A method and system for predicting the capacity decay trend of lithium batteries

Abstract

This invention relates to the field of battery health management and prediction technology, specifically a method and system for predicting the capacity degradation trend of lithium batteries. The method includes: generating a deeply fused feature by combining data-driven features and electrochemical features through a dual calibration mechanism; inputting this feature sequence into an electrochemical process sensing model, encoding it as a potential state vector characterizing personalized degradation; then inputting this vector as an initial condition into a neural differential equation model, learning the degradation dynamics and solving it through integration to ultimately generate a continuous health state degradation trajectory and obtain the prediction result. This invention overcomes the weakness of traditional "black box" models in generalization ability by deeply integrating physical mechanisms with a data-driven model. It can achieve long-term, high-precision, and continuous prediction of the degradation trend throughout the entire battery lifecycle using only weak early-stage signals, exhibiting high reliability and practical value.

Description

Technical Field

[0001] This invention relates to the field of battery health management and prediction technology, specifically to a method and system for predicting the capacity decay trend of lithium batteries. Background Technology

[0002] Lithium-ion batteries are core components of server backup power systems such as uninterruptible power supplies (UPS) in data centers, and their capacity degradation is a key factor limiting their lifespan and safety. Therefore, accurately predicting the remaining battery life is crucial for ensuring reliable system operation and reducing costs. Existing prediction methods mostly employ data-driven models such as recurrent neural networks. These models, as "black boxes," lack physical mechanism constraints, have weak generalization ability, and struggle to capture nonlinear degradation. Furthermore, they are mostly discrete-point predictions, leading to significant error accumulation during long-term extrapolation, and are difficult to make reliable predictions based solely on weak early data.

[0003] Therefore, how to effectively integrate physical mechanisms and improve the reliability and generalization ability of long-term, continuous and high-precision life prediction based solely on weak early signals of the battery has become an urgent problem to be solved.

[0004] To address this, a method and system for predicting the capacity decay trend of lithium batteries are proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting the capacity decay trend of lithium batteries, which combines physical mechanism-guided feature engineering with decay dynamics modeling to achieve accurate early life prediction.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for predicting the capacity decay trend of lithium batteries, comprising:

[0008] Temporal features are learned from raw monitoring data using convolutional neural networks;

[0009] Incremental capacity analysis is performed on the original monitoring data to extract electrochemical features; a dual calibration mechanism is introduced, using the electrochemical features as guiding signals to generate deep fusion features;

[0010] The deep fusion features are input into the electrochemical process sensing model, which compresses and encodes the feature sequences of multiple early operating cycles of the lithium battery into a low-dimensional latent state vector that can characterize the personalized degradation mode of the lithium battery.

[0011] The low-dimensional potential state vector is used as an initial condition and input into the neural differential equation model; the neural differential equation model characterizes the continuous differential dynamics of the battery health state determined by the initial condition over time.

[0012] By integrating and solving the neural differential equation model, a health state decay trajectory of the lithium battery from the current moment to the end of its lifespan is generated, and the prediction result of the remaining lifespan is obtained from the health state decay trajectory.

[0013] Preferably, the dual calibration mechanism is implemented as follows:

[0014] When training the convolutional neural network, a composite loss function consisting of a weighted sum of the main task loss and the auxiliary physical loss is used for optimization. The main task loss is calculated based on the error between the decay trajectory of the healthy state and the actual decay trajectory. The auxiliary physical loss is calculated based on the error between the predicted value of the electrochemical feature by the temporal feature and the actual value of the electrochemical feature. The auxiliary physical loss is used as a physical regularization term during training to constrain the representation of the temporal feature.

[0015] Preferably, the electrochemical process sensing model includes a physical sensing input layer and a periodic encoder; the physical sensing input layer receives each deep fusion feature vector in the sequence composed of the deep fusion features, and sets independent embedding channels for feature components of different physical types in the vector, concatenating the embedded vectors processed by each channel to generate an enhanced periodic feature vector; the periodic encoder receives a sequence composed of multiple periodic feature vectors, and processes it through a one-way decaying attention mechanism to generate the low-dimensional latent state vector; the one-way decaying attention mechanism introduces a bias term that changes unidirectionally with the number of cycles when calculating the attention weights, and applies asymmetric weights to the enhanced periodic feature vectors at different positions in the sequence.

[0016] Preferably, the neural differential equation model incorporates a neural network for parameterizing the derivative function of the health state with respect to time; the input of the neural network includes the current health state, the current time t, and the low-dimensional latent state vector generated by the electrochemical process sensing model; the output of the neural network is the instantaneous value of the derivative function at the current time; the neural differential equation model uses an ordinary differential equation numerical solver to integrate the derivative function defined by the neural network with the state represented by the low-dimensional latent state vector as the initial value to generate the health state decay trajectory.

[0017] Preferably, the electrochemical characteristics include the peak position, peak height, peak area, and peak width in the incremental capacity curve.

[0018] Preferably, the step of obtaining the prediction result of the remaining service life from the health state decay trajectory includes: preset a health state threshold for the end of battery life; determine the moment when the generated health state decay trajectory first crosses the threshold, and take the difference between the current moment and the threshold crossing moment as the prediction result of the remaining service life.

[0019] Preferably, the feature sequence is generated based on monitoring data within the first 15% of the total expected cycle life of the lithium battery.

[0020] A lithium battery capacity degradation trend prediction system, comprising:

[0021] Feature engineering module: Learns time-series features from raw monitoring data through convolutional neural networks; performs incremental capacity analysis on the raw monitoring data to extract electrochemical features; introduces a dual calibration mechanism, using the electrochemical features as guiding signals to generate deep fusion features;

[0022] Degradation mode encoding module: The deep fusion features are input into the electrochemical process sensing model, which compresses and encodes the feature sequences of multiple early operating cycles of the lithium battery into a low-dimensional latent state vector that can characterize the personalized degradation mode of the lithium battery.

[0023] The decay trajectory generation module takes the low-dimensional potential state vector as an initial condition and inputs it into the neural differential equation model. The neural differential equation model represents the continuous differential dynamic law of the battery health state evolution over time determined by the initial condition. By integrating and solving the neural differential equation model, the health state decay trajectory of the lithium battery from the current moment to the end of its lifespan is directly generated, and the prediction result of the remaining service life is obtained from the health state decay trajectory.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. The dual calibration mechanism proposed in this invention uses physical mechanisms as guiding signals to force the convolutional neural network model to learn physically meaningful features. This overcomes the problems of poor generalization ability and inaccurate prediction when operating conditions change, which are caused by traditional black-box models relying solely on statistical correlation, making the model prediction more robust and reliable.

[0026] 2. Employing an "encoder-neural differential equation" architecture, the model is no longer a simple pattern match, but rather learns and represents the underlying differential dynamics of battery degradation. This enables the model to simulate the real physical degradation process when making long-term extrapolation predictions, effectively avoiding the problem of severe distortion in long-term prediction results caused by error accumulation in traditional sequence models.

[0027] 3. This invention directly generates a complete, mathematically continuous health state decay curve, rather than several discrete prediction points. This not only provides the health state at any future time, but also allows for precise identification of key information such as decay rate and performance knee point through analysis such as differentiating the trajectory. This enables more refined predictive maintenance and proactive safety management, greatly enhancing the practical value of the prediction results. Attached Figure Description

[0028] Figure 1 This is a flowchart of a method for predicting the capacity decay trend of a lithium battery according to an embodiment of the present invention.

[0029] Figure 2 This is a flowchart illustrating the internal workflow of the feature engineering module proposed in an embodiment of the present invention.

[0030] Figure 3 This is a system structure diagram of a lithium battery capacity decay trend prediction system proposed in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example 1

[0033] Please see Figure 1 This invention provides a method and system for predicting the capacity degradation trend of lithium batteries. To address the challenge of accurately predicting battery life in key application scenarios such as large-scale energy storage systems, the technical solution is as follows:

[0034] A method for predicting the capacity degradation trend of lithium batteries, such as Figure 1 As shown, it includes:

[0035] Temporal features are learned from raw monitoring data using convolutional neural networks;

[0036] Incremental capacity analysis is performed on the original monitoring data to extract electrochemical features; a dual calibration mechanism is introduced, using the electrochemical features as guiding signals to generate deep fusion features;

[0037] The deep fusion features are input into the electrochemical process sensing model, which compresses and encodes the feature sequences of multiple early operating cycles of the lithium battery into a low-dimensional latent state vector that can characterize the personalized degradation mode of the lithium battery.

[0038] The low-dimensional potential state vector is used as an initial condition and input into the neural differential equation model; the neural differential equation model characterizes the continuous differential dynamics of the battery health state determined by the initial condition over time.

[0039] By integrating and solving the neural differential equation model, a health state decay trajectory of the lithium battery from the current moment to the end of its lifespan is generated, and the prediction result of the remaining lifespan is obtained from the health state decay trajectory.

[0040] Specifically, the structure of a convolutional neural network is as follows: an input layer that receives a voltage / current sequence of length N; followed by two one-dimensional convolutional layers, the first layer having 16 filters, a kernel size of 5, and a stride of 1; the second layer having 32 filters, a kernel size of 3, and a stride of 1; each convolutional layer is followed by an activation function (such as ReLU) and a max-pooling layer with a window size of 2; then a flattening layer connected to a hidden fully connected layer (ReLU activation) containing 64 neurons; finally, the temporal features are output through a fully connected layer with an output dimension of M (e.g., 32).

[0041] Furthermore, the raw monitoring data includes time-series data of voltage, current, and battery temperature collected by sensors in the battery management system during one or more charge-discharge cycles of the lithium battery.

[0042] Further, the incremental capacity analysis specifically involves performing a differential transformation on the charging or discharging curves in the original monitoring data to obtain an ICA curve of incremental capacity (dQ / dV) versus voltage (V); and identifying characteristic peaks corresponding to the phase transition of the electrode material from the ICA curve, extracting the position, height, area, or width of the characteristic peaks as the electrochemical characteristics. Before performing the differential transformation, a Savitzky-Golay filter is first used to smooth the voltage and capacity data, with a filter window length of 21 and a polynomial order of 2, to effectively remove noise while preserving peak shape characteristics. Subsequently, the ICA curve is obtained by calculating the ratio of the capacity increment (dQ) to the voltage increment (dV). The main characteristic peaks on the ICA curve are located using a peak-finding algorithm (e.g., the find_peaks function in the SciPy library), setting parameters such as minimum peak height and minimum peak distance. For each identified peak, its peak voltage is used as the peak position, and the peak dQ / dV value is used as the peak height; the peak area is calculated by numerical integration over the peak region. Through data filtering and precise peak identification, noise is effectively suppressed, ensuring the accuracy, stability, and repeatability of the extracted electrochemical features.

[0043] Furthermore, the parameters of the peak finding algorithm, such as minimum peak height and minimum peak distance, are adaptively determined. Specifically, the minimum peak height threshold can be set to a specific percentage (e.g., 10%) of the current maximum peak value of the ICA curve, rather than a fixed absolute value, to adapt to the overall changes in peak values ​​at different aging stages. The rule used to determine these parameters is itself a function associated with the battery's current state of health (SOH). For example, a higher percentage threshold is used when the battery's SOH is high; when the SOH decreases, a lower percentage threshold is automatically used, allowing the sensitivity of peak identification to be dynamically adjusted according to the battery's aging state. This significantly improves the robustness and accuracy of key feature extraction throughout the battery's entire lifespan.

[0044] Furthermore, the specific implementation of the dual calibration mechanism is as follows:

[0045] When training the convolutional neural network, a composite loss function consisting of a weighted sum of the main task loss and the auxiliary physical loss is used for optimization. The main task loss is calculated based on the error between the decay trajectory of the healthy state and the actual decay trajectory. The auxiliary physical loss is calculated based on the error between the predicted value of the electrochemical feature by the temporal feature and the actual value of the electrochemical feature. The auxiliary physical loss is used as a physical regularization term during training to constrain the representation of the temporal feature.

[0046] The weighting method for the composite loss function can be implemented in several ways, including the following:

[0047] A fixed weighting method is adopted. The weighting relationship between the main task loss and the auxiliary physical loss is adjusted by a preset, fixed hyperparameter λ. The value of the hyperparameter λ can be determined by performing a grid search on the validation set or by selecting it empirically, and its preferred value range is between 0.1 and 1.0.

[0048] Dynamic weighting or annealing methods are employed. The weight λ is not a fixed value, but a function that changes with the training process (e.g., the number of training epochs). For example, a higher λ value can be set at the beginning of training to force the model to prioritize learning physical representations; as training progresses, the λ value is gradually reduced, allowing the model to shift its optimization focus more towards improving the final prediction accuracy once it has a physical foundation.

[0049] An adaptive weighting based on uncertainty is employed. Both the primary and auxiliary tasks are treated as probabilistic models. When the model outputs its prediction, it also outputs a measure of the uncertainty of that prediction. The loss function dynamically and adaptively adjusts its weights based on the uncertainty of each task, assigning higher weights to tasks that the model is "more confident" in, thus achieving automatic weight adjustment.

[0050] Both the primary task loss and the auxiliary physical loss are calculated using the mean squared error (MSE) function.

[0051] The network structure used to predict the electrochemical features from the temporal features is specifically a multilayer perceptron. This multilayer perceptron receives the temporal features output by the convolutional neural network as its input and outputs predicted values ​​for the electrochemical features. The multilayer perceptron includes two hidden layers and one output layer; wherein the number of neurons in the two hidden layers can be 128 and 64, respectively, and a modified linear unit (ReLU) is used as the activation function; the number of neurons in the output layer matches the number of electrochemical features to be predicted, and a linear activation function is used to output continuous predicted values.

[0052] By employing a dual calibration mechanism, this invention deeply integrates data-driven models with physical mechanisms. This method forces convolutional neural networks to learn physically meaningful features, overcoming the shortcomings of existing black-box models, such as weak generalization ability and reliance solely on statistical correlation. Therefore, this invention can achieve long-term, high-precision prediction of the degradation trend throughout the entire battery lifecycle using only weak signals from the early stages of battery life, significantly improving the reliability and stability of the prediction results.

[0053] Furthermore, the electrochemical process sensing model includes a physical sensing input layer and a periodic encoder. The physical sensing input layer receives each deep fusion feature vector in the sequence composed of the deep fusion features, and sets independent embedding channels for feature components of different physical types in the vector. The embedded vectors processed by each channel are concatenated to generate an enhanced periodic feature vector. The periodic encoder receives a sequence composed of multiple periodic feature vectors and processes it through a one-way decaying attention mechanism to generate the low-dimensional latent state vector. When calculating the attention weights, the one-way decaying attention mechanism introduces a bias term that changes unidirectionally with the number of cycles, and applies asymmetric weights to the enhanced periodic feature vectors at different positions in the sequence.

[0054] The physical perception input layer processes the sequence of deep fusion features vector by vector. Its internal structure and function are as follows:

[0055] The deep fusion features are divided into at least two different types based on their physical origin and properties, including: (1) the dynamic temporal feature group, i.e. the features derived from the convolutional neural network; and (2) the electrochemical index feature group, i.e. the features derived from the incremental capacity analysis.

[0056] The physical perception input layer configures an independent embedding channel for each of the aforementioned feature types. Each embedding channel is an independent multilayer perceptron used to map the corresponding type of features to a high-dimensional embedding space. Finally, the embedding vectors output from all independent channels are concatenated to form an enhanced periodic feature vector with better information representation.

[0057] The periodic encoder receives a sequence consisting of multiple enhanced periodic feature vectors as described above and processes it through a one-way decaying attention mechanism to generate the low-dimensional latent state vector.

[0058] The unidirectional decay attention mechanism introduces a pre-defined, non-learnable unidirectional decay bias matrix during the standard scaled dot product attention calculation process. Standard attention score matrix After calculation, the new attention score matrix pass Perform calculations, and then Input to the Softmax function.

[0059] The bias matrix The matrix elements are used to embed prior physical knowledge of the battery degradation process into attention computation. The value is determined by the position of the query vector. and the position of the key vector A joint decision, representing the [number]th The loop iterates through the first... Attention bias for each cycle. In a preferred embodiment, The calculation method is as follows:

[0060] when > Time (i.e., focusing on future cycles), It is set to negative infinity to achieve causal constraints;

[0061] when < Time (i.e., focusing on the past or current cycle), Set as , where λ is a positive decay coefficient. This bias term causes the model to apply a larger negative bias to past cycles with a more distant time distance when calculating attention, thus prioritizing the state of recent cycles.

[0062] The positive decay coefficient λ is a preset fixed hyperparameter. The value of λ can range from 0.01 to 0.5. The specific value of λ can be determined by cross-validation or grid search on the validation set to achieve optimal prediction performance.

[0063] An electrochemical process sensing model was constructed through in-depth modification of the encoder. This model, through its hierarchical structure and physical sensing input layer, effectively handles the multi-scale and heterogeneous nature of battery degradation data in terms of features and time. Most importantly, its "one-way degradation attention mechanism" embeds the cumulative physical priors of the degradation process into the core of the model, enabling it to learn more fundamental degradation patterns from early data.

[0064] Furthermore, the neural differential equation model incorporates a neural network for parameterizing the derivative function of the health state with respect to time; the input of the neural network includes the current health state, the current time t, and the low-dimensional latent state vector generated by the electrochemical process sensing model; the output of the neural network is the instantaneous value of the derivative function at the current time; the neural differential equation model uses an ordinary differential equation numerical solver, taking the state represented by the low-dimensional latent state vector as the initial value, to integrate the derivative function defined by the neural network to generate the health state decay trajectory.

[0065] Specifically, in this embodiment, the neural network can be a multilayer perceptron. The input of the multilayer perceptron is used to input the current health status. The current time t and the low-dimensional latent state vector are concatenated to form a combined input vector. This multilayer perceptron includes at least two fully connected hidden layers, each containing 128 neurons, and uses the hyperbolic tangent function (Tanh) as the activation function to ensure that the derivative function has good nonlinear fitting ability and smoothness.

[0066] The output layer of this multilayer perceptron is a linear layer containing a single neuron, and its output value is the derivative of the health state at the current moment. The instantaneous value.

[0067] The specific type of numerical solver for ordinary differential equations is an adaptive step-size solver, which can be the Dormon-Prince method or its equivalent algorithm. This type of solver dynamically adjusts its integration step size based on the local truncation error calculated in each step.

[0068] Furthermore, the initial conditions for the integral solution process of the neural differential equation model are subject to the following constraints:

[0069] The starting time of the integral solution ( Initial health status value It is calculated or determined directly based on the monitoring data of the last cycle among the early multiple operating cycles.

[0070] The low-dimensional latent state vector generated by the electrochemical process sensing model is not directly used as the starting value for integration, but rather as the derivative function. A constant condition input. It remains unchanged throughout the integration process and is used to parameterize and determine the specific form of the kinetic law, thereby ensuring that the generated healthy state decay trajectory can reflect the personalized decay mode of the battery.

[0071] The neural differential equation model adopted by the present invention does not directly predict future state points, but learns and simulates the underlying differential kinetic law of battery degradation. It makes the long-term extrapolation prediction extremely stable and provides a richer information basis for refined diagnosis.

[0072] Furthermore, the electrochemical characteristics include the peak position, peak height, peak area, and peak width in the incremental capacity curve.

[0073] Specifying the electrochemical characteristics as peak parameters provides high-quality inputs with clear physical meanings for the model, thereby significantly improving the reliability and accuracy of the model prediction.

[0074] Furthermore, the step of obtaining the prediction result of the remaining useful life from the healthy state decay trajectory includes: presetting a healthy state threshold at the end of battery life; determining the moment when the generated healthy state decay trajectory first crosses the threshold, and taking the difference between the current moment and the threshold crossing moment as the prediction result of the remaining useful life.

[0075] Specifically, the healthy state threshold (EOL) pre-configures a failure threshold of the healthy state (SOH) in the system according to the specific application scenario; for electric vehicle applications, this threshold can be set to 80%; for energy storage or other applications, it can be set to 70%.

[0076] The threshold crossing moment traverses the discrete time point sequence ( , ) generated by the neural differential equation model to find the first adjacent data point pair ( >=EOL threshold and <EOL threshold) ( , ) and ( , ). Then, by performing linear interpolation calculation between these two data points, the moment corresponding to when the healthy state decay trajectory is exactly equal to the EOL threshold is solved.

[0077] Obtain the current moment for this prediction. The final remaining useful life RUL prediction result is obtained through the formula RUL = - The calculation yields the result. The unit of RUL is consistent with the unit of time t, and can be either the number of cycles or calendar days.

[0078] This invention provides a clear and reproducible method for calculating the relative lifetime (RUL). By using interpolation to accurately calculate the threshold crossing time on a generated continuous decay trajectory, it overcomes the problems of fuzzy and low-precision RUL estimation caused by traditional discrete-point prediction. This method ensures that the final lifetime prediction result is no longer a rough estimate, but an objective and accurate numerical value, significantly improving the practicality and reliability of the prediction results.

[0079] Furthermore, the feature sequence is generated based on monitoring data within the first 15% of the total expected cycle life of the lithium battery.

[0080] The total expected cycle life is a known baseline number of cycles predetermined based on the application scenario.

[0081] When this method is applied to the R&D testing or production quality control stages of batteries, the "total expected cycle life" can be the design life or nominal life provided by the battery manufacturer in the product specifications. For example, if the design life of a certain model of battery is 2000 cycles, then the data for the first 15% is the data for the first 300 cycles.

[0082] When this method is applied to a battery management system (BMS) already deployed in an electric vehicle or energy storage system, the “total expected cycle life” can be the rated cycle life of the battery pack or the manufacturer’s warranty cycle life.

[0083] By limiting the required data for prediction to the very early stages of the battery's nominal lifespan (e.g., the first 15%), the lifespan assessment cycle is significantly shortened. In the R&D and production stages, this method enables rapid screening and grading of battery cells, significantly accelerating development and improving quality control efficiency. In practical applications, it can provide reliable, personalized lifespan predictions as early as possible, possessing significant engineering and economic value.

[0084] Furthermore, this embodiment may also include a degradation mode diagnostic step. The extracted electrochemical characteristics (such as the change in peak parameters of the ICA curve over time) are input into a preset diagnostic model or rule base to qualitatively or quantitatively output the dominant degradation mode of the current battery, such as lithium-ion loss, positive electrode active material loss, or negative electrode active material loss, thereby providing in-depth health insights beyond remaining lifespan. This step enhances the diagnostic function of degradation modes, enabling the system to analyze the root causes of failure and providing deeper health insights beyond lifespan prediction.

[0085] This invention firstly overcomes the weakness of traditional data-driven models in generalization ability by deeply integrating physical mechanisms into feature engineering through a dual calibration mechanism, thereby improving model reliability. Secondly, it adopts a novel architecture combining an electrochemical process sensing model and neural differential equations. This allows the model to learn and simulate the underlying dynamics of battery degradation, rather than simply pattern matching, which significantly improves the accuracy and stability of long-term predictions. Finally, this method can generate a complete and continuous degradation trajectory based solely on weak early-stage battery signals, achieving accurate and reliable early lifetime prediction, and possessing significant engineering application value.

[0086] Example 2

[0087] Please see Figures 2 to 3 , Figure 2 This is a flowchart illustrating the workflow within the feature engineering module. Figure 3 This is a block diagram of a lithium battery capacity degradation trend prediction system. This embodiment applies the system to a data center server backup power supply health management scenario. In this scenario, the system adopts an "edge-cloud" collaborative architecture, where the on-site battery management system acts as the data acquisition end, and the cloud server deploys the core prediction model. The system aims to utilize data collected during the early operating cycles of the backup power supply to provide maintenance personnel and the back-end management platform with high-precision predictions of the remaining battery life, effectively ensuring data center power supply security and achieving predictive maintenance. Specifically, as... Figure 2 As shown, the core prediction model deployed on a cloud server can be functionally divided into the following collaborative modules:

[0088] A lithium battery capacity degradation trend prediction system, such as Figure 2 , Figure 3 As shown, it includes:

[0089] Feature engineering module: Learns time-series features from raw monitoring data through convolutional neural networks; performs incremental capacity analysis on the raw monitoring data to extract electrochemical features; introduces a dual calibration mechanism, using the electrochemical features as guiding signals to generate deep fusion features;

[0090] Degradation mode encoding module: The deep fusion features are input into the electrochemical process sensing model, which compresses and encodes the feature sequences of multiple early operating cycles of the lithium battery into a low-dimensional latent state vector that can characterize the personalized degradation mode of the lithium battery.

[0091] The decay trajectory generation module takes the low-dimensional potential state vector as an initial condition and inputs it into the neural differential equation model. The neural differential equation model represents the continuous differential dynamic law of the battery health state evolution over time determined by the initial condition. By integrating and solving the neural differential equation model, the health state decay trajectory of the lithium battery from the current moment to the end of its lifespan is directly generated, and the prediction result of the remaining service life is obtained from the health state decay trajectory.

[0092] Specifically, the backup power supply's battery management system continuously collects and stores raw monitoring data such as voltage, current, and temperature during routine float charging or periodic charge-discharge maintenance. Under specific conditions, such as when the device is connected to a data center network or after charging is completed, the battery management system packages the collected early cycle data and securely uploads it to a cloud server via its built-in communication module.

[0093] After receiving data from a specific power device, the cloud server immediately activates its feature engineering module to process the data and generate a deep fusion feature sequence for the battery. Subsequently, the degradation mode encoding module compresses and encodes this feature sequence into a low-dimensional latent state vector that represents the battery's unique "degradation genes."

[0094] The decay trajectory generation module receives the potential state vector as a key initial condition, drives the neural differential equation model to perform integral solutions, thereby generating a complete health state decay trajectory "tailor-made" for the backup power supply from the current moment to the end of its lifespan, and calculates the specific remaining service life prediction value from it.

[0095] The calculated remaining lifespan prediction results can be pushed to the local monitoring system to update the battery health or estimated backup time information on the local monitoring interface; on the other hand, they can be stored in the cloud management platform for monitoring the health status of the entire fleet, early warning of high-risk batteries, and intelligent scheduling of predictive maintenance work orders.

[0096] The capacity degradation trend prediction system proposed in this invention features a modular design that enables in-depth analysis layer by layer from features and patterns to trajectories. The feature engineering module, through a dual calibration mechanism, deeply integrates physical mechanisms and data-driven methods, ensuring the quality and reliability of input features. Its degradation pattern encoding module accurately captures the unique degradation patterns of each battery. Finally, its trajectory generation module simulates the dynamics of degradation, achieving stable and accurate long-term lifetime prediction and generating more informative continuous degradation curves. Overall, this system solves the industry challenge of reliable prediction using only early data.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the capacity decay trend of a lithium battery, characterized in that, include: Temporal features are learned from raw monitoring data using convolutional neural networks; Incremental capacity analysis was performed on the raw monitoring data to extract electrochemical characteristics; A dual calibration mechanism is introduced, using the electrochemical features as guiding signals to generate deep fusion features; The dual calibration mechanism is specifically implemented as follows: when training the convolutional neural network, a composite loss function consisting of a weighted average of the main task loss and the auxiliary physical loss is used for optimization; the main task loss is calculated based on the error between the decay trajectory of the healthy state and the actual decay trajectory; the auxiliary physical loss is calculated based on the error between the predicted value of the electrochemical feature by the temporal features and the actual value of the electrochemical feature, and the auxiliary physical loss is used as a physical regularization term during training to constrain the representation of the temporal features; The deep fusion features are input into the electrochemical process sensing model, which compresses and encodes the feature sequences of multiple early operating cycles of the lithium battery into a low-dimensional latent state vector that can characterize the personalized degradation mode of the lithium battery. The low-dimensional potential state vector is used as an initial condition and input into the neural differential equation model; the neural differential equation model characterizes the continuous differential dynamics of the battery health state determined by the initial condition over time. By integrating and solving the neural differential equation model, a health state decay trajectory of the lithium battery from the current moment to the end of its lifespan is generated, and the prediction result of the remaining lifespan is obtained from the health state decay trajectory.

2. The method for predicting the capacity decay trend of a lithium battery according to claim 1, characterized in that: The electrochemical process sensing model includes a physical sensing input layer and a periodic encoder. The physical sensing input layer receives each deep fusion feature vector in the sequence composed of the deep fusion features, and sets independent embedding channels for feature components of different physical types in the vector. The embedded vectors processed by each channel are concatenated to generate an enhanced periodic feature vector. The periodic encoder receives a sequence composed of multiple periodic feature vectors and processes it through a one-way decaying attention mechanism to generate the low-dimensional latent state vector. When calculating the attention weights, the one-way decaying attention mechanism introduces a bias term that changes unidirectionally with the number of cycles, and applies asymmetric weights to the enhanced periodic feature vectors at different positions in the sequence.

3. The method for predicting the capacity decay trend of a lithium battery according to claim 1, characterized in that: The neural differential equation model incorporates a neural network for parameterizing the derivative function of the health state with respect to time. The inputs to the neural network include the current health state, the current time t, and the low-dimensional latent state vector generated by the electrochemical process sensing model. The output of the neural network is the instantaneous value of the derivative function at the current time. The neural differential equation model uses an ordinary differential equation numerical solver, taking the state represented by the low-dimensional latent state vector as the initial value, to integrate the derivative function defined by the neural network, generating the health state decay trajectory.

4. The method for predicting the capacity decay trend of a lithium battery according to claim 1, characterized in that: The electrochemical characteristics include the peak position, peak height, peak area, and peak width in the incremental capacity curve.

5. The method for predicting the capacity decay trend of a lithium battery according to claim 1, characterized in that, The step of obtaining the prediction result of the remaining service life from the health state decay trajectory includes: preset a health state threshold for the end of battery life; determine the moment when the generated health state decay trajectory first crosses the threshold, and take the difference between the current moment and the threshold crossing moment as the prediction result of the remaining service life.

6. The method for predicting the capacity decay trend of a lithium battery according to claim 1, characterized in that: The feature sequence is generated based on monitoring data from the first 15% of the total expected cycle life of the lithium battery.

7. A lithium battery capacity decay trend prediction system, characterized in that, include: Feature engineering module: Learns temporal features from raw monitoring data using convolutional neural networks; Incremental capacity analysis was performed on the raw monitoring data to extract electrochemical characteristics; A dual calibration mechanism is introduced, using the electrochemical features as guiding signals to generate deep fusion features; The dual calibration mechanism is specifically implemented as follows: when training the convolutional neural network, a composite loss function consisting of a weighted average of the main task loss and the auxiliary physical loss is used for optimization; the main task loss is calculated based on the error between the decay trajectory of the healthy state and the actual decay trajectory; the auxiliary physical loss is calculated based on the error between the predicted value of the electrochemical feature by the temporal features and the actual value of the electrochemical feature, and the auxiliary physical loss is used as a physical regularization term during training to constrain the representation of the temporal features; Degradation mode encoding module: The deep fusion features are input into the electrochemical process sensing model, which compresses and encodes the feature sequences of multiple early operating cycles of the lithium battery into a low-dimensional latent state vector that can characterize the personalized degradation mode of the lithium battery. The decay trajectory generation module takes the low-dimensional potential state vector as an initial condition and inputs it into the neural differential equation model. The neural differential equation model represents the continuous differential dynamic law of the battery health state evolution over time determined by the initial condition. By integrating and solving the neural differential equation model, the health state decay trajectory of the lithium battery from the current moment to the end of its lifespan is directly generated, and the prediction result of the remaining service life is obtained from the health state decay trajectory.

Images