A multi-task driven lithium battery state of health joint prediction method

Abstract

The application provides a multi-task driven lithium battery health state joint prediction method, and belongs to the technical field of battery management. The method first acquires multi-modal operation time sequence data of a battery to be predicted, and then inputs the multi-modal operation time sequence data into a pre-trained joint prediction model to output a final coordinated joint prediction result. The application fundamentally avoids feature competition and gradient conflict among multiple tasks, and forcibly guarantees physical logic consistency among prediction results, thereby significantly improving the accuracy, reliability and generalization ability of lithium battery health state and remaining life joint prediction.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and more specifically to a multi-task-driven method for joint prediction of the health status of lithium batteries. Background Technology

[0002] Driven by the global energy structure transformation and the "dual-carbon" strategic goal, lithium-ion batteries, as efficient and flexible energy storage units, have been widely used in large-scale energy storage power stations, distributed energy systems, backup power supplies for communication base stations, and various industrial energy storage scenarios. The Battery Management System (BMS) is the core control system that ensures the safe, stable, and efficient operation of battery packs. Accurately assessing the battery's State of Health (SOH) and predicting its Remaining Life (RUL) is crucial for preventing battery failure, optimizing operation and maintenance strategies, and reducing total lifecycle costs.

[0003] With the rapid development of artificial intelligence technology, data-driven SOH and RUL prediction methods have become a research hotspot. Early prediction techniques typically treated SOH estimation and RUL prediction as two completely independent tasks, building and training models separately. This independent modeling paradigm has a significant drawback: there is a lack of information interaction and coordination between the two prediction tasks, and the prediction results may be isolated from each other, or even contradictory in physical logic.

[0004] To address these issues, researchers introduced Multi-Task Learning (MTL), a machine learning paradigm aimed at simultaneously handling the two related tasks of SOH estimation and RUL prediction using a unified model framework. Existing joint prediction methods employing MTL frameworks generally adopt a design pattern of "shared low-level feature encoder + independent task prediction heads." This architecture is based on an idealized assumption: the existence of a common set of low-level features that optimally serves both SOH estimation and RUL prediction simultaneously. However, SOH and RUL differ fundamentally in their physical meaning and prediction objectives. Forcing a shared feature encoder leads to feature competition and gradient conflicts during the optimization process. A gradient update beneficial to one task may harm the performance of the other. This inherent conflict makes it difficult for the model to learn feature representations optimal for both tasks, often resulting in a compromise, suboptimal solution, failing to achieve truly meaningful joint prediction.

[0005] Therefore, there is an urgent need for a novel multi-task joint prediction method that can fundamentally avoid task feature conflicts and enforce physical and logical consistency between SOH and RUL prediction results, in order to overcome the inherent defects of the existing "shared-branch" architecture and improve the intelligence level and decision reliability of the lithium battery health management system. Summary of the Invention

[0006] In view of the above-mentioned problems in the prior art, the present invention provides a solution.

[0007] To achieve the above objectives, embodiments of the present invention provide a multi-task-driven joint prediction method for the state of health of lithium batteries, the method comprising:

[0008] Obtain multimodal runtime sequence data of the battery to be predicted;

[0009] The multimodal runtime sequence data is input into a pre-trained joint prediction model; the joint prediction model includes a first feature extractor and a second feature extractor with mutually exclusive parameters, as well as a relation constraint.

[0010] The first feature extractor is used to process the multimodal runtime sequence data to obtain a first health state prediction value;

[0011] The second feature extractor is used to process the multimodal runtime sequence data to obtain a parameterized capacity decay trajectory function;

[0012] The first remaining lifetime prediction value is determined based on the parameterized capacity decay trajectory function and the preset capacity failure threshold.

[0013] The relationship constraint is used to verify and correct the consistency between the first health status prediction value and the first remaining life prediction value, so as to obtain the final joint prediction result.

[0014] Preferably, the method further includes obtaining a pre-trained joint prediction model, wherein obtaining the pre-trained joint prediction model includes:

[0015] Obtain a training sample set, wherein each training sample in the training sample set includes multimodal runtime sequence data of the battery, a real health status label, and a real remaining life label;

[0016] Construct an initial model that includes the first feature extractor, the second feature extractor, and the relation constraint;

[0017] For any sample in the training sample set, perform the following operations:

[0018] The multimodal runtime sequence data of the sample is input into the initial model to obtain the corresponding intermediate health status prediction value, intermediate remaining life prediction value and at least one intermediate consistency constraint signal.

[0019] Calculate the first prediction loss based on the intermediate health status prediction value and the actual health status label;

[0020] The second prediction loss is calculated based on the intermediate remaining lifetime prediction value and the actual remaining lifetime label;

[0021] Calculate the relation constraint loss based on at least one intermediate consistency constraint signal;

[0022] Based on the first prediction loss, the second prediction loss, and the relational constraint loss, a joint loss function is constructed;

[0023] With the goal of minimizing the joint loss function, the parameters of the initial model are iteratively updated to obtain the joint prediction model.

[0024] Preferably, the computational relation constraint loss includes the computational starting point alignment loss;

[0025] The calculation of the starting point alignment loss includes:

[0026] Based on the parameterized capacity decay trajectory function output by the second feature extractor, the function value of the parameterized capacity decay trajectory function at the starting time is obtained as the trajectory starting point prediction value.

[0027] Calculate the first difference between the intermediate health state prediction value and the trajectory starting point prediction value;

[0028] The starting point alignment loss is determined based on the first difference.

[0029] Preferably, the calculation of relational constraint loss includes the calculation of cycle consistency loss;

[0030] The calculation of cycle consistency loss includes:

[0031] The first implicit remaining lifetime value is calculated based on the intermediate health status prediction value and the preset capacity failure threshold.

[0032] The second implicit remaining lifetime value is obtained by solving the parameterized capacity decay trajectory function and the preset capacity failure threshold.

[0033] Calculate the second difference between the first implicit remaining lifetime value and the second implicit remaining lifetime value;

[0034] The cycle consistency loss is determined based on the second difference.

[0035] Preferably, obtaining the parameterized capacity decay trajectory function using the second feature extractor includes:

[0036] The multimodal runtime sequence data is processed using the feature encoding network in the second feature extractor to obtain a trend feature vector;

[0037] The parameter generation network in the second feature extractor generates dynamic parameters based on the trend feature vector.

[0038] The feedforward neural network is instantiated using the dynamic parameters, wherein the input of the feedforward neural network is a time variable and the output is a predicted battery capacity value, and the feedforward neural network constitutes the parameterized capacity decay trajectory function.

[0039] Preferably, the method further includes applying physical constraints to the feedforward neural network;

[0040] The physical constraints imposed on the feedforward neural network include:

[0041] Configure a monotonically non-increasing activation function for the output layer of the feedforward neural network; and / or,

[0042] Configure a range constraint function for the output layer of the feedforward neural network, the range constraint function being used to limit the output value of the feedforward neural network within the theoretical capacity range of the battery.

[0043] Preferably, the consistency verification and correction using relational constraint includes:

[0044] The compatibility score between the first health status prediction value and the first remaining lifespan prediction value is calculated using the relation constraint.

[0045] Determine whether the compatibility score is lower than a preset confidence threshold;

[0046] If the compatibility score is lower than the confidence threshold, then proceed to the calibration cycle;

[0047] The calibration cycle includes:

[0048] Keep the parameters of the joint prediction model unchanged;

[0049] With the goal of improving the compatibility score, the internal feature representations generated by the first feature extractor and the second feature extractor for the current input data are adjusted;

[0050] Based on the adjusted internal feature representation, the forward calculation is re-performed to obtain updated health status predictions and remaining life predictions, and the compatibility score is recalculated.

[0051] The calibration cycle is repeated until the compatibility score reaches or exceeds the confidence threshold, and the last predicted value is used as the final joint prediction result.

[0052] Preferably, the acquisition of the pre-trained joint prediction model further includes domain adaptation training; wherein, the training sample set contains data samples from multiple different domains, and each sample has a domain label representing its domain.

[0053] The domain adaptation training includes:

[0054] A gradient inversion layer and a domain classifier are connected to the back end of the first feature extractor and / or the second feature extractor;

[0055] Using the domain classifier, the domain to which a sample belongs is predicted based on the features output by the gradient inversion layer;

[0056] Based on the prediction results of the domain classifier and the domain labels of the samples, calculate the domain adversarial loss;

[0057] The domain adversarial loss is added to the joint loss function, and the gradient reversal layer is used to reverse the gradient direction during backpropagation.

[0058] The beneficial effects of this invention are as follows: By employing a first feature extractor and a second feature extractor with mutually exclusive parameters, the competition and conflict between the two tasks of health status estimation and remaining lifetime prediction during the feature learning process are fundamentally avoided, enabling the model to learn the optimal feature representation for each task. Simultaneously, by introducing a relation constraint and applying relation constraint loss during training, it is enforced that the health status values ​​and remaining lifetime values ​​output by the two independent prediction paths conform to the physical laws of battery degradation, achieving logically self-consistent joint prediction. This overcomes the inherent defects of existing shared encoder architectures and improves prediction accuracy and reliability. Attached Figure Description

[0059] Figure 1 A flowchart of a multi-task-driven joint prediction method for the health status of lithium batteries provided in an embodiment of the present invention. Detailed Implementation

[0060] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0061] As mentioned earlier, the core problem with existing joint prediction schemes based on multi-task learning stems from their reliance on a shared underlying feature encoder. This architecture cannot reconcile the inherent contradiction between battery health status and remaining lifespan's feature requirements, leading to optimization conflicts and preventing true joint prediction.

[0062] Please refer to Figure 1 Based on the above reasons, embodiments of the present invention provide a multi-task-driven joint prediction method for the health status of lithium batteries. The method includes: acquiring multimodal runtime sequence data of the battery to be predicted; inputting the multimodal runtime sequence data into a pre-trained joint prediction model; the joint prediction model includes a first feature extractor and a second feature extractor with mutually exclusive parameters, and a relation constraint; processing the multimodal runtime sequence data using the first feature extractor to obtain a first health status prediction value; processing the multimodal runtime sequence data using the second feature extractor to obtain a parameterized capacity decay trajectory function; determining a first remaining lifetime prediction value based on the parameterized capacity decay trajectory function and a preset capacity failure threshold; and using the relation constraint to perform consistency verification and correction between the first health status prediction value and the first remaining lifetime prediction value to obtain the final joint prediction result.

[0063] The joint prediction model provided in this embodiment completely abandons the traditional shared underlying design and instead adopts two parallel learning paths with completely independent parameters.

[0064] First, multimodal runtime timing data of the battery to be predicted within a time window is acquired from the data acquisition system. This data typically includes, but is not limited to, voltage sequences, current sequences, temperature sequences, and incremental capacity curve features and differential voltage curve features derived from them, which together constitute a high-dimensional time series input.

[0065] Subsequently, this data is fed in parallel into two structurally independent feature extractors with non-shared parameters: the first feature extractor is used to quickly and stably extract features strongly correlated with the current battery health state. For example, multi-scale convolution and channel attention mechanisms can be used to capture local patterns and short-term dependencies, ultimately regressing a scalar value, i.e., the "first health state prediction value," through a fully connected layer, typically ranging from 0 to 1. The second feature extractor is used to capture the long-term trend and dynamic patterns of capacity decay, generating a mathematical function that describes future capacity changes. For example, dilated convolution and temporal attention mechanisms can be used to process the data to determine a parameterized capacity decay trajectory function, e.g. ,in It is a relative time in the future. It is a dynamic parameter set generated by the network corresponding to the second feature extractor. This is the predicted capacity value. This function is solved by setting a capacity failure threshold (e.g., 80% of rated capacity). =Capacity failure threshold, and the corresponding "first remaining lifetime prediction value" can be obtained. Finally, the two are fed into the relation constraint unit, which, as an independent logical unit, evaluates the degree of consistency between the two initial prediction values ​​under the physical laws of battery degradation. If the consistency is high, it is directly output as the final result; if a contradiction is detected, an internal correction mechanism is triggered to generate a new set of physically more consistent prediction values ​​as the final joint prediction result.

[0066] This embodiment successfully decouples the feature learning of SOH and RUL by constructing a parameter-independent dual feature extractor, eliminating the task conflict in the traditional shared encoder. The two extractors can be optimized independently without interfering with each other. The introduction of the relation constraint establishes a mandatory physical and logical association at the output level, unifying the two independent paths into a coordinated prediction framework, realizing true joint prediction and improving prediction accuracy.

[0067] In this embodiment of the invention, the method further includes obtaining a pre-trained joint prediction model, wherein obtaining the pre-trained joint prediction model includes: obtaining a training sample set, wherein each training sample in the training sample set includes multimodal runtime sequence data of the battery, a true health status label, and a true remaining life label; constructing an initial model including a first feature extractor, a second feature extractor, and the relation constraint; for any sample in the training sample set, performing the following operations: inputting the multimodal runtime sequence data of the sample into the initial model to obtain a corresponding intermediate health status prediction value, an intermediate remaining life prediction value, and at least one intermediate consistency constraint signal; calculating a first prediction loss based on the intermediate health status prediction value and the true health status label; calculating a second prediction loss based on the intermediate remaining life prediction value and the true remaining life label; calculating a relation constraint loss based on the at least one intermediate consistency constraint signal; constructing a joint loss function based on the first prediction loss, the second prediction loss, and the relation constraint loss; and iteratively updating the parameters of the initial model with the goal of minimizing the joint loss function to obtain the joint prediction model.

[0068] In practical applications, loss-weighted summation can be used to train a joint prediction model. However, this approach has significant drawbacks: First, the weight coefficients need to be carefully adjusted and are sensitive to data distribution, resulting in high tuning costs; second, this linear weighting does not explicitly model the intrinsic relationship between the outputs of the two tasks, and the model is not strongly constrained by the requirement that the prediction results must conform to physical common sense during the learning process, which may lead to the trained model performing reasonably well on independent tasks, but with poor logical coherence in the joint output.

[0069] To address this issue, embodiments of the present invention construct a joint loss function that includes relational constraint loss for model training.

[0070] Specifically, firstly, a large amount of lithium battery lifecycle data, from healthy state to failure, is collected to construct a training sample set. Next, an initial model is initialized, containing a network structure of a first feature extractor, a second feature extractor, and a relation constraint generator. After model initialization, forward propagation is performed on each training sample, yielding three key outputs: 1) intermediate health state predictions generated by the first feature extractor; 2) intermediate remaining lifetime predictions generated by the second feature extractor; and 3) one or more intermediate consistency constraint signals generated by the relation constraint generator. These signals are scalars or vectors calculated internally by the relation constraint generator based on the two initial predictions, used to quantify their inconsistency. The loss is calculated based on these values, and the loss calculation is divided into three parts: the first prediction loss... The difference between the predicted intermediate health status and the actual data is usually calculated using mean squared error or mean absolute error. Second prediction loss. The same remaining useful life prediction value is used to calculate the difference between the intermediate remaining useful life prediction value and the actual data. Relationship constraint loss. The joint loss function is calculated based on the intermediate consistency constraint signal output by the relation constraint generator, and its purpose is to penalize physically inconsistent predictions. The final joint loss function is constructed as follows: ,in This is a hyperparameter used to balance prediction accuracy and logistic consistency. It is minimized using a gradient descent algorithm (such as Adam). Meanwhile, the parameters of the first feature extractor, the second feature extractor, and the relation constraint are iteratively updated, so that the model not only learns to fit the data, but also learns to generate joint predictions that conform to physical laws.

[0071] The joint loss function and training strategy proposed in this embodiment unify the two objectives of prediction accuracy and logical consistency within a single optimization framework. Compared to simple weighted loss, this training method, which incorporates explicit relational constraints, can more effectively prevent the model from getting trapped in local optima, improve the overall robustness and generalization ability of the model, and make the final deployed model more reliable.

[0072] In this embodiment of the invention, the calculation of relational constraint loss includes the calculation of starting point alignment loss; the calculation of starting point alignment loss includes: obtaining the function value of the parameterized capacity decay trajectory function at the starting time according to the parameterized capacity decay trajectory function output by the second feature extractor, as the trajectory starting point prediction value; calculating the first difference between the intermediate health state prediction value and the trajectory starting point prediction value; and determining the starting point alignment loss based on the first difference.

[0073] In the physical process of battery degradation, the current health state should numerically equal to the starting point of the capacity degradation trajectory from that moment. However, in existing frameworks, even those based on joint learning, the predictions of the current state often differ because health state prediction and remaining lifetime prediction are optimized relatively independently.

[0074] To address this issue, embodiments of the present invention introduce a start-point alignment loss to force alignment of this critical point that must be physically equal.

[0075] Specifically, in the parameterized d-capacity decay trajectory function output by the second feature extractor, the time variable is set to the starting time (i.e., 0), and the predicted capacity value of the trajectory at the starting time is calculated, which is used as the predicted trajectory starting point value. Simultaneously, the intermediate health state prediction value is obtained from the first feature extractor, and the difference between the two, i.e., the first difference, is calculated and used as a core component of the starting point alignment loss. In one possible implementation, this first difference can be directly used as the starting point alignment loss.

[0076] By introducing the starting point alignment loss, this invention establishes a robust and shared current state benchmark for two independent task predictions. This allows the accuracy of health state prediction to directly improve the accuracy of trajectory starting point. Conversely, the rationality of trajectory starting point will also constrain the range of health state prediction, greatly improving the coordination of joint prediction and enhancing the consistency of the model's internal representation.

[0077] In this embodiment of the invention, the calculation of relational constraint loss includes calculating cycle consistency loss; the calculation of cycle consistency loss includes: calculating a first implicit remaining lifetime value based on the intermediate health state prediction value and the preset capacity failure threshold; solving for a second implicit remaining lifetime value based on the parameterized capacity decay trajectory function and the preset capacity failure threshold; calculating a second difference between the first implicit remaining lifetime value and the second implicit remaining lifetime value; and determining the cycle consistency loss based on the second difference.

[0078] Given a current health status value and a hypothetical capacity failure threshold, a projected remaining lifetime can theoretically be derived. Similarly, given a predicted capacity decay trajectory, the remaining lifetime can also be calculated. In actual predictions, the remaining lifetime values ​​calculated through these two independent paths should be the same. However, due to model errors, they often differ.

[0079] To address the aforementioned issues, embodiments of the present invention introduce a cycle consistency loss.

[0080] Specifically, a first health state prediction and a preset capacity failure threshold are known. A first implicit remaining lifetime value is calculated based on a preset decay model. For example, in one implementation, the preset decay model may employ linear decay, calculated using a formula incorporating the average decay rate. More generally, this calculation can be performed by a small learnable network module. The implicit remaining lifetime is derived from the capacity decay trajectory function. For the parameterized capacity decay trajectory function, the equation is numerically solved to equal the preset capacity failure threshold; the resulting variable is the second implicit remaining lifetime value. To ensure differentiability, the solution process is implemented using a differentiable numerical method. The difference between the first and second implicit remaining lifetime values, i.e., the second difference, is calculated, and the cycle consistency loss is determined based on the second difference. By incorporating the cycle consistency loss into the joint loss function, the model is forced to ensure that the lifetime predicted from the health state matches the lifetime solved from the decay trajectory.

[0081] This embodiment introduces a cycle consistency loss, which adds a strong constraint of dynamic process consistency on the basis of static consistency of starting point alignment constraint. This greatly improves the physical credibility of joint prediction, making the model's behavior closer to a simulator that conforms to basic physical laws, rather than two independent guessers. This significantly improves the logical rigor and overall reliability of the prediction results.

[0082] In this embodiment of the invention, obtaining the parameterized capacity decay trajectory function using the second feature extractor includes: processing the multimodal runtime sequence data using the feature encoding network in the second feature extractor to obtain a trend feature vector; generating dynamic parameters based on the trend feature vector using the parameter generation network in the second feature extractor; and instantiating a feedforward neural network using the dynamic parameters, wherein the input of the feedforward neural network is a time variable, and the output is a predicted battery capacity value, and the feedforward neural network constitutes the parameterized capacity decay trajectory function.

[0083] Specifically, multimodal runtime sequence data is first processed by a feature encoding network. This network uses a deep structure including convolutional layers, recurrent layers, or attention layers to abstract and fuse the raw data layer by layer, ultimately extracting and outputting a fixed-dimensional trend feature vector. This vector encapsulates the key information contained in the current battery's historical operating data that determines its future capacity degradation trend. Subsequently, this trend feature vector is input into a parameter generation network, which is typically designed as a multilayer perceptron. The output of the parameter generation network is not a direct capacity prediction value, but a set of dynamic parameters. The dimensions and specific meanings of these parameters are predefined by the target function architecture to be instantiated. Finally, instantiation is achieved based on the dynamic parameters. Instantiation requires predefining a fixed-structure feedforward neural network as an architecture template, specifying, for example, the number of layers, the number of nodes, and the types of activation functions. All weights and bias parameters of this template network are dynamically assigned values ​​by the dynamic parameters output by the aforementioned parameter generation network. Therefore, for each different input sample, a feedforward neural network instance with a unique set of parameters is generated in real time. The input of this instantiated feedforward neural network is any future relative time point, and its output is the predicted capacity value at that relative time point. Ultimately, this feedforward neural network, defined by the dynamic parameter θ, constitutes the required parameterized capacity decay trajectory function.

[0084] This invention, through the generation of a continuous function, enables continuous querying and prediction of capacity values ​​at any future time, overcoming the limitations of discrete-time point prediction. Secondly, by dynamically controlling the expression of the feedforward neural network, it can flexibly fit various complex trajectories, from exponential decay to compound decay with inflection points, demonstrating strong pattern adaptability. Finally, because the main structure of the function is fixed, the complexity and capacity of the model are effectively controlled, ensuring the stability of the training process. Furthermore, the generated trajectory function is smooth and differentiable, facilitating subsequent accurate calculation of remaining lifetime and consistency constraint calculations. Compared to directly regressing fixed sequences or using static parameter models, this scheme exhibits significant advantages in prediction flexibility, accuracy, and reliability.

[0085] To further enhance the precision and physical plausibility of trajectory modeling, the dynamic parameters provided in this embodiment of the invention employ a hierarchical generation method. Specifically, the trend feature vector is fed into two sub-networks: one generates global parameters controlling the basic decay pattern, and the other generates local correction parameters adapted to subtle individual characteristics. These are then merged into dynamic parameters. This design enables the model to separate the common patterns of battery aging from individual differences. Furthermore, physically guided trajectory initialization can be implemented. For example, when instantiating the feedforward neural network, prior knowledge can be used to constrain the generation range of the dynamic parameters or set initial biases, making them tend to generate monotonically decreasing curves with reasonable starting points, thereby accelerating training and reducing the generation of non-physical trajectories.

[0086] In an embodiment of the invention, the method further includes applying physical constraints to the feedforward neural network; applying physical constraints to the feedforward neural network includes: configuring a monotonically non-increasing activation function for the output layer of the feedforward neural network; and / or configuring a range constraint function for the output layer of the feedforward neural network, the range constraint function being used to limit the output value of the feedforward neural network within the theoretical capacity range of the battery.

[0087] While neural networks possess powerful fitting capabilities, the functions they generate may violate basic physical principles, such as predicting capacity growth over time or exceeding theoretical limits. Such predictions, which defy physical laws, are not only erroneous but also undermine the logical foundation for subsequent consistency checks.

[0088] To address this problem, this invention proposes a constraint method to impose physical constraints on feedforward neural networks. Specifically, it configures the output layer of the feedforward neural network with a monotonically non-increasing activation function or employs a monotonic network structure design. For example, by ensuring that the network weights are non-negative and using a monotonically increasing activation function, and by taking a negative sign on the final output, the capacity value of the output is mathematically guaranteed to be monotonically non-increasing over time, preventing the anomalous situation where the capacity increases with use. Alternatively, a range constraint function can be added after the final output of the feedforward neural network. For example, a scaled sigmoid function can be used. ,in, Given the rated capacity, and the final output of the feedforward neural network, the output of the feedforward neural network can be forcibly mapped to... Interval.

[0089] This embodiment applies physical constraints, limiting all possible capacity decay trajectory functions generated by the second feature extractor to a reasonable function space that conforms to physical common sense. This significantly reduces the possibility of the model generating absurd predictions and improves the security and reliability of the output.

[0090] In this embodiment of the invention, the consistency verification and correction using a relation constraint includes: calculating a compatibility score between the first health status prediction value and the first remaining lifespan prediction value using the relation constraint; determining whether the compatibility score is lower than a preset confidence threshold; if the compatibility score is lower than the confidence threshold, entering a calibration loop; the calibration loop includes: keeping the parameters of the joint prediction model unchanged; adjusting the internal feature representations generated by the first feature extractor and the second feature extractor for the current input data with the goal of improving the compatibility score; recalculating based on the adjusted internal feature representations to obtain updated health status prediction values ​​and remaining lifespan prediction values, and recalculating the compatibility score; repeating the calibration loop until the compatibility score reaches or exceeds the confidence threshold, and using the last obtained prediction value as the final joint prediction result.

[0091] Specifically, after the multimodal runtime sequence data of the battery to be predicted is processed by the first and second feature extractors to obtain the first health state prediction value and the first remaining lifetime prediction value, these two initial prediction values ​​are jointly input into the relation constraint unit to initiate the following process:

[0092] Calculating the compatibility score: The relation constraint includes a built-in evaluation module to quantify the degree of consistency between the two predictions. This score can be calculated in several ways, such as calculating the joint likelihood probability of the two predictions under a simplified battery degradation physics model; or calculating their deviation from the internally implicit starting point alignment and cyclic consistency constraints. A higher score indicates that the two predictions are more physically consistent.

[0093] Judgment and Decision: The calculated compatibility score is compared with a preset confidence threshold. This threshold can be set according to the reliability requirements of the actual application.

[0094] If the score is greater than or equal to the threshold, it indicates that the initial prediction has sufficient internal consistency, and the relation constraint will directly output the first health state prediction value and the first remaining life prediction value as the final joint prediction result.

[0095] If the score is less than the threshold, it indicates a logical conflict in the initial prediction, and the relation constraint will activate an internal calibration loop. The core principle of this loop is to keep the weight parameters of all network layers in the joint prediction model (the first and second feature extractors and the relation constraint itself) fixed, and only optimize the intermediate feature representations generated by the two feature extractors in reverse for the current input sample.

[0096] The core principle of the calibration loop is to keep the weight parameters of all network layers (the first and second feature extractors and the relation constraint itself) in the joint prediction model constant, and only optimize the internal feature representations generated by the two feature extractors in reverse for the current input sample. Here, "internal feature representation" specifically refers to the output of an intermediate layer in the feature extractor before generating the first health state prediction or the first remaining lifetime prediction (e.g., the activation value of the last fully connected layer or convolutional layer).

[0097] The specific steps of the calibration loop are as follows: Feature back-adjustment: With maximizing the compatibility score as the optimization objective, the gradient of this objective with respect to a selected intermediate layer (i.e., its internal feature representation) of the first and second feature extractors is calculated using the backpropagation algorithm. Feature update: Based on the calculated gradient, the internal feature representation of the selected intermediate layer is updated along the gradient direction. This step is equivalent to fine-tuning the model's "understanding" or "feature extraction focus" for the current specific sample under fixed model "knowledge" (weights), aiming to find a new set of features that can derive more consistent prediction results. Re-prediction: Using the new features, the forward computation of subsequent network layers (such as the final fully connected output layer) is continued to obtain a new set of adjusted health status predictions and remaining lifespan predictions. Re-evaluation: Based on the new predictions, the compatibility score is recalculated. Iterative loop: The above process of "feature back-adjustment-re-evaluation" is repeated until the new compatibility score reaches or exceeds the confidence threshold, or the preset maximum number of iterations is reached. Output the final result: The set of predicted values ​​obtained when the calibration cycle ends (score reaches the target or iteration ends) is output as the final joint prediction result after consistency verification and correction.

[0098] The online calibration loop mechanism introduced in this invention can effectively correct prediction logic contradictions caused by input data noise, distribution shift, or model randomness bias, significantly improving the robustness and output reliability of the model in real complex application scenarios.

[0099] To further improve the efficiency and intelligence of the calibration cycle, inconsistency source diagnosis and selective calibration strategies can be introduced.

[0100] In this embodiment of the invention, obtaining the pre-trained joint prediction model further includes domain adaptation training; wherein, the training sample set contains data samples from multiple different domains, each sample having a domain label representing its domain; the domain adaptation training includes: connecting a gradient inversion layer and a domain classifier at the back end of the first feature extractor and / or the second feature extractor; using the domain classifier, predicting the domain to which the sample belongs based on the features output by the gradient inversion layer; calculating a domain adversarial loss based on the prediction result of the domain classifier and the domain label of the sample; adding the domain adversarial loss to the joint loss function, and inverting the gradient direction using the gradient inversion layer during backpropagation.

[0101] To enable the model to generalize across different operating conditions, this embodiment integrates domain adaptation training into the training process.

[0102] Specifically, at the end of the feature encoding network of the first and / or second feature extractors, a gradient inversion layer is added. This layer performs an identity mapping during forward propagation, but multiplies the incoming gradient by a negative coefficient during backpropagation. Following the gradient inversion layer is a lightweight domain classifier, typically a small MLP, whose task is to receive features and predict the domain from which the sample originates. Loss calculation and backpropagation: After obtaining the domain prediction results through forward propagation, the cross-entropy loss of the domain classifier, i.e., the domain adversarial loss, is calculated. During backpropagation, the gradient flows back from the domain classifier and is inverted and scaled in the gradient inversion layer, making its output features so that the domain classifier cannot distinguish which domain the sample originates from. In other words, the feature extractor is encouraged to discard information related to specific operating conditions, retaining only cross-domain shared information relevant to the nature of battery degradation. Adding the adversarial loss to the joint loss function yields... ,in, Indicating resistance to loss, This represents the hyperparameters corresponding to the adversarial loss.

[0103] This embodiment introduces a gradient reversal layer and domain adversarial loss, which forces the model to learn the degradation features that remain unchanged during the training phase, greatly enhancing the model's adaptability and prediction stability when facing new and unknown operating conditions.

[0104] Those skilled in the art will understand that although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the machine equivalents of the claims, the invention also intends to include these modifications and modifications.

Claims

1. A multi-task-driven joint prediction method for the state of health of lithium batteries, characterized in that, include: Obtain multimodal runtime sequence data of the battery to be predicted; The multimodal runtime sequence data is input into a pre-trained joint prediction model; the joint prediction model includes a first feature extractor and a second feature extractor with mutually exclusive parameters, as well as a relation constraint. The first feature extractor is used to process the multimodal runtime sequence data to obtain a first health state prediction value; The second feature extractor is used to process the multimodal runtime sequence data to obtain a parameterized capacity decay trajectory function; The first remaining lifetime prediction value is determined based on the parameterized capacity decay trajectory function and the preset capacity failure threshold. The relationship constraint is used to verify and correct the consistency between the first health status prediction value and the first remaining life prediction value to obtain the final joint prediction result. The method further includes obtaining a pre-trained joint prediction model, wherein obtaining the pre-trained joint prediction model includes: Obtain a training sample set, wherein each training sample in the training sample set includes multimodal runtime sequence data of the battery, a real health status label, and a real remaining life label; Construct an initial model that includes the first feature extractor, the second feature extractor, and the relation constraint; For any sample in the training sample set, perform the following operations: The multimodal runtime sequence data of the sample is input into the initial model to obtain the corresponding intermediate health status prediction value, intermediate remaining life prediction value and at least one intermediate consistency constraint signal. Calculate the first prediction loss based on the intermediate health status prediction value and the actual health status label; The second prediction loss is calculated based on the intermediate remaining lifetime prediction value and the actual remaining lifetime label; Calculate the relation constraint loss based on at least one intermediate consistency constraint signal; Based on the first prediction loss, the second prediction loss, and the relational constraint loss, a joint loss function is constructed; With the goal of minimizing the joint loss function, the parameters of the initial model are iteratively updated to obtain the joint prediction model.

2. The method according to claim 1, characterized in that, The computational relation constraint loss includes the computational starting point alignment loss; The calculation of the starting point alignment loss includes: Based on the parameterized capacity decay trajectory function output by the second feature extractor, the function value of the parameterized capacity decay trajectory function at the starting time is obtained as the trajectory starting point prediction value. Calculate the first difference between the intermediate health state prediction value and the trajectory starting point prediction value; The starting point alignment loss is determined based on the first difference.

3. The method according to claim 2, characterized in that, The computational relation constraint loss also includes the computational cycle consistency loss; The calculation of cycle consistency loss includes: The first implicit remaining lifetime value is calculated based on the intermediate health status prediction value and the preset capacity failure threshold. The second implicit remaining lifetime value is obtained by solving the parameterized capacity decay trajectory function and the preset capacity failure threshold. Calculate the second difference between the first implicit remaining lifetime value and the second implicit remaining lifetime value; The cycle consistency loss is determined based on the second difference.

4. The method according to claim 1, characterized in that, The step of processing the multimodal runtime sequence data using the second feature extractor to obtain the parameterized capacity decay trajectory function includes: The multimodal runtime sequence data is processed using the feature encoding network in the second feature extractor to obtain a trend feature vector; The parameter generation network in the second feature extractor generates dynamic parameters based on the trend feature vector. The feedforward neural network is instantiated using the dynamic parameters, wherein the input of the feedforward neural network is a time variable and the output is a predicted battery capacity value, and the feedforward neural network constitutes the parameterized capacity decay trajectory function.

5. The method according to claim 4, characterized in that, The method further includes applying physical constraints to the feedforward neural network; The physical constraints imposed on the feedforward neural network include: Configure a monotonically non-increasing activation function for the output layer of the feedforward neural network; and / or, Configure a range constraint function for the output layer of the feedforward neural network, the range constraint function being used to limit the output value of the feedforward neural network within the theoretical capacity range of the battery.

6. The method according to claim 1, characterized in that, The step of using the relation constraint to perform consistency verification and correction between the first health status prediction value and the first remaining life prediction value includes: The compatibility score between the first health status prediction value and the first remaining lifespan prediction value is calculated using the relation constraint. Determine whether the compatibility score is lower than a preset confidence threshold; If the compatibility score is lower than the confidence threshold, then proceed to the calibration cycle; The calibration cycle includes: Keep the parameters of the joint prediction model unchanged; With the goal of improving the compatibility score, the internal feature representations generated by the first feature extractor and the second feature extractor for the current input data are adjusted; Based on the adjusted internal feature representation, the forward calculation is re-performed to obtain updated health status predictions and remaining life predictions, and the compatibility score is recalculated. The calibration cycle is repeated until the compatibility score reaches or exceeds the confidence threshold, and the last predicted value is used as the final joint prediction result.

7. The method according to claim 1, characterized in that, The process of obtaining the pre-trained joint prediction model also includes domain adaptation training; wherein, the training sample set contains data samples from multiple different domains, and each sample has a domain label representing its domain. The domain adaptation training includes: A gradient inversion layer and a domain classifier are connected to the back end of the first feature extractor and / or the second feature extractor; Using the domain classifier, the domain to which a sample belongs is predicted based on the features output by the gradient inversion layer; Based on the prediction results of the domain classifier and the domain labels of the samples, calculate the domain adversarial loss; The domain adversarial loss is added to the joint loss function, and the gradient reversal layer is used to reverse the gradient direction during backpropagation.

Images