Battery health real-time diagnosis method in battery replacement process based on voltage internal resistance correlation analysis

Abstract

The invention relates to the technical field of battery science, in particular to a battery health real-time diagnosis method based on voltage internal resistance correlation analysis in the battery replacement process. The method comprises the following specific implementation steps: firstly, capturing a high-frequency transient signal at the moment of battery connection and disconnection, and carrying out real-time evaluation and error correction on the high-frequency transient signal based on a signal symmetry evaluation mechanism; thirdly, a battery health model is designed, the model generates a nonlinear internal resistance spectrum through an AI model part, and a health state physical model is used as priori knowledge for constraint; and the health state physical model outputs an initial health state according to the internal resistance spectrum adjustment parameters. And finally, training the battery health verification model by utilizing meta learning, outputting a health state predicted value by the model, and obtaining a fusion result of an error of the battery health verification model and an error of the health state predicted value through an error intelligent fusion method. And performing weighted average on the health state prediction value and the initial battery health state according to a fusion result to determine a final battery health state.

Description

Technical Field

[0001] The present invention relates to the field of battery science and technology, and specifically to a real-time diagnosis method for battery health during a battery replacement process based on voltage-internal resistance correlation analysis. Background Art

[0002] With the rapid development of electric vehicles, managing the performance and health of batteries, a core component, has become increasingly important. Especially in battery swapping scenarios, quickly and accurately assessing the health of battery packs is key to ensuring operational safety and optimizing asset management.

[0003] Currently, the main methods for assessing battery health include internal resistance analysis, ampere-hour integration, and open-circuit voltage. Diagnosing health status based on internal resistance is a common and effective approach. However, these methods have some practical challenges. Traditional diagnostic models are typically fixed and lack adaptability, resulting in poor diagnostic performance for batteries of different models and chemistries.

[0004] To address these issues, some research has attempted to incorporate machine learning. However, purely data-driven models suffer from reliance on large amounts of historical data, poor generalization, and a lack of physical interpretability. Furthermore, during the battery swap process, the transient signals generated at the moment of battery connection and disconnection contain rich electrochemical information, but this signal is often overlooked, and existing technologies struggle to capture high-quality transient signals, preventing them from being fully utilized for in-depth analysis.

[0005] To this end, a real-time diagnosis method for battery health during the battery replacement process based on voltage-internal resistance correlation analysis is proposed. Summary of the Invention

[0006] The purpose of the present application is to provide a battery health real-time diagnosis method for battery swapping process based on voltage internal resistance correlation analysis. In order to solve the problems existing in the prior art, the present application first constructs a transient signal capture model, learns and dynamically adjusts the capture duration by obtaining the battery operating parameters. The model is triggered by the control instruction of the battery swapping station, captures the high-frequency transient signal at the moment of connecting and disconnecting the battery, and performs real-time evaluation and error correction on the high-frequency transient signal based on the signal symmetry evaluation mechanism to ensure data quality. Then, a battery health model is designed, and the signal is input into the model. The model generates a nonlinear internal resistance spectrum through the AI part, and uses the health state physical model as prior knowledge for constraint. The health state physical model adjusts the parameters according to the internal resistance spectrum, thereby accurately compensating the OCV, and outputs the initial health state of the battery. Finally, a battery health verification model that can quickly adapt to new batteries is trained using meta-learning. The model takes the nonlinear internal resistance spectrum and the initial health state as input, where the initial health state is used as a guide signal. The model outputs a health state prediction value, and the error of the battery health verification model and the error of the health state prediction value are fused through an error intelligent fusion method, and the final battery health state is determined by weighted average of the health state prediction value and the initial battery health state according to the fusion result.

[0007] To achieve the above purpose, the present application provides the following technical scheme: a battery health real-time diagnosis method for battery swapping process based on voltage internal resistance correlation analysis, the method comprising:

[0008] Obtain high-frequency transient signals during the battery swapping process, and simultaneously perform real-time evaluation and error correction on the high-frequency transient signals based on the signal symmetry evaluation mechanism;

[0009] Construct a battery health diagnosis model including an AI model and a health state physical model, analyze the nonlinear relationship between the transient signal and the battery parameters through the AI model, generate a nonlinear internal resistance spectrum, adjust the internal parameters of the health state physical model according to the nonlinear internal resistance spectrum, and constrain the AI model based on the physical loss of the health state physical model and compensate the open circuit voltage; analyze the internal parameters and the compensated open circuit voltage through voltage internal resistance correlation, and output the initial health state;

[0010] Train a battery health verification model through meta-learning, use the nonlinear internal resistance spectrum and the initial health state as input, and obtain the health state prediction value and the fusion error through an error intelligent fusion method, and determine the final health state by weighted average of the health state prediction value and the initial health state according to the fusion error.

[0011] Preferably, the specific implementation process of obtaining the high-frequency transient signal of the battery during the battery replacement process includes: when the time length selected by the transient signal capture model captures the complete transient signal, a positive reward is obtained; if the capture time length is too long, resulting in data redundancy and the capture time length is too short, resulting in incomplete signal, a negative reward is obtained. Through continuous interaction with the environment and trial and error, the transient signal capture model eventually obtains a capture time length selection method that can maximize the reward under various conditions; collects battery model, real-time temperature, and connection status data, and inputs the data into the deep learning transient signal capture model in the form of a multidimensional time series. The transient signal capture model determines the optimal time length for capturing the battery's high-frequency transient signal based on the input data and adjusts the capture time length.

[0012] Preferably, high-frequency transient signals are evaluated and corrected in real time based on a signal symmetry evaluation mechanism, and the specific implementation process of capturing high-frequency transient signals includes: using the contactor closing and disconnecting control instructions of the battery swap station as trigger conditions, identifying the initial contact and separation moments of the battery pack connecting and disconnecting with the external electrical system, and capturing high-frequency voltage and current transient signals according to the optimal capture time; at the same time, the transient signal capture model analyzes the voltage jump amplitude and attenuation time constant characteristics of the two transient processes based on the transient data at the moment of connection and disconnection. When these parameters are significantly asymmetric at the moment of connection and disconnection, it indicates that there is an abnormality in the data acquisition process, and the transient signal capture is immediately re-triggered.

[0013] Preferably, the specific implementation process of designing a battery health model includes: the battery health model is composed of an AI model and a health status physical model, the AI ​​model processes time series data and learns the complex nonlinear relationship between transient signals and battery parameters; the health status physical model is based on a second-order RC equivalent circuit.

[0014] Preferably, the AI ​​model analyzes the nonlinear relationship between the transient signal and the battery parameters to generate a nonlinear internal resistance spectrum, and constrains it through the health state physical model. The specific implementation process includes: performing battery replacement transient tests on batteries in different health states, charge states and temperatures, collecting high-frequency voltage and current data, and using an electrochemical workstation to measure the internal resistance spectrum of the battery under the same operating conditions as a "true value" label; by fitting the internal resistance spectrum data, the true parameter value of the health state physical model is obtained; a composite loss function consisting of data loss and physical loss is constructed, the data loss measures the difference between the internal resistance spectrum predicted by the AI ​​model and the true value, and the physical loss measures whether the parameters predicted by the AI ​​model conform to the behavior of the health state physical model; the transient signal and health state, charge state, and temperature data are used as input to train the AI ​​model.

[0015] Preferably, the health state physical model adjusts internal parameters according to the nonlinear internal resistance spectrum, compensates for the open circuit voltage, and outputs the initial health state through voltage-internal resistance correlation analysis of the internal parameters and the compensated open circuit voltage. The specific implementation process includes: the AI ​​model generates a complete nonlinear internal resistance spectrum in real time according to the input data, and inputs the generated nonlinear internal resistance spectrum into the health state physical model. The health state physical model adjusts and calibrates the internal ohmic resistance, charge transfer resistance and double layer capacitance parameters in real time according to the high-frequency semicircle diameter and low-frequency slope characteristics in the internal resistance spectrum; the health state physical model uses the adjusted parameters to decompose the total voltage change into three parts caused by ohmic impedance, electrochemical polarization and open circuit voltage; by separating the voltage drop caused by ohmic impedance and polarization from the total voltage, the voltage-internal resistance correlation analysis outputs the initial health state of the battery pack.

[0016] Preferably, meta-learning is used to train a battery health verification model, with the nonlinear internal resistance spectrum and the initial battery health state as inputs, and the initial battery health state provides prior knowledge as a guiding signal to obtain a health state prediction value. The specific implementation process includes: using the meta-learning MAML algorithm to perform alternating training on multiple different tasks, and the Bayesian neural network-based model learns the common laws in different battery aging processes, and takes the nonlinear internal resistance spectrum and the initial health state as input; the initial health state provides a priori information to help the prediction model quickly focus on the historical battery data closest to it, and finally the battery health verification model learns a priori knowledge of the weight distribution that can adapt to the new task; when the model faces a new task, the battery health verification model uses the learned prior knowledge of the weight distribution to fine-tune through data to obtain a weight distribution that adapts to the new task. During prediction, the battery health verification model obtains prediction results through Monte Carlo sampling, the standard deviation of these results is used as a quantitative indicator of the model error, and the average value of the prediction results is output as the health state prediction value.

[0017] Preferably, a fusion result of the error of the battery health verification model and the error of the health status prediction value is obtained by an error intelligent fusion method, and the health status prediction value and the initial battery health status are weighted averaged according to the fusion result to determine the final battery health status. The specific implementation process includes: comparing the prediction result of the battery health verification model with the initial health status, taking the absolute value of the difference between the health status prediction value and the initial health status as the data error, and weighted averaging the battery health verification model error and the data error to obtain the final error; weighted averaging the health status prediction value and the initial health status according to the error size, when the error is less than the preset threshold, it indicates that the prediction result is highly credible, and the weight of the health status prediction value is increased; when the error is greater than the preset threshold, it indicates that there is an abnormality in the data, and the weight of the initial health status is increased, and the final health status value is calculated through the weighted average.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] 1. In the prior art, during the battery replacement process, the transient signals generated at the moment of battery connection and disconnection contain rich electrochemical information, but it is usually difficult to capture high-quality transient signals. The present invention adopts a transient signal capture model based on reinforcement learning. This model can dynamically adjust the capture time according to the battery operating parameters to ensure that complete, high-quality transient signals can be captured under various complex operating conditions. In addition, based on the signal symmetry evaluation mechanism, diagnostic errors caused by data anomalies are avoided, and the reliability of the data is improved.

[0020] 2. In the prior art, the estimation of internal resistance is easily affected by temperature, state of charge, and dynamic working conditions, which greatly reduces the accuracy and universality of the diagnostic results. The present invention designs a battery health model. By taking the physical model of the health status as the prior knowledge of the AI ​​model, it constrains the learning process of the AI ​​model and solves the problems of poor generalization ability and lack of physical interpretability of pure data-driven models. Through this two-way collaboration, the model can more accurately separate the parts of the voltage change caused by ohmic impedance, electrochemical polarization, and open-circuit voltage changes, and achieve accurate compensation for the open-circuit voltage, thereby providing a more reliable basis for health status diagnosis.

[0021] 3. Existing diagnostic models are typically fixed, lack adaptability, and rely on large amounts of historical data, making them difficult to adapt to the diversity of battery models and chemistries. This paper utilizes meta-learning techniques to train a diagnostic model that can rapidly adapt to new battery types. This model requires only a small amount of sample data for rapid fine-tuning, addressing the limitations of traditional models. Furthermore, by integrating multiple sources of error through intelligent error fusion, it is possible to more comprehensively assess the errors in the current diagnostic results, leading to more robust decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A flowchart of a battery health real-time diagnosis method based on voltage-internal resistance correlation analysis in a battery replacement process;

[0023] Figure 2 A transient signal capture and real-time evaluation mechanism based on signal symmetry evaluation flowchart are provided in the embodiments of the present application.

[0024] Figure 3 A design diagram of a battery health model is provided in the embodiments of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0026] Please refer to Figures 1 to 3 The present application provides a battery health real-time diagnosis method based on voltage-internal resistance correlation analysis in a battery replacement process, and the technical solutions are as follows:

[0027] The battery health real-time diagnosis method based on voltage-internal resistance correlation analysis in a battery replacement process is described with reference to Figure 1 The specific implementation steps of the method proposed by the present application include:

[0028] S1, a transient signal capture model is constructed, battery operation parameters are obtained, the parameters are input into the model, and the capture duration is learned and dynamically adjusted;

[0029] S2, triggered by a battery replacement station control instruction, high-frequency transient signals at the moment of battery connection and disconnection are identified and captured, and the signals are real-time evaluated and corrected based on a signal symmetry evaluation mechanism;

[0030] S3, a battery health model is designed, an AI model generates a nonlinear internal resistance spectrum, and a health state physical model is used as prior knowledge for constraint, and the health state physical model adjusts internal parameters according to the nonlinear internal resistance spectrum;

[0031] S4, a battery health verification model is trained by meta-learning, the nonlinear internal resistance spectrum and the initial battery health state are used as inputs, the initial battery health state provides prior knowledge as a guide signal, and a health state prediction value is obtained;

[0032] S5, a fusion error is obtained by an error intelligent fusion method, and the health state prediction value and the initial health state are weighted and averaged according to the fusion error to determine the final health state.

[0033] Example 1

[0034] This embodiment provides a specific application of a real-time diagnosis method for battery health during the battery replacement process based on voltage-internal resistance correlation analysis. Its typical application scenario is a batch of lithium iron phosphate battery packs operating at a new energy large and medium-sized passenger bus battery replacement station, ensuring that the battery packs can complete an accurate assessment of their health status within a few minutes of rapid battery replacement at the station.

[0035] Furthermore, a transient signal capture model is constructed to obtain battery operating parameters, input the parameters into the model, learn and dynamically adjust the capture duration, corresponding to the above step S1, the specific process is as follows:

[0036] The battery operating parameters are obtained in real time from the on-board battery management system and the battery swap station control unit. These parameters include battery model (such as LF-100Ah lithium iron phosphate battery pack), real-time temperature, and control instructions for closing and opening the battery swap station contactor. These parameters are then input into the constructed deep learning transient signal capture model in the form of multi-dimensional time series.

[0037] The deep learning transient signal capture model combines a hybrid architecture of convolutional neural networks and deep Q networks. The CNN part is used to extract working condition features from time series data, and the DQN part serves as a decision network to select the best action based on these features. The "action" of the model is to select different capture durations. For example, the capture duration can be discretized into a series of options, such as 0.5 seconds, 1.0 seconds, 1.5 seconds, etc. A reward mechanism is designed to guide the learning of the model. When the duration selected by the model just captures the complete transient signal and enables the subsequent health status diagnosis accuracy to reach the preset target, the model receives a positive reward. If the capture duration is too long, resulting in data redundancy, or the duration is too short, resulting in incomplete signals, the model receives a negative reward. The model optimizes its decision-making strategy through continuous interaction and trial and error with the actual battery swapping environment. For example, at room temperature, the model found through multiple attempts that a capture duration of 0.5 seconds was sufficient to obtain a complete signal with minimal data redundancy, and thus used this as the optimal action under this working condition. At low temperatures, due to the slower electrochemical response inside the battery, the model will learn that a longer capture time is required to capture the complete polarization decay curve and use it as the optimal strategy.

[0038] By building a transient signal capture model based on reinforcement learning, the problem of fixed capture time and difficulty in adapting to different working conditions is solved. The capture time can be dynamically adjusted to ensure the acquisition of complete and high-quality transient signals, realizing the transformation from "passive rules" to "active intelligent decision-making". This not only improves the efficiency of data collection, but also ensures that the most valuable information can be quickly obtained in time-sensitive scenarios such as battery swap stations, providing a reliable data foundation for subsequent accurate diagnosis.

[0039] Furthermore, the high-frequency transient signals at the moment of battery connection and disconnection are identified and captured using the control command of the battery swap station as a trigger, and the signals are evaluated and corrected in real time based on the signal symmetry evaluation mechanism. Corresponding to the above step S2, the specific process is as follows:

[0040] The battery swap station's contactor closing and disconnection control commands are used as trigger conditions to accurately identify the initial moments of connection and disconnection between the battery pack and the external electrical system. Based on the optimal capture duration obtained by dynamically adjusting the model in the previous step, the high-frequency voltage and current transient signals at the moment of connection and disconnection are captured. The transient data at the moment of connection and disconnection are analyzed based on the signal symmetry assessment mechanism, extracting key characteristic parameters such as the voltage jump amplitude and decay time constant. If the data quality is judged to be poor, corrective measures are immediately taken, the command is reissued, and the transient signal capture is repeated at the next moment until data that meets the symmetry requirements is obtained.

[0041] Specifically, the signal is evaluated in real time based on the signal symmetry evaluation mechanism to determine the signal quality. The processing process is as follows:

[0042] The voltage jump amplitude primarily reflects the ohmic internal resistance and contact resistance, while the decay time constant is related to the electrochemical polarization process. Ideally, the connection and disconnection transient processes should be symmetrical. The model compares the characteristic parameters of these two processes. When the model finds significant asymmetry in these characteristic parameters, such as the voltage jump amplitude and decay time constant, between the connection and disconnection moments, it indicates an anomaly in the data acquisition process. These anomalies stem from a variety of factors, resulting in poor data quality.

[0043] By making full use of the transient data at the moment of connection and disconnection, and conducting symmetry analysis on it, it can automatically determine whether the data is interfered with by abnormalities, providing a guarantee for the subsequent correct triggering of information recapture and ensuring that the data input into the diagnostic model is of high quality.

[0044] The signal symmetry assessment mechanism provides self-checking and self-correction capabilities, resolving the issue of inability to assess data acquisition quality in real time. It automatically determines if data is subject to abnormal interference and immediately recaptures it, ensuring high-quality data input into the battery health model and significantly improving diagnostic robustness and reliability.

[0045] Furthermore, a battery health model is designed. The AI ​​model generates a nonlinear internal resistance spectrum and uses the health state physical model as a priori knowledge for constraints. The health state physical model adjusts internal parameters based on the nonlinear internal resistance spectrum. Corresponding to step S3 above, the specific process is as follows:

[0046] A battery health model is constructed. The AI ​​model consists of a convolutional neural network, which processes time series data and learns the complex nonlinear relationships between transient signals and battery parameters. The health state physical model is based on a second-order RC equivalent circuit model, consisting of key components such as ohmic resistance, charge transfer resistance, and double-layer capacitance, providing the fundamental physical laws governing battery behavior.

[0047] In a laboratory environment, lithium iron phosphate battery packs in different health states, states of charge, and temperatures were subjected to battery replacement transient tests, and high-frequency voltage and current data were collected. At the same time, under the same operating conditions, the internal resistance spectrum of the battery was measured using an electrochemical workstation as the "true value" label of the model. By fitting the internal resistance spectrum data, the true parameter values ​​of the equivalent circuit model under different operating conditions, such as ohmic resistance and charge transfer resistance, were obtained. A composite loss function consisting of data loss and physical loss was constructed. Data loss measures the difference between the internal resistance spectrum predicted by the AI ​​model and the true EIS value. Physical loss measures whether the parameters predicted by the AI ​​model conform to the behavior of the physical model in a healthy state. Transient signals as well as health status, state of charge, and temperature data are used as input to train the AI ​​model. During the training process, the model must minimize the gap with the true internal resistance spectrum while meeting physical constraints.

[0048] The captured high-quality high-frequency transient signal, along with the current battery operating parameters, is input into the trained battery health model in real time. Based on the input data, the AI ​​model generates a complete nonlinear internal resistance spectrum in real time. This internal resistance spectrum is the battery electrochemical "fingerprint" learned by the model that best matches the current operating conditions. The health status physical model will adjust and calibrate parameters such as internal ohmic resistance, charge transfer resistance, and double-layer capacitance in real time based on features such as the high-frequency semicircle diameter and low-frequency slope in the internal resistance spectrum. The health status physical model uses the adjusted parameters to decompose the total voltage change into three parts caused by ohmic impedance, electrochemical polarization, and open-circuit voltage. By separating the voltage drop caused by ohmic impedance and polarization from the total voltage, the true open-circuit voltage is obtained, thereby achieving compensation for the open-circuit voltage. Based on the adjusted internal resistance parameters and the compensated open-circuit voltage, the initial health status of the battery pack is finally output.

[0049] The battery health model enables the model to make predictions consistent with physical laws even when training data is insufficient, thus addressing the problem of poor generalization. It also empowers the model with adaptability, enabling it to maintain high accuracy in dynamic and complex environments. The combination of these two makes diagnostic results not only accurate but also physically interpretable, enhancing their credibility.

[0050] By designing a composite loss function, the AI ​​model is able to generate a nonlinear internal resistance spectrum and be constrained by the physical model of the health state, solving the problems of pure data-driven models' lack of physical interpretability and poor generalization ability. This allows the model to make reasonable predictions that conform to physical laws when data is scarce or facing new working conditions.

[0051] The two-way real-time correction of the health status physical model and the AI ​​model achieves precise compensation for the open-circuit voltage, can more accurately decompose the total voltage change into parts from different physical sources, eliminates interference from factors such as ohmic impedance and polarization, and makes health status diagnosis based on open-circuit voltage and internal resistance more reliable, significantly improving the accuracy of the diagnostic results.

[0052] Furthermore, meta-learning is used to train a battery health verification model, with the nonlinear internal resistance spectrum and the initial battery health state as input. The initial battery health state provides prior knowledge as a guidance signal to obtain a health state prediction value. Corresponding to the above step S4, the specific process is as follows:

[0053] Meta-learning algorithms, such as MAML, need to be trained on multiple different tasks to learn common patterns. Each task corresponds to a specific battery pack model or chemical system, and the Bayesian neural network-based model is trained alternately on the datasets of these different tasks. In this way, the model learns the common patterns in the aging process of different batteries, rather than the unique characteristics of a specific battery. The dynamic nonlinear internal resistance spectrum and the initial health state are input into the meta-learning model. The initial health state serves as a guiding signal, providing prior information for the model. The meta-learning model uses the general knowledge learned in the offline stage to quickly fine-tune with a small amount of sample data to adapt to the characteristics of the new battery. For example, the model may only need one or two iterations to adjust from the general aging pattern to the aging pattern of this specific battery. The fine-tuned model outputs a health state prediction value.

[0054] After training, the battery health verification model learns prior knowledge about the weight distribution that is suitable for new tasks. When the model is faced with a new task, it uses this prior knowledge to fine-tune the model using data to obtain a weight distribution that is suitable for the new task. During prediction, the battery health verification model uses Monte Carlo sampling to obtain prediction results. The standard deviation of these results is used as a quantitative indicator of model error, and the average of these multiple prediction results is output as the health status prediction value.

[0055] Specifically, the initial battery health status provides prior knowledge as a guiding signal, and the prediction result is obtained based on the physical loss of the battery. The processing process is as follows:

[0056] The dynamic nonlinear internal resistance spectrum generated by the battery health model and the initial health state are taken as inputs of the battery health verification model, and the initial health state is taken as a guide signal to provide prior information for the battery health verification model; the initial health state information is used by the battery health verification model to quickly focus on the closest historical battery data, so that the model can quickly identify the aging mode and parameters most relevant to the current battery state, and the model is made to comply with physical constraints and make accurate predictions by adding physical loss in the loss function; through focusing and guiding, the battery health verification model uses the general knowledge learned in the offline stage to quickly fine-tune through a small amount of sample data, adapts to the characteristics of new batteries, and the battery health verification model after quick fine-tuning outputs a health state prediction value;

[0057] The initial health state is not only a result of diagnosis, but also a key input for the next diagnosis link. Through this guiding mechanism, the model can more accurately capture the aging mode and parameters of new batteries, rather than blindly learning in all historical data, which enables the model to output more accurate health state prediction values and significantly improves the adaptation speed and prediction accuracy of the battery health verification model.

[0058] The battery health verification model is trained through meta-learning to quickly adapt to new batteries, solving the problem that traditional models are difficult to cope with the diversity of different batteries, while also avoiding dependence on a large amount of historical data, so that after a new battery is connected to the system, it only needs a small amount of data for quick and accurate diagnosis, greatly reducing the cost of model deployment and maintenance.

[0059] Further, a fusion error is obtained through an error intelligent fusion method, and the health state prediction value and the initial health state are weighted and averaged according to the fusion error to determine the final health state, corresponding to step S5, the specific process is as follows:

[0060] The prediction result of the battery health verification model is compared with the initial health state, and the absolute value of the difference between the health state prediction value and the initial health state is taken as the data error, and the final error is obtained by weighted averaging the battery health verification model error and the data error. When the final error is large, it means that the current working condition is special or the collected data is abnormal. According to the uncertainty, the prediction result and the initial health state are weighted and averaged, and finally the health state of the battery pack is output through the above weighted average calculation.

[0061] Specifically, the prediction result and the initial health state are weighted and averaged, and the processing process is as follows:

[0062] When the error is less than the preset threshold, the prediction result is considered to be highly credible. At this time, the prediction result of the battery health verification model is given a higher weight, and the initial health status is given a lower weight; when the error is greater than the preset threshold, there are special working conditions or abnormal data. At this time, the initial health status is given a higher weight to utilize the robustness of its physical prior information to correct the prediction of the AI ​​model, and finally the health status of the battery pack is output through weighted average calculation.

[0063] By giving a higher weight to the initial health state provided by the health state physical model, which has more physical prior information, its robustness can be effectively utilized to correct the prediction of the AI ​​model, avoiding the deterioration of the final result when the AI ​​model prediction is highly inaccurate, while also making full use of the AI ​​model's ability to process nonlinear and complex data.

[0064] By analyzing the errors in the battery health verification model and comparing the errors between the predicted results and the initial health status, and taking a weighted average based on the errors, the problem of unreliable diagnostic results when the errors are high is solved. By dynamically adjusting the weights, the powerful adaptability of AI prediction and the robust prior information provided by the physical model of the health status are effectively integrated, so that a final health status value that has been verified by multiple times, is more accurate and reliable, can be output in various situations.

[0065] Example 2

[0066] This embodiment takes as an example a batch of mixed new and old ternary lithium battery packs operated by a logistics company in a cold region under complex working conditions such as severe ambient temperature fluctuations. The specific implementation method can be:

[0067] The battery model, temperature, and other data are acquired in real time using the closing and disconnecting instructions of the battery swap station contactor as trigger conditions. The reinforcement learning model dynamically adjusts the capture time from 0.6 seconds at room temperature to 1.5 seconds based on the current low-temperature environment to ensure that the slow electrochemical polarization response in the low-temperature environment is captured. The transient signals at the moment of connection and disconnection are captured. In one capture, due to slight contamination on the connection port between the battery pack and the battery swap station, there was a significant asymmetry between the voltage jump amplitude at the moment of connection (1.8V) and the moment of disconnection (-1.5V). Based on the signal symmetry assessment mechanism, the data quality is immediately judged to be poor, and the capture is retriggered, ultimately obtaining a set of high-quality data with good symmetry.

[0068] Subsequently, the acquired high-quality transient signal is input into the battery health model. The AI model generates a nonlinear internal resistance spectrum in real time. Since the battery pack health state is 90%, the internal resistance spectrum shows that the high-frequency semicircle diameter is slightly larger than that of the new battery. The health state physical model adjusts and calibrates the internal parameters according to the high-frequency semicircle diameter and low-frequency slope in the internal resistance spectrum generated by the AI model. For example, according to the increased high-frequency semicircle diameter, the health state physical model adjusts and calibrates the ohmic resistance to 0.85 mΩ and the charge transfer resistance to 1.75 mΩ.

[0069] Subsequently, the acquired high-quality transient signal is input into the battery health model. The AI model generates a nonlinear internal resistance spectrum in real time. Since the battery pack health state is 90%, the internal resistance spectrum shows that the high-frequency semicircle diameter is slightly larger than that of the new battery. The health state physical model adjusts and calibrates the internal parameters according to the high-frequency semicircle diameter and low-frequency slope in the internal resistance spectrum generated by the AI model. For example, according to the increased high-frequency semicircle diameter, the health state physical model adjusts and calibrates the ohmic resistance to 0.85 mΩ and the charge transfer resistance to 1.75 mΩ.

[0070] The embodiment shows, through specific working conditions and data, how the method of the application guarantees the accuracy, robustness and adaptability of the diagnosis result in the face of complex scenes such as mixed use of new and old batteries and low temperature environment through multiple intelligent mechanisms.

[0071] Embodiment Three

[0072] In the embodiment, a certain battery swap station captures a high-quality high-frequency transient signal in a one-time battery swap process. Since the battery has a slight aging problem, its internal impedance characteristics have changed. The diagnosis process of the physical-AI fusion model for the battery is as follows:

[0073] The captured high-quality transient signal is input into the trained battery health model together with the current battery operating parameters. The convolutional neural network AI model in the battery health model generates a complete nonlinear internal resistance spectrum in real time according to the input data. The internal resistance spectrum reflects the impedance characteristics of the battery at different frequencies, and the high-frequency semicircle diameter and low-frequency slope change slightly due to aging. The health state physical model of the equivalent circuit model adjusts and calibrates its internal parameters in real time according to the high-frequency semicircle diameter and low-frequency slope in the internal resistance spectrum generated by the AI model. For example, according to the increased high-frequency semicircle diameter, the health state physical model adjusts and calibrates the ohmic resistance to 0.85 mΩ and the charge transfer resistance to 1.75 mΩ.

[0074] The health state physical model decomposes the total voltage change of the battery into three parts caused by ohmic impedance, electrochemical polarization and OCV using adjusted parameters. By separating the voltage drop caused by ohmic impedance and polarization from the total voltage, accurate compensation of open circuit voltage is achieved. For example, the total voltage drop is 0.25V, of which the ohmic voltage drop is 0.085V and the polarization voltage drop is 0.165V. By subtracting these two parts of voltage drop, the model obtains the true open circuit voltage. According to the adjusted internal resistance parameter and the compensated open circuit voltage, the initial health state of the battery pack is finally output as 90.2%.

[0075] This embodiment shows, through detailed data and processes, how the battery health model can utilize the synergistic advantages of physics and AI to achieve accurate calibration of battery electrochemical parameters and accurate compensation of open circuit voltage even in the case of slight battery aging, thereby obtaining an accurate initial health state value.

[0076] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A real-time battery health diagnosis method during battery replacement based on voltage-internal-resistance correlation analysis, characterized in that: include: Acquire high-frequency transient signals during the battery swapping process, and perform real-time evaluation and error correction on the high-frequency transient signals based on the signal symmetry evaluation mechanism; Construct a battery health diagnostic model, including an AI model and a health status physical model. The AI ​​model analyzes the nonlinear relationship between transient signals and battery parameters to generate a nonlinear internal resistance spectrum. The internal parameters of the health status physical model are adjusted based on the nonlinear internal resistance spectrum. The health status physical model constrains the AI ​​model based on physical losses and compensates for the open-circuit voltage. The internal parameters and the compensated open-circuit voltage are analyzed through the voltage-resistance correlation to output the initial health status. The battery health verification model is trained through meta-learning, using the nonlinear internal resistance spectrum and the initial health state as input. The health state prediction value and fusion error are obtained through the error intelligent fusion method. The health state prediction value and the initial health state are weighted averaged according to the fusion error to determine the final health state.

2. The method for real-time battery health diagnosis during battery replacement based on voltage-internal-resistance correlation analysis according to claim 1 is characterized in that: The specific implementation process of obtaining the high-frequency transient signal of the battery during the battery replacement process includes: When the duration selected by the transient signal capture model captures a complete transient signal, a positive reward is obtained; if the capture duration is too long, resulting in data redundancy, and if the capture duration is too short, resulting in an incomplete signal, a negative reward is obtained. Through continuous interaction with the environment and trial and error, the transient signal capture model eventually obtains a capture duration selection method that can maximize rewards under various conditions; battery model, real-time temperature, and connection status data are collected, and the data is input into the deep learning transient signal capture model in the form of a multidimensional time series. The transient signal capture model determines the optimal duration for capturing the battery's high-frequency transient signal based on the input data and adjusts the capture duration.

3. The method for real-time battery health diagnosis during battery replacement based on voltage-internal-resistance correlation analysis according to claim 1 is characterized in that: Real-time evaluation and error correction of high-frequency transient signals are performed based on the signal symmetry evaluation mechanism. The specific implementation process of capturing high-frequency transient signals includes: The contactor closing and disconnecting control instructions of the battery swap station are used as trigger conditions to identify the initial contact and separation moments of the battery pack connecting and disconnecting with the external electrical system, and capture high-frequency voltage and current transient signals according to the optimal capture time; at the same time, the transient signal capture model analyzes the voltage jump amplitude and attenuation time constant characteristics of the two transient processes based on the transient data at the connection and disconnection moments. When these parameters are significantly asymmetric at the connection and disconnection moments, it indicates that there is an abnormality in the data acquisition process, and the transient signal capture is immediately re-triggered.

4. The method for real-time battery health diagnosis during battery replacement based on voltage-internal-resistance correlation analysis according to claim 1 is characterized in that: The specific implementation process of designing a battery health model includes: The battery health model consists of an AI model and a health status physical model. The AI ​​model processes time series data and learns the complex nonlinear relationship between transient signals and battery parameters; the health status physical model is based on a second-order RC equivalent circuit.

5. The method for real-time battery health diagnosis during battery replacement based on voltage-internal-resistance correlation analysis according to claim 1 is characterized in that: The AI ​​model analyzes the nonlinear relationship between transient signals and battery parameters to generate a nonlinear internal resistance spectrum, and then constrains it using a health state physical model. The specific implementation process includes: Perform battery replacement transient tests on batteries in different health states, charge states, and temperatures, collect high-frequency voltage and current data, and use an electrochemical workstation to measure the internal resistance spectrum of the battery under the same operating conditions as the "true value" label; obtain the true parameter values ​​of the health state physical model by fitting the internal resistance spectrum data; construct a composite loss function consisting of data loss and physical loss, where data loss measures the difference between the internal resistance spectrum predicted by the AI ​​model and the true value, and physical loss measures whether the parameters predicted by the AI ​​model conform to the behavior of the health state physical model; use transient signals and health state, charge state, and temperature data as input to train the AI ​​model.

6. The method for real-time battery health diagnosis during battery replacement based on voltage-internal-resistance correlation analysis according to claim 1 is characterized in that: The health state physical model adjusts internal parameters according to the nonlinear internal resistance spectrum, compensates the open circuit voltage, and outputs the initial health state through voltage-resistance correlation analysis of the internal parameters and the compensated open circuit voltage. The specific implementation process includes: The AI ​​model generates a complete nonlinear internal resistance spectrum in real time based on the input data, and inputs the generated nonlinear internal resistance spectrum into the health status physical model. The health status physical model adjusts and calibrates the internal ohmic resistance, charge transfer resistance and double layer capacitance parameters in real time based on the high-frequency semicircle diameter and low-frequency slope characteristics in the internal resistance spectrum; the health status physical model uses the adjusted parameters to decompose the total voltage change into three parts caused by ohmic impedance, electrochemical polarization and open circuit voltage; by separating the voltage drop caused by ohmic impedance and polarization from the total voltage, the voltage-internal resistance correlation analysis is performed on the adjusted internal resistance parameters and the compensated open circuit voltage to output the initial health status of the battery pack.

7. The method for real-time battery health diagnosis during battery replacement based on voltage-internal-resistance correlation analysis according to claim 1 is characterized in that: Meta-learning is used to train a battery health verification model. The nonlinear internal resistance spectrum and the initial battery health state are used as inputs. The initial battery health state provides prior knowledge as a guidance signal to obtain a health state prediction value. The specific implementation process includes: The meta-learning MAML algorithm is used to perform alternating training on multiple different tasks. The Bayesian neural network-based model learns the common laws in different battery aging processes, taking the nonlinear internal resistance spectrum and initial health state as input; the initial health state provides a priori information to help the prediction model quickly focus on the historical battery data closest to it. Finally, the battery health verification model learns a priori knowledge of the weight distribution that can adapt to the new task; when the model faces a new task, the battery health verification model uses the learned prior knowledge of the weight distribution to fine-tune through data to obtain a weight distribution that adapts to the new task. During prediction, the battery health verification model obtains prediction results through Monte Carlo sampling. The standard deviation of these results is used as a quantitative indicator of model error, and the average value of the prediction results is output as the health state prediction value.

8. The method for real-time battery health diagnosis during battery replacement based on voltage-internal-resistance correlation analysis according to claim 1 is characterized in that: The error intelligent fusion method is used to obtain the fusion result of the battery health verification model error and the health state prediction value error. Based on the fusion result, the health state prediction value and the initial battery health state are weighted averaged to determine the final battery health state. The specific implementation process includes: The prediction result of the battery health verification model is compared with the initial health state, and the absolute value of the difference between the health state prediction value and the initial health state is used as the data error. The battery health verification model error and the data error are weighted averaged to obtain the final error; the health state prediction value and the initial health state are weighted averaged according to the error size. When the error is less than the preset threshold, it indicates that the prediction result is highly credible, and the weight of the health state prediction value is increased; when the error is greater than the preset threshold, it indicates that there is an abnormality in the data, and the weight of the initial health state is increased. The final health state value is calculated through the weighted average.

Images