Power battery fault diagnosis method and system based on voltage prediction residual analysis

Abstract

The invention discloses a power battery fault diagnosis method and system based on voltage prediction residual analysis, and the method comprises the steps: predicting the voltage of each single battery in a vehicle operation process through employing a deep learning model according to the vehicle operation data and the battery operation data, and taking the predicted voltage as a prediction reference voltage; measuring the voltage of each single battery in the running process of the vehicle, and generating a voltage residual sequence of each single battery by calculating the difference value between the actually measured voltage and the predicted reference voltage; performing multi-dimensional feature extraction on the voltage residual error sequence by using a sliding window, extracting a variable coefficient and a sample entropy, and constructing a residual error feature vector; and calculating a local abnormal factor according to the residual feature vector, and identifying whether the single battery is abnormal or not according to the local abnormal factor. The method can improve the fault diagnosis accuracy of the power battery of the new energy automobile.

Description

Technical Field

[0001] This invention belongs to the field of new energy vehicle technology, and relates to power battery fault diagnosis, and particularly to a power battery fault diagnosis method and system based on predictive residual analysis. Background Technology

[0002] With the global energy structure transformation and the advancement of "dual-carbon" goals, electric vehicles have become an important direction for sustainable development in the transportation sector. Lithium-ion batteries, due to their high energy density and long cycle life, are widely used in electric vehicles, portable electronics, and other fields. However, various forms of abuse under complex operating conditions and manufacturing defects can lead to lithium-ion battery failures, causing safety issues. The power system of an electric vehicle typically consists of several battery cells connected in series or parallel to meet its capacity and voltage requirements. Failure of any battery cell can, at best, reduce the electric vehicle's range, and at worst, trigger thermal runaway, threatening the safety of passengers and their property. Therefore, timely and accurate diagnosis of battery faults and location of the faulty cell, followed by safety warnings before thermal runaway occurs, is crucial for improving the safety of battery-powered vehicles and promoting the further development of electric vehicles.

[0003] Lithium-ion battery failures typically manifest as changes in internal parameters. Data-driven methods can be used to diagnose battery failures by collecting battery parameters through sensors, extracting relevant feature parameters, and establishing machine learning or deep learning models for fault diagnosis. However, existing fault diagnosis methods still suffer from drawbacks such as insufficient sensitivity to fault features, poor real-time performance, and a high false positive rate. Summary of the Invention

[0004] To address the shortcomings of existing electric vehicle power battery fault diagnosis technologies, such as insufficient sensitivity to fault features, poor real-time performance, and high false positive rate, this invention provides a power battery fault diagnosis method and system based on voltage prediction residual analysis, thereby improving the accuracy of power battery fault diagnosis.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0006] A power battery fault diagnosis method based on voltage prediction residual analysis includes:

[0007] A deep learning model is used to predict the voltage of each individual battery cell during vehicle operation based on vehicle operation data and battery operation data, and this voltage is used as a prediction reference voltage.

[0008] The voltage of each individual battery cell is measured during vehicle operation, and the voltage residual sequence of each individual battery cell is generated by calculating the difference between the actual measured voltage and its predicted reference voltage.

[0009] A sliding window is used to extract multidimensional features from the voltage residual sequence, extracting the coefficient of variation and sample entropy, and constructing a residual feature vector.

[0010] The local anomaly factor is calculated based on the residual feature vector, and the anomaly factor is used to identify whether a single cell is abnormal.

[0011] Furthermore, the vehicle operating data includes: vehicle speed, vehicle status, charging status, cumulative mileage, drive motor voltage, and drive motor torque.

[0012] Furthermore, the battery operating data includes: total voltage, total current, temperature, SOC, SOH, and insulation status of the battery pack.

[0013] Furthermore, extracting the coefficient of variation using a sliding window includes:

[0014] First, the voltage residual sequence is obtained using a sliding window, and represented as a matrix:

[0015]

[0016]

[0017] in, The residual sequence matrix for a single sliding window; The length of the sliding window. This refers to the number of individual cells included in the battery pack. , , These represent the voltage residual value, predicted reference voltage, and actual measured voltage of cell n at time t, respectively.

[0018] Then, a weighting factor is introduced to weight the voltage residual values ​​at each time point within the sliding window, resulting in a weighted voltage residual sequence matrix:

[0019]

[0020] in, This is the weighted voltage residual sequence matrix. , Let be the weighted residual voltage value of cell n at time t. Let be the weight at time t. It is a weighting factor. ;

[0021] Finally, based on the weighted voltage residual sequence matrix, the coefficient of variation of the voltage residual sequence for each individual cell in the residual matrix is ​​calculated:

[0022]

[0023] in, These are the coefficients of variation for cell number n; and These are the mean and standard deviation of the voltage residuals of cell number n within the sliding window, respectively.

[0024] Furthermore, extracting sample entropy using a sliding window includes:

[0025] First, the voltage residual sequence is obtained using a sliding window, and represented as a matrix:

[0026]

[0027]

[0028] in, The residual sequence matrix for a single sliding window; The length of the sliding window. This refers to the number of individual cells included in the battery pack. , , These represent the voltage residual value, predicted reference voltage, and actual measured voltage of cell n at time t, respectively.

[0029] Then, a weighting factor is introduced to weight the voltage residual values ​​at each time point within the sliding window, resulting in a weighted voltage residual sequence matrix:

[0030]

[0031] in, This is the weighted voltage residual sequence matrix. , Let be the weighted residual voltage value of cell n at time t. Let be the weight at time t. It is a weighting factor. ;

[0032] right Each column of voltage residual sequence , build A length of Vector sequence:

[0033] ,

[0034] For two vectors and Distance between:

[0035]

[0036] Set tolerance threshold Statistics in all China satisfies Number of and define intermediate variables Then take all vector sequences corresponding to mean :

[0037]

[0038] Increase m to m+1, and follow the... Recalculate using the method ;

[0039] Final utilization and Calculate the sample entropy of cell number n:

[0040]

[0041] in, Let be the sample entropy of cell number n.

[0042] Furthermore, local outlier factors are calculated based on the residual eigenvectors, including:

[0043] First, for the feature vector of cell number n... The distance to its k-th nearest neighbor is defined as:

[0044]

[0045] Will The k corresponding neighboring points form the nearest neighbor set:

[0046]

[0047] in, For two individual cells, based on their feature vectors Calculated Euclidean distance:

[0048]

[0049] Then calculate To its neighbor feature vector reachable distance :

[0050]

[0051] Then calculate The local reachability density is the reciprocal of the sum of the average reachability distances to its nearest neighbors:

[0052]

[0053] in for The number of nearest neighbors;

[0054] The final calculation of cell number n is based on Local anomaly factor values That is, the ratio of the local reachability density of its neighbors to the average local reachability density of itself:

[0055] .

[0056] Furthermore, identifying whether a single cell is abnormal based on local anomaly factors includes:

[0057] Battery data for each individual cell during its normal operating period are selected, and the local anomaly factor within each sliding window is calculated to obtain the anomaly factor matrix for all individual cells across C sliding windows during their normal operating period. :

[0058]

[0059] Calculate the maximum local anomaly value within each sliding window:

[0060]

[0061] in, This represents the maximum anomaly factor value of the c-th sliding window, which is the maximum value of the anomaly factor values ​​among all individual cells, and T represents the matrix transpose.

[0062] Calculate the average value of the largest local anomaly factor With variance :

[0063]

[0064]

[0065] According to the average With variance Calculate the adaptive threshold :

[0066]

[0067] The local anomaly factor values ​​of each nth cell With adaptive threshold In comparison, if If the cell is normal, it is considered to be normal; otherwise, it is considered to be faulty.

[0068] Furthermore, the deep learning model employs LSTM.

[0069] A power battery fault diagnosis system based on voltage prediction residual analysis includes:

[0070] The reference voltage prediction module is used to: use a deep learning model and based on vehicle operation data and battery operation data to predict the voltage of each individual battery cell during vehicle operation, and use it as a prediction reference voltage;

[0071] The voltage acquisition module is used to measure the voltage of each individual battery cell during vehicle operation.

[0072] The residual sequence generation module is used to generate voltage residual sequences for each individual cell by calculating the difference between the actual measured voltage and its predicted reference voltage.

[0073] The residual feature extraction module is used to: extract multidimensional features from the voltage residual sequence using a sliding window, extract the coefficient of variation and sample entropy, and construct a residual feature vector.

[0074] The local anomaly factor calculation module is used to calculate the local anomaly factor based on the residual feature vector.

[0075] The fault diagnosis module is used to identify whether a single cell is abnormal based on local anomaly factors.

[0076] Compared with the prior art, the technical effects of the present invention are as follows:

[0077] This invention utilizes real-vehicle data and performs data processing. Pearson correlation coefficients are used to filter potential features strongly correlated with voltage, avoiding unnecessary feature parameters from affecting the deep learning model and improving the reliability of the prediction model. An LSTM deep learning model is then built and trained. The filtered feature parameters are then input into the trained deep learning model to obtain the predicted normal operating voltage of the battery. A voltage residual sequence is generated based on the predicted and measured voltages. A weighted sliding window mechanism is used to extract two-dimensional feature parameters from the residual sequence, avoiding the impact of single-point voltage fluctuations on fault diagnosis and reducing the false positive rate. Simultaneously, the weight of the latest data within the window is increased to advance the fault diagnosis time. Finally, a local anomaly factor algorithm and adaptive thresholds are used to diagnose and locate faulty cells, further improving battery safety. Attached Figure Description

[0078] Figure 1 This is a flowchart of the power battery fault diagnosis method described in the embodiments of this application. Detailed Implementation

[0079] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.

[0080] This embodiment provides a power battery fault diagnosis method based on voltage prediction residual analysis, including:

[0081] Step 1: Using a deep learning model and based on vehicle operation data and battery operation data, predict the voltage of each individual battery cell during vehicle operation, and use this voltage as a prediction reference voltage.

[0082] The vehicle operating data includes the following characteristic parameters: vehicle speed, vehicle status, charging status, cumulative mileage, drive motor voltage, and drive motor torque. The battery operating data includes the following characteristic parameters: total battery pack voltage, total current, and temperature; state of charge (SOC), state of equilibrium (SOH), and insulation status.

[0083] In this embodiment, the acquired vehicle operation data and battery operation data are preprocessed, and the correlation between each characteristic parameter and the battery cell voltage is evaluated using the Pearson correlation coefficient. The formula for the Pearson correlation coefficient is as follows:

[0084]

[0085] Where r is the Pearson correlation coefficient between the two variables, X i Here are the sequences of characteristic parameters, and Y is the sequence of individual unit voltages. and Let |r| be the average value of variables X and Y, respectively. The closer |r| is to 1, the stronger the positive or negative correlation between the two variables. Conversely, the closer |r| is to 0, the less correlated the two variables are. Generally, a correlation between the two variables is considered strong when |r| > 0.7. Therefore, feature parameters with |r| > 0.7 related to the unit voltage are chosen as inputs to the subsequent prediction model.

[0086] This embodiment uses a Long Short-Term Memory (LSTM) network as the core algorithm of the prediction model. As a special type of recurrent neural network (RNN), LSTM can solve the gradient vanishing or exploding problems that traditional RNNs encounter when processing long sequences. Therefore, LSTM can deeply capture the long-term dynamic dependence characteristics of battery states, thereby achieving higher accuracy and stronger robustness in single-cell voltage prediction. The functional expression of the LSTM algorithm is:

[0087]

[0088]

[0089]

[0090]

[0091]

[0092]

[0093] in, The input variable consists of the filtered sequence of feature parameters; Represents the sigmoid function; This is an input gate used to control the inflow of new information. Forget gates control the degree to which old information is retained, determining whether parameters from the previous moment need to be forgotten; The cell state is determined by first acquiring and retaining previous data, then adding the current state input and paranoia term, and finally determining the cell state. As an output gate, it controls the proportion of the current memory output to the outside world, determining how much information of the current cell state needs to be output as a hidden state. The hidden state is the current hidden state, which is passed to the next time step to update the cell parameters, and is also input to the output layer to obtain the predicted value at the current time step. This is the output layer, which outputs the predicted voltage at the current moment; tanh is the hyperbolic tangent activation function. These are the input weights for the input gate, forget gate, cell state, output gate, and output layer, respectively. These are the recursive weight matrices for the input gate, forget gate, cell state, and output gate, respectively. These are the cell state weight matrices for the input gate, forget gate, and output gate, respectively. These are the bias terms for the input gate, forget gate, cell state, output gate, and output layer, respectively.

[0094] An LSTM network is trained using pre-existing battery cell normal operation data as the training set to construct a voltage prediction model. After successfully building the prediction model, the voltage of each battery cell during vehicle operation is predicted using real-time feature parameters provided by the signal extraction module—namely, real-time vehicle operation data and battery operation data collected during the current vehicle operation—and used as the reference voltage for the current moment. Subsequently, the residual sequence between the real-time voltage data and the reference voltage is analyzed to identify whether any voltage anomalies have occurred.

[0095] Step 2: Measure the voltage of each individual battery cell during vehicle operation, and generate a voltage residual sequence for each individual battery cell by calculating the difference between the actual measured voltage and its predicted reference voltage.

[0096]

[0097] in, , , These represent the voltage residual value, predicted reference voltage, and actual measured voltage of cell n at time t, respectively.

[0098] Step 3: Use a sliding window to extract multidimensional features from the voltage residual sequence, extract the coefficient of variation and sample entropy, and construct the residual feature vector.

[0099] Step 3.1: Obtain the voltage residual sequence using a sliding window. The voltage residual sequence within a single sliding window is represented by a matrix as follows:

[0100]

[0101] in, For a single sliding window, the voltage residual sequence matrix is ​​represented. The length of the sliding window. This refers to the number of individual cells included in the battery pack.

[0102] By using a sliding window, the latest residual data can be analyzed in real time, and various feature parameters of the sequence within the window can be extracted to avoid misjudgments caused by single-point fluctuations.

[0103] Step 3.2: Weight the voltage residual sequence matrix.

[0104] Traditional sliding window methods, due to the equal weighting of data within the window, can lead to prolonged fault diagnosis time and even the overlooking of minor short-term anomalies when a fault first appears because historical normal data dominates the window. Therefore, this invention introduces a weighting factor to increase the weight of the most recent data within the window, reducing the influence of historical data. The weighted voltage residual sequence matrix is ​​represented as follows:

[0105]

[0106] in, This is the weighted voltage residual sequence matrix. , Let be the weighted residual voltage value of cell n at time t. Let be the weight at time t. It is a weighting factor. .

[0107] Step 3.3: Based on the weighted voltage residual sequence matrix, calculate the coefficient of variation of the voltage residual sequence for each individual cell in the residual matrix:

[0108]

[0109] in, These are the coefficients of variation for cell number n; and These are the mean and standard deviation of the voltage residuals of cell number n within the sliding window, respectively. , .

[0110] Therefore, the coefficient of variation of all N battery cells within a single sliding window can be expressed as:

[0111] .

[0112] Step 3.4: Based on the weighted voltage residual sequence matrix, calculate the sample entropy of the voltage residual sequence of each individual cell in the residual matrix.

[0113] First of all, Each column of voltage residual sequence , build A length of Vector sequence:

[0114] ,

[0115] Then, for two vectors and Distance between:

[0116]

[0117] Set tolerance threshold Statistics in all China satisfies Number of and define intermediate variables Then take all vector sequences corresponding to mean :

[0118]

[0119] Then increase m to m+1, and follow... Recalculate using the method ;

[0120] Ultimately, utilizing and Calculate the sample entropy of cell number n:

[0121]

[0122] in, Let be the sample entropy of cell number n.

[0123] Therefore, the sample entropy of each battery cell within a single sliding window can be expressed as:

[0124] .

[0125] Step 4: Calculate the local anomaly factor based on the residual feature vector, and identify whether a single cell is abnormal based on the local anomaly factor.

[0126] Local anomaly factor (LOF) is a density-based anomaly detection algorithm. Its core idea is that if the local density of a sample point is significantly less than the average density of its neighbors, that point is likely an outlier. LOF can detect outliers in local regions and is suitable for data with uneven density distribution. Combining the two-dimensional feature parameters obtained above, the LOF method is used to calculate the anomaly score of each battery cell within each window. Based on previous normal operation data, a threshold is set using the 3δ rule. Then, by judging whether the anomaly score of a single cell is greater than the set threshold, abnormal cells can be located and identified, thereby achieving power battery fault diagnosis.

[0127] Step 4.1, calculate the local anomaly factor based on the residual eigenvector, including:

[0128] First, for the feature vector of cell number n... The distance to its k-th nearest neighbor is defined as:

[0129]

[0130] Will The k corresponding neighboring points form the nearest neighbor set:

[0131]

[0132] in, For two individual cells, based on their feature vectors Calculated Euclidean distance:

[0133]

[0134] Then calculate To its neighbor feature vector reachable distance :

[0135]

[0136] Then calculate The local reachability density is the reciprocal of the sum of the average reachability distances to its nearest neighbors:

[0137]

[0138] in for The number of its nearest neighbors.

[0139] The final calculation of cell number n is based on Local anomaly factor values That is, the ratio of the local reachability density of its neighbors to the average local reachability density of itself:

[0140] .

[0141] Based on the above process, the anomaly factor value of each battery cell within each sliding window can be calculated.

[0142] Step 4.2: Calculate the adaptive threshold for local anomaly factors.

[0143] First, battery data for each individual cell during its normal operating period are selected, and the local anomaly factor within each sliding window is calculated to obtain the anomaly factor matrix for all individual cells across C sliding windows during their normal operating period. :

[0144]

[0145] Then, calculate the maximum local anomaly factor value within each sliding window:

[0146]

[0147] in, This represents the maximum anomaly factor value of the c-th sliding window, which is the maximum value of the anomaly factor values ​​among all individual cells, and T represents the matrix transpose.

[0148] Then calculate the average value of the maximum local anomaly factor. With variance :

[0149]

[0150]

[0151] Finally, based on the average value With variance Calculate the adaptive threshold :

[0152]

[0153] Step 4.3: Identify whether a single cell is abnormal based on local anomaly factors.

[0154] The local anomaly factor values ​​of each nth cell With adaptive threshold In comparison, if If the cell is normal, it is considered to be normal; otherwise, it is considered to be faulty.

[0155] Example 2

[0156] This embodiment provides a power battery fault diagnosis system based on voltage prediction residual analysis, including:

[0157] The reference voltage prediction module is used to: use a deep learning model and based on the vehicle's operating data and its battery operating data during normal operation to predict the voltage of each individual battery cell in the vehicle, and use this as a prediction reference voltage.

[0158] The voltage acquisition module is used to measure the voltage of each individual battery cell during vehicle operation.

[0159] The residual sequence generation module is used to generate voltage residual sequences for each individual cell by calculating the difference between the actual measured voltage and its predicted reference voltage.

[0160] The residual feature extraction module is used to: extract multidimensional features from the voltage residual sequence using a sliding window, extract the coefficient of variation and sample entropy, and construct a residual feature vector.

[0161] The local anomaly factor calculation module is used to calculate the local anomaly factor based on the residual feature vector.

[0162] The fault diagnosis module is used to identify whether a single cell is abnormal based on local anomaly factors.

[0163] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, these changes or improvements should fall within the scope of protection claimed in this application.

Claims

1. A power battery fault diagnosis method based on voltage prediction residual analysis, characterized in that, include: A deep learning model is used to predict the voltage of each individual battery cell during vehicle operation based on vehicle operation data and battery operation data, and this voltage is used as a prediction reference voltage. The voltage of each individual battery cell is measured during vehicle operation, and the voltage residual sequence of each individual battery cell is generated by calculating the difference between the actual measured voltage and its predicted reference voltage. A sliding window is used to extract multidimensional features from the voltage residual sequence, extracting the coefficient of variation and sample entropy, and constructing a residual feature vector. The local anomaly factor is calculated based on the residual feature vector, and the anomaly factor is used to identify whether a single cell is abnormal.

2. The power battery fault diagnosis method according to claim 1, characterized in that, The vehicle operating data includes: vehicle speed, vehicle status, charging status, cumulative mileage, drive motor voltage, and drive motor torque.

3. The power battery fault diagnosis method according to claim 1, characterized in that, The battery operating data includes: total voltage, total current, temperature, SOC, SOH, and insulation status of the battery pack.

4. The power battery fault diagnosis method according to claim 1, characterized in that, Extracting the coefficient of variation using a sliding window includes: First, the voltage residual sequence is obtained using a sliding window, and represented as a matrix: ； ； in, The residual sequence matrix for a single sliding window; The length of the sliding window. This refers to the number of individual cells included in the battery pack. , , These represent the voltage residual value, predicted reference voltage, and actual measured voltage of cell n at time t, respectively. Then, a weighting factor is introduced to weight the voltage residual values ​​at each time point within the sliding window, resulting in a weighted voltage residual sequence matrix: ； in, This is the weighted voltage residual sequence matrix. , Let be the weighted residual voltage value of cell n at time t. Let be the weight at time t. It is a weighting factor. ; Finally, based on the weighted voltage residual sequence matrix, the coefficient of variation of the voltage residual sequence for each individual cell in the residual matrix is ​​calculated: ； in, These are the coefficients of variation for cell number n; and These are the mean and standard deviation of the voltage residuals of cell number n within the sliding window, respectively.

5. The power battery fault diagnosis method according to claim 1, characterized in that, Extracting sample entropy using a sliding window includes: First, the voltage residual sequence is obtained using a sliding window, and represented as a matrix: ； ； in, The residual sequence matrix for a single sliding window; The length of the sliding window. This refers to the number of individual cells included in the battery pack. , , These represent the voltage residual value, predicted reference voltage, and actual measured voltage of cell n at time t, respectively. Then, a weighting factor is introduced to weight the voltage residual values ​​at each time point within the sliding window, resulting in a weighted voltage residual sequence matrix: ； in, This is the weighted voltage residual sequence matrix. , Let be the weighted residual voltage value of cell n at time t. Let be the weight at time t. It is a weighting factor. ; right Each column of voltage residual sequence , build A length of Vector sequence: ， ； For two vectors and Distance between: ； Set tolerance threshold Statistics in all China satisfies Number of and define intermediate variables Then take all vector sequences corresponding to mean : ； Increase m to m+1, and follow the... Recalculate using the method ; Final utilization and Calculate the sample entropy of cell number n: ； in, Let be the sample entropy of cell number n.

6. The power battery fault diagnosis method according to claim 1, characterized in that, The local anomaly factor is calculated based on the residual eigenvector, including: First, for the feature vector of cell number n... The distance to its k-th nearest neighbor is defined as: ； Will The k corresponding neighboring points form the nearest neighbor set: ； in, For two individual cells, based on their feature vectors Calculated Euclidean distance: ； Then calculate To its neighbor feature vector reachable distance : ； Then calculate The local reachability density is the reciprocal of the sum of the average reachability distances to its nearest neighbors: ； in for The number of nearest neighbors; The final calculation of cell number n is based on Local anomaly factor values That is, the ratio of the local reachability density of its neighbors to the average local reachability density of itself: 。 7. The power battery fault diagnosis method according to claim 1, characterized in that, Identifying whether a single cell is abnormal based on local anomaly factors includes: Battery data for each individual cell during its normal operating period are selected, and the local anomaly factor within each sliding window is calculated to obtain the anomaly factor matrix for all individual cells across C sliding windows during their normal operating period. : ； Calculate the maximum local anomaly value within each sliding window: ； in, This represents the maximum anomaly factor value of the c-th sliding window, which is the maximum value of the anomaly factor values ​​among all individual cells, and T represents the matrix transpose. Calculate the average value of the largest local anomaly factor With variance : ； ； According to the average With variance Calculate the adaptive threshold : ； The local anomaly factor values ​​of each nth cell With adaptive threshold In comparison, if If the cell is normal, it is considered to be normal; otherwise, it is considered to be faulty.

8. The power battery fault diagnosis method according to claim 1, characterized in that, The deep learning model uses LSTM.

9. A power battery fault diagnosis system based on voltage prediction residual analysis, characterized in that, include: The reference voltage prediction module is used to: use a deep learning model and based on vehicle operation data and battery operation data to predict the voltage of each individual battery cell during vehicle operation, and use it as a prediction reference voltage; The voltage acquisition module is used to measure the voltage of each individual battery cell during vehicle operation. The residual sequence generation module is used to generate voltage residual sequences for each individual cell by calculating the difference between the actual measured voltage and its predicted reference voltage. The residual feature extraction module is used to: extract multidimensional features from the voltage residual sequence using a sliding window, extract the coefficient of variation and sample entropy, and construct a residual feature vector. The local anomaly factor calculation module is used to calculate the local anomaly factor based on the residual feature vector. The fault diagnosis module is used to identify whether a single cell is abnormal based on local anomaly factors.

Images