional methods in complex acoustic environments;

[0047] The specific steps include: Data acquisition: Obtain the time - series signals of the acoustic sensor array; Cross - validation preliminary judgment: Quickly locate sensors with fixed deviations or trend anomalies; LSTM feature extraction: Capture the time - series dependencies of the signals; Auto - encoder reconstruction error analysis: Quantify the anomaly degree of the sensor signals; Comprehensive determination: Combine cross - validation and reconstruction error to improve the detection accuracy; The following beneficial effects are produced by the above method:

[0048] Improve the determination efficiency: The initial judgment method of the acoustic signal cross - correlation coefficient method combined with the anomaly judgment methods of long short - term memory network and auto - encoder can quickly judge the abnormal situation of some sensors through statistical methods;

[0049] Improve the accuracy of anomaly detection: Combining the advantages of long short - term memory network and auto - encoder, the long short - term memory network can effectively capture the time - series features and long - term dependencies of acoustic data, and the auto - encoder can learn the normal patterns of data and reconstruct them. Anomaly detection is realized by comparing the reconstruction errors, greatly improving the accuracy of anomaly detection;

[0050] Adapt to complex acoustic environments: This method learns the complex patterns of acoustic data in normal and faulty situations of on - site power equipment, reducing false positives caused by environmental interference and data fluctuations;

[0051] Strong real - time performance: Adopting real - time data processing and anomaly determination mechanisms can timely detect anomalies of acoustic sensors, provide support for taking timely measures, and avoid monitoring errors and potential risks caused by sensor anomalies;

[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, with good scalability. Brief Description of the Drawings

[0053] Figure 1 It is the overall flowchart of the self - anomaly detection method of the present invention;

[0054] Figure 2 It is a data sample of a normal 8 - sensor array;

[0055] Figure 3 It is the training result of the reconstruction error;

[0056] Figure 4 It is the faulty signal of poor sensor contact;

[0057] Figure 5is the comprehensive judgment coefficient for each sensor;

[0058] Figure 6 is the reconstruction error when sensor 1 is abnormal;

[0059] Figure 7 is the reconstruction error when none of the sensors are abnormal. Specific implementation manner

[0060] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0061] The purpose of the present invention is to provide a method for detecting self - anomalies of an acoustic sensor array by combining cross - validation with long short - term memory networks and autoencoders. Using cross - validation for initial judgment can quickly determine fixed deviations and trend anomalies of some sensors in the array. For problems such as possible accuracy degradation during the long - term use of acoustic sensors, the advantages of long short - term memory networks in processing time - series data and the feature - learning ability of autoencoders are utilized to improve the accuracy and reliability of self - anomaly detection of acoustic sensors, timely discover abnormal situations of acoustic sensors, reduce monitoring errors and potential risks caused by sensor anomalies. Essentially, it is a method for an acoustic sensor array anomaly determination module, which is the self - anomaly monitoring of acoustic sensors. As Figure 1 shown in the overall flowchart of the method for detecting self - anomalies of the power grid acoustic sensor array, the auto - encoding anomaly judgment model therein is a judgment model trained through data, which can detect the self - faults of acoustic sensors in real - time and accurately in a complex acoustic environment, and has important practical significance for ensuring the safety and reliability of power converter valve equipment monitoring.

[0062] The specific embodiments of this method are as follows:

[0063] (1) Data acquisition and pre - processing

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

[0065] (2) Feature extraction of long short - term memory network (LSTM)

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

[0067]

[0068] Among them, Sig(x) = (1 + e -x ) -1 , tanh(x) = (e x - e -x ) / (e x + e -x ), ⊙ is element-wise multiplication, x -1 is defined as the sampling value of the nth discrete point of any one discrete-time sequence in an acoustic array with M elements, h n is the hidden state of the (n - 1)th discrete point, W n−1 , W ni , W hi , W nc , W hc , W nf , W hf , W no , W ho are weights respectively, b i , b c , b f , b o are bias vectors, i n is the input gate function; n is the candidate cell state function; f n is the 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 the output gate function; h n is the hidden state of the nth discrete point, from which the feature set of the sampling value of any one discrete-time sequence in the acoustic array can be obtained .

[0069] Construct a three-layer LSTM network with 20,000 neurons in the input layer, 128 units in both the first and second LSTM hidden layers, and 64 neurons in the output layer. Use the training set data to train the LSTM model, select the mean squared error (MSE) as the loss function, select the Adam optimizer, set the learning rate to 0.001, and the number of training epochs to 50. During the training process, use the validation set data for model evaluation and parameter adjustment. After training, input all data sequences into the LSTM model and extract the output of the hidden layer as the temporal features of the acoustic data.

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

[0071] Input 75% of the data of the number of normal samples of the acoustic sensor obtained from data preparation into the features extracted by the long short-term memory network, and obtain the feature set H of each acoustic sensor. s to form the acoustic sensor feature set H. Ts Map it 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. where W d is the weight and b d is the bias vector:

[0074] ; Define the reconstruction error as:

[0075] where is the i-th value of the original feature set, is the i-th eigenvalue of the reconstructed original feature space. When the reconstruction error converges, output the trained model, and use 25% of the training preparation data to verify the trained model, and output the autoencoder anomaly judgment model.

[0076] Construct an autoencoder model. The encoder consists of two fully connected layers, which map the 64-dimensional LSTM features to 32-dimensional and 16-dimensional latent spaces respectively; the decoder consists of two fully connected layers to reconstruct the 16-dimensional latent space features into 64-dimensional original features. Use the long short-term memory network features of the training set to train the autoencoder. Similarly, the mean squared error (MSE) is used as the loss function, and the Adam optimizer is used to update the parameters. The learning rate is set to 0.001, and the number of training epochs is 300. Calculate the reconstruction error using the validation set data, and the reconstruction error graphs of the training set and the validation set are as shown in Figure 3 shown. Add the mean of the reconstruction error plus 3 times the standard deviation as the anomaly threshold.

[0077] (4) Real-time anomaly detection

[0078] Set the threshold of the reconstruction error , and 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 collect data for the acoustic sensors of the converter valves in the power grid substation. First, perform cross-validation. When setting the first acoustic sensor connection wire to have a poor contact fault, the detected data is as Figure 4 shown. Perform cross-validation. When setting to 0.5, to 0.7, the comprehensive judgment coefficient graph as Figure 5 can be obtained, and it can be determined that sensor 1 is abnormal.

[0081] When there is a problem of the accuracy decline of sensor 1, the comprehensive judgment coefficient method cannot solve it, and the method of using long short-term memory network and autoencoder is used for legacy judgment. Input the data sequence into the trained long short-term memory network to extract the time series features, and then input the features into the trained autoencoder for reconstruction. Calculate the reconstruction error of the real-time data and compare it with the pre-determined anomaly threshold, which is 0.05. It is determined that sensor 1 is faulty. The reconstruction errors of 8 sensors are as Figure 6 shown. When all 8 sensors are fault-free, the reconstruction errors of 8 sensors are as Figure 7 shown.

[0082] The following beneficial effects are produced by the above method:

[0083] Using the initial judgment of cross-validation, it can quickly judge the abnormality of individual sensors with the simplest mathematical model, improving the efficiency of sensor self-checking.

[0084] Combining long short-term memory network and autoencoder: The present invention innovatively combines cross-validation with long short-term memory network (LSTM) and autoencoder, giving full play to the advantages of LSTM in processing time series data and the feature learning and reconstruction ability of the autoencoder, and being able to more accurately capture the time series features and normal patterns of acoustic data, improving the accuracy of anomaly detection.

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

[0086] In terms of feature fusion, in addition to the time series features extracted by the long short-term 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 features, logarithmic power spectrum can be preferably used for feature extraction.

[0087] Following the above-mentioned part of the acoustic sensor array data acquisition, assume a set of acoustic arrays with the number of M. Continuously collect the acoustic signals at a fixed sampling frequency, convert the collected analog signals into digital signals. Through the sound signal data within a period of time, M discrete time series, x 1 [n], x2 [n], ···, x M [n], and the length of each sequence is N.

[0088] Perform a discrete Fourier transform on x[n], and the frequency-domain representation X(k) is expressed as follows, where k represents frequency and j is the imaginary unit:

[0089] Calculate the power spectrum P(k):

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

[0091]

[0092] From this, M sets of logarithmic power spectrum features of M discrete-time sequences can be obtained.

[0093] In terms of optimizing the anomaly detection strategy, a more complex anomaly detection strategy can be adopted, such as combining the reconstruction errors of multiple time windows for comprehensive judgment, or introducing a dynamic threshold adjustment mechanism to dynamically adjust the anomaly threshold according to the real-time changes of the data.

[0094] For comprehensive judgment of the reconstruction errors of multiple time windows, three different-length time windows are selected, a short time window T 1 , a medium time window T 2 and a long time window T 3 . T 1 is 10 sampling periods in this detection method, T 2 is 100 sampling periods, and T 3 is 1000 sampling periods.

[0095] Divide the original discrete signal according to the selected time window to obtain a series of overlapping or non-overlapping data segments. For the short time window T 1 Starting from the marked start time of the signal, intercept data segments with a length of T 1 , and perform the same interception operations for the medium time window T 2 and the long time window T 3 in turn.

[0096] After extracting the features of the long short-term memory network for the data of the intercepted short time window, medium time window, and long time window respectively, perform an auto-encoding operation module to calculate the reconstruction errors of discrete sequences in different time windows. Set weights for the reconstruction errors of different time windows to reflect their importance in anomaly detection. Different from acoustic signal fault diagnosis, the self-anomaly detection of acoustic sensors can reflect the long-term trend and overall changes of the sensors, and the weights can be relatively high, while the medium and short time windows are relatively low. Let the weight of the short time window T 1 be w1 Medium time window T 2 with weight w 2 Long time window T 3 with weight w 3 . Suggested value

[0097] w 1 + w 2 + w 3 = 1. The suggested value of weight w 1 is 0.1, w 2 is 0.2, w 3 is 0.7.

[0098] Then the reconstruction error of the short time window can be obtained as , 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, and the convergence of this reconstruction error is used to determine whether there is an abnormal situation in the acoustic sensor array itself.

Claims

1. A method for detecting abnormality of an acoustic sensor array for a power grid, characterized in that: The method comprises the following steps: Step 1: Acoustic sensor array data acquisition: Assume a set of M acoustic arrays, continuously acquire acoustic signals at a fixed sampling frequency, convert the acquired analog signals into digital signals, and obtain M discrete time series, x1[n], x2[n], ···, x, through the sound signal data over a period of time. M [n], the length of each sequence is N; Step 2: Initial judgment using the acoustic signal cross-correlation coefficient method; for the kth acoustic sensor to be verified, calculate the cross-correlation coefficient ρ between it and the remaining M−1 acoustic sensor signals ki (i≠k, i=1,2…M), M-1 cross-correlation coefficients can be obtained. When it is greater than the set correlation coefficient threshold ρ th hour, The count is 1, and the M-1 cross-correlation coefficients are counted and summed to obtain the comprehensive fault judgment coefficient p ki , set the initial fault judgment threshold p th , when the comprehensive fault judgment coefficient p ki Greater than the initial fault judgment threshold p th When the kth acoustic sensor is judged to be faulty, the formula is as follows: ， ， ； Step 3: Feature extraction of LSTM network: Use the LSTM network model to capture the dependent features and complex patterns in the data of the acoustic sensor, extract the features of the input data through the gating mechanism, and obtain the feature set H of any discrete time series sampling value in the acoustic array. s ; Step 4: Autoencoder abnormality judgment model; use the long short-term memory network to obtain the feature set H of any discrete time series sampling value in the acoustic array s Input the autoencoder abnormality judgment model to obtain the acoustic sensor feature set H s The reconstruction error ; Step 5: Determine whether the acoustic sensor is abnormal; set the threshold of reconstruction error , set the abnormal decision function Anomaly(x): ， When Anomaly(x) is 1, the acoustic sensor is determined to be abnormal, and when Anomaly(x) is 0, the acoustic sensor is determined to be normal.

2. A reliable routing method for wireless sensor network node failure according to claim 1, characterized in that: The feature set H of the sampled values ​​in step 3 s Obtained through the following function: ， ， ， ， ， ， Where Sig(x)=(1+e -x ) -1 , tanh(x)=(e x -e -x ) (e x + e -x ) -1 , ⊙ is element-by-element multiplication, x n It is defined as the sampling value of the nth discrete point in any discrete time sequence of the M-number acoustic array, h n−1 is the hidden state of the n-1th 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 is the bias vector, i n is the input gate function; n is the candidate cell state function; f n is the forget gate function; c n−1 is the cell state of the n-1th discrete point; c n is the cell state of the nth discrete point; o n is the output gate function; h n is the hidden state of the nth discrete point, from which the feature set of any discrete time series sampling value in the acoustic array can be obtained .

3. A method for detecting abnormality of an acoustic sensor array for a power grid according to claim 2, characterized in that: The long short-term memory network model in step 3 can adjust the number of layers and hidden units of the long short-term memory network according to the application scenario and data characteristics.

4. A method for detecting abnormality of an acoustic sensor array for a power grid according to any one of claims 1 to 3, characterized in that: The steps for training the autoencoder anomaly judgment model in step 4 are as follows: Step 4-1, acoustic sensor array data acquisition; Suppose a group of M acoustic arrays are used to continuously collect acoustic signals at a fixed sampling frequency, and the collected analog signals are converted into digital signals. Through the sound signal data over a period of time, M discrete time series can be obtained, x1[n], x2[n], ···, x M [n], the length of each sequence is N; Step 4-2, training data preparation; Preparing training data for normal acoustic array acoustic sensor data, the training data including a normal acoustic sensor object non-fault scenario data group and a normal acoustic sensor object fault scenario data group; Step 4-3, feature extraction of long short-term memory network; The training data set prepared in step 4-2 is used to capture the dependent features and complex patterns in the data of the acoustic sensor using the long short-term memory network model. The input data is feature extracted through the gating mechanism to obtain the feature set H of any discrete time series sampling value in the acoustic array. s ; Step 4-4, autoencoder model training; Input 75% of the normal samples of acoustic sensors obtained in data preparation into the features extracted by the long short-term memory network to obtain the feature set H of each acoustic sensor. s , forming the acoustic sensor feature set H Ts Mapping to a low-dimensional latent space , the expression is: ， Among them, ELU is the activation function , The potential space Reconstruct the original feature space , where W d is the weight, b d is the bias vector: ； The reconstruction error ζ is defined as: ， in is the i-th value of the original feature set, To reconstruct the i-th eigenvalue of the original feature space, when the reconstruction error When converged, the training model is output, 25% of the training preparation data is used to verify the training model, and the autoencoder anomaly judgment model is output.

5. A method for detecting abnormality of an acoustic sensor array for a power grid according to claim 4, characterized in that: The correlation coefficient threshold The value ranges from 0.4 to 0.

6.

6. A method for detecting abnormality of an acoustic sensor array for a power grid according to claim 5, characterized in that: In step 3, the logarithmic power spectrum is also used for feature extraction. The discrete Fourier transform is performed on x[n] to obtain the frequency domain representation X(k) as shown below, where k represents the frequency and j is the imaginary unit: ， Calculate the power spectrum P(k): ， Thus we get the logarithmic power spectrum L(k): L(k)=log(P(k)), and M groups of logarithmic power spectrum features of M discrete time series are obtained.

7. A method for detecting abnormality of an acoustic sensor array for a power grid according to claim 6, characterized in that: The reconstruction error ζ can also be combined with the reconstruction errors of multiple time windows for comprehensive judgment, and 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 start time of the signal and intercepts the data segment with T1 as the length. The medium time window T2 and the long time window T3 are also intercepted in sequence. The data of the short time window T1, medium time window T2, and long time window T3 are respectively subjected to feature extraction of the long short memory network and then autoencoder operation, and the reconstruction error of the discrete sequence of different time windows is calculated. The weights are set for the reconstruction errors of different time windows. 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, where the weight w1 is 0.1, w2 is 0.2, and w3 is 0.7; The short time window reconstruction error is obtained as , the medium time window reconstruction error is , the long time window reconstruction error is , calculate the comprehensive reconstruction error , , by comprehensive reconstruction error The convergence of the acoustic sensor array can be used to determine whether the acoustic sensor array has any abnormality.

Images