Battery fault diagnosis method based on multi-dimensional feature fusion

Abstract

The invention relates to a multi-dimensional feature fusion-based battery fault diagnosis method, which comprises the following steps of: acquiring basic data of a battery, performing cleaning processing, extracting voltage data, judging whether an overcharge or overdischarge fault occurs according to the voltage data, if not, performing sliding time window processing on the voltage data to obtain corresponding feature data, comprising standard deviation, Shannon entropy and stack type auto-encoder data; determining a threshold value corresponding to each feature data by combining each feature data and adopting a quartile method for comparing and judging whether a battery fault occurs or not, if so, correcting each feature data by utilizing a correction coefficient, and determining a current fault type according to a change trend of the corrected feature data; and a local outlier factor algorithm is adopted, and the current fault level is determined by calculating outlier factors of the feature data. Compared with the prior art, the method has the advantages that the accuracy and the economical efficiency of the faulty battery can be ensured, and meanwhile, the fault type is determined and the fault level is divided.

Description

Technical Field

[0001] The present invention relates to the technical field of energy storage battery safety management, and in particular to a battery fault diagnosis method based on multi-dimensional feature fusion. Background Art

[0002] New energy batteries meet the modern requirements of energy conservation, but on the other hand, they also pose a huge challenge to the safe and reliable operation of new energy batteries. In order to ensure the safe, stable and reliable operation of the battery system, it is necessary to perform fault diagnosis on the battery. If timely diagnosis and positioning are not performed, the battery will fail and cause thermal runaway in the final stage, resulting in major accidents.

[0003] Existing battery fault diagnosis technologies mainly include fault diagnosis methods based on battery models, model-free fault diagnosis methods, and research on multi-fault joint diagnosis. Through the application of these technologies, possible battery failures can be identified and prevented earlier, thereby reducing the risk of accidents.

[0004] Patent CN114801751B discloses a car battery fault prediction system based on data analysis, which includes a historical fault analysis unit, a fault threshold acquisition unit, an instant data prediction unit, and a performance trend prediction unit. By acquiring the operating data of the car battery to perform fault diagnosis data analysis, the impact of each parameter of the car battery and the qualified prediction range of the parameters can be understood, thereby improving the accuracy of battery fault diagnosis, and also improving the accuracy of battery fault prediction, and providing strong support for the fault diagnosis results of the battery, real-time monitoring of the battery operation quality, and preventing the car battery from failing during operation; predicting the long-term operating trend of the battery is conducive to real-time control of the battery's operating efficiency. This solution can meet the requirements for battery fault diagnosis, but it cannot guarantee low computing costs, effectively divide the fault level, analyze and capture the real fault, and omit or erroneously record some fault information, resulting in the risk of false alarms or missed alarms.

[0005] Patent CN113093016B provides a method for fault diagnosis of a power battery management system based on uncertain noise filtering, which includes establishing an electrothermal coupling model of the power battery system, expanding the output vector of the system through the state constraints of the power battery, and the state vector of the battery system fault expansion system, obtaining the augmented system of the power battery system, and using a fully symmetric polytope Kalman filter to obtain the estimated interval of the power battery sensor fault, and judging whether the power battery management system has a fault according to the threshold range of the fault estimate, and if there is a fault, determining the fault type and fault time according to the result. This method solves the problem of fault diagnosis of a system with state constraints by extending the system state constraints to the system output vector. However, this solution uses a model-based method to perform fault diagnosis, and first needs to obtain accurate battery modeling, which will have a great impact on the accuracy of fault diagnosis, and if there is no corresponding model, it is impossible to diagnose other faults. This problem also leads to certain limitations, so it is difficult to be uniformly applied to energy storage systems.

[0006] Patent CN116381511B provides a battery status detection method, which obtains the standard deviation and mean of battery data, calculates the first median through the Chebyshev algorithm, obtains the output data of the battery at the last prediction, the input data of the current prediction and the battery parameters, obtains the actual intermediate state matrix and the actual output matrix, predicts the intermediate state matrix and the predicted output matrix, determines the first residual based on the actual intermediate state matrix and the predicted intermediate state matrix, and determines the second residual based on the actual output matrix and the predicted output matrix; in response to the first residual being greater than the first threshold or the second residual being greater than the second threshold, generates a fault residual spectrum based on the first residual and the second residual; determines the amplitude square correlation spectrum based on the fault residual spectrum and the normal state spectrum, and determines the correlation spectrum median of the amplitude square correlation spectrum; based on the correlation spectrum median and the preset judgment threshold, increases the accuracy and traceability of the detection results, thereby determining the fault state of the battery. However, in actual applications, multiple fault types may occur at the same time. This solution does not identify and divide the battery fault type and fault level while ensuring accurate identification of the fault result.

[0007] In summary, existing solutions mainly focus on realizing battery fault diagnosis, and none of them classifies and confirms battery fault levels and fault types. Summary of the invention

[0008] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a battery fault diagnosis method based on multi-dimensional feature fusion, which can determine the fault type and classify the fault level while ensuring the accuracy and economy of the faulty battery.

[0009] The object of the present invention can be achieved by the following technical solution: A battery fault diagnosis method based on multi-dimensional feature fusion, comprising the following steps:

[0010] S1, collect basic battery data and perform cleaning processing to extract voltage data;

[0011] S2. Determine whether an overcharge or over-discharge fault occurs based on the voltage data. If the judgment is no, execute step S3.

[0012] S3, performing sliding time window processing on the voltage data to obtain corresponding feature data, including standard deviation, Shannon entropy and stacked autoencoder data;

[0013] S4, combining each characteristic data, using the quartile method to determine the threshold value corresponding to each characteristic data, for comparison and judgment on whether a battery failure occurs, if the judgment is yes, executing step S5;

[0014] S5. Correct each characteristic data using the correction coefficient to obtain corresponding corrected characteristic data, and determine the current fault type according to the change trend of the corrected characteristic data;

[0015] The local outlier factor algorithm is used to determine the current fault level by calculating the outlier factor of each feature data.

[0016] Furthermore, the step S1 specifically includes the following steps:

[0017] S11, arranging the collected basic battery data in ascending order of time;

[0018] S12, replacing a single missing value or invalid value with normal data from the previous moment;

[0019] S13, replacing the abnormal voltage value beyond the normal working range with the upper limit or lower limit of the working voltage range;

[0020] S14, delete the row with constant voltage;

[0021] S15. Perform min-max normalization on the data:

[0022]

[0023] Among them, U mn,min-max is the voltage data after min-max standardization, U mn is the voltage data of the nth battery at the mth acquisition moment, min{U n} is the minimum value of the nth battery in the whole collection time, max{U n} is the maximum value of the nth battery at all sampling moments.

[0024] Furthermore, the process of determining whether an overcharge or over-discharge fault occurs in step S2 includes:

[0025] S21, obtaining voltage data of each battery at different times, and establishing a corresponding time-voltage (TV) curve;

[0026] S22. Determine the normal operating voltage range according to the battery type, count the number of times each battery exceeds the normal operating voltage range, and if overcharging or overdischarging occurs repeatedly, determine that the battery has an overcharging fault or an overdischarging fault.

[0027] Furthermore, in step S22, the normal operating voltage range includes a normal voltage upper limit and a normal voltage lower limit, and if the battery voltage data is greater than the normal voltage upper limit, it is determined that overcharging occurs;

[0028] If the battery voltage data is lower than the normal voltage lower limit, it is judged that over-discharge has occurred.

[0029] Furthermore, the step S3 specifically includes the following steps:

[0030] S31, extract the voltage data of the single battery, obtain the voltage matrix, let U mn represents the voltage value of cell n at the mth sampling time, then the average voltage of n battery cells at the mth sampling time is for:

[0031]

[0032] S32, define the difference between the sampled voltage of each battery cell n at sampling time m in any sliding window and the average voltage of the battery pack as The length of the sliding window is M;

[0033] S33, calculating the standard deviation of the voltage difference of a single cell n in any sliding window, specifically the standard deviation of the difference between the voltage value of the single cell at each sampling moment and the average voltage of n battery cells:

[0034]

[0035] S34, according to G n (m), the probability matrix P is calculated:

[0036]

[0037] Among them, G n (i) is the difference between the sampled voltage of each battery cell n at sampling time i in any sliding window and the average voltage of the battery pack;

[0038] S35. Using the elements in the obtained probability matrix P as the calculation input of the information entropy, the information entropy, i.e., the Shannon entropy, is calculated:

[0039]

[0040] S36, use the pre-trained stacked autoencoder to extract G n The characteristic parameters of (m) are used to obtain the corresponding stacked autoencoder data.

[0041] Furthermore, the stacked autoencoder is composed of multiple autoencoders superimposed, the output of the first-level autoencoder is used as the input of the second-level autoencoder, the output of the second-level autoencoder is used as the input of the third-level autoencoder, and so on to complete the training of the deep neural network.

[0042] Furthermore, the step S4 specifically includes the following steps:

[0043] S41. According to the quartile method, the threshold corresponding to each feature data is obtained:

[0044] IQR=Q 3 -Q 1

[0045] F 1 =Q 3 +1.5IQR

[0046] F 2 =Q 3 +3IQR

[0047] Among them, F 1 The fault threshold is 1.5 times the IQR fault threshold, F 2 The fault threshold is 3 times the IQR fault threshold, Q 1 is the lower quartile, Q 3 is the upper quartile, and IQR is the interquartile range;

[0048] S42, when the standard deviation, Shannon entropy and stacked autoencoder feature data exceed the corresponding threshold value, it is determined that the battery has fault information;

[0049] When a battery detects fault information in two or more characteristics, the battery is determined to be a faulty battery.

[0050] Furthermore, in step S5, the correction coefficient is -1, 0 or 1. When the voltage at the mth sampling point is less than the average voltage in the current sliding window, the correction coefficient is -1.

[0051] When the voltage at the mth sampling point is greater than the average voltage in the current sliding window, the correction coefficient is 1;

[0052] In other cases, the correction factor is 0.

[0053] Furthermore, in step S5, when the corrected characteristic data suddenly drops, it is determined that a short circuit fault exists; and when the corrected characteristic data suddenly rises, it is determined that an open circuit fault exists.

[0054] Furthermore, the process of determining the current fault level in step S5 includes:

[0055] (1) Obtaining the local reachable distance. According to the definition, if the actual distance to point x is less than the kth distance, it is the kth distance. If it is greater than the kth distance, it is the actual distance, as shown in the following formula:

[0056]

[0057] (2) Determination of local reachability density. The local reachability density is defined as the reciprocal of the average reachability distance from a point in the k-distance neighborhood to point x, as shown in the following formula:

[0058]

[0059] (3) Calculate the local outlier factor of each battery, which is equal to the average of the local reachable density of the points in the k-distance neighborhood divided by the local reachable density of point x, as shown in the following formula:

[0060]

[0061] in, is the local reachable density of data point x in the sliding window M; N k (x,M) is the set of k nearest neighbors of data point x in the sliding window; reach_dist k (x, y, M) is the reachable distance from data point x to its nearest neighbor y in the sliding window; is the local anomaly factor of data point x; is the local reachability density of data point y; d(x,y) is the actual distance between point x and point y; i k,t is the farthest point among the k nearest neighbors of point x; d(x,x k,t ) is the distance from point x to the farthest point among its k nearest neighbors;

[0062] (4) When the local outlier factor of the battery is lower than F 1 , then the fault level is slight;

[0063] When the local outlier factor of the battery is higher than F 1 , then the fault level is medium;

[0064] When the local outlier factor of the battery is higher than F 2, the fault level is serious.

[0065] Compared with the prior art, the present invention has the following advantages:

[0066] The present invention extracts voltage data from the collected basic battery data, firstly judges whether an overcharge or over-discharge fault occurs according to the voltage data, then performs sliding time window processing on the voltage data to obtain corresponding feature data, including standard deviation, Shannon entropy and stacked autoencoder data, then combines each feature data, uses the quartile method to determine the corresponding threshold of each feature data, which is used to compare and judge whether a battery fault occurs, and then uses the correction coefficient to correct each feature data respectively to obtain the corresponding corrected feature data, and determines the current fault type according to the change trend of the corrected feature data; on the other hand, uses the local outlier factor algorithm to determine the current fault level by calculating the outlier factor of each feature data. Thus, the multi-dimensional features are combined with the threshold obtained by the quartile method to perform fault analysis on the battery historical data, and the fault type is determined and the fault level is divided while ensuring the accuracy and economy of the faulty battery.

[0067] The present invention normalizes the collected battery data, applies a sliding time window to calculate the eigenvalues, obtains the standard deviation within the sliding time window, and uses Shannon entropy and stacked autoencoder to reduce the dimension to obtain compressed data. Battery fault diagnosis is performed from three dimensions: statistical analysis, information volume, and machine learning, which support each other to ensure the accuracy of the fault diagnosis results.

[0068] The present invention analyzes abnormal data in voltage signals from multiple dimensions, converts the fault diagnosis problem into a voltage signal outlier detection problem, and improves the ability to diagnose early battery faults. The voltage trend, which is highly sensitive to faults, is used to characterize faults, which improves the speed of fault diagnosis. The voltage trend is extracted online through recursive calculation of normalized battery voltage, reducing the amount of calculation. The feature data is input into the local outlier factor algorithm, and the faulty battery detected by multi-feature fusion is verified, which can accurately and quickly diagnose battery faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0070] Figure 2a to 2c It is a schematic diagram of each characteristic data in Example 1;

[0071] Figure 3 is the outlier factor graph calculated in Example 1;

[0072] Figure 4 This is a schematic diagram of the sound and light alarm system in Example 2;

[0073] Figure 5This is a schematic diagram of the application process of Example 2. DETAILED DESCRIPTION

[0074] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0075] Embodiment 1

[0076] Power batteries are core components of electric vehicles and energy storage systems. Their fault diagnosis and alarm are of great significance to ensure the safety and reliability of battery packs. This scheme intends to adopt a battery fault diagnosis strategy based on standard deviation, Shannon entropy and stacked autoencoder multi-feature fusion, obtain the fault threshold through the dynamic quartile method, locate the faulty battery cell and the corresponding fault time, and diagnose the fault level of abnormal batteries with the local outlier factor algorithm. Therefore, a battery fault diagnosis method based on multi-dimensional feature fusion is proposed, such as Figure 1 As shown, the following steps are included:

[0077] S1, collect basic battery data and perform cleaning processing to extract voltage data;

[0078] S2. Determine whether an overcharge or over-discharge fault occurs based on the voltage data. If the judgment is no, execute step S3.

[0079] S3, performing sliding time window processing on the voltage data to obtain corresponding feature data, including standard deviation, Shannon entropy and stacked autoencoder data;

[0080] S4, combining each characteristic data, using the quartile method to determine the threshold value corresponding to each characteristic data, for comparison and judgment on whether a battery failure occurs, if the judgment is yes, executing step S5;

[0081] S5. Correct each characteristic data using the correction coefficient to obtain corresponding corrected characteristic data, and determine the current fault type according to the change trend of the corrected characteristic data;

[0082] The local outlier factor algorithm is used to determine the current fault level by calculating the outlier factor of each feature data.

[0083] Specifically, in step S1, the basic battery data including time, voltage, battery SOC, temperature, power, etc. are collected through the battery management system, wherein the battery voltage can be collected through a voltage sensor, and the time is converted into a Unix timestamp format and recorded in text;

[0084] After that, the acquired raw data is cleaned and preprocessed. For the data cleaning part, this solution adopts a five-step cleaning method, which specifically includes:

[0085] (1) The timestamp is used as the horizontal axis and the voltage of each battery is used as the vertical axis to form a matrix; and the cells are arranged in ascending order according to the timestamp;

[0086] (2) For missing values ​​and invalid values ​​that may appear during the actual operation process, use the normal data at the previous moment to replace a single missing value or invalid value;

[0087] (3) Replace the battery that exceeds the normal operating voltage range with the upper or lower limit of the range to prevent the normalized data results from being affected during data preprocessing;

[0088] (4) Delete the rows with repeated information (i.e., delete the rows with constant voltage, which is usually caused by too high data sampling frequency). When the data at the same time is repeated multiple times, keep one of the rows to reduce the calculation cost;

[0089] (5) Apply min-max normalization processing to the cleaned data to convert data of different ranges to the same scale to facilitate comparison and comprehensive analysis and accelerate the convergence speed of the algorithm. The specific calculation formula is shown in formula (1).

[0090]

[0091] Among them, U mn,min-max is the voltage data after min-max standardization, U mn is the voltage data of the nth battery at the mth acquisition moment, min{U n} is the minimum value of the nth battery at all sampling moments, max{U n} is the maximum value of the nth battery at all sampling moments.

[0092] In step S2, the first level fault diagnosis is performed based on the original battery data to determine whether the battery has an overcharge fault or an over-discharge fault, and the diagnosis and location of such faults are performed through the frequency, specifically including:

[0093] Obtaining voltage data of each battery at different times to establish a corresponding time-voltage (TV) curve;

[0094] Determine the normal operating voltage range (normal operating voltage upper limit and normal operating voltage lower limit) according to the battery type, and count the number of times each battery exceeds the normal operating voltage range (i.e., greater than the normal operating voltage upper limit or less than the normal operating voltage lower limit). When signs of overcharging or overdischarging appear repeatedly, it is determined that the battery has an overcharging fault or an overdischarging fault.

[0095] In step S3, the data processed in step S1 is input into the second-layer fault diagnosis algorithm to obtain rolling features under the sliding time window, including rolling standard deviation, rolling Shannon entropy and rolling stacked autoencoder data, specifically including:

[0096] (1) Extract the voltage data of the single cell and obtain the voltage matrix. Let U mn represents the voltage value of single battery n at the mth sampling time, then the average voltage of the nth battery cell at the mth sampling time is It can be calculated by formula (2):

[0097]

[0098] (2) Define the difference between the sampled voltage of each battery cell n at sampling time m in any sliding window and the average voltage of the battery pack as The length of the sliding window is M.

[0099] (3) The standard deviation of the voltage difference of cell n in any sliding window is the standard deviation of the difference between the voltage value of the cell at each sampling moment and the average voltage of n battery cells, which is calculated by formula (3):

[0100]

[0101] (4) According to G n (m) The probability matrix P is calculated by formula (4):

[0102]

[0103] (5) The elements in the obtained probability matrix P are used as the calculation input of information entropy. The calculation formula is shown in formula (5):

[0104]

[0105] (6) The stacked autoencoder is composed of multiple autoencoders stacked together. The output of the first-level autoencoder is used as the input of the second-level autoencoder, and the output of the second-level autoencoder is used as the input of the third-level autoencoder, and so on to complete the training of the deep neural network.

[0106] (7) Use the trained stacked autoencoder to extract G n (m) characteristic parameters;

[0107] (8) Thus, the feature sequence of all batteries is obtained, where the sliding window size is M, which is determined based on the feature fluctuation and the mean square error corresponding to the stacked autoencoder.

[0108] In step S4, the quartile method is used in combination with the feature data to obtain the fault threshold of the battery, thereby determining whether the battery has other types of faults besides overcharge and overdischarge. Specifically, after determining the sliding window size and obtaining the rolling feature sequence, the fault threshold of each feature is located by the quartile method, and the threshold of each feature is obtained by formulas (6) to (8).

[0109] IQR = Q 3 -Q 1 (6)

[0110] F 1 = Q 3 + 1.5 IQR (7)

[0111] F 2 = Q 3 + 3 IQR (8)

[0112] Among them, IQR is the interquartile range, Q 1 is the lower quartile, Q 3 is the upper quartile, F 1 The fault threshold is 1.5 times the IQR fault threshold, F 2 The boundary value is 3 times the IQR fault threshold.

[0113] When the standard deviation, Shannon entropy and stacked autoencoder features exceed the corresponding thresholds, the battery is judged to have fault information. When two or more feature data of a battery are detected to have fault information, the battery is judged to be a faulty battery.

[0114] In step S5, the corresponding correction coefficient is obtained through standard deviation, Shannon entropy and stacked autoencoder to assist in determining the type of battery fault, which mainly includes the location of open circuit fault and short circuit fault, specifically including:

[0115] (1) Compare the rolling feature sequence corresponding to each battery with the corresponding fault threshold. If the threshold is exceeded, the battery is judged to have fault information. If multiple features diagnose that the battery has fault information at the same time, it is judged to be a faulty battery and the fault type is further determined.

[0116] (2) Setting of the correction coefficient: when the voltage at the mth sampling point is less than the average voltage in the current sliding window, the correction coefficient is -1; when the voltage at the mth sampling point is greater than the average voltage in the current sliding window, the correction coefficient is 1; otherwise, it is 0.

[0117] (3) Each feature is corrected accordingly according to its correction coefficient. When the corrected feature data drops suddenly, it indicates that a short circuit fault exists. When the corrected feature data rises suddenly, it indicates that an open circuit fault exists.

[0118] In practical applications, the corrected feature sequence is obtained by multiplying the feature sequence by the corresponding correction coefficient, and a differential operation is performed to obtain the fluctuation of the corrected feature. When the fluctuation amplitude is too large, that is, the feature suddenly increases or decreases, it is determined that the battery has an open circuit or short circuit fault.

[0119] In step S5, the third level fault diagnosis is also performed, and the characteristic data of the second level fault diagnosis is input into the local outlier factor algorithm to calculate the outlier factor for classifying the fault level, specifically including:

[0120] (1) Obtaining the local reachable distance. According to the definition, if the actual distance to point x is less than the kth distance, it is the kth distance. If it is greater than the kth distance, it is the actual distance, as shown in formula (9):

[0121]

[0122] (2) Determination of local reachability density: The local reachability density is defined as the reciprocal of the average reachability distance from point to point x in the k-distance neighborhood, as shown in formula (10):

[0123]

[0124] (3) Calculate the local outlier factor of each battery, which is equal to the average of the local reachable density of the points in the k-distance neighborhood divided by the local reachable density of point x, as shown in formula (11):

[0125]

[0126] in, is the local reachability density of data point x in the sliding window M; N k (x,M) is the set of k nearest neighbors of data point x in the sliding window; reach_dist k (x, y, M) is the reachable distance from data point x to its nearest neighbor y in the sliding window; is the local anomaly factor of data point x; is the local reachability density of data point y; d(x,y) is the actual distance between point x and point y; i k,t is the farthest point among the k nearest neighbors of point x; d(x,x k,t ) is the distance from point x to the farthest point among its k nearest neighbors.

[0127] (4) According to formulas (7) and (8), the level threshold of fault level classification is obtained. When the local outlier factor of the battery is lower than F 1 , then the fault level is slight;

[0128] When the local outlier factor of the battery is higher than F1 , then the fault level is medium;

[0129] When the local outlier factor of the battery is higher than F 2 , the fault level is serious.

[0130] This embodiment performs fault detection on a lithium iron phosphate battery actually running in a certain energy storage power station, wherein the rated voltage of the battery is 3.2V, the rated capacity is 70Ah, the data sampling frequency is one sample every 5 minutes, and the connection mode is series connection.

[0131] After the first level of fault diagnosis (recording the number of times each battery exceeds the normal operating voltage range at all acquisition times, this embodiment verifies the battery data of the energy storage power station in actual operation, and records when the voltage is lower than 2.50V or higher than 3.65V; when the phenomenon that the voltage of a certain battery is lower than 2.50V at all acquisition times is not accidental, it is determined that there is an over-discharge fault. And when the voltage exceeds 3.65V for many times, it is determined that there is an over-charge fault), it is found that the voltage maximum value of battery #181 exceeds the specified range, but the probability is only 0.05%. Therefore, battery #181 did not have an over-charge fault. The remaining batteries may exceed the operating voltage range by at least 5%, which indicates that batteries #182-#184 and #192 have over-charge faults. And batteries #181-#192 all have over-discharge faults. The battery voltage data after cleaning is normalized and input into the second level of fault diagnosis.

[0132] Figure 2a to 2c They correspond to the characteristic values ​​corresponding to each battery. When the corresponding threshold is exceeded, it will respond that the battery has fault information. When multiple characteristic data of the battery respond to fault information at the same time, the battery is a faulty battery. In this embodiment, the standard deviation, Shannon entropy, and stacked autoencoder feature corresponding thresholds are 0.2699, 0.4736, and 0.8444, respectively.

[0133] Table 1 shows the faulty batteries diagnosed in this embodiment, where "√" indicates that fault information is detected in the feature, and "△" indicates that fault information is not detected in the feature. Among them, #83, #86, #90, #158, and #218 batteries have a sudden drop in features, which means they have a short circuit fault, while #1, #6, #7, #9, #14, #15, #69, #92, and #94 batteries have a sudden increase in features, which means they have an open circuit fault.

[0134] Table 1 Faulty batteries diagnosed by the second-layer fault diagnosis strategy

[0135] Battery ID Stacked Autoencoders Standard Deviation Shannon Entropy #1 √ √ △ #6 √ △ √ #7 √ √ △ #9 √ √ √ #14 √ △ √ #15 √ △ √ #69 √ △ √ #83 √ △ √ #86 √ △ √ #90 √ △ √ #92 √ √ √ #94 √ △ √ #158 √ √ △ #218 √ √ △

[0136] In the third-level fault diagnosis, the feature sequence is fused according to the proportion of faulty batteries detected to form a fusion feature that is input into the local outlier factor algorithm to obtain the local outlier factor value corresponding to each battery (such as Figure 3 As shown in Figure 2, the thresholds corresponding to the medium fault and severe fault levels are obtained according to the quartile method. While achieving the fault level classification, the accuracy of the first two levels of fault diagnosis is verified.

[0137] According to the battery fault diagnosis technology based on multi-dimensional feature fusion involved in this embodiment, on the basis of multi-dimensional features, combined with sliding time window and local outlier factor algorithm, a three-layer fault diagnosis strategy is formed to realize the determination and classification of battery fault types. The voltage data of the lithium iron phosphate battery in actual operation is used for method verification. The results show that the strategy can accurately locate the faulty battery while determining the fault type and classifying the fault level. This provides a possibility for online fault diagnosis of batteries in engineering. The sampling frequency of this embodiment is 5 minutes per sample. The lower the sampling frequency, the fewer data points can be collected during the test, corresponding to less calculation, and low sampling frequency does not require expensive test units, and the price of BMS (battery management system) will also be lower, while the higher the sampling frequency, the finer the data and the more accurate the estimation model. However, the results of this embodiment show that the diagnostic results of this strategy are very effective, which is very helpful for engineering applications of online fault diagnosis, because the sampling frequency of battery data is not necessarily high, and the voltage window of battery operation is not necessarily wide.

[0138] Embodiment 2

[0139] Based on the battery fault diagnosis solution described in the first embodiment, in order to facilitate users to obtain the fault detection results in a timely and accurate manner, this embodiment proposes an audible and visual alarm system, such as Figure 4 As shown, the sound and light alarm system is used to feed back the alarm signal of the corresponding fault to the user to ensure the safety of life and property of the user and surrounding personnel.

[0140] The application process of this embodiment is as follows Figure 5 As shown, the fault level obtained in the third-level fault diagnosis is sent to the user with different alarm signals. After the fault level is confirmed, in order to ensure that relevant personnel can be notified immediately when the battery system fails, an audible and visual alarm system is used to send different signals for each level so that the staff can respond quickly when a fault occurs, thereby reducing the risk of accidents and protecting the safety of passengers. When the fault is resolved, the alarm system will pause and restart the fault diagnosis for detection.

[0141] Specifically include:

[0142] (1) Based on the results of the third-level fault diagnosis and the fault level, the system activates the corresponding sound and light alarm. A minor fault will display a yellow warning light and a short alarm sound. A moderate fault will display an orange flashing light and a continuous alarm sound. A serious fault will light up a red emergency light accompanied by a long alarm sound, indicating an emergency situation.

[0143] (2) After the alarm, the system will continue to monitor the fault status to determine whether the alarm needs to be updated. If the fault persists, the system will increase the intensity of the audible or visual alarm to ensure the user's continued attention.

[0144] (3) When the fault condition is resolved, the system will release the alarm after user confirmation and return to normal monitoring status, waiting for the next fault detection.

[0145] In summary, this solution first collects the time and voltage data of lithium iron phosphate batteries through the battery management system; then cleans and preprocesses the battery data; then extracts features from the three dimensions of statistical analysis, information volume, and data dimension reduction of the preprocessed voltage data; then combines the sliding window to obtain a rolling feature sequence; uses the quartile method to determine the feature threshold, and when the feature exceeds the threshold, it is considered that the battery has fault information, and when multiple features are displayed, it is determined to be a faulty battery; then the correction coefficient is obtained through the feature sequence to assist in determining the type of battery failure. Finally, the feature data is fused and input into the local outlier factor algorithm to obtain the outlier factor, and is divided into the medium fault level and severe fault level thresholds obtained by the quartile method, thereby realizing the fault diagnosis of lithium-ion batteries.

[0146] The working data in the multi-dimensional feature fusion battery fault diagnosis technology proposed in the present invention can be directly collected in real time by the battery management system, without the need for offline experiments on the battery and additional data collection, thus reducing the workload. In addition, this method is insensitive to the sampling frequency, and accurate diagnostic results can still be obtained at low sampling frequencies, which is suitable for real-time online fault diagnosis of battery systems. Therefore, this method is more advantageous for BMS of large energy storage power stations with limited computing and storage capabilities, and is easier to implement.

Claims

1. A battery fault diagnosis method based on multi-dimensional feature fusion, characterized in that: The following steps are involved: S1, collect basic battery data and perform cleaning processing to extract voltage data; S2. Determine whether an overcharge or over-discharge fault occurs based on the voltage data. If the judgment is no, execute step S3. S3, performing sliding time window processing on the voltage data to obtain corresponding feature data, including standard deviation, Shannon entropy and stacked autoencoder data; S4, combining each characteristic data, using the quartile method to determine the threshold value corresponding to each characteristic data, for comparison and judgment on whether a battery failure occurs, if the judgment is yes, executing step S5; S5. Correct each characteristic data using the correction coefficient to obtain corresponding corrected characteristic data, and determine the current fault type according to the change trend of the corrected characteristic data; The local outlier factor algorithm is used to determine the current fault level by calculating the outlier factor of each feature data.

2. A battery fault diagnosis method based on multi-dimensional feature fusion according to claim 1, characterized in that: The step S1 specifically includes the following steps: S11, arranging the collected basic battery data in ascending order of time; S12, replacing a single missing value or invalid value with normal data from the previous moment; S13, replacing the abnormal voltage value beyond the normal working range with the upper limit or lower limit of the working voltage range; S14, delete the row with constant voltage; S15. Perform min-max normalization on the data: Among them, U mn,min-max is the voltage data after min-max standardization, U mn is the voltage data of the nth battery at the mth acquisition moment, min{U n } is the minimum value of the nth battery in the whole collection time, max{U n } is the maximum value of the nth battery at all sampling moments.

3. The battery fault diagnosis method based on multi-dimensional feature fusion according to claim 1 is characterized in that: The process of determining whether an overcharge or over-discharge fault occurs in step S2 includes: S21, obtaining voltage data of each battery at different times, and establishing a corresponding time-voltage curve; S22. Determine the normal operating voltage range according to the battery type, count the number of times each battery exceeds the normal operating voltage range, and if overcharging or overdischarging occurs repeatedly, determine that the battery has an overcharging fault or an overdischarging fault.

4. A battery fault diagnosis method based on multi-dimensional feature fusion according to claim 3, characterized in that: In step S22, the normal working voltage range includes a normal voltage upper limit and a normal voltage lower limit. If the battery voltage data is greater than the normal voltage upper limit, it is determined that overcharging occurs. If the battery voltage data is lower than the normal voltage lower limit, it is judged that over-discharge has occurred.

5. The battery fault diagnosis method based on multi-dimensional feature fusion according to claim 2 is characterized in that: The step S3 specifically comprises the following steps: S31, extract the voltage data of the single battery, obtain the voltage matrix, let U mn represents the voltage value of cell n at the mth sampling time, then the average voltage of n battery cells at the mth sampling time is for: S32, define the difference between the sampled voltage of each battery cell n at sampling time m in any sliding window and the average voltage of the battery pack as The length of the sliding window is M; S33, calculating the standard deviation of the voltage difference of a single cell n in any sliding window, specifically the standard deviation of the difference between the voltage value of the single cell at each sampling moment and the average voltage of n battery cells: S34, according to G n (m), the probability matrix P is calculated: Among them, C n (i) is the difference between the sampled voltage of each battery cell n at sampling time i in any sliding window and the average voltage of the battery pack; S35. Using the elements in the obtained probability matrix P as the calculation input of the information entropy, the information entropy, i.e., the Shannon entropy, is calculated: S36, use the pre-trained stacked autoencoder to extract G n The characteristic parameters of (m) are used to obtain the corresponding stacked autoencoder data.

6. A battery fault diagnosis method based on multi-dimensional feature fusion according to claim 5, characterized in that: The stacked autoencoder is composed of multiple autoencoders superimposed on each other. The output of the first-level autoencoder is used as the input of the second-level autoencoder, and the output of the second-level autoencoder is used as the input of the third-level autoencoder, and so on to complete the training of the deep neural network.

7. The battery fault diagnosis method based on multi-dimensional feature fusion according to claim 5 is characterized in that: The step S4 specifically comprises the following steps: S41. According to the quartile method, the threshold corresponding to each feature data is obtained: IQR=Q3-Q1 F1=Q3+1.5IQR F2=Q3+3IQR Among them, F1 is the fault threshold of 1.5 times the boundary value IQR, F2 is the fault threshold of 3 times the boundary value IQR, Q1 is the lower quartile, Q3 is the upper quartile, and IQR is the interquartile range; S42, when the standard deviation, Shannon entropy and stacked autoencoder feature data exceed the corresponding threshold value, it is determined that the battery has fault information; When a battery detects fault information in two or more characteristics, the battery is determined to be a faulty battery.

8. The battery fault diagnosis method based on multi-dimensional feature fusion according to claim 1 is characterized in that: In step S5, the correction coefficient takes a value of -1, 0 or 1. When the voltage at the mth sampling point is less than the average value of the voltage in the current sliding window, the correction coefficient takes a value of -1; When the voltage at the mth sampling point is greater than the average voltage in the current sliding window, the correction coefficient is 1; In other cases, the correction factor is 0.

9. The battery fault diagnosis method based on multi-dimensional feature fusion according to claim 8, characterized in that: In the step S5, when the corrected characteristic data suddenly drops, it is determined that a short circuit fault exists; when the corrected characteristic data suddenly rises, it is determined that an open circuit fault exists.

10. The battery fault diagnosis method based on multi-dimensional feature fusion according to claim 7, characterized in that: The process of determining the current fault level in step S5 includes: (1) Obtaining the local reachable distance. According to the definition, if the actual distance to point x is less than the kth distance, it is the kth distance. If it is greater than the kth distance, it is the actual distance, as shown in the following formula: (2) Determination of local reachability density. The local reachability density is defined as the reciprocal of the average reachability distance from a point in the k-distance neighborhood to point x, as shown in the following formula: (3) Calculate the local outlier factor of each battery, which is equal to the average of the local reachable density of the points in the k-distance neighborhood divided by the local reachable density of point x, as shown in the following formula: in, is the local reachable density of data point x in the sliding window M; N k (x,M) is the set of k nearest neighbors of data point x in the sliding window; reach_dist k (x, y, M) is the reachable distance from data point x to its nearest neighbor y in the sliding window; is the local anomaly factor of data point x; is the local reachability density of data point y; d(x,y) is the actual distance between point x and point y; i k,t is the farthest point among the k nearest neighbors of point x; d(x,x k,t ) is the distance from point x to the farthest point among its k nearest neighbors; (4) When the local outlier factor of the battery is lower than F1, the fault level is minor; When the local outlier factor of the battery is higher than F1, the fault level is medium; When the local outlier factor of the battery is higher than F2, the fault level is severe.

Images