New energy automobile battery multi-fault early warning method based on unbalanced data

Abstract

The invention belongs to the technical field of data mining, and particularly relates to a new energy automobile battery multi-fault early warning method based on unbalanced data, which comprises the following steps: acquiring new energy automobile battery monitoring data to be early warned, and inputting the new energy automobile battery monitoring data into a trained new energy automobile battery multi-fault early warning model, a new energy automobile battery multi-fault identification result is obtained; the new energy automobile battery multi-fault early warning model comprises a feature enhancement module, a prediction module, a sample minority mixing enhancement module, a label fine granularity management module and a label correction module. According to the method, data enhancement and loss correction strategies are combined, and the robustness and accuracy of the model to new energy automobile battery fault tag imbalance and noise data are remarkably enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of data mining, and in particular relates to a new energy vehicle battery multiple fault early warning method based on unbalanced data. Background Art

[0002] With the rapid growth of production and sales of new energy vehicles in my country, the quality problems of new energy vehicles have become increasingly prominent, and safety accidents such as fire and power loss control have occurred frequently. The safety and quality level of new energy vehicles urgently needs to be improved. According to statistics, in recent years, more than 90% of new energy vehicle fire accidents in the country occurred in the operation link. The main cause of the accident was the abnormal working state of the power battery, and its abnormal state often presents a mixed failure mode. Each battery failure type has its own unique expression characteristics, and some modes have cross-features. The specific rules are complex and changeable, and it is difficult to summarize.

[0003] There are three common methods for power battery fault diagnosis at present, mainly including methods based on physical models, methods based on signal processing, and methods based on data-driven. Among them, the data-driven fault diagnosis method uses the real-time operation data of new energy vehicles for analysis and processing, and has good accuracy and reliability. However, the fault warning mechanism of this method mainly combines expert experience to design a fault identification algorithm to label the data set, so the battery data presents uncertainty and multi-label characteristics. In addition, because the fault phenomenon accounts for a small proportion of the overall data and the proportion of each fault mode is uneven, the battery data label has class imbalance and positive and negative imbalance. This unbalanced noise data dominates the optimization process, increases the risk of noise accumulation caused by model miscalibration, and leads to prominent model overfitting, which brings challenges to multi-label classification in automotive fault diagnosis.

[0004] Therefore, the uncertainty, multi-label nature, class imbalance, and positive-negative imbalance of battery data will bring the risk of noise accumulation to data-driven methods, thereby reducing the effectiveness of identifying potential battery failures. Summary of the invention

[0005] In order to solve the above-mentioned problems in the prior art, the present invention adopts a new energy vehicle battery multiple fault early warning method based on unbalanced data, including: obtaining new energy vehicle battery monitoring data to be warned, inputting it into a trained new energy vehicle battery multiple fault early warning model, and obtaining a new energy vehicle battery multiple fault identification result; the new energy vehicle battery multiple fault early warning model includes: a feature enhancement module, a prediction module, a sample minority mixed enhancement module, a label fine-grained management module and a label correction module;

[0006] The training process of the new energy vehicle battery multi-fault warning model includes:

[0007] S1: Obtain new energy vehicle battery monitoring data X and its fault category label Input the data X into the feature enhancement module to obtain the enhanced feature X′;

[0008] S2: Enhance the feature X′ and the fault category label Combine them to get the complete data set X all ; The complete data set X all Input the prediction module to get the fault category probability; complete data set X all Includes multiple samples, each sample includes enhanced features and its fault category label;

[0009] S3: The complete dataset X all and its fault category probability input sample minority hybrid enhancement module, to obtain an enhanced data set;

[0010] S4: Input the enhanced data set into the label fine-grained management module to obtain the clean label probability of each sample in the enhanced data set; divide the enhanced data set into a non-clean label set U and a clean label set H according to the clean label probability;

[0011] S5: Input the unclean label set U into the label correction module to obtain the refurbished label set R;

[0012] S6: Calculate the total loss function value according to the fault category probability and clean label probability of the samples in the refurbished label set R and the clean label set H, optimize the model parameters according to the total loss function value, and obtain the trained new energy vehicle battery multi-fault warning model when the total loss function value is minimized.

[0013] data Fault Category Label Among them, T+1 is the number of time steps, F is the number of sensor indicators at each time step, and K is the number of fault categories;

[0014] The process of processing data X by the feature enhancement module includes:

[0015] S11: Calculate the initialization query vector Q for each time step t based on the data X t and initialize the key vector K t and the initialization query vector Q for each sensor indicator f f and initialize the key vector K f ;

[0016] S12: According to the query vector Q t and the key vector K t Calculate the spatial attention weight matrix α according to the query vector Q f and the key vector K f Calculate the channel attention weight matrix β;

[0017] S13: Perform feature fusion on the data X according to the spatial attention weight matrix α and the channel attention weight matrix β to obtain enhanced features.

[0018] Full dataset in, is the sample at time step t, x t ∈X′ is the enhanced feature at time step t, is the fault category label at time step t, T+1 is the number of time steps; the sample minority hybrid enhancement module is used for the complete dataset X all The processing process includes:

[0019] S31: For the complete dataset X all The samples are randomly sampled m times with the same probability, and the samples obtained from the m samplings are grouped into a set S;

[0020] S32: According to the probability of fault category, the complete data set X all Perform m minority sampling on the samples in , and form a set M of samples obtained by m sampling;

[0021] S33: Draw m random variables from the Beta distribution B(v,ν) and denote them as ν i , select v i and 1-v i The larger one is taken as the mixing ratio λ i , i∈{1,…,m};

[0022] S34: Take two corresponding samples from set S and set M respectively and samples Calculate features Fault category label Rarity score of and the entropy of the fault class probability Among them, i∈{1,……,m}, represents the i-th sample in the sample set S, Represents the i-th sample in the sample set M;

[0023] S35: Score based on rarity and entropy Calculate the adjustment factor δ i , using the adjustment factor δ i Adjust the mixing ratio λ i , and obtain the new mixing ratio λ′ i ;

[0024] S36: According to λ′ i For samples and samples Perform mixed enhancement to obtain mixed enhanced samples Mixing enhanced samples Add the full dataset X all ;

[0025] S37: Repeat steps SS34 to S36 until all samples in set S and set M are taken out to obtain an enhanced data set.

[0026] For the complete dataset X all The process of performing m minority sampling on the samples in includes:

[0027] S321: Calculate each sample separately based on the probability of the fault category The presence confidence score C(k,t) and absence prediction confidence D(k,t) for each fault category k;

[0028] S322: Setting weight w k , combined with the weight w k , score C(k,t) and score D(k,t) are calculated for the corresponding samples Weighted confidence score of

[0029] S323: Scoring based on weighted confidence Calculate the corresponding sample The resampling probability

[0030] S324: Based on resampling probability For the complete dataset X all The samples in are sampled with probability.

[0031] The formula for calculating the resampling probability is:

[0032]

[0033] Among them, ε 1 is a very small constant to prevent the denominator from being zero, [] is an Iverson bracket used to represent a conditional expression, j is the index of the time step, k is the index of the fault category, and K is the number of fault categories.

[0034] According to λ′ i For samples and samples Performing hybrid enhancements includes:

[0035]

[0036] The samples of the augmented dataset are represented as t′∈[0,T′], T′+1 is the number of time steps of the enhanced dataset; the process of processing each sample in the enhanced dataset by the label fine-grained management module includes:

[0037] S41: construct a Gaussian mixture model, which includes multiple Gaussian distributions, and set parameters for each Gaussian distribution; the parameters include: mixture weight, mean, and covariance;

[0038] S42: Calculate each sample Focal Loss function on fault category k Among them, Focal Loss is the focal loss function;

[0039] S43: Calculate the loss function based on the parameters of the Gaussian distribution The posterior probability under each Gaussian distribution;

[0040] S44: update the parameters of each Gaussian distribution using the posterior probability;

[0041] S45: Repeat steps S43 to S44 until the changes in all parameters before and after the update are less than the threshold ε 2 , and obtain the adjusted Gaussian mixture model;

[0042] S46: Calculate each sample based on the adjusted Gaussian mixture model The clean label probability for each class k

[0043] In step S2, the prediction module also outputs the predicted fault category label; the label correction module processes the samples in the non-clean label set U including:

[0044] S51: Get the fault category labels predicted by the previous q trainings for all samples in the non-clean label set U, and calculate The labels of the fault category k predicted by the first q trainings are consistent with the fault category labels The proportion of the same tags

[0045] S52: According to the proportion Calculate the prediction uncertainty value for each sample

[0046]

[0047] in, is the sample at time step t′ in the augmented dataset, x t′ is the enhanced feature at time step t′, is the fault category label at time step t′, For sample The set of fault category labels predicted by the first q training runs, β represents the normalization term, K represents the number of fault categories, and k represents the index of the fault category;

[0048] S53: Setting threshold ε 3 and threshold ε 4 , set the sample The refurbished sample is like and Then the fault category label of the refurbished sample is The label of fault category k in otherwise

[0049] S54: The refurbished sample Classified in refurbished label set R.

[0050] The total loss function is:

[0051]

[0052] in, is a sample in the clean label set H, is the sample in the refurbished label set R, Representation sample The loss function on fault category k is, Representation sample Loss function on fault category k, K is the number of fault categories and k is the index of the fault category.

[0053] The loss function L is:

[0054]

[0055] in, Representation sample Focal Loss function on fault category k, Representation sample The clean label probability on fault class k, ρ is a hyperparameter.

[0056] The beneficial effects of the present invention are:

[0057] 1. The present invention combines data enhancement and loss correction strategies to significantly enhance the robustness and accuracy of the model of the present invention to the imbalanced and noisy data of new energy vehicle battery fault labels; 2. Compared with the method of oversampling minority class labels and adjusting loss values, the present invention uses interpolation method to perform data enhancement through random sampling and minority sampling, ensuring sample diversity while effectively alleviating the class imbalance problem; 3. Compared with the method of treating the entire sample as a noise correction object, the present invention constructs a Gaussian mixture model based on the Focal Loss loss function for fine-grained label management, which increases the reliability of the division of clean and unclean sample sets; compared with directly discarding unclean samples, the present invention selectively corrects and renovates each label in the unclean sample set, and fully explores the data. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A flow chart of a new energy vehicle battery multi-fault warning model training provided by an embodiment of the present invention;

[0059] Figure 2 A schematic diagram of the structure of a new energy vehicle battery multi-fault warning model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] The present invention proposes a new energy vehicle battery multiple fault early warning method based on unbalanced data, the method comprising the following contents: obtaining new energy vehicle battery monitoring data to be warned, inputting the data into a trained new energy vehicle battery multiple fault early warning model, and obtaining a new energy vehicle battery multiple fault identification result; the new energy vehicle battery multiple fault early warning model comprises: a feature enhancement module, a prediction module, a sample minority mixed enhancement module, a label fine-grained management module and a label correction module;

[0062] like Figure 1 , Figure 2 As shown in the figure, the training process of the new energy vehicle battery multi-fault warning model includes:

[0063] S1: Obtain new energy vehicle battery monitoring data X and its fault category label Input the data X into the feature enhancement module to obtain the enhanced feature X′;

[0064] New energy vehicle battery monitoring data is collected every 10 seconds. The input data includes new energy vehicle battery monitoring data and its tags The new energy vehicle battery monitoring data X includes the data X of each sensor indicator f at each time step t tf , the label Y includes the fault category at each time step t T+1 is the number of time steps, accumulated from the 0th time step, corresponding to the number of data, F is the number of sensor indicators at each time step, and K is the fault category;

[0065] The feature enhancement module uses the spatial and channel dual attention mechanism to process the data X. The specific process includes:

[0066] S11: Calculate the initialization query vector Q for each time step t based on the new energy vehicle battery monitoring data X t and the key vector K t and the initialization query vector Q for each sensor indicator f f and the key vector K f ;

[0067] According to the empirical formula Determine the initialization range and use a uniform distribution to generate d from the above range k values ​​and fill to d k ×1 matrix; where d k is the number of features of the parameter matrix. Based on the above method, the initialization parameter matrix W is obtained q , W k , W v ;

[0068] The query vector Q is obtained by multiplying the parameter matrix with the matrix items in the data X along the direction of the number of sensor indicators and time steps. t , Q f and the key vector K t , K f ,Right now

[0069] S12: According to the query vector Q t and the key vector K t Calculate the spatial attention weight matrix α according to the query vector Q f and the key vector K f Calculate the channel attention weight matrix β;

[0070] Calculate the channel attention weight score Score(f) and the spatial attention weight score Score(t). Specifically, Score(t) = Q t T·K t , Score(f)=Q f T ·K f . Further, Score(t) and Score(f) are Softmax normalized to obtain α t and β f . Specifically, Among them, t′ is the index of the time step and f′ is the index of the sensor indicator.

[0071] S13: Feature fusion is performed on the new energy vehicle battery monitoring data according to the spatial attention weight matrix α and the channel attention weight matrix β to obtain enhanced features.

[0072] α T =(α 0 , ..., α t , ..., α T ), β=(β 1 , ..., β f , ..., β F ), which is the spatial attention weight matrix Channel attention weight matrix Copy α in the column direction F-1 times to form a matrix A, where each column of the matrix A is in the form of α, that is Copy β T times in the row direction to form a matrix B. Each row of the matrix B is in the form of β, that is Furthermore, element-wise multiplication is used to perform feature fusion to obtain the enhanced feature X′=A⊙B⊙X.

[0073] S2: Enhance the feature X′ and the fault category label Perform horizontal splicing to obtain a complete data set; the complete data set X all Input the prediction module to obtain the probability of fault category;

[0074] Full dataset x t ∈X′ is the enhanced feature at each time step t, is the label of time step t; for each time step t, the enhanced feature x t , if there is a fault category k, then its corresponding label otherwise

[0075] The prediction module includes a fully connected layer (Linear), which transforms each sample Input the fully connected layer separately to obtain the fault category probability of each sample Specifically, the calculation formula for the fault category probability is:

[0076]

[0077] Among them, x t The shape and size is 1×F, The shape size is 1×K,W fc is the weight matrix of the fully connected layer, with a shape of F×K, b fc is the bias matrix of the fully connected layer, with a shape size of 1×K.

[0078] S3: The complete dataset X all and its fault category probability input sample minority hybrid enhancement module, to obtain an enhanced data set;

[0079] The sample minority hybrid enhancement module combines random sampling and minority sampling based on the probability of fault categories to perform a complete dataset X. all Perform sample minority mixed enhancement to obtain an enhanced data set; the specific steps include:

[0080] S31: For the complete dataset X all The samples of are randomly sampled m times with the same probability, and the samples obtained from the m times are grouped into a set S;

[0081] Among them, each sample The probability of being selected is

[0082] S32: According to the probability of fault category, the complete data set X all Perform m minority sampling on the samples in , and form a set M of samples obtained by m sampling;

[0083] S321: Calculate each sample separately based on the probability of the fault category The presence confidence score C(k,t) and absence prediction confidence D(k,t) for each fault category k;

[0084] Set the threshold ξ∈(0.5,0.7), if the sample The probability of fault category k is Then the sample The specific category k of Classify into set Φ C(k,t) ; otherwise it is considered non-existent and is included in the set Φ D(k,t) .

[0085] Specifically,

[0086] Further,

[0087] in, express The fault class probability of fault class k in .

[0088] S322: Setting weight w based on the degree of class imbalance k , combined with the weight w k , confidence scores C(k,t) and D(k,t) are the corresponding samples Calculate a weighted confidence score

[0089] Add a dynamic weighting factor w k The confidence score of each fault category is adjusted to ensure that categories with fewer samples receive more attention during training. Specifically, Where N is the complete data set X all The total number of samples in is the label of fault category k The number of samples, ε 0 is a very small constant used to prevent the denominator from being zero.

[0090] Furthermore, combined with w k Calculate the sample weighted confidence score. Specifically, the weighted confidence score Among them, [] is Iverson brackets, which are used to represent conditional expressions, that is, when For true, is equal to 1;

[0091] S323: Calculate sample based on inverted weighted confidence score The resampling probability Furthermore, based on P′ sampling Sort the samples in descending order based on the resampling probability P′ sampling For samples Perform m probability samplings from the high probability posterior.

[0092] Specifically, Resampling probability The calculation formula is:

[0093]

[0094]

[0095] Among them, ε 1 is a very small constant to prevent the denominator from being zero, and j is the index of the time step.

[0096] S33: Draw m random variables from the Beta distribution B(ν,ν) and denote them as v i , select v iand 1-v i The larger one is taken as the mixing ratio λ i , i∈{1,…,m};

[0097] S34: Take two corresponding samples from set S and set M respectively and samples Statistical characteristics Fault category label In the complete dataset X all Number of times it appears in And according to the number of Calculation Tags Rarity score of

[0098] Specifically, where i∈{1,…,m}, represents the i-th sample in the sample set S obtained by random sampling, represents the i-th sample in the sample set M obtained by minority sampling;

[0099] S35: Calculate samples The entropy of the probability distribution of fault categories Score by rarity and entropy Calculate the adjustment factor δ i ;

[0100] Specifically, the entropy calculation formula of probability distribution is:

[0101]

[0102] Further, Among them, ω 1 Controlling the effect of category rarity, ω 2 Controls the effect of the entropy of the probability distribution.

[0103] S36: Use adjustment factor δ i Adjust the mixing ratio λ i , forming a new weight λ′ i ;

[0104] Specifically, the new weight λ′ i =λ i ×(1+δ i ).

[0105] S37: According to λ′ i By interpolating the samples and samples Perform hybrid enhancement to obtain a hybrid enhancement instance Mixing Enhanced Instances Add the full dataset X all ;

[0106] Specifically, the calculation formula for sample minority hybrid enhancement is:

[0107]

[0108] in, Indicates that in the complete data set X all The i-th sample added later.

[0109] S38: Repeat steps SS34 to S37 until all samples in set S and set M are taken out to obtain an enhanced data set.

[0110] The samples of the augmented dataset are represented as t′∈[0,T′], T′=T+m, T′+1 is the number of samples in the enhanced dataset.

[0111] S4: Input the enhanced data set into the label fine-grained management module to obtain the clean label probability of each sample in the enhanced data set; divide the enhanced data set into a non-clean label set U and a clean label set H according to the clean label probability;

[0112] The label fine-grained management module uses a Gaussian mixture model based on the Focal Loss loss function to calculate the clean label probability of each sample in the enhanced dataset; the specific steps include:

[0113] S41: Construct a Gaussian mixture model, which includes multiple Gaussian distributions n, and set the same mixture weight for each Gaussian distribution on fault category k Randomly select a set of samples from each Gaussian distribution for fault category k in the augmented dataset Initialize the mean of Focal Loss based on the sample set Around the sample set The mean Form the initial cluster and get the sample set According to the sample set Get the covariance

[0114] Around the sample set The mean Forming the initial clustering includes: In the example, the Focal Loss of all samples on fault category k is calculated, and the Focal Loss value and mean of each sample are calculated. The absolute value of the difference is arranged in ascending order, and the first The samples corresponding to the difference are included in the sample set The range can be

[0115] Specifically,

[0116] Among them, n∈{1,2}, represents the nth Gaussian distribution, the sample set and sample set The shape and size of the samples are all 1×F. Index Index For the sample set The number of samples, For the sample set The number of samples.

[0117] S42: Based on Focal Loss, Bayesian optimization technology is used to balance the weights of positive and negative samples, increase the attention level of difficult-to-classify samples, and calculate each sample Focal Loss function on fault category k

[0118] Specifically, the loss function is calculated as:

[0119]

[0120] Among them, entropy As a sample Criteria for classifying the degree of difficulty; Is the weight factor for balancing positive and negative samples, which is used to adjust the impact of class imbalance. The value is greater than that of negative samples. In the present invention, positive samples are samples whose corresponding label of the fault category is 1, and negative samples are samples whose corresponding label of the fault category is 0. γ is an adjustment factor used to reduce the weight of easy-to-classify samples and increase the weight of difficult-to-classify samples.

[0121] S43: Calculate the loss function based on the mixed weight, mean and covariance of Gaussian distribution n The posterior probability under the nth Gaussian distribution

[0122] Specifically, the calculation formula is:

[0123]

[0124] in, Representation sample The Focal Loss value belongs to the probability density of the nth Gaussian distribution, znk =1 indicates that the current consideration belongs to the fault category k under the nth Gaussian distribution, and n′ indicates the index of the Gaussian distribution.

[0125] S44: Use the posterior probability to update the mixture weight, mean and covariance of each Gaussian distribution.

[0126] The formula for updating the mixed weight is:

[0127]

[0128] The mean update formula is:

[0129]

[0130] The covariance update formula is:

[0131]

[0132] S45: Iterate S43 and S44 until the changes in the mixed weight, mean, and covariance before and after the update Δ are all less than the threshold ε 2 Preferably, ε 2 =0.001.

[0133] S46: Sample calculation of mixture weights, mean and covariance based on adjusted Gaussian mixture model The posterior probability of class k The clean label probability is obtained by summing the posterior probabilities under the two Gaussian distributions

[0134] Specifically, If exists Then the sample It is classified into the non-clean label set U, otherwise it is classified into the clean label set H.

[0135] S5: Input the unclean label set U into the label correction module to obtain the refurbished label set R;

[0136] In step S2, the fully connected layer also outputs the predicted fault category label; the label correction module selectively modifies the label based on the low prediction uncertainty value and gradually increases the number of valid training samples. The specific steps include:

[0137] S51: Get the fault category labels predicted by the previous q trainings for all samples in the non-clean label set U, and calculate The labels of the fault category k predicted by the first q trainings are consistent with the fault category labels The proportion of the same tags

[0138] S52: According to the proportion Calculate the prediction uncertainty value for each sample

[0139]

[0140] in, is the sample at time step t′ in the augmented dataset, x t′ is the enhanced feature at time step t′, is the fault category label at time step t′, For sample The set of fault category labels predicted by the first q training runs, Representation sample The qth label prediction of fault categories 1 to K, z is the index of the first q trainings, β represents a normalization term, which is used to scale the prediction uncertainty value to [0,1]. Indicated in the sample The labels of the fault category k predicted by the first q trainings are consistent with the fault category labels The proportion of identical labels;

[0141] S53: Consider samples with consistent label predictions as refurbishable. Specifically, set the threshold ε 3 and threshold ε 4 , set the sample The refurbished sample is like and Then the fault category label of the refurbished sample is The label of fault category k in otherwise Among them, ε 3 ∈{0.05,0.10,0.15,0.20}, ε 4 ∈[0.5,1);

[0142] S54: The refurbished sample Classified in refurbished label set R.

[0143] S6: Calculate the total loss function value according to the fault category probability and clean label probability of the samples in the refurbished label set R and the clean label set H, and perform back propagation to optimize the model parameters according to the total loss function value. When the total loss function value is minimized, the trained new energy vehicle battery multi-fault warning model is obtained.

[0144] The total loss function L all for:

[0145]

[0146] in, is a sample in the clean label set H, is the sample in the refurbished label set R, Representation sample The loss function on fault category k is, Representation sample Loss function on fault class k.

[0147] The loss function L is:

[0148]

[0149] in, Representation sample Focal Loss function on fault category k, Representation sample The clean label probability on fault class k, ρ is a hyperparameter with a value of [0,1] that controls the degree of attenuation.

[0150] The back propagation calculation formula is:

[0151]

[0152] Among them, θ is the model parameter set and η is the learning rate.

[0153] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A new energy vehicle battery multiple fault early warning method based on unbalanced data, characterized in that: include: Obtain the new energy vehicle battery monitoring data to be warned, input it into the trained new energy vehicle battery multi-fault warning model, and obtain the new energy vehicle battery multi-fault identification result; The new energy vehicle battery multi-fault warning model includes: feature enhancement module, prediction module, sample minority mixed enhancement module, label fine-grained management module and label correction module; The training process of the new energy vehicle battery multi-fault warning model includes: S1: Obtain new energy vehicle battery monitoring data X and its fault category label Input data X into the feature enhancement module to obtain enhanced features X ′ ; S2: Enhance feature X ′ and fault category labels Combine them to get the complete data set X all ; The complete data set X all Input the prediction module to get the fault category probability; complete data set X all Includes multiple samples, each sample includes enhanced features and its fault category label; S3: The complete dataset X all and its fault category probability input sample minority hybrid enhancement module, to obtain an enhanced data set; S4: Input the enhanced dataset into the label fine-grained management module to obtain the clean label probability of each sample in the enhanced dataset; divide the enhanced dataset into a non-clean label set U and a clean label set H according to the clean label probability; S5: Input the unclean label set U into the label correction module to obtain the refurbished label set R; S6: Calculate the total loss function value according to the fault category probability and clean label probability of the samples in the refurbished label set R and the clean label set H, optimize the model parameters according to the total loss function value, and obtain the trained new energy vehicle battery multi-fault warning model when the total loss function value is minimized.

2. According to claim 1, a new energy vehicle battery multiple fault early warning method based on unbalanced data is characterized in that: data Fault Category Label Among them, T+1 is the number of time steps, F is the number of sensor indicators at each time step, and K is the number of fault categories; The process of processing data X by the feature enhancement module includes: S11: Calculate the initialization query vector Q for each time step t based on the data X t and initialize the key vector K t and the initialization query vector Q for each sensor indicator f f and initialize the key vector K f ; S12: According to the query vector Q t and the key vector K t Calculate the spatial attention weight matrix α according to the query vector Q f and the key vector K f Calculate the channel attention weight matrix β; S13: Perform feature fusion on the data X according to the spatial attention weight matrix α and the channel attention weight matrix β to obtain enhanced features.

3. According to claim 1, a new energy vehicle battery multiple fault early warning method based on unbalanced data is characterized in that: Full dataset in, is the sample at time step t, x t ∈X ′ is the enhanced feature at time step t, is the fault category label at time step t, T+1 is the number of time steps; the sample minority hybrid enhancement module is used for the complete dataset X all The processing process includes: S31: For the complete dataset X all The samples are randomly sampled m times with the same probability, and the samples obtained from the m samplings are grouped into a set S; S32: According to the probability of fault category, the complete data set X all Perform m minority sampling on the samples in , and form a set M of samples obtained by m sampling; S33: Draw m random variables from the Beta distribution B(v,v) and record them as v i , select v i and 1-v i The larger one is taken as the mixing ratio λ i , i∈{1,…,m}; S34: Take two corresponding samples from set S and set M respectively and samples Calculate features Fault category label Rarity score of and the entropy of the fault class probability Among them, i∈{1,……,m}, represents the i-th sample in the sample set S, Represents the i-th sample in the sample set M; S35: Score based on rarity and entropy Calculate the adjustment factor δ i , using the adjustment factor δ i Adjust the mixing ratio λ i , and get the new mixing ratio λ ′ i ; S36: According to λ ′ i For samples and samples Perform mixed enhancement to obtain mixed enhanced samples Mixing enhanced samples Add the full dataset X all ; S37: Repeat steps SS34 to S36 until all samples in set S and set M are taken out to obtain an enhanced data set.

4. A new energy vehicle battery multiple fault early warning method based on unbalanced data according to claim 3, characterized in that: For the complete dataset X all The process of performing m minority sampling on the samples in includes: S321: Calculate each sample separately based on the probability of the fault category The presence confidence score C(k,t) and absence prediction confidence D(k,t) for each fault category k; S322: Setting weight w k , combined with the weight w k , score C(k,t) and score D(k,t) are calculated for the corresponding samples Weighted confidence score of S323: Scoring based on weighted confidence Calculate the corresponding sample The resampling probability S324: Based on resampling probability For the complete dataset X all The samples in are sampled with probability.

5. A new energy vehicle battery multiple fault early warning method based on unbalanced data according to claim 4, characterized in that: The formula for calculating the resampling probability is: Among them, ε1 is a very small constant to prevent the denominator from being zero, [] is the Iverson brackets used to represent the conditional expression, j is the index of the time step, k is the index of the fault category, and K is the number of fault categories.

6. A new energy vehicle battery multiple fault early warning method based on unbalanced data according to claim 3, characterized in that: According to λ ′ i For samples and samples Performing hybrid enhancements includes:

7. The method for early warning of multiple faults of new energy vehicle batteries based on unbalanced data according to claim 1 is characterized in that: The samples of the augmented dataset are represented as t ′ ∈[0,T ′ ], T ′ +1 is the number of time steps of the augmented dataset; The process of processing each sample in the enhanced dataset by the label fine-grained management module includes: S41: construct a Gaussian mixture model, which includes multiple Gaussian distributions, and set parameters for each Gaussian distribution; the parameters include: mixture weight, mean, and covariance; S42: Calculate each sample Focal Loss function for each fault category k Among them, FocalLoss is the focal loss function; S43: Calculate the loss function based on the parameters of the Gaussian distribution The posterior probability under each Gaussian distribution; S44: update the parameters of each Gaussian distribution using the posterior probability; S45: repeat steps S43 to S44 until the changes of all parameters before and after the update are less than the threshold ε2, and obtain the adjusted Gaussian mixture model; S46: Calculate each sample based on the adjusted Gaussian mixture model The clean label probability for each fault class k 8. The method for early warning of multiple faults of new energy vehicle batteries based on unbalanced data according to claim 1 is characterized in that: In step S2, the prediction module also outputs the predicted fault category label; the label correction module processes the samples in the non-clean label set U including: S51: Get the fault category label set predicted by the previous q trainings for all samples in the non-clean label set U, and calculate The labels of the fault category k predicted by the first q trainings are consistent with the fault category labels The proportion of the same tags S52: According to the proportion Calculate the prediction uncertainty value for each sample in, is the time step t in the augmented dataset ′ A sample of x t′ is the time step t ′ The enhanced features of is the time step t ′ The fault category label, For sample The set of fault category labels predicted by the first q training runs, β represents the normalization term, K represents the number of fault categories, and k represents the index of the fault category; S53: Set threshold ε3 and threshold ε4, set sample The refurbished sample is like and Then the fault category label of the refurbished sample is The label of fault category k in otherwise S54: The refurbished sample Classified in refurbished label set R.

9. The method for early warning of multiple faults of new energy vehicle batteries based on unbalanced data according to claim 1 is characterized in that: The total loss function L all for: in, is a sample in the clean label set H, is the sample in the refurbished label set R, Representation sample The loss function on fault category k is, Representation sample The loss function on fault category k, K is the number of fault categories and k is the index of the fault category.

10. A new energy vehicle battery multiple fault early warning method based on unbalanced data according to claim 9, characterized in that: The loss function L is: in, Representation sample Focal Loss loss function on fault category k, FocalLoss is the focal loss function, Representation sample The clean label probability on fault class k, ρ is a hyperparameter.

Images