A method, system, device, and storage medium for predicting battery life.

Abstract

This invention discloses a battery life prediction method, system, device, and storage medium. The method includes acquiring real-time multidimensional time-series data and ontological attribute data of a target battery; preprocessing the real-time multidimensional time-series data based on an adaptive constraint threshold determined by the ontological attribute data to obtain target multidimensional time-series data; extracting features from the target multidimensional time-series data to obtain a multidimensional feature dataset; determining an electrochemical physical model matching the target battery, and fusing the constructed first battery life prediction model and the electrochemical physical model to obtain a second battery life prediction model; extracting individual mechanism features of the target battery, and optimizing the second battery life prediction model based on the individual mechanism features to obtain a battery life prediction model; inputting the multidimensional feature dataset into the battery life prediction model, and outputting the battery life prediction result. This invention improves the accuracy of predicting the battery life of smart meters.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, and in particular to a battery life prediction method, system, device, and storage medium. Background Technology

[0002] As the core power supply unit for clock synchronization and data storage in the event of a power outage in a smart grid, the health status of the smart meter's clock battery directly determines the accuracy of metering and the reliability of power supply services.

[0003] The power operation and maintenance field is currently upgrading from the traditional model of regular replacement and emergency repair after a failure to data-driven predictive maintenance. Battery life prediction has become a key technology direction for smart meter status monitoring.

[0004] Existing battery life prediction methods are mostly designed for high-current charging and discharging scenarios of power batteries. For example, the XGBoost gradient boosting tree life assessment method, which relies on a large number of fault samples, is based on the characteristics of high-frequency charging and discharging and rapid capacity decay of power batteries in its modeling logic, feature system and physical constraints. It does not build a special solution to adapt to the weak signal and slow degradation characteristics of smart meter clock batteries. Existing methods, when used for life prediction of smart meter clock batteries, generally suffer from poor scenario adaptability and weak early warning capabilities.

[0005] Furthermore, the existing pure data-driven black box models lack feasible explanations and cannot solve the problems of cold start of small batches of batteries and adaptation to individual differences. The prediction results are difficult to support accurate and intelligent operation and maintenance decisions. Summary of the Invention

[0006] This invention provides a battery life prediction method, system, device, and storage medium to address the technical problem of how to improve existing battery life prediction methods, thereby achieving the technical effect of improving the accuracy of battery life prediction.

[0007] To address the aforementioned technical problems, this invention provides a battery life prediction method, comprising: Acquire real-time multidimensional time-series data of the target battery during static and pulse alternating operation processes, and correlate and acquire the intrinsic attribute data of the target battery. Based on the ontological attribute data, the adaptive constraint threshold of the target battery is determined. Based on the adaptive constraint threshold, the real-time multidimensional time series data is sequentially subjected to anomaly identification and missing data imputation processing to obtain the target multidimensional time series data. Feature extraction is performed on the target multidimensional time-series data to obtain a multidimensional feature dataset; wherein, the multidimensional feature dataset includes at least a subset of aging stage identification features; The aging stage of the target battery is determined based on the subset of features identified by the aging stage, and an electrochemical physical model matching the aging stage is determined. A first battery life prediction model is constructed, and the first battery life prediction model and the electrochemical physical model are fused to obtain a second battery life prediction model; Extract the individual mechanism features of the target battery, and optimize the second battery life prediction model based on the individual mechanism features to obtain the battery life prediction model; The multidimensional feature dataset is input into the battery life prediction model, and the battery life prediction result of the target battery is output.

[0008] As one preferred embodiment, the physical attribute data includes at least one of the following: battery model, production batch, nominal capacity, and initial internal resistance. The step of determining the adaptive constraint threshold of the target battery based on the ontological attribute data includes: Based on the ontological attribute data, an equivalent circuit model and a self-discharge physical model of the target battery are constructed. The theoretically reasonable range of the instantaneous rate of change of voltage and current during pulsed load is calculated based on the equivalent circuit model. Based on the theoretically reasonable range and the target battery's health status, ambient temperature, and historical statistical characteristics of pulse load, an adaptive constraint threshold is dynamically set.

[0009] As one preferred embodiment, the step of sequentially performing anomaly identification and missing data imputation processing on the real-time multidimensional time-series data according to the adaptive constraint threshold to obtain the target multidimensional time-series data includes: The real-time multidimensional time-series data is screened based on the adaptive constraint threshold to obtain the first multidimensional time-series data; wherein, the data screening is designed to remove abnormal data points that are not true battery responses; Based on the self-discharge physical model and the historical static data characteristics of the first multidimensional time series data, the missing data of the first multidimensional time series data is filled to obtain the target multidimensional time series data.

[0010] As one preferred embodiment, the multidimensional feature dataset further includes a physical feature subset, a statistical feature subset, and a deep feature subset; The step of extracting features from the target multidimensional time-series data to obtain a multidimensional feature dataset includes: The nonlinear features of the target multidimensional time-series data are extracted using a convolutional neural network to obtain the deep feature subset. The physical indicator features of battery operation are extracted from the target multidimensional time series data to obtain the physical feature subset; Statistical features of battery operation are extracted from the target multidimensional time-series data to obtain the statistical feature subset; The physical feature subset, the statistical feature subset, the deep feature subset, and the aging stage identification feature subset are integrated to obtain a multidimensional feature dataset.

[0011] As one preferred embodiment, the aging stage identification feature subset includes at least the rate of change of the voltage recovery time constant after the pulse and the long-term drift acceleration of the average voltage during the static period. Determining the aging stage of the target battery based on the subset of features identified by the aging stage includes: The degradation kinetic characteristic values ​​of the target battery are calculated based on the rate of change of the voltage recovery time constant after the pulse and the long-term drift acceleration of the average voltage during the static period. The degradation dynamics eigenvalues ​​are divided into stages based on a clustering algorithm to obtain the stage division results; The aging stage of the target battery is determined based on the stage division results. The aging stage includes the initial stable period, the intermediate accelerated degradation period, and the final critical failure period.

[0012] As one preferred embodiment, the extraction of individual mechanistic features of the target battery includes: Obtain a statistical summary of the operating data and key physical characteristics of the target battery during its initial operation phase; The statistical summary and key physical features are input into a lightweight auxiliary neural network for inference, and the individual mechanism features of the target battery are output, wherein the individual mechanism features are probability vectors characterizing the battery degradation mechanism type.

[0013] As one preferred embodiment, the first battery life prediction model is constructed based on a deep learning temporal network; The optimization of the second battery life prediction model based on the individual mechanism characteristics to obtain the battery life prediction model includes: The fusion weights of the electrochemical physics model and the deep learning temporal network are dynamically allocated according to the attention mechanism. The electrochemical physics model and the deep learning temporal network are fused based on the fusion weights to obtain the second battery lifetime prediction model. Based on the probability vector, the physical parameters of the second battery life prediction model are transferred to obtain the battery life prediction model.

[0014] Another aspect of the present invention provides a battery life prediction system, comprising: The acquisition module is used to acquire real-time multi-dimensional time-series data of the target battery during static and pulse alternating operation processes, and to associate and acquire the physical attribute data of the target battery. The preprocessing module is used to determine the adaptive constraint threshold of the target battery based on the ontological attribute data, and to perform anomaly identification and missing data imputation processing on the real-time multidimensional time series data according to the adaptive constraint threshold to obtain the target multidimensional time series data. An extraction module is used to extract features from the target multidimensional time-series data to obtain a multidimensional feature dataset; wherein, the multidimensional feature dataset includes at least a subset of aging stage identification features; An identification module is used to determine the aging stage of the target battery based on a subset of the aging stage identification features, and to determine an electrochemical physical model that matches the aging stage. A construction module is used to construct a first battery life prediction model, and to fuse the first battery life prediction model and the electrochemical physical model to obtain a second battery life prediction model. An optimization module is used to extract the individual mechanism features of the target battery and optimize the second battery life prediction model based on the individual mechanism features to obtain a battery life prediction model. The prediction module is used to input the multidimensional feature dataset into the battery life prediction model and output the battery life prediction result of the target battery.

[0015] Another embodiment of the present invention provides a battery life prediction device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the battery life prediction method as described above.

[0016] In another embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, the battery life prediction method described above is implemented.

[0017] Compared with the prior art, the beneficial effects of the present invention are at least one of the following: 1) This invention completes the anomaly identification and missing data imputation of time series data by using an adaptive constraint threshold driven by ontological attributes, purifying battery operation data under static-pulse alternating conditions from the source, significantly improving the reliability of the original data and the accuracy of subsequent modeling; at the same time, it accurately divides the battery degradation stage by relying on the feature subset identified by the aging stage and matches the corresponding electrochemical physical model, deeply integrating the physical mechanism with the data-driven model, overcoming the defects of poor interpretability and unreasonable boundary scenario prediction of traditional pure data black box models, and greatly improving the stability and accuracy of lifetime prediction.

[0018] 2) This invention optimizes the fusion model by extracting the individual mechanistic features of the target battery, effectively solving the industry pain points of cold start of small batches of batteries and difficulty in adapting to individual process differences. It achieves an upgrade from modeling general laws to individualized and accurate prediction. The model can dynamically switch mechanistic constraints according to the battery aging stage, adapt to the degradation characteristics of slow self-discharge and internal resistance growth throughout the process, significantly enhance the ability to capture early weak attenuation signals, and finally output stable and reliable life results, providing accurate and practical technical support for predictive maintenance of smart meter batteries. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of a battery life prediction method in one embodiment of the present invention; Figure 2 This is a structural block diagram of a battery life prediction system in one embodiment of the present invention; Figure 3 This is a structural block diagram of a battery life prediction device in one embodiment of the present invention; Figure label: The module comprises: 11. Acquisition module; 12. Preprocessing module; 13. Extraction module; 14. Recognition module; 15. Construction module; 16. Optimization module; 17. Prediction module; 21. Processor; 22. Memory. Detailed Implementation

[0020] 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. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] In the description of this invention, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0022] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0023] One embodiment of the present invention provides a battery life prediction method; for details, please refer to [link to specific documentation]. Figure 1 , Figure 1 The diagram shown is a flowchart of a battery life prediction method according to one embodiment of the present invention, which includes steps S1-S7: S1: Obtain real-time multi-dimensional time-series data of the target battery during static and pulse alternating operation, and correlate it with the target battery's intrinsic attribute data.

[0024] It should be noted that this embodiment is applicable to the life prediction scenario of smart meter clock batteries. The target battery is a lithium manganese battery or lithium thionyl chloride battery that provides power for clock maintenance and data storage in the smart meter during power outages. It operates under special conditions of alternating microampere-level static holding current and milliampere-level instantaneous pulse current. The real-time multidimensional time series data is continuous sampling data reflecting the battery's operating status, and the intrinsic attribute data is the physical and specification parameters inherent to the battery at the time of manufacture.

[0025] In this embodiment, the real-time multidimensional timing data includes battery terminal voltage timing data, pulse load current timing data, and ambient temperature timing data. The voltage timing data is collected separately in a static phase and a pulse phase. The static phase represents a low-power maintenance state without pulse triggering, while the pulse phase represents an instantaneous load state triggered by meter communication. The current timing data is used to distinguish between the static operating point and the pulse peak value, and the ambient temperature timing data is used to characterize the external stress conditions experienced by the battery.

[0026] Preferably, voltage time-series data is collected once per minute during the static phase and 100 times per second during the pulse phase to fully capture the voltage drop and recovery process during the pulse. Current time-series data is collected simultaneously for static holding current and pulse peak current. Ambient temperature data is collected every 10 minutes for temperature compensation in subsequent data processing and model building. This invention does not limit the collection frequency and can be adjusted according to the meter power consumption, battery type, and deployment environment.

[0027] In this embodiment, the collected real-time multidimensional time-series data is associated and bound with the target battery's intrinsic attribute data. The intrinsic attribute data includes at least one of the following: battery model, production batch, nominal capacity, and initial internal resistance. Specifically, the battery model is used to distinguish between lithium manganese systems and lithium thionyl chloride systems; the production batch is used to characterize the consistency of the battery manufacturing process; the nominal capacity is the battery's factory rated capacity; and the initial internal resistance is the battery's internal resistance parameter before it is put into use.

[0028] The association method involves storing real-time multidimensional time-series data and ontological attribute data under the same battery unique identifier in the same data unit. This ensures that subsequent steps such as adaptive constraint threshold calculation, anomaly identification, missing data imputation, feature extraction, and model prediction are all based on the complete data after association, thus avoiding prediction deviations due to data misalignment.

[0029] It should be noted that the static and pulse alternating operation process in this embodiment refers to a working mode in which the battery powers the clock circuit with a microampere-level current most of the time, and outputs a milliampere-level instantaneous pulse current only when the meter data is uploaded or a command is responded to. The two states alternate periodically. Real-time multidimensional time-series data refers to a numerical sequence that is continuously collected in chronological order and contains multiple physical dimensions. It has the characteristics of temporal sequence, numerical continuity, and operating condition correlation, providing raw data support for subsequent aging stage identification and lifespan prediction.

[0030] This invention does not limit the hardware precision of data acquisition, but preferably adopts an acquisition method that meets the requirements of metrological accuracy to ensure that the sampling errors of voltage, current and temperature are within a preset range, so as to ensure the reliability of subsequent data processing and prediction results.

[0031] S2: Determine the adaptive constraint threshold of the target battery based on the ontological attribute data, and perform anomaly identification and missing data imputation processing on the real-time multidimensional time series data according to the adaptive constraint threshold to obtain the target multidimensional time series data.

[0032] It should be noted that this step is implemented based on the real-time multidimensional time-series data and ontological attribute data acquired in step S1, and is applicable to the data governance stage of smart meter clock batteries under static-pulse alternating operating conditions. Because the target battery is susceptible to factors such as grid impacts, sensor drift, and data acquisition interruptions during operation, real-time multidimensional time-series data often exhibits problems such as false pulses, abnormal jumps, and missing local time periods. Directly using this data for feature extraction and lifetime prediction will lead to a significant decrease in accuracy. Therefore, it is necessary to construct a physical model based on the inherent properties of the battery and generate adaptive constraint thresholds to achieve accurate purification and completion of the original data.

[0033] In this embodiment, an equivalent circuit model and a self-discharge physical model of the target battery are first constructed based on the battery's intrinsic attribute data. The intrinsic attribute data includes at least one of the following: battery model, production batch, nominal capacity, and initial internal resistance. Different battery models and batches correspond to different electrochemical characteristics and initial internal resistance values. The equivalent circuit model constructed accordingly consists of an ideal voltage source, an ohmic internal resistance, a polarization resistor, and a polarization capacitor connected in series, used to characterize the battery's electro-response characteristics under pulsed load. The self-discharge physical model characterizes the slow voltage decay law of the battery during the static maintenance phase, including the self-discharge rate and temperature correction coefficient. This invention does not limit the specific structure of the equivalent circuit model or the expression of the self-discharge physical model, as long as they match the electrochemical characteristics of the target battery.

[0034] Preferably, based on the constructed equivalent circuit model, the theoretically reasonable range of the instantaneous voltage change rate and current of the target battery during pulse load is calculated. The instantaneous voltage change rate reflects the rate of voltage drop at the moment of pulse triggering, and the theoretically reasonable range of current corresponds to the actual pulse current range generated by the meter's communication action. Values ​​exceeding this range can be determined as non-real battery responses. Furthermore, combining the target battery's currently estimated health status, real-time ambient temperature, and historical statistical characteristics of pulse loads within a preset sliding time window, an adaptive constraint threshold is dynamically set using a combination of statistical outlier detection and physical allowable range. This allows the threshold to adaptively adjust with battery aging and environmental changes, improving the accuracy of anomaly identification.

[0035] In this embodiment, real-time multidimensional time-series data is screened based on a determined adaptive constraint threshold to obtain the first multidimensional time-series data. Specifically, the data screening involves comparing the real-time collected voltage and current time-series values ​​with the adaptive constraint threshold, removing abnormal pulse points and voltage jump points caused by instantaneous grid impacts, sensor malfunctions, or poor line contact, and retaining only valid time-series data that conforms to the battery's actual physical response, thus avoiding interference from abnormal data with subsequent feature extraction and model calculation.

[0036] Furthermore, based on the self-discharge physical model and the historical static data characteristics in the first multi-dimensional time-series data, missing data completion processing is performed on the first multi-dimensional time-series data to obtain the target multi-dimensional time-series data. The missing data mainly occurs during the static acquisition period between pulse load intervals. The completion process is achieved using the mechanism-constrained interpolation method. Based on the static voltage and voltage drift rate before and after the missing period, the theoretical voltage value at the missing time is calculated in conjunction with the self-discharge physical model, rather than using traditional linear interpolation, to ensure that the completed data conforms to the electrochemical degradation law of the battery itself.

[0037] It should be noted that the target multidimensional time-series data obtained after anomaly identification and missing data imputation is a pure, continuous time-series dataset that conforms to physical mechanisms. It can be directly used for subsequent feature extraction and aging stage identification, and is a key data foundation for improving the accuracy and reliability of lifetime prediction. This invention does not limit the specific judgment rules for anomaly identification or the calculation details of missing data imputation. As long as a physical model is built based on ontological attributes and data governance is completed with adaptive constraint thresholds, it falls within the protection scope of this invention.

[0038] S3: Extract features from the target multidimensional time series data to obtain a multidimensional feature dataset; wherein, the multidimensional feature dataset includes at least a subset of aging stage identification features.

[0039] It should be noted that this step is implemented based on the target multidimensional time-series data obtained in step S2, and belongs to the feature engineering link of fusion aging stage identification. The degradation signal of the smart meter clock battery is weak and highly nonlinear, and it is difficult to accurately characterize its health status by relying on a single type of feature. Therefore, it is necessary to construct a complete feature system from four dimensions: physical mechanism, statistical law, deep nonlinear expression, and aging dynamics, so as to provide high-quality input for subsequent adaptive modeling and individualized life prediction.

[0040] In this embodiment, firstly, multi-type feature extraction is performed on the target multi-dimensional time-series data. The multi-dimensional feature dataset includes a physical feature subset, a statistical feature subset, a deep feature subset, and an aging stage identification feature subset. The physical feature subset directly extracts core physical indicators of battery operation from the target multi-dimensional time-series data, including peak current under pulsed load, voltage recovery amplitude after pulse, average voltage during the static period, battery internal resistance, average ambient temperature, pulse duration, and pulse interval. These features are all directly related to the battery's electrochemical degradation mechanism and can intuitively reflect changes in battery performance. This invention does not limit the specific dimensions of the physical features and can adapt them to different battery models and operating conditions.

[0041] Preferably, statistical calculations are performed on the target multidimensional time-series data to obtain a subset of statistical features. These features include statistical values ​​of the mean, variance, kurtosis, skewness, voltage drift rate, and pulse interval time of the voltage time-series data, used to characterize the distribution characteristics and fluctuation patterns of the time-series data. Simultaneously, a one-dimensional convolutional neural network is used to perform nonlinear feature mining on the target multidimensional time-series data. By setting convolutional kernels of different sizes, local and global time-dependent features are extracted, outputting a fixed-dimensional deep feature subset to capture implicit degradation information that cannot be represented by traditional manual features, thereby improving the completeness of feature representation.

[0042] In this embodiment, a separate subset of aging stage identification features is constructed. This subset includes at least two key indicators: the rate of change of the voltage recovery time constant after the pulse and the long-term drift acceleration of the average voltage during the static period. The rate of change of the voltage recovery time constant after the pulse is used to characterize the degradation rate of the battery's pulse response capability, while the long-term drift acceleration of the average voltage during the static period is used to characterize the cumulative effect of battery self-discharge and internal resistance growth. The combination of these two indicators can accurately reflect the transformation of the battery degradation kinetic mode, providing a direct basis for the division of aging stages.

[0043] Furthermore, the extracted subsets of physical features, statistical features, deep features, and aging stage identification features are normalized and concatenated to form a unified multidimensional feature dataset. Normalization eliminates differences in units and numerical ranges between different features, ensuring a balanced contribution of each feature during model training. The integrated multidimensional feature dataset possesses mechanistic interpretability, statistical stability, and deep expressive power, and can be directly input into the battery life prediction model.

[0044] It should be noted that this invention does not limit the specific structure of the convolutional neural network, the calculation method of statistical features, or the feature integration method. As long as the above four types of features are extracted based on the target multidimensional time-series data and a fused feature set is formed, they all fall within the protection scope of this invention. The multidimensional feature dataset constructed through this step can significantly improve the model's ability to identify weak early-stage degradation signals of the battery, laying the foundation for subsequent staged and individualized lifespan prediction.

[0045] S4: Determine the aging stage of the target battery based on the feature subset identified by the aging stage, and determine the electrochemical physical model that matches the aging stage.

[0046] It should be noted that this step is implemented based on the multidimensional feature dataset constructed in step S3. The degradation process of the smart meter clock battery throughout its entire life cycle exhibits significant nonlinear and stage-based characteristics. A single electrochemical physical model is difficult to fully match the degradation pattern throughout the entire cycle. Therefore, it is necessary to accurately classify the battery stage based on the feature subset identified by the aging stage, and then match the dominant degradation physical model of the corresponding stage to provide a mechanistic basis for adaptive lifetime prediction in subsequent stages.

[0047] In this embodiment, the aging stage identification feature subset includes at least two core indicators: the rate of change of the voltage recovery time constant after the pulse and the long-term drift acceleration of the average voltage during the static period. The rate of change of the voltage recovery time constant after the pulse characterizes the rate of decay of the battery's response capability under pulsed load, while the long-term drift acceleration of the average voltage during the static period characterizes the slow voltage shift trend caused by battery self-discharge and internal resistance growth. These two indicators together constitute the core features that reflect the transformation of the battery's degradation kinetic mode.

[0048] Based on the above two characteristic indicators, the degradation kinetic characteristic value of the target battery is calculated. This characteristic value is used as input data for stage division, and a clustering algorithm is used to classify the degradation kinetic characteristic value to obtain the corresponding stage division result. Preferably, the clustering algorithm used includes at least one of the K-means algorithm, K-means++ algorithm, or semi-supervised clustering algorithm. This invention is not limited to this, as long as it can achieve stable classification of the degradation kinetic characteristic value.

[0049] Based on the stage division results, the aging stages of the target battery are divided into the initial stable period, the intermediate accelerated degradation period, and the final critical failure period. In this embodiment, the initial stable period corresponds to the initial stage of battery operation, where the degradation trend is gradual and the performance is stable; the intermediate accelerated degradation period corresponds to a significant increase in the battery's internal resistance and self-discharge rate, and the performance enters a rapid decline stage; the final critical failure period corresponds to the battery approaching the failure boundary, where the voltage and response capability deteriorate sharply.

[0050] Furthermore, based on the determined aging stages, appropriate electrochemical physical models are matched for the target battery. In the initial stable phase, the self-discharge mechanism model and the SEI film growth model are preferentially matched; in the mid-stage accelerated degradation phase, the internal resistance growth model and the electrolyte loss model are preferentially matched; and in the final critical failure phase, the internal resistance surge model and the electrode aging model are preferentially matched. The electrochemical physical models for different stages correspond to the dominant degradation mechanisms of that stage, ensuring that subsequent model fusion and prediction processes conform to the actual electrochemical changes of the battery.

[0051] It should be noted that this invention does not limit the specific threshold division of each aging stage, the parameter settings of the clustering algorithm, or the specific expression of the electrochemical physical model. These can be adaptively adjusted according to the battery model, usage environment, and operating conditions. Any implementation method that completes the stage division and matches the corresponding electrochemical physical model based on the feature subset identified by the aging stage falls within the protection scope of this invention.

[0052] S5: Construct a first battery life prediction model, and then merge the first battery life prediction model with the electrochemical physics model to obtain a second battery life prediction model.

[0053] It should be noted that this step is implemented based on the aforementioned aging stage division and electrochemical-physical model matching results, and is used to complete the construction, fusion, and individual optimization of the battery life prediction model. Because smart meter clock batteries exhibit batch differences, operating condition differences, and degradation mechanism differences, relying solely on deep learning time-series networks or a single electrochemical-physical model cannot achieve high-precision predictions. Therefore, it is necessary to fuse the data-driven model with the physical mechanism model, and optimize the model based on the individual battery mechanism characteristics, so that the model possesses both generalization ability and individual adaptability.

[0054] In this embodiment, a first battery life prediction model is first constructed based on a deep learning temporal network. The deep learning temporal network is preferably a long short-term memory network (LSTM) used to learn the temporal dependencies and nonlinear degradation trends of the target multidimensional temporal data, thereby capturing the implicit changes in the battery during long-term operation. The constructed first battery life prediction model is then initially fused with the electrochemical physical model matched in step S4 to obtain a second battery life prediction model. During the fusion process, an attention mechanism is used to dynamically allocate the weights of the electrochemical physical model and the deep learning temporal network, ensuring that different aging stages are matched with the corresponding dominant modeling approach.

[0055] Preferably, the attention mechanism adaptively adjusts the weights according to the aging stage of the battery: increasing the weight of the electrochemical physical model during the initial stable period, maintaining a balance between the weights of the two models during the mid-stage accelerated degradation period, and increasing the weight of the deep learning temporal network during the final critical failure period, so that the second battery life prediction model maintains stable prediction accuracy throughout the entire life cycle. This invention does not limit the specific structure of the deep learning temporal network or the calculation method of the attention mechanism, and can be adaptively adjusted according to the battery type and operating conditions.

[0056] S6: Extract the individual mechanism features of the target battery, optimize the second battery life prediction model based on the individual mechanism features, and obtain the battery life prediction model.

[0057] Furthermore, the individual mechanism characteristics of the target battery are extracted. First, the statistical summary of the operating data and key physical characteristics of the target battery in the early stage of operation are obtained. The early stage of operation is the period of 1 to 3 months after the battery is put into operation. The statistical summary includes statistical quantities such as average voltage, current fluctuation range, and temperature distribution. The key physical characteristics include indicators that are strongly related to the mechanism, such as battery internal resistance, pulse recovery rate, and voltage drift rate.

[0058] The statistical summary and key physical features are input into a lightweight auxiliary neural network to perform forward inference, outputting individual mechanistic features of the target battery. These individual mechanistic features are probability vectors characterizing the battery degradation mechanism type, used to distinguish degradation types such as self-discharge-dominated, internal resistance growth-dominated, electrolyte loss-dominated, or interface polarization-dominated. The lightweight auxiliary neural network contains only a few hidden layers, enabling rapid inference under small sample conditions and avoiding overfitting.

[0059] In this embodiment, the second battery lifetime prediction model is optimized through transfer learning based on the probability vector to obtain the final battery lifetime prediction model. Specifically, according to the individual degradation type indicated by the probability vector, the electrochemical physical model parameters in the second battery lifetime prediction model are fine-tuned. The fine-tuning parameters include the activation energy in the Arrhenius equation and the initial values ​​of the components in the equivalent circuit model, so that the model parameters are consistent with the individual characteristics of the target battery.

[0060] It should be noted that the battery life prediction model optimized through transfer learning integrates staged physical mechanisms, deep learning temporal features, and individual battery characteristics, effectively addressing issues such as small sample cold start, poor adaptation due to individual differences, and unreasonable predictions in boundary scenarios. This invention does not limit the specific algorithm or parameter fine-tuning range of transfer learning; any method that optimizes the model based on individual mechanism features and achieves physical-data fusion prediction falls within the scope of this invention.

[0061] S7: Input the multidimensional feature dataset into the battery life prediction model and output the battery life prediction result of the target battery.

[0062] It should be noted that this step is implemented based on the aforementioned model construction, feature extraction, and aging stage segmentation. It is a crucial step in inputting multi-dimensional feature data into the individualized optimized battery life prediction model, completing inference, and outputting the final prediction result. Since the target battery is in a static-pulse alternating operating condition, and the degradation patterns of different aging stages and different individuals vary significantly, it is necessary to rely on a physical-data hybrid model with dynamic weighted fusion using an attention mechanism, combined with parameters optimized by transfer learning, to achieve accurate predictions, ensuring that the results have both physical plausibility and data accuracy.

[0063] In this embodiment, the multidimensional feature dataset obtained in step S3 is normalized and preprocessed before being input into the battery life prediction model optimized by individual mechanism features for forward inference. During inference, the model automatically allocates the fusion weights of the electrochemical physics model and the deep learning temporal network based on the attention mechanism. These weights are adaptively adjusted according to the aging stage of the battery: the weight of the electrochemical physics model is increased in the initial stable period, the weights of the two models are balanced in the middle accelerated degradation period, and the weight of the deep learning temporal network is increased in the final critical failure period, so that the model can match the actual degradation mode of the battery throughout its entire life cycle. This invention does not limit the specific allocation ratio of the attention weights and can be adaptively set according to the battery model and environmental conditions.

[0064] Preferably, the battery life prediction model used in the inference process has been optimized by transfer learning based on the probability vector representing the degradation mechanism type. The key physical parameters inside the model, including the activation energy in the Arrhenius equation and the initial values ​​of the components in the equivalent circuit model, have been fine-tuned according to the individual characteristics of the target battery. This enables precise matching of the degradation mechanism bias of a single battery, such as self-discharge dominance, internal resistance growth dominance, or electrolyte loss dominance, effectively solving the prediction bias problem caused by individual differences and cold start of small samples.

[0065] In this embodiment, the battery life prediction model performs inference calculations based on the input multidimensional feature dataset and outputs the battery life prediction result for the target battery. The prediction result includes the remaining life value, the remaining life probability distribution, and the health status level. Among them, the remaining life probability distribution is used to characterize the uncertainty of the prediction result, and the health status level is used to intuitively reflect the current performance level of the battery, which can be directly used for subsequent operation and maintenance decisions and maintenance work order generation.

[0066] It should be noted that this invention does not limit the specific calculation process of model inference, the format and dimensions of the output results. As long as the multi-dimensional feature dataset is input into the model after stage adaptive fusion and individual optimization, and the output life prediction results conform to the battery electrochemical degradation law, it falls within the protection scope of this invention. The prediction results obtained through this step have the characteristics of strong early warning capability, high physical interpretability, and good individual adaptability, and can provide reliable support for the predictive maintenance of smart meter clock batteries.

[0067] The present invention also provides several specific embodiments of the above-mentioned battery life prediction method to illustrate the beneficial effects of the solution.

[0068] Example 1: Single-phase smart meter based on CR2032 battery (normal environment) example.

[0069] This embodiment applies to a single-phase smart meter deployed in a smart grid in a plain area (ambient temperature -5℃ to 35℃, humidity 40% to 70%). Its clock battery uses a CR2032 lithium-manganese battery with a nominal capacity of 220mAh and a static holding current of 30kWh. The instantaneous pulse current peak is 8mA, and the duration is 30ms. It is mainly used for electricity metering for residential users. The requirements are that the battery life prediction error is ≤8% and the early warning accuracy is ≥95%.

[0070] Step 1: Mechanism-enhanced data acquisition and management.

[0071] By utilizing the high-precision voltage, current, and ambient temperature sensors built into the smart meter, combined with an edge acquisition module, multi-dimensional time-series data acquisition is completed. Specific parameters are as follows: Voltage data: In the static stage (no pulse load), the sampling frequency is 1 time / minute, and the sampling accuracy is ±0.001V; in the pulse load stage, the sampling frequency is 100 times / second, and the sampling accuracy is ±0.0005V. Current data: Focus on collecting microampere-level static sustaining current (collection range 10~50μA, accuracy ±1μA) and milliampere-level instantaneous pulse current (collection range 5~10mA, accuracy ±0.1mA). Ambient temperature data: Data collection frequency: once every 10 minutes, collection range: -10℃ to 40℃, accuracy: ±0.5℃; Battery attribute data: Collect basic information such as battery model (CR2032), production batch (202401), nominal capacity (220mAh), initial internal resistance (80Ω), and manufacturer.

[0072] Data governance employs dedicated rules tailored to the specific operating mode of CR2032 batteries: a) Pulse data anomaly identification: Based on the equivalent circuit model of the CR2032 battery (composed of an ideal voltage source, ohmic internal resistance, polarization resistor, and polarization capacitor connected in series), the theoretical reasonable range for the instantaneous voltage drop rate during pulse load is calculated to be 0.01~0.05V / ms, and the theoretical reasonable range for the current is 5~10mA; an adaptive threshold is designed, using a 5-minute sliding time window, combined with 3 Statistical outlier detection is performed within the allowable range of the aforementioned physical equations to dynamically identify and eliminate abnormal pulse data points—including voltage drops caused by instantaneous power grid impacts (drop rate > 0.08V / ms), false pulses caused by sensor mis-triggers (current < 2mA or > 15mA), and pulse signal distortion data caused by poor contact.

[0073] b) Mechanistic interpolation of missing data: For static data loss during pulse load gaps (such as 10-30 minutes of data loss due to temporary sensor failure), a mechanism-constrained interpolation method is adopted. Based on the CR2032 battery self-discharge physical model (self-discharge rate 0.5% / year @ 25℃, the self-discharge rate doubles for every 10℃ increase in temperature), combined with the static voltage data characteristics (average voltage, voltage drift rate) of 30 minutes before and after the missing period, the theoretical voltage value of the missing period is calculated, avoiding the system deviation caused by traditional linear interpolation. The interpolation error is ≤0.002V.

[0074] Step 2: Feature engineering for aging stage identification.

[0075] Construct physical feature subsets, statistical feature subsets, and deep feature subsets, and add a new aging stage identification feature subset, as detailed below: 1) Physical feature subset (8 dimensions): peak current of pulsed load, voltage recovery amplitude after pulse, average voltage during static period, battery internal resistance (calculated by pulse voltage-current response), mean ambient temperature, variance of ambient temperature, pulse duration, and pulse interval time. 2) Statistical feature subset (6 dimensions): mean, variance, kurtosis, skewness of voltage time series data, statistical values ​​of voltage drift rate, and statistical values ​​of pulse interval time; 3) Deep feature subset (16-dimensional): Features are extracted from voltage and current time series data using a one-dimensional convolutional neural network (CNN). The network contains two convolutional layers (with kernel sizes of 3 and 5 respectively), one pooling layer, and one fully connected layer, outputting a 16-dimensional deep feature vector. 4) Aging stage identification feature subset (2D): the rate of change of the voltage recovery time constant after the pulse (calculation period 7 days, unit: s / day), and the long-term drift acceleration of the average voltage during the static period (calculation period 30 days, unit: V / day). 2 ).

[0076] By identifying feature subsets based on aging stages, and employing the K-means clustering algorithm, the current aging stage of the battery can be dynamically determined. Initial stabilization period: recovery time constant change rate < 0.01 s / day, voltage drift acceleration < 1e-6 V / day 2 ; Mid-term accelerated degradation period: 0.01 s / day ≤ rate of change of recovery time constant < 0.05 s / day, 1e-6V / day 2 ≤Voltage drift acceleration<5e-6V / day 2 ; Critical period of final failure: rate of change of recovery time constant ≥ 0.05 s / day, voltage drift acceleration ≥ 5e-6 V / day 2 .

[0077] Step 3: Physical Mechanisms of Stage Adaptation and Individualization - Data-Driven Collaborative Modeling.

[0078] A multi-stage adaptive physical information neural network model is constructed. The overall structure includes an input layer, a feature fusion layer, a stage-adaptive mechanism fusion layer, a deep learning temporal sub-network (LSTM network), and an output layer. The specific settings are as follows: a) Stage-adaptive mechanism fusion: This is achieved through a trainable attention mechanism. The attention mechanism receives the aging stage identification results and features from step S2 and automatically assigns appropriate attention scores to the electrochemical degradation physical model for different aging stages. Initial stabilization period: Activation of the self-discharge mechanism model (exponential self-discharge formula) and the SEI membrane growth model (parabolic growth model), with attention score accounting for 60%; deep learning temporal sub-network attention score accounting for 40%. Mid-term accelerated degradation phase: Activate the internal resistance growth model (linear growth formula) and the electrolyte loss model (linear loss formula), with attention score accounting for 50%; deep learning temporal sub-network attention score accounts for 50%; End-stage failure critical period: activation internal resistance surge model (exponential growth formula) and electrode aging model (power function degradation formula), attention score accounts for 40%; deep learning temporal sub-network attention score accounts for 60%.

[0079] b) Individual Mechanism Feature Extraction and Adaptation: Within the framework of transfer learning and meta-learning, a lightweight auxiliary neural network (12-dimensional input layer, 2 hidden layers, and 6-dimensional output layer) is constructed. This network takes the statistical summary of physical features and key physical features of the battery in the first 3 months as input and outputs a 6-dimensional probability vector representing the main degradation mechanism types of the battery (corresponding to the dominant probabilities of three degradation mechanisms: electrolyte loss, interface polarization, and self-discharge). Based on this probability vector, the key parameters of the general physical model are fine-tuned—the activation energy in the Arrhenius equation (adjustment range ±5%) and the initial value of the ohmic internal resistance in the equivalent circuit model (adjustment range ±8%), to achieve dual adaptation between the general degradation law and the differences in individual mechanisms.

[0080] Model training: Pre-training was performed using 5 years of operational data (including the complete life cycle) from 1000 CR2032 batteries of the same model. Fine-tuning was performed using the initial 3 months of operational data from batteries of the currently deployed electricity meters. The loss function used was mean squared error (MSE) combined with physical constraints (to ensure that the model output conforms to the physical laws of battery degradation). The training iterations were 500 times, the learning rate was 0.001, and the model fit R² ≥ 0.98 after training was completed.

[0081] Step 4: Continuously evolving online prediction and adaptation.

[0082] A two-layer adaptive mechanism is adopted to ensure continuous model evolution. The specific settings are as follows: 1) First layer (gradual fine-tuning): Every 7 days, the model parameters are slightly fine-tuned using the latest 7-day battery runtime data (learning rate 0.0005) to track the natural aging process of the battery and avoid model performance drift over time. 2) Second layer (abnormal trigger recalibration): Triggering conditions include three categories: Key physical indicator deviation: The absolute value of the deviation between the predicted and actual values ​​of the voltage recovery rate is ≥10% and lasts for 3 consecutive days; Aging stage transition warning: If the battery is detected to transition from the initial stable period to the mid-accelerated degradation period, or from the mid-accelerated degradation period to the final failure critical period within 15 days; Significant deviation in individual mechanism characteristics: The degradation mechanism probability vector output by the auxiliary neural network deviates from the model preset value (average level of batteries in the same batch) by ≥20%.

[0083] When any trigger condition is met, the model recalibration process is initiated, and the model is fully calibrated using battery operation data from the past 30 days. The calibration time is ≤2 hours to ensure that the model can quickly adapt to sudden changes in battery state. After calibration, the model prediction error is ≤8%.

[0084] Step 5: Decision outputs that drive predictive maintenance.

[0085] The model output includes: the current health status level of the battery (based on the probability distribution of remaining lifespan: excellent ≥80%, good 60%~80%, average 40%~60%, poor 20%~40%, and failure <20%), and the probability distribution of remaining lifespan (mean, variance, and 95% confidence interval). By combining the geographical location information of electricity meters, the distribution of inspection personnel and vehicle resources in the GIS system, a genetic algorithm is used to optimize maintenance work orders: for batteries with a health status level of "poor" or "failed", maintenance work orders are generated first; for batteries with a health status level of "medium", work orders are generated in order of remaining life from shortest to longest; at the same time, the inspection route is optimized to shorten the inspection mileage and improve the operation and maintenance efficiency.

[0086] In this embodiment, the lifespan of the CR2032 clock battery of 100 single-phase smart meters is predicted. Compared with the actual battery failure time, the prediction error is ≤7.5%, and the early warning accuracy rate is ≥96.2%. Compared with the traditional periodic inspection mode, the operation and maintenance cost is reduced by 42%, and the loss of power data and billing anomalies caused by battery failure are effectively avoided.

[0087] Example 2: Three-phase smart meter based on ER14505 battery (high temperature environment) example.

[0088] This embodiment applies to a three-phase smart meter deployed in a smart grid in a high-temperature region (ambient temperature 15℃~45℃, extreme summer temperatures reaching 50℃, humidity 60%~85%). Its clock battery uses a lithium thionyl chloride battery, model ER14505, with a nominal capacity of 2400mAh and a static holding current of 50... The instantaneous pulse current peak is 15mA, and the duration is 50ms. It is mainly used for electricity metering for industrial users. It requires battery life prediction error ≤10% and early warning accuracy ≥93%. It can adapt to the accelerated battery degradation scenario in high temperature environment.

[0089] Step 1: Mechanism-enhanced data acquisition and management.

[0090] A high-temperature resistant sensor (operating temperature -10℃~60℃) and an enhanced edge acquisition module are used to complete multi-dimensional time-series data acquisition. Specific parameters are as follows: Voltage data: In the static stage, the sampling frequency is 1 time / 30 seconds (intensified sampling under high temperature environment), and the sampling accuracy is ±0.001V; in the pulse load stage, the sampling frequency is 200 times / second, and the sampling accuracy is ±0.0005V. Current data: Static sustaining current acquisition range 30~70μA (accuracy ±1μA), instantaneous pulse current acquisition range 10~20mA (accuracy ±0.1mA); Ambient temperature data: collected once every 5 minutes (increased to once every 2 minutes during high-temperature periods), with a collection range of -10℃ to 60℃ and an accuracy of ±0.3℃; additionally, the internal temperature of the meter is collected (linked to the ambient temperature to reflect the actual operating temperature of the battery). Battery attribute data: Collect information such as battery model (ER14505), production batch (202312), nominal capacity (2400mAh), initial internal resistance (50Ω), high temperature resistance rating, and manufacturer.

[0091] Data governance rules optimized for high-temperature environments: a) Pulse data anomaly identification: Based on the equivalent circuit model of the ER14505 battery and combined with the internal resistance change law under high temperature environment (internal resistance decreases by 5% for every 10℃ increase in temperature), the theoretical reasonable range of voltage instantaneous drop rate during pulse load is calculated to be 0.02~0.06V / ms, and the theoretical reasonable range of current is 10~20mA; the adaptive threshold combined with the high temperature correction coefficient (the threshold is relaxed by 15% when the temperature is ≥40℃), using a 3-minute sliding time window, combined with 3σ statistical outlier detection and the allowable range of physical equations, eliminates abnormal pulse data caused by high temperature interference (such as sensor drift and pulse signal attenuation caused by high temperature).

[0092] b) Missing data mechanism interpolation: For static data loss (10~60 minutes) caused by sensor failure under high temperature environment, an improved mechanism constraint interpolation method is adopted. Based on the high temperature self-discharge physical model of ER14505 battery (self-discharge rate 1% / year @ 25℃, the self-discharge rate increases by 1.2 times for every 10℃ increase in temperature), combined with the static voltage and temperature data characteristics of 60 minutes before and after the missing period, the theoretical voltage value of the missing period is calculated. The interpolation error is ≤0.003V, which is suitable for the rapid self-discharge characteristics under high temperature environment.

[0093] Step 2: Feature engineering for aging stage identification.

[0094] Based on Example 1, the feature subset is optimized to adapt to high-temperature environments and the characteristics of ER14505 batteries: 1) Physical feature subset (10 dimensions): Two new features have been added: the difference between the internal temperature of the meter and the ambient temperature, and the duration of high temperature (duration of daily average temperature ≥ 35℃), to adapt to the battery degradation characteristics under high temperature environment. 2) Statistical feature subset (7 dimensions): Added statistical values ​​of voltage drift rate during high temperature periods, optimized the calculation weights of variance and kurtosis, and highlighted the impact of high temperature on battery performance; 3) Deep feature subset (16-dimensional): Adjust the kernel size of the CNN network (4, 6), add a high-temperature feature attention mechanism, and improve the extraction accuracy of high-temperature related features; 4) Aging stage identification feature subset (3-dimensional): Added high temperature accelerated degradation coefficient (calculated based on daily average temperature, unit: (°C)) -1 The threshold for dividing the aging stages has been optimized (the threshold for the mid-term accelerated degradation period and the final failure critical period is relaxed by 20% under high temperature conditions) to adapt to the rapid aging pattern of batteries under high temperature conditions.

[0095] By identifying the feature subset of the optimized aging stage, and using the K-means++ clustering algorithm, the battery aging stage is dynamically determined, with a stage identification accuracy of ≥94%.

[0096] Step 3: Physical Mechanisms of Stage Adaptation and Individualization - Data-Driven Collaborative Modeling.

[0097] Based on the model in Example 1, high-temperature adaptation optimization was performed: a) Stage-adaptive mechanism fusion: A new high-temperature degradation physical model (electrolyte decomposition acceleration model at high temperatures) is added, and the attention score for different aging stages is adjusted to adapt to the high-temperature environment. Initial stable phase (high-temperature environment): Activation of the self-discharge mechanism model, SEI film growth model, and high-temperature degradation physical model, with attention score accounting for 65%; deep learning temporal sub-network attention score accounting for 35%. Mid-term accelerated degradation period (high-temperature environment): Activation of the internal resistance growth model, electrolyte loss model, and high-temperature degradation physical model, with attention score accounting for 55%; deep learning temporal sub-network attention score accounting for 45%. End-stage failure critical period (high temperature environment): activation internal resistance surge model, electrode aging model and high temperature degradation physical model, attention score accounts for 45%; deep learning temporal sub-network attention score accounts for 55%.

[0098] b) Individual mechanism feature extraction and adaptation: The auxiliary neural network adds high temperature response feature input (voltage recovery curve and internal resistance change rate at high temperature), the output probability vector adds high temperature degradation dominant probability, and the fine-tuned general physical model parameters add high temperature correction terms (such as temperature coefficient optimization in the Arrhenius equation) to adapt to the individual degradation differences of different batteries under high temperature environment.

[0099] Model training: Pre-training was performed using 4 years of high-temperature environment operation data (including the complete life cycle) of 800 ER14505 batteries of the same model. Fine-tuning was performed using the initial 2 months of operation data of batteries in the currently deployed electricity meters (high-temperature period accounts for ≥60%). The training iterations were 600 times, the learning rate was 0.001, and the model fit was R²≥0.97 after training, which is suitable for the battery degradation pattern under high temperature environment.

[0100] Step 4: Continuously evolving online prediction and adaptation.

[0101] A dual-layer adaptive mechanism is optimized to adapt to sudden changes in battery state under high-temperature environments: 1) First layer (gradual fine-tuning): Every 5 days, the model parameters are slightly fine-tuned using the latest 5 days of battery operation data (with a focus on retaining data from high-temperature periods) to track the rapid aging process of the battery under high-temperature conditions. 2) Second layer (abnormal trigger recalibration): Add high temperature trigger condition (daily average temperature ≥ 45℃ for 5 consecutive days). When any trigger condition is met, the model recalibration process is started. The battery operation data of the past 20 days (with a focus on high temperature period data) is used for comprehensive calibration. The calibration time is ≤ 1.5 hours and the model prediction error after calibration is ≤ 9.8%.

[0102] Step 5: Decision outputs that drive predictive maintenance.

[0103] The model output now includes a new high-temperature degradation risk level (low, medium, high). Combined with the electricity usage priority of industrial users (key enterprises, general enterprises), the priority of maintenance work orders is optimized: for key enterprise meters with a high high-temperature degradation risk level and a health status level of medium or below, maintenance work orders are generated first; for general enterprise meters, work orders are generated according to the health status level and remaining lifespan.

[0104] In this embodiment, the lifespan of the ER14505 clock battery of 80 three-phase smart meters is predicted. Compared with the actual battery failure time, the prediction error is ≤9.5%, and the early warning accuracy is ≥93.5%. It can effectively warn of accelerated battery failure under high temperature environment. Compared with the traditional passive fault replacement mode, it reduces industrial power metering anomalies caused by battery failure by more than 80%, and improves operation and maintenance response efficiency by 60%.

[0105] Example 3: A novel smart meter based on CR2450 battery (small sample cold start) example.

[0106] This embodiment applies to a newly deployed smart meter (using a new metering chip with lower power consumption) of a power company. Its clock battery is a CR2450 lithium-manganese battery with a nominal capacity of 620mAh and a static holding current of 20... The instantaneous pulse current peak is 6mA and the duration is 20ms. This model of meter and its matching battery are being deployed in batches for the first time. There is a lack of complete life cycle operation data (only 1 month of initial operation data is available). It belongs to a small sample cold start scenario. The battery life prediction error is required to be ≤12%, the early warning accuracy rate is ≥90%, and it can quickly achieve accurate prediction in small sample scenarios.

[0107] Step 1: Mechanism-enhanced data acquisition and governance.

[0108] A new high-precision data acquisition module is adopted, adapted to the low-power characteristics of the new smart meter. The data acquisition parameters are as follows: Voltage data: In the static phase, the sampling frequency is 1 time / 2 minutes (adapted to low power consumption), and the sampling accuracy is ±0.001V; in the pulse load phase, the sampling frequency is 100 times / second, and the sampling accuracy is ±0.0005V. Current data: Static sustaining current acquisition range 15~25 (Accuracy ±1) The instantaneous pulse current acquisition range is 4~8mA (accuracy ±0.1mA). Ambient temperature data: Data collection frequency: once every 15 minutes, collection range: -10℃~40℃, accuracy: ±0.5℃; Battery attribute data: Collect information such as battery model (CR2450), production batch (202403), nominal capacity (620mAh), initial internal resistance (70Ω), and new meter compatibility parameters (in small sample scenarios, focus on supplementing battery production process parameters).

[0109] Data governance rules optimized for small sample scenarios: a) Pulse data anomaly identification: Based on the equivalent circuit model of the CR2450 battery and combined with the low-power pulse characteristics of the new type of meter, the theoretical reasonable range of the instantaneous voltage drop rate during pulse load is calculated to be 0.008~0.04V / ms, and the theoretical reasonable range of the current is 4~8mA; the adaptive threshold adopts a 2-minute sliding time window, and combined with the statistical characteristics of small sample data (small sample size, relax the outlier detection threshold to 4σ), abnormal pulse data points are removed to reduce abnormal data interference in small sample scenarios.

[0110] b) Missing data mechanism interpolation: To address the issue of high data missing rate (up to 15%) in small sample scenarios, the mechanism constraint interpolation method is optimized. Combined with battery production process parameters (such as electrode materials and electrolyte concentration), the prior information for missing data is supplemented to reduce the impact of missing small sample data on interpolation accuracy. The interpolation error is ≤0.003V.

[0111] Step 2: Feature engineering for aging stage identification.

[0112] To adapt to small-sample cold start scenarios, optimize feature subsets and reduce reliance on large amounts of historical data: 1) Physical feature subset (7 dimensions): retain core physical features (pulse peak current, static average voltage, battery internal resistance, and average ambient temperature), and remove features that are highly dependent on historical data (such as pulse interval time statistics). 2) Statistical feature subset (5 dimensions): Optimize the calculation method of statistical features, use short-term time series data (1 week) to calculate statistical values, and reduce the dependence on long-term historical data; 3) Deep feature subset (12-dimensional): Simplifies the CNN network structure (1 convolutional layer, 1 pooling layer, 1 fully connected layer), adapts to feature extraction of small sample data, and avoids overfitting; 4) Aging stage identification feature subset (2D): retain the rate of change of voltage recovery time constant after pulse and the long-term drift acceleration of the average voltage during the static period, and adjust the calculation cycle (3 days for the calculation cycle of the rate of change of recovery time constant and 15 days for the calculation cycle of voltage drift acceleration) to adapt to early stage identification in small sample scenarios.

[0113] By transferring learning, the experience in identifying aging stages in Examples 1 and 2 is transferred to this example. Combined with small initial sample data, a semi-supervised clustering algorithm is used to dynamically determine the battery aging stage, with a stage identification accuracy of ≥90%.

[0114] Step 3: Physical Mechanisms of Stage Adaptation and Individualization - Data-Driven Collaborative Modeling.

[0115] The focus is on optimizing the cold start capability with few samples, based on a transfer learning and meta-learning framework. The specific settings are as follows: a) Stage Adaptive Mechanism Fusion: Transfer the physical model parameters from Examples 1 and 2 (fine-tuning them in conjunction with the characteristics of the CR2450 battery), adopt the MAML (Model Independent Meta-Learning) algorithm in meta-learning, use a small amount of life cycle data of similar batteries (CR2032, ER14505) to pre-train the general degradation law of the model, and then use the initial one-month operating data of the new meter battery in this example to quickly fine-tune the model parameters to adapt to the degradation characteristics of the CR2450 battery.

[0116] b) Individual mechanism feature extraction and adaptation: The auxiliary neural network adopts a lightweight structure (10-dimensional input layer, 1 hidden layer, and 4-dimensional output layer), focusing on extracting individual difference features (such as the rate of change of internal resistance and temperature sensitivity) in the initial small sample data. The output probability vector is simplified to 4-dimensional (corresponding to the core degradation mechanism). Based on this feature, the parameters of the general physical model are fine-tuned to achieve individualized adaptation in small sample scenarios.

[0117] Model training: Meta-pre-training was performed using lifecycle data from 500 similar lithium manganese batteries (CR2032, ER14505). Fine-tuning was performed using one month of initial operation data from the new type of meter battery in this embodiment. The training iterations were 300 times, with a learning rate of 0.002. After training, the model fit R2 ≥ 0.95, solving the problem of cold start with small samples and achieving effective prediction in the early stage of operation.

[0118] Step 4: Continuously evolving online prediction and adaptation.

[0119] To adapt to small sample cold start scenarios, the two-layer adaptive mechanism is optimized to accelerate model evolution: 1) First layer (gradual fine-tuning): Every 3 days, the model parameters are slightly fine-tuned using the latest 3 days of battery operation data (learning rate 0.001) to quickly accumulate operation data of the new type of meter battery and optimize the model prediction accuracy. 2) Second layer (abnormal trigger recalibration): Lower the trigger condition threshold (absolute deviation of key physical indicators ≥12% and lasting for 2 days), add a small sample data accumulation trigger condition (actively start recalibration once for every month of accumulated running data), and ensure that the model is continuously optimized as data accumulates, gradually improving the prediction accuracy.

[0120] Step 5: Decision outputs that drive predictive maintenance.

[0121] The model output now includes prediction accuracy levels (high, medium, and low). Combined with the deployment batches of the new meters (first batch and second batch), the maintenance work order generation strategy is optimized: for batteries with a prediction accuracy level of "high" and a health status level of "poor," an emergency maintenance work order is generated; for batteries with a prediction accuracy level of "medium," prediction results are updated regularly based on data accumulation progress, and the maintenance plan is gradually optimized; for the first batch of deployed meter batteries, the inspection frequency is appropriately increased, and preventative maintenance is carried out based on prediction results.

[0122] In this embodiment, the lifespan prediction of the CR2450 clock battery of 60 new smart meters (small sample cold start scenario) was carried out. The initial prediction error was ≤11.8%. With the accumulation of data (after 6 months of operation), the prediction error dropped to ≤8.5%, and the early warning accuracy was ≥91%. This successfully solved the problem of small sample cold start prediction of new meter batteries, achieved effective early warning in the early stage of operation, and promoted the transformation of the operation and maintenance mode of new meters to predictive maintenance.

[0123] Another embodiment of the present invention provides a battery life prediction system; for details, please refer to [link to relevant documentation]. Figure 2 , Figure 2 The diagram shown is a structural block diagram of a battery life prediction system according to one embodiment of the present invention.

[0124] The acquisition module 11 is used to acquire real-time multi-dimensional time-series data of the target battery during static and pulse alternating operation processes, and to associate and acquire the target battery's intrinsic attribute data. Preprocessing module 12 is used to determine the adaptive constraint threshold of the target battery based on the ontology attribute data, and to perform anomaly identification and missing data imputation processing on the real-time multidimensional time series data according to the adaptive constraint threshold to obtain the target multidimensional time series data. Extraction module 13 is used to extract features from the target multidimensional time series data to obtain a multidimensional feature dataset; wherein, the multidimensional feature dataset includes at least a subset of aging stage identification features; The identification module 14 is used to determine the aging stage of the target battery based on the feature subset of the aging stage identification, and to determine the electrochemical physical model that matches the aging stage. Module 15 is used to build a first battery lifetime prediction model and fuse the first battery lifetime prediction model with the electrochemical physics model to obtain a second battery lifetime prediction model. The optimization module 16 is used to extract the individual mechanism features of the target battery and optimize the second battery life prediction model based on the individual mechanism features to obtain the battery life prediction model. The prediction module 17 is used to input the multidimensional feature dataset into the battery life prediction model and output the battery life prediction result of the target battery.

[0125] Another embodiment of the present invention provides a battery life prediction device, specifically, see [link to details]. Figure 3 This is a structural block diagram of a battery life prediction device provided in an embodiment of the present invention. The battery life prediction device includes a processor 21, a memory 22, and a computer program stored in the memory 22 and configured to be executed by the processor 21. When the processor 21 executes the computer program, it implements the steps described in the above-described battery life prediction method embodiment, for example... Figure 1 The steps S1 to S7 described above; or, when the processor 21 executes the computer program, it implements the functions of each module in the above-described device embodiments, such as the acquisition module 11.

[0126] For example, the computer program may be divided into one or more modules, which are stored in the memory 22 and executed by the processor 21 to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the battery life prediction device.

[0127] The battery life prediction device may include, but is not limited to, a processor 21 and a memory 22. Those skilled in the art will understand that the schematic diagram is merely an example of a battery life prediction device and does not constitute a limitation on the device. It may include more or fewer components than illustrated, or combine certain components, or use different components. For example, the battery life prediction device may also include input / output devices, network access devices, buses, etc.

[0128] The processor 21 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor 21 is the control center of the battery life prediction device, connecting all parts of the device via various interfaces and lines.

[0129] The memory 22 can be used to store the computer program and / or modules. The processor 21 implements various functions of the battery life prediction device by running or executing the computer program and / or modules stored in the memory 22 and calling the data stored in the memory 22. The memory 22 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0130] If the module integrated into the battery life prediction device is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0131] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0132] Accordingly, embodiments of the present invention provide a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform steps in the battery life prediction method of the above embodiments, for example... Figure 1 Steps S1 to S7 as described above.

[0133] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for predicting battery life, characterized in that, include: Acquire real-time multidimensional time-series data of the target battery during static and pulse alternating operation processes, and correlate and acquire the intrinsic attribute data of the target battery. Based on the ontological attribute data, the adaptive constraint threshold of the target battery is determined. Based on the adaptive constraint threshold, the real-time multidimensional time series data is sequentially subjected to anomaly identification and missing data imputation processing to obtain the target multidimensional time series data. Feature extraction is performed on the target multidimensional time-series data to obtain a multidimensional feature dataset; wherein, the multidimensional feature dataset includes at least a subset of aging stage identification features; The aging stage of the target battery is determined based on the subset of features identified by the aging stage, and an electrochemical physical model matching the aging stage is determined. A first battery life prediction model is constructed, and the first battery life prediction model and the electrochemical physical model are fused to obtain a second battery life prediction model; Extract the individual mechanism features of the target battery, and optimize the second battery life prediction model based on the individual mechanism features to obtain the battery life prediction model; The multidimensional feature dataset is input into the battery life prediction model, and the battery life prediction result of the target battery is output.

2. The battery life prediction method as described in claim 1, characterized in that, The physical attribute data includes at least one of the following: battery model, production batch, nominal capacity, and initial internal resistance. The step of determining the adaptive constraint threshold of the target battery based on the ontological attribute data includes: Based on the ontological attribute data, an equivalent circuit model and a self-discharge physical model of the target battery are constructed. The theoretically reasonable range of the instantaneous rate of change of voltage and current during pulsed load is calculated based on the equivalent circuit model. Based on the theoretically reasonable range and the target battery's health status, ambient temperature, and historical statistical characteristics of pulse load, an adaptive constraint threshold is dynamically set.

3. The battery life prediction method as described in claim 2, characterized in that, The step of sequentially performing anomaly identification and missing data imputation processing on the real-time multidimensional time series data according to the adaptive constraint threshold to obtain the target multidimensional time series data includes: The real-time multidimensional time-series data is screened based on the adaptive constraint threshold to obtain the first multidimensional time-series data; wherein, the data screening is designed to remove abnormal data points that are not true battery responses; Based on the self-discharge physical model and the historical static data characteristics of the first multidimensional time series data, the missing data of the first multidimensional time series data is filled to obtain the target multidimensional time series data.

4. The battery life prediction method as described in claim 1, characterized in that, The multidimensional feature dataset also includes a physical feature subset, a statistical feature subset, and a deep feature subset; The step of extracting features from the target multidimensional time-series data to obtain a multidimensional feature dataset includes: The nonlinear features of the target multidimensional time-series data are extracted using a convolutional neural network to obtain the deep feature subset. The physical indicator features of battery operation are extracted from the target multidimensional time series data to obtain the physical feature subset; Statistical features of battery operation are extracted from the target multidimensional time-series data to obtain the statistical feature subset; The physical feature subset, the statistical feature subset, the deep feature subset, and the aging stage identification feature subset are integrated to obtain a multidimensional feature dataset.

5. The battery life prediction method as described in claim 4, characterized in that, The aging stage identification feature subset includes at least the rate of change of the voltage recovery time constant after the pulse and the long-term drift acceleration of the average voltage during the static period. Determining the aging stage of the target battery based on the subset of features identified by the aging stage includes: The degradation kinetic characteristic values ​​of the target battery are calculated based on the rate of change of the voltage recovery time constant after the pulse and the long-term drift acceleration of the average voltage during the static period. The degradation dynamics eigenvalues ​​are divided into stages based on a clustering algorithm to obtain the stage division results; The aging stage of the target battery is determined based on the stage division results. The aging stage includes the initial stable period, the intermediate accelerated degradation period, and the final critical failure period.

6. The battery life prediction method as described in claim 1, characterized in that, The extraction of individual mechanistic features of the target battery includes: Obtain a summary of operational data and key physical characteristics of the target battery during its initial operation phase; The statistical summary and key physical features are input into a lightweight auxiliary neural network for inference, and the individual mechanism features of the target battery are output, wherein the individual mechanism features are probability vectors characterizing the battery degradation mechanism type.

7. The battery life prediction method as described in claim 6, characterized in that, The first battery life prediction model is built based on a deep learning temporal network; The optimization of the second battery life prediction model based on the individual mechanism characteristics to obtain the battery life prediction model includes: The fusion weights of the electrochemical physics model and the deep learning temporal network are dynamically allocated according to the attention mechanism. The electrochemical physics model and the deep learning temporal network are fused based on the fusion weights to obtain the second battery lifetime prediction model. Based on the probability vector, the physical parameters of the second battery life prediction model are transferred to obtain the battery life prediction model.

8. A battery life prediction system, characterized in that, include: The acquisition module is used to acquire real-time multi-dimensional time-series data of the target battery during static and pulse alternating operation processes, and to associate and acquire the physical attribute data of the target battery. The preprocessing module is used to determine the adaptive constraint threshold of the target battery based on the ontological attribute data, and to perform anomaly identification and missing data imputation processing on the real-time multidimensional time series data according to the adaptive constraint threshold to obtain the target multidimensional time series data. An extraction module is used to extract features from the target multidimensional time-series data to obtain a multidimensional feature dataset; wherein, the multidimensional feature dataset includes at least a subset of aging stage identification features; An identification module is used to determine the aging stage of the target battery based on a subset of the aging stage identification features, and to determine an electrochemical physical model that matches the aging stage. A construction module is used to construct a first battery life prediction model, and to fuse the first battery life prediction model and the electrochemical physical model to obtain a second battery life prediction model. An optimization module is used to extract the individual mechanism features of the target battery and optimize the second battery life prediction model based on the individual mechanism features to obtain a battery life prediction model. The prediction module is used to input the multidimensional feature dataset into the battery life prediction model and output the battery life prediction result of the target battery.

9. A battery life prediction device, characterized in that, The device includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the battery life prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the battery life prediction method as described in any one of claims 1 to 7.

Images