Multi-fault detection and grading method for near space aircraft battery pack
By using a deep learning-based battery aging neural network and a coupled-comparison fault detection framework, the problem of early fault detection in near-space lithium battery packs has been solved, achieving accurate and rapid fault detection and classification, and improving the safety and reliability of the battery management system.
Patent Information
- Application Number
- CN202511162795.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-19
- Publication Date
- 2025-12-05
AI Technical Summary
Existing technologies struggle to efficiently detect early faults in lithium-ion battery packs in near-space environments, and traditional methods perform poorly when data is scarce, resulting in insufficient safety and reliability.
A deep learning-based battery aging neural network and a coupled comparative fault detection framework are adopted. By utilizing the current and voltage data of the battery pack, and through simulated operating conditions and training of the battery aging neural network, multiple fault detection and classification are achieved.
It enables accurate and rapid fault detection and classification with limited data, improves the safety and reliability of the battery management system, reduces reliance on large datasets, and provides real-time early warning of faults.
Smart Images
Figure CN121069189A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of battery fault diagnosis, in particular to a multi-fault detection and grading method for a near-space aircraft battery pack. BACKGROUND
[0002] With the continuous progress of aerospace technology, near-space unmanned aircrafts are increasingly widely used in communication relay and earth observation fields due to their long endurance and flexible deployment characteristics. Lithium battery packs, as the core energy components of near-space unmanned aircrafts, are responsible for ensuring their stable operation. However, lithium batteries are easily affected by factors such as vibration and overcharging in the near-space environment, which can cause problems such as faults and thermal runaway. Lithium battery systems have high nonlinearity and are affected by aging and inconsistency, making it difficult to directly measure their internal state. Early faults often have weak features and are easily masked by standard signals, and due to safety considerations, actual fault data is scarce, causing data imbalance.
[0003] In recent years, significant progress has been made in the study of lithium battery fault detection, mainly including threshold-based methods, model-based methods, and data-driven methods. Threshold-based methods monitor the deviation of key parameters such as current and voltage to identify faults by determining whether they exceed the predetermined threshold. However, early faults may not exceed the set threshold, and a low threshold may lead to false positives. Model-based methods use equivalent circuit models to capture the underlying mechanisms of battery faults, but this method relies on accurate parameter identification and is difficult to adapt to battery aging, and performs poorly in cases of limited actual data. To overcome these limitations, data-driven methods have gradually gained favor, as they directly utilize external measurement data for fault detection. Although data-driven methods can effectively model the relationship between parameters and fault types, they require a large amount of data, limiting their application in cases of limited data.
[0004] Therefore, there is an urgent need for a multi-fault detection and grading method for near-space aircraft battery packs to improve the safety of near-space unmanned aircraft energy systems. SUMMARY
[0005] The present application aims to provide a multi-fault detection and grading method for near-space aircraft battery packs. This method only uses current and voltage data of the battery pack, utilizes limited sensor data, and can accurately and quickly detect battery pack faults and fault grading, providing accurate fault information for the battery management system (BMS), thereby implementing corresponding fault-tolerant control and significantly improving the safety and reliability of the BMS.
[0006] To achieve the above-mentioned purpose, the present application provides a multi-fault detection and grading method for near-space aircraft battery packs, comprising the following steps: S1, mimic the working condition of near space vehicles, establish a random charging and discharging data set of the battery pack under normal conditions, and preprocess the data; S2, use the current and voltage data of the battery pack random charging and discharging data set to train the battery aging neural network; S3, input the battery pack current data of a dynamic test condition into the trained battery aging neural network to obtain the predicted voltage, and then subtract the predicted voltage from the actual voltage to obtain the predicted voltage error; S4, perform tensor slicing and splicing operations on the predicted voltage error and the actual current to train the coupled comparative fault detection framework to obtain a reference feature data set; S5, perform fault injection under the dynamic condition of the battery pack, and perform sliding window slicing processing on the fault data to obtain a to-be-detected slice; S6, calculate the minimum Euclidean distance between the to-be-detected slice and the reference feature data set to obtain the fault score of the corresponding slice, and realize battery pack fault detection and grading according to the fault score.
[0007] Further, in step S1, the random charging and discharging data set of the battery pack under normal conditions is established, and the specific process is as follows: S11, perform battery monomer capacity testing, select battery monomers with similar even block capacities, and divide them into two series-connected battery packs, which are divided into a random constant current group and a random constant power group; S12, perform several rounds of constant current charging and discharging cycles on the two battery packs obtained in S11, collect battery pack total voltage, total current, and voltage and current data of each battery monomer, until the battery pack produces obvious inconsistency; S13, perform random charging and discharging experiments on the battery packs obtained in S12, with m charging and discharging cycles as one large cycle. In the random constant current group, the first m-1 charging and discharging cycles of each large cycle are ordinary constant current charging and discharging; the discharge current is fixed and changes once in the last cycle, and the discharge current value is randomly selected. In the random constant power group, the first m-1 charging and discharging cycles of each large cycle are ordinary constant power charging and discharging; the discharge power is fixed and changes once in the last cycle, and the discharge power value is randomly selected.
[0008] Further, in step S2, the battery aging neural network comprises: S21, the battery pack is equivalent to a second-order RC equivalent circuit model with time-varying parameters, which is added to the network training process as a physical constraint, and the predicted terminal voltage is:
[0009] wherein, represents the terminal voltage of the battery, Vc1, Vc2, Vc3 represent the voltages of the three capacitors of the battery model, calculated from Kirchhoff's equation; S22, the battery aging neural network is composed of a data-driven model and an aging model; the input of the data-driven model is the battery pack current and the time stamp, and the output is the predicted terminal voltage under the premise of physical constraints:
[0010] wherein, represents the predicted value of the terminal voltage, represents the data-driven model, represents the input current, represents the time stamp; The aging model corrects the artificial input parameters through the number of aging cycles, and the specific formula is:
[0011] wherein, represents the initial artificial input parameter, represents the parameter corrected by aging, represents the number of cycles of the battery pack, represents the aging neural network; S23, the loss function of the battery aging neural network is composed of a logarithmic weighted loss and a monotonicity loss; wherein the logarithmic weighted loss is:
[0012] wherein, is the decay weight for the early time step, smoothly penalizes major errors while maintaining sensitivity to small deviations, denotes a weight factor, denotes the true value of the i-th time step, denotes the predicted value of the i-th time step; The monotonicity loss is: . wherein, denotes the voltage measurement value of the i-th time step, denotes the voltage measurement value of the i-1-th time step, and the loss function penalizes the voltage rising behavior at t, so that the battery aging neural network follows the physical law; The total loss function is obtained by linear combination of the logarithmic weighted loss and the monotonicity loss:
[0013] wherein represents the total loss, represents the logarithmic weighted loss, represents the monotonicity loss, Represents the weighting factor.
[0014] Furthermore, step S4 specifically includes: S41. Perform tensor slicing and splicing of the predicted voltage error and the actual current. Use a sliding window to perform tensor slicing operation and splice the error and current slices into tensors. S42. Using the data slices obtained in S41 as training samples, train the coupled comparison fault detection framework; the formulas for generating positive samples, negative samples, and benchmark samples for the coupled comparison fault detection framework are as follows:
[0015] in, Represents the baseline sample. Represents positive samples. Representing negative samples, This represents Gaussian noise added to the current data. Represents the mean of the noise. The standard deviation of noise This represents the set of all negative samples, excluding the baseline sample; In the coupled contrastive fault detection framework, the contrastive loss of a single opposing pair is defined as:
[0016] in, Cosine similarity represents the similarity between two samples. Represents exponentiation. It is a temperature coefficient used to control the sensitivity of similarity; S43. Based on the coupled comparison fault detection framework trained in S42, input battery pack data slices from several normal operating modes into the network to construct a reference feature dataset. .
[0017] Further, in step S5, the fault injection includes: S51. Based on the aged battery pack obtained in S13, perform fault injection. Select n basic test condition segments, one of which is not injected with faults, and inject common battery faults into the remaining n-1 segments respectively. S52. For each segment containing a fault, the fault injection time and fault duration are different to prevent the fault detection framework from learning the location and length information of the fault.
[0018] Furthermore, step S6 specifically includes: S61. The fault score of the sample to be tested is defined as:
[0019] wherein, a fault score representing the sample to be detected, a reference feature dataset obtained in step S43, representing the sample to be detected, representing the Euclidean distance between the sample to be detected and the reference feature dataset; S62, calculate the fault score of the reference feature dataset, and the fault threshold is obtained by Principle, the fault score of the sample to be detected is considered to have a fault in the battery pack when it exceeds the threshold, and the larger the fault score represents the higher the fault level.
[0020] The technical effects of the present application are as follows: 1. The present application is based on deep learning, which can accurately and quickly detect various faults of lithium battery packs, and realize grading according to the fault score, thereby improving the safety of the battery management system.
[0021] 2. By simulating the working condition and battery aging neural network, the fault can be effectively detected under limited data, which reduces the dependence on large data sets and enhances the adaptability in practical application.
[0022] 3. The coupling contrast learning framework is adopted, which can realize real-time early warning of early faults and improve the reliability and safety of the battery system. BRIEF DESCRIPTION OF DRAWINGS
[0023] Figure 1 is a flowchart of the multi-fault detection and grading method for near space vehicle battery packs of the present application; Figure 2 is a schematic diagram of the experimental topology in the embodiment of the present application; Figure 3 is a schematic diagram of the second-order RC equivalent circuit model of the present application; Figure 4 is a schematic diagram of the battery aging neural network structure in the embodiment of the present application; Figure 5 is a schematic diagram of the coupling contrast fault detection framework structure in the embodiment of the present application; Figure 6 is a schematic diagram of the fault data in the embodiment of the present application; Figure 7 is a detection result diagram in the embodiment of the present application. DETAILED DESCRIPTION
[0024] The technical solutions of the present application are further described below through the drawings and embodiments.
[0025] As Figure 1 shown, a multi-fault detection and grading method for near space vehicle battery packs, the method comprises the following specific steps: S1, first imitate near space vehicle working condition, establish battery pack normal condition random charging and discharging data set, and pre-process data; S2, the current voltage data of the battery pack random charging and discharging data set is used to train the battery aging neural network; S3, the battery pack current data of a dynamic test working condition is input into the trained battery aging neural network, the predicted voltage is obtained, the predicted voltage error is obtained by subtracting the actual voltage from the predicted voltage; S4, the predicted voltage error and the actual current are subjected to tensor slicing and splicing operation, the coupling comparison fault detection framework is trained, and the reference feature data set is obtained; S5, fault injection under the dynamic working condition of the battery pack, the fault data is subjected to sliding window slicing processing, and the detection slice is obtained; S6, the minimum Euclidean distance between the detection slice and the reference feature data set is calculated, the fault score of the corresponding slice is obtained, and the battery pack fault detection and grading are realized according to the fault score. Embodiment
[0026] S1, first imitate near space vehicle working condition, establish battery pack normal condition random charging and discharging data set, and pre-process data, the experimental topology is as shown in Figure 2 , wherein the host computer is used to control the charging and discharging program of the battery monomer tester and the battery module tester, and the battery tester is used to store data; S11, battery monomer capacity test is carried out, 8 battery packs with similar capacity are selected, two battery packs composed of four battery monomers in series are formed, and they are divided into random constant current group and random constant power group; S12, 40 rounds of constant current charging and discharging cycles are carried out on the two battery packs obtained in S11, wherein the charging and discharging current is 1C, the total voltage, total current and voltage and current data of each battery monomer of the battery pack are collected, and the inconsistency of the battery pack is generated.
[0027] S13, random charging and discharging experiment is carried out on the battery pack obtained in S12, five charging and discharging cycles are taken as one large cycle, in the random constant current group, the first four charging and discharging cycles of each large cycle are ordinary constant current charging and discharging, and the charging and discharging current is 1C; the discharging current changes every 60s in the last cycle, and the discharging current value is randomly selected from [2, 2.5, 3, 3.5, 4]A; in the random constant power group, the first four charging and discharging cycles of each large cycle are ordinary constant power charging and discharging, and the charging and discharging power is 40W; the discharging power changes every 60s in the last cycle, and the discharging power value is randomly selected from [20, 25, 30, 35]W; the termination condition of the two experimental groups is that the capacity of the battery pack is less than 90%.
[0028] S2, use the current-voltage data of the random charge-discharge of the battery pack for training the battery aging neural network; S21, equivalent the battery pack to a second-order RC equivalent circuit model with time-varying parameters, as shown in Figure 3 The model is added to the network training process as a physical constraint, and the predicted terminal voltage is:
[0029] wherein, represents the terminal voltage of the battery, respectively represent the voltages of the three capacitors of the battery model, which can be calculated by Kirchhoff's equation; S22, as shown in Figure 4 The battery aging neural network is composed of a data-driven model and an aging model; the input of the data-driven network is the current of the battery pack and the time stamp, and the output is the predicted terminal voltage under the premise of physical constraint:
[0030] wherein, represents the predicted value of the terminal voltage, represents the data-driven model, represents the input current, represents the time stamp.
[0031] The aging model corrects the artificial input parameters through the number of aging cycles, and the specific formula is:
[0032] wherein, represents the initial artificial input parameter, represents the parameter corrected by aging, represents the number of cycles of the battery pack, represents the aging neural network; S23, the loss function of the battery aging neural network is composed of a logarithmic weighted loss and a monotonicity loss. The logarithmic weighted loss is:
[0033] wherein, is the decay weight for the early time steps, smoothly punishes major errors while maintaining sensitivity to small deviations, refers to a weight factor, refers to the true value of the i-th time step, refers to the predicted value of the i-th time step; The monotonicity loss is: . wherein, denotes the voltage measurement at the i-th time step, denotes the voltage measurement at the i-1-th time step, the loss function penalizes the voltage rising behavior at the i-th time step, so that the battery aging neural network follows the physical law; The total loss function is obtained by linear combination of the logarithmic weighted loss and the monotonicity loss:
[0034] wherein denotes the total loss, denotes the logarithmic weighted loss, denotes the monotonicity loss, denotes the weight factor.
[0035] S3, input the battery pack current data of a dynamic test working condition into the trained battery aging neural network to obtain a predicted voltage, subtract the predicted voltage from the actual voltage to obtain a predicted voltage error; S4, perform tensor slicing and splicing operation on the predicted voltage error and the actual current to train the coupling comparison fault detection framework to obtain a reference feature data set; S41, perform tensor slicing and splicing on the predicted voltage error and the actual current, take the size of the sliced sliding window as 50 sampling points, and the sliding step as 25 sampling points, perform tensor splicing on the error and the current slices, and the format of each spliced tensor is (50, 2); S42, use the data slices obtained in S41 as training samples to train the coupling comparison fault detection framework; the formula for generating positive samples, negative samples and reference samples of the coupling comparison fault detection framework is as shown in the following formula (1), and the structure of the coupling comparison fault detection framework is as shown in the following formula (2): Figure 5 In the embodiment, the encoder selects a long short-term memory network (LSTM):
[0036] wherein, is the output of the forget gate, is the output of the input gate, is the candidate memory cell, is the memory cell, is the output of the output gate, is the hidden state, is a Sigmoid activation function, is a hyperbolic tangent activation function, are weight matrices of the forget gate, the input gate, the candidate memory cell and the output gate, respectively, is a bias term, is the hidden state at the previous time, is the input at the current time; The formula for generating positive samples, negative samples and benchmark samples of contrast learning is:
[0037] wherein, represents the benchmark sample, represents the positive sample, represents the negative sample, represents the Gaussian noise added to the current data, in the embodiment, the mean of the noise , the standard deviation of the noise ; represents all negative sample sets, all other samples except the benchmark sample; The contrast loss of a single positive pair in the coupled contrast fault detection framework is defined as:
[0038] wherein, represents the cosine similarity of the two samples, represents the exponential operation, is a temperature coefficient used to control the sensitivity of the similarity; S43, on the basis of the coupled contrast fault detection framework trained in S42, input the battery pack data slices in several normal working modes into the network to construct a reference feature data set.
[0039] S5, fault injection in the dynamic working condition of the battery pack, and the fault data is processed by sliding window slicing to obtain the detection slice; S51, on the basis of the aged battery pack obtained in S13, fault injection is performed, the benchmark test working condition is selected as the urban road cycle working condition, and 4 sections of 1400s of urban road cycle working condition are selected, one section is not injected with faults, and the remaining three sections are injected with short circuit faults, connection faults and current sensor offset faults, respectively; S52, the short circuit fault is realized by connecting a small resistance in parallel with a certain battery monomer, and is injected between 1000-1200s, the resistance values are set to ; the connection fault is realized by connecting a resistance in series between two batteries, and is injected between 600-720s, the resistance values are set to ; the current sensor offset fault is realized by subtracting a fixed offset value from the actual current in a certain time period, and is injected between 800-900s, the current offset value is set to ; the fault data schematic diagram is shown in Figure 6 , wherein, Figure 6 (a) represents the connection fault, Figure 6 (b) represents the current sensor offset fault, Figure 6(c) represents a short circuit fault.
[0040] S6, calculate the minimum Euclidean distance between the to-be-detected slice and the reference feature data set, obtain the fault score of the corresponding slice, and realize battery pack fault detection and grading according to the size of the fault score S61, the fault score of the to-be-detected sample is defined as:
[0041] wherein, represents the fault score of the to-be-detected sample, represents the reference feature data set obtained in S43, represents the to-be-detected sample, represents the Euclidean distance between the to-be-detected sample and the reference feature data set; S62, calculate the fault score of the reference feature data set, and the fault threshold is determined by In this embodiment, the fault threshold is 0.01, and if the fault score of the to-be-detected sample exceeds the threshold, it is considered that a fault occurs in the battery pack, and the greater the fault score, the higher the fault level; the fault detection result corresponding to the fault sample generated in S5 is as shown in Figure 7 wherein, Figure 7 (a) represents a connection fault, and the greater the resistance value, the more serious the fault; Figure 7 (b) represents a current sensor offset fault, and the greater the offset current value, the more serious the fault; Figure 7 (c) represents a short circuit fault, and the smaller the resistance value, the more serious the fault. It can be observed that Figure 7 The fault scores in the fault interval all exceed the threshold, and the size of the fault score corresponds to the severity of the fault, which shows that the fault detection method proposed in the patent has good detection and grading effect on various small faults in the battery pack.
[0042] Therefore, the present application aims to provide a multi-fault detection and grading method for a near-space vehicle battery pack. This method only uses the current and voltage data of the battery pack, and uses limited sensor data to accurately and quickly detect battery pack faults and fault grading, providing accurate fault information for the BMS, thereby realizing corresponding fault-tolerant control and significantly improving the safety and reliability of near-space unmanned vehicles. Unlike traditional threshold-based and model-based fault detection methods, the present application models the normal working mode of the battery pack using a small amount of normal data to accurately detect early faults, overcoming the shortcomings of traditional deep learning methods that rely on large amounts of data, and significantly reducing the cost of fault detection.
[0043] It should be pointed out finally that the above examples are only used to illustrate the technical solutions of the present application but not to limit it, and although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can still be modified or replaced equivalently, and these modifications or equivalent replacements should not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.
Claims
1. A method for multiple fault detection and classification for a near space vehicle battery pack, comprising: The method comprises the following steps: S1, simulating the working condition of a near space vehicle, establishing a random charging and discharging data set of the battery pack under normal conditions, and preprocessing the data; S2, using the current and voltage data of the battery pack random charging and discharging data set to train the battery aging neural network; S3, inputting the battery pack current data of a dynamic test working condition into the trained battery aging neural network to obtain predicted voltage, and then subtracting the predicted voltage from the actual voltage to obtain the predicted voltage error; S4, performing tensor slicing and splicing operations on the predicted voltage error and the actual current to train the coupled comparative fault detection framework to obtain a reference feature data set; S5, performing fault injection in the dynamic working condition of the battery pack, performing sliding window slicing processing on the fault data to obtain a to-be-detected slice; S6, calculating the minimum Euclidean distance between the to-be-detected slice and the reference feature data set to obtain the fault score of the corresponding slice, and realizing battery pack fault detection and grading according to the fault score.
2. A method for multiple fault detection and classification for a near space vehicle battery pack as recited in claim 1, wherein, In step S1, the random charging and discharging data set of the battery pack under normal conditions is established, and the specific process is as follows: S11, performing battery monomer capacity testing, selecting battery monomers with similar even block capacities, and dividing them into two series-connected battery packs, which are divided into a random constant current group and a random constant power group; S12, performing several rounds of constant current charging and discharging cycles on the two battery packs obtained in S11, collecting battery pack total voltage, total current, and voltage and current data of each battery monomer, until the battery packs show obvious inconsistency; S13, performing random charging and discharging experiments on the battery packs obtained in S12, with m charging and discharging cycles as one large cycle, wherein in the random constant current group, the first m-1 charging and discharging cycles of each large cycle are ordinary constant current charging and discharging; the discharge current is fixed and changes once in the last cycle, and the discharge current value is randomly selected; in the random constant power group, the first m-1 charging and discharging cycles of each large cycle are ordinary constant power charging and discharging; the discharge power is fixed and changes once in the last cycle, and the discharge power value is randomly selected.
3. A method for multiple fault detection and classification for a near space vehicle battery pack as recited in claim 2, wherein, In step S2, the battery aging neural network comprises: S21, the battery pack is equivalent to a second-order RC equivalent circuit model with time-varying parameters, which is added to the network training process as a physical constraint, and the predicted terminal voltage is: ， wherein, represents the battery terminal voltage, represents the voltage of the three capacitors of the battery model, respectively, calculated from Kirchhoff's equation; S22, the battery aging neural network is composed of a data-driven model and an aging model; the input of the data-driven model is the battery pack current and the time stamp, and the output is the predicted terminal voltage under the premise of physical constraint: ， wherein, representative end voltage prediction value, representative data driven model, representative input current, representative timestamp; The aging model corrects the artificial input parameters through the number of aging cycles, and the specific formula is: ， wherein, represents an initial human input parameter, represents an aged corrected parameter, represents a battery cycle count, represents an aging neural network; S23, the loss function of the battery aging neural network is composed of a logarithmic weighted loss and a monotonicity loss; wherein the logarithmic weighted loss is: ， wherein, is a decay weight for an early time step, smoothly penalizes major errors while maintaining sensitivity to minor deviations, denotes a weight factor, denotes a true value for the i-th time step, denotes a predicted value for the i-th time step; The monotonicity loss is: . wherein, denotes the voltage measurement at the i-th time step, denotes the voltage measurement at the i-1-th time step, the loss function penalizes the voltage rise behavior at time t, so that the battery aging neural network follows the physical laws; The total loss function is obtained by linear combination of the logarithmic weighted loss and the monotonicity loss: ， wherein represents the total loss, represents the log-weighted loss, represents the monotonicity loss, represents the weight factor.
4. A method for multiple fault detection and classification for a near space vehicle battery pack as recited in claim 3, wherein, The step S4 specifically comprises: S41, performing tensor slicing and splicing on the predicted voltage error and the actual current, using a sliding window to perform tensor slicing operation, and performing tensor splicing on the error and current slices; S42, the several data slices obtained in S41 are taken as training samples to perform training of the coupled contrast fault detection framework; and a formula for generating positive samples, negative samples and benchmark samples of the coupled contrast fault detection framework is as follows: ， wherein, represents a reference sample, represents a positive sample, represents a negative sample, represents Gaussian noise added to the current data, represents the mean of the noise, represents the standard deviation of the noise, represents the set of all negative samples, all other samples except the reference sample; The contrast loss of a single positive pair in the coupled contrast fault detection framework is defined as: ， wherein, represents a cosine similarity of two samples, represents an exponential operation, is a temperature coefficient to control the sensitivity of the similarity; S43, on the basis of the coupling contrast fault detection framework trained in S42, inputting a plurality of battery pack data slices in a normal working mode into the network to construct a reference feature data set .
5. A method for multiple fault detection and classification for a near space vehicle battery pack as recited in claim 1, wherein, In step S5, the fault injection includes: S51, on the basis of the aged battery pack obtained in S13, fault injection is performed, and n basic test working condition segments are selected, wherein one segment is not injected with a fault, and the remaining n-1 segments are respectively injected with common faults of the battery; S52, for each segment containing a fault, the fault injection time is different from the fault duration time, so as to prevent the coupled contrast fault detection framework from learning the position and length information of the fault.
6. A method for multiple fault detection and classification for a near space vehicle battery pack as recited in claim 5, wherein, The step S6 specifically includes: S61, the fault score of the to-be-detected sample is defined as: ， wherein, a failure score representative of the sample to be tested, a reference feature dataset obtained at step S43, a sample to be tested, a Euclidean distance between the sample to be tested and the reference feature dataset; S62, calculate a fault score for the reference feature dataset, the fault threshold being passed by which It is determined in principle that a fault has occurred in the battery pack if the fault score of the sample to be examined exceeds the threshold value, the greater the fault score representing a higher fault level.