Battery health state prediction method and device, vehicle and storage medium

Abstract

The invention discloses a battery health state prediction method and device, a vehicle and a computer readable storage medium, and the method comprises the steps: obtaining a plurality of pieces of battery detection data, the data comprising a sequence of voltage, current and capacity changing with time; organizing the battery detection data into a meta-training task set containing a support set and a query set; constructing a neural network model containing a shared parameter basic network and a task specific prediction network; in the meta-training stage, the task specific prediction network is updated through inner ring adaptation, and the shared parameter basic network is updated and optimized through an outer ring; in the deployment stage, a small amount of monitoring data of a target battery is used for inner ring adaptation and health state prediction. By applying the technical scheme of the invention, the problem of poor model generalization ability caused by data scarcity under a new battery model or a new working condition can be solved, and rapid adaptation and high-precision prediction are realized.

Description

Technical Field

[0001] This application relates to the field of vehicle battery management technology, specifically to a method, device, vehicle, and computer-readable storage medium for predicting the health status of a battery. Background Technology

[0002] With the rapid development of new energy vehicles, accurate prediction of battery health status has become a key technology for ensuring safe system operation and extending service life. Against the backdrop of accelerated battery technology iteration and diversified application scenarios, deep learning-based battery health status prediction methods are gradually becoming the mainstream technical solution. These technologies employ supervised learning, relying on a large amount of complete lifecycle data from similar batteries for model training, ultimately achieving the prediction of battery health status.

[0003] However, the model's insufficient adaptability becomes increasingly apparent under new battery models or new operating conditions, making it difficult to meet the prediction accuracy requirements of actual engineering. Summary of the Invention

[0004] In view of the above problems, this application provides a method, device, vehicle and computer-readable storage medium for predicting the health status of a battery, which can solve the problems of poor model generalization ability, long deployment cycle and slow adaptation speed caused by the scarcity of training data in traditional deep learning methods under new battery models or new operating conditions.

[0005] According to one aspect of the embodiments of this application, a method for predicting the health status of a battery is provided, the method comprising: During the battery charge-discharge cycle, multiple battery detection data are acquired, including sequences of voltage, current, and capacity changes over time. Each battery detection data is used as a set of meta-training tasks, and each meta-training task includes a support set and a query set. Construct a neural network model based on meta-learning; wherein the neural network model includes a shared parameter base network and a task-specific prediction network; During the meta-training phase, inner loop adaptation and outer loop update are performed for each meta-training task; wherein, the inner loop adaptation uses support set data to update the task-specific prediction network through gradient descent algorithm, and the outer loop update updates the shared parameter base network based on query set loss through second-order gradient optimization algorithm. During the deployment phase, monitoring data of a preset number of target batteries are acquired as a support set. An adapted prediction model is obtained through an inner-loop adaptation process, and the health status of the target batteries is predicted.

[0006] In an optional embodiment, the method further includes: The voltage and current in the battery detection data are standardized using a sliding window, and the capacity in the battery detection data is normalized; wherein the size of the sliding window is a preset time step.

[0007] In an optional embodiment, the method further includes the step of using each of the battery detection data as a set of meta-training tasks: The meta-training task is constructed as follows: from the complete lifecycle data of the same battery, two consecutive segments that do not overlap in time are selected. One segment is used as the support set, and the other segment is used as the query set. The number of loops corresponding to the support set is earlier than the number of loops corresponding to the query set. The length of each segment is a preset number of consecutive loops.

[0008] In an optional embodiment, during the meta-training phase, the method further includes: A gradient alignment regularization term is added to the meta-training loss function. The gradient alignment regularization term is calculated by taking the average value of the cosine similarity of the gradient vectors between each pair of tasks. Maximizing the cosine similarity makes the adaptation direction of the shared parameter base network tend to be consistent across different tasks.

[0009] In an optional embodiment, the shared parameter base network adopts a deep temporal convolutional structure, and each temporal convolutional block structure is composed of a causal convolutional layer, a weight normalization layer, and a ReLU activation function layer connected in sequence.

[0010] In an optional embodiment, the task-specific prediction network adopts a linear regression layer structure, the input dimension of which is consistent with the output feature dimension of the shared parameter base network, and each of the meta-training tasks includes corresponding task-specific prediction network parameters.

[0011] In an optional embodiment, the method further includes: During the inner loop adaptation process for each of the meta-training tasks, features of the support set data are extracted from the shared parameter base network, and a task-specific prediction network is used to calculate the prediction result based on the feature representation of the support set data. The difference between the prediction result and the actual health status value is calculated based on the mean squared error loss function, and the parameters of the task-specific prediction network are updated by performing a preset gradient descent. After the inner loop adaptation is completed, the adapted task-specific prediction network is used to calculate the prediction loss on the query set, and the second gradient of the prediction loss with respect to the initial parameters of the shared parameter base network is calculated. The shared parameter base network is then updated based on the second gradient.

[0012] According to another aspect of the embodiments of this application, a battery health status prediction device is provided, the device comprising: The acquisition module is used to acquire multiple battery detection data during battery charge-discharge cycles, the battery detection data including a sequence of voltage, current and capacity changes over time; A construction module is used to construct a meta-learning-based neural network model by taking each of the battery detection data as a set of meta-training tasks; wherein each of the meta-training tasks includes a support set and a query set, and the neural network model includes a shared parameter base network and a task-specific prediction network; The training module is used to perform inner loop adaptation and outer loop update for each of the meta-training tasks during the meta-training phase; wherein the inner loop adaptation uses support set data to update the task-specific prediction network through gradient descent algorithm, and the outer loop update updates the shared parameter base network based on query set loss through second-order gradient optimization algorithm. The prediction module is used during the deployment phase to acquire monitoring data of a preset number of target batteries as a support set, obtain an adapted prediction model through an inner-loop adaptation process, and predict the health status of the target batteries.

[0013] According to another aspect of the embodiments of this application, a vehicle is provided, comprising: Controller; The memory is used to store one or more programs, which, when executed by the controller, enable the controller to implement the battery health state prediction method described above.

[0014] According to another aspect of the embodiments of this application, a computer-readable storage medium is provided, wherein a computer program is stored in the computer program, the computer program including at least one executable instruction, which, when executed on a battery health state prediction device / vehicle, causes the battery health state prediction device / vehicle to perform the operation of the battery health state prediction method as described above.

[0015] The battery health state prediction method in this application acquires multiple battery detection data sets to collect the time-varying sequences of voltage, current, and capacity, thus providing basic data support for model training. A meta-training task set is constructed based on the battery detection data and divided into a support set and a query set, enabling the model to simulate data conditions under new battery scenarios. Through the collaborative design of a shared parameter base network and a task-specific prediction network, the shared parameter base network focuses on extracting general degradation features across batteries, while the task-specific prediction network is rapidly adjusted for individual tasks. During the meta-training phase, an inner-loop adaptation process is executed, using the support set data to update the task-specific prediction network via gradient descent, allowing the model to quickly fit the current task features with limited data. Simultaneously, an outer-loop update mechanism is used to update the shared parameter base network based on the query set loss using a second-order gradient optimization algorithm, enabling the base network to quickly adapt to new tasks. During the deployment phase, a preset number of monitoring data sets of the target battery are acquired as the support set, and the adapted prediction model is obtained through the inner-loop adaptation process, ultimately achieving accurate prediction of the target battery's health state. The technical solution of this application avoids the reliance on a large amount of data on the same type of battery, significantly improves the generalization ability of the model under new battery models or new operating conditions, shortens the deployment cycle, and improves the practicality and reliability of battery health status prediction.

[0016] The above description is merely an overview of the technical solutions of the embodiments of this application. In order to better understand the technical means of the embodiments of this application and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of this application more obvious and understandable, specific implementation methods of this application are described below. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0018] Figure 1 A flowchart illustrating an embodiment of the battery health status prediction method of this application applied to remote terminal devices is shown.

[0019] Figure 2 A schematic diagram of an embodiment of the battery health status prediction device applied to a vehicle according to this application is shown.

[0020] Figure 3 A schematic diagram of the structure of an embodiment of the vehicle provided in this application is shown. Detailed Implementation

[0021] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0022] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0023] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily need to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0024] In this application, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0025] With the rapid development of electric vehicles and large-scale energy storage industries, accurate prediction of battery state of health (SOH) has become crucial for ensuring safe system operation and extending service life. The current diversity of battery materials, significant differences in manufacturing processes, and complex operating conditions place higher demands on SOH prediction methods. Existing deep learning methods mostly employ supervised learning paradigms, relying on a large amount of labeled data from the complete lifecycle of similar batteries for training. However, they exhibit poor adaptability to new battery models or new operating conditions: early data for new batches of batteries is scarce, preventing timely model deployment; traditional transfer learning struggles to cope with individual battery differences and complex operating conditions; a lack of rapid adaptation mechanisms necessitates the collection of large amounts of data and model retraining each time, resulting in long deployment cycles and high costs; and limited ability to extract early battery degradation features prevents the full utilization of limited early data for accurate prediction.

[0026] In view of this, to address the aforementioned problems, this application proposes a battery health state prediction method to solve the problems of poor model generalization ability, long deployment cycle, and slow adaptation speed caused by the scarcity of training data in traditional deep learning methods under new battery models or new operating conditions. The execution entity of the battery health state prediction method can be a terminal device, server, vehicle domain controller, cockpit domain controller, or other processing device. The terminal device can be user equipment (UE), computer, mobile device, user terminal, terminal, cellular phone, personal digital assistant (PDA), handheld device, computing device, in-vehicle device, wearable device, etc. In some possible implementations, the battery health state prediction method can be implemented by the processor calling computer-readable instructions stored in memory.

[0027] Specifically, please refer to, for example Figure 1 As shown, the battery health status prediction method of this embodiment includes the following steps: Step S100: During the battery charge-discharge cycle, acquire multiple battery detection data, including a sequence of voltage, current, and capacity changes over time.

[0028] In this embodiment, multiple battery detection data refer to the raw monitoring data collected from batteries from different source domains (e.g., lithium-ion batteries from different production batches, different chemical systems, or different usage scenarios) throughout their complete life cycle. Each set of data is a three-dimensional time-series signal synchronously recorded by the same battery during charge-discharge cycles. The sequences of voltage, current, and capacity changes over time represent discrete-time sequences collected at a fixed sampling frequency (e.g., a sampling frequency of 1 Hz). The voltage and current sequences are real-valued floating-point number sequences, and the capacity sequence is a monotonically non-decreasing sequence composed of cumulative ampere-hour (Ah) changes. These three types of sequences collectively characterize the electrochemical response characteristics of the battery at different aging stages, providing basic data support for the subsequent construction of meta-training tasks.

[0029] Furthermore, the voltage and current in the battery detection data are standardized using a sliding window, and the capacity in the battery detection data is normalized; wherein, the size of the sliding window is a preset time step.

[0030] In this embodiment, sliding window normalization refers to sliding a fixed-length time window point by point across the voltage and current sequences, calculating the mean and standard deviation of the data within the window, and then subtracting the mean of the window from the data point corresponding to the center position of the current window and dividing by the standard deviation of the window to obtain the locally normalized voltage and current values. The voltage and current sequences are components of the battery detection data defined in the above embodiment, and serve as the original input sequences for subsequent meta-training tasks and model training. The preset time step is the number of time points covered by the pre-set sliding window, used to control the granularity of local statistical characteristic extraction. Its value can be configured according to the dynamic response characteristics of the battery charging and discharging process, for example, 10–20 time steps, with a typical value of 15 time steps. The normalization formula involved is as follows: In the standardized formula Represents the original voltage (unit: V) or current reading (unit: A). It is the standardized value. It is the mean of the data within the window (reflecting the baseline level within that time window). It is the standard deviation (characterizing the degree of data fluctuation); this processing method can eliminate the dimensional effects caused by individual differences and working conditions of different batteries, while preserving the relative variation patterns in the data.

[0031] Capacity normalization refers to the process of mapping a capacity sequence to a uniform numerical range. This involves linearly scaling the capacity decay trajectory recorded throughout the entire lifespan of the same battery, so that the minimum value corresponds to 0 and the maximum value to 1. The normalization formula involved is: In the normalization formula This is the original capacity value (unit: ampere-hour). It is the normalized value. and These are the maximum and minimum capacity values ​​of the battery throughout its entire lifespan. The capacity sequence reflects the overall degradation trend of the battery. Maximum and minimum value normalization can map the capacity values ​​of different batteries to a unified numerical range, which facilitates model comparison and learning of degradation patterns. This differentiated preprocessing strategy fully considers the physical meaning and mechanism of action of different types of signals in battery health state prediction.

[0032] The technical solution in this embodiment is responsible for cleaning and standardizing the raw battery data, and collaboratively performing sliding window standardization and capacity normalization processing. On the one hand, the sliding window mechanism is used to preserve the local dynamic characteristics of voltage and current, and on the other hand, a unified health status reference system is established through capacity normalization. The sliding window mechanism and capacity normalization together constitute a robust preprocessing link for the meta-learning paradigm, which not only alleviates the adverse effects of the original data distribution offset on model training, but also provides a structured guarantee for the subsequent temporal consistency modeling between the support set and the query set, and finally achieves fast and accurate health status prediction with a small amount of target battery data.

[0033] Step S200: Use each of the battery detection data as a set of meta-training tasks to construct a neural network model based on meta-learning; wherein each of the meta-training tasks includes a support set and a query set, and the neural network model includes a shared parameter base network and a task-specific prediction network; Using each battery's detection data as a meta-training task set means constructing an independent meta-training task based on the complete lifecycle monitoring data of a single battery. This task structure satisfies that each task contains a support set for rapid parameter adjustment and a query set for evaluating generalization performance. The support set and query set are two non-overlapping time segments divided from the same battery data, which are strictly separated in the time dimension. The support set corresponds to the earlier aging stage (e.g., the first 20–50 cycles), and the query set corresponds to the later aging stage (e.g., the last 20–50 cycles), thereby simulating the task objective of predicting subsequent degradation trends using early data in real deployments.

[0034] The shared-parameter base network is used to extract time-series feature representations with cross-task generalization capabilities from voltage, current, and capacity sequences in the support and query sets. Its parameters are fully shared across all meta-training tasks. The task-specific prediction network is a lightweight mapping module whose parameters are independently initialized and updated in each meta-training task. It is responsible for mapping the feature vectors output by the shared network to health status predictions (SOH, usually expressed as a percentage). Together, they form a dual-path modeling paradigm that ensures the versatility of the underlying feature extraction capabilities while retaining the task adaptability of the high-level prediction logic.

[0035] Furthermore, the construction method of the meta-training task is as follows: from the complete life cycle data of the same battery, two consecutive segments that do not overlap in time are selected, one segment is used as the support set, and the other segment is used as the query set, and the number of loops corresponding to the support set is earlier than the number of loops corresponding to the query set; wherein, the length of each segment is a preset number of consecutive loops.

[0036] Among them, the complete life cycle data refers to all the charge-discharge cycle monitoring data recorded during the period from the initial use of the battery until it reaches the end-of-life threshold (for example, the capacity degrades to 80% of the rated capacity); the non-overlapping continuous segments in time mean that two segments have no overlap and no gap on the time axis, and the internal cycle numbers are continuous within each segment; the cycle number corresponding to the support set is earlier than the cycle number corresponding to the query set, which means that the set of cycle numbers covered by the support set and the set of cycle numbers covered by the query set satisfy max(support set cycle number) < min(query set cycle number), so as to ensure that the support set only contains data from the early or middle degradation stages of the battery, and the query set corresponds to the subsequent more significant degradation stage; this time sequence relationship simulates the logical constraint of using historical data to predict future states in the actual prediction scenario, preventing model learning bias caused by future information leakage.

[0037] The preset number of continuous cycles means that when constructing any meta-training task, the number of charge-discharge cycles contained in the support set and the query set respectively is a fixed integer value, which is uniformly set before meta-training. For example, it can be 20, 30, 40, or 50 continuous cycles; in an optional implementation, this preset number is dynamically configured according to the battery type: for lithium iron phosphate batteries, it is set to 20 - 30 cycles; for ternary lithium ion batteries, it is set to 30 - 40 cycles; further, this preset number can also be adaptively adjusted based on the current health state range of the target battery. For example, when the SOH is in the range of 95% - 100%, a smaller value is taken, and when the SOH is in the range of 85% - 95%, a larger value is taken to match the data change rate in different degradation stages.

[0038] This application ensures that each meta-training task conforms to using early degradation data to predict subsequent degradation states by defining the temporal position relationship and the consistency of segment lengths between the support set and the query set in the complete life cycle; on this basis, the support set provides a basis for the task-specific prediction network to quickly adapt to the individual battery, and the query set is used to evaluate the generalization performance of the shared parameter base network under this task; the non-overlap and sequence of the support set and the query set in the time dimension together constitute the authenticity constraint of the meta-training task, enabling the model to learn a robust and transferable degradation pattern recognition ability during the inner loop adaptation and outer loop update processes, thereby improving the response accuracy and stability to a small amount of early data of the target battery in the deployment stage.

[0039] Step S300: In the meta-training phase, inner-loop adaptation and outer-loop update are performed for each meta-training task. Inner-loop adaptation is a fast parameter tuning process performed within a single meta-training task, updating only the parameters of the task-specific prediction network without changing the initial weights of the shared parameter base network. The method of inner-loop adaptation is as follows: the support set data of a certain task is forward propagated to the shared parameter base network to obtain the hidden state features, and then the SOH prediction value is output by the task-specific prediction network. The mean square error (MSE) between the prediction value and the true SOH label is calculated, and 1-3 steps of SGD are performed to update the parameters of the task-specific prediction network. Outer loop update is a global parameter optimization process performed across all meta-training tasks. It calculates the second gradient (i.e., the gradient of gradients) of the initial parameters of the shared parameter base network based on the prediction loss of each task on the query set, and drives the update of the shared network parameters through this second gradient, enabling the network to learn how to learn. The outer loop update method is as follows: after completing the inner loop adaptation, the task-specific prediction network parameters are frozen, the query set data is input into the adapted model, the query loss is calculated, and the computation graph is preserved. Backpropagation is used to obtain the first gradient of the loss with respect to the initial weights of the shared parameter base network, and then the second gradient is obtained by differentiating the first gradient again. Finally, the Adam optimizer is used to update the shared network parameters.

[0040] This dual-loop mechanism ensures that the model can quickly fit the current task from a small amount of support set data, while continuously improving its adaptability to new tasks.

[0041] In this embodiment, the inner loop adaptation uses support set data to update the task-specific prediction network through gradient descent algorithm, and the outer loop update uses query set loss to update the shared parameter base network through second-order gradient optimization algorithm; the shared parameter base network adopts a deep temporal convolutional structure, and each temporal convolutional block structure is composed of a causal convolutional layer, a weight normalization layer and a ReLU activation function layer connected in sequence.

[0042] In this embodiment, the deep temporal convolutional structure refers to a deep neural network structure formed by stacking multiple temporal convolutional blocks, used to extract hierarchical features from time-varying sequence data of voltage, current, and capacity. This structure can progressively model the local dynamic changes and long-term dependencies in the battery degradation process, improving the ability to represent nonlinear degradation modes. The deep temporal convolutional structure can be composed of 4–8 cascaded temporal convolutional blocks, with the input of each temporal convolutional block connected to the output of the previous temporal convolutional block. The first temporal convolutional block receives the preprocessed battery detection data sequence as input, and the last temporal convolutional block outputs a high-dimensional temporal feature representation.

[0043] Causal convolutional layers utilize data from the current and historical moments for feature computation without incorporating future information, thus satisfying the temporal causality constraint of battery charging and discharging. Their kernel size is 3–5, and the number of kernels is 64–128, balancing local detail capture and global trend modeling capabilities. Weight normalization layers normalize the scale of the weight tensors output by the causal convolutional layers; they are weight normalization or its equivalent variants, transforming the weight vector... Decomposed into amplitude and direction Two components, namely This decomposition allows the model to independently optimize the magnitude and direction of the weights, accelerating the training convergence process. Its role is to stabilize gradient propagation, accelerate model convergence, and mitigate internal covariate bias; ReLU activation function. The network is used to introduce nonlinear mapping capabilities, with its output being the maximum value between the input and zero. This enhances the model's ability to fit nonlinear phases (such as accelerated decay) during battery health degradation. The residual connections after each convolutional block enable the network to learn identity mappings, mitigating the vanishing gradient problem in deep networks. The entire shared parameter base network extracts generalizable temporal features from the raw battery data, capturing common degradation patterns across different batteries.

[0044] The task-specific prediction network employs a linear regression layer structure, with its input dimension consistent with the output feature dimension of the shared parameter base network. Each meta-training task includes corresponding task-specific prediction network parameters. The expression for the linear regression layer is: In the linear regression layer It is a feature vector (256 dimensions) extracted by the shared parameter base network. It is a weight matrix (dimension [1,256]). It is a bias term. It is a health status prediction value (range) - (This represents the relative degree of health status). Each meta-training task has independent task-specific prediction network parameters. These parameters are rapidly adjusted during the inner-loop adaptation phase. Parameter initialization employs the Xavier initialization method to ensure that signal strength remains constant during forward propagation and gradient variance remains consistent during backward propagation. The core idea of ​​this design is to divide the model parameters into two parts: a shared-parameter base network learns general feature representations across batteries, while a task-specific prediction network is responsible for adapting to individual differences. This separation design allows the model to quickly adapt to new batteries by adjusting only a small number of parameters (the task-specific prediction network), significantly reducing the requirement for new data.

[0045] The linear regression layer structure refers to a single-layer fully connected mapping structure consisting of a weight matrix and a bias vector, without any non-linear activation functions or hidden layers. This structure is used to directly map the high-dimensional temporal feature vectors output by the shared parameter base network to scalar health state prediction values. Input dimension matching means that the number of columns in the weight matrix of the task-specific prediction network is strictly equal to the feature dimension of the last layer output by the shared parameter base network, ensuring that the feature vectors can be accessed without loss. This matching relationship is verified during the model initialization phase and remains fixed during training.

[0046] In this embodiment, a gradient alignment regularization term is added to the meta-training loss function. The gradient alignment regularization term is calculated by taking the average value of the cosine similarity of the gradient vectors between each pair of tasks. Maximizing the cosine similarity makes the adaptation direction of the shared parameter base network tend to be consistent across different tasks.

[0047] During the meta-training phase, after performing inner-loop adaptation for each meta-training task, the updated gradient vector of the task-specific prediction network on the support set is obtained. The gradient vector is the gradient of the task-specific prediction network parameters with respect to the initial parameters of the shared-parameter base network, and its dimension is consistent with the total dimension of the learnable parameters of the shared-parameter base network. The gradient vector is obtained by backpropagating the first-order gradient of the support set loss with respect to the initial parameters of the shared-parameter base network during the inner-loop adaptation process. The gradient vector represents the local optimization direction of the shared-parameter base network parameter update for the meta-training task.

[0048] In the current batch of multiple meta-training tasks participating in the outer ring update, any two tasks are selected, and the cosine similarity between their corresponding gradient vectors is calculated. The cosine similarity is defined as the ratio of the dot product of the two gradient vectors to the product of their magnitudes, and the value range can be [-1, 1]. The gradient vector between each pair of tasks refers to the gradient vector generated by meta-training tasks from different battery individuals and different life cycle stages but with structural isomorphism. The cosine similarity is used to quantify the degree of consistency of the update direction applied by different tasks to the shared parameter base network. The gradient alignment regularization term is obtained by calculating the arithmetic mean of the cosine similarities of all pairwise combinations of tasks in the current batch. The calculation of the average value covers all unordered task pairs in the current batch without repetition. The gradient alignment regularization term is a scalar, and the larger its value, the more similar the update directions of the shared parameter base network of each task pair are. The gradient alignment regularization term is introduced into the meta-training total loss function in a weighted form (e.g., the weighting coefficient can be 0.1–0.5) to form a joint optimization objective. The weighting coefficient is used to balance the optimization priority between the original query set prediction loss and the gradient direction consistency constraint. The joint optimization objective is minimized in each outer loop update, thereby achieving synergistic optimization of query set prediction accuracy and cross-task gradient direction consistency.

[0049] When performing outer loop updates based on the joint optimization objective, the second gradient of the total loss with respect to the initial parameters of the shared parameter base network is calculated through backpropagation, and the shared parameter base network is updated. The second gradient simultaneously carries the query set prediction error signal and the gradient direction alignment guidance signal. The update process makes the parameter adjustment path of the shared parameter base network more common and stable when facing different battery degradation modes. This application introduces a gradient alignment regularization term into the meta-training loss function. By averaging and maximizing the cosine similarity of gradient vectors between pairs of tasks, it encourages the shared parameter base network to generate gradient update paths with more consistent directions during the inner loop adaptation process of different meta-training tasks. On this basis, the second-order gradient on which the outer loop update depends not only reflects the curvature information of the prediction error, but also implies the consensus of feature extraction directions between tasks. This improves the generalization ability of the shared parameter base network to diverse battery degradation modes and enhances the model's rapid adaptation stability and convergence robustness in new battery deployment scenarios.

[0050] Furthermore, during the inner-loop adaptation process for each of the meta-training tasks, features of the support set data are extracted from the shared parameter base network, and a task-specific prediction network is used to calculate the prediction result based on the feature representation of the support set data. The difference between the prediction result and the actual health state value is calculated based on the mean squared error loss function, and the parameters of the task-specific prediction network are updated by performing a preset gradient descent. The preset gradient can be a 1-3 step gradient, that is, the parameters of the task-specific prediction network are updated by 1-3 step gradient descent, and the inner-loop learning rate can be set to 0.01-0.1.

[0051] After the inner loop adaptation is completed, the adapted task-specific prediction network is used to calculate the prediction loss on the query set, and the second gradient of the prediction loss with respect to the initial parameters of the shared parameter base network is calculated. The shared parameter base network is then updated based on the second gradient.

[0052] For the above embodiments, the mean squared error loss function The mean squared error loss function is used to measure prediction accuracy. It is the true health status value (calculated based on capacity measurement). It is a predicted value. This supports a set number of samples. Parameter updates use the gradient descent algorithm, with the update rule being... In updating rules These are the parameters of a task-specific prediction network. It is the inner loop learning rate (which can be set to 0.05). It is the gradient of the loss function with respect to the parameters.

[0053] The choice of the inner loop learning rate is based on multiple experimental verifications, achieving a good balance between adaptation speed and stability; the inner loop adaptation step is typically performed... The choice of this number of iterations ensures sufficient adaptability while avoiding the risk of overfitting. The entire process simulates the scenario of fine-tuning the model using a small amount of data from the target battery during actual deployment, enabling the model to learn how to quickly adjust its parameters to adapt to new tasks.

[0054] The outer-loop update process is responsible for optimizing the initial state of the shared parameter base network. After the inner-loop adaptation is complete, this process uses the adapted task-specific prediction network to calculate the prediction loss on the query set. The outer-loop update employs a second-order gradient optimization method, and the update rule for the shared parameter base network is as follows: ,in These are the parameters of the shared parameter base network. It is the outer loop learning rate (the outer loop learning rate can be 0.005). Indicates the first Individual training tasks This represents the model after inner loop adaptation.

[0055] The calculation of the second gradient involves the product of the Hessian matrix and the vector, which is achieved through automatic differentiation. Preserving the complete computation graph is the key to calculating the second gradient, which allows the model to consider the impact of the inner loop adaptation process on the final performance. The core of the outer loop update is to optimize the initial parameters of the shared parameter base network, so that the model obtained after inner loop adaptation from this initial state can show good performance on various battery prediction tasks.

[0056] Furthermore, the aforementioned gradient alignment mechanism improves the effectiveness of meta-learning by constraining the optimization directions of different tasks. This gradient alignment mechanism calculates the cosine similarity between the gradient vectors generated by different meta-training tasks during the inner-loop adaptation process. The mathematical expression for cosine similarity is: ,in and Representing tasks and tasks The gradient vector; the gradient alignment loss is defined as... ,in This refers to the number of tasks in a small batch (e.g., the number of tasks is set to...). The total loss function is: ,in This is the regularization coefficient (e.g., a regularization coefficient set to 0.3). If different tasks optimize in the same direction in the parameter space, it indicates that these tasks share certain common features that are universally applicable to battery health prediction. By maximizing gradient similarity, the model can identify and reinforce these common features while weakening task-specific noise features. The regularization coefficient... The choice of needs to strike a balance between meta-learning loss and gradient alignment loss, and 0.3 was determined to be the optimal value through grid search.

[0057] In this embodiment, extracting features from the support set data from the shared parameter base network refers to inputting the voltage, current, and capacity time series data from the support set into the shared parameter base network, processing them layer by layer through its deep temporal convolutional structure, and outputting a fixed-dimensional temporal feature vector. This feature vector represents the high-level semantic information of the support set data in the degradation mode. The features of the support set data are the output feature vector, and its dimension is strictly consistent with the input dimension of the task-specific prediction network. Calculating the prediction result based on the feature representation of the support set data using the task-specific prediction network refers to directly feeding the feature vector into a linear regression layer, and obtaining a single-dimensional scalar output through weighted summation and superposition with the bias term, which serves as the predicted health status value of the support set for the corresponding number of iterations under the current meta-training task.

[0058] Step S400: During the deployment phase, monitoring data of a preset number of target batteries are acquired as a support set. An adapted prediction model is obtained through an inner-loop adaptation process, and the health status of the target batteries is predicted.

[0059] In this embodiment, the target battery's preset quantity of monitoring data refers to the collection of voltage, current, and capacity sequences for the initial few charge-discharge cycles (e.g., 5–15 cycles) of a new battery to be predicted (e.g., a new model of power battery or a spare battery replaced on-site) in a real-world application scenario. The adapted prediction model obtained through the inner loop adaptation process refers to the reuse of the shared parameter base network weights obtained from the meta-training. Based on this small amount of target battery data, only the parameter updates of the task-specific prediction network are performed without modifying the shared network structure or parameters.

[0060] In the above embodiments, during the deployment phase, a learning rate decay strategy is also employed to optimize the adaptation process during deployment; the learning rate update rule is as follows: ,in This is the initial learning rate (set to 0.04). It is the attenuation coefficient (for example, the attenuation coefficient can be 0.6). This represents the current update count. The first four gradient updates use the initial learning rate, and subsequent updates gradually decrease the learning rate by 0.6. Based on the learning rate scheduling principle in optimization theory, a larger learning rate is used in the early stages of adaptation to quickly approach the optimum, while a smaller learning rate is used in the later stages for fine-tuning.

[0061] This application organizes historical detection data from multiple source domain batteries into a meta-training task containing support and query sets. It constructs a dual-path meta-learning model with a shared parameter base network and a task-specific prediction network. During the meta-training phase, a dual-loop parameter learning mechanism is collaboratively executed: the inner loop adapts and updates the task-specific network, while the outer loop updates and optimizes the shared base network. This allows the shared parameter base network to learn general degradation characteristics across batteries, while the task-specific prediction network can quickly complete personalized adaptation with a limited amount of target battery data. Furthermore, during deployment, only a preset number of early monitoring data points from the target batteries are needed to initiate the inner-loop adaptation process, rapidly generating the adapted model and performing health status prediction. This effectively solves the problems of poor generalization ability, long deployment cycles, and slow adaptation speed caused by the scarcity of training data in traditional methods, significantly improving prediction accuracy and engineering practicality.

[0062] Based on the above embodiments, the following example illustrates a practical application scenario for predicting battery health status. In this scenario, the goal is to accurately predict the health status of a new batch of lithium batteries using a small amount of early cycle data: First, complete lifecycle data, including voltage, current, and capacity sequences, are obtained from multiple source domain batteries (e.g., batches A, B, and C) from the battery testing platform. The voltage and current sequences are then standardized using a sliding window method with a window size of 15 time steps. For each window, the mean is calculated. and standard deviation Then apply the standardized formula. For the capacity sequence, determine the capacity of each cell. and And apply the normalization formula The preprocessed data is organized into meta-training tasks, with the support set for each task taken from consecutive segments within the first 100 cycles of the battery lifecycle, and the query set taken from consecutive segments in subsequent cycles.

[0063] Then, the meta-learning model is trained on the server; during the inner-loop adaptation phase, for each meta-training task... It will support data collection Input shared parameters base network Extract features and use task-specific prediction networks Calculate the predicted value The prediction error is calculated using the mean squared error loss function, and the task-specific parameters are updated using gradient descent. ,in During the outer ring update phase, the adapted model is used on the query set. Calculate the meta-learning loss Simultaneously calculate gradient alignment loss The total loss function is ,in The shared parameter base network parameters are updated through second-order gradient optimization. ,in .

[0064] Next, when a new batch of lithium batteries (e.g., batch D) needs to be predicted, only the data from the first 50 cycles of that batch are used as the support set to perform inner-loop adaptation on the pre-trained model; the adaptation process employs a learning rate decay strategy. The first update uses a learning rate of 0.04, and subsequent updates gradually decrease the learning rate by a increment of 0.6. The learning rate is updated every time. , No. The learning rate is updated every time. And so on; in each update, the loss function on the support set is calculated and the task-specific prediction network parameters are updated until convergence or the maximum number of updates is reached.

[0065] Finally, the adapted model receives real-time detection data from the new batch of batteries and outputs accurate health status predictions. During the prediction process, the model continuously detects prediction uncertainty. When the uncertainty exceeds a threshold, it automatically triggers a re-adaptation process. The prediction results are transmitted to the battery management system via an interface to optimize charging strategies and lifespan predictions.

[0066] Through the above embodiments, by introducing the meta-learning framework, the model can not only learn specific battery degradation patterns, but also learn how to quickly adapt to new degradation patterns. This gives it a significant advantage in the face of constantly updated battery technologies and diverse usage scenarios. The entire system maintains prediction accuracy while significantly shortening the adaptation time for new battery models, and only requires a small amount of early data from the target battery to achieve good prediction performance.

[0067] Figure 2 A schematic diagram of an embodiment of the battery health status prediction device applied to a vehicle according to this application is shown. Please refer to... Figure 2 As shown, the battery health status prediction device 500 includes an acquisition module 510, a construction module 520, a training module 530, and a prediction module 540. The acquisition module 510 is used to acquire multiple battery detection data during a battery charge-discharge cycle, wherein the battery detection data includes a sequence of voltage, current and capacity changes over time. The construction module 520 is used to construct a neural network model based on meta-learning by using each of the battery detection data as a set of meta-training tasks; wherein each of the meta-training tasks includes a support set and a query set, and the neural network model includes a shared parameter base network and a task-specific prediction network. Training module 530 is used to perform inner loop adaptation and outer loop update for each of the meta-training tasks during the meta-training phase; wherein, the inner loop adaptation uses support set data to update the task-specific prediction network through gradient descent algorithm, and the outer loop update updates the shared parameter base network based on query set loss through second-order gradient optimization algorithm. The prediction module 540 is used during the deployment phase to acquire monitoring data of a preset number of target batteries as a support set, obtain an adapted prediction model through an inner-loop adaptation process, and predict the health status of the target batteries.

[0068] It should be noted that the battery health status prediction device 500 provided in the above embodiments and the battery health status prediction method provided in the foregoing embodiments belong to the same concept. The specific way in which each module and unit performs operations has been described in detail in the method embodiments, and will not be repeated here.

[0069] Figure 3 The diagram illustrates the structure of an embodiment of the vehicle described in this application, and also shows the structure of a computer system suitable for implementing the vehicle in this application. The specific embodiments of this application do not limit the specific implementation of the vehicle.

[0070] Please see Figure 3 As shown, the vehicle includes: a controller; and a memory for storing one or more programs, which, when executed by the controller, perform the battery health state prediction method described above.

[0071] Please continue reading. Figure 3As shown, the vehicle's computer system 600 includes a Central Processing Unit (CPU) 601, which can perform various appropriate actions and processes, such as executing the methods described in the above embodiments, based on programs stored in Read-Only Memory (ROM) 602 or programs loaded from storage portion 608 into Random Access Memory (RAM) 603. The RAM 603 also stores various programs and data required for system operation. The CPU 601, ROM 602, and RAM 603 are interconnected via a bus 604. An Input / Output (I / O) interface 605 is also connected to the bus 604.

[0072] The following components are connected to I / O interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to I / O interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 610 as needed so that computer programs read from it can be installed into storage section 608 as needed.

[0073] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 609, and / or installed from removable medium 611. When the computer program is executed by central processing unit (CPU) 601, it performs various functions defined in the system of this application.

[0074] Another aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the battery health state prediction method as described above. This computer-readable storage medium may be included in the vehicle described in the above embodiments, or it may exist independently and not installed in the vehicle.

[0075] Another aspect of this application provides a computer program product or computer program including at least one executable instruction that, when executed on a battery health state prediction device / vehicle, causes the battery health state prediction device / vehicle to perform the battery health state prediction method as described below: During the battery charge-discharge cycle, multiple battery detection data are acquired, including sequences of voltage, current, and capacity changes over time. Each of the battery detection data is used as a set of meta-training tasks to construct a neural network model based on meta-learning; wherein each of the meta-training tasks includes a support set and a query set, and the neural network model includes a shared parameter base network and a task-specific prediction network; During the meta-training phase, inner loop adaptation and outer loop update are performed for each meta-training task; wherein, the inner loop adaptation uses support set data to update the task-specific prediction network through gradient descent algorithm, and the outer loop update updates the shared parameter base network based on query set loss through second-order gradient optimization algorithm. During the deployment phase, monitoring data of a preset number of target batteries are acquired as a support set. An adapted prediction model is obtained through an inner-loop adaptation process, and the health status of the target batteries is predicted.

[0076] In an alternative approach, the executable instructions can also be used to cause the battery health prediction device / vehicle to perform the following operations: The voltage and current in the battery detection data are standardized using a sliding window, and the capacity in the battery detection data is normalized; wherein the size of the sliding window is a preset time step.

[0077] In an alternative approach, the executable instructions can also be used to cause the battery health prediction device / vehicle to perform the following operations: The meta-training task is constructed as follows: from the complete lifecycle data of the same battery, two consecutive segments that do not overlap in time are selected. One segment is used as the support set, and the other segment is used as the query set. The number of loops corresponding to the support set is earlier than the number of loops corresponding to the query set. The length of each segment is a preset number of consecutive loops.

[0078] In an alternative approach, the executable instructions can also be used to cause the battery health prediction device / vehicle to perform the following operations: A gradient alignment regularization term is added to the meta-training loss function. The gradient alignment regularization term is calculated by taking the average value of the cosine similarity of the gradient vectors between each pair of tasks. Maximizing the cosine similarity makes the adaptation direction of the shared parameter base network tend to be consistent across different tasks.

[0079] In an alternative approach, the executable instructions can also be used to cause the battery health prediction device / vehicle to perform the following operations: The shared parameter base network adopts a deep temporal convolutional structure, and each temporal convolutional block structure is composed of a causal convolutional layer, a weight normalization layer, and a ReLU activation function layer connected in sequence.

[0080] In an alternative approach, the executable instructions can also be used to cause the battery health prediction device / vehicle to perform the following operations: The task-specific prediction network adopts a linear regression layer structure, and its input dimension is consistent with the output feature dimension of the shared parameter base network. Each meta-training task includes corresponding task-specific prediction network parameters.

[0081] In an alternative approach, the executable instructions can also be used to cause the battery health prediction device / vehicle to perform the following operations: During the inner loop adaptation process for each of the meta-training tasks, features of the support set data are extracted from the shared parameter base network, and a task-specific prediction network is used to calculate the prediction result based on the feature representation of the support set data. The difference between the prediction result and the actual health status value is calculated based on the mean squared error loss function, and the parameters of the task-specific prediction network are updated by performing a preset gradient descent. After the inner loop adaptation is completed, the adapted task-specific prediction network is used to calculate the prediction loss on the query set, and the second gradient of the prediction loss with respect to the initial parameters of the shared parameter base network is calculated. The shared parameter base network is then updated based on the second gradient.

[0082] The battery health state prediction method in this application acquires multiple battery detection data sets to collect the time-varying sequences of voltage, current, and capacity, thus providing basic data support for model training. A meta-training task set is constructed based on the battery detection data and divided into a support set and a query set, enabling the model to simulate data conditions under new battery scenarios. Through the collaborative design of a shared parameter base network and a task-specific prediction network, the shared parameter base network focuses on extracting general degradation features across batteries, while the task-specific prediction network is rapidly adjusted for individual tasks. During the meta-training phase, an inner-loop adaptation process is executed, using the support set data to update the task-specific prediction network via gradient descent, allowing the model to quickly fit the current task features with limited data. Simultaneously, an outer-loop update mechanism is used to update the shared parameter base network based on the query set loss using a second-order gradient optimization algorithm, enabling the base network to quickly adapt to new tasks. During the deployment phase, a preset number of monitoring data sets of the target battery are acquired as the support set, and the adapted prediction model is obtained through the inner-loop adaptation process, ultimately achieving accurate prediction of the target battery's health state. The technical solution of this application avoids the reliance on a large amount of data on the same type of battery, significantly improves the generalization ability of the model under new battery models or new operating conditions, shortens the deployment cycle, and improves the practicality and reliability of battery health status prediction.

[0083] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable computer program. The transmitted data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0084] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0085] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0086] According to one aspect of the embodiments of this application, a computer system is also provided, including a Central Processing Unit (CPU), which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) or a program loaded from storage into random access memory (RAM), such as performing the methods described above. Various programs and data required for system operation are also stored in the RAM. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0087] The following components are connected to the I / O interface: input components including keyboards, mice, etc.; output components including cathode ray tubes (CRTs), liquid crystal displays (LCDs), and speakers; storage components including hard drives; and communication components including network interface cards such as LAN (Local Area Network) cards and modems. The communication components perform communication processing via networks such as the Internet. Drives are also connected to the I / O interface as needed. Removable media, such as disks, optical discs, magneto-optical discs, semiconductor memories, etc., are installed on the drive as needed so that computer programs read from them can be installed into the storage components as required.

[0088] The above description is merely a preferred exemplary embodiment of this application and is not intended to limit the implementation of this application. Those skilled in the art can easily make corresponding modifications or alterations based on the main concept and spirit of this application. Therefore, the scope of protection of this application should be determined by the scope of protection claimed in the claims.

[0089] In practice, the collection and processing of data in this application should strictly comply with the requirements of relevant national laws and regulations, obtain the informed consent or separate consent of the data subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the data subject.

Claims

1. A method for predicting the health status of a battery, characterized in that, The method includes: During the battery charge-discharge cycle, multiple battery detection data are acquired, including sequences of voltage, current, and capacity changes over time. Each of the battery detection data is used as a set of meta-training tasks to construct a neural network model based on meta-learning; wherein each of the meta-training tasks includes a support set and a query set, and the neural network model includes a shared parameter base network and a task-specific prediction network; During the meta-training phase, inner loop adaptation and outer loop update are performed for each meta-training task; wherein, the inner loop adaptation uses support set data to update the task-specific prediction network through gradient descent algorithm, and the outer loop update updates the shared parameter base network based on query set loss through second-order gradient optimization algorithm. During the deployment phase, monitoring data of a preset number of target batteries are acquired as a support set. An adapted prediction model is obtained through an inner-loop adaptation process, and the health status of the target batteries is predicted.

2. The battery health status prediction method according to claim 1, characterized in that, The method further includes: The voltage and current in the battery detection data are standardized using a sliding window, and the capacity in the battery detection data is normalized; wherein the size of the sliding window is a preset time step.

3. The battery health status prediction method according to claim 1, characterized in that, In the step of using each of the battery detection data as a meta-training task set, the method further includes: The meta-training task is constructed as follows: from the complete lifecycle data of the same battery, two consecutive segments that do not overlap in time are selected. One segment is used as the support set, and the other segment is used as the query set. The number of loops corresponding to the support set is earlier than the number of loops corresponding to the query set. The length of each segment is a preset number of consecutive loops.

4. The battery health status prediction method according to claim 1, characterized in that, During the meta-training phase, the method further includes: A gradient alignment regularization term is added to the meta-training loss function. The gradient alignment regularization term is calculated by taking the average value of the cosine similarity of the gradient vectors between each pair of tasks. Maximizing the cosine similarity makes the adaptation direction of the shared parameter base network tend to be consistent across different tasks.

5. The battery health status prediction method according to claim 1, characterized in that, The shared parameter base network adopts a deep temporal convolutional structure, and each temporal convolutional block structure is composed of a causal convolutional layer, a weight normalization layer, and a ReLU activation function layer connected in sequence.

6. The battery health status prediction method according to claim 1, characterized in that, The task-specific prediction network adopts a linear regression layer structure, and its input dimension is consistent with the output feature dimension of the shared parameter base network. Each meta-training task includes corresponding task-specific prediction network parameters.

7. The battery health status prediction method according to claim 1, characterized in that, The method further includes: During the inner loop adaptation process for each of the meta-training tasks, features of the support set data are extracted from the shared parameter base network, and a task-specific prediction network is used to calculate the prediction result based on the feature representation of the support set data. The difference between the prediction result and the actual health status value is calculated based on the mean squared error loss function, and the parameters of the task-specific prediction network are updated by performing a preset gradient descent. After the inner loop adaptation is completed, the adapted task-specific prediction network is used to calculate the prediction loss on the query set, and the second gradient of the prediction loss with respect to the initial parameters of the shared parameter base network is calculated. The shared parameter base network is then updated based on the second gradient.

8. A battery health status prediction device, characterized in that, The device includes: The acquisition module is used to acquire multiple battery detection data during battery charge-discharge cycles, the battery detection data including a sequence of voltage, current and capacity changes over time; A construction module is used to construct a meta-learning-based neural network model by taking each of the battery detection data as a set of meta-training tasks; wherein each of the meta-training tasks includes a support set and a query set, and the neural network model includes a shared parameter base network and a task-specific prediction network; The training module is used to perform inner loop adaptation and outer loop update for each of the meta-training tasks during the meta-training phase; wherein the inner loop adaptation uses support set data to update the task-specific prediction network through gradient descent algorithm, and the outer loop update updates the shared parameter base network based on query set loss through second-order gradient optimization algorithm. The prediction module is used during the deployment phase to acquire monitoring data of a preset number of target batteries as a support set, obtain an adapted prediction model through an inner-loop adaptation process, and predict the health status of the target batteries.

9. A vehicle, characterized in that, include: Controller; A memory for storing one or more programs that, when executed by a controller, cause the controller to implement the battery health state prediction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including at least one executable instruction, which, when executed on the battery health state prediction device / vehicle, causes the battery health state prediction device / vehicle to perform the operation of the battery health state prediction method as described in any one of claims 1 to 7.

Images