A method and device for identifying battery operating status defects

Abstract

The present invention provides a method and device for identifying defects in the operating state of a battery. The method first collects measurable variables during the battery's charge and discharge operations, performs normalization preprocessing on them, and then converts these variables into matrix data using a sliding window technique. A convolutional autoencoder (CAE) is then established to extract deep features of the battery's measurable variables, and a loss function is designed to optimize CAE training performance. Furthermore, a structural similarity evaluation index is used to cluster the deep features, and the clustering results are divided into defect-free clusters and defective clusters using the permutation entropy of the deep features. Finally, a quantitative evaluation index is established to calculate the probability of battery defects, and a long short-term memory (LSTM) neural network is further established to predict battery operating state defects. The present invention can efficiently and accurately identify battery operating state defects and quantify their probability of occurrence, providing support for battery health management.

Description

Technical Field

[0001] The present invention relates to the technical field of battery status detection, and in particular to a method and device for identifying battery operating status defects. Background Art

[0002] Batteries, as a crucial component of energy storage, are widely used in power systems, communication base stations, and other fields. However, as batteries age, they can experience performance degradation and capacity loss, seriously impacting their lifespan and safety. Traditional battery testing methods typically rely on manual inspection and threshold judgment. These methods have significant limitations in practical applications and are unable to accurately reflect the complex operating conditions of batteries. Therefore, it is crucial to develop a real-time, intelligent sliding window-based method for identifying battery operating status defects.

[0003] In recent years, deep learning technology has made significant progress in data processing and pattern recognition. In particular, models such as autoencoders and recurrent neural networks have been rapidly adopted in deep feature learning and time series analysis. However, given the operational characteristics of batteries, traditional deep learning techniques still have shortcomings in defect identification efficiency and accuracy. Summary of the Invention

[0004] In view of the problems existing in the prior art, the present invention provides a method for identifying battery operating status defects, which specifically includes the following steps:

[0005] Step S101: real-time acquisition of measurable variable data of the battery within a certain period of time;

[0006] Step S103: converting the measurable variable data into a first time series data set in matrix form based on a sliding window method;

[0007] Step S105: Input the time series data set into a first convolutional autoencoder (CAE) model to extract a first deep feature set;

[0008] Step S107: dividing the first deep feature set into non-defective clusters and defective clusters based on similarity between features;

[0009] Step S109: evaluating the probability of defects occurring in the first deep feature set within the sliding window based on a quantitative evaluation index model;

[0010] Step S1011: input the first deep feature set into a first long short-term memory neural network (LSTM) model, and output the defect probability of each sliding window.

[0011] The measurable variable data include voltage, current, temperature and internal resistance.

[0012] The step S103 further includes calculating the sample entropy of the data segments in each sliding window to select the optimal window size, which specifically includes:

[0013] Step S1031: Initialize a series of window sizes W and step sizes S;

[0014] Step S1033: For each combination of W and S, calculate and record the sample entropy SampEn=ln A m (r)-ln A m+1 (r), where A m (r) and A m+1 (r) represents the matching probability of the phase space vector formed by the data in the window, m represents the dimension of the vector taken from the first time series data set, and r is the tolerance threshold;

[0015] Step S1035 : draw a two-dimensional surface graph of sample entropy versus W and S, set a second threshold T, and find the minimum window whose sample entropy of all windows is greater than the threshold.

[0016] Among them, A m (r) and A m+1 (r) represents the probability that the phase space vector satisfies the similarity condition when the embedding dimension of the data series is m and m + 1, respectively. The embedding dimension m means that the data points in the time series are combined as an m-dimensional vector, that is, the analysis is performed at m consecutive time points. It is usually chosen to be a small positive number, such as 2 or 3, to capture the short-term dynamic characteristics of the data.

[0017] The tolerance threshold r is used to judge the similarity between vectors. It is usually used as part of data standardization to control the matching requirements between data. That is, if the Euclidean distance between two m-dimensional vectors is less than r, they are considered to match.

[0018] Probability A m (r) represents the probability that all possible m-dimensional vector pairs in the sequence meet the similarity requirement. Specifically, it is the number of m-dimensional vector pairs that meet the similarity requirement divided by the number of all possible vector pairs.

[0019] Probability A m+1 (r) represents the proportion of vector pairs that still meet the similarity requirements when the embedding dimension is increased to m+1.

[0020] The steps for calculating sample entropy include:

[0021] Construct the phase space vector, take out the m-dimensional vector from the time series, and form an m-dimensional phase space vector set.

[0022] Calculate the number of matches. For each vector, calculate the distance between it and other vectors. If the distance is less than the tolerance threshold r, it is considered a match. Count the total number of matches to obtain A. m (r).

[0023] Increase the embedding dimension, add one dimension to each vector to form an m+1 dimensional vector set, and repeat the above steps to obtain A m+1 (r).

[0024] Take the logarithmic difference and take the natural logarithm of the ratio of the two probabilities to get the sample entropy.

[0025] The convolutional autoencoder CAE model in step S105 adopts the following loss function: L CAE =αL M (x,x rec )+βL S (x,x rec ),in x represents the input data, x rec Represents the reconstruction data, x ij and x rec,ij Represent the values ​​at position (i, j) in the input data and the reconstructed data respectively, α and β represent weight coefficients, and O and P represent the number of rows and columns of the matrix.

[0026] Step S107 includes: performing cluster analysis on the first deep feature based on a first structural similarity index model to form two independent clusters, wherein the first structural similarity index model is specifically defined as:

[0027]

[0028] Where, and represents any two eigenvectors, represents the feature dimension, The value range is [0, 1].

[0029] The cluster analysis of the first deep feature set is performed based on the first structural similarity index model to form two independent clusters, which specifically includes the following steps:

[0030] S1071. Randomly select two features from the first deep feature set as the initial centroids of the two clusters;

[0031] S1073, calculating the structural similarity indexes of all other deep features and the two centroids, and reallocating all features in the first deep feature layer to two clusters accordingly;

[0032] S1075. Calculate the mean of each dimension of the features in the two clusters to update the centroids of the two clusters;

[0033] S1077. Repeat steps S1071 to S1075 until the centroids of the two clusters no longer change or the limited number of updates is reached.

[0034] Among them, the clusters with permutation entropy less than the set threshold are defect-free clusters, and the clusters with permutation entropy greater than the set threshold are defective clusters.

[0035] Among them, the first quantitative evaluation index model is established based on the centroid and the distribution of samples within the cluster.

[0036] The quantitative evaluation index model is expressed by the following formula: in represents the defect probability of the t-th deep feature, ξθ(t) represents the position value of the deep feature θ; represents the position value of the defect-free cluster center θ; Indicates the position value of the defect cluster center θ.

[0037] The value range is 0 to 1, where 0 indicates no defect, 1 indicates a defect, and 0.5 is the dividing line between the no-defect state and the defect state.

[0038] Among them, step S1011 includes, based on the output of LSTM, if the defect probability exceeds a second set threshold, determining that a defect exists; otherwise, it is considered that the battery state is normal.

[0039] The present invention further provides a battery operating status defect identification device, which adopts the above-mentioned battery operating status defect identification method based on sliding window, and the device includes:

[0040] A data acquisition module, which is used to collect real-time measurable variable data of the battery within a certain period of time;

[0041] a data conversion module, configured to convert the measurable variable data into a first time series data set in a matrix form based on a sliding window method;

[0042] A feature extraction module, configured to input the first time series dataset into a first convolutional autoencoder (CAE) model to extract a first deep feature set;

[0043] a feature partitioning module, configured to partition the first deep feature set into a non-defective cluster and a defective cluster based on similarities between the features;

[0044] A defect assessment module, configured to assess the probability of defects occurring in the first deep feature set within the sliding window based on a quantitative evaluation index model;

[0045] A probability output module is used to input the first deep feature set into a first long short-term memory neural network (LSTM) model and output the defect probability of each sliding window.

[0046] Expected effects of implementing the present invention

[0047] The method proposed in this paper can efficiently extract deep features from battery charge and discharge data and quantify the probability of battery operating status defects based on these features, achieving high-precision battery defect identification. Compared with traditional defect detection methods, the method proposed in this paper has the following advantages:

[0048] 1. Improved accuracy: By extracting deep features through convolutional autoencoders (CAEs) and combining them with the powerful time series modeling capabilities of long short-term memory neural networks (LSTMs), the battery's operating status can be analyzed more comprehensively and deeply, significantly improving the accuracy of defect identification.

[0049] 2. Enhanced adaptability: This invention innovatively adopts a clustering method based on structural similarity evaluation indicators and permutation entropy analysis to effectively distinguish between defect-free and defective battery states, and establishes a defect evaluation system, which enhances the system's adaptability to different types of batteries and operating environments;

[0050] 3. Wide range of applications: This method is not only applicable to various types of batteries, but also can maintain good performance in a variety of operating scenarios. It has broad application prospects, especially in industries such as power supply and communication security. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The present invention is a flow chart of a method for identifying defects in the operating state of a battery.

[0052] Figure 2 This is a flow chart of preprocessing data of measurable variables of a storage battery according to the present invention.

[0053] Figure 3 It is a schematic diagram of the convolutional autoencoder structure of the present invention.

[0054] Figure 4 It is a flow chart of cluster analysis based on structural similarity index of the present invention.

[0055] Figure 5 It is a schematic diagram of the long short-term memory network structure of the present invention.

[0056] Figure 6 It is a schematic diagram of a battery operating status defect identification device of the present invention. DETAILED DESCRIPTION

[0057] The technical solution of the present invention will be further explained below with reference to the accompanying drawings.

[0058] Example 1

[0059] like Figure 1 The present invention provides a method for identifying battery operating status defects, which specifically includes the following steps:

[0060] Step S101: real-time acquisition of measurable variable data of the battery within a certain time period; wherein the certain time period is a time period when the battery is in a charging and discharging operation state;

[0061] Step S103: converting the measurable variable data into a first time series data set in matrix form based on a sliding window method;

[0062] Step S105: Input the time series data set into a first convolutional autoencoder (CAE) model to extract a first deep feature set;

[0063] Step S107: dividing the first deep feature set into non-defective clusters and defective clusters based on similarity between features;

[0064] Step S109: evaluating the probability of defects occurring in the first deep feature set within the sliding window based on a quantitative evaluation index model;

[0065] Step S1011: input the first deep feature set into a first long short-term memory neural network (LSTM) model, and output the defect probability of each sliding window.

[0066] The collected data includes multiple measurable variables such as voltage, current, temperature and internal resistance.

[0067] During the battery's charge and discharge processes, multiple measurable variables (such as voltage, current, temperature, and internal resistance) are collected in real time. This data is recorded continuously over a certain timeframe. These variables change over time, typically forming a multidimensional dataset.

[0068] This data can be considered as time series data, which is usually collected at specific time intervals. For example, data is recorded once per second. This forms a multidimensional dataset containing multiple time points over time.

[0069] For analysis and feature extraction, a sliding window technique is used. Specifically:

[0070] Set a fixed window size (e.g., 10 seconds) and extract a data segment in each time period.

[0071] The window is moved to cover subsequent time periods, possibly overlapping (e.g., the window slides by 5 seconds), thus generating multiple window segments.

[0072] Each window segment will contain the data of all measurable variables within that time period, forming a multi-dimensional feature vector.

[0073] Each window segment is processed through deep feature extraction (e.g., using a convolutional autoencoder) to generate a feature vector that represents the state of the battery during that time period.

[0074] When these window segments are input into a convolutional autoencoder (CAE), CAE extracts features through multiple convolutional and pooling layers. Ultimately, CAE outputs a deep feature representation of each window segment.

[0075] These deep features are usually a combination of multiple features to form a new feature vector. For example, if the feature vector output by each window segment after CAE processing has 10 dimensions, then each window of new data will correspond to a 10-dimensional feature vector.

[0076] Based on the deep feature vectors extracted from all window segments, clustering algorithms can be applied to analyze these features.

[0077] By clustering, similar feature vectors in the feature space can be grouped to form multiple feature clusters. Each cluster represents a class of similar states (such as normal or defective states).

[0078] After clustering, each cluster will contain feature vectors of multiple windows, which together describe a specific state.

[0079] Centroid positioning: After clustering, the centroid of each feature cluster is calculated. The centroid is the average value of the feature vectors of all samples in the cluster and represents the main features of the cluster.

[0080] The feature vector formed by the collected data after deep feature extraction can be compared with the centroid of each cluster.

[0081] Based on the distance between the centroid and the feature vector, combined with a pre-established quantitative evaluation index model, the probability of a new sample containing a defect can be assessed. For example, a sample farther from the centroid may be more likely to be judged as a defect.

[0082] Example 2

[0083] The present invention provides a method for identifying battery operating status defects based on a sliding window, which specifically includes the following steps:

[0084] Step S101: real-time acquisition of measurable variable data of the battery within a certain period of time;

[0085] Step S103: converting the measurable variable data into a first time series data set in matrix form based on a sliding window method;

[0086] Step S105: Input the time series data set into a first convolutional autoencoder (CAE) model to extract a first deep feature set;

[0087] Step S107: dividing the first deep feature set into non-defective clusters and defective clusters based on similarity between features;

[0088] Step S109: evaluating the probability of defects occurring in the first deep feature set within the sliding window based on a quantitative evaluation index model;

[0089] Step S1011: input the first deep feature set into a first long short-term memory neural network (LSTM) model, and output the defect probability of each sliding window.

[0090] The measurable variable data include voltage, current, temperature and internal resistance.

[0091] Among them, such as Figure 2 The example shown is a flow chart of the preprocessing of measurable variable data of a battery according to the present invention. Step S103 further includes calculating the sample entropy of the data segments within each sliding window to select the optimal window size, which specifically includes:

[0092] Step S1031: Initialize a series of window sizes W and step sizes S;

[0093] Step S1033: For each combination of W and S, calculate and record the sample entropy SampEn=ln A m (r)-ln A m+1 (r), where Am(r) and A m+1 (r) represents the matching probability of the phase space vector formed by the data in the window, m represents the dimension of the vector taken from the first time series data set, and r is the tolerance threshold;

[0094] Step S1035 : draw a two-dimensional surface graph of sample entropy versus W and S, set a second threshold T, and find the minimum window whose sample entropy of all windows is greater than the threshold.

[0095] The convolutional autoencoder CAE model in step S105 adopts the following loss function: L CAE =αL M (x,x rec )+βL S (x,x rec ),in x represents the input data, x rec Represents the reconstruction data, x ij and x rec,ijThey represent the values ​​at position (i, j) in the input data and the reconstructed data respectively, and α and β represent weight coefficients.

[0096] like Figure 3 As shown, it is the structure of the convolutional autoencoder CAE, and the step S105 specifically includes:

[0097] Step S1051: Matrix data formed by the sliding window is used as input data x, with a dimension of O×P;

[0098] Step S1053: The input data x is sent to the convolutional encoder e1 for processing, which specifically includes: first, feature extraction by Q1 convolutional filters C1 with a c×c window size and a q-step length, then feature activation by a ReLU activation function [y(λ)=max(0,λ), where λ represents the input neuron], and feature dimensionality reduction by a maximum pooling layer P1 with an o×o window size and a l-step length, and finally outputting a deep feature h1 with a dimension of a1×b1×c1;

[0099] In step S1055, the deep feature h1 is processed by the convolutional encoders e2, e3, and e4 in sequence. The data processing process is similar to step S3.2, thereby outputting the deep feature h4 with the dimension of a4×b4×c4.

[0100] Step S1057: The deep feature h4 is sent to the convolution decoder d1 for processing. First, the feature dimension is increased by the upsampling layer U1 (g×g) with a g×g window size, and then processed by the convolution filter C5 (Q5@c×c+q) and the ReLU activation function to obtain the deep feature h5.

[0101] In step S1059, the deep feature h5 is processed by the convolution decoders d2, d3, and d4 in turn. The data processing process is similar to step 3.4, except that the activation function of the convolution decoder d4 is different, that is, the Sigmoid function [y(λ) = 1 / 1 + exp(-λ), λ represents the input neuron, and the final output dimension is O×P reconstructed data xrec.

[0102] Step S107 includes: performing cluster analysis on the first deep feature based on a first structural similarity index model to form two independent clusters, wherein the first structural similarity index model is specifically defined as:

[0103]

[0104] Where, and represents any two eigenvectors, represents the feature dimension, The value range is [0, 1].

[0105] Among them, such as Figure 4 As shown, cluster analysis is performed on the first deep feature set based on the first structural similarity index model to form two independent clusters, specifically including the following steps:

[0106] S1071. Randomly select two features from the first deep feature set as the initial centroids of the two clusters;

[0107] S1073, calculating the structural similarity indexes of all other deep features and the two centroids, and reallocating all features in the first deep feature layer to two clusters accordingly;

[0108] S1075. Calculate the mean of each dimension of the features in the two clusters to update the centroids of the two clusters;

[0109] S1077. Repeat steps S1071 to S1075 until the centroids of the two clusters no longer change or the limited number of updates is reached.

[0110] Among them, the clusters with permutation entropy less than the set threshold are defect-free clusters, and the clusters with permutation entropy greater than the set threshold are defective clusters.

[0111] The specific steps for obtaining the permutation entropy are as follows:

[0112] For the deep features {h t}, construct the phase space vector sequence H t =[h t , h t+τ ,...,h t+(m-1)τ ], where H t represents the phase space vector at time t, τ represents the delay time, and m represents the embedding dimension;

[0113] For each phase space vector H t , determine its sequential pattern τ t , that is, H t The order of the elements in ;

[0114] Count the frequency or probability P(π) of all possible sequential patterns π, and calculate the permutation entropy PermEn=-∑ π P(π)log P(π).

[0115] Among them, the first quantitative evaluation index model is established based on the centroid and the distribution of samples within the cluster.

[0116] The quantitative evaluation index model is expressed by the following formula: in represents the defect probability of the t-th deep feature, ξ θ (t) represents the position value of deep feature θ; represents the position value of the defect-free cluster center θ; Indicates the position value of the defect cluster center θ.

[0117] The value range is 0 to 1, where 0 indicates no defect, 1 indicates a defect, and 0.5 is the dividing line between the no-defect state and the defect state.

[0118] In step S109, the goal is to evaluate the state of the new sample in the feature space by calculating the centroid and establishing quantitative evaluation indicators, focusing primarily on the relationship between the current sample and the known clusters. Based on the results of the cluster analysis, the centroids of the defect-free cluster and the defective cluster are calculated. These centroids represent the center point of each cluster and reflect the average characteristics of the samples within the cluster.

[0119] The calculation of the centroid provides a reference point for subsequent defect probability assessment.

[0120] Based on the centroid and the feature vector of the new sample, a quantitative evaluation index is established. This index is used to evaluate the probability of defects in the new sample within the sliding window.

[0121] Calculate the distance between the new sample and the centroid: The distance between the feature vector of the new sample and the centroid can reflect its similarity to the normal state or defect state. The closer the distance, the higher the similarity between the sample and the cluster.

[0122] Distribution density of samples in the cluster: By analyzing the distribution of samples within the cluster, the status of new samples can be further evaluated.

[0123] Step S1011 includes determining, based on the LSTM output, that a defect exists if the defect probability exceeds a second set threshold; otherwise, the battery is considered normal. Defect probability prediction using the LSTM model focuses on the dynamic characteristics of time series data and can leverage historical information to improve prediction accuracy.

[0124] like Figure 5 The figure shows the LSTM structure of the long short-term memory network, which consists of an input gate i, a forget gate f, and an output gate o. The specific calculation process is as follows:

[0125]

[0126]

[0127]

[0128]

[0129]

[0130] r t =o t ⊙tanhc t (10)

[0131] Where, represents the deep features at time t, r t represents the time series characteristics at time t, r t-1 represents the time series characteristics at time t-1, i t 、f t and o t Represent the activation vectors of the input gate, forget gate, and output gate respectively, σ represents the Sigmoid function, Tanh represents the hyperbolic sine function, Indicates the candidate storage state, c t Represents the internal state; U and W both represent weight matrices, b represents the bias, and ⊙ represents the dot product of the matrices.

[0132] LSTM is a network that can capture long-term and short-term dependencies in time series data. It is designed to memorize previously input information and use it to make more accurate predictions.

[0133] Based on the input sliding window features and the patterns learned during training, LSTM outputs the defect probability for each window. This means that for each time window, LSTM will evaluate whether the battery is likely to be defective within that window and give a probability value.

[0134] LSTM predicts the defect probability for each window based on the input feature vector and its temporal dependencies. Crucially, LSTM not only considers the characteristics of the current window but also captures the dynamic changes of historical windows in the time series, thereby improving the accuracy of defect identification.

[0135] Example 3

[0136] like Figure 6 As shown, the present invention provides a battery operating status defect identification device, which adopts the above-mentioned battery operating status defect identification method based on sliding window, and the device may include:

[0137] A data acquisition module, which is used to collect real-time measurable variable data of the battery within a certain period of time;

[0138] a data conversion module, configured to convert the measurable variable data into a first time series data set in a matrix form based on a sliding window method;

[0139] A feature extraction module, configured to input the first time series dataset into a first convolutional autoencoder (CAE) model to extract a first deep feature set;

[0140] a feature partitioning module, configured to partition the first deep feature set into a non-defective cluster and a defective cluster based on similarities between the features;

[0141] A defect assessment module, configured to assess the probability of defects occurring in the first deep feature set within the sliding window based on a quantitative evaluation index model;

[0142] A probability output module is used to input the first deep feature set into a first long short-term memory neural network (LSTM) model and output the defect probability of each sliding window.

Claims

1. A method for identifying battery operating status defects, characterized in that: It includes the following steps: Step S101: real-time acquisition of measurable variable data of the battery within a certain period of time; Step S103: converting the measurable variable data into a first time series data set in matrix form based on a sliding window method; Step S105: Input the first time series data set into a first convolutional autoencoder (CAE) model to extract a first deep feature set. The convolutional autoencoder (CAE) model uses the following loss function: ,in , , Represents input data, Reconstructing data, and Represent the input data and the reconstructed data respectively The value at the position, 、 represents the weight coefficient, O and P represent the number of rows and columns of the matrix; Step S107: Based on the similarity between the features, the first deep feature set is divided into a non-defective cluster and a defective cluster, including: performing cluster analysis on the first deep feature based on the first structural similarity index model to form two independent clusters, specifically including the following steps: S1071: randomly selecting two features in the first deep feature set as the initial centroids of the two clusters; S1073: calculating the structural similarity index of all other deep features and the two centroids, and redistributing all features in the first deep feature to the two clusters accordingly; S1075: calculating the mean of each dimension of the features in the two clusters to update the centroids of the two clusters; S1077: repeating steps S1071 to S1075 until the centroids of the two clusters no longer change or the limited number of updates is reached; Among them, based on the distribution of centroids and samples within the cluster, a quantitative evaluation index model is established, which is expressed by the following formula: ,in Indicates the The defect probability of deep features, Representing deep features Position value; Denotes the center of the defect-free cluster Position value; Denotes the defect cluster center Position value; The value range is 0~1, where 0 indicates no defect, 1 indicates defect, and 0.5 is the dividing value between the no defect state and the defect state; Step S109: evaluating the probability of defects occurring in the first deep feature set within the sliding window based on a quantitative evaluation index model; Step S1011: input the first deep feature set into a first long short-term memory neural network (LSTM) model, and output the defect probability of each sliding window.

2. The method according to claim 1, wherein: The step S103 further includes calculating the sample entropy of the data segments in each sliding window to select the optimal window size, which specifically includes: Step S1031: Initialize a series of window sizes W and step sizes S; Step S1033: For each combination of W and S, calculate and record the sample entropy ,in and Both represent the matching probability of the phase space vector formed by the data in the window, m represents the dimension of the vector taken from the first time series dataset, and r is the tolerance threshold; Step S1035 : draw a two-dimensional surface graph of sample entropy versus W and S, set a second threshold T, and find the minimum window whose sample entropy of all windows is greater than the threshold.

3. The method according to claim 1, wherein: The first structural similarity index model is specifically defined as: ; Where, and represents any two eigenvectors, represents the feature dimension, The value range is [0,1].

4. The method according to claim 1, wherein: The clusters whose permutation entropy is less than the set threshold are defect-free clusters, and the clusters whose permutation entropy is greater than the set threshold are defective clusters.

5. The method according to claim 1, wherein: The measurable variable data include voltage, current, temperature and internal resistance.

6. The method according to claim 1, wherein: Step S1011 includes determining that a defect exists based on the output of the LSTM if the defect probability exceeds a second set threshold; otherwise, the battery is considered to be in normal condition.

7. A battery operating status defect identification device, characterized in that: The method for identifying battery operating status defects according to any one of claims 1 to 6 is adopted, and the device includes: A data acquisition module, which is used to collect real-time measurable variable data of the battery within a certain period of time; a data conversion module, configured to convert the measurable variable data into a first time series data set in a matrix form based on a sliding window method; A feature extraction module, configured to input the first time series dataset into a first convolutional autoencoder (CAE) model to extract a first deep feature set; a feature partitioning module, configured to partition the first deep feature set into a non-defective cluster and a defective cluster based on similarities between the features; A defect assessment module, configured to assess the probability of defects occurring in the first deep feature set within the sliding window based on a quantitative evaluation index model; A probability output module is used to input the first deep feature set into a first long short-term memory neural network (LSTM) model and output the defect probability of each sliding window.

Images