Electric vehicle battery life prediction method based on deep learning

Abstract

The invention discloses an electric vehicle battery life prediction method based on deep learning, and relates to the technical field of electric vehicle power battery management and health state evaluation. The adaptive starting point of the target battery is converted into directional retrieval associated with the degradation mode from global model search, and the drift risk caused by cross-system mismatch is reduced; only a small number of expert sub-networks are activated through sparse gating, and an initial model is formed, so that sub-structures participating in modeling are consistent with target conditions, and interference of uncorrelated sub-networks on output is reduced; in addition, by freezing a feature encoder and a shared output head and limiting an update range in an activated expert sub-network, input characterization and an output aperture are kept stable, parameter modification is concentrated in a local degradation mode branch, and controllability of an adaptation process is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric vehicle power battery management and health state evaluation, and in particular to an electric vehicle battery life prediction method based on deep learning. BACKGROUND

[0002] Currently, the health management of electric vehicle power batteries needs to estimate the state of health (SOH) and the remaining useful life (RUL) of the batteries to support safety warning and operation and maintenance decision-making. In recent years, with the improvement of the data collection capability of vehicle-mounted sensors and battery management systems (BMS) on operating data (such as voltage, current, temperature, etc. during charging and discharging), data-driven methods based on deep learning are gradually used for SOH / RUL prediction, and research and engineering exploration for online prediction and life inference using sparse fragment data have appeared.

[0003] At the same time, new energy vehicle products and battery technology routes continue to evolve, and some new battery systems (such as sodium-ion batteries, semi-solid / quasi-solid routes) have begun to enter the demonstration or mass production application stage. For newly listed vehicle models or new system battery packs, there is often a lack of historical data covering the entire life cycle, making it difficult to rely solely on the data of the target battery to train a deep model to form a reliable prediction in a timely manner.

[0004] To solve the above-mentioned data scarcity problem, existing technologies use transfer learning or domain adaptation ideas: after training the model on source batteries / source working conditions with sufficient historical data, the parameters are fine-tuned using a small amount of early data of the target battery, or the feature alignment is used to reduce the data distribution difference between the source domain and the target domain, in order to improve the prediction accuracy in the target scenario.

[0005] However, when the target object is a battery with a large difference in chemical system, structural design or use condition from the source domain, the degradation mechanism and data pattern may change significantly, and the applicability of the source domain knowledge to the target domain is difficult to guarantee. The existing fine-tuning or distribution alignment-based solutions are prone to problems such as decreased prediction accuracy and insufficient result stability, thereby affecting the timely evaluation and management of new system battery life. SUMMARY

[0006] In view of the above-mentioned existing problems, the present application is proposed.

[0007] The present application provides an electric vehicle battery life prediction method based on deep learning to solve the problems of data scarcity of new system batteries, instability of transfer fine-tuning and alignment, and poor generalization.

[0008] To solve the above-mentioned technical problems, the present application provides the following technical solutions:

[0009] In a first aspect, the present application provides an electric vehicle battery life prediction method based on deep learning, which comprises:

[0010] Step S1, construct an attenuation knowledge graph and a sparse modular neural network, the attenuation knowledge graph stores the association of the condition feature vector and the degradation representation vector extracted from the source battery full life operation data; the sparse modular neural network includes a feature encoder, a priori index layer, a sparse gating layer, a plurality of expert sub-networks, a shared output head and a fine-tuning controller;

[0011] Step S2, obtain the static parameters and the first N complete charge and discharge cycle data of the target battery; form a target condition feature vector from the static parameters and the cycle data statistics, and calculate the similarity by the priori index layer and the attenuation knowledge graph to obtain a sparse gating vector; the cycle data statistics is an operation stress statistics calculated from the first N complete charge and discharge cycle data;

[0012] Step S3, activate at least one expert sub-network according to the sparse gating vector to form an initial model;

[0013] Step S4, update only the parameters of the activated expert sub-network under the control of the fine-tuning controller to obtain a final model; and output the remaining service life of the target battery by using the final model.

[0014] As a preferred scheme of the battery life prediction method based on deep learning, wherein: the construction of the attenuation knowledge graph comprises: segmenting the source battery full life operation data according to the attenuation stage, extracting the condition feature vector and the degradation representation vector of each segment, and establishing a graph structure storage with the condition feature vector as the index and the degradation representation vector as the content.

[0015] As a preferred scheme of the battery life prediction method based on deep learning, wherein: the condition feature vector at least includes one or more of the average charge and discharge rate, the state of charge interval, the temperature statistics, and the static time statistics.

[0016] As a preferred scheme of the battery life prediction method based on deep learning, wherein: the similarity is obtained from the correlation measurement of the target condition feature vector and each index vector in the graph, and the K index vectors with the highest correlation measurement scores are selected according to a predetermined number K, or the correlation measurement scores are screened according to a predetermined threshold to generate a sparse gating vector; wherein K is a predetermined positive integer.

[0017] As a preferred scheme of the battery life prediction method based on deep learning, wherein: when the threshold screening does not select an expert sub-network, the expert sub-network corresponding to the index vector with the largest similarity is selected as the activated expert sub-network.

[0018] As a preferred scheme of the battery life prediction method for electric vehicles based on deep learning, the plurality of expert sub-networks share the feature encoder and the shared output head, and each has an independent intermediate hidden layer; the sparse gating vector is used to control the participation weight or on-off of the output of each expert sub-network.

[0019] As a preferred scheme of the battery life prediction method for electric vehicles based on deep learning, the fine-tuning controller freezes the parameters of the feature encoder and the shared output head, and only performs no more than T rounds of iterative updates on the parameters of the activated expert sub-networks, and determines to stop early based on the performance of the validation set divided by the first N cycles of data.

[0020] As a preferred scheme of the battery life prediction method for electric vehicles based on deep learning, the first N complete charging and discharging cycle data are the first N complete cycles obtained in chronological order after the target battery is put into use, and N is an integer from 3 to 50.

[0021] The output caliber of the remaining useful life is one of the remaining cycle number or the remaining drivable mileage, and the corresponding uncertainty measure can be output at the same time.

[0022] In a second aspect, an embodiment of the present application provides a computer device, including a memory and a processor, and the memory stores a computer program, wherein when the computer program is executed by the processor, any step of the battery life prediction method for electric vehicles based on deep learning according to the first aspect of the present application is implemented.

[0023] In a third aspect, an embodiment of the present application provides a computer readable storage medium, which stores a computer program, wherein when the computer program is executed by a processor, any step of the battery life prediction method for electric vehicles based on deep learning according to the first aspect of the present application is implemented.

[0024] The above technical scheme can at least achieve the following beneficial effects:

[0025] In view of the problem that the source domain knowledge of new vehicle models or new chemical systems is weak and blind migration lacks guidance, by constructing a decay knowledge graph and performing prior matching on the target condition features, the starting point of adapting the target battery is changed from global model search to directional retrieval associated with the degradation mode, and the drift risk caused by cross-system mismatch is reduced.

[0026] In view of the problem that the distribution difference between the source domain and the target domain leads to unconstrained updating of the overall model parameters and large prediction fluctuations, by sparse gating, only a small number of expert sub-networks are activated to form an initial model, the sub-structures participating in modeling are consistent with the target conditions, and the interference of irrelevant sub-networks on the output is reduced.

[0027] Early data fine-tuning for small sample faces the problem of catastrophic forgetting and unstable generalization. By freezing the feature encoder and sharing the output head and limiting the update range within the activated expert subnetwork, the input representation and output caliber are kept stable, and the parameter changes are concentrated in the local degradation mode branch, enhancing the controllability of the adaptation process.

[0028] Faced with the problem that threshold screening may cause process interruption due to empty activation, the expert subnetwork with the highest similarity is activated by candidate sequence rollback, so that the gating selection still maintains a closed loop under extreme conditions, avoiding unpredictable interruption state.

[0029] Faced with the problem that early cycle data quality is uneven, missing and abnormal, introducing statistical bias, through complete cycle judgment, invalid cycle elimination, missing completion and abnormal point processing, the input statistics form a consistent caliber, improving the stability of atlas retrieval and gating selection.

[0030] Faced with the problem that the prediction credibility of the new system is difficult to express and difficult to use for health management decision-making, by outputting uncertainty metrics in the reasoning phase, the life prediction results are accompanied by stability characterization, which is convenient for risk grading and threshold triggering in management strategies. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as limiting the scope of the present application.

[0032] Figure 1 A flowchart of an embodiment of a deep learning-based electric vehicle battery life prediction method. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the present application clearer, the following will further illustrate the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0034] All terms used in the present application (including technical and scientific terms) have meanings generally understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted to have meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner.

[0035] Embodiment 1:

[0036] As shown in Figure 1 The present application proposes a deep learning-based electric vehicle battery life prediction method, comprising the following steps:

[0037] In step S1, an attenuation knowledge graph and a sparse modular neural network are constructed. The attenuation knowledge graph stores the association between the condition feature vector and the degradation representation vector extracted from the source battery full-life operation data. The sparse modular neural network includes a feature encoder, a prior index layer, a sparse gating layer, a plurality of expert sub-networks, a shared output head, and a fine-tuning controller.

[0038] The sparse modular neural network is supervised trained on the source battery data. The training data is composed of the source battery full-life operation data, the input is the cycle-level operation data and the corresponding condition feature vector, and the output supervision signal is the remaining useful life label value corresponding to each input sample. The remaining useful life label value is generated according to the health state label and the life termination criterion, and the life termination criterion is determined by the health state label reaching a predetermined termination state. During the training process, the network parameters are updated using the regression supervision target, and the regression error and stability indicators are monitored on the validation data to determine the training termination time.

[0039] The attenuation knowledge graph is composed of index nodes, content nodes, and association edges. Each index node records a condition feature vector and its source identifier, including the battery chemical system identifier, the vehicle model identifier, and the working condition identifier. Each content node records the degradation representation vector and the health state label corresponding to the index node. The association edge is used to record the one-to-one association relationship between the index node and the content node, and the similar adjacency relationship between the index nodes. The similar adjacency relationship is established by similarity calculation on the condition feature vector of the index node, and is used for candidate expansion during retrieval.

[0040] In step S2, the static parameters and the first N complete charge-discharge cycle data of the target battery are obtained. The target condition feature vector is formed from the static parameters and the cycle data statistics, and the sparse gating vector is obtained by calculating the similarity between the prior index layer and the attenuation knowledge graph. The cycle data statistics are the operation stress statistics calculated from the first N complete charge-discharge cycle data.

[0041] The static parameters are mapped and normalized after discretization to form the static parameter features. The chemical system identifier is mapped to a fixed-length embedding vector through table lookup; the vehicle model identifier is mapped to a fixed-length embedding vector through table lookup; the rated capacity and the upper limit of the charge rate are mapped to continuous features using the same dimension normalization method. The cycle data statistics are obtained from the calculation process of the aforementioned operation stress statistics features, and are concatenated with the static parameter features to form the target condition feature vector.

[0042] In step S3, at least one expert sub-network is activated according to the sparse gating vector to form an initial model.

[0043] Step S4, only update the activated expert sub-network parameters under the control of the fine-tuning controller to obtain the final model; and output the target battery remaining service life by using the final model;

[0044] In this embodiment, constructing the attenuation knowledge graph comprises: segmenting the source battery full-life operation data according to the attenuation stages, extracting the condition feature vector and the degradation representation vector of each segment, and establishing a graph structure storage with the condition feature vector as the index and the degradation representation vector as the content;

[0045] The segmentation is completed according to the change rule of the health state label. The health state label is determined by the capacity-related indicators and the internal resistance-related indicators. When the capacity-related indicators show a continuous downward trend and the downward rate changes abruptly, the corresponding position is determined as the stage boundary. When the internal resistance-related indicators show a continuous upward trend and the upward rate changes abruptly, the corresponding position is determined as the stage boundary. When the stage boundaries of the capacity-related indicators and the internal resistance-related indicators are inconsistent, the boundary that appears first is taken as the stage boundary and the adjacent segments are merged and verified. The merging and verifying are based on the monotonicity of the degradation representation vector in the segment.

[0046] The condition feature vector is formed by splicing the static parameter features and the operation stress statistical features. The static parameter features include the chemical system identifier, the rated capacity identifier and the upper limit of the charge rate. The operation stress statistical features are calculated from the sequences of the charging and discharging current, the terminal voltage, the temperature and the state of charge in the segment. The degradation representation vector is composed of the capacity change representation, the internal resistance change representation and the voltage curve shape representation. The capacity change representation includes the difference between the first and last capacity-related indicators and the change rate thereof. The internal resistance change representation includes the difference between the first and last internal resistance-related indicators and the change rate thereof. The voltage curve shape representation includes the voltage platform position feature and the platform slope feature of the representative cycle in the segment.

[0047] In this embodiment, the condition feature vector at least includes one or more of the average charging and discharging rate, the state of charge interval, the temperature statistical quantity and the static time statistical quantity.

[0048] The average charging and discharging rate is obtained by normalizing the current sequence in the cycle and the rated capacity identifier. The charging phase and the discharging phase in the cycle are counted and merged into the segment statistical value. The state of charge interval is determined by the maximum and minimum values of the state of charge sequence in the cycle, and the charging interval and the discharging interval are counted respectively. The temperature statistical quantity is calculated from the temperature sequence in the cycle, including the mean value, the extreme value and the fluctuation amplitude. The static time statistical quantity is obtained by accumulating the continuous time pieces with the absolute value of the current sequence under the static judgment condition. The static judgment condition is determined by the noise level of the current sequence and remains consistent in the same battery data.

[0049] In this embodiment, the similarity is obtained by measuring the correlation between the target conditional feature vector and each index vector in the graph, and the K index vectors with the highest correlation measurement scores are selected according to a predetermined number K, or the correlation measurement scores are filtered according to a predetermined threshold to generate sparse gating vectors; where K is a predetermined positive integer.

[0050] The target conditional feature vector and the conditional feature vectors of each index node are mapped to a unified representation vector by the same feature encoder and then normalized. The relevance measure is obtained by a weighted combination of one or both of dot product relevance and distance relevance, where dot product relevance is calculated after normalization, and distance relevance is calculated using a distance metric within the unified representation space. The relevance measure results of all index nodes are sorted to form a candidate sequence.

[0051] The sparse gating vector is generated from the candidate sequence. When selecting a predetermined number of nodes, the top-ranking index nodes in the candidate sequence are selected, and their corresponding gating positions in the expert subnetwork are activated; unselected gating positions are deactivated. When selecting a threshold, the set of index nodes whose relevance measurement results meet predetermined filtering criteria are selected, and their corresponding gating positions in the expert subnetwork are activated; gating positions that do not meet the filtering criteria are deactivated. The gating vector simultaneously generates a weight vector, which is obtained by normalizing the relevance measurement results of the activated index nodes and is used for weighting during expert fusion.

[0052] The output of the prior index layer is used to transform the correlation measurement results between the target conditional feature vector and each index vector in the decaying knowledge graph into sparse gating vectors, so as to control the activation of the expert subnetwork.

[0053] In one implementation, the process can be organized into links based on alignment, measurement, normalization, and sparsification; specifically as follows:

[0054] When the target conditional feature vector is denoted as And in the atlas the first The index vectors are denoted as To reduce the impact of differences in the units of different statistical measures on similarity, the prior index layer performs consistent linear alignment on both:

[0055] ,

[0056] in, This represents the aligned target conditional embedding vector; Indicates the aligned first 1 indexed embedding vector; This represents the projection matrix of the prior index layer; This represents the bias vector of the prior index layer; denotes the index vector number and is one-to-one corresponding to the expert subnetwork number.

[0057] The similarity can adopt cosine similarity to express the direction consistency:

[0058] ,

[0059] wherein, denotes the similarity score of the target and the th index; denotes vector transposition; denotes two-norm.

[0060] The dot product attention can also be adopted and a temperature coefficient is introduced to adjust the concentration degree of the score distribution:

[0061] ,

[0062] wherein, denotes the dot product correlation score; denotes the normalized attention weight; denotes the temperature coefficient; denotes the total number of graph index vectors and is equal to the total number of expert subnetworks; denotes the summation number.

[0063] When there is a significant correlation between conditional statistics, Mahalanobis distance can be used to suppress the virtual high similarity in the covariance direction, and the distance is mapped to the similarity:

[0064] ,

[0065] wherein, denotes the square of the Mahalanobis distance of the target and the th index; denotes the covariance matrix of the index embedding; denotes the inverse matrix of the covariance matrix. It can be estimated by the graph index embedding:

[0066] ,

[0067] wherein, denotes the mean vector of all ; denotes the diagonal perturbation coefficient; denotes the unit matrix.

[0068] After obtaining a set of scores or weights , the sparse gating vector can be generated by Top-K hard gating.

[0069] In order to For example, first take the set: ,

[0070] where, represents the selected index number set; represents the selected index number set with the largest weight; represents the Top-K number.

[0071] Then construct the sparse gating and do normalization:

[0072] ,

[0073] where, represents the non-normalized sparse gating component; represents the normalized gating component; represents the sparse gating vector.

[0074] When screening by threshold, the selected set can be defined as:

[0075] ,

[0076] where, represents the index number set that meets the threshold; represents the gating threshold.

[0077] If the case of occurs, the expert subnetwork corresponding to the largest weight can be activated according to the weight fallback logic:

[0078] ,

[0079] where, represents the index number with the largest weight; represents the number corresponding to the maximum value. Then the sparse gating vector is obtained in the same way as and in the above formula.

[0080] When it is necessary to keep the sparsification differentiable in the training phase in order to jointly optimize the prior index layer, the generation of can be replaced by a sparse normalization operator, such as Sparsemax, to obtain naturally sparse weights:

[0081] ,

[0082] where, represents the sparse normalized weight vector; represents the sparse maximization mapping operator; and take the same vector as the sparse gating vector.​

[0083] Specifically, in the above embodiments, the condition information of the target battery and the index information stored in the atlas are established to have a calculable correspondence, so that the gating result is driven by data correlation rather than fixed allocation. The prior index layer is used to map different statistical quantities to the same embedding space, reducing the bias caused by the difference in dimension and scale. The similarity measure can be selected between direction consistency, dot product correlation and covariance correlation, so that the matching under different operating conditions has an interpretable scoring basis. After normalization, the score forms a weight distribution, and then the Top-K or threshold screening is used to generate a sparse gating, so that only a small number of expert sub-networks participate in subsequent modeling, thereby reducing the interference introduced by irrelevant experts. The empty set case of threshold screening is kept at least one expert activated by the maximum weight fallback, so that the process can still be closed loop under extreme conditions. Temperature regulation is used to control the concentration of weights, so as to form a adjustable compromise between the number of experts and the dispersion of gating.

[0084] In this embodiment, when the threshold screening does not select an expert sub-network, the expert sub-network corresponding to the index vector with the largest similarity is selected as the activated expert sub-network;

[0085] When the activation set generated by the threshold screening method is empty, the expert sub-network corresponding to the first index node in the candidate sequence is selected as the only activated expert sub-network, and its gating weight is set to full weight state; when the candidate sequence contains multiple index nodes with the same correlation measure result, the unique index node is selected according to the chemical system identification priority matching rule in the source identification, and the unique index node is selected according to the vehicle model identification priority matching rule when the chemical system identification is consistent.

[0086] In this embodiment, the plurality of expert sub-networks share the feature encoder and the shared output head, and each has an independent intermediate hidden layer; the sparse gating vector is used to control the participation weight or on-off of the output of each expert sub-network;

[0087] Each expert sub-network receives a unified representation vector from the feature encoder as input and outputs an expert representation vector. The expert sub-network whose sparse gating vector is in the inactivated state sets its output expert representation vector to a zero vector; the expert sub-network whose sparse gating vector is in the activated state multiplies its output expert representation vector by the corresponding gating weight to obtain a weighted expert representation vector. The shared output head receives the sum of all weighted expert representation vectors as input and outputs the remaining service life prediction value.

[0088] In this embodiment, the fine-tuning controller freezes the parameters of the feature encoder and the shared output head, only iteratively updates the parameters of the activated expert sub-networks for no more than T rounds, and decides to stop early based on the performance of the validation set divided by the first N cycles of data;

[0089] The frozen parameter set contains all learnable parameters of the feature encoder and all learnable parameters of the shared output head. The activated expert subnetwork parameter set contains all learnable parameters of its independent intermediate hidden layers. The fine-tuning iteration is performed with the training subset obtained from the initial loop data partition as input, and the regression error is calculated on the validation subset after each iteration; when the validation subset regression error no longer decreases in continuous iterations, the iteration update is stopped and the final prediction model is output. The fine-tuning target contains both the life regression error term and the parameter shift penalty term, which is determined by the change of the activated expert subnetwork parameters relative to the initial model parameters.

[0090] In an embodiment, only updating the activated expert subnetwork parameters can be limited by the gating result of the fine-tuning controller to the trainable parameter set, and the update target and the early stopping condition are bound to the validation set obtained from the previous complete loop data partition; as follows:

[0091] The non-zero components of the sparse gating vector corresponding to the activated expert subnetwork number set can be recorded as:

[0092] ,

[0093] wherein, represents the number set of the activated expert subnetwork; represents the expert subnetwork number; represents the first component of the sparse gating vector .

[0094] The fine-tuning controller constructs the target battery complete loop data into a sample set , and divides the training set and the validation set in chronological order:

[0095] ,

[0096] wherein, represents the target sample set formed by the previous complete loop data; represents the training set; represents the validation set; represents the empty set.

[0097] In the parameter update range, the fine-tuning controller freezes the feature encoder parameters and the shared output head parameters , and freezes the parameters of the expert subnetworks that are not activated, allowing only the expert subnetwork parameters in the set to be updated. In the forward calculation, for the sample The gated weighted expert output can be written as:

[0098] ,

[0099] where, denotes the input features of the target battery sample ; denotes the latent representation obtained by the feature encoder; denotes the forward mapping of the feature encoder; denotes the gated weighted expert fusion representation; denotes the forward mapping of the th expert subnetwork; denotes the remaining useful life prediction value of the sample ; denotes the forward mapping of the shared output head; denotes the th expert subnetwork parameter.

[0100] The target update can adopt the remaining useful life regression loss and superimpose the parameter change constraint to suppress the bias to the source domain knowledge. Taking the mean square error as an example:

[0101] ,

[0102] where, denotes the remaining useful life regression loss; denotes the number of training set samples; denotes the corresponding reference remaining useful life value of the sample .

[0103] The regularization term introduced by only updating the activated expert subnetwork parameters can be constrained by the two-norm of the expert parameters relative to the initial value:

[0104] ,

[0105] where, denotes the expert parameter change constraint term; denotes the initial value of the th expert subnetwork parameter at the beginning of fine-tuning.

[0106] When it is necessary to further suppress catastrophic forgetting, the source domain importance weight can be introduced into the preservation term:

[0107] ,

[0108] where, denotes the parameter preservation term with importance weight; denotes the The parameter importance weight matrix obtained by the expert sub-network in the source domain training stage.

[0109] The fine-tuning total target can be written as:

[0110] ,

[0111] wherein, represents the fine-tuning optimization target; represents the parameter change constraint term weight coefficient; represents the parameter reservation term weight coefficient.

[0112] In the iterative update, the fine-tuning controller only performs gradient descent on the expert parameters of , and the remaining parameters remain frozen:

[0113] ,

[0114] wherein, represents the parameter of the th expert sub-network in the th iteration; represents the fine-tuning learning rate; represents the gradient operator with respect to ; represents the fine-tuning target value corresponding to the th iteration; represents the maximum number of iterations.

[0115] The early stopping based on the validation set performance can adopt a sliding window smoothing validation loss and introduce a waiting round number mechanism. The validation loss can be calculated in the same caliber as the training loss:

[0116] ,

[0117] wherein, represents the validation loss of the th iteration; represents the number of validation set samples; represents the predicted value of the th iteration on the validation sample ; represents the smoothed validation loss of the sliding window; represents the sliding window length; represents the summation number within the window.

[0118] Under the waiting round number mechanism, the effective drop threshold and the waiting count can be defined, and the iteration is stopped when the count reaches the upper limit:

[0119] ,

[0120] wherein, represents the th iteration; represents the minimum validation loss up to the th iteration; represents the effective drop threshold; represents the operation of taking the smaller value. The fine-tuning controller ends the update when the or , where represents the maximum number of allowed waiting rounds.

[0121] When the remaining useful lifetime output is given with an uncertainty measure, the validation set uncertainty convergence can be used as an additional stopping condition:

[0122] ,

[0123] where represents the average uncertainty on the validation set for the th iteration; represents the uncertainty measure output for the validation sample for the th iteration. The fine-tuning controller can stop the iteration when continuously satisfies , where represents the uncertainty change threshold; represents the uncertainty convergence waiting round number.

[0124] In practice, the parameter settings can be organized around small sample size, few rounds, and low learning rate: take 5 to 20; take to ; take 2 to 4; take 2 to 6; take to , when , let take to . The frozen layer list contains the feature encoder, shared output head, and unactivated expert subnetworks; the prior index layer and the sparse gating layer remain un-updated at this stage, so that the gating selection and the fine-tuning update keep the same boundary of action.

[0125] Specifically, the fine-tuning process of the above-mentioned embodiments limits the updateable range to a small number of expert sub-networks selected by gating, so that the small sample data of the target battery only locally corrects the degradation mode similar to its operating conditions. After freezing the feature encoder and sharing the output head, the input representation and the output caliber remain stable, and the update is concentrated in the intermediate hidden layer of the expert, reducing the fluctuations caused by parameter linkage. The regression loss provides a fitting direction for the remaining life output, and the parameter change constraint pulls the update amplitude back to the fine-tuning starting point, thereby incorporating both the target battery adaptation and the source domain knowledge preservation into the optimization goal. The validation set is obtained by time slicing the first N complete cycle data, and the validation loss is used to describe the error change of the subset not involved in the training during the iteration process. The sliding window smooths the single round fluctuation into a trend, and the waiting mechanism converts the state of continuous improvement into a stop signal, so that the number of iteration rounds converges adaptively within the upper limit of T. If the output also provides uncertainty, the change in average uncertainty can be used as an additional stopping basis to identify the stable interval of the model on the target data.

[0126] Embodiment 2:

[0127] Based on embodiment 1, the first N complete charge and discharge cycle data of the target battery are the first N complete cycles obtained in time sequence after the target battery is put into use, and N is an integer from 3 to 50; the determination of complete cycle is based on the coverage range of state of charge sequence and the change completion of current direction. The cycle that meets the predetermined minimum state of charge span and contains the charging phase and the discharging phase is determined as a complete cycle; the cycle that does not meet the predetermined minimum state of charge span is determined as an invalid cycle and is excluded from the initial cycle data; the cycle that has a missing segment in the cycle and the missing duration exceeds the predetermined upper limit is determined as an invalid cycle and is excluded; the cycle that has a missing duration not exceeding the predetermined upper limit is interpolated and completed on the time axis. The sampling point that has an abnormal peak point in the cycle and the peak amplitude exceeds the predetermined abnormal determination condition is determined as an abnormal point and is excluded, and after the exclusion, it is smoothed and completed on the time axis.

[0128] The output caliber of the remaining useful life is one of the remaining cycle number or the remaining drivable distance, and the corresponding uncertainty measure can be output simultaneously;

[0129] The uncertainty measure is generated in the inference stage. The inference stage keeps the random inactivation mechanism in the network enabled, performs multiple forward inferences on the same input sample to obtain a set of remaining useful life prediction values, and calculates the uncertainty measure from the dispersion of the set of prediction values, which is output together with the prediction mean. In another implementation, multiple initialized and structurally identical sparse modular neural networks are used to form an ensemble model, which outputs multiple sets of prediction values for the same input sample, and generates an uncertainty measure from the dispersion of the multiple sets of prediction values.

[0130] The embodiment also provides a computer device, comprising a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the deep learning based electric vehicle battery life prediction method provided in the above embodiment.

[0131] The computer device can be a terminal, and the computer device comprises a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, an operator network, NFC (Near Field Communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0132] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to realize the deep learning based electric vehicle battery life prediction method provided in the above embodiment. The storage medium can be realized by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0133] Finally, it should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, rather than limit the present application; even though the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some or all of the technical features thereof can be replaced equivalently; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

[0134] In addition, those skilled in the art will appreciate that a combination of features of different embodiments means that such combination is within the scope of the present application and forms a different embodiment. For example, all of the above embodiments can be used in any combination. The information disclosed in this section is merely intended to deepen the understanding of the general background of the present application, and should not be regarded as acknowledging or implying in any form that the information constitutes the prior art known to those skilled in the art.

Claims

1. A method for predicting the lifespan of electric vehicle batteries based on deep learning, characterized in that, Includes the following steps: Step S1: Construct a decay knowledge graph and a sparse modular neural network. The decay knowledge graph stores the association between conditional feature vectors and degradation representation vectors extracted from the full lifespan operation data of the source battery. The sparse modular neural network includes a feature encoder, a prior index layer, a sparse gating layer, multiple expert subnetworks, a shared output head, and a fine-tuning controller. Step S2: Obtain the static parameters of the target battery and the data of the first N complete charge-discharge cycles; form a target condition feature vector from the static parameters and the statistics of the cycle data, and obtain a sparse gating vector by calculating the similarity between the prior index layer and the attenuation knowledge graph; the statistics of the cycle data are the operating stress statistics calculated from the data of the first N complete charge-discharge cycles. Step S3: Activate at least one expert subnetwork according to the sparse gating vector to form an initial model; Step S4: Under the control of the fine-tuning controller, only the parameters of the activated expert sub-network are updated to obtain the final model; the remaining lifespan of the target battery is output using the final model.

2. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 1, characterized in that, The construction of the degradation knowledge graph includes: dividing the full life operation data of the source battery into segments according to the degradation stage, extracting the conditional feature vector and degradation representation vector of each segment, and establishing a graph structure for storage with the conditional feature vector as the index and the degradation representation vector as the content.

3. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 2, characterized in that, The conditional feature vector includes at least one or more of the following: average charge / discharge rate, state of charge range, temperature statistics, and resting time statistics.

4. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 1, characterized in that, The similarity is obtained by measuring the correlation between the target conditional feature vector and each index vector in the graph, and the K index vectors with the highest correlation measurement scores are selected according to a predetermined number K, or the correlation measurement scores are filtered according to a predetermined threshold to generate sparse gating vectors; where K is a predetermined positive integer.

5. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 4, characterized in that, When the threshold screening fails to select an expert subnetwork, the expert subnetwork corresponding to the index vector with the highest similarity is selected as the activated expert subnetwork.

6. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 1, characterized in that, The multiple expert subnetworks share the feature encoder and the shared output head, and each has an independent intermediate hidden layer; the sparse gating vector is used to control the participation weight or on / off state of the output of each expert subnetwork.

7. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 1, characterized in that, The fine-tuning controller freezes the parameters of the feature encoder and the shared output head, only iteratively updates the parameters of the activated expert subnetwork for no more than T rounds, and stops early based on the performance of the validation set divided by the first N cycles of data.

8. The method for predicting the lifespan of electric vehicle batteries based on deep learning according to claim 1, characterized in that, The first N complete charge-discharge cycle data are the N complete cycles obtained first in chronological order after the target battery is put into use, and N is an integer from 3 to 50; The output of the remaining service life is either the remaining number of cycles or the remaining mileage, and the corresponding uncertainty metric can be output simultaneously.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the deep learning-based electric vehicle battery life prediction method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the deep learning-based electric vehicle battery life prediction method according to any one of claims 1 to 8.

Images