Lithium-ion battery health prediction method and equipment based on machine learning

Abstract

This invention discloses a method and device for predicting lithium battery health based on machine learning, relating to the field of battery health prediction technology. The method inputs the operating data of the lithium battery to be predicted into a Light-GRU model for health prediction, evaluates the noise level, and outputs the lithium battery health prediction result. Based on this, it determines whether to perform fault prediction verification and fault verification feedback optimization. Then, it issues an early warning based on the lithium battery health prediction result and evaluates the accuracy of the model prediction result to decide whether to call the CNN-GRU model. If called, during the model's operation, model efficiency analysis is performed to determine whether battery sequence optimization is needed; otherwise, the Light-GRU model continues to be used. This improves the accuracy of lithium battery health prediction and solves the problem in existing technologies where the similarity between normal battery load fluctuations and initial fault signals leads to a decrease in the accuracy of lithium battery health prediction.

Description

Technical Field

[0001] This invention relates to the field of battery health prediction technology, and in particular to a method and device for predicting the health of lithium batteries based on machine learning. Background Technology

[0002] To predict potential problems in lithium batteries in advance, reduce the probability of failure, and extend the lifespan of lithium batteries, existing technologies often incorporate machine learning for lithium battery health prediction. This has significantly improved prediction accuracy and enabled real-time online monitoring, while also supporting group management and lifecycle optimization. It is a key technology for future intelligent BMS (Battery Management System).

[0003] For example, Chinese invention patent CN116381536B discloses a method and system for predicting the health status of lithium batteries based on regression meta-learning. The method includes: dividing charging data from multiple lithium batteries into multiple training datasets according to charging protocols; each training dataset includes a support set and a query set; establishing a regression meta-learning model and training the model using each training dataset; training includes: calculating the similarity between the features of each query set sample and the features of all support set samples using cosine distance; matching each query set sample with several support set samples based on the similarity to obtain the predicted health status of each query set sample; correcting the regression meta-learning model based on the predicted and actual health status; and predicting the health status of lithium batteries with different charging protocols using the trained regression meta-learning model.

[0004] For example, the Chinese invention patent with announcement number CN112946499B discloses a method for jointly estimating the state of health and state of charge of a lithium battery based on machine learning. This method includes: determining the battery model; fitting a V-SOC curve based on detailed data from the charging and discharging process; establishing an equivalent circuit model for the lithium battery; identifying parameters of the voltage rebound characteristic curve within a charging and discharging cycle to obtain a machine learning model; performing an initialization test upon power-on; fitting the voltage rebound curve within that cycle to obtain the current battery ohmic resistance and polarization resistance, while simultaneously measuring the ambient temperature, reading battery charging and discharging cycle data from the storage chip, and calculating the battery's state of health (SOH); updating the model parameters with the identified ohmic resistance, polarization resistance, polarization capacitance, and maximum available capacity; estimating the battery's SOC using the UKF or EKF algorithm; and recording the SOC value in the storage chip.

[0005] Existing technologies for predicting the health of lithium batteries using machine learning models primarily improve the performance of the machine learning itself and optimize the model building process to enhance the accuracy of the prediction. They also quickly output the health status by inputting real-time collected voltage, temperature, and current sequences.

[0006] Therefore, it is evident that using machine learning models for lithium battery health prediction is both necessary and reliable. The current mainstream implementation process generally involves data acquisition, data preprocessing, feature extraction, input and output model construction, model training, model validation, online prediction deployment, and dynamic updates and optimization. Input and output model construction refers to building the machine learning model. Currently, mainstream models fall into three main categories: traditional machine learning models, deep learning models, and hybrid models. Commonly used deep learning models include Light-GRU (Lightweight Gated Recurrent Unit) or CNN-GRU (Convolutional Neural Network–Gated Recurrent Unit). Online prediction deployment involves integrating the model into the BMS, vehicle system, or energy storage system. The input to online prediction is typically the most recent charge / discharge curves, real-time voltage, current, and other measurement data. The output includes the current SOH (State of Health), Remaining Useful Life (RUL), and a determination of whether alarms and maintenance are needed.

[0007] The above-mentioned technology has at least the following technical problems: Dynamic load and temperature fluctuations can mask abnormal features, making it difficult for models to distinguish between normal fluctuations and initial fault signals, leading to misjudgments. The factors that cause dynamic load and temperature fluctuations to mask abnormal features and make it difficult for models to distinguish between normal fluctuations and initial fault signals mainly come from operating conditions, such as EV (Electric Vehicle) acceleration and regenerative braking, which will cause the current to be in a pulse form and the voltage to drop or rise transiently. These normal fluctuations are very similar to the initial fault features such as "increased internal resistance" and "increased polarization", making it impossible for the model to distinguish between voltage fluctuations caused by normal load and voltage anomalies caused by increased internal resistance. Summary of the Invention

[0008] This invention provides a method and device for predicting the health of lithium batteries based on machine learning. This method enables multiple optimizations for predicting the health of lithium batteries, thereby improving the accuracy of lithium battery health prediction. The technical solution provided by this application is as follows: Firstly, a machine learning-based method for predicting lithium battery health is provided. The specific implementation of this method is as follows: The operating data of the lithium battery to be predicted is monitored through a data acquisition system and input into a Light-GRU model for health prediction. Simultaneously, noise level assessment is performed to determine the impact of noise levels in the operating data on lithium battery fault prediction, and the corresponding lithium battery health prediction result is output. Based on the lithium battery health prediction result, it is determined whether fault prediction verification should be performed, and corresponding fault verification feedback optimization is conducted. The lithium battery health prediction result is transmitted to a preset terminal for early warning, and the accuracy of the Light-GRU model prediction result is evaluated to determine whether to call the CNN-GRU model for lithium battery health prediction. If the CNN-GRU model is called, model efficiency analysis is performed during the CNN-GRU model's operation to determine whether battery sequence optimization is necessary; otherwise, the Light-GRU model continues to be used for lithium battery health prediction.

[0009] Secondly, a machine learning-based lithium battery health prediction device is provided. This device includes: a processor for executing machine learning model inference, data preprocessing, feature extraction, and control logic; a memory for storing the processor's executable instructions, intermediate tensors in the machine learning model inference, machine learning model parameters, program firmware, system configuration, and real-time data; a battery status acquisition module, including sensors and sampling circuits, which is the core data source for machine learning prediction; and a data acquisition and conversion module for simulating front-end circuits, multi-channel ADCs, controlling the sampling rate, synchronizing timestamps, and providing raw signals to the processor. The processor is configured to execute the steps of a machine learning-based lithium battery health prediction method.

[0010] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. The system monitors the operating data of the lithium battery to be predicted through a data acquisition system and inputs it into the Light-GRU model for health prediction. Simultaneously, noise level assessment is performed to determine the impact of noise in the operating data on lithium battery fault prediction. This prevents frequent occurrences of unstable predictions and misjudgments of SOH and RUL due to noise and large fluctuations in battery operating data under real-world conditions. Corresponding lithium battery health prediction results are output, ensuring the stability of predictions under complex conditions such as acceleration, deceleration, and high-power impact. Based on the lithium battery health prediction results, it is determined whether fault prediction verification is necessary to ensure the accuracy of lithium battery fault prediction. Corresponding fault verification feedback optimization is performed, solving the problem that a single prediction result cannot guarantee the reliability of fault warnings. This ensures the prediction stability of the model during long-term operation. Finally, the lithium battery health prediction results are transmitted to a preset terminal for early warning, and the accuracy of the Light-GRU model prediction results is evaluated. Deciding whether to use the CNN-GRU model for lithium battery health prediction solves the problem that traditional systems cannot achieve intelligent early warning and model scheduling, leading to excessive prediction errors or high latency. This helps improve prediction accuracy and overall model precision and stability. If the CNN-GRU model is used, model efficiency analysis is performed during its operation to determine whether battery sequence optimization is necessary. This helps improve the real-time performance of the CNN-GRU model for lithium battery health prediction, preventing issues that may arise when processing high-frequency battery data and reducing invalid sequences without compromising prediction accuracy, thus improving computational speed. Otherwise, the Light-GRU model continues to be used for lithium battery health prediction, improving the accuracy, stability, and reliability of lithium battery health prediction. This also solves the problem that the accuracy of lithium battery health prediction decreases because the normal fluctuations in battery load and the trends of initial fault signals are similar.

[0011] 2. By acquiring noise level data for various operating conditions, and calculating the deviation of each noise level data from the preset battery operating data noise limit, the method compares this deviation with the corresponding noise level error range obtained from the preset database. This prevents the problem of traditional algorithms using a single absolute threshold failing to adapt to noise differences across multiple operating conditions and battery models, achieving adaptive noise baseline correction across batteries or environments. Noise level data whose deviation falls outside the noise level error range is marked as abnormal noise data; otherwise, it is marked as normal noise data. This not only provides a concise and traceable secondary identifier but also facilitates rapid statistics and decision-making. If abnormal noise data exists in the noise level data, it indicates that the current noise level of the operating data may mask initial fault signals. Therefore, a mapping based on the number of abnormal noise data is performed. By obtaining the initial fault masking percentage and simultaneously performing noise masking compensation, the problem of high false negative rates caused by the small amplitude of initial fault signals and their susceptibility to noise in existing technologies is addressed. By mapping the number of high-level indicators to the masking percentage, the degree to which current data may mask faults is quantified, thereby achieving risk quantification. Subsequently, corresponding compensation strategies are automatically triggered based on the masking percentage, which helps reduce the probability of missing early faults. If the noise level data are all normal noise data, it means that the current noise level working data does not have the possibility of masking the initial fault signal, and lithium battery health prediction continues. This avoids the problem of blindly triggering complex compensation or high-cost detection under low noise conditions, which leads to resource waste and delays. It also reduces unnecessary secondary verification or model switching, saves computing and communication resources, and improves the overall efficiency of lithium battery health prediction.

[0012] 3. First, the feedback results after obtaining the prediction results of the Light-GRU model are obtained. Through the feedback verification mechanism, the prediction results of Light-GRU can be automatically verified, erroneous predictions can be detected in time, and the model or prediction method can be adjusted. If the feedback result is that the fault prediction is correct, the working data of the lithium battery to be predicted is continued to be input into the Light-GRU model to predict the lithium battery health. This avoids the computational overhead and delay caused by frequent model switching, ensuring the stability of battery health prediction. In the case of accurate prediction, the model continues to use the Light-GRU model, maintaining efficient prediction and avoiding unnecessary waste of resources. If the feedback result is that the fault prediction is incorrect, the CNN-GRU model is called and the working data of the lithium battery to be predicted is input into the model to predict the lithium battery health. This achieves higher prediction accuracy through a more complex network structure when the fault prediction is incorrect, enhancing the system's ability to identify faults early. At the same time, the efficiency of the CNN-GRU model is analyzed to help to discover in time the problem that the CNN-GRU model may require a lot of computational resources during operation, which may lead to a decrease in real-time performance. By optimizing the battery sequence, the model inference speed is improved, ensuring higher accuracy in lithium battery health prediction and reducing the impact on the real-time performance of lithium battery health prediction. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating the lithium battery health prediction method based on machine learning provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the architecture of the Light-GRU model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the noise level assessment process provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the process of calling the CNN-GRU model for determination provided in an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0016] Before providing a detailed explanation of the embodiments of this application, the application scenarios of these embodiments will be described first.

[0017] Machine learning-based lithium battery health prediction methods are primarily applied in electric vehicles, energy storage systems, portable electronic devices, and industrial-grade battery management systems to proactively identify performance degradation and potential failure risks during battery use. This method analyzes multi-dimensional operating data such as battery voltage, current, temperature, capacity, and internal resistance, utilizing lightweight prediction models and deep learning models to achieve real-time prediction of battery health status and failure risks. It is suitable for battery systems operating under complex conditions, with high data noise, and requiring high safety and real-time performance. This approach can effectively improve battery lifespan, reduce maintenance costs, and ensure system operational safety.

[0018] This invention provides a machine learning-based method for predicting the health of lithium batteries, which can be implemented using a machine learning-based lithium battery health prediction device. Figure 1 The flowchart shown in this embodiment of the invention illustrates a machine learning-based lithium battery health prediction method. The processing flow of this method may include the following steps: The system monitors the operating data of the lithium battery to be predicted through a data acquisition system and inputs it into the Light-GRU model for health prediction. At the same time, the noise level is assessed to determine the impact of the noise level of the operating data on the lithium battery fault prediction. The corresponding lithium battery health prediction results are output, including but not limited to fault prediction results, SOH prediction, and RUL prediction.

[0019] The data acquisition system includes, but is not limited to, voltage sampling units, current sampling units, temperature sampling units, and auxiliary sensors. The voltage sampling unit is used to collect the voltage of individual cells and module voltage (high-precision ADC). The current sampling unit includes Hall current sensors, shunt resistors, amplifier circuits, etc. The temperature sampling unit includes NTC and PTC thermistors, etc. The auxiliary sensors include ambient temperature sensors, pressure or gas sensors (for early bulging or venting monitoring), and acceleration sensors, etc.

[0020] Based on the lithium battery health prediction results, determine whether to conduct fault prediction verification to ensure the accuracy of lithium battery fault prediction, and perform corresponding fault verification feedback optimization.

[0021] The lithium battery health prediction results are transmitted to a preset terminal for early warning, and the accuracy of the Light-GRU model prediction results is evaluated to determine whether to call the CNN-GRU model for lithium battery health prediction.

[0022] It's important to explain that the Light-GRU model is a lightweight time-series prediction model derived from the traditional GRU (Gated Recurrent Unit) through structural simplification and parameter compression. This model significantly reduces computational complexity and inference latency while maintaining the time-series modeling capabilities of GRUs by reducing the number of parameters in the gating structure, optimizing the hidden layer size, and employing more efficient computational methods. Light-GRU is particularly suitable for embedded BMS, automotive terminals, and real-time prediction scenarios. It can quickly process dynamic sequence data such as battery voltage, current, and temperature, enabling rapid prediction of SOH, RUL, and fault risk, while maintaining good robustness and real-time performance under high noise and complex operating conditions.

[0023] like Figure 2 The diagram shows the architecture of the Light-GRU model provided in this embodiment of the invention. It illustrates a battery state prediction model framework based on a GRU (Gated Recurrent Unit). The core of this model is the use of a GRU to process the battery's time-series data, enabling tasks such as SOH (State of Health), RUL (Relative Lifetime), and fault detection. The input is the battery's time-series features (dimension T×F), including parameters such as voltage (V), temperature (T), current (I), and SOC (State of Charge) that change over time. The output includes three categories: SOH prediction (outputting a percentage representing the battery's health status), RUL prediction (outputting a numerical value representing the battery's remaining lifespan), and fault detection (outputting a multi-class probability vector representing the probability distribution of different fault types). The GRU in the diagram is a type of recurrent neural network that uses "gates" to control information transmission. The core gates here are the Update Gate and the Reset Gate. The first gate (using the Sigmoid function) determines whether to forget the previous state. It multiplies the previous hidden state to filter out historical information that needs to be retained. The second gate (using the tanh function) combines the historical information processed by the reset gate with the current input features to generate the current candidate state (temporary state). The third gate (using the Sigmoid function) determines how many historical states to retain and how many candidate states to add. It multiplies the previous hidden state (to retain history) and multiplies the update gate with the candidate state (to add new information). Finally, the two gates are summed to obtain the new hidden state at the current moment, thus enabling health prediction of lithium batteries.

[0024] The CNN-GRU model integrates a convolutional neural network (CNN) with a gated recurrent unit (GRU), combining the local feature extraction capability of CNN with the temporal memory capability of GRU to form a high-precision hybrid deep learning model. The CNN part is used to extract key morphological features (such as trends, local fluctuations, and implicit degradation information) from the battery charge-discharge curve, while the GRU part is used to capture long-term dependencies in the battery aging process. Therefore, CNN-GRU has a stronger modeling capability for nonlinear degradation patterns under complex dynamic conditions, making it suitable for high-precision health prediction, early fault identification, and battery scenarios with high noise or drastic changes. However, it has a relatively high computational cost and is more suitable for situations where prediction accuracy is the priority.

[0025] If the CNN-GRU model is used, model efficiency analysis is performed during the CNN-GRU model's operation to determine whether battery sequence optimization is necessary, thereby improving the real-time performance of the CNN-GRU model in predicting lithium battery health. Otherwise, the Light-GRU model is used for lithium battery health prediction.

[0026] In this embodiment, by real-time acquisition of lithium battery operating data and inputting it into the Light-GRU model for health prediction, rapid prediction of key health indicators such as faults, state of health (SOH), and raw uptime (RUL) can be achieved under lightweight computing conditions, effectively meeting the requirements of battery management systems for real-time performance and low computational cost. Simultaneously, by incorporating a noise level assessment mechanism, the degree to which multi-dimensional noise such as voltage, current, and temperature masks early fault signals can be identified before prediction, thereby improving the effectiveness of input data and avoiding misjudgments caused by noise interference. The fault verification and feedback mechanism based on the preliminary health prediction results enables the model to dynamically calibrate its prediction accuracy, achieving continuous improvement of the prediction strategy. Furthermore, by transmitting the prediction results to a preset terminal for early warning, potential risks can be alerted in advance, improving system safety. When the prediction accuracy of the Light-GRU model decreases, the system automatically switches to the CNN-GRU model, utilizing convolutional feature extraction capabilities to improve prediction accuracy under complex operating conditions. Model efficiency analysis is performed during the operation of the CNN-GRU model, and battery sequence optimization helps improve its real-time performance and computational efficiency. The method provided in this embodiment combines the advantages of lightweight, robustness, and high precision, enabling more reliable lithium battery health prediction under complex operating conditions and noisy environments, thus helping to improve the safety and lifespan of the battery system.

[0027] It should be added that the working data includes voltage data, current data, rate and power data, and impedance data; voltage data includes transient voltage change amplitude, voltage change rate, voltage recovery time, and voltage noise; current data includes current change, current change rate, and current pulse frequency; rate and power data includes real-time discharge rate and power; impedance data includes internal resistance estimate and dynamic polarization value.

[0028] like Figure 3 The diagram shown is a flowchart illustrating the noise level assessment process provided in this embodiment of the invention. The specific logic is as follows: Noise level data for various operating data types are acquired; the noise level data is then compared with the noise limit value of the battery operating data to obtain the deviation of each noise level, and this deviation is compared with the corresponding noise level error range. Noise level data whose deviation is outside the noise level error range is marked as abnormal noise data; otherwise, it is marked as normal noise data. If abnormal noise data exists, the initial fault masking percentage is obtained based on the number of abnormal noise data, and noise masking compensation is performed simultaneously. If all noise level data are normal noise data, lithium battery health prediction continues. Through the above process, the accuracy and reliability of fault prediction, SOH prediction, and RUL prediction for lithium batteries are improved.

[0029] As a further embodiment, the specific procedure for noise level assessment is as follows: The first step is to obtain noise level data for various types of work data, including voltage noise, internal resistance estimation error, temperature measurement error, and current noise.

[0030] Specifically, early anomalies in lithium batteries (such as slight increases in internal resistance, increased polarization, and micro-short circuits) typically manifest as changes in voltage signals, internal resistance, and temperature rise. These minute amplitudes are easily masked by noise. Therefore, voltage noise can directly drown out early voltage anomalies, and the internal resistance estimation error is equivalent to the amplitude of early faults. This makes it impossible to distinguish between normal temperature-induced changes and fault-induced changes. Temperature measurement errors cause internal resistance to change with temperature; for every 1°C decrease, the internal resistance can increase by 1–3 mΩ, which is the same as the amplitude of the fault signal. This may mask early fault signals. Current noise can cause voltage estimation distortion, rendering internal resistance calculation meaningless. Voltage noise is represented by the voltage standard deviation, internal resistance estimation error is represented by the internal resistance fluctuation standard deviation, temperature measurement error is represented by the temperature drift rate, and current noise is represented by the current standard deviation.

[0031] The second step is to perform deviation calculations on the noise level data and the preset battery operating data noise limit values ​​to obtain the deviation of each noise level. This involves calculating the ratio between the difference between the noise level data and the battery operating data noise limit value and the battery operating data noise limit value, and then comparing it with the corresponding noise level error range obtained from the preset database.

[0032] The noise level error range is preset by staff and stored in a preset database. It is generally based on data from historical time periods and empirical rules.

[0033] The third step is to mark noise level data whose deviation is outside the noise level error range as abnormal noise data, and otherwise mark them as normal noise data.

[0034] The fourth step is to find out if there are abnormal noise data in the noise level data. This means that the current noise level data may have the potential to mask the initial fault signal. The percentage of initial fault masking is then obtained based on the number of abnormal noise data, and noise masking compensation is performed at the same time.

[0035] Specifically, the number of noise anomaly data is input into the noise masking compensation map set, and the corresponding initial fault masking percentage is output. The noise masking compensation map set is used to fit the mapping relationship between the number of noise anomaly data and the initial fault masking percentage. The initial dataset constructed by the logistic regression algorithm, as well as the least squares criterion and the statsmodels framework, are used to complete the training to obtain the noise masking compensation map set. The training data used are the number of noise anomaly data obtained in the historical time period and the initial fault masking percentage set according to empirical rules.

[0036] Fifth, if all noise level data are within normal range, it means that the current noise level data does not have the potential to mask initial fault signals, and the lithium battery health prediction will continue.

[0037] In this embodiment, by quantifying and evaluating the noise levels of multi-dimensional working data such as voltage noise, internal resistance estimation error, temperature measurement error, and current noise, the quality of battery sampling data can be effectively identified before health prediction. Simultaneously, by calculating the deviation between the noise level and a preset noise limit value, and comparing it with the noise error range in the database, refined identification of different noise sources is achieved. Marking data with deviations within the error range as high-noise levels helps to more quickly locate key data that may mask early fault characteristics, improving the ability to judge the sensitivity of initial fault signals. Furthermore, the percentage of initial fault masking is mapped based on the quantity of abnormal noise data, quantifying the noise impact into an assessable risk level and automatically triggering corresponding noise masking compensation. This helps reduce the risk of missed fault detection due to noise interference. If the noise level is low, health prediction can be performed directly, improving the efficiency and real-time performance of the overall prediction process. Overall, this noise judgment and compensation mechanism enhances the robustness of the health prediction system in complex operating conditions and high-noise environments, improving the accuracy and reliability of fault prediction, SOH prediction, and RUL prediction.

[0038] Furthermore, the specific details of noise masking compensation are as follows: The ratio of the number of abnormal noise data to the total number of noise level data is used to obtain the corresponding percentage of noise exceeding the standard.

[0039] Based on the percentage of noise exceeding the standard, a mapping is performed in the noise masking compensation mapping table to obtain a compensation data set, which includes temperature compensation value and sampling frequency compensation value.

[0040] It should be added that the proportion of noise exceeding the standard is input into the noise masking compensation mapping table after training, and the corresponding compensation data group is obtained. The noise masking compensation mapping table is used to reflect the correlation between the proportion of noise exceeding the standard and the compensation data group. The proportion of noise exceeding the standard in the historical time period, as well as the compensation data group set by professionals based on experience rules, are input into the initial data table constructed by the logistic regression algorithm. The table is trained based on the least squares criterion and the statsmodels framework to obtain the corresponding noise masking compensation mapping table.

[0041] The temperature compensation and sampling frequency compensation are calculated by performing compensation operations on the temperature compensation and sampling frequency respectively, which means that the temperature compensation and sampling frequency compensation are multiplied by the temperature compensation and sampling frequency respectively to obtain the adjusted temperature compensation and the adjusted sampling frequency.

[0042] Temperature compensation and data sampling are performed by adjusting the temperature compensation and the sampling frequency, respectively.

[0043] In this embodiment, the method can quantify noise issues in battery operating data and provide a basis for noise masking compensation. By combining the proportion of noise exceeding the standard with a noise masking compensation mapping table, temperature compensation values ​​and sampling frequency compensation values ​​can be mapped according to the actual noise level. This provides dynamic adjustment parameters for subsequent compensation calculations, helping to solve the problem of initial fault signal masking caused by high noise levels, and improving data reliability and prediction accuracy. Specifically, temperature compensation can eliminate measurement errors caused by temperature changes, while adjusting the sampling frequency can reduce the impact of noise on data accuracy, ensuring more accurate and detailed information during the sampling process. By compensating and sampling the operating data with adjusted temperature compensation and sampling frequency, the data becomes more stable and accurate, thereby improving the accuracy and robustness of lithium battery health prediction and reducing prediction errors caused by noise.

[0044] Furthermore, based on the lithium battery health prediction results, it is determined whether to conduct fault prediction verification. The specific process is as follows: If the output fault prediction result indicates that the lithium battery has an initial fault risk, then obtain the operating condition description parameters of the lithium battery to be predicted during operation. The operating condition description parameters include temperature, SOC value, and current load.

[0045] The operating condition description parameters are compared with the reference operating condition parameters read from the preset database. The reference operating condition parameters are preset values, which are generally set by preset operators. Specifically, they include the operating temperature range, the reference operating condition SOC value, and the reference load error range.

[0046] If all the operating condition description parameters and the corresponding reference operating condition parameters meet the fault verification conditions, it means that the fault prediction results output by the Light-GRU model have been verified and the corresponding initial fault masking percentage and the lithium battery health prediction results output by the Light-GRU model are fed back together.

[0047] Specifically, the fault verification conditions include the temperature not being within the operating temperature range, the SOC value being lower than the reference operating condition SOC value, and the current load not being within the reference load error range.

[0048] If there are operating condition description parameters that do not meet the fault verification conditions, it means that the fault prediction results output by the Light-GRU model have failed verification. This may be due to normal fluctuations in dynamic load, and fault verification feedback optimization should be performed.

[0049] In this embodiment, introducing a working condition verification mechanism after fault prediction helps improve the accuracy and robustness of initial fault identification. When the Light-GRU model outputs that the battery has an initial fault risk, the working condition description parameters are first obtained and compared item by item with the reference working condition parameters in the preset database. This comparison process helps to identify whether the prediction result is affected by extreme working conditions, dynamic load fluctuations, or short-term disturbances, avoiding misjudgments caused by abnormal working conditions. If there are parameters that do not meet the verification conditions, it indicates that the fluctuation may be caused by normal dynamic working conditions rather than actual faults. At this time, the fault verification feedback optimization process is automatically triggered to adjust the sensitivity of the prediction model and avoid over-responding to normal load changes. Through this mechanism, a dual verification system is built between model prediction and actual working conditions, which not only effectively reduces the false alarm rate caused by noise and load disturbances, but also improves the credibility of initial battery fault detection. Therefore, this method realizes dynamic fault verification based on working condition consistency, enabling the health prediction system to have higher stability, reliability, and engineering adaptability in complex working conditions and dynamic load environments.

[0050] Furthermore, the specific process for fault verification feedback optimization is as follows: The initial fault coverage percentage is matched with the preset fault coverage level range read from the preset database. The preset fault coverage level range includes a first coverage risk level range, a second coverage risk level range, and a third coverage risk level range. The risk of the initial fault being covered gradually decreases. The preset fault coverage level range is generally set by preset staff based on historical data and empirical rules within a historical time period.

[0051] If the initial fault masking percentage falls within the first or third masking risk level range, and the fault prediction verification result indicates normal fluctuations in dynamic load, then the initial fault may be masked or a false alarm may occur. In this case, a warning will be issued according to the preset high feedback warning frequency.

[0052] If the initial fault masking percentage falls within the second masking risk level range, and the fault prediction verification result shows normal fluctuations in dynamic load, then it indicates that the current initial fault has a certain possibility of being masked, and an early warning will be issued according to the preset low feedback warning frequency.

[0053] Among them, the high feedback warning frequency indicates a higher level of urgency in the warning, while the low feedback warning frequency corresponds to a lower level of urgency in the warning. Both are read from a preset database and are generally set in advance by preset staff.

[0054] In this embodiment, by combining the initial fault masking percentage with preset fault masking level ranges (high, medium, and third-level masking risk ranges), the risk level of battery fault masking is quantified more accurately. Adjusting the warning frequency according to this risk level helps improve the responsiveness and accuracy of the fault detection system. When the initial fault masking percentage falls into the high or third-level masking risk range, and the fault verification result shows normal fluctuations in dynamic load, the system is more likely to identify fault masking. At this time, a high-feedback warning frequency is triggered, quickly issuing a high-urgency warning to prompt operators to conduct in-depth inspections or take appropriate measures. If the initial fault masking percentage is in the second-level masking risk range, it indicates a lower risk of fault masking, but there is still a certain possibility. At this time, a low-feedback warning frequency is triggered, issuing a warning with a lower urgency level to ensure timely response without excessively interfering with normal operation. By dynamically adjusting the warning frequency, early fault identification is ensured while avoiding misleading warnings and resource waste caused by excessive warnings. Therefore, this method not only improves the intelligence and flexibility of fault warnings but also enhances the adaptability and accuracy of the battery management system, especially in complex operating conditions and load variations, ensuring timely detection and response to fault risks.

[0055] like Figure 4 The diagram illustrates the process of calling the CNN-GRU model for judgment according to an embodiment of the present invention. The specific logic is as follows: after obtaining the prediction result of the Light-GRU model, the feedback result includes whether the fault prediction is correct or incorrect; if the feedback result is that the fault prediction is correct, the working data of the lithium battery to be predicted is input into the Light-GRU model to predict the health of the lithium battery; if the feedback result is that the fault prediction is incorrect, the CNN-GRU model is called and the working data of the lithium battery to be predicted is input into the model to predict the health of the lithium battery, while the efficiency of the CNN-GRU model is analyzed. Through the above process, not only is the adaptive switching between lightweight and complex models realized, but the prediction reliability of the battery management system is also improved.

[0056] Furthermore, the accuracy of the Light-GRU model's prediction results is evaluated to determine whether to use the CNN-GRU model for lithium battery health prediction. The specific method is as follows: The feedback results obtained after obtaining the prediction results of the Light-GRU model include whether the fault prediction is correct or incorrect, which are used to verify the accuracy of the lithium battery fault prediction results output by the Light-GRU model. The feedback results are obtained from the human terminal. Generally, the preset staff will perform detection and maintenance based on the warning situation and then input the corresponding results into the human terminal.

[0057] If the feedback result indicates that the fault prediction is correct, the operating data of the lithium battery to be predicted will continue to be input into the Light-GRU model to predict the health of the lithium battery.

[0058] If the feedback result indicates a fault prediction error, the CNN-GRU model is invoked and the working data of the lithium battery to be predicted is input into the model to perform lithium battery health prediction. At the same time, the efficiency of the CNN-GRU model is analyzed.

[0059] In this embodiment, after the Light-GRU model completes lithium battery fault prediction, a feedback result verification mechanism is introduced. By judging two types of feedback—"fault prediction correct" and "fault prediction incorrect"—dynamic evaluation of the model's prediction accuracy is achieved. Furthermore, when the prediction result is verified as correct, the system continues to use the Light-GRU model for lithium battery health prediction, maintaining its prediction efficiency under low computational load conditions and ensuring the speed and continuity of the health assessment process. When the feedback result indicates a prediction error, the system automatically calls the higher-performance, more feature-extraction-capable CNN-GRU model for health assessment. The method provides a model collaborative prediction framework based on feedback verification, achieving a balance between performance and accuracy, and effectively improving the intelligence level and prediction reliability of the battery management system under real-world conditions. Furthermore, it analyzes the efficiency of the CNN-GRU model's operation, enabling adaptive switching between lightweight and complex models. This mechanism not only improves the overall accuracy of fault and health predictions but also maintains higher computational efficiency and reliability under different load environments and noise conditions. Overall, the method provided in this embodiment constructs a model collaborative prediction framework based on feedback verification, achieving a balance between performance and accuracy, and effectively improving the intelligence level and prediction reliability of the battery management system under real-world conditions.

[0060] Furthermore, the specific process for performing efficiency analysis on the CNN-GRU model is as follows: First, the analysis time is obtained by monitoring the time taken by the CNN-GRU model to make health predictions for each lithium battery to be predicted in real time, and the average analysis time taken by the Light-GRU model to make health predictions for all lithium batteries to be predicted is used as the reference analysis time.

[0061] Next, the analysis time of the CNN-GRU model is compared with the reference analysis time to obtain the model analysis time efficiency ratio.

[0062] Finally, if the model analysis time-efficiency ratio is higher than the time-efficiency threshold used to reflect the excessive health prediction time, battery sequence optimization is performed; otherwise, health prediction is continued for each lithium battery to be predicted based on the CNN-GRU model.

[0063] It should be explained that the time efficiency threshold is extracted from the preset database, and is generally preset by the preset staff based on historical data and experience rules within a historical time period, and is entered into the preset database for storage for later use.

[0064] In this embodiment, a dynamic model optimization mechanism based on operational efficiency is introduced by comparing the real-time analysis time of the CNN-GRU model with the reference analysis time of the Light-GRU model. First, the actual analysis time required by the CNN-GRU model to predict the health of each lithium battery to be predicted is recorded in real time, and the average analysis time of the Light-GRU model under the same task is calculated as the reference analysis time. By calculating the ratio of the two, the model analysis time efficiency ratio is obtained, which can more accurately quantify the operational efficiency of high-complexity models under the current operating conditions. When the efficiency ratio is higher than a preset time efficiency threshold, it indicates that the prediction time of the CNN-GRU model has significantly deviated from the normal range, and further optimization is needed. This can lead to prediction delays, excessive computational resource consumption, or task backlog. In such cases, battery sequence optimization is automatically triggered to adjust the task execution order or reallocate computational resources, preventing the decline in model efficiency from affecting the overall prediction process. If the efficiency ratio does not exceed the threshold, it indicates that the CNN-GRU model is running stably and can continue to complete health predictions under this model. Furthermore, by introducing model time monitoring, efficiency comparison, and dynamic optimization strategies, this method effectively improves the operating efficiency, response speed, and system scalability of the battery health prediction system in multi-battery, multi-task real-time monitoring scenarios. At the same time, this mechanism ensures the stability and reliability of health predictions under complex working conditions or high load conditions, avoiding prediction delays caused by large model computation.

[0065] Furthermore, the specific steps for battery sequence optimization are as follows: Preset battery sequence features are extracted from a preset database. The preset battery sequence features include the preset number of convolutional layers and the preset battery sequence calculated convolution. The preset battery sequence features are preset data set by preset staff. Typically, the preset number of convolutional layers uses 1×3 or 1×5 one-dimensional convolution. The preset battery sequence calculated convolution is usually depthwise conv or pointwise conv.

[0066] The initial battery sequence features of the CNN-GRU model are replaced with preset battery sequence features to reduce the size of the CNN convolution kernel. The initial battery sequence features include the number of convolutional layers of the initial battery sequence and the convolution calculated from the initial battery sequence.

[0067] The CNN-GRU model after the initial battery sequence feature replacement is called the optimized CNN-GRU model. The real-time working data is input into the optimized CNN-GRU model to continue health prediction, so as to output the corresponding lithium battery health prediction results.

[0068] In this embodiment, by introducing preset battery sequence features, the structure of the CNN-GRU model is dynamically optimized, which helps improve the model's efficiency and adaptability in real-time health prediction tasks. Battery sequence features, including a preset number of convolutional layers and a preset number of computational convolutions, are extracted from a preset database. These preset features then replace the initial battery sequence features of the CNN-GRU model, helping to reduce the size and computational depth of the CNN convolutional kernels. Furthermore, this structural adjustment significantly reduces the number of model parameters and the computational complexity of convolutions, improving the model's running speed and resource utilization efficiency while maintaining feature extraction capabilities. The model after replacement is designated as the optimized CNN-GRU model. This model can continue to perform health predictions after receiving real-time working data, outputting more efficient and stable prediction results. Through this optimization mechanism, not only is the computational bottleneck caused by excessively large convolution size avoided under high load or multi-battery parallel prediction conditions, but the model's ability to respond quickly to data under different operating conditions is also enhanced. In summary, this method achieves structural dynamic simplification and adaptive optimization of deep learning models, improving the real-time performance, scalability, and overall system performance of lithium battery health prediction, and providing more reliable and efficient prediction support for battery management in complex operating scenarios.

[0069] This invention also provides a machine learning-based lithium battery health prediction device, including a processor for executing machine learning model inference, data preprocessing, feature extraction, and control logic; a memory for storing the processor's executable instructions, intermediate tensors in the machine learning model inference, machine learning model parameters, program firmware, system configuration, and real-time data; a battery status acquisition module, including sensors and sampling circuits, which is the core data source for machine learning prediction; and a data acquisition and conversion module for simulating front-end circuits, multi-channel ADCs, controlling the sampling rate, synchronizing timestamps, and providing raw signals to the processor. The processor is configured to execute the steps of a machine learning-based lithium battery health prediction method. The CNN-GRU model after feature replacement is used to optimize the CNN-GRU model, and the real-time working data is input into the optimized CNN-GRU model to continue health prediction, outputting the corresponding lithium battery health prediction result.

[0070] The above-disclosed embodiments are merely some examples of the present invention and should not be construed as limiting the scope of the present invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A lithium battery health prediction method based on machine learning, characterized in that, The method includes: The system monitors the operating data of the lithium battery to be predicted through a data acquisition system and inputs it into the Light-GRU model for health prediction. At the same time, the noise level is assessed to determine the impact of the noise level of the operating data on the lithium battery failure prediction, and the corresponding lithium battery health prediction results are output. Based on the lithium battery health prediction results, determine whether to perform fault prediction verification and perform corresponding fault verification feedback optimization. The lithium battery health prediction results are transmitted to a preset terminal for early warning, and the accuracy of the Light-GRU model prediction results is evaluated to determine whether to call the CNN-GRU model for lithium battery health prediction. If the CNN-GRU model is called, model efficiency analysis is performed during the operation of the CNN-GRU model to determine whether battery sequence optimization is necessary; otherwise, the Light-GRU model is used to predict the health of lithium batteries. The specific procedure for assessing noise levels is as follows: Acquire noise level data for various types of working data, including voltage noise, internal resistance estimation error, temperature measurement error, and current noise; The noise level data is calculated by deviating from the preset battery operating noise limit value to obtain the deviation of each noise level, and then compared with the corresponding noise level error range. Noise level data whose deviation is outside the noise level error range is marked as abnormal noise data; otherwise, it is marked as normal noise data. If there are abnormal noise data in the noise level data, the initial fault masking percentage is obtained by mapping the number of abnormal noise data, and noise masking compensation is performed at the same time. If all noise level data are within normal range, then continue with the lithium battery health prediction.

2. The lithium battery health prediction method based on machine learning as described in claim 1, characterized in that, The working data includes voltage data, current data, multiplier and power data, and impedance data; The voltage data includes the transient change amplitude of the terminal voltage, the rate of change of voltage, the voltage recovery time, and the voltage noise. The current-related data includes the amount of current change, the rate of current change, and the frequency of current pulses. The rate and power data include real-time discharge rate and power; The impedance data includes internal resistance estimates and dynamic polarization values.

3. The lithium battery health prediction method based on machine learning as described in claim 1, characterized in that, The specific details of the noise masking compensation are as follows: The ratio of the number of abnormal noise data to the total number of noise level data is used to obtain the corresponding percentage of noise exceeding the standard. Based on the proportion of noise exceeding the standard, the noise masking compensation mapping table is used to obtain the temperature compensation value and the sampling frequency compensation value. Based on the temperature compensation value and the sampling frequency compensation value, compensation calculations are performed on the temperature compensation and the sampling frequency respectively to obtain the adjusted temperature compensation and the adjusted sampling frequency. Temperature compensation and data sampling are performed by adjusting the temperature compensation and the sampling frequency, respectively.

4. The lithium battery health prediction method based on machine learning as described in claim 1, characterized in that, The specific process for determining whether to perform fault prediction verification based on lithium battery health prediction results is as follows: If the output fault prediction result indicates that the lithium battery has an initial fault risk, then obtain the operating condition description parameters of the lithium battery to be predicted when it is working. The operating condition description parameters include temperature, SOC value and current load. The operating condition description parameters are compared with the read reference operating condition parameters: If all the operating condition description parameters and the corresponding reference operating condition parameters meet the fault verification conditions, it means that the fault prediction results output by the Light-GRU model have been verified and the corresponding initial fault masking percentage and the lithium battery health prediction results output by the Light-GRU model are fed back together. If there are operating condition description parameters that do not meet the fault verification conditions, it means that the fault prediction results output by the Light-GRU model have failed verification, and fault verification feedback optimization will be performed.

5. The lithium battery health prediction method based on machine learning as described in claim 4, characterized in that, The specific process for fault verification feedback optimization is as follows: The initial fault coverage percentage is matched with the read preset fault coverage level range, which includes a first coverage risk level range, a second coverage risk level range, and a third coverage risk level range, representing a gradual decrease in the risk of the initial fault being covered. If the initial fault coverage percentage falls within the first or third coverage risk level range, and the fault prediction verification result indicates normal fluctuations in dynamic load, then an early warning will be issued according to the preset high feedback warning frequency. If the initial fault coverage percentage falls within the second coverage risk level range, and the fault prediction verification result indicates normal fluctuations in dynamic load, an early warning will be issued according to the preset low feedback warning frequency. The urgency level of the warning corresponding to the low feedback warning frequency is lower than that of the high feedback warning frequency.

6. The lithium battery health prediction method based on machine learning as described in claim 1, characterized in that, The accuracy evaluation of the Light-GRU model prediction results to determine whether to use the CNN-GRU model for lithium battery health prediction is performed as follows: Feedback results are obtained after the Light-GRU model prediction results are obtained, including whether the fault prediction is correct or incorrect. If the feedback result indicates that the fault prediction is correct, then continue to input the working data of the lithium battery to be predicted into the Light-GRU model to predict the health of the lithium battery. If the feedback result indicates a fault prediction error, the CNN-GRU model is invoked, and the operating data of the lithium battery to be predicted is input into the model to perform lithium battery health prediction. At the same time, the efficiency of the CNN-GRU model is analyzed.

7. The lithium battery health prediction method based on machine learning as described in claim 6, characterized in that, The specific process for performing efficiency analysis on the CNN-GRU model is as follows: The analysis time is obtained by real-time monitoring of the time taken by the CNN-GRU model to make health predictions for each lithium battery to be predicted, and the average analysis time of the Light-GRU model to make health predictions for all lithium batteries to be predicted is used as the reference analysis time. The ratio of the analysis time of the CNN-GRU model to the reference analysis time is calculated to obtain the model analysis time efficiency ratio. If the model analysis time-efficiency ratio is higher than the time-efficiency threshold used to reflect the excessive health prediction time, then battery sequence optimization is performed; otherwise, the health prediction of each lithium battery to be predicted continues based on the CNN-GRU model.

8. The lithium battery health prediction method based on machine learning as described in claim 7, characterized in that, The specific steps for battery sequence optimization are as follows: Extract features from a preset battery sequence, the preset battery sequence features including the preset number of convolutional layers in the battery sequence and the preset battery sequence convolution calculation; The initial battery sequence features of the CNN-GRU model are replaced with preset battery sequence features, which include the number of convolutional layers in the initial battery sequence and the initial battery sequence convolution. The CNN-GRU model after the initial battery sequence feature replacement is called the optimized CNN-GRU model. The real-time working data is input into the optimized CNN-GRU model to continue health prediction, so as to output the corresponding lithium battery health prediction results.

9. A lithium battery health prediction device based on machine learning, employing the lithium battery health prediction method based on machine learning as described in any one of claims 1-8, characterized in that, The machine learning-based lithium battery health prediction device includes: The processor is used to perform machine learning model inference, data preprocessing, feature extraction, and control logic. Memory is used to store the processor's executable instructions; Battery status acquisition module, including sensors and sampling circuits; The data acquisition and conversion module is used to simulate the front-end circuit, multi-channel ADC, control the sampling rate, synchronize the timestamp, and provide the raw signal to the processor; The processor is configured to perform the steps of the machine learning-based lithium battery health prediction method according to any one of claims 1 to 8.

Images