Battery charging and discharging method and system based on large current test

Abstract

The application provides a battery charge-discharge method and system based on a large current test, and relates to the technical field of battery testing. The method dynamically generates a large current test waveform through a reinforcement learning agent, and synchronously collects acoustic emission signals and volume expansion data of the battery. The system analyzes acoustic emission characteristics to identify thermal runaway precursors, and judges lithium extraction by comparing real-time expansion rate with expected rate. The detection result is fed back to the agent in real time, which is used to adaptively adjust the test waveform parameters, so as to adaptively apply test stress within the safety boundary. The application combines multi-dimensional physical perception and intelligent decision-making, and balances the test efficiency and safety.

Description

Technical Field

[0001] This invention relates to the field of battery testing technology, and in particular to a battery charging and discharging method and system based on high current testing. Background Technology

[0002] In order to accurately assess and verify the performance boundaries and safety margins of batteries under extreme operating conditions, high-current testing (such as high-rate charge and discharge, fast charging, high-power pulse testing, etc.) has become an indispensable key link in battery research and development, design verification and quality control.

[0003] However, traditional high-current testing methods have significant limitations. These methods typically employ preset, fixed test conditions (such as constant current, constant voltage, or standardized dynamic conditions). Due to individual differences in battery performance and the complex evolution of their internal electrochemical states during aging, fixed test conditions are often too conservative, failing to fully unlock the battery's potential performance, resulting in lengthy testing cycles and low R&D efficiency. Conversely, if the test conditions are too aggressive, especially under high-current stress, the internal polarization of the battery intensifies, easily inducing side reactions such as lithium plating and internal short circuits, which can lead to thermal runaway and serious safety risks.

[0004] Specifically, during high-current charging, the negative electrode potential drops rapidly. When the negative electrode potential falls below the deposition potential of lithium metal, lithium ions will deposit on the negative electrode surface to form metallic lithium, a process known as lithium plating. Lithium plating not only leads to capacity decay and increased internal resistance, but the formed lithium dendrites may also puncture the separator, causing an internal short circuit. Simultaneously, if the Joule heat generated by the high current cannot dissipate in time, it will cause the battery temperature to rise rapidly, accelerating a series of exothermic side reactions and ultimately leading to a chain reaction and thermal runaway.

[0005] To monitor these risks, traditional battery management systems (BMS) primarily rely on external macroscopic parameters such as voltage, current, and surface temperature. However, these parameters often lag behind changes in the battery's internal microstructure. By the time an abnormal temperature rise or voltage fluctuation is detected, the internal failure process may have already progressed to an irreversible stage. Therefore, there is an urgent need to develop advanced sensing technologies capable of real-time, in-situ, and non-destructive monitoring of the battery's internal state.

[0006] In recent years, acoustic emission (AE) technology and volume expansion measurement technology have shown great potential in battery research. AE technology, by monitoring the elastic waves generated by stress release in the battery's internal materials, can sensitively capture microscopic physical changes associated with precursors of thermal runaway, such as electrode particle rupture and gas generation. Volume expansion measurement, by monitoring changes in battery thickness, can effectively assess the battery's health and identify anomalies, particularly abnormal expansion caused by lithium plating.

[0007] While these advanced sensing technologies provide a wealth of information, they still present challenges in practical applications. On one hand, accurately extracting characteristic parameters representing precursors to thermal runaway and lithium plating from complex acoustic emission signals and expansion data, and setting reasonable safety thresholds, remains a difficult problem. On the other hand, a more critical challenge lies in how to utilize this real-time monitoring information to dynamically adjust test conditions, maximizing test stress while ensuring safety, thereby accelerating the testing process. Most existing test systems rely on simple threshold control logic, lacking adaptive adjustment capabilities and unable to perform refined exploration and optimization near safety boundaries.

[0008] Therefore, there is an urgent need for an intelligent high-current testing method and system that can integrate multi-dimensional physical field perception information, identify the risk state inside the battery in real time, and dynamically optimize the test waveform through intelligent decision-making algorithms, so as to maximize testing efficiency and testing depth while ensuring testing safety. Summary of the Invention

[0009] One of the objectives of this invention is to provide a battery charging and discharging method and system based on high-current testing, which aims to solve the problems in the prior art where it is difficult to balance efficiency and safety in high-current battery testing, and the lack of real-time monitoring and adaptive closed-loop control of the battery's internal failure mechanism.

[0010] In a first aspect, embodiments of the present invention provide a battery charging and discharging method based on high-current testing, comprising: The test waveform is dynamically generated and applied to the target battery by a reinforcement learning RL agent; During the application of the high-current test waveform, the acoustic emission signal and volume expansion data of the target battery are simultaneously acquired; Analyze the spectral fingerprint and energy release rate of the acoustic emission signal to determine whether there are precursors to thermal runaway; The real-time expansion rate of the volume expansion data is calculated and compared with the expected electrochemical expansion rate to determine whether lithium plating exists. The results of the judgment on the precursor to thermal runaway and the results of the judgment on lithium plating are fed back into the RL agent; The RL agent adjusts the parameters of the high-current test waveform in real time based on the feedback to adaptively apply test stress within the safety boundary.

[0011] Optionally, the high-current test waveform includes a high-rate pulse sequence for evaluating battery power performance, wherein the peak current of the high-rate pulse sequence is greater than 2C.

[0012] Optionally, the high-current test waveform is used to simulate the dynamic load under actual working conditions, and the goal of the RL agent is to make the test stress applied by the generated high-current test waveform approximate the stress mode of the dynamic load.

[0013] Optionally, during the application of the high-current test waveform, the following further includes: Collect the terminal voltage and current of the target battery; The DC internal resistance DCR is calculated based on the terminal voltage and current, and the trend of the DCR change is used to assess the state of health (SOH) of the battery.

[0014] Optionally, the steps of analyzing the spectral fingerprint and energy release rate of the acoustic emission signal include: Time-frequency domain analysis was performed on the acoustic emission signal to extract the characteristic frequency components in the high-frequency band; Monitor the energy release rate of the high-frequency band; When the energy release rate suddenly changes and exceeds a preset safety threshold, it is determined that there are signs of impending thermal runaway.

[0015] Optionally, the step of calculating the real-time expansion rate of the volume expansion data includes: Strain signals are acquired using high-resolution strain sensors, and these strain signals characterize the volume expansion data. The real-time expansion rate is obtained by taking the first derivative of the strain signal.

[0016] Optionally, the steps for determining whether lithium plating exists include: Calculate the expected electrochemical expansion rate based on the current current density and state of charge (SOC). When the real-time expansion rate exceeds the expected electrochemical expansion rate and reaches a preset lithium plating threshold, lithium plating is determined to exist.

[0017] Secondly, an embodiment of the present invention provides a battery charging and discharging system based on high-current testing, comprising: A reinforcement learning (RL) agent module is used to dynamically generate high-current test waveforms; The test execution module is used to apply the high-current test waveform to the target battery; A multi-dimensional sensing module, including an acoustic sensor and a strain sensor, is used to simultaneously acquire the acoustic emission signal and volume expansion data of the target battery; The real-time analysis module is used to analyze the spectral fingerprint and energy release rate of the acoustic emission signal to determine the precursors of thermal runaway, and to calculate the real-time expansion rate of the volume expansion data to determine lithium plating. The feedback control module is used to feed back the judgment results of the thermal runaway precursor and the lithium plating to the RL intelligent agent module, so that the RL intelligent agent module can adjust the parameters of the high current test waveform in real time and adaptively apply test stress within the safety boundary.

[0018] The present invention has achieved the following beneficial effects: This invention proposes an innovative battery testing paradigm that deeply integrates the adaptive decision-making capability of reinforcement learning with the real-time monitoring capability of multi-dimensional physical perception (acoustic emission and volume expansion), enabling efficient and safe exploration of battery performance boundaries in high-current testing.

[0019] First, adaptive optimization of test stress was achieved. The RL agent can dynamically adjust the test waveform based on real-time feedback from the battery, breaking free from the constraints of traditional fixed test conditions. This allows the testing process to consistently approach the battery's performance limits within safe boundaries, significantly improving testing efficiency and accelerating battery performance evaluation and verification.

[0020] Secondly, it significantly improves testing safety. By introducing acoustic emission and volume expansion—two physical signals highly sensitive to the internal failure mechanisms of batteries—real-time, in-situ, and non-destructive detection of precursors to thermal runaway and lithium plating is achieved. This overcomes the lag inherent in traditional electrical and temperature parameter monitoring, enabling early warnings at the early stages of internal failure and providing a valuable time window for safety intervention.

[0021] Furthermore, it achieves an optimal balance between testing efficiency and safety. This invention constructs a complete system from multi-dimensional perception to intelligent decision-making and closed-loop control. The RL agent can learn and adapt to the characteristics of different individual batteries and aging states, ensuring that the testing process is always finely controlled within safety boundaries, effectively avoiding the occurrence of safety accidents.

[0022] Finally, this invention achieves a deep integration and synergistic effect of multidimensional physical perception and intelligent decision-making. On the one hand, acoustic emission and volume expansion monitoring provide the RL agent with key information reflecting the real-time physical mechanism inside the battery that traditional electrical parameters cannot provide, enabling the agent to safely approach its performance limits. On the other hand, the RL agent provides decision-making capabilities for dynamic optimization and refined control of multidimensional physical perception data.

[0023] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0025] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a battery charging and discharging method based on high current testing in an embodiment of the present invention; Figure 2 This is a schematic diagram of a battery charging and discharging system based on high current testing in an embodiment of the present invention. Detailed Implementation

[0026] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0027] This invention aims to construct an intelligent closed-loop testing system that, while ensuring safety, autonomously explores and applies optimal high-current stress to comprehensively evaluate the performance boundaries of a battery. Its core idea lies in utilizing the autonomous decision-making capabilities of reinforcement learning (RL), combined with real-time feedback from acoustic emission (AE) signals and volume expansion data—which are highly sensitive to the internal physical state of the battery—to achieve dynamic adaptive control of the testing process.

[0028] Example 1

[0029] This embodiment provides a battery charging and discharging method based on high-current testing. The method details how to achieve efficient and safe evaluation of battery performance through an intelligent closed-loop control strategy. (Refer to...) Figure 1 The method in this embodiment mainly includes the following steps: Step S101: Dynamically generate and apply a large current test waveform to the target battery through a reinforcement learning RL agent.

[0030] In this step, the RL agent plays the role of the core decision-maker. Unlike traditional preset test procedures that rely on fixed operating conditions, the RL agent can autonomously decide the current waveform parameters to be applied next based on the current environmental conditions (including the battery's electrochemical state, thermal state, and historical test data). Here, "high current" usually refers to a current exceeding 1C of the battery's nominal capacity, such as 2C, 3C, or even higher. This type of current can apply significant stress to the battery, accelerating performance evaluation or the aging process.

[0031] The design of the RL agent is a key aspect of implementing this method. In a specific implementation, the RL agent can employ advanced reinforcement learning algorithm frameworks, such as Proximal Policy Optimization (PPO) or Deep Deterministic Policy Gradient (DDPG). These algorithms can handle continuous state and action spaces, making them suitable for scenarios requiring fine-grained control, such as battery testing. The agent's goal is to learn an optimal policy that maximizes some form of test objective while satisfying safety constraints, such as maximizing accumulated test stress, minimizing test time, or most accurately simulating a specific operating condition. In this invention, test stress is a quantitative indicator used to measure the intensity of the load applied to the battery. In a preferred embodiment, to accelerate testing, test stress is defined as a function of the current rate (C-rate), such as the square of the current rate, to emphasize the nonlinear effect of high current on the battery.

[0032] The state space of an agent needs to comprehensively characterize the current state of the battery. The selection of input features for the state space is crucial to the performance of the agent. In this invention, the state space should contain information that comprehensively reflects the current electrochemical, thermal, and mechanical state of the battery. Specifically, this includes, but is not limited to: the target battery's current state of charge (SOC), terminal voltage, current, surface temperature, internal temperature (if measurable or estimable), recent current and voltage historical records (e.g., rate of change), and battery state of health (SOH) indicators (e.g., capacity decay rate or internal resistance). By inputting these features into the agent's neural network model (typically composed of a policy network and / or a value network), the agent can perceive the complex environment in which the battery exists. In this invention, the state vector St specifically includes: electrochemical features (current SOC, terminal voltage, current, DC internal resistance DCR); thermal features (surface temperature, estimated internal temperature, rate of temperature rise); acoustic features (real-time energy release rate in key frequency bands); and mechanical features (real-time volume expansion rate, deviation from the expected electrochemical expansion rate). To capture the dynamic characteristics of time series data, historical current and voltage records from the past N time steps (e.g., N=10) are also incorporated into the state vector.

[0033] The action space of an agent defines the test waveform parameters it can control. To ensure that the RL agent can effectively handle multi-source heterogeneous state information, the input features need to be preprocessed. Specifically, the Z-score normalization method is used to process all state features (such as SOC, voltage, temperature, acoustic emission energy release rate, real-time expansion rate, etc.) into dimensionless values ​​with a mean of 0 and a variance of 1, in order to accelerate neural network convergence. The agent's action space adopts a continuous action space, with action vector A... t It includes three dimensions: the current amplitude at the next time step (e.g., continuously adjustable within the range of 0C to 5C), the current rise slope, and the pulse frequency (if pulse testing is used).

[0034] The neural network model structure of the RL agent is as follows: Taking the Proximal Policy Optimization (PPO) algorithm as an example, both its policy network (Actor) and value network (Critic) adopt a fully connected multilayer perceptron (MLP) structure. In a preferred embodiment, the policy network contains two hidden layers, each containing 256 neurons, using the ReLU activation function. The output layer outputs the Gaussian distribution mean and variance of actions (such as current amplitude). The value network structure is similar.

[0035] During training, the key hyperparameters were set as follows: the Adam optimizer was used, with a learning rate of 0.0003, a discount factor of 0.99, an epsilon parameter for the PPO algorithm of 0.2, and a batch size of 2048. The training process typically lasted for hundreds of thousands of rounds until the cumulative reward converged.

[0036] The construction of a digital twin model is fundamental to offline training. In this embodiment, an electro-thermal-mechanical multiphysics coupled model is constructed: Electrochemical Model: A pseudo-two-dimensional (P2D) porous electrode model or a higher-order equivalent circuit model (such as a second-order RC model) is employed to accurately predict polarization effects under high current. Model parameters are calibrated using electrochemical impedance spectroscopy (EIS) and galvanostatic titration (GITT). To ensure high fidelity of the digital twin model across a wide operating range, model parameter calibration is based on a comprehensive and diverse experimental dataset. This dataset covers the entire operating window of the target battery and includes: Temperature coverage: From low temperature (e.g. -20°C) to high temperature (e.g. 60°C), multiple temperature points can be set (e.g., at 10°C intervals).

[0037] State of charge (SOC) coverage: from 0% to 100% SOC, measured at high resolution (e.g., in 2% SOC intervals).

[0038] Current rate coverage: from low rates (e.g., C / 50, used for calibrating thermodynamic parameters) to high rates (e.g., 5C, used for calibrating kinetic parameters).

[0039] By using this comprehensive dataset for multi-objective optimization and cross-validation, the robustness and predictive power of the digital twin model are ensured.

[0040] Thermal model: A lumped-parameter thermal model was used, taking into account Joule heating and reaction heat. Heat capacity and thermal conductivity were measured using an accelerating calorimeter (ARC).

[0041] Mechanical model: A constitutive model based on stress-strain relationship to describe the expansion behavior during lithium intercalation.

[0042] Model validation utilizes independent dynamic operating condition datasets (such as the DST operating condition). A model is considered to have sufficient fidelity and suitable for pre-training of the RL agent when the predicted terminal voltage error is less than 10mV and the surface temperature error is less than 1℃. To ensure sufficient robustness and generalization ability of the pre-trained RL agent, diverse training datasets need to be generated using the digital twin model. The dataset generation strategy includes simulating batteries at different states of health (SOH) (e.g., simulating different aging stages by adjusting model parameters) and different ambient temperatures (e.g., from -20℃ to 60℃). During training, a high exploration rate strategy (e.g., an ε-greedy strategy or injecting Gaussian noise into the action space) is employed to encourage the RL agent to fully explore the state space, especially critical data near the safety boundary. Collected experience data is stored in an experience replay buffer, and a Priority Experience Replay (PER) technique is used to give higher sampling weights to "dangerous experiences" that lead to safety penalties, thereby improving learning efficiency and safety.

[0043] In a typical implementation, the action space can include the current amplitude (e.g., charge rate or discharge rate), pulse width, pulse frequency, or slope of the current change to be applied in the next time step (e.g., 1 second or 10 seconds). For fine-grained control, a continuous action space is preferred, allowing the agent to arbitrarily select parameter values ​​within a preset range. For example, the current amplitude can be continuously adjusted between 0C and 5C. The RL agent outputs an action vector at each decision moment, which is sent to a high-precision battery testing device (such as a high-power charge / discharge machine) to execute the corresponding current loading.

[0044] To ensure the RL agent learns effective strategies, sufficient training is necessary. Since extensive trial-and-error training on real batteries is costly and poses safety risks, a strategy combining offline training and online fine-tuning based on digital twin models can be adopted. First, a high-fidelity digital twin model of the target battery is established, including electrochemical, thermal, and mechanical models. Preliminary experimental data is used to identify and calibrate the model's parameters. Then, the RL agent is pre-trained in a simulation environment. During training, the agent continuously interacts with the simulation environment, optimizing its strategy through trial and error to maximize cumulative rewards. After pre-training, the RL agent is deployed to an actual testing system, and online fine-tuning is performed based on measured data during actual testing to adapt to the individual differences and time-varying characteristics of real batteries.

[0045] Step S102: During the application of the high current test waveform, the acoustic emission signal and volume expansion data of the target battery are acquired simultaneously.

[0046] To monitor the microscopic changes and potential risks inside the battery under high current stress in real time, this embodiment employs a multi-dimensional physical sensing strategy, particularly integrating information from both acoustic and mechanical dimensions. This is a key feature that distinguishes this invention from traditional electrochemical testing, because these physical signals typically reflect the battery's internal failure mechanisms more directly than electrical signals.

[0047] Acoustic emission signal acquisition: Acoustic emission refers to the transient elastic waves generated by the rapid release of energy due to stress redistribution within a material. In lithium-ion batteries, microstructural changes during charging and discharging, such as phase transitions in electrode materials, particle cracking, formation and rupture of the SEI film, growth and fracture of lithium dendrites, and gas production from electrolyte decomposition, all generate acoustic emission signals.

[0048] To acquire high-quality acoustic emission signals, a highly sensitive acoustic sensor, such as a piezoelectric ceramic sensor (PZT), is required. The sensor selection should consider its frequency response range, typically covering a frequency band from tens of kHz to several MHz. A broadband acoustic emission sensor is preferred. The sensor needs to be tightly bonded to a specific location on the battery surface (e.g., near the tabs or the center of the battery) using a coupling agent (such as silicone oil or epoxy resin) to minimize signal attenuation caused by acoustic impedance mismatch. Acoustic emission signal acquisition requires a high-speed data acquisition card, with a sampling rate typically exceeding 1 MHz. A preamplifier is also needed to amplify the weak signal, and a bandpass filter is required to remove ambient noise and electromagnetic interference.

[0049] Volume expansion data acquisition: During the charging and discharging process of lithium-ion batteries, the insertion and extraction of lithium ions cause changes in the crystal structure of the electrode materials, resulting in the expansion and contraction of the battery volume. Volume expansion data can reflect the electrochemical processes and mechanical stress state inside the battery.

[0050] Volumetric expansion data can be acquired using high-resolution strain or displacement sensors. For example, resistance strain gauges or fiber Bragg grating (FBG) sensors can be used to measure strain changes on the battery surface. FBG sensors are preferred in this invention due to their strong electromagnetic interference resistance, small size, and high sensitivity. Alternatively, high-precision displacement sensors, such as laser displacement sensors or linear variable differential transformers (LVDTs), can be used to measure displacement changes in the battery thickness direction, requiring a resolution at the micrometer level. It is important to note that volumetric expansion data is highly sensitive to temperature changes; therefore, the battery surface temperature needs to be recorded simultaneously during acquisition, and temperature compensation should be applied to the data to eliminate the effects of thermal expansion.

[0051] The key to this step is achieving strict time synchronization of electrochemical data, acoustic emission data, and volume expansion data. This requires all data acquisition devices to use a unified clock source or triggering mechanism to ensure that the causal relationship between the electrochemical process and the acoustic and mechanical responses can be accurately correlated in subsequent data analysis.

[0052] To accurately eliminate the impact of temperature fluctuations on volume expansion measurements, this embodiment employs a dynamic temperature compensation strategy. First, the coefficient of thermal expansion (CTE) of the target battery at different states of charge (SOC) is experimentally calibrated. Then, based on real-time temperature changes and the calibrated CTE model, the volume change caused by thermal effects is calculated, and this amount is subtracted from the total volume expansion data to obtain the net expansion data purely caused by the electrochemical process. This ensures the accuracy of lithium plating determination.

[0053] Step S103: Analyze the spectral fingerprint and energy release rate of the acoustic emission signal to determine whether there are signs of impending thermal runaway.

[0054] To ensure the accuracy of acoustic emission signal analysis in complex testing environments, this embodiment employs multiple anti-interference measures. At the hardware level, differential input and shielded cables are used to suppress electromagnetic interference. At the signal processing level, in addition to bandpass filtering, intelligent noise reduction algorithms based on deep learning (such as autoencoders) can be introduced to separate background noise from actual acoustic emission events. Furthermore, multi-sensor arrays and sound source localization technology can be used to include only signals with sound sources located inside the battery within the analysis scope, thereby significantly improving the reliability of the monitoring system.

[0055] The acquired raw acoustic emission signals need to be analyzed for their physical meaning using advanced signal processing and feature extraction methods, and used to identify precursors to thermal runaway. Thermal runaway is a complex chain reaction process that typically begins with microscopic damage inside the battery, which generates acoustic emission signals with specific characteristics.

[0056] Spectral fingerprint analysis: Different acoustic emission sources possess different frequency characteristics, i.e., spectral fingerprints. To extract spectral fingerprints, time-frequency domain analysis of the acoustic emission signal is required. Commonly used methods include Short-Time Fourier Transform (STFT) and Wavelet Transform. Wavelet Transform, due to its multi-resolution analysis capabilities, is more suitable for handling non-stationary acoustic emission signals. Through wavelet transform, the signal can be decomposed into different frequency bands, and the energy distribution in each band can be analyzed over time. In this invention, the "spectral fingerprint" refers not only to the frequency distribution of the signal but, more specifically, to a unique pattern or combination of characteristic parameters characterizing a specific failure mode on a time-frequency graph. For example, a spectral fingerprint may include the dominant frequency band, peak frequency, bandwidth, signal duration, and rise time. For instance, a spectral fingerprint associated with early gas production might appear as a continuous signal in the 150-250 kHz frequency band, while one associated with electrode particle cracking might appear as a burst signal in the 300-400 kHz frequency band.

[0057] In this invention, we focus particularly on characteristic frequency components in the high-frequency range (e.g., above 100 kHz). Studies have shown that acoustic emission signals associated with microstructural damage (e.g., particle cracking) and severe side reactions (e.g., gas generation) are mainly concentrated in the high-frequency range. Through analysis of a large amount of experimental data, a correlation can be established between specific spectral fingerprints and specific failure modes. For example, if energy concentration occurs in a specific frequency band, it may indicate a specific precursor to failure.

[0058] Energy Release Rate Analysis: The energy release rate is an important indicator for measuring the intensity of acoustic emission activity, reflecting the rate of energy dissipation within the battery. Under normal operating conditions, the energy release rate remains at a low level. When abnormal changes occur within the battery, acoustic emission activity intensifies, and the energy release rate increases significantly. The energy release rate can be obtained by calculating the root mean square (RMS) of the acoustic emission signal or the cumulative energy. We are more concerned with abrupt changes in the energy release rate; by monitoring its trend over time, we can promptly detect the occurrence of anomalies.

[0059] Identifying Precursors to Thermal Runaway: Based on the analysis results of spectral fingerprints and energy release rates, criteria for identifying precursors to thermal runaway can be established. This invention employs a method combining threshold and trend analysis. First, preset safety thresholds are set. These thresholds are determined based on historical experimental data (e.g., calibrated through abuse testing) and safety standards. For example, absolute thresholds and growth rate thresholds for energy release rates can be set. The specific calibration process for the preset safety thresholds and spectral fingerprints is as follows: First, acoustic emission data related to thermal runaway were collected by subjecting the target battery to abuse tests (e.g., overcharge or nail penetration tests). Through wavelet packet decomposition and unsupervised clustering analysis (e.g., K-means clustering), we identified characteristic frequency bands associated with early precursors of thermal runaway (e.g., SEI film decomposition, gas generation). For example, we found that signals in the 150–250 kHz band were highly correlated with early gas generation, while signals in the 300–400 kHz band may be associated with cracking of electrode particles.

[0060] Secondly, determine the irreversible initiation time of thermal runaway (T). onset For example, it can be defined based on the rate of temperature change dT / dt exceeding 1℃ / s.

[0061] Then, the analysis is performed on T. onset Previous warning time window (e.g., T) onset Energy release rate data for this characteristic frequency band within the first 60 seconds.

[0062] Finally, a preset safety threshold is set. Based on statistical analysis, the safety threshold is set to be significantly higher than the baseline noise level under normal operating conditions. Specifically, the 6-sigma principle from statistical process control is preferably adopted, setting the threshold to the baseline mean plus 6 times the standard deviation to ensure an extremely low false alarm rate. When the real-time monitored energy release rate exceeds this threshold and the duration exceeds 100 milliseconds, it is determined that there are signs of impending thermal runaway.

[0063] During testing, the energy release rate in the high-frequency band is monitored in real time. When the energy release rate undergoes a sudden change—that is, its growth rate exceeds a preset growth rate threshold, and its absolute value exceeds a preset safety threshold—it is determined that there is a precursor to thermal runaway. To improve the accuracy of the judgment, changes in spectral fingerprints can also be considered. If a sudden change in the energy release rate is accompanied by the appearance of a spectral fingerprint of a specific failure mode, the risk of thermal runaway can be determined with greater certainty.

[0064] Step S104: Calculate the real-time expansion rate of the volume expansion data and compare it with the expected electrochemical expansion rate to determine whether lithium plating exists.

[0065] Volume expansion analysis is a key method for identifying lithium plating. Lithium plating causes irreversible expansion of the battery volume, and its expansion rate differs significantly from the normal electrochemical expansion rate.

[0066] Calculation of real-time expansion rate: First, the real-time expansion rate needs to be calculated from the acquired volume expansion data (e.g., strain signals obtained through high-resolution strain sensors). To obtain the real-time expansion rate, the first derivative of the strain signal needs to be taken. In practice, due to the presence of noise in the signal, direct derivative calculation will amplify the noise. Therefore, the strain signal needs to be filtered first, such as using a Savitzky-Golay filter for smoothing. This filter can smooth the signal while more accurately estimating the derivative of the signal, and then differential calculation is performed to obtain the expansion rate.

[0067] Calculation of the expected electrochemical expansion rate: To determine whether the real-time expansion rate is abnormal, it is necessary to know what the normal electrochemical expansion rate should be under the current operating conditions. Normal electrochemical expansion is mainly caused by lattice expansion resulting from lithium-ion insertion into the anode material (such as graphite). This expansion is reversible, and its rate is closely related to the current density and the state of charge (SOC).

[0068] The expected electrochemical expansion rate needs to be modeled through experimental calibration. The calibration process is as follows: the target battery is charged and discharged at low rates (e.g., C / 10 or lower) to ensure that lithium plating does not occur. During cycling, the battery's volume expansion data, current, and state of charge (SOC) are measured simultaneously. By analyzing these data, the volume expansion caused by a unit change in charge at different SOCs can be obtained, i.e., the differential expansion coefficient (DEC). Based on the DEC, a lookup table or empirical model can be built to describe the relationship between the expected electrochemical expansion rate and the current current density and SOC.

[0069] In this embodiment, a multidimensional lookup table is used to implement the DEC model. The calibration process is as follows: The battery was charged and discharged at multiple constant temperatures (e.g., 0°C, 25°C, 45°C) using ultra-low rate currents (e.g., C / 50 or lower) to ensure the process approached thermodynamic equilibrium and prevented lithium plating. Volume expansion and charge change were precisely measured. The DEC value was obtained by differentiating the volume expansion relative to the charge change. For example, for a certain type of NCM / graphite battery, the DEC value reached its peak (e.g., 1.5 μm / Ah) in the SOC range of 20%-50% due to the graphite phase transition.

[0070] The lookup table uses SOC and temperature as input dimensions. During real-time testing, the system uses a bilinear interpolation algorithm to calculate the current expected DEC value based on the current real-time SOC and temperature.

[0071] Regarding the setting of the lithium plating threshold, we adopted an adaptive threshold mechanism. This threshold is adjusted based on the confidence level of the DEC model in different SOC intervals. In the interval where changes in SOC cause rapid changes in DEC values ​​(phase transition region), the model prediction error may be large. In this case, the system will automatically and appropriately increase the adaptive threshold (e.g., set to 20% of the expected rate) to avoid false alarms. In the interval where DEC values ​​are relatively flat, the system will appropriately decrease the threshold (e.g., set to 10% of the expected rate) to improve detection sensitivity.

[0072] In actual testing, the expected electrochemical expansion rate can be obtained by consulting the model based on the current current density and SOC.

[0073] Determining Lithium Plating: The presence of lithium plating can be determined by comparing the real-time expansion rate with the expected electrochemical expansion rate. When lithium plating occurs, the deposition of metallic lithium causes additional volume expansion. Therefore, the real-time expansion rate will be significantly higher than the expected electrochemical expansion rate.

[0074] Set a preset lithium plating threshold. This threshold represents the tolerance level for the real-time expansion rate to exceed the expected expansion rate. For example, the lithium plating threshold can be set to 10% or 20% of the expected expansion rate, or an adaptive threshold can be used. When the real-time expansion rate exceeds the expected electrochemical expansion rate to the preset lithium plating threshold, lithium plating is determined to exist. It should be noted that the determination of lithium plating also needs to consider the temperature factor. The influence of temperature should be taken into account when building the model, and temperature compensation should be performed during the determination.

[0075] Step S105: The judgment results of the thermal runaway precursor and the judgment results of lithium plating are fed back to the RL agent.

[0076] In this step, the risk assessment results obtained in steps S103 and S104 are fed back to the RL agent to form closed-loop control. Feedback information is a crucial basis for the RL agent's decision-making. The feedback information can be binary (whether a risk exists) or continuous (risk level). In this embodiment, the judgment results of thermal runaway precursors and lithium plating can be encoded as risk indicators and input into the RL agent's state space, or more importantly, directly used to construct the reward function.

[0077] To ensure real-time feedback, the entire process from data acquisition to risk assessment and feedback input requires extremely low latency. This necessitates a system with high-performance real-time computing capabilities, preferably employing an edge computing architecture. This involves deploying signal processing and risk assessment algorithms on edge devices close to the sensors to reduce data transmission latency. The system's response time should be controlled at the millisecond level to ensure timely intervention when a risk is detected.

[0078] Step S106: The RL agent adjusts the parameters of the high-current test waveform in real time based on the feedback to adaptively apply test stress within the safety boundary.

[0079] The adaptive closed-loop control system implemented in this invention has strict timing requirements and a clear data flow process. The system operates based on discrete time steps (e.g., set to Δt = 1 second). At each time step, the system sequentially performs state observation, intelligent decision-making, test execution, real-time monitoring, and reward feedback. To ensure the effectiveness and safety of the closed-loop control, the entire system achieves high-precision time synchronization (e.g., via the IEEE 1588 Precision Time Protocol), ensuring strict alignment of electrochemical, acoustic, and mechanical data in timestamps. The response delay from risk assessment to action adjustment is controlled at the millisecond level, ensuring timely intervention when a risk is detected.

[0080] In this step, the RL agent adjusts its testing strategy based on feedback information to achieve adaptive stress application. This is one of the core innovations of this invention. The goal of the RL agent is to maximize test stress (e.g., current rate, power density, or cumulative energy throughput) while ensuring safety. To achieve this goal, the reward function of the RL agent needs to be carefully designed.

[0081] The reward function is used to evaluate the quality of the actions taken by the agent at each decision-making time. In this invention, the reward function should include two main parts: performance reward and safety penalty.

[0082] Performance rewards: These are used to encourage agents to apply greater testing stress. For example, performance rewards can be set to be proportional to the current amplitude; the greater the current amplitude, the higher the reward.

[0083] Safety Penalty: Used to punish actions by which the agent causes the battery to enter a dangerous state. A significant negative reward (penalty) should be given to the agent when signs of impending thermal runaway or lithium plating are detected. The severity of the penalty should be adjusted according to the severity of the risk.

[0084] To achieve a quantitative balance between testing efficiency and safety, this embodiment constructs a multi-objective reward function R, whose specific structure includes three components: the test stress reward term R.p Lithium plating penalty item R Li And thermal runaway penalty item R th The total reward R is the weighted sum of these three items.

[0085] R p The aim is to encourage the agent to apply a larger current. In this embodiment, R p Set to the current multiplier C within the current time step. rate It is proportional to the square of the value, in order to encourage the exploration of high magnification.

[0086] R Li Designed to suppress lithium plating. Triggered when the real-time expansion rate exceeds the lithium plating threshold. R Li It is set to be proportional to the excess magnitude (excess inflation rate) and multiplied by a large negative weighting coefficient W. Li (For example, set to -50).

[0087] R th Designed to prevent thermal runaway. Triggered when the acoustic emission energy release rate exceeds a preset safety threshold. th Set to a value much larger than R p The fixed negative value W of the maximum possible value th (For example, set it to -5000) to ensure that security is the highest priority.

[0088] In order to systematically determine R Li Weighting coefficient W Li An automatic optimization method based on simulation was adopted. Specifically, in the digital twin environment, the Bayesian optimization algorithm was used to optimize W. Li Perform a parameter search (e.g., within the range of -10 to -100). Find the W value that maximizes the cumulative test stress while satisfying the lithium plating constraint. Li This systematic approach ensures that the reward function achieves an optimal balance between efficiency and safety. Through extensive training and optimization in a simulation environment, the weight coefficients of each item were determined, enabling the RL agent to learn the optimal strategy for balancing various metrics under different conditions.

[0089] Specifically, the reward function of the RL agent is set to maximize the test stress, and the judgment results of the thermal runaway precursor and the lithium plating in the feedback are used as penalty items.

[0090] When signs of impending thermal runaway are detected, the risk level is extremely high. In this case, a very severe penalty should be imposed on the agent immediately, and an emergency stop mechanism should be triggered to terminate the test immediately to ensure safety.

[0091] When lithium plating is detected, the risk level is relatively low, but long-term accumulation can lead to performance degradation and safety hazards. In this case, an appropriate penalty should be imposed on the agent to guide it to adjust the testing strategy and avoid further lithium plating.

[0092] Upon receiving a penalty signal, the RL intelligence automatically adjusts its policy network parameters through a learning process to avoid taking actions that would result in penalties in the future. Real-time adjustment of test waveform parameters includes adjusting the amplitude, rise slope, and pulse frequency of the high-current test waveform.

[0093] When lithium plating is detected, the RL agent preferentially reduces the amplitude of the high-current test waveform. Reducing the current amplitude directly lowers the overpotential of the negative electrode, thereby suppressing lithium plating. Furthermore, the agent can learn to increase the pulse interval or decrease the current rise slope to provide the battery with more relaxation time, promoting a more uniform distribution of lithium ions and reducing the risk of lithium plating caused by localized high concentrations.

[0094] Through this adaptive adjustment mechanism, the RL agent can conduct refined exploration near the safety boundary. When the battery is in good condition, the agent gradually increases the test stress, approaching the battery's performance limits; when a risk signal is detected, the agent quickly reduces the test stress, restoring the battery to a safe state. This dynamic balance makes the testing process both efficient and safe.

[0095] The battery charging and discharging method based on high-current testing provided in this embodiment integrates multiple advanced technologies such as reinforcement learning, acoustic emission monitoring, and volume expansion monitoring. This enables adaptive optimization of test conditions and real-time monitoring of internal battery risks, effectively resolving the contradiction between efficiency and safety in traditional high-current testing methods.

[0096] Example 2

[0097] Based on Example 1, this embodiment specifically defines the type of high-current test waveform, aiming to evaluate the power performance of the battery.

[0098] According to the method described in this embodiment, the high-current test waveform includes a high-rate pulse sequence for evaluating battery power performance. The high-rate pulse sequence typically consists of a series of short-duration, high-amplitude charge and discharge pulses, used to simulate the battery power requirements under conditions such as acceleration and braking of an electric vehicle.

[0099] The peak current of the high-rate pulse sequence is greater than 2C. Here, C represents the rated capacity rate of the battery. For example, for a 10Ah battery, 2C represents a current of 20A. A peak current greater than 2C can apply significant stress to the battery, fully testing its power output capability and internal polarization characteristics.

[0100] In this embodiment, the task of the RL agent is to dynamically generate parameters for a high-rate pulse sequence, including pulse amplitude, pulse width (e.g., 1 to 30 seconds), and pulse interval (relaxation time). The agent's goal is to maximize pulse power (or amplitude) to explore the battery's peak power capability while ensuring that thermal runaway and lithium plating do not occur.

[0101] Acoustic emission monitoring and volume expansion monitoring are particularly important during the application of high-rate pulse sequences. High-rate pulses can cause rapid changes in internal stress within the battery, which can easily induce microstructural damage and lithium plating.

[0102] Acoustic emission monitoring systems pay particular attention to the acoustic response at the moment of pulse application. If the pulse causes microscopic damage to the electrode material, such as the fracture of active particles due to stress concentration during rapid lithium insertion / extraction, the acoustic emission system will capture the sudden acoustic event. By analyzing the intensity and frequency of these events, the stability of the battery structure can be assessed.

[0103] The volume expansion monitoring system focuses on the instantaneous expansion behavior during the pulse. Under high-rate charging pulses, if lithium plating occurs, it will cause rapid volume expansion. The calculation of the real-time expansion rate and comparison with the expected value can sensitively detect this phenomenon.

[0104] The RL agent adjusts pulse parameters in real time based on this feedback. For example, when the RL agent attempts to apply a pulse with a higher amplitude (e.g., increasing from 3C to 3.5C), if the real-time analysis module reports signs of lithium plating (abnormally increased expansion rate) or intensified acoustic emission activity, the agent immediately receives a penalty signal. In the next time step, the agent adjusts its strategy, such as reducing the pulse amplitude back to 3C, shortening the pulse width, or increasing the relaxation time to allow the battery's internal state to recover.

[0105] Through this adaptive exploration process, the RL agent can quickly and safely determine the maximum pulse current and corresponding pulse width that the battery can withstand under current SOC and temperature conditions. Compared to traditional manual trial-and-error methods or fixed HPPC testing, the method in this embodiment significantly improves the efficiency and safety of power performance evaluation and can obtain a more refined performance profile. For example, the agent can learn that under low temperature and high SOC conditions, the battery is more prone to lithium plating, thus requiring limitation of the charging pulse amplitude.

[0106] Example 3

[0107] This embodiment expands the application scenarios of high current test waveforms based on Embodiment 1, aiming to simulate dynamic loads under actual working conditions.

[0108] According to the method described in this embodiment, the high-current test waveform is used to simulate the dynamic load under actual operating conditions. Actual operating conditions are usually highly dynamic and random, such as the driving conditions of electric vehicles on urban roads or highways (e.g., WLTC, CLTC, etc.), or the operating conditions of grid-side energy storage systems participating in frequency regulation and peak shaving.

[0109] In this embodiment, the goal of the RL agent is to make the test stress applied by the generated high-current test waveform approximate the stress mode of the dynamic load. Here, the stress mode refers not only to the magnitude of the current or power, but also to its dynamic characteristics over time, such as frequency distribution, amplitude distribution, and current change rate distribution.

[0110] To achieve this goal, the reward function of the RL agent needs to be redesigned. The reward function is no longer simply maximizing the test stress, but rather minimizing the difference between the generated test waveform and the target dynamic load (e.g., differences in characteristic parameters such as power spectral density and current rate of change distribution), while satisfying safety constraints. The smaller the difference, the higher the reward.

[0111] Meanwhile, safety penalties remain in place. When signs of impending thermal runaway or lithium plating are detected, the agent is penalized to ensure the safety of the testing process.

[0112] During testing, the RL agent continuously learns how to generate test waveforms that both conform to the target dynamic load characteristics and ensure battery safety. For example, the agent may learn that during periods of high power demand, the current waveform needs to be adjusted appropriately (e.g., limiting the rise rate) to avoid triggering safety risks; while during periods of low power demand, the stress intensity can be appropriately increased to accelerate testing.

[0113] The method provided in this embodiment can reproduce the stress on batteries under actual operating conditions with high fidelity in a laboratory environment, thereby more accurately evaluating the battery's performance and lifespan in real-world applications. Furthermore, the introduction of real-time safety monitoring and adaptive adjustment mechanisms can prevent safety accidents from occurring when simulating extreme operating conditions.

[0114] Example 4

[0115] This embodiment adds an online assessment function for battery state of health (SOH) based on embodiment one, realizing the integration of testing and diagnosis.

[0116] According to the method described in this embodiment, during the application of the high-current test waveform, the following steps are further included: The terminal voltage and current of the target battery are collected. This electrochemical data is the basis for battery testing and can be acquired with high precision through the measurement module built into the test execution device, and is synchronized with acoustic emission and volume expansion data.

[0117] The DC internal resistance (DCR) is calculated based on the terminal voltage and current. DCR is an important indicator of battery internal resistance, reflecting the ohmic and electrochemical polarization characteristics within the battery. DCR calculation is typically based on the voltage relaxation method. The terminal voltage changes when a current pulse is applied or after the pulse ends. The DCR can be calculated by analyzing the ratio of the voltage change to the current change at a specific time point (e.g., 1 second or 10 seconds after the pulse is applied).

[0118] The trend of DCR change is used to assess the state of health (SOH) of the battery. As the battery ages, its internal resistance gradually increases. The trend of DCR change can reflect the rate and mechanism of battery aging.

[0119] In this embodiment, the high-current test waveforms (especially high-rate pulse sequences) dynamically generated by the RL agent provide ample opportunities for DCR calculation. The system can calculate DCR during each charge-discharge cycle or periodically at specific SOC points.

[0120] The trend of DCR changes can be used as feedback information input into the RL agent. The agent can adjust the testing strategy based on the changes in DCR. For example, if the DCR increases too quickly, indicating accelerated battery aging, the agent may appropriately reduce the testing stress to slow down the aging rate.

[0121] Furthermore, the trend of DCR changes can be correlated with the results of acoustic emission monitoring and volume expansion monitoring to gain a deeper understanding of the battery aging mechanism. For example, if DCR increases along with increased acoustic emission activity, it may indicate that aging is caused by damage to the microstructure; if it is accompanied by abnormal volume expansion, it may indicate that aging is caused by side reactions such as lithium plating or gas production.

[0122] The method provided in this embodiment can not only accelerate the battery testing process, but also monitor the battery's health status in real time during the test, thus achieving a comprehensive evaluation of battery performance, safety, and lifespan.

[0123] Example 5

[0124] This embodiment, based on Embodiment 1, provides a detailed explanation of the acoustic emission signal analysis method to improve the accuracy of judging precursors to thermal runaway.

[0125] According to the method described in this embodiment, the steps of analyzing the spectral fingerprint and energy release rate of the acoustic emission signal include: Step 1: Perform time-frequency domain analysis on the acoustic emission signal to extract the characteristic frequency components in the high-frequency band.

[0126] Time-frequency domain analysis is a crucial step in extracting spectral fingerprints. In this embodiment, continuous wavelet transform (CWT) or wavelet packet decomposition (WPD) is preferably used for time-frequency domain analysis. These methods can provide high-resolution time-frequency plots, clearly showing the changes in signal frequency components over time.

[0127] Analysis of CWT or WPD results can identify different types of acoustic emission events and their corresponding frequency ranges. Particular attention is paid to high-frequency signal components (e.g., above 100 kHz). Statistical analysis of extensive experimental data can determine characteristic frequency components closely related to precursors of thermal runaway. For example, signals in the 150 kHz–250 kHz band may be associated with the rupture and repair process of the SEI film, while signals in the 300 kHz–400 kHz band may be associated with the cracking of electrode particles or gas escape.

[0128] The process of extracting characteristic frequency components also needs to consider noise suppression. Background noise can be filtered out and the true acoustic emission events can be extracted by setting an adaptive threshold or using a deep learning-based noise reduction algorithm.

[0129] Step 2: Monitor the energy release rate of the high-frequency band.

[0130] After identifying the characteristic frequency components, it is necessary to monitor the energy release rate of these frequency bands in real time. The frequency bands of interest can be extracted using bandpass filters, and then their energy release rate can be calculated. The energy release rate can be calculated using the sliding time window method, which involves calculating the cumulative energy or root mean square (RMS) value of the signal within each time window.

[0131] Step 3: When the energy release rate suddenly changes and exceeds the preset safety threshold, it is determined that there are signs of impending thermal runaway.

[0132] Sudden changes in energy release rate are a crucial indicator for identifying precursors to thermal runaway. To accurately detect these changes, statistical process control (SPC) methods, such as cumulative sum (CUSUM) control charts or exponentially weighted moving average (EWMA) control charts, can be employed. These methods can sensitively detect minute trend changes and reduce false alarm rates.

[0133] Setting a preset safety threshold requires determining the characteristic parameter levels of the acoustic emission signal before thermal runaway occurs through abuse testing experiments, thereby establishing a reasonable safety threshold. For example, the safety threshold can be set as a several times the baseline level, or it can be set based on a statistical distribution.

[0134] In this embodiment, by using refined time-frequency domain analysis and statistical process control methods, abnormal acoustic emission activities related to thermal runaway can be identified more accurately, thereby improving the sensitivity and reliability of thermal runaway precursor judgment.

[0135] Example 6

[0136] This embodiment, based on Embodiment 1, provides a detailed explanation of the method for processing volume expansion data in order to improve the accuracy of real-time expansion rate calculation.

[0137] According to the method described in this embodiment, the step of calculating the real-time expansion rate of the volume expansion data includes: Step 1: Acquire strain signals using a high-resolution strain sensor, the strain signals representing the volume expansion data.

[0138] In this embodiment, a fiber Bragg grating (FBG) strain sensor is preferably used. FBG sensors have high sensitivity and high resolution, capable of detecting micro-strain (με) levels. By arranging an FBG sensor array at different locations on the battery surface, the distribution information of the strain field on the battery surface can be obtained.

[0139] When acquiring strain signals, it is necessary to eliminate cross-sensitivity caused by temperature changes. FBG sensors are sensitive to both strain and temperature. Therefore, temperature compensation techniques are required, such as using an additional FBG temperature sensor placed close to the strain sensor to measure temperature changes and subtract temperature-induced wavelength drift from the strain signal.

[0140] The placement of strain sensors also requires careful design. For pouch or prismatic batteries, sensors are typically placed in the central region of the battery surface, as this area experiences the most significant expansion. Finite element analysis can be used to simulate the battery's expansion behavior to determine the optimal sensor placement.

[0141] Step 2: Take the first derivative of the strain signal to obtain the real-time expansion rate.

[0142] Directly differentiating the strain signal amplifies noise. In this embodiment, a method based on local polynomial fitting is used to calculate the first derivative, namely the Savitzky-Golay (SG) filter. This method can smooth the signal while more accurately estimating its derivative.

[0143] The parameters of the SG filter (such as window size and polynomial order) need to be optimized based on the characteristics of the signal. The window size determines the smoothness; a larger window results in better smoothing, but some high-frequency information (transient information) will be lost. The polynomial order determines the complexity of the fit; a lower order (such as 2nd or 3rd order) is usually chosen.

[0144] In its implementation, the system acquires strain signals in real time and applies the SG filtering algorithm within a sliding window. SG filtering directly outputs the fitted signal value and its derivative. This method yields a smooth and real-time expansion rate curve. The real-time expansion rate curve clearly reflects the dynamic process of volume change during battery charging and discharging.

[0145] The refined data processing method provided in this embodiment can obtain more accurate and reliable real-time expansion rate data, providing a solid foundation for subsequent lithium plating judgment.

[0146] Example 7

[0147] This embodiment, based on Embodiment Six, provides a detailed explanation of the specific method for determining lithium plating, in order to improve the sensitivity and accuracy of lithium plating detection.

[0148] According to the method described in this embodiment, the step of determining whether lithium plating exists includes: Step 1: Calculate the expected electrochemical expansion rate based on the current current density and state of charge (SOC).

[0149] The expected electrochemical expansion rate model serves as a benchmark for determining lithium plating. In this embodiment, a model based on the differential expansion coefficient (DEC) is used to calculate the expected electrochemical expansion rate.

[0150] DEC is defined as the volume expansion caused by a unit change in charge. DEC can be calibrated through low-rate charge-discharge experiments. DEC is a function of SOC, and its value varies in different SOC ranges, reflecting the expansion characteristics of the negative electrode material at different lithium intercalation stages (e.g., the staged expansion characteristics of graphite).

[0151] The expected electrochemical expansion rate can be calculated by multiplying the current current density by the DEC value corresponding to the current state of charge (SOC). It is important to note that the current density refers to the current per unit area and needs to be normalized based on the electrode area of ​​the battery.

[0152] To improve model accuracy, the effect of temperature on DEC needs to be considered. The expansion characteristics of the negative electrode material differ at different temperatures. Therefore, calibration experiments need to be conducted at different temperatures to establish a DEC model that incorporates temperature factors. Furthermore, the impact of battery aging on expansion characteristics needs to be considered; the model may require periodic updates or online calibration using adaptive algorithms.

[0153] Step 2: When the real-time expansion rate exceeds the expected electrochemical expansion rate and reaches the preset lithium plating threshold, lithium plating is determined to exist.

[0154] The additional expansion caused by lithium plating can result in a real-time expansion rate higher than expected. Setting the lithium plating threshold requires a trade-off between detection sensitivity and false alarm rate. Setting the threshold too low may misinterpret normal electrochemical expansion fluctuations as lithium plating; setting it too high may miss early lithium plating phenomena.

[0155] In this embodiment, an adaptive lithium plating threshold can be used. The threshold is no longer a fixed value, but is dynamically adjusted based on the current operating conditions (e.g., SOC range, temperature) and historical data. For example, in the low SOC range, the expansion characteristics of the anode material change rapidly, and the error of the expected model may be large, so the threshold can be appropriately increased; in the high SOC range, the expansion characteristics of the anode material are relatively stable, so the threshold can be appropriately decreased to improve detection sensitivity.

[0156] Furthermore, the reversibility of lithium plating can be used to make a judgment. The expansion caused by lithium plating is irreversible to some extent. If the battery volume shrinks less than the previous expansion during subsequent discharge or resting, it indicates that irreversible expansion (dead lithium formation) has occurred, which can further confirm the occurrence of lithium plating.

[0157] The lithium plating detection method based on the differential expansion coefficient model and adaptive threshold provided in this embodiment can detect lithium plating more accurately and sensitively, providing a reliable basis for the decision-making of RL agents.

[0158] Example 8

[0159] Based on Example 1, this embodiment provides a detailed explanation of the specific strategies for adjusting the test waveform parameters of the RL agent to achieve more refined control.

[0160] According to the method described in this embodiment, the parameters of the high-current test waveform adjusted in real time by the RL agent include: adjusting the amplitude, rise slope, and pulse frequency of the high-current test waveform. These three parameters have different effects on the stress level and internal state of the battery, and the RL agent needs to coordinately adjust these parameters to achieve optimal control.

[0161] Amplitude: The current amplitude is the most important factor determining the test stress. A larger amplitude results in greater test stress, but also increases the risk of lithium plating and thermal runaway. The RL agent controls the overall stress level by adjusting the amplitude. When a risk is detected, reducing the amplitude is the most direct and effective intervention.

[0162] Slew Rate: The slew rate represents the rate of current change (di / dt). A larger slew rate indicates a faster current change, resulting in a greater impact on the battery. Rapid current changes lead to uneven distribution of lithium-ion concentration within the battery, exacerbating local polarization and increasing the risk of lithium plating. The RL agent can control the dynamic characteristics of stress by adjusting the slew rate. When the risk of lithium plating is detected, reducing the slew rate can make the current change smoother, promoting uniform lithium-ion diffusion and mitigating local polarization.

[0163] Pulse Frequency: Pulse frequency represents the number of pulses (or pulse intervals) per unit time. Pulse frequency affects the battery's thermal accumulation effect and relaxation process. Higher frequencies result in shorter pulse intervals, shorter relaxation times, and easier heat accumulation, leading to increased temperature. The RL agent can control the battery's thermal state by adjusting the pulse frequency. When excessively high temperatures or signs of impending thermal runaway are detected, reducing the pulse frequency can increase the relaxation time, which is beneficial for heat dissipation and lowers the temperature.

[0164] During training, the RL agent learns the influence of these three parameters on the battery state and adjusts these parameters collaboratively based on the current feedback information.

[0165] For example, in the initial stages of testing, when the battery is in good condition, the RL agent might simultaneously increase the amplitude, rise time slope, and pulse frequency to rapidly increase the test stress. Upon detecting slight signs of lithium plating, the agent might preferentially reduce the rise time slope to improve lithium-ion distribution while maintaining a high amplitude and frequency to sustain test efficiency. If the signs of lithium plating worsen, the agent will further reduce the amplitude. If excessive temperature is detected, the agent will reduce the pulse frequency to enhance heat dissipation.

[0166] Through this multi-parameter collaborative adjustment strategy, the RL agent can achieve fine-grained control of test stress, maximizing the battery's performance potential within safety boundaries. When implementing the RL agent, algorithmic frameworks supporting continuous multi-dimensional action spaces, such as DDPG or PPO, can be employed. The agent can directly output a continuous action vector containing target values ​​for amplitude, slope, and frequency.

[0167] Example 9

[0168] Based on Example 1, this embodiment provides a detailed explanation of the reward function design and security response mechanism for the RL agent to ensure the safety and efficiency of the testing process.

[0169] According to the method described in this embodiment, the reward function of the RL agent is set to maximize the test stress, and the judgment results of the thermal runaway precursor and the judgment results of lithium plating in the feedback are used as penalty items.

[0170] The design details of the reward function are as follows: Performance reward: Set to be proportional to the test stress applied at the current moment. The definition of test stress can be determined based on the test objective. For example, in accelerated life testing, it can be defined as the square of the current multiplier or the cumulative charge throughput. To encourage the agent to explore higher stress levels, a non-linear reward mechanism can be introduced.

[0171] Safety penalties: consist of two parts, corresponding to precursors of thermal runaway and lithium plating, respectively.

[0172] Thermal runaway penalty: When a precursor to thermal runaway is detected, a very large fixed penalty value is applied. This value should be much larger than the maximum possible performance reward to ensure that the agent prioritizes avoiding thermal runaway.

[0173] Lithium plating penalty: When lithium plating is detected, a moderate penalty value is applied. For finer control over the degree of lithium plating, the penalty value can be set proportional to the severity of the plating. The severity of lithium plating can be measured by the extent to which the real-time expansion rate exceeds the expected value. The greater the exceedance, the greater the penalty.

[0174] The sum of the reward functions is the performance reward term minus the safety penalty term. The goal of the RL agent is to maximize the cumulative reward.

[0175] The security response mechanism is designed as follows: This invention employs a hierarchical security response mechanism, taking different intervention measures based on the severity of the risk.

[0176] When signs of impending thermal runaway are detected, the risk level is highest. At this point, the system immediately triggers an emergency stop mechanism, terminating the test, cutting off current, and activating the cooling system to prevent an accident. Simultaneously, the RL agent is subjected to the maximum penalty to enhance its ability to avoid thermal runaway risks. This logic can be implemented within a standalone safety monitoring system with the highest control authority, exceeding that of the RL agent.

[0177] When lithium plating is detected, the risk level is considered medium. In this case, the system does not immediately terminate the test but guides the RL agent to adjust the test strategy through a reward function. The RL agent prioritizes reducing the amplitude of the high-current test waveform. Reducing the amplitude can directly suppress lithium plating. If the lithium plating indications disappear after reducing the amplitude, the agent may gradually try to increase the amplitude to continue exploring the safety boundary. If the lithium plating indications still exist after reducing the amplitude, the agent will further reduce the amplitude or adjust other parameters (such as rise time slope and pulse frequency). The test is only terminated when all corrective measures are ineffective.

[0178] Through this carefully designed reward function and hierarchical safety response mechanism, RL agents can strictly adhere to safety constraints while pursuing efficient testing, ensuring the safety and reliability of the testing process.

[0179] Example 10

[0180] This embodiment provides a battery charging and discharging system based on high-current testing, used to implement the methods described in embodiments one through nine. (Refer to...) Figure 2 The system includes the following modules: Reinforcement Learning (RL) Agent Module 201: This module is the decision-making core of the system, used to dynamically generate high-current test waveforms. It is typically deployed on a high-performance computing server or workstation, equipped with dedicated AI acceleration hardware (such as a GPU). The RL agent module runs a trained reinforcement learning model. It receives state information and risk assessment results from the feedback control module and calculates the optimal test waveform parameters through the policy network. This module also needs data storage and management functions to store training data, model parameters, and test records.

[0181] Test Execution Module 202: This module is used to apply the high-current test waveform to the target battery. This module mainly includes a high-power, high-precision bidirectional DC power supply (battery charging / discharging equipment) and environmental control equipment (such as a constant temperature chamber). The battery charging / discharging equipment needs to have fast response capabilities (e.g., millisecond-level current rise time) and high-precision current control capabilities, capable of generating the required current waveform in real time according to the instructions sent by the RL agent module. To meet the requirements of the highly dynamic test waveform dynamically generated by the RL agent, the battery charging / discharging equipment needs to meet the following key performance indicators: current response speed (rise time) should be less than 5 milliseconds, preferably less than 1 millisecond; current control accuracy should be better than 0.05% of full scale; data sampling rate should reach 1 kHz or higher, and maintain strict synchronization with the multi-dimensional sensing module (synchronization error less than 1 millisecond).

[0182] Multidimensional Sensing Module 203: This module is used to synchronously acquire multidimensional physical field information of the target battery. This module includes acoustic sensors and strain sensors. Acoustic sensors (such as PZT sensors) are used to acquire acoustic emission signals. Strain sensors (such as FBG sensors or laser displacement sensors) are used to acquire volume expansion data. In addition, this module also includes temperature sensors and voltage / current sensors. The multidimensional sensing module requires a high-speed, high-precision data acquisition (DAQ) device to ensure data quality and synchronization.

[0183] Real-time Analysis Module 204: This module processes and analyzes the acquired data in real time to assess the battery's internal state and risk level. It is typically deployed on edge computing devices (such as FPGAs or DSPs) to meet real-time requirements. The real-time analysis module runs advanced signal processing algorithms and risk assessment models.

[0184] Specifically, the real-time analysis module is used to analyze the spectral fingerprint and energy release rate of the acoustic emission signal to determine precursors of thermal runaway. This includes time-frequency domain analysis (such as wavelet transform), feature extraction, and threshold-based judgment logic.

[0185] Simultaneously, the real-time analysis module is used to calculate the real-time expansion rate of the volume expansion data to determine lithium plating. This includes signal filtering (such as SG filtering), first-order differentiation, calculation of the expected electrochemical expansion rate, and threshold-based judgment logic.

[0186] Feedback Control Module 205: This module feeds back the risk assessment results obtained by the real-time analysis module to the RL agent module, forming a closed-loop control. This module is responsible for data transmission and communication protocol implementation, ensuring low latency and reliable transmission of feedback information.

[0187] The feedback control module feeds back the judgment results of the thermal runaway precursor and the lithium plating to the RL agent module, enabling the RL agent module to adjust the parameters of the high-current test waveform in real time and adaptively apply test stress within the safety boundary. This module is also responsible for executing safety protocols, such as directly sending an emergency stop command to the test execution module when a thermal runaway precursor is detected.

[0188] The battery charging and discharging system based on high-current testing provided in this embodiment achieves intelligent, adaptive, and safe high-current testing through the coordinated work of various modules, providing powerful tool support for battery research and development and quality control.

[0189] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A battery charging and discharging method based on high current testing, characterized in that, include: The test waveform is dynamically generated and applied to the target battery by a reinforcement learning RL agent; During the application of the high-current test waveform, the acoustic emission signal and volume expansion data of the target battery are simultaneously acquired; Analyze the spectral fingerprint and energy release rate of the acoustic emission signal to determine whether there are precursors to thermal runaway; The real-time expansion rate of the volume expansion data is calculated and compared with the expected electrochemical expansion rate to determine whether lithium plating exists. The results of the judgment on the precursor to thermal runaway and the results of the judgment on lithium plating are fed back into the RL agent; The RL agent adjusts the parameters of the high-current test waveform in real time based on the feedback, so as to adaptively apply test stress within the safety boundary; The steps for analyzing the spectral fingerprint and energy release rate of the acoustic emission signal include: Time-frequency domain analysis was performed on the acoustic emission signal to extract the characteristic frequency components in the high-frequency band; Monitor the energy release rate of the high-frequency band; When the energy release rate suddenly changes and exceeds a preset safety threshold, it is determined that there are signs of impending thermal runaway. The steps for calculating the real-time expansion rate of the volume expansion data include: Strain signals are acquired using high-resolution strain sensors, and these strain signals characterize the volume expansion data. The real-time expansion rate is obtained by taking the first derivative of the strain signal; The steps to determine whether lithium plating is present include: Calculate the expected electrochemical expansion rate based on the current current density and state of charge (SOC). When the real-time expansion rate exceeds the expected electrochemical expansion rate and reaches a preset lithium plating threshold, lithium plating is determined to exist. The reward function of the RL agent is set to maximize the test stress, and the judgment results of the thermal runaway precursor and the judgment results of lithium plating in the feedback are used as penalty items. If signs of impending thermal runaway are detected, the test should be terminated immediately. When lithium plating is detected, the RL agent preferentially reduces the amplitude of the high-current test waveform.

2. The method according to claim 1, characterized in that, The high-current test waveform includes a high-rate pulse sequence for evaluating battery power performance, wherein the peak current of the high-rate pulse sequence is greater than 2C.

3. The method according to claim 1, characterized in that, The high-current test waveform is used to simulate the dynamic load under actual working conditions. The goal of the RL agent is to make the test stress applied by the generated high-current test waveform approximate the stress mode of the dynamic load.

4. The method according to claim 1, characterized in that, During the application of the high-current test waveform, the following is also included: Collect the terminal voltage and current of the target battery; The DC internal resistance DCR is calculated based on the terminal voltage and current, and the trend of the DCR change is used to assess the state of health (SOH) of the battery.

5. The method according to claim 1, characterized in that, The parameters of the high-current test waveform that are adjusted in real time by the RL agent include: adjusting the amplitude, rise slope, and pulse frequency of the high-current test waveform.

6. A battery charging and discharging system based on high-current testing, characterized in that, include: A reinforcement learning (RL) agent module is used to dynamically generate high-current test waveforms; The test execution module is used to apply the high-current test waveform to the target battery; A multi-dimensional sensing module, including an acoustic sensor and a strain sensor, is used to simultaneously acquire the acoustic emission signal and volume expansion data of the target battery; The real-time analysis module is used to analyze the spectral fingerprint and energy release rate of the acoustic emission signal to determine the precursors of thermal runaway, and to calculate the real-time expansion rate of the volume expansion data to determine lithium plating. The feedback control module is used to feed back the judgment results of the thermal runaway precursor and the judgment results of the lithium plating to the RL intelligent agent module, so that the RL intelligent agent module adjusts the parameters of the high current test waveform in real time and adaptively applies test stress within the safety boundary. The steps for analyzing the spectral fingerprint and energy release rate of the acoustic emission signal include: Time-frequency domain analysis was performed on the acoustic emission signal to extract the characteristic frequency components in the high-frequency band; Monitor the energy release rate of the high-frequency band; When the energy release rate suddenly changes and exceeds a preset safety threshold, it is determined that there are signs of impending thermal runaway. The steps for calculating the real-time expansion rate of the volume expansion data include: Strain signals are acquired using high-resolution strain sensors, and these strain signals characterize the volume expansion data. The real-time expansion rate is obtained by taking the first derivative of the strain signal; The steps to determine whether lithium plating is present include: Calculate the expected electrochemical expansion rate based on the current current density and state of charge (SOC). When the real-time expansion rate exceeds the expected electrochemical expansion rate and reaches a preset lithium plating threshold, lithium plating is determined to exist. The reward function of the RL agent is set to maximize the test stress, and the judgment results of the thermal runaway precursor and the judgment results of lithium plating in the feedback are used as penalty items. If signs of impending thermal runaway are detected, the test should be terminated immediately. When lithium plating is detected, the RL agent preferentially reduces the amplitude of the high-current test waveform.

Images