Battery state of health prediction type charging method fusing deep learning

Abstract

This invention discloses a battery health state prediction charging method integrating deep learning, belonging to the field of battery management technology. The method constructs structured temporal feature inputs by synchronously collecting multi-source data during the charging process. The core lies in deploying a deep learning model composed of a convolutional neural network, a bidirectional long short-term memory network, and an attention mechanism to predict the current and future health state evolution trend of the battery online. Furthermore, the predicted trend is quantified into a lifespan degradation target, which, together with safety and efficiency targets, constitutes a dynamic optimization function to solve for the optimal charging parameters in real time. Finally, closed-loop control is executed, and an elastic weight solidification algorithm is used to incrementally learn and update the model, achieving self-adaptation throughout the entire lifespan. This invention, through the deep integration of deep learning prediction and real-time optimization, achieves forward-looking and personalized intelligent charging control, significantly improving the safety, efficiency, and lifespan maintenance level of battery management.

Description

Technical Field

[0001] This invention relates to the field of battery management technology, and more specifically to a battery health state prediction charging method that integrates deep learning. Background Technology

[0002] As an electrochemical energy storage device, the internal aging of lithium-ion batteries is essentially the cumulative result of a series of irreversible side reactions during lithium-ion insertion / extraction (such as solid electrolyte interfacial film growth, active lithium loss, and cathode material phase transition). State of Health (SOH), as an indicator of this aging degree, is a key internal state variable determining the battery's usable capacity, power capability, and safety boundary. Traditional battery management systems (BMS) rely on threshold values ​​of externally measurable parameters (such as terminal voltage, current, and temperature) for charging control (e.g., constant current-constant voltage charging). This control method is open-loop or based on instantaneous state feedback control, and its control law does not include estimation and response to the battery's internal SOH and its future evolution trend. Therefore, when facing batteries at different aging stages and with individual differences, it is impossible to dynamically adjust charging parameters to find the optimal trade-off between charging speed, temperature rise, and aging rate, leading to either overly conservative approaches (sacrificing charging efficiency) or accelerated aging (increasing safety risks).

[0003] To improve the intelligence level of battery management, existing technologies have introduced data-driven methods such as deep learning for State of Harshness (SOH) prediction. These methods typically learn the nonlinear mapping relationship between SOH and external features (voltage, current, temperature, etc.) from historical operating data by constructing complex neural networks (such as CNN, LSTM, and attention mechanisms). However, the existing technical approaches generally suffer from the following drawbacks:

[0004] Most existing solutions position deep learning models as independent "state observers" or "health monitoring instruments." Their high-precision SOH (State of Health) predictions are only used for status display, lifespan warnings, or system log recording, and are not used as decision inputs for the charging controller. For example, patents such as CN116299005B disclose high-precision SOH prediction models, but their outputs are not fed back to the charging control loop to change the charging current, voltage, or switching thresholds. The prediction module and the control module are serially connected but uncoupled in the system architecture, forming an "open-loop observation, fixed-rule control" system, where the value of the prediction information is not fully utilized.

[0005] A few attempts to combine prediction and control have significant limitations in utilizing predictive information. A common approach is to set a simple SOH threshold (e.g., SOH < 80%), and when the predicted value falls below this threshold, switch to a set of preset, more conservative charging parameters (e.g., reducing current). This "if-then" rule-based control is inherently static and piecewise linear, unable to perform smooth, dynamic optimization based on continuous changes and future trends in SOH. More importantly, it fails to establish a quantitative mathematical relationship between the predicted SOH value and the optimal charging parameters, making it impossible to achieve continuous Pareto front optimization for multiple objectives (safety, efficiency, lifetime).

[0006] To obtain richer dynamic characteristics and improve prediction accuracy, some approaches (such as the perturbation injection-dependent methods mentioned in the background section) require actively injecting diagnostic signals such as current or voltage pulses during charging. This active perturbation introduces additional electrochemical stress, potentially accelerating battery aging and increasing the complexity of power hardware and control logic, thus reducing the long-term reliability and usability of the system. An ideal technical approach should fully utilize the natural excitation response data during normal charging.

[0007] The aging process of a battery is accompanied by a slow drift in its internal electrochemical properties, causing the accuracy of a well-trained static prediction model to decline in the later stages of the battery's lifespan. Existing solutions generally lack a low-cost, sustainable online model update mechanism, making it impossible for the prediction model to maintain accuracy matching the current aging state throughout the battery's entire service life. This leads to the gradual failure of control decisions based on outdated prediction information.

[0008] Existing technological systems suffer from gaps or inefficient coupling in the "perception-prediction-decision-execution" chain. There is an urgent need for a technology that can directly embed high-precision, low-latency health status prediction capabilities into the charging decision core without introducing additional disturbances, thereby constructing an end-to-end intelligent charging closed loop and achieving a dynamic balance and global optimization between safety, efficiency, and lifespan loss. Summary of the Invention

[0009] The purpose of this invention is to provide a battery health state prediction charging method that integrates deep learning. This method solves the technical problems in the prior art where battery health state prediction and charging control strategies are disconnected, and static models cannot adapt to battery aging drift. As a result, the prediction information cannot guide the charging process in real time and in an optimized manner, making it difficult to coordinate and ensure safety, efficiency and lifespan throughout the entire battery life cycle.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0011] A battery health state prediction charging method incorporating deep learning includes the following steps:

[0012] Step 1: Collect multi-source operating data of the battery in real time during the charging process. Simultaneously acquire the battery's terminal voltage, charging and discharging current, surface temperature, ambient temperature and cumulative charging and discharging amount through high-precision sensors. Perform timestamp alignment and noise filtering on the raw data to form a structured time-series feature sequence.

[0013] Step 2: Construct and update the deep learning health status prediction model online. Based on the multi-source operation data accumulated in the historical charging cycle and the corresponding health status labels, train a composite neural network model composed of convolutional neural network, bidirectional long short-term memory network and attention mechanism. The model takes the time sequence of the current charging cycle as input and outputs the current health status value of the battery and the health status evolution trend of the next few charging cycles.

[0014] Step 3: Embed the health status prediction results into the charging decision core. Based on the health status values ​​and their evolution trends, combined with preset safety boundaries, efficiency targets and lifespan loss constraints, dynamically generate the optimal charging parameter set for the next stage, including the current amplitude of the constant current stage, the voltage threshold of the constant voltage stage, the triggering conditions for stage switching and the charging termination criteria.

[0015] Step 4: Execute adaptive charging control, send the optimal charging parameter set to the charging controller, adjust the output characteristics of the charging power supply in real time, and feed back actual operating data in each control cycle for model fine-tuning and parameter correction, forming a closed-loop optimization loop;

[0016] Step 5: Continuous evaluation and iterative optimization. After each charge, the actual operation data and changes in health status of the entire charging process are included in the model training set. An incremental learning strategy is used to update the deep learning model online to ensure that it maintains high accuracy and strong generalization ability throughout the battery's entire life cycle.

[0017] Furthermore, the high-precision sensor includes a voltage sensor, a current sensor, and a temperature sensor. All sensor signals are synchronously sampled via the same data acquisition card, with a time synchronization error less than a preset time synchronization error threshold. Moreover, the measurement accuracy of the voltage sensor, current sensor, and temperature sensor is better than the preset voltage accuracy threshold, preset current accuracy threshold, and preset temperature accuracy threshold, respectively.

[0018] Furthermore, the health status tag is obtained through offline capacity testing. After each predetermined number of charging cycles, a standard constant current-constant voltage discharge test is performed. The ratio of the actual discharge capacity to the initial nominal capacity is used as the health status value at that moment. The health status tag for the intermediate cycle is generated by linear interpolation or spline fitting.

[0019] Furthermore, the convolutional neural network includes several one-dimensional convolutional layers, each configured with a number of convolutional kernels, kernel length, and stride, for extracting local temporal patterns; the bidirectional long short-term memory network includes several layers, each configured with a number of hidden units, for capturing long-term dependencies; the attention mechanism adopts an additive attention structure, with weights calculated based on the similarity between the hidden states of the current time step and historical time steps, to enhance attention to key aging features.

[0020] Furthermore, the safety boundary includes the maximum allowable charging current, the maximum allowable terminal voltage, and the upper limit of the temperature rise rate. The efficiency target is defined as maximizing the amount of charge received per unit time. The lifespan loss constraint is modeled as a quadratic function of the health state decay rate. The three are weighted and summed to form a multi-objective optimization function, and the weight coefficients are dynamically adjusted according to the current health state of the battery.

[0021] Furthermore, when the health status value is lower than the preset health status threshold, the constant current stage current amplitude is automatically reduced to a predetermined proportion of the rated value, and the constant voltage stage voltage threshold is lowered by a preset voltage adjustment amount, while the stage switching trigger condition is brought forward.

[0022] Furthermore, the charging controller adopts digital pulse width modulation technology, with a response delay less than a preset response delay threshold. It can stably adjust the output current or voltage within a predetermined time period based on the optimal charging parameter set issued, and has hardware-level protection functions for overvoltage, overcurrent, and overtemperature.

[0023] Furthermore, the incremental learning strategy employs an elastic weight solidification algorithm, which retains the important parameter weights of the learned aging mode while adaptively adjusting to changes in feature distribution in new data to avoid catastrophic forgetting, and the model update cycle does not exceed the predetermined number of charging cycles.

[0024] Furthermore, the method supports collaborative charging management of multi-battery parallel systems. By sharing a unified health status prediction model and a distributed charging controller, it achieves balanced regulation of the health status of each individual battery, and controls the maximum health status difference within a preset health status difference threshold.

[0025] Furthermore, in a multi-battery parallel system, when the maximum difference in the system's health status exceeds a preset threshold for the difference in health status, an equalization term is added to the multi-objective optimization function. This involves allocating lower charging current and voltage to batteries with lower health status, while allowing batteries with higher health status to be charged close to their rated parameters.

[0026] Furthermore, the method is deployed on an embedded edge computing platform equipped with a dedicated neural network accelerator, which can support real-time inference of composite neural network models, and the time taken for a single health status prediction is less than a preset inference time threshold.

[0027] Furthermore, the embedded edge computing platform adopts a lightweight neural network structure, including a compressed convolutional backbone network, a single-layer bidirectional long short-term memory network, and a simplified attention mechanism, to adapt to resource-constrained embedded hardware environments.

[0028] Furthermore, in home energy storage application scenarios, the efficiency target is implicitly defined by the charging completion time window set by the user. Within this time window, the system dynamically optimizes charging parameters to minimize lifespan loss.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] This invention achieves deep integration and coupling of sensing, prediction, decision-making, and execution. By directly embedding a deep learning-based health status prediction model into the charging decision core, an end-to-end intelligent charging closed loop is constructed. It abandons the traditional path that relies on external disturbance signals, utilizing multi-source operational data during normal charging to achieve millisecond-level adaptive regulation. This not only eliminates the risk of introducing additional electrochemical stress and simplifies system hardware complexity, but more importantly, it establishes a direct and dynamic mapping from health status prediction to charging parameter generation. This allows the charging strategy to proactively respond to the microscopic evolution of the battery's internal aging state, rather than passively reacting to current external physical quantities.

[0031] This invention quantifies the predicted health status evolution trend into a lifetime degradation target and embeds it into a real-time multi-objective optimization function, achieving coordinated dynamic optimization of safety, efficiency, and lifespan. Unlike existing methods that use fixed rules or simple threshold triggers, this scheme uses the future health status degradation trend as one of the core optimization variables and performs refined trade-offs through dynamic weighting coefficients linked to the current health status value. This allows the system to automatically and smoothly adjust its strategy focus at different stages of battery aging; for example, prioritizing charging efficiency when the battery is in good health and delaying capacity degradation as aging intensifies, thereby seeking the optimal Pareto balance point globally.

[0032] This invention introduces an incremental learning strategy based on an elastic weight solidification algorithm, endowing the system with the crucial capability of self-evolution throughout its entire lifecycle. This allows the deep learning prediction model to continuously update online using newly generated charging data without forgetting learned aging patterns, thus adaptively tracking the slow drift of feature distribution during battery aging. It fundamentally overcomes the inherent problem of declining prediction accuracy over time due to battery performance degradation in static models, ensuring that the matching degree between prediction accuracy and control strategy is maintained throughout the entire lifecycle from initial battery service to end of life, achieving sustained, personalized, and precise management.

[0033] This invention supports multi-scale hardware platforms ranging from embedded edge devices to large-scale energy storage systems, and provides corresponding lightweight modeling and distributed collaborative management schemes. This approach exhibits good scalability and engineering practicality. The architecture supports balanced control of multi-battery parallel systems, proactively reducing differences in the health status of individual cells, improving the overall available capacity and reliability of the system, and providing a practical intelligent charging solution for large-scale applications such as new energy vehicles and grid energy storage. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0035] Figure 1 This is an overall flowchart of the battery health state prediction charging method that integrates deep learning as described in this invention.

[0036] Figure 2 This is a schematic diagram of the CNN-BiLSTM-attention mechanism composite prediction model used in this invention.

[0037] Figure 3 This is a schematic diagram of the architecture for collaborative charging management when the present invention is applied to a multi-battery parallel system. Detailed Implementation

[0038] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0039] The following is in conjunction with the appendix Figures 1-3 The embodiments of the present invention will be described in detail below.

[0040] Example 1: This example discloses a battery health state prediction charging method that integrates deep learning, specifically including the following steps:

[0041] Step 1: Collect multi-source operational data of the battery in real time during the charging process, specifically including the following operations:

[0042] A high-precision sensor array is deployed, consisting of a voltage sensor, a current sensor, and a temperature sensor. The voltage sensor has a measurement accuracy better than 0.5% of full scale, the current sensor has a measurement accuracy better than 0.2% of full scale, and the temperature sensor has a measurement accuracy better than ±0.3℃.

[0043] All sensor signals are connected to the same model of data acquisition card, which has 8-channel synchronous sampling capability, a sampling frequency set to 100Hz, and a time synchronization error controlled within 1 millisecond.

[0044] The raw signals collected include battery terminal voltage. Charging and discharging current Battery surface temperature Ambient temperature and cumulative charge / discharge amount Each signal is accompanied by a high-precision hardware timestamp;

[0045] The original time-series signal is preprocessed. First, timestamp alignment is performed. Using the voltage signal as a reference, other signals are synchronized to a unified time axis through linear interpolation, with a time resolution of 10 milliseconds (i.e., 0.01 seconds). Then, a fifth-order Butterworth low-pass filter is used for noise filtering, with a cutoff frequency of 40Hz, to eliminate high-frequency electromagnetic interference and sensor jitter.

[0046] The filtered multidimensional signal is sliced ​​into time windows, each 600 seconds long, with a sliding step of 30 seconds, forming a structured time-series feature sequence. :

[0047] ;

[0048] in:

[0049] : Structured time-series feature sequences, which contain a set of time-series slices of multi-source operational data during the charging process;

[0050] : The multidimensional feature vector (t=1,2,…,T) at the t-th time step corresponds to the multi-source running data at a single time point;

[0051] The total number of time steps in a single timing window, in this embodiment ;

[0052] : A T-row, 5-column real number space, with 5 corresponding to 5 types of multi-source operating data (terminal voltage, charging and discharging current, surface temperature, ambient temperature, and cumulative charging and discharging amount).

[0053] Each row vector The basic unit that constitutes the input of the model.

[0054] Step 2: Build and update the deep learning health status prediction model online, which includes the following:

[0055] Health status label The data is obtained by performing a standard constant current-constant voltage discharge test after every 20 charging cycles. This involves first discharging at a constant current rate of 1C to the cutoff voltage of 3.0V, then discharging at a constant voltage of 3.0V until the current decays to 0.05C, and recording the actual discharge capacity. and calculate ,in This refers to the initial nominal capacity of the battery. This represents the battery health status value corresponding to the k-th offline capacity test (range: [0,1]); for intermediate loops that have not undergone offline testing, a continuous health status label sequence is generated using cubic spline interpolation.

[0056] Construct a composite neural network model, whose input is the generated time-series feature sequence. The output is an estimate of the battery's current health status. and the health status evolution trend sequence over the next 5 charging cycles .

[0057] The front end of the model is a convolutional neural network (CNN) module, which contains three one-dimensional convolutional layers. The first layer has 64 convolutional kernels, a length of 128, and a stride of 8. The second layer has 128 convolutional kernels, a length of 64, and a stride of 4. The third layer has 256 convolutional kernels, a length of 32, and a stride of 2. Each layer is followed by batch normalization and ReLU activation functions to extract local temporal patterns such as voltage plateau fluctuations and temperature rise inflection points.

[0058] The CNN output, after being flattened, is fed into a Bidirectional Long Short-Term Memory (BiLSTM) network module. This module consists of two stacked layers, each with 512 hidden units. The forward LSTM captures the aging cumulative effect from the initial charging stage to the current time step, while the backward LSTM models the dependency of future states on current features. The two are concatenated to output a context-aware sequence of hidden states. .

[0059] The attention mechanism employs an additive attention structure, which applies the hidden state to the output of the BiLSTM at each time step. Calculate attention weights First, calculate the global average of the hidden states across all time steps. ,in Output sequence length for BiLSTM.

[0060] The calculation formula is as follows:

[0061] ;

[0062] in,

[0063] Hidden state at time step t Attention weights (range: [0,1]);

[0064] Learnable query vectors (dimensions: ,in Hidden state (dimensions)

[0065] , Learnable weight matrix (dimension: );

[0066] The hidden state vector of a Bidirectional Long Short-Term Memory (BiLSTM) network at time step t;

[0067] BiLSTM outputs a globally average pooled feature vector of the hidden state sequence (dimension: );

[0068] : Summation variable, iterating through all time steps (τ=1,2,…,T′) of the BiLSTM output;

[0069] : The length of the hidden state sequence output by BiLSTM (the total number of time steps after CNN dimensionality reduction).

[0070] exp(·): Exponential function;

[0071] tanh(·): Hyperbolic tangent activation function.

[0072] The final weighted summation yields the context vector. Context vector It is fed into two parallel fully connected output layers: the first output layer is used for the current SOH estimation, and outputs a scalar. The second output layer is used for future SOH trend prediction, outputting a 5-dimensional vector. The loss function is a weighted combination of the mean squared error of the current value prediction and the smoothed L1 loss of the future trend prediction. The Adam optimizer is used during training, and the initial learning rate is set to 0.001.

[0073] Step 3: Embed the health status prediction results into the charging decision core, specifically including the following:

[0074] Define the set of safety boundary constraints Including the maximum allowable charging current Maximum allowable terminal voltage Upper limit of temperature rise rate .

[0075] A multi-objective optimization problem considering safety constraints is constructed. The optimization objective is to maximize charging efficiency and minimize lifetime loss while satisfying safety constraints. This can be transformed into a single-objective minimization problem through weighted summation:

[0076] Minimize: ;

[0077] in: This refers to the total charging capacity.

[0078] ,in , which is the amount of SOH decay that the current charging cycle will cause, as predicted by the model; This is the lifespan loss coefficient (preset as a positive real number). For dynamic weighting coefficients, satisfying ,and .

[0079] Constraints: , , , .

[0080] Based on this optimization problem, the Sequential Quadratic Programming (SQP) algorithm is used to solve for the optimal charging parameter set for the next stage. ,in This represents the current amplitude during the constant current phase. This is the voltage threshold during the constant voltage phase. This is the switching point from constant current to constant voltage. This is the criterion for determining the state of charge when charging is terminated.

[0081] When predicting health status values When the value is below 0.9, a protective strategy is triggered: Automatically reduces to 70% of the rated value (e.g., 1.0C), i.e., 0.7C; The voltage was reduced by 0.03V from 4.20V to 4.17V; and the stage switching condition was advanced from the default SOC=80% to SOC=70% to slow down the loss of lattice oxygen in the positive electrode material and the thickening of the SEI film in the negative electrode.

[0082] Step 4: Execute adaptive charging control, which includes the following operations:

[0083] The generated optimal charging parameter set The data is transmitted to the charging controller via CAN bus or EtherCAT industrial communication protocol, with a communication cycle of 100 milliseconds.

[0084] The charging controller uses digital pulse width modulation (DPWM) technology. Its main control chip is a TIC2000 series DSP with a built-in high-resolution PWM module, a switching frequency of 20kHz, and a response delay of less than 500 microseconds.

[0085] Controller according to Adjust the PWM duty cycle to stabilize the charging power supply output current to the target value within 200 milliseconds, with a steady-state ripple of less than ±1%; during the constant voltage stage, according to... Adjust the output voltage, with an adjustment time not exceeding 300 milliseconds.

[0086] Within each control cycle (10 milliseconds), the controller collects the actual output current, voltage, and battery feedback signals, and uploads them to the edge computing platform for model fine-tuning. If the actual current deviates from the command value by more than 3%, the parameter correction subroutine is triggered to adjust the DPWM dead time and the PI controller gain.

[0087] The controller has built-in hardware-level protection circuitry, including an overvoltage comparator (threshold 4.35V), an overcurrent detector (threshold 2.5C), and a temperature fuse (operating temperature 65℃). Once any of these protections is triggered, the power MOSFET drive signal is immediately cut off, and a fault code is reported.

[0088] Step 5: Continuous evaluation and iterative optimization, specifically including the following operations:

[0089] After each charge is completed, the system automatically packages the actual operating data of the entire charging process, including a complete time-series feature sequence. Actual charging parameters After the system completes a predetermined number of charging cycles (e.g., 20) and performs an offline capacity test, it obtains the accurate SOH label at that moment. This label is then associated with historical charging data within the corresponding time period to form new labeled training samples. The data is then incorporated into the model training set. During intermediate loops, the system records data but does not immediately use it for supervised training.

[0090] The Elastic Weight Fixation (EWC) algorithm is used for online model updates: First, based on the old model parameters... and their corresponding parameter importance metrics (For example, this importance metric can be approximated by the diagonal elements of the Fisher information matrix calculated from historical data.) The importance metric The calculation formula can be:

[0091] ;

[0092] in, It is the loss function for model training (such as mean squared error). It is the total number of historical data samples. For the first historical training samples ( ), Represents the model parameter vector The first in One parameter.

[0093] Then, optimize the loss function on the new data as follows: ; where the summation symbol Represents all parameters of the model Perform summation. For the standard loss function on the new data, This is the regularization strength coefficient of the EWC algorithm, which is set to 1000 in this embodiment.

[0094] The model update cycle is set to be performed once every 10 charging cycles to ensure that the model can adapt to the feature distribution drift caused by battery aging in a timely manner.

[0095] The updated model version number is incremented, and its inference stability is verified through a secure startup mechanism. It is only allowed to be deployed to the production environment if the time taken for a single prediction is less than 50 milliseconds and the output range is within [0.7, 1.05].

[0096] In practical implementation, consider an electric passenger vehicle equipped with a 60kWh lithium iron phosphate battery pack. Upon initial vehicle use, the system is initialized to a health state of 1.0, and an initial training set is established. During the first to 20th charges, the system employs a standard 1C constant current - 4.2V constant voltage strategy, while continuously collecting data and training the initial model in the background. Before the 21st charge, the system performs an offline capacity test, measuring a State of Health (SOH) of 0.998, and updates the label accordingly.

[0097] Thereafter, the method of this invention is performed for each charging:

[0098] At the 50th charge, the model predicted a current SOH of 0.96, and the evolution trend showed an average SOH decay of 0.0015 per cycle over the next 5 cycles. Based on this, the system generated the following charging parameters: constant current 0.95C (slightly below 1C to balance efficiency and lifespan), constant voltage 4.19V, switching SOC = 78%, and termination SOC = 98%. During charging, the controller provided feedback on actual data every 10 milliseconds, and the model fine-tuned the output to ensure stability.

[0099] At the 100th charge, the predicted State of Harmony (SOH) dropped to 0.88, and the system automatically activated its protection strategy: the constant current was reduced to 0.7C, the constant voltage to 4.17V, and the SOC switching was brought forward to 70%. Simulation / actual measurements show that, compared to the fixed strategy, this method keeps the SOH at 0.82 after 800 cycles, while the control group's SOH has decayed to 0.75, verifying its lifespan extension effect. The entire process runs on the NVIDIA Jetson AGX Orin embedded platform, equipped with a dedicated TensorRT acceleration engine, with a single SOH prediction time of 38 milliseconds, meeting real-time control requirements.

[0100] Example 2:

[0101] In another application scenario, this invention is applied to the collaborative charging management of a 10MWh lithium-ion energy storage power station on the grid side. The system consists of 200 parallel 50kWh battery clusters, each cluster containing 100 individual batteries. The system deploys a unified health status prediction model, but configures an independent distributed charging controller for each cluster.

[0102] In step 1, the local data acquisition unit of each cluster synchronously acquires the voltage, current and temperature of all its internal cells at a frequency of 200Hz, forming a cluster-level time-series characteristic sequence, and uploads it to the central edge server through the fiber optic ring network.

[0103] In step 2, the central server aggregates historical data from all clusters to train a global model, but retains a personalized fine-tuning layer for each cluster to accommodate manufacturing tolerances and differences in local thermal environments.

[0104] In step 3, the charging decision not only considers the state of health (SOH) of a single cluster, but also introduces an equalization objective: defining the maximum difference in the state of health among all individual cells in the system. for:

[0105] ;

[0106] in: This represents the total number of individual cells in the system. The difference between the maximum and minimum SOH values ​​among all K individual cells; : Health status value of the kth individual cell; (·): Function to find the maximum value; (·): The function that takes the minimum value.

[0107] when In this case, an equilibrium term is added to the multi-objective optimization function. ,in These are the preset equilibrium weight coefficients.

[0108] During optimization, lower charging current and voltage are assigned to clusters with lower SOH, while clusters with higher SOH are allowed to be charged close to the rated parameters, thereby actively reducing the difference.

[0109] In step 4, each distributed controller receives a personalized parameter set and independently performs DPWM regulation, while reporting a local SOH estimate every 50 milliseconds.

[0110] In step 5, incremental learning employs a federated learning framework: each cluster calculates its gradient locally, and the central server aggregates the gradients to update the global model, avoiding the uploading of raw data and ensuring data privacy. After six months of testing, the maximum SOH difference within the system was stably controlled within 0.025, significantly better than the 0.065 difference of the traditional current sharing charging strategy, effectively improving the overall available capacity and security of the system.

[0111] Example 3:

[0112] In a home energy storage scenario, this invention is deployed in a wall-mounted charging terminal that supports V2H (Vehicle-to-Home) functionality. The terminal integrates a 7kWh ternary lithium battery pack, which needs to be charged during off-peak hours at night and supports supplying power to household loads during the day. Due to space and cost constraints, the system adopts a lightweight model.

[0113] The CNN-BiLSTM-Attention model in step 2 is compressed into a one-dimensional convolutional backbone network in the style of MobileNetV3. The number of BiLSTM layers is reduced to 1, the number of hidden units is reduced to 256, and the attention mechanism is simplified to channel attention. The number of model parameters is reduced from 1.2M to 0.35M, and it can run on an ARM Cortex-A55-based embedded SoC.

[0114] In step 1, the sampling frequency is reduced to 50Hz, but the time synchronization error is still guaranteed to be less than 20 microseconds.

[0115] In step 3, the multi-objective optimization function is simplified to a dual objective: only the safety boundary and lifetime loss constraints are retained, while the efficiency objective is implicitly defined by the user-defined charging completion time window. For example, if the user sets "fully charged within 6 hours," the system maximizes lifetime protection within this window. In step 5, incremental learning employs a knowledge distillation strategy: the terminal periodically uploads local data summaries (not raw data) to the cloud platform. After the cloud platform trains the large model, it sends soft labels back to the terminal for fine-tuning the small model, achieving "cloud-edge" co-evolution. This solution reduces hardware costs by 40% while ensuring a SOH prediction error of <2%, making it suitable for large-scale civilian deployment.

[0116] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0117] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A battery health state prediction charging method integrating deep learning, characterized in that, Includes the following steps: Step 1: Collect multi-source operating data of the battery in real time during the charging process. Simultaneously acquire the battery's terminal voltage, charging and discharging current, surface temperature, ambient temperature and cumulative charging and discharging amount through high-precision sensors. Perform timestamp alignment and noise filtering on the raw data to form a structured time-series feature sequence. Step 2: Construct and update the deep learning health status prediction model online. Based on the multi-source operation data accumulated in the historical charging cycle and the corresponding health status labels, train a composite neural network model composed of convolutional neural network, bidirectional long short-term memory network and attention mechanism. The composite neural network model takes the time sequence of the current charging cycle as input and outputs the current health status value of the battery and the health status evolution trend of the next few charging cycles. Step 3: Embed the health status prediction results into the charging decision core. Based on the health status values ​​and their evolution trends, combined with preset safety boundaries, efficiency targets and lifespan loss constraints, dynamically generate the optimal charging parameter set for the next stage, including the current amplitude of the constant current stage, the voltage threshold of the constant voltage stage, the triggering conditions for stage switching and the charging termination criteria. Step 4: Execute adaptive charging control, send the optimal charging parameter set to the charging controller, adjust the output characteristics of the charging power supply in real time, and feed back actual operating data in each control cycle for model fine-tuning and parameter correction, forming a closed-loop optimization loop; Step 5: Continuous evaluation and iterative optimization. After each charge, the actual operation data and changes in health status of the entire charging process are included in the model training set. An incremental learning strategy is used to update the deep learning model online to ensure that it maintains high accuracy and strong generalization ability throughout the battery's entire life cycle.

2. The battery health state prediction charging method integrating deep learning according to claim 1, characterized in that, The high-precision sensor includes a voltage sensor, a current sensor, and a temperature sensor. All sensor signals are synchronously sampled via the same data acquisition card. The time synchronization error is less than a preset time synchronization error threshold, and the measurement accuracy of the voltage sensor, current sensor, and temperature sensor is better than the preset voltage accuracy threshold, preset current accuracy threshold, and preset temperature accuracy threshold, respectively.

3. The battery health state prediction charging method integrating deep learning according to claim 1, characterized in that, The health status label was obtained through offline capacity testing.

4. The battery health state prediction charging method integrating deep learning according to claim 1, characterized in that, The convolutional neural network contains several one-dimensional convolutional layers, each configured with the number of convolutional kernels, kernel length, and stride, used to extract local temporal patterns.

5. The battery health state prediction charging method integrating deep learning according to claim 1, characterized in that, The safety boundary includes the maximum allowable charging current, the maximum allowable terminal voltage, and the upper limit of the temperature rise rate. The efficiency target is defined as maximizing the amount of charge charged per unit time. The lifetime loss constraint is modeled as a quadratic function of the health state decay rate. The three are weighted and summed to form a multi-objective optimization function, and the weight coefficients are dynamically adjusted according to the current health state of the battery.

6. The battery health state prediction charging method integrating deep learning according to claim 5, characterized in that, When the health status value is lower than the preset health status threshold, the constant current stage current amplitude is automatically reduced to a predetermined percentage of the rated value, and the constant voltage stage voltage threshold is lowered by a preset voltage adjustment amount. At the same time, the stage switching trigger condition is brought forward.

7. The battery health state prediction charging method integrating deep learning according to claim 1, characterized in that, The charging controller adopts digital pulse width modulation technology, with a response delay less than the preset response delay threshold. It can complete the stable adjustment of output current or voltage within a predetermined time period according to the optimal charging parameter set issued, and has hardware-level protection functions for overvoltage, overcurrent and overtemperature.

8. The battery health state prediction charging method integrating deep learning according to claim 1, characterized in that, The incremental learning strategy employs an elastic weight solidification algorithm, which retains the important parameter weights of the learned aging patterns while adaptively adjusting to changes in feature distribution in new data.

9. The battery health state prediction charging method integrating deep learning according to claim 1, characterized in that, The method supports collaborative charging management of multi-battery parallel systems. By sharing a unified health status prediction model and a distributed charging controller, it achieves balanced regulation of the health status of each individual battery, and controls the maximum health status difference within a preset health status difference threshold.

10. The battery health state prediction charging method integrating deep learning according to claim 9, characterized in that, In a multi-battery parallel system, when the maximum difference in the system's health status exceeds a preset threshold for the difference in health status, an equalization term is added to the multi-objective optimization function. This involves allocating lower charging current and voltage to batteries with lower health status, while allowing batteries with higher health status to be charged close to their rated parameters.

Images