Method for constructing performance prediction model of nickel-hydrogen battery based on machine learning

Abstract

The application discloses a nickel-hydrogen battery performance prediction model construction method based on machine learning, aligns electrochemical test data, environment monitoring data, operation log data and material characterization data through a unified time reference, forms a synchronous original data stream, generates a structured data matrix through missing value filling, abnormal correction and dynamic range compression, improves data quality and comparability, extracts voltage platform duration, temperature rise rate and principal component score and gives semantic labels, realizes physical interpretable feature construction, fuses multiple source features by using a latent space projection network, generates a unified latent representation, combines a recurrent neural network, an attention mechanism and a multi-scale fusion strategy, mines historical dependence, generates a comprehensive state representation, synchronously outputs performance prediction values and deduces degradation trajectories, cooperates independent verification and dynamic calibration, and ensures prediction stability. Multi-source information is deeply fused, and the accuracy and adaptability of the nickel-hydrogen battery performance degradation trend prediction are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of battery technology, specifically relating to a method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning. Background Technology

[0002] In practical applications of nickel-metal hydride batteries, accurately predicting their performance degradation trends is crucial for ensuring safe system operation and optimizing maintenance strategies. Existing methods typically rely on single-type data, such as modeling based solely on charge-discharge curves or temperature changes, which fails to comprehensively reflect the complex physicochemical evolution processes within the battery. Some methods attempting to incorporate multi-source information, while collecting data from electrochemical, environmental, and operational logs, often employ simple splicing or weighted fusion methods, failing to adequately consider differences in sampling frequencies, inconsistencies in units, and semantic hierarchical separation among different data types. This results in ineffective synergy of heterogeneous information, limiting the improvement of prediction accuracy. Especially during long-term use, the battery aging path exhibits nonlinear cumulative characteristics, making it difficult for traditional linear extrapolation or empirical formulas to capture its dynamic evolution patterns, easily leading to trend misjudgments.

[0003] The above processing method exposes a prominent technical problem: how to construct a predictive system that can deeply integrate various types of information and accurately characterize the performance degradation trend of nickel-metal hydride batteries under the condition of multi-source heterogeneous data with asynchronous time, different dimensions, and semantic layering. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning. This method can effectively integrate multi-source heterogeneous data from nickel-metal hydride batteries, overcome interference caused by time asynchrony and dimensional differences, and significantly improve the accuracy of characterizing performance degradation trends. In particular, it exhibits stronger stability and adaptability in trend direction identification and long-term trajectory extrapolation, thereby providing a reliable basis for battery health management and solving the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning, comprising the following steps:

[0006] Collect electrochemical test data, environmental monitoring data, operation log data and material characterization data, and map multi-source heterogeneous data to a unified time base to form a time-aligned raw data stream;

[0007] Missing value imputation and outlier correction are performed based on the original data stream. Numerical variables are normalized and their dynamic range is compressed to generate a structured data matrix.

[0008] The voltage plateau duration of the electrochemical sequence, the temperature rise rate of the thermodynamic data, and the principal component score of the material image are extracted from the structured data matrix to form a feature set with semantic labels.

[0009] The feature set is input into the latent spatial projection network, and the nonlinear mapping process is optimized by self-reconstruction loss and contrastive regularization term to generate a unified latent representation.

[0010] By feeding a unified latent representation into a recurrent neural network structure and combining an attention mechanism with a multi-scale temporal information fusion strategy, a comprehensive state representation containing historical dependencies is generated.

[0011] Based on the comprehensive state representation, the predicted values ​​of remaining capacity, internal resistance growth trend and health index are output synchronously, and the future degradation trajectory is generated recursively.

[0012] The error and trend consistency of the predicted values ​​are evaluated using an independent validation set, and calibration is performed on the new battery model to optimize inference efficiency and establish a dynamic update mechanism.

[0013] Preferably, the electrochemical test data includes constant current charge-discharge records and internal resistance change sequences; environmental monitoring data includes temperature sensor readings and humidity monitoring data; operation log data includes cycle count logs and package deformation images; and material characterization data includes X-ray diffraction patterns. The timestamps of all data are linearly mapped with the first test time as the origin.

[0014] Preferably, the missing value imputation adopts the sliding window local mean interpolation method, the outlier correction adopts the three standard deviation criterion to identify and replace outliers, the dimension normalization maps the variables to the zero-one interval, and the dynamic range compression applies a logarithmic transformation to the long-tailed distribution variables.

[0015] Preferably, the voltage plateau duration is calculated by the start and end times of the charge-discharge cycle, the temperature rise rate is obtained by temperature difference processing, the principal component score is extracted by material image vectorization and covariance matrix decomposition, and all features are aligned by time and concatenated into a feature set.

[0016] Preferably, the latent spatial projection network contains three fully connected layers. The self-reconstruction loss recovers the original features through the decoder network, and the contrast regularization term calculates the distribution difference of positive and negative sample pairs through cosine similarity. The optimized unified latent representation retains the discriminative structure of multi-source information.

[0017] Preferably, the recurrent neural network structure is a long short-term memory network, the attention mechanism focuses on key historical segments through weighted context vectors, and multi-scale temporal information is achieved by splicing hidden states with different sampling granularities. The final output comprehensive state representation contains complete evolutionary path information.

[0018] Preferably, the prediction of remaining capacity is generated synchronously through a multi-task output structure to generate the internal resistance growth trend and health index, and the future degradation trajectory is generated by recursively predicting the state at future time steps. The prediction results are accompanied by confidence intervals and ranking of influencing factors.

[0019] Preferably, it also includes verifying the model's adaptability, specifically including hierarchical cross-validation and transferability testing, and optimizing inference efficiency, including weight quantization, low-rank decomposition, and locality-sensitive hash pre-screening.

[0020] Preferably, the dynamic update mechanism triggers retraining by periodically collecting new data, the version management mechanism maintains system stability through difference analysis and rollback strategies, and the update process adjusts the parameters in a limited number of rounds starting from the current parameters.

[0021] Preferably, it also includes: developing a visualization platform to display an overview of the health status, historical evolution curves and future prediction paths, with confidence ranges superimposed; based on the content displayed on the visualization platform, supporting users to modify future usage assumptions and observe changes in the prediction trajectory in real time; and setting input integrity checks based on user interaction behavior, pausing prediction and issuing alerts for data with excessively high missing rates or physically unreasonable data.

[0022] Technical effects and advantages of the present invention: The method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning proposed in this invention has the following advantages compared with the prior art:

[0023] This invention aligns electrochemical test data, environmental monitoring data, operational log data, and material characterization data using a unified time reference to form a synchronous raw data stream, thus resolving the time misalignment problem of multi-source heterogeneous data. A structured data matrix is ​​generated through missing value imputation, anomaly correction, and dynamic range compression, improving data quality and comparability. Voltage plateau duration, temperature rise rate, and principal component scores are extracted and assigned semantic labels to construct physically interpretable features. A latent spatial projection network is used to fuse multi-source features, generating a unified latent representation. By combining recurrent neural networks, attention mechanisms, and multi-scale fusion strategies, historical dependencies are mined to generate a comprehensive state representation. Performance prediction values ​​are output synchronously, and degradation trajectories are deduced. Independent verification and dynamic calibration ensure prediction stability. This deep fusion of multi-source information significantly improves the accuracy and adaptability of predicting the performance degradation trend of nickel-metal hydride batteries. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning, as described in this invention. Detailed Implementation

[0025] 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 specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention provides, for example Figure 1 This paper presents a machine learning-based method for constructing a performance prediction model for nickel-metal hydride (NiMH) batteries. This method systematically integrates data from various sources, including electrochemical testing, environmental monitoring, operational logs, and material characterization, to establish a unified data processing framework. Furthermore, it utilizes nonlinear mapping and dynamic weight allocation techniques to enhance the ability to characterize the internal state evolution trends of NiMH batteries. Specifically:

[0027] In this embodiment, the method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning includes the following steps:

[0028] Electrochemical test data, environmental monitoring data, operation log data, and material characterization data are collected and mapped to a unified time base to form a time-aligned raw data stream. Furthermore, the electrochemical test data includes constant current charge-discharge records and internal resistance change sequences; the environmental monitoring data includes temperature sensor readings and humidity monitoring data; the operation log data includes cycle count logs and package deformation images; and the material characterization data includes X-ray diffraction patterns. The timestamps of all data are linearly mapped with the first test time as the origin.

[0029] Missing value imputation and outlier correction are performed based on the original data stream. Numerical variables are normalized and their dynamic range is compressed to generate a structured data matrix. Furthermore, missing value imputation uses the sliding window local mean interpolation method, outlier correction uses the three-standard-deviation criterion to identify and replace outliers, dimensional normalization maps variables to the zero-one interval, and dynamic range compression applies a logarithmic transformation to long-tailed distributed variables.

[0030] The voltage plateau duration of the electrochemical sequence, the temperature rise rate of the thermodynamic data, and the principal component score of the material image are extracted from the structured data matrix to form a feature set with semantic labels. Furthermore, the voltage plateau duration is calculated by the start and end time of the charge-discharge cycle, the temperature rise rate is obtained by temperature difference processing, and the principal component score is extracted by material image vectorization and covariance matrix decomposition. All features are aligned by time and concatenated into a feature set.

[0031] The feature set is input into the latent spatial projection network, and the nonlinear mapping process is optimized by self-reconstruction loss and contrastive regularization term to generate a unified latent representation. Specifically, the latent spatial projection network contains three fully connected layers. The self-reconstruction loss recovers the original features through the decoder network, and the contrastive regularization term calculates the distribution difference of positive and negative sample pairs through cosine similarity. The optimized unified latent representation retains the discriminative structure of multi-source information.

[0032] A unified latent representation is fed into a recurrent neural network structure, and an attention mechanism and a multi-scale temporal information fusion strategy are combined to generate a comprehensive state representation containing historical dependencies. Specifically, the recurrent neural network structure is a long short-term memory network, the attention mechanism focuses on key historical segments through weighted context vectors, and the multi-scale temporal information is achieved by splicing hidden states with different sampling granularities. The final output comprehensive state representation contains complete evolutionary path information.

[0033] Based on the comprehensive state representation, the predicted values ​​of remaining capacity, internal resistance growth trend and health index are output synchronously, and the future degradation trajectory is generated recursively. Furthermore, the prediction of remaining capacity is generated synchronously through the multi-task output structure to generate the internal resistance growth trend and health index, and the future degradation trajectory is generated by recursively predicting the state of future time steps. The prediction results are accompanied by confidence intervals and ranking of influencing factors.

[0034] The error and trend consistency of the predicted values ​​are evaluated using an independent validation set. Calibration is performed on the new battery model to optimize inference efficiency and establish a dynamic update mechanism. The dynamic update mechanism triggers retraining by periodically collecting new data. The version management mechanism maintains system stability through difference analysis and rollback strategies. The update process starts with the current parameters and makes adjustments in a limited number of rounds.

[0035] This invention effectively integrates multi-source heterogeneous data from nickel-metal hydride batteries, overcoming interference caused by time asynchrony and dimensional differences, and significantly improving the accuracy of characterizing performance degradation trends. It exhibits stronger stability and adaptability, particularly in trend direction identification and long-term trajectory extrapolation, thus providing a reliable basis for battery health management. The following detailed description of the specific steps is as follows:

[0036] Step 1: Multi-source heterogeneous data acquisition and preprocessing

[0037] This step aims to collect and initially organize various types of data generated during the actual use of nickel-metal hydride (NiMH) batteries, providing a clear and reliable foundational information flow for subsequent analysis. Because the operating state of NiMH batteries is influenced by multiple factors such as electrochemical reactions, thermodynamic processes, mechanical stress, and external loads, their operational data exhibits significant multi-source and heterogeneity, characterized by inconsistent sampling frequencies, different unit systems, large dimensional spans, and uneven distribution of missing values. Therefore, a standardized data access process must be designed to effectively clean and normalize the format of the raw data.

[0038] Define data acquisition channels and time base:

[0039] First, the data sources involved in performance prediction are identified, including but not limited to constant current charge / discharge records, temperature sensor readings, voltage fluctuation curves, internal resistance change sequences, cycle count logs, ambient humidity monitoring, battery package deformation images, and material X-ray diffraction patterns. Each type of data is generated by independent hardware devices or software systems, and its original output format varies, such as numerical sequences, two-dimensional images, text logs, or spectrograms. To achieve cross-source alignment, a unified time base is established. Typically, the time point is taken as the first time the battery is put into testing, and the timestamps of all subsequent data acquisition actions are linearly mapped to this origin. For discrete events (such as an increase in the number of cycles), the time of their occurrence is recorded as... and introduce indicator variables Indicates the time of the event Has it already happened? ;

[0040] This expression is used to embed discontinuous events into a continuous timeline, facilitating subsequent synchronization with other time series data. The technical benefit of this operation is that it enables asynchronous data that were previously incomparable to be correlated within a unified timeframe, providing a consistent temporal reference for subsequent feature extraction.

[0041] Perform missing value imputation and outlier correction:

[0042] During actual data acquisition, some sensors may experience temporary malfunctions or communication interruptions, leading to data loss for certain time periods. Furthermore, electromagnetic interference or hardware drift may cause individual values ​​to deviate significantly from the normal range, forming outliers. To ensure data integrity, a sliding window-based local mean interpolation method is used to handle missing values. Let a certain time series... In the interval Memory is missing, select before and after. Each valid sampling point constitutes a neighborhood set. Calculate their arithmetic mean as the filler value: ;

[0043] in This represents the number of valid samples within the neighborhood. Outliers are identified using the three-standard-deviation criterion: if a sampling point... ,in If the mean and standard deviation of the sequence within the nearest window are found to be outliers, they are replaced using the imputation method described above. The principle behind this process is to restore data continuity using local statistical properties while avoiding biases in subsequent analysis caused by extreme values, thereby improving overall data quality.

[0044] Implement dimensional normalization and dynamic range compression:

[0045] Different sensors output data with different physical units and numerical ranges. For example, voltage is measured in volts (V), temperature in degrees Celsius (°C), and diffraction intensity is a dimensionless value. Directly merging these data would cause larger variables to dominate subsequent analyses, masking smaller but crucial trends. Therefore, min-max normalization is applied to all numerical variables, mapping them to... Interval: ;

[0046] in and These represent the minimum and maximum values ​​of the variable over the entire observation period. For variables exhibiting a long-tailed distribution (such as the rate of internal resistance growth), a logarithmic transformation is further introduced to compress the dynamic range: ;

[0047] This transformation effectively mitigates the impact of high-amplitude samples on the overall distribution, making it easier to capture low-frequency but important changes. The underlying principle is to balance the contribution weights of each variable in subsequent analysis, preventing any single data type from dominating the overall structure learning process due to numerical advantage.

[0048] Establish a multi-source data alignment matrix:

[0049] After completing the above preprocessing, all data streams need to be aligned by timestamps to form a unified data structure, facilitating subsequent joint analysis. Construct a two-dimensional matrix. ,in This represents the total number of time steps. This represents the total number of feature dimensions. Each row corresponds to a time point. Each column represents a preprocessed sequence of variables. For variables with an original sampling frequency lower than that of the main time axis (such as weekly material testing), forward fill is used to extend them to the full time domain. ;

[0050] The most recent valid measurement value is used to fill the current time step. For image-type data (such as encapsulated deformation maps), the principal component scores are extracted as vector representations and added as multiple columns to the matrix. The final alignment matrix is ​​obtained. This serves as the foundational input for subsequent feature extraction. The technical advantage of this structure lies in its ability to achieve structured integration of heterogeneous data, transforming the originally dispersed information flow into a unified representation that can be used for mathematical operations, thus laying the foundation for subsequent in-depth analysis.

[0051] Step 2: Hierarchical Feature Extraction and Semantic Encoding:

[0052] After completing the structured integration of multi-source data, the next step is to align the matrix. Each type of information is analyzed in depth to extract representative features that reflect the internal state evolution of nickel-metal hydride batteries. Since different data sources contain different levels of information—for example, electrochemical data reflects instantaneous reaction kinetics, material images reveal microstructural degradation, and environmental data reflects the effects of external disturbances—a hierarchical feature extraction strategy is needed to process each type of data separately and assign them corresponding semantic labels for subsequent targeted fusion.

[0053] Extracting dynamic response features from electrochemical sequences:

[0054] For high-frequency sampling sequences such as voltage, current, and internal resistance, the focus is on their variation patterns within the charge-discharge cycle. First, the start and end times of each complete charge-discharge cycle are identified, and boundaries are determined based on current direction changes. Within each cycle, the duration of the voltage plateau period is calculated. This value decreases with battery aging and is an important indicator of capacity degradation. Furthermore, the integral of the rate of voltage change is defined as a measure of polarization: ;

[0055] A higher integral value indicates higher resistance to charge transfer within the battery, reflecting electrolyte degradation or electrode passivation. Additionally, the voltage drop at the end of each discharge cycle is calculated. Its growth trend is closely related to the decline in the hydrogen storage capacity of the negative electrode. These characteristics together constitute the primary representation of the electrochemical response layer. ,in This represents the number of features extracted. This representation preserves the dynamic behavior information of the battery during the energy conversion process, providing a direct basis for subsequent state deduction.

[0056] Identifying energy dissipation patterns from thermodynamic data:

[0057] Temperature data not only reflects the battery's heat dissipation efficiency but also indirectly reveals the activity level of internal side reactions. The normalized temperature series... Differential processing is performed to obtain the temperature rise rate. During the constant current discharge stage, if A reading significantly higher than the historical average may indicate an internal short circuit or accelerated electrolyte decomposition. Furthermore, the ratio of temperature change to the square of the current per unit time is calculated as an estimate of the equivalent thermal resistance. ;

[0058] This ratio increases with the length of service life, reflecting insulation aging or poor contact. Simultaneously, the percentage of high-temperature duration is also statistically analyzed. This means the temperature exceeds the safe threshold. The percentage of time spent on this task relative to the total running time: ;

[0059] This indicator is negatively correlated with battery life. The above combination of characteristics forms a thermodynamic characteristic set. This study captures the loss characteristics of batteries during energy conversion, supplementing the physical degradation pathways that are not directly reflected in electrochemical data.

[0060] Extracting structural degradation indices from material images:

[0061] For periodically acquired images of encapsulation deformation or X-ray diffraction patterns, principal component analysis (PCA) is used to extract the main directions of change. Let the image matrix be... Vectorize it as And form a matrix from the vectors at all time points. Calculate the covariance matrix. And seek its predecessor. The eigenvectors corresponding to the largest eigenvalues Projecting each image onto these principal components yields a low-dimensional representation: ;

[0062] Each component Reflects the first The activation intensity of typical deformation modes, such as electrode expansion, diaphragm displacement, or weld cracking, was determined. These scores were interpolated across the entire time axis to form the material evolution profile. The principle behind this process is to compress high-dimensional visual information into quantifiable degradation trajectories, transforming image data that is difficult to model directly into numerical features that can be used for trend prediction.

[0063] Construct a feature set with semantic labels:

[0064] The aforementioned three types of features Concatenate columns to form a comprehensive feature matrix. ,in To enhance the interpretability of subsequent processing, semantic labels are assigned to each feature dimension, such as "voltage plateau duration," "temperature rise rate," and "principal component 1 score." These labels not only record the physical meaning of the features but also indicate their source category and extraction method, forming a metadata dictionary. This dictionary plays a crucial role in subsequent weight allocation and result interpretation. The technical effect of this operation is to transform the raw data into an intermediate representation with clear physical meaning, enabling subsequent modeling processes to not only capture statistical correlations but also maintain consistency with the battery's working mechanism.

[0065] Step 3: Cross-domain feature mapping and latent space construction

[0066] In obtaining the original feature set containing semantic information The next step is to map these features to a unified latent space, allowing features from different sources and with different dimensions to be compared and fused within the same geometric structure. Since the rates of change, response sensitivities, and noise levels of various features differ, direct splicing may lead to information distortion. Therefore, a shared implicit representation space needs to be established through nonlinear transformations, in which similar battery states correspond to nearby points, while different degradation paths are distinguished.

[0067] Initialize the latent spatial projection network:

[0068] Design a feedforward network structure to process the feature vectors at each time point. Mapping to a low-dimensional latent space ,in The network consists of three fully connected layers, using the rectified linear unit (ReLU) activation function. Let the network parameters be... The mapping relationship is then: ;

[0069] During initialization, parameters The Xavier method is used for randomization to ensure that the outputs of each layer have a similar variance distribution. The input of this network is... Each row of the output forms the latent representation matrix. The purpose of this step is to establish a learnable nonlinear transformation pathway, providing the infrastructure for subsequent spatial optimization.

[0070] Introducing self-reconstruction loss to guide spatial learning:

[0071] To ensure potential representation To fully preserve key information from the original features, a reconstruction task is introduced as a supervision signal. A decoder network is designed. Try to start from Restore original features ;

[0072] Define mean squared error as the reconstruction loss: ;

[0073] By minimizing this loss, the latent space is forced to retain sufficient information to support inverse reconstruction. The principle behind this mechanism is to utilize the concept of information bottlenecks to select the most representative structural patterns while reducing dimensionality, thereby removing redundant and noisy components.

[0074] Adding contrastive regularization terms enhances state discriminability:

[0075] Relying solely on reconstruction loss may lead to excessive clustering of different battery states in the latent space, which is detrimental to subsequent classification or regression tasks. Therefore, a contrastive learning mechanism is introduced to encourage similar states to cluster together and dissimilar states to separate. Positive sample pairs are defined as the states of the same battery at two adjacent time points. Negative sample pairs represent states from different batteries or at time points separated by a considerable distance. InfoNCE loss is used. ;

[0076] in For cosine similarity, For temperature parameters, The total number of samples is represented by this loss. This loss enables the network to learn more discriminative representations, improving its sensitivity to subtle state changes.

[0077] Joint optimization yields the final potential representation:

[0078] The reconstruction loss and the contrast loss are weighted and combined to form the overall optimization objective: ;

[0079] in The balance coefficients are updated iteratively using gradient descent. and This continues until the loss converges. Finally, the optimized latent representation is obtained. each of the lines That is, the first The comprehensive state code at each time point. This representation integrates multi-source information and has undergone geometric structure optimization, making it an ideal input for subsequent relational modeling.

[0080] Step 4: Temporal Dependency Modeling and Dynamic Relationship Mining

[0081] Obtaining the optimized latent state representation The next step is to analyze its evolution over time and uncover the long-term dependencies in battery performance degradation. The aging process of nickel-metal hydride batteries is cumulative and non-linear; the state at a single point in time is insufficient to predict future performance, and the influence of historical trajectories must be considered. Therefore, a time-series analysis structure capable of capturing long-range dependencies needs to be constructed to reveal the intrinsic dynamics of state transitions.

[0082] Constructing a recurrent neural network structure:

[0083] Long Short-Term Memory (LSTM) networks were chosen as the temporal modeling tool because their gating mechanism effectively mitigates the vanishing gradient problem, making them suitable for processing sequence data with long time spans. Let the hidden states of an LSTM unit be... The update process is as follows:

[0084] ;

[0085] ;

[0086] ;

[0087] ;

[0088] ;

[0089] ;

[0090] in These are the input gate, forget gate, and output gate, respectively. In cellular state, This represents element-wise multiplication. This is a sigmoid function. This structure can selectively remember or forget historical information, adapting to the phased characteristics of battery degradation.

[0091] Introducing attention mechanisms to focus on key historical segments:

[0092] While LSTMs can handle long sequences, not all historical states are equally important for the current prediction. For example, recent deep discharges may have a greater impact than earlier, milder uses. To address this, an attention mechanism is introduced into the LSTM output, calculating the weights for each time step:

[0093] ;

[0094] ;

[0095] in These are learnable parameters. The weighted context vector is: ;

[0096] This vector focuses on reflecting the past experiences that have the greatest influence on the current state, thus improving the accuracy of predictions.

[0097] Fusion of multi-scale temporal information:

[0098] To balance short-term fluctuations and long-term trends, a multi-layer LSTM structure is adopted, with each layer processing information at different time granularities. The bottom layer processes the raw time step, and the middle layers process information at intervals of... Each layer is sampled once, with the top layer further downsampled. The outputs of each layer are then combined by concatenation or weighted averaging. ;

[0099] in For the number of floors, To incorporate weights, this design enhances the model's adaptability to complex temporal patterns.

[0100] Generate time-enhanced state representations:

[0101] Finally, the attention-weighted context vector With the hidden state of the last LSTM layer By piecing them together, a comprehensive representation containing complete historical dependencies is formed. This representation, serving as input for the next stage, carries information about the complete evolutionary path of the battery from its initial state to the current moment.

[0102] Step 5: Performance Index Deduction and Degradation Trajectory Generation

[0103] To obtain a comprehensive representation that includes complete historical dependencies The next step is to transform these results into concrete performance predictions, including quantifiable indicators such as remaining capacity, internal resistance growth trends, and failure probability. The core task at this stage is to establish a mapping from abstract state encoding to specific physical quantities, while simultaneously generating degradation path predictions for a future period, providing a basis for maintenance decisions.

[0104] Establish a multi-task output structure:

[0105] Design a multi-task network with a shared hidden layer to simultaneously predict multiple relevant performance metrics. Let the output of the shared layer be... Each of the three independent regression heads is connected to predict the remaining capacity. Internal resistance With health index The loss function for each task is the mean squared error:

[0106] ;

[0107] ;

[0108] ;

[0109] The total loss is the weighted sum: Multi-task learning can promote feature sharing and improve generalization ability.

[0110] Introducing uncertainty estimation improves reliability:

[0111] To assess the reliability of the prediction results, the Monte Carlo Dropout method is used to estimate the prediction variance. Dropout is kept enabled during the network inference phase, and multiple forward propagations are performed to obtain the distribution of the predicted values: ;

[0112] Calculate the mean and variance: High variance areas indicate significant forecast uncertainty and should be treated with caution.

[0113] Generate future degradation trajectory:

[0114] Based on the current state By combining environmental assumptions (such as constant temperature and periodic load), future predictions are recursively made. The hidden state is updated using an LSTM structure at each step, and the corresponding performance value is output to form a degradation curve. This trajectory can be used for end-of-life (EOL) estimation.

[0115] Output a forecast report with confidence intervals:

[0116] The final report generates a complete report containing point predictions and 95% confidence intervals, supplemented by a ranking of key influencing factors (based on gradient analysis) to help users understand the basis for the predictions.

[0117] Step Six: Model Validation and Deployment Configuration

[0118] After completing the performance index extrapolation and degradation trajectory generation process, the resulting prediction system needs to undergo systematic verification and engineering adaptation to ensure stable output capabilities in real-world application scenarios. This stage not only focuses on the statistical consistency of prediction results but also comprehensively considers the diversity of operating environments, hardware resource constraints, and long-term maintenance requirements. By constructing a closed-loop verification mechanism and a flexible deployment architecture, the entire system can maintain functional integrity and responsive reliability in application scenarios of varying scales and complexities.

[0119] Design a multi-dimensional cross-validation scheme:

[0120] To comprehensively evaluate the performance of the prediction system on unknown data, a hierarchical cross-validation strategy was adopted to avoid misjudgments caused by sample distribution bias. First, all available battery samples were classified according to usage history, manufacturing batch, and working environment, ensuring that each category contained a sufficient number of individuals. Then, time-series segmentation was performed within each category, retaining the most recent 20% of the time period as an independent test set, with the remainder used for training and parameter tuning. This segmentation method simulates the logical scenario of "predicting future states based on historical data" in real-world applications, preventing information leakage. During the validation process, not only was the error between the predicted values ​​and actual observations (such as root mean square deviation) calculated, but a trend consistency index was also introduced. This is used to measure how well the predicted degradation path matches the actual trajectory. ;

[0121] in The length of the verification period. and The first Predicted and measured values ​​at each time point Returns the sign of the numerical value. This indicator reflects whether the model can correctly capture the trend direction of performance changes. Even if there is a deviation in magnitude, as long as the trend is consistent, it still has decision-making reference value. The technical effect of this operation is to break through the limitations of a single error metric, and to comprehensively evaluate predictive ability from two levels: "numerical approximation" and "trend identification," thereby improving the practicality of the evaluation system.

[0122] Perform cross-device migration capability testing:

[0123] Because different models or batches of nickel-metal hydride batteries differ in material ratios, packaging processes, and electrode structures, directly applying prediction logic trained on a particular batch to new devices may result in performance degradation. To address this, a transfer adaptability testing process was designed: a group of new battery models that did not participate in any training process was selected, their historical operating data was input, and the initial deviation and convergence speed of the prediction results were observed. If a systematic bias was found (such as generally overestimating capacity predictions), a lightweight calibration mechanism was introduced. Specifically, an adjustable affine transformation layer was added to the model output. ;

[0124] in and These are undetermined coefficients. Using a small amount of initial measured data from the new equipment (such as the first three complete cycles), the optimal parameters are solved using the least squares method.

[0125] ;

[0126] To calibrate the sample size, this transformation can quickly adapt to the static offset characteristics of new devices without modifying the backbone structure. The principle behind this process is to separate the "common evolutionary laws" from the "individual initial biases." The former is learned by the master model, while the latter is compensated for through post-processing, thereby enhancing the system's generalization boundary.

[0127] Optimize inference efficiency to adapt to edge operating environments:

[0128] In practical deployments, prediction systems may need to run on embedded controllers or mobile terminals. These devices have limited computing resources, making it difficult to support real-time inference of highly complex neural networks. Therefore, lightweight modifications to the model are necessary. First, weight quantization is implemented on the latent spatial projection network and LSTM structure, compressing the 32-bit floating-point parameters into 8-bit integer representations, significantly reducing memory usage while ensuring controllable accuracy loss. Second, for matrix multiplication operations in LSTM, low-rank decomposition techniques are employed to reduce the size of the large weight matrix. It can be approximated as the product of two smaller matrices: ;

[0129] This decomposition can be achieved through singular value decomposition (SVD), taking the first... The left and right singular vectors corresponding to the maximal singular values ​​constitute and This operation will restore the original The number of multiply-accumulate operations is reduced to This significantly reduces the computational load. Furthermore, for similarity calculation in the attention mechanism, a Locality Sensitive Hash (LSH) pre-screening mechanism is employed, performing fine-grained matching only on potentially relevant historical segments to avoid full-sequence scanning. These optimization techniques work together to enable the model to complete a single prediction within milliseconds, meeting the real-time requirements of online monitoring. The technical effect of this processing is to balance prediction accuracy and operational efficiency, allowing complex analytical capabilities to be applied on resource-constrained devices.

[0130] Establish a dynamic update and version management mechanism:

[0131] Battery aging is a long-term, cumulative process. Over time, new failure modes may emerge, and existing predictive logic may gradually become ineffective. To maintain system effectiveness, a periodic update mechanism needs to be established. A fixed time interval should be set. (For example, every three months), collect newly accumulated experimental data and re-evaluate the model's performance on the latest samples. If the average error exceeds a preset threshold... If the model fails to perform well, a retraining process is triggered. During retraining, the system doesn't start from scratch but rather uses the current model parameters as a starting point, performing a limited number of parameter adjustments on newly added data to avoid catastrophic forgetting. Simultaneously, a new version number is generated after each update, and a copy of the old version is saved, forming a version chain. When a new version performs poorly on the validation set, it can be quickly rolled back to the previous stable version. Inter-version difference analysis can also reveal trends in data distribution, assisting in improving data collection strategies. The principle behind this mechanism is to treat the prediction system as a continuously evolving entity, rather than a one-time product, achieving self-improvement through closed-loop feedback and extending its lifespan.

[0132] Build a visual interactive interface to support decision interpretation:

[0133] To enhance system usability, a supporting visualization platform was developed to transform abstract prediction results into intuitive graphical information. The platform's homepage displays an overview of the current battery health status, including estimated remaining capacity, expected number of usable cycles, and a ranking of major risk factors. Clicking on any indicator expands a detailed evolution curve, overlaying historical measured values ​​with future predicted paths, and using shaded areas to indicate confidence levels. For critical events (such as abnormal temperatures and deep discharge), the system automatically labels their occurrence time and provides a quantitative assessment of their impact on battery life. Users can view replays of different stages by sliding the timeline, and can also modify future usage assumptions (such as increasing the discharge rate or changing the ambient temperature) to observe changes in the predicted trajectory in real time, enabling a "hypothesis analysis" function. All charts can be exported to standard formats for easy archiving or reporting. The technical effect of this interface is to bridge the gap between professional analysis and practical operation, allowing non-technical personnel to understand the basis of predictions and make reasonable decisions.

[0134] Develop safety redundancy and contingency plans:

[0135] In critical application scenarios, the failure of the prediction system can have serious consequences, therefore, comprehensive safety protection measures must be in place. First, an integrity verification mechanism should be implemented at the data input end. If a sensor data missing rate exceeds 50% or a clearly physically unreasonable value appears (such as a sudden voltage drop to a negative value), the system will pause prediction and issue an alarm. Second, a reasonableness boundary must be set at the prediction output end. If the generated capacity prediction is lower than 10% of the nominal value or higher than 105% of the initial value, the result is considered abnormal, and a backup simplification rule (such as linear extrapolation) is activated to provide a conservative estimate.

[0136] In addition, the system should periodically self-check the operational status of core components and record relevant log files. Upon detecting anomalies such as memory overflow or computation timeout, the system will immediately restart the service and notify the administrator. All prediction request and response information must be stored encrypted to support post-event traceability and auditing. The principle behind this contingency plan is to build a multi-layered fault-tolerant system, implementing rigorous monitoring throughout the entire process from front-end data input and mid-level computation processing to back-end result output. This ensures that even if some components fail, the system can still provide basic services or perform security degradation responses, thereby guaranteeing the continuity and reliability of overall operation.

[0137] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning, characterized in that, Includes the following steps: Collect electrochemical test data, environmental monitoring data, operation log data and material characterization data, and map multi-source heterogeneous data to a unified time base to form a time-aligned raw data stream; Missing value imputation and outlier correction are performed based on the original data stream. Numerical variables are normalized and their dynamic range is compressed to generate a structured data matrix. The voltage plateau duration of the electrochemical sequence, the temperature rise rate of the thermodynamic data, and the principal component score of the material image are extracted from the structured data matrix to form a feature set with semantic labels. The feature set is input into the latent spatial projection network, and the nonlinear mapping process is optimized by self-reconstruction loss and contrastive regularization term to generate a unified latent representation. By feeding a unified latent representation into a recurrent neural network structure and combining an attention mechanism with a multi-scale temporal information fusion strategy, a comprehensive state representation containing historical dependencies is generated. Based on the comprehensive state representation, the predicted values ​​of remaining capacity, internal resistance growth trend and health index are output synchronously, and the future degradation trajectory is generated recursively. The error and trend consistency of the predicted values ​​are evaluated using an independent validation set, and calibration is performed on the new battery model to optimize inference efficiency and establish a dynamic update mechanism.

2. The method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning according to claim 1, characterized in that: The electrochemical test data includes constant current charge-discharge records and internal resistance change sequences; environmental monitoring data includes temperature sensor readings and humidity monitoring data; operation log data includes cycle count logs and package deformation images; material characterization data includes X-ray diffraction patterns, and the timestamps of all data are linearly mapped with the first test time as the origin.

3. The method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning according to claim 1, characterized in that: The missing value imputation method uses the sliding window local mean interpolation method, the outlier correction method uses the three standard deviation criterion to identify and replace outliers, the dimension normalization maps the variables to the zero-one interval, and the dynamic range compression applies a logarithmic transformation to the long-tailed distribution variables.

4. The method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning according to claim 1, characterized in that: The voltage plateau duration is calculated by the start and end times of the charge-discharge cycle, the temperature rise rate is obtained by temperature difference processing, the principal component score is extracted by material image vectorization and covariance matrix decomposition, and all features are aligned by time and concatenated into a feature set.

5. The method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning according to claim 1, characterized in that: The latent spatial projection network contains three fully connected layers. The self-reconstruction loss recovers the original features through the decoder network, and the comparison regularization term calculates the distribution difference of positive and negative sample pairs through cosine similarity. The optimized unified latent representation retains the discriminative structure of multi-source information.

6. The method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning according to claim 1, characterized in that: The recurrent neural network structure is a long short-term memory network. The attention mechanism focuses on key historical segments through weighted context vectors. Multi-scale temporal information is achieved by splicing hidden states with different sampling granularities. The final output comprehensive state representation contains complete evolutionary path information.

7. The method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning according to claim 1, characterized in that: The prediction of remaining capacity is generated synchronously through a multi-task output structure, which generates the internal resistance growth trend and health index. The future degradation trajectory is generated by recursively predicting the state at future time steps. The prediction results include confidence intervals and ranking of influencing factors.

8. The method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning according to claim 1, characterized in that: It also includes verifying model adaptability, specifically including hierarchical cross-validation and transferability testing, and optimizing inference efficiency, including weight quantization, low-rank decomposition, and locality-sensitive hash pre-screening.

9. The method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning according to claim 1, characterized in that: The dynamic update mechanism triggers retraining by periodically collecting new data, and the version management mechanism maintains system stability through difference analysis and rollback strategies. The update process starts with the current parameters and makes adjustments in a limited number of rounds.

10. The method for constructing a performance prediction model for nickel-metal hydride batteries based on machine learning according to claim 1, characterized in that: Also includes: Develop a visualization platform to display an overview of health status, historical evolution curves, and future prediction paths, overlaid with confidence ranges; Based on the content displayed on the visualization platform, users can modify future usage assumptions and observe changes in predicted trajectories in real time. Based on user interaction behavior, input integrity checks are set up, and predictions are paused and alerts are issued for data with excessively high missing rates or physically unreasonable data.

Images