AI-based power bank and battery health degree prediction system and method thereof

Abstract

This invention proposes an AI-based power bank and its battery health prediction system and method. It belongs to the interdisciplinary field of intelligent power management and artificial intelligence. The method includes: integrating a high-precision sensor array within the power bank to collect multi-dimensional macroscopic and microscopic feature data during the battery charging and discharging process in real time, generating a multi-modal raw dataset of the battery; performing dynamic feature engineering processing based on the multi-modal raw dataset to extract microscopic feature parameters reflecting the electrode aging mechanism, and constructing a battery feature vector integrating electrochemical mechanisms; by integrating high-precision sensors to collect multi-dimensional macroscopic and microscopic feature data, the comprehensiveness and accuracy of battery health status assessment are improved, allowing users to understand the battery condition in greater detail.

Description

Technical Field

[0001] This invention proposes an AI-based power bank and its battery health prediction system and method, belonging to the interdisciplinary field of intelligent power management and artificial intelligence. Background Technology

[0002] With the widespread use of consumer electronic devices, power banks, as important mobile energy supply devices, have their battery health directly affecting device safety and user experience. However, current power bank battery health monitoring technology has many limitations.

[0003] Traditional monitoring methods mostly focus only on macroscopic parameters such as voltage, current, and temperature, and lack effective capture of microscopic characteristics that reflect electrode aging mechanisms, such as relaxation behavior and interface impedance. This results in a single data dimension, making it difficult to comprehensively and accurately reflect the true health status of the battery.

[0004] In terms of prediction models, pure data-driven models, such as LSTM and SVM, will experience a sharp decline in performance when there is insufficient training data or distribution bias, and cannot guarantee that the prediction results conform to the basic laws of electrochemistry, lacking physical consistency.

[0005] At the same time, different users have vastly different usage habits, and different charging modes such as frequent fast charging and slow charging will cause battery degradation paths to be very different, making it difficult for a general model to adapt to these individual differences.

[0006] Furthermore, existing technologies can only provide a binary judgment of "healthy / abnormal," lacking an assessment of predictive uncertainty and failing to offer effective decision-making guidance to users. Therefore, developing a battery health prediction method that can deeply integrate physical models, multi-scale perception, and reliable AI to achieve high accuracy, personalization, and risk quantification mechanisms has become a critical issue that urgently needs to be addressed in the consumer electronics field. Summary of the Invention

[0007] This invention provides an AI-based power bank and its battery health prediction system and method to solve the problems mentioned in the background section above: The present invention proposes an AI-based method for predicting the battery health of power banks, the method comprising: S1. Integrate a high-precision sensor array into the power bank to collect multi-dimensional macroscopic and microscopic feature data during the battery charging and discharging process in real time, and generate a multi-modal raw dataset of the battery; perform dynamic feature engineering processing based on the multi-modal raw dataset to extract microscopic feature parameters that reflect the electrode aging mechanism and construct a battery feature vector that integrates electrochemical mechanism. S2. Construct a lightweight physical information neural network model with embedded electrochemical conservation equations based on battery feature vectors, load pre-trained parameters and adapt to individual differences of power banks through transfer learning technology; use historical data and real-time data to jointly train the PINN model to generate a dynamic health assessment model with physical consistency. S3. Through a dynamic health assessment model, multi-scale feature fusion processing is performed on the real-time collected battery feature vectors to generate battery capacity decay trend data and interface impedance anomaly data; combined with user usage habit data, degradation path correction is performed on the capacity decay trend data to generate personalized health prediction data. S4. Based on personalized health prediction data, perform uncertainty quantification processing, assess prediction confidence interval, and generate battery health uncertainty distribution data; use uncertainty distribution data to perform risk weighting processing on battery capacity degradation trend data to generate a 7-14 day fault warning index. S5. Calculate the weighted health index based on the fault warning index to generate a comprehensive battery health score that integrates physical constraints and data-driven factors; combine the comprehensive health score with uncertainty distribution data to generate multi-level risk warning signals, and output highly reliable battery health prediction results through the local display module of the power bank.

[0008] The present invention proposes a system for implementing the AI-based battery health prediction method for power banks as described above, the system comprising: Vector construction module: A high-precision sensor array is integrated into the power bank to collect multi-dimensional macroscopic and microscopic feature data during the battery charging and discharging process in real time, generating a multi-modal raw dataset of the battery; dynamic feature engineering processing is performed based on the multi-modal raw dataset to extract microscopic feature parameters that reflect the electrode aging mechanism and construct a battery feature vector that integrates electrochemical mechanism. Model generation module: Constructs a lightweight physical information neural network model embedding electrochemical conservation equations based on battery feature vectors, loads pre-trained parameters and adapts them to individual differences in power banks through transfer learning technology; and jointly trains the PINN model using historical and real-time data to generate a dynamic health assessment model with physical consistency. Fusion processing module: Performs multi-scale feature fusion processing on real-time collected battery feature vectors through a dynamic health assessment model to generate battery capacity degradation trend data and interface impedance anomaly data; Combines user usage habit data to correct the degradation path of capacity degradation trend data and generate personalized health prediction data. Fault warning module: Based on personalized health prediction data, uncertainty quantification is performed to assess the prediction confidence interval and generate battery health uncertainty distribution data; the battery capacity degradation trend data is risk-weighted through uncertainty distribution data to generate a 7-14 day fault warning index; The results output module calculates a weighted health index based on the fault warning index, generating a comprehensive battery health score that integrates physical constraints and data-driven approaches. It then combines the comprehensive health score with uncertainty distribution data to generate multi-level risk warning signals and outputs highly reliable battery health prediction results through the power bank's local display module.

[0009] This invention proposes an AI-based power bank, the power bank comprising: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0010] The beneficial effects of this invention are as follows: By integrating high-precision sensors to collect multi-dimensional macroscopic and microscopic feature data, the comprehensiveness and accuracy of battery health status assessment are improved, allowing users to understand the battery condition in greater detail. Utilizing a lightweight PINN model embedded with electrochemical mechanisms, combined with transfer learning, prediction errors caused by insufficient data or distribution shifts are reduced, enhancing the physical consistency of prediction results and making predictions more scientific and reliable. Consideration of user habits enables personalized health predictions, reducing misjudgments caused by individual differences in general models. Uncertainty quantification allows users to clearly understand the reliability of the prediction, avoiding the blindness of making decisions based solely on binary judgments. This method can provide early warnings of battery malfunctions 7-14 days in advance, giving users more time to address the issue, and outputs highly reliable battery health prediction results, guiding users to use power banks appropriately, effectively preventing safety accidents caused by battery health problems, and ensuring user safety and convenience. Attached Figure Description

[0011] Figure 1 This is a diagram illustrating the steps of the method described in this invention; Figure 2 This is a system module diagram of the present invention; Figure 3 This is a flowchart of step S1 as described in this invention; Figure 4 This is a flowchart of step S13 of the present invention. Detailed Implementation

[0012] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0013] One embodiment of the present invention, such as Figure 1 As shown, an AI-based method for predicting the battery health of a power bank includes: S1. Integrate a high-precision sensor array within the power bank to collect multi-dimensional macroscopic and microscopic feature data during the battery charging and discharging process in real time. The multi-dimensional macroscopic and microscopic feature data includes voltage, current, temperature, internal resistance, and relaxation behavior, generating a multi-modal raw dataset of the battery. Based on the multi-modal raw dataset, perform dynamic feature engineering processing to extract microscopic feature parameters that reflect the electrode aging mechanism and construct a battery feature vector that integrates electrochemical mechanisms. S2. Construct a lightweight physical information neural network (PINN) model with embedded electrochemical conservation equations based on battery feature vectors. Load pre-trained parameters and adapt them to individual differences in power banks through transfer learning technology. Jointly train the PINN model using historical and real-time data to generate a dynamic health assessment model with physical consistency. S3. Through a dynamic health assessment model, multi-scale feature fusion processing is performed on the real-time collected battery feature vectors to generate battery capacity decay trend data and interface impedance anomaly data; combined with user usage habit data (such as charging frequency, fast charging / slow charging mode), the capacity decay trend data is corrected for degradation path to generate personalized health prediction data. S4. Based on the personalized health prediction data, uncertainty quantification is performed, and the Monte Carlo dropout method is used to evaluate the prediction confidence interval to generate battery health uncertainty distribution data. The uncertainty distribution data is then used to perform risk weighting on the battery capacity degradation trend data to generate a 7-14 day fault warning index. S5. Calculate the weighted health index based on the fault warning index to generate a comprehensive battery health score that integrates physical constraints and data-driven factors; combine the comprehensive health score with uncertainty distribution data to generate multi-level risk warning signals, and output highly reliable battery health prediction results through the local display module of the power bank.

[0014] The working principle and effects of the above technical solution are as follows: Through multi-dimensional feature data collection and dynamic feature engineering processing, the accuracy of battery health prediction can be effectively improved, avoiding prediction bias caused by single feature analysis and reducing usage inconvenience caused by misjudgment of battery status. A lightweight model integrating electrochemical mechanisms, combined with transfer learning, enhances the adaptability to individual differences in different power banks, reducing model training costs and operating power consumption. This ensures both the physical consistency of prediction results and adaptability to the limited hardware resources of power banks. Predictive data is corrected based on user habits, reducing the interference of different usage scenarios on the results and making health assessments more aligned with actual usage needs. Uncertainty quantification and risk weighting processing can generate fault warning indices in advance, avoiding charging interruptions or safety hazards caused by sudden battery failure and reducing the shortened lifespan caused by excessive battery wear. Multi-level risk warnings and local display output improve the timeliness of battery status feedback, enhance users' control over battery health, prevent inconvenience in charging during travel due to the inability to predict status, and balance prediction reliability with ease of use.

[0015] One embodiment of the present invention, such as Figure 3 As shown, S1 includes: S11. Integrate a high-precision sensor array within the power bank to collect multi-dimensional macroscopic and microscopic feature data during the battery charging and discharging process in real time, generating a multi-modal raw dataset of the battery. S12. Perform noise filtering and outlier removal on the original battery multimodal dataset to generate a clean battery multimodal dataset. S13. Conduct dynamic feature engineering processing based on the clean battery multimodal dataset to mine hidden temporal correlations and nonlinear features in the data; S14. Extract microscopic feature parameters reflecting the electrode aging mechanism from the processed feature data to form a set of microscopic feature parameters; S15. By integrating the electrochemical mechanism, the microscopic feature parameter set is dimensionally integrated and feature-selected to construct a battery feature vector based on the integrated electrochemical mechanism.

[0016] The working principle and effects of the above technical solution are as follows: A high-precision sensor array collects multi-dimensional feature data in real time, combined with noise filtering and outlier removal operations, effectively improving the purity of the original battery data, avoiding interference from impurity data in subsequent feature processing, and reducing feature extraction deviations caused by data distortion. Dynamic feature engineering deeply mines hidden temporal correlations and nonlinear features, enhancing the comprehensiveness of feature coverage and reducing the one-sidedness caused by single-dimensional analysis. Microscopic parameters reflecting the electrode aging mechanism are extracted, focusing on core aging factors, avoiding irrelevant features from occupying processing resources, and making features more consistent with the actual aging patterns of the battery. Dimensional integration and screening are carried out by incorporating electrochemical mechanisms, which not only simplifies feature dimensions and reduces subsequent computational pressure, but also ensures the physical logic of feature vectors, providing a high-quality foundation for subsequent model construction. The entire process progressively optimizes data quality, avoiding subsequent evaluation inaccuracies caused by feature redundancy or missing features, while strengthening the correlation between features and battery aging, improving the reliability of subsequent health predictions, and reducing interference from invalid features on model training.

[0017] One embodiment of the present invention, such as Figure 4 As shown, S13 includes: Temporal dimension alignment processing is performed on the clean battery multimodal dataset to unify the acquisition time benchmark of each feature data and generate a temporally aligned feature dataset. Based on the time-aligned feature dataset, feature derivation processing is performed to generate derived features, which include the change in charge and discharge rate and the feature difference sequence, forming a multi-dimensional derived feature set. Nonlinear feature transformation is performed on the multi-dimensional derived feature set to weaken the interference of linear redundancy in the data and generate a nonlinear transformed feature set. Temporal correlation analysis is performed based on nonlinear transformation feature sets to capture the intrinsic correlation between features at different times and generate temporal correlation feature sets. The temporal correlation feature set is subjected to feature aggregation processing to integrate the temporal information of similar features and generate an optimized dynamic feature set, which provides support for the subsequent extraction of micro-feature parameters.

[0018] The working principle and effects of the above technical solution are as follows: First, aligning the time series dimension unifies the time base for feature acquisition, effectively improving the consistency of multimodal data, avoiding inaccurate feature correlation analysis due to time misalignment, and reducing unnecessary processing overhead. Second, feature derivation generates dimensions such as charge / discharge rate changes and feature difference sequences, enriching feature coverage, filling information gaps in the original data, and enhancing the ability to capture changes in battery state. Third, nonlinear feature transformation weakens linear redundancy interference, reduces the obscuring of core features by redundant data, and makes features more consistent with the nonlinear laws of battery operation. Fourth, time series correlation analysis delves into the inherent correlation between features at different times, avoiding the one-sidedness caused by isolated analysis of single-time data and improving the temporal logic of features. Fifth, feature aggregation integrates similar time series information, which not only simplifies the data volume and reduces the pressure on subsequent processing but also retains key time series features, laying a solid foundation for the extraction of micro-feature parameters. Sixth, the entire process optimizes feature quality layer by layer, avoiding invalid features from interfering with subsequent operations, while strengthening the correlation between features and battery state, further improving the accuracy of subsequent health assessments.

[0019] In one embodiment of the present invention, S15 includes: By integrating electrochemical mechanisms to adapt the set of microscopic characteristic parameters, and linking electrode reaction patterns with parameter characteristics, a set of mechanism-adapted parameters is generated. The mechanism adaptation parameter set is subjected to dimension merging processing, merging similar feature dimensions, eliminating parameter redundancy, and generating a dimension merged parameter set; Based on the electrochemical mechanism, the dimensional merging parameter set is subjected to feature purification processing to remove parameters that are not related to battery aging, and a mechanism-guided purified parameter set is generated. The mechanism-guided purified parameter set is numerically balanced to unify the differences in parameter magnitudes and generate a balanced feature parameter set. The core information of the balanced feature parameter set is integrated to construct a battery feature vector that incorporates the electrochemical mechanism, providing a foundation for subsequent model construction.

[0020] The working principle and effects of the above technical solution are as follows: It integrates electrochemical mechanisms for mechanism adaptation, correlates electrode reaction patterns with parameter characteristics, making microscopic parameters more closely match the essential mechanism of battery aging, avoiding parameter decoupling from actual electrochemical processes, and improving parameter effectiveness. It merges similar feature dimensions to eliminate redundant parameters, reducing the computational load of subsequent data processing and minimizing ineffective resource consumption. Based on electrochemical mechanisms, it performs feature purification, eliminating parameters unrelated to battery aging, enhancing the specificity of feature vectors, avoiding irrelevant information from interfering with subsequent model construction, and making core features more prominent. It unifies parameter magnitude differences through numerical balancing, avoiding the weakening of core features due to magnitude imbalance, improving the stability of feature vectors, and providing balanced input for subsequent model training. It integrates core information to construct feature vectors, retaining key parameters strongly correlated with battery aging while incorporating electrochemical mechanisms to ensure physical rationality, laying a solid foundation for subsequent models. The entire process optimizes parameter quality layer by layer, avoiding insufficient model accuracy caused by feature redundancy, imbalance, or deviation from the mechanism, further improving the reliability of battery health prediction.

[0021] In one embodiment of the present invention, S2 includes: S21. Based on the dimension and distribution characteristics of battery feature vectors, construct a lightweight physical information neural network model architecture that embeds electrochemical conservation equations. S22. Load pre-trained parameters using transfer learning techniques to initialize the weights and biases of the physical information neural network model; S23. Based on the individual hardware differences of the power bank battery, the model parameters are adaptively adjusted to complete the adaptation of the model to the individual power bank. S24. Integrate historical charging and discharging data of power bank batteries with real-time collected data to construct a joint training sample set for the model; S25. Use the joint training sample set to perform multiple rounds of iterative training on the physical information neural network model to optimize the model's feature extraction and prediction capabilities; based on the trained model, generate a dynamic health assessment model with physical consistency.

[0022] The working principle and effects of the above technical solution are as follows: Embedding electrochemical conservation equations to construct a lightweight model architecture balances the physical logic of model predictions with the ability to control model size to adapt to power bank hardware conditions, reducing power consumption and resource consumption during operation, and avoiding operational lag or hardware overload caused by an overly heavy model. Transfer learning loads pre-trained parameters to initialize weights and biases, reducing the number of iterations for training from scratch, shortening the model training cycle, and avoiding slow training convergence caused by random parameter initialization. Adaptive parameter adjustment based on individual battery differences enhances the model's adaptability to different power banks, avoiding prediction bias caused by general models ignoring individual characteristics, and making the evaluation results more consistent with the actual situation of a single device. Integrating historical and real-time data to construct a joint training sample set enriches the coverage of training data scenarios, improves model generalization ability, and reduces the risk of overfitting from a single data source. Multi-round iterative training optimizes feature extraction and prediction capabilities, improves model evaluation accuracy, and generates a dynamically consistent health assessment model, avoiding a disconnect between model predictions and actual battery aging patterns. The entire process balances training efficiency and model performance, ensuring both prediction reliability and adaptability to the limitations of power bank hardware, thus laying a solid foundation for subsequent health assessments.

[0023] In one embodiment of the present invention, step S3 includes: S31. Input the real-time collected battery feature vector into the dynamic health assessment model and start the model's multi-scale feature fusion processing flow. S32. Extract battery capacity decay-related features at different time scales through the multi-scale feature fusion module of the model; generate battery capacity decay trend data based on the extracted multi-scale features. S33. Simultaneously extract battery interface impedance related features to generate interface impedance anomaly data; collect user usage habit data such as charging frequency and fast / slow charging mode during the use of power banks to form a user usage habit dataset. S34. Combining user usage habit datasets, the degradation path of battery capacity decay trend data is corrected to eliminate prediction bias caused by usage habits; based on the corrected capacity decay trend data, personalized health prediction data is generated.

[0024] The working principle and effects of the above technical solution are as follows: Multi-scale feature fusion extracts capacity decay characteristics at different time scales, capturing both short-term charging and discharging fluctuations and long-term aging trends, effectively improving the comprehensiveness of capacity decay trend data and avoiding trend misjudgments caused by single-time-scale analysis. Simultaneously, it extracts interface impedance anomaly data, enhancing the ability to detect potential battery faults, reducing the problem of early impedance anomalies being overlooked, and preventing small faults from gradually worsening and affecting battery life. It collects user usage habit data such as charging frequency and fast / slow charging modes, and combines this data to correct capacity decay trends, eliminating prediction biases caused by differences in usage habits, making health predictions more aligned with individual usage scenarios, and avoiding discrepancies between general prediction results and actual conditions. The corrected decay trend data generates personalized health prediction results, further improving prediction accuracy and helping users accurately understand the status of their power bank batteries. The entire process considers both the characteristics of the battery itself and the user's usage scenario, ensuring the comprehensiveness of trend analysis while strengthening prediction reliability through personalized correction, preventing inaccurate assessments due to ignoring usage habits or local anomalies, and providing high-quality data support for subsequent risk warnings and health scoring.

[0025] In one embodiment of the present invention, S32 includes: The input battery feature vector is split into three scale intervals: short-term charge-discharge, medium-term cycle, and long-term aging, to generate a multi-scale feature subset. Targeted feature enhancement processing is performed on feature subsets at each scale to amplify feature signals closely related to capacity decay and generate enhanced multi-scale feature sets. The enhanced multi-scale feature set is subjected to cross-scale correlation fusion processing to eliminate redundant interference between features of different scales and generate a cross-scale fused feature set. Core decay correlation features are extracted from the cross-scale fusion feature set, invalid feature information is filtered out, and a core feature set of capacity decay is generated. The core feature set of capacity decay is subjected to time-series trend fitting processing to capture the changing pattern of features over time and generate battery capacity decay trend data.

[0026] The working principle and effects of the above technical solution are as follows: Battery feature vectors are split according to three time scales, accurately dividing short-term charge / discharge, mid-term cycling, and long-term aging intervals, generating multi-scale feature subsets. This not only covers battery state changes across different cycles but also avoids missing key information in single-scale analysis, reducing the bias in trend judgment. Targeted feature enhancement amplifies signals closely related to capacity decay, weakens irrelevant interference, enhances the recognizability of core features, avoids weakly correlated features masking decay patterns, and improves feature effectiveness. Cross-scale correlation fusion eliminates redundant interference from features at different scales, integrates core information from various dimensions, making features more coherent and logical, reducing the impact of redundant data on subsequent processing, and strengthening the intrinsic correlation between states at different cycles. Extracting core decay-related features filters out invalid information, simplifies feature size, avoids invalid data consuming computational resources, and improves subsequent processing efficiency. Time-series trend fitting captures the changing patterns of features over time, making the generated capacity decay trend data more consistent with the actual aging process, avoiding misjudgments of health status due to trend fitting bias. The entire process optimizes feature quality at each stage, ensuring the accuracy and comprehensiveness of the decay trend data, providing a reliable basis for subsequent degradation path correction, and further enhancing the credibility of personalized health prediction.

[0027] In one embodiment of the present invention, step S4 includes: S41. Perform uncertainty quantification on personalized health prediction data to identify random and systematic errors in the prediction process; S42. Using the Monte Carlo dropout method, perform prediction confidence interval assessment on the personalized health prediction data to quantify the reliability of the prediction results; based on the confidence interval assessment results, generate battery health uncertainty distribution data; S43. Using the uncertainty distribution data of battery health as weight, risk-weighted processing is applied to the battery capacity decay trend data to strengthen the prediction weight of high-risk periods. S44. Based on the capacity decay trend data after risk-weighted processing, generate a 7- to 14-day fault warning index.

[0028] The working principle and effects of the above technical solution are as follows: Uncertainty quantification is performed on personalized health prediction data to accurately identify random and systematic errors, avoiding the distortion of prediction results due to hidden errors and reducing the risk omission caused by misjudgment of reliability. The Monte Carlo dropout method is used to assess confidence intervals, quantify the reliability of prediction results, and generate uncertainty distribution data. This provides a clear basis for prediction credibility while avoiding decision-making bias caused by blindly relying on a single prediction value. Risk weighting is applied using the uncertainty distribution data as weights, strengthening the prediction weight for high-risk periods. This allows capacity degradation trend analysis to focus more on key nodes, reducing the interference of low-risk information on core judgments and preventing high-risk hazards from being masked. A 7- to 14-day fault warning index is generated based on the weighted degradation trend, capturing potential battery fault signals in advance to avoid charging interruptions or equipment damage caused by sudden failures, providing users with sufficient maintenance time. The entire process balances prediction accuracy and risk control, improving prediction rigor through error quantification while enhancing practicality through risk weighting and early warning, reducing the inconvenience caused by the inability to predict faults, and further strengthening the reference value and safety assurance capabilities of battery health prediction.

[0029] In one embodiment of the present invention, step S5 includes: S51. Based on the 7-14 day fault warning index, a weighted health index is calculated, and a comprehensive index is formed by integrating the warning intensity and the predicted trend. S52. Based on the weighted health index, generate a comprehensive battery health score that integrates physical constraints and data-driven factors; S53. Combining the comprehensive battery health score with uncertainty distribution data, construct a multi-level risk warning judgment logic and divide the warning thresholds for different risk levels. S54. Based on the early warning judgment logic, generate multi-level risk warning signals corresponding to the risk level; transmit the multi-level risk warning signals and the battery health comprehensive score to the local display module of the power bank to complete the local data transmission; S55. Through the local display module of the power bank, output a highly reliable battery health prediction result to complete the battery health prediction task.

[0030] The working principle and effects of the above technical solution are as follows: A weighted health index is calculated by combining a fault warning index with the overall health index. This index integrates warning intensity and predicted trends to form a comprehensive index, taking into account both short-term risk signals and long-term degradation patterns, thus improving the comprehensiveness of health assessment and avoiding biased results caused by single-dimensional scoring. A score that integrates physical constraints and data-driven approaches is generated based on the comprehensive index, enhancing the scientific rigor and rationality of the results and avoiding scoring biases that deviate from the actual operating mechanism of the battery, making the health status assessment more closely reflect reality. A multi-level warning logic is constructed by combining the comprehensive score with uncertainty distribution data, dividing different risk thresholds to enhance the targeting and hierarchy of warnings, avoiding ambiguous warning signals that prevent users from accurately judging the risk level. Multi-level warning signals are generated and transmitted locally to the display module, reducing data transmission latency and preventing warning failures due to remote transmission delays or interruptions, ensuring the stability of information transmission. Highly reliable results are output through the local display module, allowing users to intuitively grasp the battery health status, avoiding usage risks caused by the inability to obtain accurate information in a timely manner, while simplifying the user viewing process and improving ease of use. The entire process of closed-loop optimization, evaluation, and display not only ensures the credibility of the prediction results but also takes into account the user experience, reducing potential battery usage risks caused by information asymmetry or untimely display.

[0031] In one embodiment of the present invention, S53 includes: The battery health comprehensive score and uncertainty distribution data are synchronously integrated and processed to combine the core information of the two types of data and generate a joint risk assessment dataset. The risk assessment joint dataset is decomposed into risk dimensions, splitting it into two core dimensions: health score range and uncertainty fluctuation range, to generate a risk dimension decomposed dataset. The data for each risk dimension are hierarchically divided to initially define the basic range of low, medium and high risk, and generate a preliminary risk level classification table. Cross-validation is performed on the preliminary risk level classification table to eliminate threshold conflicts between different dimensions and generate a set of validated risk thresholds. By integrating and verifying the risk threshold set and risk dimension logic, a multi-level risk early warning judgment framework is built, forming a complete multi-level risk early warning judgment logic.

[0032] The working principle and effects of the above technical solution are as follows: It synchronously integrates the comprehensive battery health score and uncertainty distribution data, fusing these two types of core information to generate a joint dataset. This avoids the biased assessment caused by fragmented analysis of the two types of data, improving the completeness of risk judgment. It separates the health score range from the core dimension of uncertainty fluctuation amplitude, allowing risk assessment to focus on key indicators, enhancing dimensional specificity, reducing interference from irrelevant dimensions, and avoiding ambiguity in threshold division caused by mixing them up. It defines low, medium, and high-risk basic ranges for each dimension's data hierarchy, building a solid foundation for the early warning logic and avoiding judgment confusion caused by unclear risk levels. Cross-validation of the preliminary partition table eliminates threshold conflicts between dimensions, corrects contradictory nodes, improves the accuracy of risk thresholds, avoids distortion of early warning signals caused by threshold conflicts, and reduces the risk of misjudgment. Integrating and validating the threshold set and dimensional logic to build a judgment framework forms a complete early warning logic, ensuring both clear hierarchical multi-level early warnings and balancing data correlation and judgment rigor. The entire process optimizes the risk assessment criteria at each level to avoid early warning failures caused by improper data integration, threshold conflicts, or ambiguity in dimensions, thereby enhancing the reliability of multi-level risk early warnings and providing solid support for the subsequent generation of accurate early warning signals.

[0033] One embodiment of the present invention, such as Figure 2 As shown, a system for implementing the AI-based battery health prediction method for power banks as described above is provided, the system comprising: Vector construction module: A high-precision sensor array is integrated into the power bank to collect multi-dimensional macroscopic and microscopic feature data in real time during the battery charging and discharging process. The multi-dimensional macroscopic and microscopic feature data includes voltage, current, temperature, internal resistance and relaxation behavior, generating a battery multimodal raw dataset; dynamic feature engineering processing is performed based on the multimodal raw dataset to extract microscopic feature parameters reflecting the electrode aging mechanism and construct a battery feature vector that integrates electrochemical mechanism. Model generation module: Constructs a lightweight physical information neural network (PINN) model with embedded electrochemical conservation equations based on battery feature vectors, loads pre-trained parameters and adapts them to individual differences in power banks through transfer learning technology; and jointly trains the PINN model using historical and real-time data to generate a dynamic health assessment model with physical consistency. Fusion processing module: Through a dynamic health assessment model, the real-time collected battery feature vectors are processed by multi-scale feature fusion to generate battery capacity decay trend data and interface impedance anomaly data; combined with user usage habit data (such as charging frequency, fast charging / slow charging mode), the capacity decay trend data is corrected for degradation path to generate personalized health prediction data. Fault warning module: Based on personalized health prediction data, uncertainty quantification is performed, and the Monte Carlo dropout method is used to evaluate the prediction confidence interval to generate battery health uncertainty distribution data; the battery capacity degradation trend data is risk-weighted through uncertainty distribution data to generate a 7-14 day fault warning index; The results output module calculates a weighted health index based on the fault warning index, generating a comprehensive battery health score that integrates physical constraints and data-driven approaches. It then combines the comprehensive health score with uncertainty distribution data to generate multi-level risk warning signals and outputs highly reliable battery health prediction results through the power bank's local display module.

[0034] According to one embodiment of the present invention, an AI-based power bank includes: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors are made to implement the method described in any one of the above.

[0035] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An AI-based method for predicting the battery health of power banks, characterized in that, The method includes: S1. Integrate a high-precision sensor array into the power bank to collect multi-dimensional macroscopic and microscopic feature data during the battery charging and discharging process in real time, and generate a multi-modal raw dataset of the battery; perform dynamic feature engineering processing based on the multi-modal raw dataset to extract microscopic feature parameters that reflect the electrode aging mechanism and construct a battery feature vector that integrates electrochemical mechanism. S2. Construct a lightweight physical information neural network model with embedded electrochemical conservation equations based on battery feature vectors, load pre-trained parameters and adapt to individual differences of power banks through transfer learning technology; use historical data and real-time data to jointly train the PINN model to generate a dynamic health assessment model with physical consistency. S3. Through a dynamic health assessment model, multi-scale feature fusion processing is performed on the real-time collected battery feature vectors to generate battery capacity decay trend data and interface impedance anomaly data; combined with user usage habit data, degradation path correction is performed on the capacity decay trend data to generate personalized health prediction data. S4. Based on personalized health prediction data, perform uncertainty quantification processing, assess prediction confidence interval, and generate battery health uncertainty distribution data; use uncertainty distribution data to perform risk weighting processing on battery capacity degradation trend data to generate a 7-14 day fault warning index. S5. Calculate the weighted health index based on the fault warning index to generate a comprehensive battery health score that integrates physical constraints and data-driven factors; combine the comprehensive health score with uncertainty distribution data to generate multi-level risk warning signals, and output highly reliable battery health prediction results through the local display module of the power bank.

2. The AI-based battery health prediction method for power banks according to claim 1, characterized in that, S1 includes: S11. Integrate a high-precision sensor array within the power bank to collect multi-dimensional macroscopic and microscopic feature data during the battery charging and discharging process in real time, generating a multi-modal raw dataset of the battery. S12. Perform noise filtering and outlier removal on the original battery multimodal dataset to generate a clean battery multimodal dataset. S13. Conduct dynamic feature engineering processing based on the clean battery multimodal dataset to mine hidden temporal correlations and nonlinear features in the data; S14. Extract microscopic feature parameters reflecting the electrode aging mechanism from the processed feature data to form a set of microscopic feature parameters; S15. By integrating the electrochemical mechanism, the microscopic feature parameter set is dimensionally integrated and feature-selected to construct a battery feature vector based on the integrated electrochemical mechanism.

3. The AI-based battery health prediction method for power banks according to claim 2, characterized in that, S13 includes: Temporal dimension alignment processing is performed on the clean battery multimodal dataset to unify the acquisition time benchmark of each feature data and generate a temporally aligned feature dataset. Based on the time-aligned feature dataset, feature derivation processing is performed to generate derived features, forming a multi-dimensional derived feature set; Nonlinear feature transformation is performed on the multi-dimensional derived feature set to weaken the interference of linear redundancy in the data and generate a nonlinear transformed feature set. Temporal correlation analysis is performed based on nonlinear transformation feature sets to capture the intrinsic correlation between features at different times and generate temporal correlation feature sets. The temporal correlation feature set is subjected to feature aggregation processing to integrate the temporal information of similar features and generate an optimized dynamic feature set.

4. The AI-based battery health prediction method for power banks according to claim 1, characterized in that, The S2 includes: S21. Based on the dimension and distribution characteristics of battery feature vectors, construct a lightweight physical information neural network model architecture that embeds electrochemical conservation equations. S22. Load pre-trained parameters using transfer learning techniques to initialize the weights and biases of the physical information neural network model; S23. Based on the individual hardware differences of the power bank battery, the model parameters are adaptively adjusted to complete the adaptation of the model to the individual power bank. S24. Integrate historical charging and discharging data of power bank batteries with real-time collected data to construct a joint training sample set for the model; S25. Use the joint training sample set to perform multiple rounds of iterative training on the physical information neural network model to optimize the model's feature extraction and prediction capabilities; based on the trained model, generate a dynamic health assessment model with physical consistency.

5. The AI-based battery health prediction method for power banks according to claim 1, characterized in that, The S3 includes: S31. Input the real-time collected battery feature vector into the dynamic health assessment model and start the model's multi-scale feature fusion processing flow. S32. Extract battery capacity decay-related features at different time scales through the multi-scale feature fusion module of the model; generate battery capacity decay trend data based on the extracted multi-scale features. S33. Simultaneously extract battery interface impedance related features to generate interface impedance anomaly data; collect user usage habit data such as charging frequency and fast / slow charging mode during the use of power banks to form a user usage habit dataset. S34. Combining user usage habit datasets, the degradation path of battery capacity decay trend data is corrected to eliminate prediction bias caused by usage habits; based on the corrected capacity decay trend data, personalized health prediction data is generated.

6. The AI-based battery health prediction method for power banks according to claim 5, characterized in that, S32 includes: The input battery feature vector is split into three scale intervals: short-term charge-discharge, medium-term cycle, and long-term aging, to generate a multi-scale feature subset. Targeted feature enhancement processing is performed on feature subsets at each scale to amplify feature signals closely related to capacity decay and generate enhanced multi-scale feature sets. The enhanced multi-scale feature set is subjected to cross-scale correlation fusion processing to eliminate redundant interference between features of different scales and generate a cross-scale fused feature set. Core decay correlation features are extracted from the cross-scale fusion feature set, invalid feature information is filtered out, and a core feature set of capacity decay is generated. The core feature set of capacity decay is subjected to time-series trend fitting processing to capture the changing pattern of features over time and generate battery capacity decay trend data.

7. The AI-based battery health prediction method for power banks according to claim 1, characterized in that, The S4 includes: S41. Perform uncertainty quantification on personalized health prediction data to identify random and systematic errors in the prediction process; S42. Using the Monte Carlo dropout method, perform prediction confidence interval assessment on the personalized health prediction data to quantify the reliability of the prediction results; based on the confidence interval assessment results, generate battery health uncertainty distribution data; S43. Using the uncertainty distribution data of battery health as weight, risk-weighted processing is applied to the battery capacity decay trend data to strengthen the prediction weight of high-risk periods. S44. Based on the capacity decay trend data after risk-weighted processing, generate a 7- to 14-day fault warning index.

8. The AI-based battery health prediction method for power banks according to claim 1, characterized in that, The S5 includes: S51. Based on the 7-14 day fault warning index, a weighted health index is calculated, and a comprehensive index is formed by integrating the warning intensity and the predicted trend. S52. Based on the weighted health index, generate a comprehensive battery health score that integrates physical constraints and data-driven factors; S53. Combining the comprehensive battery health score with uncertainty distribution data, construct a multi-level risk warning judgment logic and divide the warning thresholds for different risk levels. S54. Based on the early warning judgment logic, generate multi-level risk warning signals corresponding to the risk level; transmit the multi-level risk warning signals and the battery health comprehensive score to the local display module of the power bank to complete the local data transmission; S55. Through the local display module of the power bank, output a highly reliable battery health prediction result to complete the battery health prediction task.

9. A system for implementing the AI-based battery health prediction method for power banks as described in claim 1, characterized in that, The system includes: Vector construction module: A high-precision sensor array is integrated into the power bank to collect multi-dimensional macroscopic and microscopic feature data during the battery charging and discharging process in real time, generating a multi-modal raw dataset of the battery; dynamic feature engineering processing is performed based on the multi-modal raw dataset to extract microscopic feature parameters that reflect the electrode aging mechanism and construct a battery feature vector that integrates electrochemical mechanism. Model generation module: Constructs a lightweight physical information neural network model embedding electrochemical conservation equations based on battery feature vectors, loads pre-trained parameters and adapts them to individual differences in power banks through transfer learning technology; and jointly trains the PINN model using historical and real-time data to generate a dynamic health assessment model with physical consistency. Fusion processing module: Performs multi-scale feature fusion processing on real-time collected battery feature vectors through a dynamic health assessment model to generate battery capacity degradation trend data and interface impedance anomaly data; Combines user usage habit data to correct the degradation path of capacity degradation trend data and generate personalized health prediction data. Fault warning module: Based on personalized health prediction data, uncertainty quantification is performed to assess the prediction confidence interval and generate battery health uncertainty distribution data; the battery capacity degradation trend data is risk-weighted through uncertainty distribution data to generate a 7-14 day fault warning index; The results output module calculates a weighted health index based on the fault warning index, generating a comprehensive battery health score that integrates physical constraints and data-driven approaches. It then combines the comprehensive health score with uncertainty distribution data to generate multi-level risk warning signals and outputs highly reliable battery health prediction results through the power bank's local display module.

10. An AI-based power bank, characterized in that, The power bank includes: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 8.

Images