Power battery fault early warning method

Abstract

The invention discloses a power battery fault early warning method, solves the problem of difficulty in timely early warning of faults caused by consistency difference of a power battery pack, and belongs to the field of power battery management. Comprising the following steps: preprocessing collected operation data to obtain target data; determining the length of a sliding window based on the evaluation function, segmenting the target data according to a time sequence by using the length of the sliding window, and extracting features for representing battery consistency and fluctuation anomaly characteristics in each sliding window to form a feature matrix; dynamically adjusting a fault detection threshold according to the deviation between the median and the absolute median of the target data in the sliding window; calculating a fluctuation anomaly score through a local outlier factor algorithm; if the fluctuation abnormity score exceeds a fault detection threshold value, positioning a target single battery corresponding to the fluctuation abnormity score, determining a fault type according to a characteristic matrix of the target single battery, and outputting graded early warning prompt information; according to the invention, high-precision detection and early warning of power battery faults are realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of power battery management and relates to a power battery fault warning method, and in particular to an early warning method for power battery faults based on dynamic threshold adjustment and machine learning algorithm, which is particularly suitable for the diagnosis and safety management of battery consistency problems in energy storage systems such as electric vehicles and electric bicycles. Background Art

[0002] With the rapid expansion of the new energy vehicle market, the safety and consistency of power batteries have gradually become key factors affecting vehicle performance and safety. Power battery packs are composed of multiple single cells connected in series or in parallel, and their consistency directly affects the life, safety and energy efficiency of the battery pack. However, due to differences in production processes and changes in the use environment, performance differences between cells will continue to accumulate over time, leading to increased inconsistency within the battery pack, which in turn causes the following problems:

[0003] The battery life is greatly shortened and cannot meet the needs of long-term and efficient operation of the vehicle; serious inconsistencies may induce thermal runaway accidents, threatening the safety of the vehicle and its passengers.

[0004] Existing power battery fault warning methods, such as abnormal voltage detection, temperature monitoring and warning, have the following shortcomings:

[0005] Fixed threshold: cannot adapt to complex working conditions, resulting in a high false alarm rate;

[0006] Single feature analysis: Ignores the multi-dimensional characteristics of the fault, and the diagnostic accuracy is limited;

[0007] Strong model dependence: Some data-driven methods are too dependent on specific models and parameters and lack universality.

[0008] Therefore, there is an urgent need for a new method that combines multidimensional feature analysis and adaptive early warning mechanism to achieve accurate identification and early warning of power battery failures. Summary of the invention

[0009] The present invention discloses a power battery fault early warning method, which aims to solve the technical problem that it is difficult to provide early warning in time due to the fault of the power battery pack caused by the consistency difference. The present invention realizes high-precision detection and early warning of power battery faults by combining the dynamic adjustment of fault detection threshold and the Local Outlier Factor (LOF) algorithm.

[0010] The purpose of the present invention is specifically achieved through the following technical solutions:

[0011] The present invention discloses a power battery failure early warning method, the method comprising:

[0012] Step 1: The operating data of the single battery is collected in real time through the power battery management system BMS, and the target data is obtained after pre-processing operations such as denoising, outlier processing and missing value filling.

[0013] Step 2: Determine the sliding window length based on the evaluation function, use the sliding window length to segment the target data in chronological order, and extract features used to characterize battery consistency and fluctuation abnormality characteristics in each sliding window to form a feature matrix;

[0014] Step 3: According to the median and absolute median deviation MAD of the target data in the sliding window, the fault detection threshold is dynamically adjusted based on the target operating conditions;

[0015] Step 4: Input the feature matrix into the local outlier factor algorithm LOF to calculate the fluctuation anomaly score of each single cell; if the fluctuation anomaly score exceeds the fault detection threshold, locate the target single cell corresponding to the fluctuation anomaly score, determine the fault type according to the feature matrix of the target single cell, and output graded warning prompt information according to the fault type.

[0016] In step 1, the operating data includes voltage, current, temperature and fault alarm status signal.

[0017] In step 1, the method of performing a denoising preprocessing operation on the operating data includes: using a Savitzky-Golay filtering method to perform denoising on a voltage signal in the operating data;

[0018] The method of preprocessing the operation data for outlier processing includes: if it is detected that the operation data at adjacent time points are completely consistent, only the first operation data is retained and the rest are deleted; if it is detected that the value in the operation data exceeds the abnormal threshold, it is determined to be an abnormal value and the abnormal value is eliminated;

[0019] The method of preprocessing the operation of filling missing values ​​in the running data includes: for single-point intermittent missing values, filling them with the average value of adjacent points; for running data segments with more than three consecutive missing points, directly eliminating them.

[0020] In step 2, the evaluation function is:

[0021]

[0022] Where E(h) is the output value of the evaluation function, w1 is the stability weight, r(h) is the autocorrelation of fault feature extraction under the sliding window size h, and a high autocorrelation indicates that the feature is stable in the time dimension and is less affected by noise; w2 is the weight of the discrimination, Δs max(h) is the maximum difference of the abnormal scores of the single cells under the sliding window size h. The higher the difference, the better the discrimination of the abnormal degree of the single cells. t(h) is the algorithm running time under the sliding window size h, and h is the sliding window size.

[0023] In step 2, the characteristic matrix includes but is not limited to discrete Fréchet distance DFD, margin factor CLF, waveform factor SF and / or Shannon entropy H; wherein,

[0024] Discrete Fréchet distance DFD, used to calculate the similarity of the voltage curve to the mean curve within the sliding window. Large DFD values ​​indicate long-term consistency problems.

[0025] Margin factor CLF is used to reflect the severity of the peak of the voltage signal;

[0026] The waveform factor SF is used to measure the overall smoothness of the signal. CLF and SF together characterize short-term fluctuation anomalies.

[0027] Shannon entropy H is used to quantify the uncertainty of the voltage signal. The higher the entropy value, the stronger the randomness and inconsistency.

[0028] In step 2, the discrete Fréchet distance DFD is calculated as: In the formula, v DFD,i is the discrete Fréchet distance output value; inf α,β∈[1,k] represents the distance infimum within the time range, V i α is the voltage value of the element at position α in the sliding window sequence i, is the voltage value of the element at position β in the evaluation voltage sequence; i is the index of the sliding window sequence, α represents the index range of the element in the sliding window sequence, which is an integer from 1 to k; β represents the index range of the element in the evaluation voltage sequence, which is an integer from 1 to k, and k is the maximum integer in the range of the sliding window sequence;

[0029] The calculation method of margin factor CLF is: Where, CLF i is the margin factor output value, v i is a voltage value within the range, i is the index of the sliding window sequence, k is the maximum integer within the sliding window sequence, v i,j Refers to the absolute value of the jth voltage data point in the i-th group, where j is the absolute value index of the voltage data point in the sliding window sequence i;

[0030] The calculation method of the form factor SF is: Where SF i is the waveform factor output value;

[0031] The calculation method of Shannon entropy H is: In the formula, H i is the Shannon entropy output value, i is the index of the sliding window sequence, p i,k is the probability of the kth subclassification occurring in the i-th data group, and k is the maximum integer in the sliding window sequence.

[0032] In step 3, the method for dynamically adjusting the fault detection threshold includes:

[0033]

[0034] Where Threshold(i) is the fault detection threshold, i is the index of the sliding window sequence, represents the median of the LOF anomaly score in the i-th sliding window, Θ is a constant used to adjust the sensitivity of anomaly detection, and MAD LOF,i is the absolute median deviation of the LOF anomaly score in the i-th sliding window.

[0035] In step 4, the characteristic matrix is ​​input into the local outlier factor algorithm LOF, and the method for calculating the fluctuation anomaly score of each single cell includes:

[0036] Normalize the feature matrix received by the local outlier factor algorithm LOF;

[0037] Calculate the K-neighborhood density and local reachability density of each feature point in the normalized feature matrix;

[0038] Based on the local reachable density of feature points and the average density of the neighborhood, the degree of outliers of feature points is quantified, and the fluctuation anomaly score of each single battery is output.

[0039] In step 4, the method for determining the fault type according to the characteristic matrix of the target single cell includes:

[0040] Set the fault interval corresponding to each eigenvalue in the feature matrix, and mark the corresponding fault type for each fault interval;

[0041] Each eigenvalue in the characteristic matrix of the target single battery is matched with the fault interval, and the fault type marked by the successfully matched fault interval is recorded.

[0042] In step 4, the method of outputting graded warning prompt information according to the fault type includes:

[0043] Set the graded warning prompt information corresponding to the fault type, and the graded warning prompt information includes ordinary warning prompt information and serious warning prompt information; wherein,

[0044] General warning information, used to indicate potential hazards, and it is recommended to check the battery status;

[0045] Serious warning information is used to indicate that there is a major safety risk and the vehicle needs to be immediately stopped and protective measures taken;

[0046] The corresponding graded warning prompt information is matched according to the determined fault type, the graded warning prompt information is output to the target user terminal, and the analysis data corresponding to the graded warning prompt information is recorded, wherein the analysis data includes the faulty single cell number, abnormal characteristic parameters and the corresponding timestamp.

[0047] The beneficial effects of the present invention are:

[0048] The present invention ensures the integrity and stability of the data curve by performing preprocessing operations such as denoising, outlier processing and missing value filling on the running data, avoids interference with the analysis results, and ensures the data quality;

[0049] The optimal sliding window length is determined based on the evaluation function, and the size of the sliding window determined by the present invention ensures the accuracy and real-time performance of fault diagnosis.

[0050] According to the median and absolute median deviation (MAD) of the target data in the sliding window, the fault detection threshold is dynamically adjusted to adapt to different operating conditions.

[0051] The DFD feature is used to analyze long-term consistency problems, while the CLF and SF features are suitable for detecting short-term fluctuation anomalies. Multiple features are combined to comprehensively evaluate the battery status.

[0052] The feature matrix is ​​input into the local outlier factor algorithm LOF to calculate the fluctuation anomaly score of each single cell; if the fluctuation anomaly score exceeds the fault detection threshold, the target single cell corresponding to the fluctuation anomaly score is located, and the fault type is determined according to the feature matrix of the target single cell, and graded warning information is provided according to the fault type, achieving high-precision detection and early warning of power battery faults, and solving the problem of difficulty in timely warning of faults caused by consistency differences in power battery packs.

[0053] The analysis data corresponding to the recorded warning information facilitates subsequent maintenance and data analysis.

[0054] Compared with traditional BMS, the present invention can detect faults in advance, buy more maintenance time for users, and reduce potential safety risks.

[0055] Efficient performance is demonstrated in different types of batteries and operating scenarios, such as electric vehicle and electric bicycle fault detection, showing the versatility of the approach.

[0056] The technical solution of the present invention will not cause false alarms and has high robustness under normal operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The present invention is further described in detail below based on the accompanying drawings and embodiments.

[0058] Figure 1 It is a schematic diagram of the operating data of EV1 before the fault alarm provided by the present invention.

[0059] Figure 2 This is a schematic diagram of EV1 DFD feature extraction provided by the present invention.

[0060] Figure 3 This is a schematic diagram of EV1 SF feature extraction provided by the present invention.

[0061] Figure 4 This is a schematic diagram of EV1 CF feature extraction provided by the present invention.

[0062] Figure 5 It is a schematic diagram of EV1 Shannon entropy feature extraction provided by the present invention.

[0063] Figure 6 It is a schematic diagram of the model warning result of EV1 provided by the present invention.

[0064] Figure 7 It is a schematic diagram of the operating data before the EV2 fault alarm provided by the present invention.

[0065] Figure 8 It is a schematic diagram of EV2DFD feature extraction provided by the present invention.

[0066] Fig. 9 It is a schematic diagram of EV2 SF feature extraction provided by the present invention.

[0067] Fig.10 This is a schematic diagram of EV2 CF feature extraction provided by the present invention.

[0068] Fig.11 It is a schematic diagram of EV2 Shannon entropy feature extraction provided by the present invention.

[0069] Fig.12 It is a schematic diagram of the model warning results of EV2 provided by the present invention.

[0070] Fig.13 It is a schematic diagram of the operating data of EV3 before the fault alarm provided by the present invention.

[0071] Fig.14 This is a schematic diagram of EV3DFD feature extraction provided by the present invention.

[0072] Fig.15 This is a schematic diagram of EV3 SF feature extraction provided by the present invention.

[0073] Fig.16 This is a schematic diagram of EV3 CF feature extraction provided by the present invention.

[0074] Fig.17 This is a schematic diagram of EV3 Shannon entropy feature extraction provided by the present invention.

[0075] Fig.18 It is a schematic diagram of the model warning result of EV3 provided by the present invention.

[0076] Fig.19 It is a schematic diagram of the operating data of the electric bicycle before the fault alarm provided by the present invention.

[0077] Fig. 20 It is a schematic diagram of DFD feature extraction of an electric bicycle provided by the present invention.

[0078] Fig.21 It is a schematic diagram of SF feature extraction of an electric bicycle provided by the present invention.

[0079] Fig. 22 It is a schematic diagram of CF feature extraction of an electric bicycle provided by the present invention.

[0080] Fig.23 It is a schematic diagram of Shannon entropy feature extraction of an electric bicycle provided by the present invention.

[0081] Fig.24 It is a schematic diagram of the model warning result of the electric bicycle provided by the present invention.

[0082] Fig.25 It is a schematic flow chart of a power battery fault early warning method provided by the present invention. DETAILED DESCRIPTION

[0083] like Fig.25 As shown, an embodiment of the present invention provides a power battery failure early warning method, the method comprising:

[0084] Step 1: The operating data of the single battery is collected in real time through the power battery management system BMS, and the operating data is pre-processed by denoising, outlier processing and missing value filling to ensure data quality and obtain target data;

[0085] For example, data collection is completed through the power battery management system BMS, which monitors the voltage, current, temperature and fault alarm status signal of the single battery in real time. The sampling interval is set to 30 seconds / time, and the battery operation data is continuously recorded. The acquisition device must have high precision (voltage measurement accuracy ±0.01V) and high stability, and be able to adapt to vibration and extreme temperature difference environments. The collected data is first preprocessed to ensure its accuracy. First, detect and clean up duplicate data. If the data at adjacent time points are found to be exactly the same, only the first data is retained and the rest are deleted; secondly, for abnormal values ​​that are physically impossible to appear, such as a sudden drop in voltage to 0.5V or an increase to 5V, they are directly removed to avoid interference with the analysis results; thirdly, for single-point intermittent missing values, the average value of adjacent points is used to fill, and the data segment with more than 3 consecutive missing points is directly removed to ensure data quality; finally, the Savitzky-Golay filtering method is used to reduce the noise of the voltage signal, and the polynomial fitting smooth curve in the sliding window is used to remove environmental interference signals to ensure the integrity and stability of the data curve.

[0086] Step 2: Determine the sliding window length based on the evaluation function, use the sliding window length to segment the target data in chronological order, and extract features used to characterize battery consistency and fluctuation abnormality characteristics in each sliding window to form a feature matrix;

[0087] The size of the sliding window directly affects the accuracy and real-time performance of fault diagnosis, so it is necessary to determine the optimal window length through experimental verification. During the verification process, the evaluation function E(h) is selected to comprehensively evaluate different window sizes. The evaluation function consists of the following three parts: 1) Feature extraction stability, which measures the autocorrelation of the feature value in the time dimension. The higher the stability, the less the feature is affected by noise; 2) Anomaly detection discrimination, which measures the amplification effect of the window size on the difference in LOF scores between normal monomers and abnormal monomers. The higher the discrimination, the more conducive it is to anomaly detection; 3) Computational efficiency, which examines the time cost of running the algorithm under different window sizes.

[0088] Step 3: According to the median and absolute median deviation (MAD) of the target data in the sliding window, the fault detection threshold is dynamically adjusted based on the target operating conditions to adapt to different operating conditions.

[0089] Step 4: Input the feature matrix into the local outlier factor algorithm LOF to calculate the fluctuation anomaly score of each single cell; if the fluctuation anomaly score exceeds the fault detection threshold, locate the target single cell corresponding to the fluctuation anomaly score, determine the fault type according to the feature matrix of the target single cell, and output graded warning prompt information according to the fault type. The ordinary warning prompt information in the graded warning prompt information is used to prompt potential hidden dangers, and the serious warning prompt information is used to prompt that protective measures need to be taken immediately.

[0090] In step 1, the operating data includes voltage, current, temperature and fault alarm status signal.

[0091] In step 1, the method of performing a denoising preprocessing operation on the operating data includes: using a Savitzky-Golay filtering method to perform denoising on a voltage signal in the operating data;

[0092] The method of preprocessing the operation data for outlier processing includes: if it is detected that the operation data at adjacent time points are completely consistent, only the first operation data is retained and the rest are deleted; if it is detected that the value in the operation data exceeds the abnormal threshold, it is determined to be an abnormal value and the abnormal value is eliminated;

[0093] The method of preprocessing the operation of filling missing values ​​in the running data includes: for single-point intermittent missing values, filling them with the average value of adjacent points; for running data segments with more than three consecutive missing points, directly eliminating them.

[0094] In step 2, the evaluation function is:

[0095]

[0096] Where, E(h) is the output value of the evaluation function, w1 is the stability weight, preferably set to 0.3; r(h) is the autocorrelation of fault feature extraction under the sliding window size h. A high autocorrelation indicates that the feature is stable in the time dimension and is less affected by noise; w2 is the discrimination weight, preferably set to 0.7; Δs max (h) is the maximum difference of the abnormal scores of the single cells under the sliding window size h. The higher the difference, the better the discrimination of the abnormal degree of the single cells. t(h) is the algorithm running time under the sliding window size h, and h is the sliding window size.

[0097] In step 2, the characteristic matrix includes but is not limited to discrete Fréchet distance DFD, margin factor CLF, waveform factor SF and / or Shannon entropy H; wherein,

[0098] Discrete Fréchet distance DFD, used to calculate the similarity of the voltage curve to the mean curve within the sliding window. Large DFD values ​​indicate long-term consistency problems.

[0099] Margin factor CLF is used to reflect the severity of the peak of the voltage signal;

[0100] The waveform factor SF is used to measure the overall smoothness of the signal. CLF and SF together characterize short-term fluctuation anomalies.

[0101] Shannon entropy H is used to quantify the uncertainty of the voltage signal. The higher the entropy value, the stronger the randomness and inconsistency.

[0102] In step 2, the discrete Fréchet distance DFD is calculated as: In the formula, v DFD,i is the discrete Fréchet distance output value,

[0103] inf α,β∈[1,k] It is expressed as the distance infimum within the time range, V i α is the voltage value of the element at position α in the sliding window sequence i, is the voltage value of the element at position β in the evaluation voltage sequence; i is the index of the sliding window sequence, which also refers to a specific sequence value in the variation range in the formula; α represents the index range of the element in the sliding window sequence, which is an integer from 1 to k; β represents the index range of the element in the evaluation voltage sequence, which is an integer from 1 to k, and k is the maximum integer in the sliding window sequence range;

[0104] The calculation method of margin factor CLF is: Where, CLF i is the margin factor output value, v i is a voltage value within the range, i is the index of the sliding window sequence, k is the maximum integer within the sliding window sequence, j is the absolute value index of the voltage data point in the sliding window sequence i, v i,j refers to the absolute value of the jth voltage data point in the i-th group;

[0105] The calculation method of the form factor SF is: In the formula, SF i is the waveform factor output value;

[0106] The calculation method of Shannon entropy H is: In the formula, H i is the Shannon entropy output value, i represents the index of the sliding window sequence, p i,k is the probability of the kth sub-classification occurring in the i-th data group, and k is the maximum integer in the sliding window sequence. In step 3, the method for dynamically adjusting the fault detection threshold includes:

[0107] The recognition accuracy is improved by using the median and MAD as the basis for outlier detection. The standardized absolute deviation Modified Z-score is used to distinguish outliers from normal points.

[0108] Θ = 1.482 is a constant used to adjust the sensitivity of anomaly detection.

[0109] If Z mad ≤3, then the data point x i is a normal point. Therefore, for the fluctuation anomaly score in each sliding window, the above formula is transformed and a fixed value f=6.5 is added on this basis. The specific formula is:

[0110]

[0111] Where Threshold(i) is the fault detection threshold, i is the index of the sliding window sequence, represents the median of the LOF anomaly score in the i-th sliding window, Θ is a constant used to adjust the sensitivity of anomaly detection, and MAD LOF,i is the absolute median deviation of the LOF anomaly score in the i-th sliding window.

[0112] In step 4, the characteristic matrix is ​​input into the local outlier factor algorithm LOF, and the method for calculating the fluctuation anomaly score of each single cell includes:

[0113] The feature matrix received by the local outlier factor algorithm LOF is normalized to ensure that the dimensions of all features are consistent and to avoid bias in the model due to differences in numerical ranges.

[0114] Calculate the K-neighborhood density and local reachability density of each feature point in the normalized feature matrix;

[0115] Based on the local reachable density of feature points and the average density of the neighborhood, the degree of outliers of feature points is quantified, and the fluctuation anomaly score of each single battery is output.

[0116] In step 4, the method for determining the fault type according to the characteristic matrix of the target single cell includes:

[0117] Set the fault interval corresponding to each eigenvalue in the feature matrix, and mark the corresponding fault type for each fault interval;

[0118] Each eigenvalue in the characteristic matrix of the target single battery is matched with the fault interval, and the fault type marked by the successfully matched fault interval is recorded.

[0119] For example, if the DFD value is significantly high, the corresponding fault interval marked with the fault type is: long-term consistency fault; if the CLF and SF values ​​are obviously abnormal, the corresponding fault interval marked with the fault type is: short-term fluctuation abnormality.

[0120] In step 4, the method of outputting graded warning prompt information according to the fault type includes:

[0121] Set the graded warning prompt information corresponding to the fault type, and the graded warning prompt information includes ordinary warning prompt information and serious warning prompt information;

[0122] General warning information, used to indicate potential hazards, and it is recommended to check the battery status;

[0123] Serious warning information is used to indicate that there is a major safety risk and the vehicle needs to be immediately stopped and protective measures taken;

[0124] According to the determined fault type, the corresponding graded warning prompt information is matched, the graded warning prompt information is output to the corresponding target user terminal, the content of the graded warning prompt information is prompted, and the analysis data corresponding to the graded warning prompt information is recorded. The analysis data includes the faulty single cell number, abnormal characteristic parameters and the corresponding timestamp, which is convenient for subsequent maintenance and data analysis.

[0125] In order to facilitate those skilled in the art to understand the technical solution of the present invention, specific examples are provided as follows:

[0126] Example 1: General fault warning

[0127] Figure 1 This is the running data segment before the EV1 fault alarm. According to the evaluation function, the optimal sliding window size is 50. Figure 2-Figure 5 The extracted features show that the battery pack of this vehicle has a progressive voltage anomaly. Its margin factor curve and Shannon entropy curve can only detect slight outlier fluctuations for abnormal batteries, while the DFD and form factor features can detect that the No. 83 single battery is in a relatively outlier position throughout the entire time period. The LOF algorithm results show that the threshold is exceeded when the number of iterations is 119, corresponding to 2023 / 7 / 12 15:05. Figure 6 As shown in the figure, the BMS alarm time is 2023 / 7 / 14 8:41:00, which is about one day and 10 hours in advance of the BMS to issue a first-level warning. At the same time, according to the threshold setting, the EV2 vehicle did not exceed the second-level warning threshold line during operation. This is in line with the actual situation.

[0128] Example 2: Serious fault warning

[0129] It can be seen from the voltage curve that there was inconsistency in cell 49 at the initial moment of the segment capture. The BMS system issued two common alarms at sampling points 151 and 849, corresponding to 2019 / 7 / 6 11:15:00 and 2019 / 7 / 9 14:44. The BMS finally issued a serious fault alarm at sampling point 1400, corresponding to 2019 / 7 / 12 5:03, affecting vehicle driving. In actual conditions, the vehicle was not maintained in time and continued to be used when the common inconsistency fault alarm was issued, and eventually a serious inconsistency fault occurred. According to the evaluation function, the optimal sliding window is 55, and the fault features are extracted from the driving data of vehicle No. 2, such as Figure 7 , we can see that all four indicators can highlight that cells 2 and 49 are faulty batteries, and the outlier phenomenon of cell 49 is the most obvious. Compared with the original data, the above feature extraction methods can expand the fault information. Figures 8 to 12 As shown in the figure, after the features were input into the model, the LOF algorithm results showed that battery No. 49 exceeded the common fault threshold line at the very beginning, and had the potential for inconsistency faults, 20 hours earlier than the first BMS alarm. Battery No. 2 also exceeded the threshold line at the 7th sliding window iteration, and there was also an obvious inconsistency. Subsequently, the model results showed that at the 169th sliding window iteration, the LOF score of cell 49 exceeded the severe alarm threshold, corresponding to the time of 2019 / 7 / 9 14:42, about 2 days and 10 hours earlier than the BMS system severe fault alarm.

[0130] Example 3: Normal operation verification

[0131] like Figures 13 to 18 As shown in the figure, the battery pack of an electric bicycle contains 92 cells, and the operating data collection time is from 15:54 to 23:25 on August 18, 2020. After sliding window feature extraction and LOF algorithm detection, the LOF scores of all cells did not exceed the dynamic threshold, and no false alarm occurred, verifying the high robustness of this method under normal operating conditions.

[0132] Example 4: In order to verify the wide applicability of the model of the present invention, the operating data of two electric bicycles on the Zhizu platform were selected for example verification.

[0133] like Figures 19 to 24 It can be seen that this electric bicycle issued an alarm at the 275th, 705th, and 1089th data sampling points, indicating that battery inconsistency faults occurred at these three moments. It can be seen that the three faults occurred from the end of charging to the relaxation stage, and the inconsistency gradually increased. According to the evaluation function E(K), the sliding window selection is 45. The method in this paper extracts four types of fault features, such as Fig.24As shown, the LOF algorithm results obtained by the model calculation can be seen in the 58th, 114th, and 226th iterations. The LOF scores of monomer 1 all exceeded the threshold, and no false alarms were generated. The corresponding time was 20, 15, and 23 minutes earlier than the BMS alarm. It can be clearly seen that the three LOF scores increase in sequence, indicating that the sudden inconsistency fault is becoming more and more serious, which is in line with the actual situation.

[0134] The beneficial effects of the present invention are:

[0135] The present invention ensures the integrity and stability of the data curve by performing preprocessing operations such as denoising, outlier processing and missing value filling on the running data, avoids interference with the analysis results, and ensures the data quality;

[0136] The optimal sliding window length is determined based on the evaluation function, and the size of the sliding window determined by the present invention ensures the accuracy and real-time performance of fault diagnosis.

[0137] According to the median and absolute median deviation (MAD) of the target data in the sliding window, the fault detection threshold is dynamically adjusted to adapt to different operating conditions.

[0138] The DFD feature is used to analyze long-term consistency problems, while the CLF and SF features are suitable for detecting short-term fluctuation anomalies. Multiple features are combined to comprehensively evaluate the battery status.

[0139] The feature matrix is ​​input into the local outlier factor algorithm LOF to calculate the fluctuation anomaly score of each single cell; if the fluctuation anomaly score exceeds the fault detection threshold, the target single cell corresponding to the fluctuation anomaly score is located, and the fault type is determined according to the feature matrix of the target single cell, and graded warning information is provided according to the fault type, achieving high-precision detection and early warning of power battery faults, and solving the problem of difficulty in timely warning of faults caused by consistency differences in power battery packs.

[0140] The analysis data corresponding to the recorded warning information facilitates subsequent maintenance and data analysis.

[0141] Compared with traditional BMS, the present invention can detect faults in advance, buy more maintenance time for users, and reduce potential safety risks.

[0142] Efficient performance is demonstrated in different types of batteries and operating scenarios, such as electric vehicle and electric bicycle fault detection, showing the versatility of the approach.

[0143] The technical solution of the present invention will not cause false alarms and has high robustness under normal operating conditions.

[0144] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A power battery failure early warning method, characterized in that: The method includes: Step 1: The operating data of the single battery is collected in real time through the power battery management system BMS, and the target data is obtained after pre-processing operations such as denoising, outlier processing and missing value filling. Step 2: Determine the sliding window length based on the evaluation function, use the sliding window length to segment the target data in chronological order, and extract features used to characterize battery consistency and fluctuation abnormality characteristics in each sliding window to form a feature matrix; Step 3: According to the median and absolute median deviation MAD of the target data in the sliding window, the fault detection threshold is dynamically adjusted based on the target operating conditions; Step 4: Input the feature matrix into the local outlier factor algorithm LOF to calculate the fluctuation anomaly score of each single cell; if the fluctuation anomaly score exceeds the fault detection threshold, locate the target single cell corresponding to the fluctuation anomaly score, determine the fault type according to the feature matrix of the target single cell, and output graded warning prompt information according to the fault type.

2. The method according to claim 1, characterized in that In step 1, the operating data includes voltage, current, temperature and fault alarm status signal.

3. The method according to claim 1 or 2, characterized in that In step 1, the method of performing a denoising preprocessing operation on the operating data includes: using a Savitzky-Golay filtering method to perform denoising on a voltage signal in the operating data; The method of preprocessing the operation data for outlier processing includes: if it is detected that the operation data at adjacent time points are completely consistent, only the first operation data is retained and the rest are deleted; if it is detected that the value in the operation data exceeds the abnormal threshold, it is determined to be an abnormal value and the abnormal value is eliminated; The method of preprocessing the operation of filling missing values ​​in the running data includes: for single-point intermittent missing values, filling them with the average value of adjacent points; for running data segments with more than three consecutive missing points, directly eliminating them.

4. The method according to claim 1, characterized in that In step 2, the evaluation function is: Where E(h) is the output value of the evaluation function, w1 is the stability weight, r(h) is the autocorrelation of fault feature extraction under the sliding window size h, and a high autocorrelation indicates that the feature is stable in the time dimension and is less affected by noise; w2 is the weight of the discrimination, Δs max (h) is the maximum difference of the abnormal scores of the single cells under the sliding window size h. The higher the difference, the better the discrimination of the abnormal degree of the single cells. t(h) is the algorithm running time under the sliding window size h, and h is the sliding window size.

5. The method according to claim 1, characterized in that In step 2, the characteristic matrix includes but is not limited to discrete Fréchet distance DFD, margin factor CLF, waveform factor SF and / or Shannon entropy H; wherein, Discrete Fréchet distance DFD, used to calculate the similarity of the voltage curve to the mean curve within the sliding window. Large DFD values ​​indicate long-term consistency problems. Margin factor CLF is used to reflect the severity of the peak of the voltage signal; The waveform factor SF is used to measure the overall smoothness of the signal. CLF and SF together characterize short-term fluctuation anomalies. Shannon entropy H is used to quantify the uncertainty of the voltage signal. The higher the entropy value, the stronger the randomness and inconsistency.

6. The method according to claim 5, characterized in that In step 2, the discrete Fréchet distance DFD is calculated as: In the formula, v DFD,i is the discrete Fréchet distance output value; inf α,β∈[1,k] represents the distance infimum within the time range, V i α is the voltage value of the element at position α in the sliding window sequence i, To evaluate the voltage value of the element at position β in the voltage sequence; i is the index of the sliding window sequence, α represents the index range of the elements in the sliding window sequence, which is an integer from 1 to k; β represents the index range of the elements in the evaluation voltage sequence, which is an integer from 1 to k, and k is the maximum integer in the sliding window sequence range; The calculation method of margin factor CLF is: Where, CLF i is the margin factor output value, v i is a voltage value within the range, i is the index of the sliding window sequence, k is the maximum integer within the sliding window sequence, v i,j Refers to the absolute value of the jth voltage data point in the i-th group, where j is the absolute value index of the voltage data point in the sliding window sequence i; The calculation method of the form factor SF is: Where SF i is the waveform factor output value; The calculation method of Shannon entropy H is: In the formula, H i is the Shannon entropy output value, i is the index of the sliding window sequence, p i,k is the probability of the kth subclassification occurring in the i-th data group, and k is the maximum integer in the sliding window sequence.

7. The method according to claim 1, characterized in that In step 3, the method for dynamically adjusting the fault detection threshold includes: Where Threshold(i) is the fault detection threshold, i is the index of the sliding window sequence, represents the median of the LOF anomaly score in the i-th sliding window, Θ is a constant used to adjust the sensitivity of anomaly detection, and MAD LOF,i is the absolute median deviation of the LOF anomaly score in the i-th sliding window.

8. The method according to claim 1, characterized in that In step 4, the characteristic matrix is ​​input into the local outlier factor algorithm LOF, and the method for calculating the fluctuation anomaly score of each single cell includes: Normalize the feature matrix received by the local outlier factor algorithm LOF; Calculate the K-neighborhood density and local reachability density of each feature point in the normalized feature matrix; Based on the local reachable density of feature points and the average density of the neighborhood, the degree of outliers of feature points is quantified, and the fluctuation anomaly score of each single battery is output.

9. The method according to claim 1, characterized in that In step 4, the method for determining the fault type according to the characteristic matrix of the target single cell includes: Set the fault interval corresponding to each eigenvalue in the feature matrix, and mark the corresponding fault type for each fault interval; Each eigenvalue in the characteristic matrix of the target single battery is matched with the fault interval, and the fault type marked by the successfully matched fault interval is recorded.

10. The method according to claim 9, characterized in that In step 4, the method of outputting graded warning prompt information according to the fault type includes: Set the graded warning prompt information corresponding to the fault type, and the graded warning prompt information includes ordinary warning prompt information and serious warning prompt information; wherein, General warning information, used to indicate potential hazards, and it is recommended to check the battery status; Serious warning information is used to indicate that there is a major safety risk and the vehicle needs to be immediately stopped and protective measures taken; The corresponding graded warning prompt information is matched according to the determined fault type, the graded warning prompt information is output to the target user terminal, and the analysis data corresponding to the graded warning prompt information is recorded, wherein the analysis data includes the faulty single cell number, abnormal characteristic parameters and the corresponding timestamp.

Images