A lithium battery health degree online detection and prevention method based on impedance and digital twinning
By superimposing a 1Hz square wave excitation signal during lithium battery charging and combining digital twin technology and LSTM-RNN model, online detection and control of lithium battery health are achieved, solving the problems of real-time performance and accuracy in lithium deposition detection. It is applicable to different chemical systems and charging strategies and has wide applicability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- KUNMING UNIV OF SCI & TECH
- Filing Date
- 2026-03-11
- Publication Date
- 2026-05-29
AI Technical Summary
Existing methods for detecting lithium deposition during fast charging of lithium batteries cannot achieve online real-time monitoring and intervention, leading to safety hazards. Furthermore, traditional impedance methods suffer from time delays and misjudgments.
By superimposing a 1Hz square wave excitation signal to monitor the lithium battery impedance, and combining digital twin technology and a long short-term memory recurrent neural network model, lithium deposition anomalies can be identified in real time, and control commands can be generated to adjust the charging strategy.
It enables non-destructive online monitoring of lithium batteries during charging, with rapid response and accurate identification of the lithium deposition initiation point. It has wide applicability and high precision, and can proactively intervene before lithium deposition risks occur.
Smart Images
Figure CN122109883A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of lithium-ion battery safety management technology, specifically to an online detection and control method for lithium battery health based on impedance and digital twins. Background Technology
[0002] With the widespread adoption of electric vehicles, fast charging technology has become crucial. However, fast charging under high current or low temperature conditions can easily trigger lithium deposition on the negative electrode, leading to a sharp decrease in battery capacity, increased internal resistance, and even safety hazards such as internal short circuits and thermal runaway. Currently, methods for detecting lithium deposition mainly include: offline detection methods, such as nuclear magnetic resonance (NMR) and neutron powder diffraction (NPD). These methods require specialized large-scale equipment and can only be performed in laboratory environments, making them unsuitable for vehicle battery management systems (BMS). Model-dependent methods simulate and predict lithium deposition behavior by establishing electrochemical models. However, the accuracy, robustness, and computational load of these models are affected by numerous factors, making it difficult to meet the real-time and reliability requirements of vehicle BMS. Relaxation voltage analysis determines whether lithium deposition has occurred by analyzing the presence of a plateau in the voltage relaxation curve after charging. This method has good accuracy, but its biggest drawback is the inability to perform online detection and real-time intervention during charging, making it a "post-event detection." Existing impedance methods have shown a correlation between battery impedance and lithium deposition. For example, lithium deposition can be identified by monitoring changes in the full-spectrum electrochemical impedance spectroscopy. However, acquiring the full-spectrum EIS is time-consuming and computationally intensive, resulting in significant time delays in detection. Furthermore, noise in real-time measurements can easily cause localized abnormal changes in the impedance curve, leading to misjudgments. Therefore, further research is needed. Summary of the Invention
[0003] The purpose of this application is to provide an online detection and control method for the health of lithium batteries based on impedance and digital twins. The specific technical solution is as follows:
[0004] A method for online detection and control of lithium battery health based on impedance and digital twin includes: S1, superimposing an excitation signal of a specific frequency onto the DC charging current, simultaneously acquiring the voltage response signals at both ends of the battery, and calculating the real part of the battery impedance at the specific frequency using a fast Fourier transform based on the excitation current and response voltage signals; S2, constructing a virtual battery model in digital space including at least one equivalent circuit, and optimizing the virtual battery model based on real-time operating data of the physical battery, with the virtual battery model predicting the operating state of the physical battery; S3, identifying the abnormal decline inflection point caused by lithium deposition based on the real part of the impedance at the specific frequency in S1, the real-time operating state data of the physical battery in S2, and / or the operating state data predicted by the virtual battery model, and outputting a lithium deposition initiation signal; S4, generating control commands based on the lithium deposition initiation signal output in S3 and sending them to the battery management system, adjusting the charging strategy according to the control commands.
[0005] The excitation signal of a specific frequency in S1 is a 1Hz square wave, and the amplitude of the excitation signal is controlled within 10% of the charging current.
[0006] When optimizing the battery virtual model in S2, the voltage, current, and temperature data of the physical battery during real-time operation are acquired. The ohmic internal resistance and polarization resistance parameters of the battery virtual model are updated online or periodically using algorithms such as the least squares method, so that the output of the virtual model is synchronized with the actual state of the physical battery. The battery virtual model and the physical battery maintain a high degree of consistency. As the amount of accumulated monitoring data increases, the training amount of the battery virtual model also increases, resulting in a stronger simulation of the battery virtual model and a better reflection of the changing trend of the electrochemical performance of the physical battery.
[0007] In S3, when identifying the abnormal decline inflection point caused by lithium deposition, the real part of the impedance Z'1Hz at a specific frequency in S1 and the SOC, temperature, and charge rate features predicted by the real-time operation of the physical battery and / or the battery virtual model in S2 are used as inputs. A pre-trained long short-term memory recurrent neural network model analyzes the temporal characteristics of the real part of the impedance curve at a specific frequency in S1 to identify the abnormal decline inflection point caused by lithium deposition and outputs the lithium deposition initiation signal. The long short-term memory recurrent neural network model includes an input layer that receives a 5-dimensional feature vector as input, specifically the slope (Kz) of Z'1Hz, the first derivative of Kz (dKz), temperature, and charge rate. The structure includes: rate and SOC; LSTM hidden layer, consisting of one or more LSTM layers, with LSTM units containing forget gates, input gates, and output gates, responsible for processing time-series data and extracting deep features of impedance curves changing over time; Dropout layer, added after the LSTM hidden layer, randomly discarding a portion of neurons during training to prevent overfitting and improve generalization ability; Fully connected layer, which non-linearly combines the features extracted by the LSTM layer to further refine the information; Output layer, using neurons with the Sigmoid activation function, outputting a probability value between 0 and 1, which is determined as the starting point of lithium deposition when the value exceeds a preset threshold.
[0008] The control commands in S4 include gradually reducing the charging current, switching charging stages, or stopping charging.
[0009] When training a long short-term memory recurrent neural network model, LSTM, through its unique gating mechanism, can learn and remember this "normal" downward trend pattern and store it in the cell state. When lithium deposition occurs, the Z'1Hz curve will show a sudden, steeper abnormal drop. LSTM compares the real-time input Z'1Hz change trend with the "normal trend" it remembers. When the prediction error of the input sequence increases significantly, or the features extracted by the hidden layer point to the abnormal pattern, the model will be activated.
[0010] When training a long short-term memory recurrent neural network model, the source of robustness is noise suppression. Ordinary noise is high-frequency, random local fluctuations. By processing the long-term dependence of time series, LSTM can "see through" these local spikes and focus on the changes in the overall trend. Once the noise subsides, the model will not give false alarms. However, the decline caused by actual lithium plating is irreversible, and the model will continue to identify it.
[0011] The training of the long short-term memory recurrent neural network model in S3 includes: S3.1, charging tests of NCM and LFP batteries at different temperatures (0-40°C) and different rates (0.25C-2C); S3.2, labeling each impedance curve with whether lithium plating has occurred and the starting point of lithium plating, using the "voltage relaxation method" as the cross-validation standard and as the "Ground Truth" of the training model; S3.3, calculating the input features: the slope (Kz) of Z'1Hz, the first derivative of Kz (dKz), temperature, rate, and SOC; S3.4, dividing the long sequence data into fixed-length sliding window samples, and dividing the dataset into training set, validation set, and test set;
[0012] S3.5 The loss function during training is binary cross-entropy, the optimizer is Adam optimizer, and the evaluation metrics are accuracy, recall, and F1 score, with a focus on detection error (<2.5% SOC) and false positive rate; S3.6 The training data is input into the LSTM network, and the network weights are updated through backpropagation algorithm and backpropagation over time (BPTT).
[0013] The beneficial effects of this application are that it achieves non-destructive, online monitoring of lithium deposition during charging without affecting normal battery use. It employs a 1Hz single-frequency impedance, resulting in low computational load and fast response. Combined with the powerful temporal pattern recognition capability of the Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) model, it can accurately capture the onset point of lithium deposition. Experiments verify that its detection error is less than 2.5% of SOC, far exceeding traditional methods. LSTM-RNN can effectively learn and remember the normal changing trend of impedance, thereby effectively distinguishing between genuine lithium deposition decreases and invalid local fluctuations caused by noise, solving the problem of misjudgment. Through real-time simulation and prediction using digital twins, it can proactively intervene before the risk of lithium deposition occurs, realizing closed-loop intelligent management from "detection" to "prediction-prevention". It is applicable to different chemical systems (such as NCM, LFP), different charging strategies (CC, MCC), and different temperature conditions, possessing broad applicability. Attached Figure Description
[0014] Figure 1 This is a schematic diagram of the overall architecture in this application;
[0015] Figure 2 This is a schematic diagram of the line impedance measurement principle in this application;
[0016] Figure 3 This is a schematic diagram illustrating the abnormal decrease in the Z'1Hz curve when lithium deposition occurs in this application;
[0017] Figure 4 This is a diagram of the LSTM-RNN network structure in this application. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to specific embodiments and accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of this application. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.
[0019] like Figure 1 As shown, an online detection and control method for lithium battery health based on impedance and digital twins includes:
[0020] like Figure 2 As shown, in step S1, a specific frequency excitation signal is superimposed on the DC charging current, and the voltage response signals at both ends of the battery are simultaneously acquired. Based on the excitation current and the response voltage signals, the real part of the battery impedance Z'1Hz at that specific frequency is calculated using a fast Fourier transform. The specific frequency excitation signal is a 1Hz square wave, and the amplitude of the excitation signal is controlled within 10% of the charging current.
[0021] S2. Construct a virtual battery model in digital space, including at least one equivalent circuit, and optimize the virtual model based on real-time operating data of the physical battery. The virtual model predicts the operating state of the physical battery. During optimization, acquire real-time voltage, current, and temperature data of the physical battery. Update the ohmic internal resistance and polarization resistance parameters of the virtual model online or periodically using algorithms such as least squares, ensuring that the output of the virtual model is synchronized with the actual state of the physical battery. The virtual model and the physical battery maintain a high degree of consistency. As the accumulated amount of monitoring data increases, the training amount of the virtual model also increases, resulting in stronger simulation of the virtual model and a better reflection of the changing trends of the electrochemical performance of the physical battery.
[0022] like Figure 3 As shown, S3, based on the real part of the impedance at a specific frequency in S1 and the real-time operating state data of the physical battery and / or the operating state data predicted by the battery virtual model in S2, identifies the abnormal decline inflection point caused by lithium deposition and outputs the lithium deposition initiation signal. When identifying the abnormal decline inflection point caused by lithium deposition, the real part of the impedance Z'1Hz at a specific frequency in S1 and the SOC, temperature, and charge rate characteristics predicted by the real-time operation of the physical battery and / or the battery virtual model in S2 are used as inputs. A pre-trained long short-term memory recurrent neural network model analyzes the temporal characteristics of the real part of the impedance curve at a specific frequency in S1 to identify the abnormal decline inflection point caused by lithium deposition and outputs the lithium deposition initiation signal; as shown... Figure 4As shown, the Long Short-Term Memory (LSTM) recurrent neural network model includes: an input layer that receives a 5-dimensional feature vector as input, specifically the slope (Kz) of Z'1Hz, the first derivative of Kz (dKz), temperature, charging rate, and SOC; an LSTM hidden layer composed of one or more LSTM layers, each containing a forget gate, input gate, and output gate, responsible for processing time-series data to extract deep features of the impedance curve over time; a Dropout layer added after the LSTM hidden layer to randomly discard a portion of neurons during training, preventing overfitting and improving generalization ability; a fully connected layer that non-linearly combines the features extracted by the LSTM layers to further refine the information; and an output layer using neurons with the Sigmoid activation function, outputting a probability value between 0 and 1, which is determined as the starting point of lithium deposition when the value exceeds a preset threshold.
[0023] When training a Long Short-Term Memory (LSTM) recurrent neural network model, LSTM, through its unique gating mechanism, can learn and remember this "normal" downward trend pattern and store it in the cell state. When lithium deposition occurs, the Z'1Hz curve exhibits a sudden, steeper, aberrant drop. LSTM compares the real-time input Z'1Hz trend with its memorized "normal trend." When the prediction error of the input sequence significantly increases, or the features extracted by the hidden layer point to the aberrant pattern, the model is activated. The robustness comes from noise suppression. Ordinary noise consists of high-frequency, random local fluctuations. By processing the long-term dependencies of the time series, LSTM can "see through" these local spikes and focus on the overall trend change. Once the noise subsides, the model will not give false alarms, but the drop caused by actual lithium deposition is irreversible, and the model will continue to identify it. Specifically, this includes:
[0024] S3.1. Charge tests were conducted on NCM and LFP batteries at different temperatures (0-40°C) and different rates (0.25C-2C).
[0025] S3.2 Label each impedance curve (whether lithium deposition occurred and the starting point of lithium deposition). Use the "voltage relaxation method" as the standard for cross-validation, that is, use the characteristics of the relaxation voltage plateau to accurately calibrate the actual occurrence time of lithium deposition, as the "Ground Truth" for training the model.
[0026] S3.3 Calculate the input features: slope (Kz) of Z'1Hz, first derivative of Kz (dKz), temperature, magnification, and SOC.
[0027] S3.4. Divide the long sequence data into fixed-length sliding window samples. Divide the dataset into training set, validation set, and test set.
[0028] S3.5 Loss Function: Binary cross-entropy is typically used because this is a binary classification problem. Optimizer: The Adam optimizer is commonly used, adaptively adjusting the learning rate. Evaluation Metrics: Accuracy, Recall, F1 score, with a focus on detection error (<2.5% SOC) and false positive rate.
[0029] S3.6 Input the training data into the LSTM network and update the network weights through the backpropagation algorithm and backpropagation over time (BPTT).
[0030] S4. Based on the lithium deposition initiation signal output in S3, a control command is generated and sent to the battery management system. The charging strategy is adjusted according to the control command. The control command includes gradually reducing the charging current, switching charging stages, or stopping charging.
[0031] When training a long short-term memory recurrent neural network model, LSTM, through its unique gating mechanism, can learn and remember this "normal" downward trend pattern and store it in the cell state. When lithium deposition occurs, the Z'1Hz curve will show a sudden, steeper abnormal drop. LSTM compares the real-time input Z'1Hz change trend with the "normal trend" it remembers. When the prediction error of the input sequence increases significantly, or the features extracted by the hidden layer point to the abnormal pattern, the model will be activated.
[0032] When training a long short-term memory recurrent neural network model, the source of robustness is noise suppression. Ordinary noise is high-frequency, random local fluctuations. By processing the long-term dependence of time series, LSTM can "see through" these local spikes and focus on the changes in the overall trend. Once the noise subsides, the model will not give false alarms. However, the decline caused by actual lithium plating is irreversible, and the model will continue to identify it.
Claims
1. A method for online detection and control of lithium battery health based on impedance and digital twin, characterized in that, include: S1. A specific frequency excitation signal is superimposed on the DC charging current of the current, and the voltage response signal at both ends of the battery is collected synchronously. Based on the excitation current and the response voltage signal, the real part of the battery impedance at this specific frequency is calculated by fast Fourier transform. S2. Construct a battery virtual model in digital space that includes at least one equivalent circuit, and optimize the battery virtual model based on the real-time operating data of the physical battery. The battery virtual model predicts the operating state of the physical battery. S3. Based on the real part of the impedance at a specific frequency in S1 and the real-time operating status data of the physical battery and / or the operating status data predicted by the battery virtual model in S2, identify the abnormal decline inflection point caused by lithium deposition and output the lithium deposition initiation signal. S4. Based on the lithium deposition start signal output in S3, generate a control command and send it to the battery management system, and adjust the charging strategy according to the control command.
2. The method for online detection and control of lithium battery health based on impedance and digital twin as described in claim 1, characterized in that, The excitation signal of a specific frequency in S1 is a 1Hz square wave, and the amplitude of the excitation signal is controlled within 10% of the charging current.
3. The method for online detection and control of lithium battery health based on impedance and digital twin as described in claim 2, characterized in that, In step S2, when optimizing the battery virtual model, the voltage, current, and temperature data of the physical battery during real-time operation are acquired. The ohmic internal resistance and polarization resistance parameters of the battery virtual model are updated online or periodically using algorithms such as the least squares method, so that the output of the virtual model is synchronized with the actual state of the physical battery. The battery virtual model and the physical battery maintain a high degree of consistency. As the amount of accumulated monitoring data increases, the training amount of the battery virtual model also increases, resulting in a stronger simulation of the battery virtual model and a better reflection of the changing trend of the electrochemical performance of the physical battery.
4. The method for online detection and control of lithium battery health based on impedance and digital twin as described in claim 3, characterized in that, In step S3, when identifying the abnormal decline inflection point caused by lithium deposition, the real part of the impedance Z'1Hz at a specific frequency in S1 and the SOC, temperature, and charge rate characteristics predicted by the real-time operation of the physical battery and / or the battery virtual model in S2 are used as inputs. A pre-trained long short-term memory recurrent neural network model is used to analyze the temporal characteristics of the real part of the impedance curve at a specific frequency in S1, identify the abnormal decline inflection point caused by lithium deposition, and output the lithium deposition initiation signal. The long short-term memory recurrent neural network model includes: The input layer receives a 5-dimensional feature vector as input, specifically the slope (Kz) of Z'1Hz, the first derivative of Kz (dKz), temperature, charging rate, and SOC; An LSTM hidden layer consists of one or more LSTM layers. An LSTM unit contains a forget gate, an input gate, and an output gate. It is responsible for processing time-series data and is used to extract deep features of impedance curves as they change over time. Dropout layer: After the LSTM hidden layer, a Dropout layer is added to randomly drop a portion of neurons during training to prevent the model from overfitting and improve generalization ability; The fully connected layer nonlinearly combines the features extracted by the LSTM layer to further refine the information; The output layer uses neurons that activate the sigmoid function to output a probability value between 0 and 1. When this value exceeds a preset threshold, it is determined to be the starting point of lithium deposition.
5. The method for online detection and control of lithium battery health based on impedance and digital twin as described in claim 4, characterized in that, The control commands in S4 include gradually reducing the charging current, switching charging stages, or stopping charging.
6. In the online detection and control method for lithium battery health based on impedance and digital twin as described in claim 4, during the training of the long short-term memory recurrent neural network model in S3, LSTM, through its unique gating mechanism, can learn and remember this "normal" downward trend pattern and store it in the cell state; when lithium deposition occurs, the Z'1Hz curve will show a sudden, steeper abnormal drop. LSTM compares the real-time input Z'1Hz change trend with the "normal trend" it remembers. When the prediction error of the input sequence increases significantly, or the features extracted by the hidden layer point to the abnormal pattern, the model will be activated.
7. In the online detection and control method for lithium battery health based on impedance and digital twin as described in claim 6, the robustness of the long short-term memory recurrent neural network model training in S3 comes from noise suppression. Ordinary noise is high-frequency, random local fluctuation. LSTM can "see through" these local spikes by processing the long-term dependence of time series and focus on the changes in the overall trend. Once the noise subsides, the model will not give false alarms, but the decline caused by actual lithium plating is irreversible, and the model will continue to identify it.
8. The online detection and control method for lithium battery health based on impedance and digital twin as described in claim 7, wherein the training of the long short-term memory recurrent neural network model in S3 includes: S3.
1. Charge tests were conducted on NCM and LFP batteries at different temperatures (0-40°C) and different rates (0.25C-2C). S3.2 Label each impedance curve with whether lithium plating has occurred and the starting point of lithium plating. Use the "voltage relaxation method" as the standard for cross-validation and as the "Ground Truth" for training the model. S3.3 Calculate the input features: the slope (Kz) of Z'1Hz, the first derivative of Kz (dKz), temperature, magnification, and SOC; S3.
4. Divide the long sequence data into fixed-length sliding window samples, and divide the dataset into training set, validation set and test set; S3.5 The loss function during training is binary cross-entropy, the optimizer is Adam optimizer, and the evaluation metrics are accuracy, recall, and F1 score. The focus is on detection error (<2.5% SOC) and false positive rate. S3.6 Input the training data into the LSTM network and update the network weights through the backpropagation algorithm and backpropagation over time (BPTT).