High-sensitivity detection method for safety of battery cell

Abstract

The invention discloses a high-sensitivity detection method for the safety of a battery cell, and the method comprises the following steps: S1, collecting the data of the battery cell in real time through a tunnel magnetoresistive sensor, and outputting a preliminary detection data set; s2, generating a micro-short circuit risk identification signal based on the preliminary detection data set; s3, generating a temperature anomaly distribution map by using the preliminary detection data set; s4, generating a thermal runaway early warning signal based on the temperature anomaly distribution map; s5, performing magnetic field scanning on the pole piece to generate a magnetic distortion data atlas; s6, constructing battery cell multi-source detection data, inputting the data into the physical information neural network, and training to obtain a final battery cell aging evolution model; s7, outputting a charging and discharging strategy optimization instruction based on the final cell aging evolution model; and S8, uniformly fusing the micro short circuit risk identification signal, the thermal runaway early warning signal and the charging and discharging strategy optimization instruction to generate a comprehensive response control instruction. According to the method, magnetic resistance sensing and physical modeling are fused, and intelligent prediction and response control optimization of the risk of the battery cell are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of battery safety monitoring and intelligent control, and particularly to a highly sensitive detection method for the safety of battery cells. Background Art

[0002] With the rapid development of industries such as new energy vehicles, renewable energy, and energy storage power stations, lithium-ion battery cells, as their core power or storage units, their safety issues are becoming the focus of global technical fields. Especially in application scenarios such as large capacity, high rate, and long cycle, complex failure modes may be triggered inside the battery cells due to factors such as material inhomogeneity, manufacturing defects, or operational degradation, such as micro-short circuits, thermal runaway, capacity degradation, abnormal internal resistance, etc., which are extremely likely to lead to safety accidents. According to statistics, in recent years, many energy storage power station and electric vehicle fire accidents worldwide can be traced back to the latent risks inside the battery cells not being detected and responded to in a timely manner. Therefore, developing a battery cell detection method with high sensitivity, strong real-time performance, and precise positioning ability has become the core topic for improving the intrinsic safety level of battery systems.

[0003] Currently, relatively common battery cell detection means in the industry include multi-point acquisition of voltage, current, and temperature, as well as data fusion and anomaly judgment relying on the battery management system (BMS). Current monitoring modules represented by Hall current sensors are widely used in existing systems. Although they have a certain measurement range and stability, they have obvious shortcomings in distinguishing micro-short circuit currents at the microampere level, and the accuracy is usually not more than 1%. In addition, contact temperature monitoring means based on thermocouples or thermistors can achieve point temperature tracking, but there are limitations in the early perception of planar heat diffusion and non-contact deployment, especially in high power density environments, where they are easily affected by factors such as noise, electromagnetic interference, and poor physical contact.

[0004] Furthermore, in actual working conditions, the micro-short circuit phenomenon often shows instantaneous current fluctuations below milliamperes, which are both lower than traditional protection thresholds and easily confused with natural self-discharge phenomena, causing difficulties in detection and discrimination. And thermal runaway, as the direct cause of battery cell safety accidents, usually has non-linear and abrupt characteristics in its development process, and the conventional single-point temperature alarm mechanism based on thresholds is difficult to respond in a timely manner. In addition, most current battery cell health state modeling uses data-driven methods, such as neural networks, decision trees, support vector machines, etc. Although they can achieve prediction and classification of some states, they are still insufficient in multi-source heterogeneous data fusion, physically interpretable modeling, and cross-cycle generalization ability.

[0005] During the manufacturing process, local permeability anomalies may occur in the electrode sheets of the battery cells due to factors such as uneven coating, residual metal impurities, and deposition of active substances. Although these defects are difficult to identify through visual inspection or static electrical testing methods, they are potential sources of micro-shorts in the later stage. Existing detection methods mainly rely on devices such as X-rays, ultrasonic waves, or high-frequency conductance analysis to complete structural or compositional screening. However, such equipment is costly, has strict environmental requirements, and high integration complexity, making it unfavorable for deployment and operation on actual production lines.

[0006] To enhance the multi-dimensional risk perception ability of battery cells, some studies have attempted to introduce a multi-source data fusion mechanism and construct a time series anomaly detection model by combining information such as voltage, current, and temperature. In recent years, there have also been studies using fluxgate sensors, GMR sensors, etc. to achieve preliminary monitoring of the magnetic field distribution of batteries. However, these methods are still difficult to meet the precise capture requirements for early micro-short signals and abnormal thermal distributions in terms of key indicators such as sensitivity, response speed, and power consumption control. In particular, the lack of a multi-scale joint perception and analysis framework that integrates battery cell manufacturing defect data, physical modeling information, and real-time sensing data during the operation stage makes it difficult for the current detection system to simultaneously achieve the system goals of "early warning, physical interpretation, and control linkage".

[0007] In addition, although the current battery management system already has a certain degree of intelligence and response mechanism, most of them still remain at the level of static logical judgment or simple rule matching. Facing complex and changing operating conditions and the gradually evolving aging process, how to construct an intelligent modeling method with physical constraints and dynamic adaptation capabilities is still one of the technical bottlenecks in the industry. Especially in scenarios where it is necessary to simultaneously integrate real-time monitoring data and historical state trajectories, and consider both short-term fluctuations and long-term evolution laws, traditional black-box deep learning models lack physical consistency constraints, and have problems such as model drift, poor interpretability, and low credibility, making it difficult to meet the application requirements of safety-critical systems.

[0008] In summary, there are still several significant deficiencies in the existing technologies for battery cell safety detection. First, the micro-short circuit current and local overheating phenomena are difficult to be accurately identified by existing low-sensitivity sensors, resulting in the absence of early warning. Second, the temperature anomaly monitoring lacks non-contact and distributed spatial perception capabilities and cannot accurately restore the heat diffusion path. Third, the defects in the battery cell manufacturing stage and the safety status in the operation stage have not been effectively connected, resulting in the inability to identify potential hidden dangers comprehensively. Fourth, the existing data modeling methods lack physical constraints when processing multi-source heterogeneous signals, affecting the prediction accuracy and result interpretability. Fifth, the control response strategy has not formed a feedback closed-loop highly coupled with the detection results, and cannot achieve the generation of strategies and command output dynamically adjusted according to the risk level.

[0009] Therefore, how to provide a highly sensitive detection method for the safety of battery cells is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0010] An object of the present invention is to propose a highly sensitive detection method for the safety of battery cells. The present invention integrates tunneling magnetoresistance sensing technology, non-contact thermal field perception mechanism, physical information neural network modeling method and reinforcement learning strategy optimization algorithm, and systematically constructs a closed-loop detection system from multi-source risk perception of battery cells to intelligent control response, and details the key steps of micro-short circuit identification, thermal runaway warning, manufacturing defect quantification and aging state prediction, with the advantages of fast response speed, high detection accuracy, strong interpretability and excellent adaptive control ability.

[0011] A highly sensitive detection method for the safety of battery cells according to an embodiment of the present invention includes the following steps:

[0012] S1. Embed a tunneling magnetoresistance sensor module inside the top cover or outside the housing of the battery cell, and collect the current data and magnetic field data of the battery cell during operation in real time, and output a preliminary detection data set;

[0013] S2. Based on the preliminary detection data set, analyze the microampere to milliampere level fluctuations of the battery cell, and generate a micro-short circuit risk identification signal;

[0014] S3. Use the three-axis component data of the external magnetic field of the battery cell in the preliminary detection data set to identify the spatial magnetic anomaly caused by the change of material magnetic permeability, and generate a temperature anomaly distribution map;

[0015] S4. Based on the temperature anomaly distribution map, judge whether there is a potential thermal runaway trend, and generate a thermal runaway warning signal;

[0016] S5. During the manufacturing stage of the battery cell, apply a control excitation magnetic field to scan the pole piece with the magnetic field, and generate a magnetic distortion data map;

[0017] S6. Combine the preliminary detection data set, micro-short circuit risk identification signal, temperature anomaly distribution map and magnetic distortion data map to construct multi-source detection data of the battery cell, input it into the physical information neural network, construct an input layer, a main prediction branch and a physical residual branch, and train to obtain the final battery cell aging evolution model;

[0018] S7. Based on the final battery cell aging evolution model, construct a charge and discharge strategy optimizer, and output a charge and discharge strategy optimization instruction;

[0019] S8. Uniformly fuse the micro-short circuit risk identification signal, thermal runaway warning signal and charge and discharge strategy optimization instruction to generate a comprehensive response control instruction.

[0020] Optionally, the preliminary detection data set includes battery cell working current data, battery cell terminal voltage data, three-axis component data of the external magnetic field of the battery cell, estimated battery cell surface temperature data, and corresponding timestamp sequences.

[0021] Optionally, S2 specifically includes:

[0022] S21. Extract the cell working current data from the preliminary detection dataset to form a time series;

[0023] S22. Perform a sliding split on the current time series with a fixed time window length τ1 to construct a set of current windows;

[0024] S23. Calculate the current mean and standard deviation for each window respectively. If the current mean of the kth current window is less than the lower limit of the preset micro-short circuit current threshold and the current standard deviation is less than the preset current fluctuation threshold, then mark it as a micro-short circuit suspicious window;

[0025] S24. If there are no less than N consecutive micro-short circuit suspicious windows, then generate a micro-short circuit risk identification signal R short (t) = 1 at the starting time point t, otherwise set it as R short (t) = 0.

[0026] Optionally, S3 specifically includes:

[0027] S31. Extract the three-axis component data of the external magnetic field of the cell from the preliminary detection dataset, and calculate the magnetic anomaly response index at time t based on the change characteristics of the magnetic field in the spatial distribution:

[0028]

[0029] where Φ(t) is the magnetic anomaly response index, Ω is the three-dimensional spatial region covered by the TMR sensor, t represents the time variable in the continuous sampling process, r is the spatial position vector, is the spatial gradient tensor of the magnetic field, κ(·) is the temperature modulation coefficient function, and T(t, r) is the temperature distribution on the cell surface at position r and time t;

[0030] S32. Perform a time window smoothing process on the magnetic anomaly response index to construct a sliding window feature curve;

[0031] S33. Perform a statistical process on the magnetic anomaly response index window Φ (k) to calculate the average value of the magnetic disturbance intensity within the window and the maximum value max(Φ (k) . When the average value of the magnetic disturbance intensity is greater than the preset average threshold Φ th or the maximum value is greater than the preset peak threshold Φ max , mark the magnetic anomaly response index window Φ (k) as a magnetic field anomaly window

[0032] S34. Stitch all magnetic anomaly windows in chronological order to generate a temperature anomaly distribution map T for the time period map .

[0033] Optionally, the generation of the thermal runaway warning signal specifically includes constructing a thermal risk index function and determining the thermal runaway warning signal;

[0034] The construction of the thermal risk index function:

[0035]

[0036] Among them, H(t) is the thermal risk index, λ1 is the temperature rise rate weight coefficient, T map is the temperature anomaly distribution map, is the first-order time derivative of the temperature anomaly map, λ2 is the magnetic anomaly response weight coefficient, Φ(t) is the magnetic anomaly response index, λ3 is the temperature spatial gradient, is the spatial gradient of the temperature field, t represents the time variable in the continuous sampling process, r is the spatial position vector, λ4 is the historical temperature rise integral weight coefficient, δ is the retrospective time window length, τ is the historical time variable, and w(·) is the time decay weight function;

[0037] The determination of the thermal runaway warning signal compares the thermal risk index H(t) with the preset threshold H th When H(t)>H th is satisfied, a thermal runaway warning signal is generated at the corresponding time point t, marked as R thermal (t)=1, otherwise marked as R thermal (t)=0.

[0038] Optionally, the magnetic distortion data map includes spatial position coordinates, magnetic response deviation intensity, and anomaly marking labels; the magnetic response deviation intensity represents the maximum difference between the actually measured magnetic response value and the reference magnetic response value at the corresponding position; the anomaly marking label makes an anomaly determination for each position based on whether the magnetic response deviation exceeds the threshold.

[0039] Optionally, the S6 specifically includes:

[0040] S61. Based on the preliminary detection data set and in combination with the micro-short circuit risk identification signal, the temperature anomaly distribution map, and the magnetic distortion data map, align the time indices of each item and construct the multi-source detection data of the battery cell;

[0041] S62. Input the multi-source detection data of the battery cell into the physics-informed neural network to construct a battery cell aging evolution model;

[0042] S63. Construct a joint loss function:

[0043]

[0044] Among them, is the total loss function, t is the time index variable, is the predicted value of the cell aging state, S true (t) is the actual labeled cell aging state value, λ phys is the weight coefficient of the physical residual term, is the residual term of the physical equation of cell aging evolution;

[0045] S64. Adopt a combination of forward propagation and backward propagation to jointly minimize the training of the physical information neural network parameter set θ;

[0046] S65. During continuous multi-round training iterations, if both the data error and the physical residual in the joint loss function satisfy that the error change rate is lower than the threshold, the error value is lower than the target lower limit, or the training error no longer decreases, it is determined that the model training process is completed, and the final cell aging evolution model is output.

[0047] Optionally, the specific content of S62 includes:

[0048] S621. Construct the input layer of the neural network, and set the input vector as the multi-source detection data of the cell;

[0049] S622. Based on the multi-source detection data of the cell, construct a feedforward main prediction branch neural network. The main neural network learns the mapping relationship between the input features and the cell aging state through multiple non-linear transformations, and outputs the predicted value of the cell aging state at the corresponding moment The predicted value of the cell aging state includes the estimated value of the capacity retention rate, the growth value of the internal resistance, and the offset value of the voltage platform;

[0050] S623. Construct a physical residual branch, add the physical model of the cell aging process as a constraint to the network training, and the physical residual branch fits the physical residual term of the cell degradation equation:

[0051]

[0052] Among them, is the residual term of the physical equation of cell aging evolution, t is the time index variable, C Li is the lithium ion concentration distribution function inside the cell, D is the lithium ion diffusion coefficient, is the second-order gradient of the lithium concentration in space, k is the lithium loss rate constant, γ is the defect modulation coefficient, f defect (r) is the defect perturbation function, r is the spatial position vector;

[0053] S624. Combine the outputs of the main branch and the residual branch, and respectively output the predicted value of the cell aging state and the residual term of the physical equation for the aging evolution of the battery cell

[0054] S625. Set the physical information neural network parameter set θ, including the main branch parameters and the physical residual branch parameters;

[0055] S626. The input layer, the main prediction branch, and the physical residual branch together constitute the battery cell aging evolution model.

[0056] Optionally, the S7 specifically includes:

[0057] S71. Based on the final battery cell aging evolution model, input the multi-source detection data of the battery cell and output the predicted value of the battery cell aging state;

[0058] S72. Extract the capacity retention rate estimate value, the internal resistance growth value, and the voltage platform offset value in the predicted value of the battery cell aging state to form a state mapping vector;

[0059] S73. Set the adjustable control action set A(t), including the maximum charging current, the maximum discharging current, the cooling air speed level, and the temperature adjustment power factor in the current time period;

[0060] S74. Construct the optimization objective function of the charge and discharge control strategy:

[0061]

[0062] where A(t) is the adjustable control action set, is to solve the optimal policy action, is the expected value of the policy return, t is the time index variable, R(t) is the policy return function at the current time step t, T is the total length of the policy optimization time domain, α1 is the capacity retention rate weight coefficient, ΔC ret (t) is the positive gain of the current action to the capacity retention rate, α2 is the internal resistance suppression weight coefficient, ΔR in (t) is the internal resistance growth amplitude caused by the current action, α3 is the energy consumption suppression weight coefficient, and E(t) is the unit time energy consumption of the current action;

[0063] S75. Use the method based on deep reinforcement learning to train the optimization objective function of the charge and discharge control strategy. The policy network takes the state mapping vector as the input and the adjustable control action set as the output, and uses the maximum expected return function as the optimization objective. When the expected value of the policy return no longer increases, the optimal control strategy is obtained;

[0064] S76. Based on the optimal control strategy, generate the charge and discharge strategy optimization instruction, and the charge and discharge strategy optimization instruction includes the optimal charging current setting value, the optimal discharging current setting value, the target cooling air speed level, and the temperature adjustment power level.

[0065] Optionally, the generation of the comprehensive response control instruction includes determining a control level label L(t) and constructing a comprehensive response control instruction C(t);

[0066] The determination of the control level label L(t) generates a control level label based on the micro-short circuit risk identification signal and the thermal runaway early warning signal. If R short (t)=1, then the control level label L(t)=2, that is, short circuit disposal is prioritized. If R thermal (t)=1 and R short (t)=0, then the control level label L(t)=1, that is, thermal control is prioritized. If R thermal (t)=0 and R short (t)=0, then the control level label L(t)=0, that is, the normal execution mode;

[0067] The construction of the comprehensive response control instruction C(t) is constructed based on the control level label L(t) in combination with the current charge and discharge strategy optimization instruction. When L(t)=2, an emergency stop instruction or a power-off control amount is output. When L(t)=1, the temperature control parameters are enhanced on the basis of the charge and discharge strategy optimization instruction. When L(t)=0, the charge and discharge and thermal management instructions set in the charge and discharge strategy optimization instruction are directly adopted.

[0068] The beneficial effects of the present invention are:

[0069] First, different from the existing cell detection methods that usually rely on single current and voltage parameter monitoring, temperature point collection, or static data model analysis, the present invention proposes an integrated high-sensitivity detection method that combines magnetoresistive sensing, thermal field inversion, multi-source modeling, and control optimization, and systematically constructs a full-process closed-loop system from "micro-short circuit identification - thermal runaway early warning - aging modeling - strategy control". In terms of real-time sensing, a tunneling magnetoresistance (TMR) sensor is used to replace the traditional Hall device to achieve current change identification at the microampere level, with the sensitivity increased to 0.1%, and by dynamically monitoring the three-axis components of the magnetic field, the spatial magnetic anomaly caused by the change of the material magnetic permeability is captured to realize non-contact temperature distribution estimation, thus significantly enhancing the early identification ability of micro-short circuits and local thermal runaway inside the cell.

[0070] Secondly, in terms of data modeling, the present invention innovatively constructs a physics-informed neural network (PINN) that integrates the magnetic distortion map in the manufacturing stage and the time-series characteristics such as voltage, current, magnetic field, and temperature in the operation stage. By introducing the lithium-ion concentration diffusion equation and the physical residual term of aging evolution into the neural network, the joint prediction of state variables such as the capacity retention rate, internal resistance growth, and voltage platform shift of the battery cell is realized. Different from the black-box modeling of traditional deep learning models, this method introduces a physically interpretable constraint mechanism in the structure, effectively reducing the overfitting risk of the model, improving the credibility and traceability of the prediction, and at the same time having strong cross-cycle generalization ability, capable of adapting to the evolution differences of battery cells under different types and working conditions.

[0071] Thirdly, in terms of risk response control, the present invention constructs a charge and discharge strategy optimizer based on the reinforcement learning framework, uses the predicted aging state mapping vector of the battery cell as the input, and combines the setting of the maximum reward function to realize the autonomous learning and output of the optimal control action. Different from the traditional fixed strategy rules, the present invention can dynamically adjust control instructions such as the maximum charging current and the cooling air volume level according to the state change, improving the flexibility and practicality of the strategy response. At the same time, the control level label is determined by combining the micro-short circuit risk identification signal and the thermal runaway early warning signal, and the control response strategy is multi-level integrated to ensure that power-off or heat dissipation intervention is preferentially carried out under high-risk conditions, thereby achieving a dynamic balance between safety guarantee and energy efficiency control.

[0072] In addition, in terms of data fusion processing, the present invention establishes a time alignment and feature encoding mechanism, uniformly standardizes multi-source signals from different time scales and spatial distributions, generates a model input vector with a consistent structure, and significantly improves the information fusion efficiency and the stability of downstream modeling. Through the unified input of the magnetic response deviation, the thermal risk index function, and the micro-short circuit identification signal, not only the fusion at the data level is realized, but also a vertical coupling mechanism from the sensor layer to the model layer and then to the control layer is constructed, solving the problem that detection, modeling, and control are isolated from each other and difficult to cooperate in traditional methods.

[0073] Overall, the present invention breaks through multiple limitations of the prior art in terms of sensitivity, response speed, interpretability, and closed-loop control, and constructs a new paradigm of battery cell safety management integrating "perception - cognition - prediction - decision-making". Compared with the existing passive detection methods that rely on single-point acquisition and rule reasoning, the present invention has comprehensive advantages of high detection accuracy, fast warning response, strong model credibility, and complete closed-loop control ability, can effectively reduce the safety risks during the use of battery cells, improve the stability and reliability of the battery system in complex environments, and has good engineering adaptability and promotion and application value. Description of the Drawings

[0074] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the accompanying drawings:

[0075] Figure 1 is a flowchart of a highly sensitive detection method for the safety of battery cells proposed by the present invention;

[0076] Figure 2 is a schematic diagram of the physical information neural network structure of the battery cell during the aging evolution in a highly sensitive detection method for the safety of battery cells proposed by the present invention;

[0077] Figure 3 is a logic diagram for constructing a comprehensive response control instruction and determining the control level in a highly sensitive detection method for the safety of battery cells proposed by the present invention. Detailed Description of the Invention

[0078] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0079] Refer to Figures 1-3 , a highly sensitive detection method for the safety of battery cells, comprising the following steps:

[0080] S1. Install a tunneling magnetoresistance sensor module inside the inner lid or outside the housing of the battery cell, and collect the current data and magnetic field data of the battery cell during operation in real time, and output a preliminary detection data set;

[0081] S2. Based on the preliminary detection data set, analyze the microampere to milliampere level fluctuations of the battery cell, and generate a micro-short circuit risk identification signal;

[0082] S3. Use the three-axis component data of the external magnetic field of the battery cell in the preliminary detection data set to identify the spatial magnetic anomaly caused by the change of material magnetic permeability, and generate a temperature anomaly distribution map;

[0083] S4. Based on the temperature anomaly distribution map, judge whether there is a potential thermal runaway trend, and generate a thermal runaway warning signal;

[0084] S5. During the manufacturing stage of the battery cell, apply a control excitation magnetic field, scan the pole piece with the magnetic field, and generate a magnetic distortion data map;

[0085] S6. Combine the preliminary detection data set, the micro-short circuit risk identification signal, the temperature anomaly distribution map and the magnetic distortion data map to construct the multi-source detection data of the battery cell, input it into the physical information neural network, construct the input layer, the main prediction branch and the physical residual branch, and train to obtain the final battery cell aging evolution model;

[0086] S7. Based on the final battery cell aging evolution model, construct a charge and discharge strategy optimizer to output charge and discharge strategy optimization instructions;

[0087] S8. Integrate the micro-short circuit risk identification signal, thermal runaway warning signal, and charge and discharge strategy optimization instructions to generate a comprehensive response control instruction.

[0088] A highly sensitive detection method for battery cell safety provided by the present invention breaks through the limitation of traditional battery cell monitoring relying on a single current and temperature parameter. By introducing a tunneling magnetoresistance sensor, high-dimensional magnetic anomaly analysis, extraction of magnetic distortion characteristics in the manufacturing stage, and physical information neural network modeling, a predictive control system integrating multi-source data is constructed. Combining the micro-short circuit identification, thermal runaway warning, and charge and discharge strategy linkage optimization mechanism not only improves the sensitivity and response speed of battery cell anomaly identification, but also realizes the interpretability of prediction results and the adaptive adjustment of control strategies. This method has the advantages of flexible deployment, expandable model, and strong industrial adaptability, and has broad engineering application value and popularization prospects.

[0089] In this embodiment, the preliminary detection data set includes battery cell working current data, battery cell terminal voltage data, three-axis components of the external magnetic field of the battery cell, battery cell surface temperature estimation data, and corresponding time stamp sequences.

[0090] The present invention realizes the multi-dimensional accurate characterization of the operating state of the battery cell by constructing a preliminary detection data set including current, voltage, three-axis components of the magnetic field, temperature estimation, and time stamps. This data set provides a high-precision input basis for micro-short circuit identification, magnetic anomaly analysis, and thermal runaway warning, significantly improving the accuracy and response speed of early risk identification of battery cells. It has the advantages of strong real-time performance, complete information, and good scalability, and is applicable to multi-type battery cell monitoring scenarios.

[0091] In this embodiment, the specific content of S2 includes:

[0092] S21. Extract the battery cell working current data from the preliminary detection data set to form a time series;

[0093] S22. Slide and segment the current time series with a fixed time window length τ1 to construct a set of current windows;

[0094] S23. Calculate the current mean and standard deviation for each window respectively. If the current mean of the k-th current window is less than the lower limit of the preset micro-short circuit current threshold and the current standard deviation is less than the preset current fluctuation threshold, it is marked as a micro-short circuit suspicious window;

[0095] S24. If there are no less than N consecutive micro-short circuit suspicious windows, generate a micro-short circuit risk identification signal R short (t)=1 at the starting time point t, otherwise set it as Rshort i(t) = 0.

[0096] By performing a sliding window analysis on the current time series of the battery cell and combining the current mean value and fluctuation characteristics to identify suspicious micro-short circuit behaviors, the present invention achieves high-sensitivity capture of milliamp-level weak abnormal signals. This method can accurately output micro-short circuit risk identification signals at the initial stage of a fault, significantly improving the timeliness and accuracy of anomaly detection. It has the advantages of high computational efficiency, strong adaptability, and can be embedded in the real-time operation of the BMS, and is applicable to intelligent safety monitoring scenarios of various lithium batteries.

[0097] In this embodiment, step S3 specifically includes:

[0098] S31. Extract the three-axis component data of the external magnetic field of the battery cell from the preliminary detection dataset, and calculate the magnetic anomaly response index at time t based on the variation characteristics of the magnetic field in spatial distribution:

[0099]

[0100] where Φ(t) is the magnetic anomaly response index, Ω is the three-dimensional spatial region covered by the TMR sensor, t represents the time variable in the continuous sampling process, r is the spatial position vector, is the spatial gradient tensor of the magnetic field, κ(·) is the temperature modulation coefficient function, and T(t, r) is the temperature distribution on the surface of the battery cell at position r and time t;

[0101] S32. Perform time-window smoothing processing on the magnetic anomaly response index to construct a sliding window feature curve;

[0102] S33. Perform statistical processing on the magnetic anomaly response index window Φ (k) to calculate the average value of the magnetic disturbance intensity within the window and the maximum value max(Φ (k) . When the average value of the magnetic disturbance intensity is greater than the preset average threshold Φ th or the maximum value is greater than the preset peak threshold Φ max , mark the magnetic anomaly response index window Φ (k) as a magnetic field anomaly window

[0103] S34. Concatenate all magnetic anomaly windows in chronological order to generate a temperature anomaly distribution map T map .

[0104] The present invention extracts the three-axis components of the external magnetic field of the battery cell, constructs a magnetic anomaly response index, and combines temperature modulation coefficient and spatial gradient analysis to achieve accurate identification of local magnetic disturbances and potential thermal anomalies. Through sliding window statistics and abnormal window splicing, a high-resolution temperature anomaly distribution map is generated, significantly improving the spatial identification ability and early warning lead time of the precursor of battery cell thermal runaway. It has the advantages of non-contact, high sensitivity, strong interpretability, etc., and is applicable to thermal risk monitoring in multiple scenarios.

[0105] In this embodiment, the generation of the thermal runaway warning signal specifically includes constructing a thermal risk index function and determining the thermal runaway warning signal;

[0106] The construction of the thermal risk index function:

[0107]

[0108] Among them, H(t) is the thermal risk index, λ1 is the weight coefficient of the temperature rise rate, T map is the temperature anomaly distribution map, is the first-order time derivative of the temperature anomaly map, λ2 is the weight coefficient of the magnetic anomaly response, Φ(t) is the magnetic anomaly response index, λ3 is the temperature spatial gradient, is the spatial gradient of the temperature field, t represents the time variable in the continuous sampling process, r is the spatial position vector, λ4 is the weight coefficient of the historical temperature rise integral, δ is the length of the retrospective time window, τ is the historical time variable, and w(·) is the time decay weight function;

[0109] The determination of the thermal runaway warning signal compares the thermal risk index H(t) with the preset threshold H th When H(t)>H th is satisfied, a thermal runaway warning signal is generated at the corresponding time point t, marked as R thermal (t)=1, otherwise marked as R thermal (t)=0.

[0110] The present invention constructs a thermal risk index function that integrates multi-dimensional thermal field information and magnetic anomaly response to systematically evaluate the potential thermal runaway trend during the operation of the battery cell. This function comprehensively considers the spatial distribution and time derivative of the temperature anomaly map, the intensity of magnetic disturbance, and the change of historical temperature rise integral, and introduces a time decay mechanism to enhance the perception ability of long-term heat accumulation effect. Compared with the traditional early warning method that only relies on the single-point temperature threshold judgment, this method can achieve dynamic and continuous identification of the early signal of thermal runaway. It is applicable to high-safety-level power batteries and energy storage systems, and has important engineering value for reducing the risk of thermal accidents and improving the overall system stability.

[0111] In this embodiment, the magnetic distortion data map includes spatial position coordinates, magnetic response deviation intensity, and anomaly marking labels; the magnetic response deviation intensity represents the maximum difference between the actually measured magnetic response value and the reference magnetic response value at the corresponding position; the anomaly marking labels perform anomaly determination on each position based on whether the magnetic response deviation exceeds a threshold.

[0112] The present invention realizes the precise quantification of the structural consistency and magnetic response anomalies of the electrode sheets during the cell manufacturing stage by constructing a magnetic distortion data map. The map includes spatial position coordinates, magnetic response deviation intensity, and anomaly marking labels, and can identify potential defects such as uneven coating and lithium deposition. Compared with traditional manual sampling inspection or contact detection methods, this method has the advantages of high resolution, non-destructive, and full coverage, and can complete the screening of structural anomalies before the cell is put into production, improving production consistency and the safety of later operation.

[0113] In this embodiment, the specific steps of S6 are as follows:

[0114] S61. Based on the preliminary detection data set and combined with the micro-short circuit risk identification signal, the temperature anomaly distribution map, and the magnetic distortion data map, construct the multi-source detection data of the cell after aligning the time indexes of each item;

[0115] S62. Input the multi-source detection data of the cell into the physics-informed neural network to construct the cell aging evolution model;

[0116] S63. Construct the combined loss function:

[0117]

[0118] Among them, is the total loss function, t is the time index variable, is the predicted value of the cell aging state, S true (t) is the actually labeled cell aging state value, λ phys is the weight coefficient of the physical residual term, is the residual term of the cell aging evolution physical equation;

[0119] S64. Adopt a method combining forward propagation and backward propagation to perform joint minimization training on the parameter set θ of the physics-informed neural network;

[0120] S65. During continuous multi-round training iterations, when both the data error and the physical residual in the combined loss function satisfy that the error change rate is lower than the threshold, the error value is lower than the target lower limit, or the training error no longer decreases, it is determined that the model training process is completed, and the final cell aging evolution model is output.

[0121] The present invention constructs a unified input sequence containing multi-source detection data such as current, voltage, magnetic anomaly, thermal distribution, and manufacturing defects, and introduces a joint loss function, incorporating both prediction error and physical residual into the optimization objective to construct an aging evolution model of the battery cell with physical consistency. This method effectively integrates data-driven and mechanism constraints, improving the prediction accuracy and stability of the model for aging characteristics such as capacity retention rate and internal resistance growth. Compared with traditional black-box neural network models, the present invention not only improves the generalization ability but also significantly enhances the credibility and interpretability of the model output results, having engineering deployment value and wide applicability.

[0122] In this embodiment, the S62 specifically includes:

[0123] S621. Construct the input layer of the neural network, and set the input vector as the multi-source detection data of the battery cell;

[0124] S622. Based on the multi-source detection data of the battery cell, construct a feedforward main prediction branch neural network. The main neural network learns the mapping relationship between the input features and the aging state of the battery cell through multiple non-linear transformations, and outputs the predicted value of the aging state of the battery cell at the corresponding moment. The predicted value of the battery cell aging state includes the estimated value of the capacity retention rate, the internal resistance growth value, and the voltage platform offset value;

[0125] S623. Construct a physical residual branch, add the physical model of the battery cell aging process as a constraint to the network training, and the physical residual branch fits the physical residual term of the battery cell degradation equation:

[0126]

[0127] Wherein, is the residual term of the physical equation of the battery cell aging evolution, t is the time index variable, C Li is the lithium ion concentration distribution function inside the battery cell, D is the lithium ion diffusion coefficient, is the second-order gradient of the lithium concentration in space, k is the lithium loss rate constant, γ is the defect modulation coefficient, f defect (r) is the defect perturbation function, and r is the spatial position vector;

[0128] S624. Combine the outputs of the main branch and the residual branch, and respectively output the predicted value of the battery cell aging state and the residual term of the physical equation of the battery cell aging evolution

[0129] S625. Set the physical information neural network parameter set as θ, including the parameters of the main branch and the physical residual branch;

[0130] S626. The input layer, the main prediction branch, and the physical residual branch together constitute the aging evolution model of the battery cell.

[0131] By constructing a physical information neural network architecture including an input layer, a backbone prediction branch, and a physical residual branch, the present invention realizes high-precision modeling of the aging state of the battery cell and physical consistency constraints. The backbone branch learns the mapping relationship between multi-source data and aging indicators such as capacity retention rate and internal resistance growth, while the residual branch introduces a lithium-ion diffusion and defect modulation model to quantify the physical deviation in the battery cell degradation process. This structure embeds interpretable physical knowledge while maintaining the flexibility of data-driven, significantly improving the credibility and engineering applicability of the prediction results, and is applicable to life prediction and health management in high-safety-level energy storage scenarios.

[0132] In this embodiment, the S7 specifically includes:

[0133] S71. Based on the final battery cell aging evolution model, input the multi-source detection data of the battery cell and output the predicted value of the battery cell aging state;

[0134] S72. Extract the capacity retention rate estimated value, internal resistance growth value, and voltage platform offset value in the predicted value of the battery cell aging state to form a state mapping vector;

[0135] S73. Set the adjustable control action set A(t), including the maximum charging current, maximum discharging current, cooling wind speed level, and temperature regulation power factor in the current time period;

[0136] S74. Construct the optimization objective function of the charge and discharge control strategy:

[0137]

[0138] where A(t) is the adjustable control action set, is to solve the optimal policy action, is the expected value of the policy return, t is the time index variable, R(t) is the policy return function at the current time step t, T is the total length of the policy optimization time domain, α1 is the capacity retention rate weight coefficient, ΔC ret (t) is the positive gain of the current action on the capacity retention rate, α2 is the internal resistance suppression weight coefficient, ΔR in (t) is the internal resistance growth amplitude caused by the current action, α3 is the energy consumption suppression weight coefficient, and E(t) is the energy consumption per unit time of the current action;

[0139] S75. Adopt a method based on deep reinforcement learning to train the optimization objective function of the charge and discharge control strategy. The policy network takes the state mapping vector as the input and the adjustable control action set as the output, uses the maximum expected return function as the optimization objective, and obtains the optimal control strategy when the expected value of the policy return no longer increases;

[0140] S76. Based on the optimal control strategy, generate charge and discharge strategy optimization instructions, where the charge and discharge strategy optimization instructions include the optimal charging current setting value, the optimal discharging current setting value, the target cooling air velocity level, and the temperature regulation power level.

[0141] In the present invention, a state mapping vector is constructed based on the prediction result of the cell aging state, and an objective function for optimizing the charge and discharge control strategy is constructed by combining capacity retention, internal resistance suppression, and energy consumption constraints. The deep reinforcement learning method is used to realize the adaptive training of control actions and strategy optimization. The policy network can dynamically output the optimal charging current, discharging current, and thermal management parameters, realizing the balanced control of charge and discharge performance and cell life. Compared with the traditional fixed strategy or rule table method, the present invention has the advantages of high strategy accuracy, strong self-learning ability, and strong adaptability to the operating environment, effectively improving the operating efficiency and use safety of the cell system.

[0142] In this embodiment, the generation of the comprehensive response control instruction includes determining the control level label L(t) and constructing the comprehensive response control instruction C(t);

[0143] The determination of the control level label L(t) generates a control level label according to the micro-short circuit risk identification signal and the thermal runaway early warning signal. If R short (t) = 1, then the control level label L(t) = 2, that is, short circuit disposal is prioritized. If R thermal (t) = 1 and R short (t) = 0, then the control level label L(t) = 1, that is, thermal control is prioritized. If R thermal (t) = 0 and R short (t) = 0, then the control level label L(t) = 0, that is, the normal execution mode;

[0144] The construction of the comprehensive response control instruction C(t) is constructed according to the control level label L(t) in combination with the current charge and discharge strategy optimization instruction. When L(t) = 2, an emergency stop instruction or a power-off control quantity is output. When L(t) = 1, the temperature control parameters are enhanced on the basis of the charge and discharge strategy optimization instruction. When L(t) = 0, the charge and discharge and thermal management instructions set in the charge and discharge strategy optimization instruction are directly adopted.

[0145] The present invention dynamically determines the control level label by comprehensively considering the micro-short circuit risk signal and the thermal runaway early warning signal, and generates a response control instruction based on the level strategy and the optimized charge and discharge instruction, realizing an intelligent hierarchical disposal mechanism under abnormal conditions. Compared with the traditional single-trigger control method, this method can give priority to power-off protection in the short circuit risk scenario and actively strengthen thermal management in the thermal risk scenario, realizing precise intervention and flexible control, and significantly improving the safety response ability and operating stability of the cell system.

[0146] Example 1:

[0147] To verify the feasibility of the present invention in implementation, the present invention is applied to the aging test link of a lithium-ion power cell production line to perform high-precision operation monitoring and anomaly identification on a batch of lithium iron phosphate cell samples. In the previous operation of this production line, problems such as sudden thermal runaway of some cells in the later stage of use and undetected micro-shorts in advance occurred, affecting product consistency and safety. The traditional scheme based on Hall current sensors and thermocouple point detection has obvious deficiencies in terms of response speed and identification accuracy. Especially in the initial stage of microampere-level fluctuations and spatial heat diffusion of cells, phenomena such as alarm lag or detection failure are likely to occur.

[0148] During the implementation process, the present invention integrates tunnel magnetoresistance (TMR) sensors in the top cover structure of each cell to form a magnetic field and current sensing channel close to the cell source, supplemented by a magnetic distortion analysis module and a physical neural network modeling platform, realizing a full-process closed-loop from real-time data acquisition, micro-short identification, thermal runaway trend judgment to aging state prediction and optimized control response. Each test cell obtains a complete working characteristic sequence through simulating the standard usage cycle, including charge and discharge cycles, static tests, and temperature rise responses.

[0149] During the test period, a total of 200 lithium iron phosphate cells were tested. Among them, the present invention's scheme successfully captured the characteristic fluctuation signals in the early stage of micro-shorts for 26 cells, triggering the high-sensitivity risk identification module. In the traditional method, only 8 cells triggered the initial current anomaly alarm, and the average warning time lagged behind the present invention's scheme by about 20 minutes or more. Through joint analysis with the temperature anomaly map, 14 cells showed obvious local heat diffusion trends within 24 hours after the formation of the magnetic anomaly window, further verifying the high correlation between magnetic anomaly and the evolution of thermal runaway.

[0150] In terms of aging state prediction, the physical information neural network model fuses and inputs data such as current, voltage, magnetic anomaly, temperature estimation, and magnetic distortion map data in the manufacturing stage, and performs output fitting on the capacity retention rate, internal resistance growth rate, and voltage platform change of each test cell. The average prediction error is less than 1.8%. The model also outputs the physical residual term during the cell degradation process to judge whether the prediction deviation is affected by external disturbances or internal structure changes, improving the interpretability of the model.

[0151] For the battery cells marked as high-risk status, the control strategy optimization module dynamically adjusts the upper limits of charge and discharge currents and the temperature control strategy. Compared with the battery cell group without intervention, after the optimization strategy is executed, the average surface temperature rise drops by about 6°C, and the number of abnormal temperature peaks drops by more than 60%. At the same time, at the system response layer, the comprehensive control instruction automatically integrates the micro-short circuit signal and the thermal runaway trend, and realizes multi-level response strategies such as power-off, load reduction, or enhanced temperature control according to the set control level label, effectively suppressing the abnormal evolution trend and improving the overall system safety level.

[0152] The application results of the present invention show that in multiple sets of comparative tests, the micro-short circuit recognition rate is increased by more than 60%, the false alarm rate is decreased by about 15%, and the early warning ability is significantly better than the traditional detection scheme. The thermal runaway suppression effect shows that the peak temperature suppression ability is increased by nearly 8°C, and the intervention response speed of the control strategy is increased by more than 2 times. The above data fully illustrate that the present invention not only has a leading advantage in sensing sensitivity, but also forms an engineering closed-loop at the intelligent response layer, providing a feasible and deployable system path for improving the intrinsic safety level of battery cell products.

[0153] Table 1 Key performance data of the high-sensitivity detection and response system for battery cells during batch testing

[0154]

[0155]

[0156] It can be seen from the comparison data in Table 1 above that in the practical application of high-sensitivity detection and response of battery cells, the present invention is significantly better than the traditional detection method in terms of micro-short circuit recognition ability, thermal runaway early warning response speed, aging state modeling accuracy, and control strategy execution effectiveness. Specifically, in the early identification of micro-short circuits, the present invention realizes the real-time capture and marking of milliamper-level fluctuations by integrating a tunneling magnetoresistance sensor and a sliding window statistical discrimination mechanism. The table shows that the present invention successfully detects micro-short circuit signals in multiple groups of samples such as C001, C014, and C048, while the traditional system fails to identify them completely, reflecting the sensitivity improvement in the dimension of weak signal identification.

[0157] In terms of thermal anomaly response, the present invention constructs a thermal risk index function based on the fusion of magnetic anomaly and temperature gradient, and cooperates with the neural network prediction output to realize the trend warning dozens of minutes in advance. In Table 1, for battery cells such as C001, C077, and C156, when the traditional system fails to give an early warning or is severely lagged, the present invention realizes an average early warning time of 26.5 minutes, effectively avoiding the potential thermal runaway risk brought by the rapid rise of the temperature peak. This ability has extremely high practical value for safety guarantee in large-scale battery cell usage scenarios.

[0158] In terms of the accuracy of aging modeling, the present invention adopts a physics-informed neural network (PINN) structure, deeply coupling multi-source detection data with the lithium-ion diffusion equation. The prediction error of the capacity retention rate output by the model is basically controlled between 1.0% and 2.0%. As shown in C023, C048, and C189, the prediction errors are 1.0%, 2.0%, and 1.4% respectively, which is better than the error performance of more than 3% of most traditional black-box models under actual working conditions. This structure integrating physical interpretability not only improves the accuracy but also enhances the credibility and traceability of the prediction results.

[0159] At the level of control response strategy, the optimal control strategy generated by the present invention through reinforcement learning can dynamically adjust key parameters such as discharge current and temperature control level. As can be seen from the table, after the implementation of strategy optimization for all the battery cells triggering the comprehensive control instructions, the temperature rise amplitude has decreased by more than 5.0°C, and the maximum can reach 7.3°C (such as C156), greatly reducing the risk of thermal aggregation. In addition, through the adaptive judgment of the control level label and the issuance of instructions, the present invention realizes hierarchical linkage response based on the risk level, which is different from the single control logic of "whether to alarm" in the traditional scheme, and the response is more hierarchical and targeted.

[0160] More importantly, the present invention demonstrates good stability and low false alarm ability in large-scale tests. Among 200 groups of test samples, only 3 groups had false triggers, and the overall false positive rate was about 1.5%, which is much lower than about 4% of the traditional system. At the same time, the charge and discharge efficiency and temperature control load after the intervention of the control strategy remain at a relatively optimal balance point, and there is no problem of system performance degradation caused by overprotection, reflecting the high adaptability of the control mechanism of the present invention.

[0161] In summary, the present invention not only realizes the early identification of micro-short circuit and thermal runaway of battery cells, but also constructs a safety detection and response control closed-loop through physical modeling and reinforcement learning optimization strategy, comprehensively improving the intelligence, response efficiency and practical value of the system, and having great potential for wide promotion in high-safety-level scenarios such as electric vehicles and grid energy storage.

[0162] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A highly sensitive detection method for the safety of battery cells, characterized in that, It includes the following steps: S1. Install a tunneling magnetoresistance sensor module inside the top cover or outside the shell of the battery cell to collect the current data and magnetic field data of the battery cell during operation in real time, and output a preliminary detection data set; S2. Analyze the microampere-to-milliampere-level fluctuations of the battery cell based on the preliminary detection data set, and generate a micro-short circuit risk identification signal; S3. Use the three-axis component data of the external magnetic field of the battery cell in the preliminary detection data set to identify the spatial magnetic anomaly caused by the change of material magnetic permeability, and generate a temperature anomaly distribution map; S4. Based on the temperature anomaly distribution map, judge whether there is a potential thermal runaway trend, and generate a thermal runaway warning signal; S5. During the manufacturing stage of the battery cell, apply a control excitation magnetic field to scan the pole piece with the magnetic field to generate a magnetic distortion data map; S6. Combine the preliminary detection data set, the micro-short circuit risk identification signal, the temperature anomaly distribution map and the magnetic distortion data map to construct the multi-source detection data of the battery cell, input it into the physics-informed neural network, construct the input layer, the main prediction branch and the physical residual branch, and train to obtain the final battery cell aging evolution model; S7. Based on the final battery cell aging evolution model, construct a charge and discharge strategy optimizer, and output a charge and discharge strategy optimization instruction; S8. Uniformly fuse the micro-short circuit risk identification signal, the thermal runaway warning signal and the charge and discharge strategy optimization instruction to generate a comprehensive response control instruction.

2. The highly sensitive detection method for the safety of battery cells according to claim 1, wherein The preliminary detection data set includes the battery cell working current data, the battery cell terminal voltage data, the three-axis component data of the external magnetic field of the battery cell, the estimated battery cell surface temperature data, and the corresponding time stamp sequence.

3. A highly sensitive detection method for the safety of battery cells according to claim 1, characterized in that, The specific content of S2 includes: S21. Extract the battery cell working current data from the preliminary detection data set to form a time series; S22. Perform sliding segmentation on the current time series with a fixed time window length τ1 to construct a current window set; S23. Calculate the current mean and standard deviation for each window respectively. If the current mean of the kth current window is less than the preset lower limit of the micro-short circuit current threshold and the current standard deviation is less than the preset current fluctuation threshold, it is marked as a micro-short circuit suspicious window; S24. If there are no less than N continuously occurring micro-short circuit suspicious windows, generate a micro-short circuit risk identification signal R at the starting time point t short (t) = 1, otherwise set it as R short (t) = 0.

4. A highly sensitive detection method for the safety of battery cells according to claim 1, characterized in that, The specific content of S3 includes: S31. Extract the three-axis component data of the external magnetic field of the battery cell from the preliminary detection data set, and calculate the magnetic anomaly response index at time t based on the change characteristics of the magnetic field in the spatial distribution: where, Φ(t) is the magnetic anomaly response index, Ω is the three-dimensional spatial region covered by the TMR sensor, t represents the time variable in the continuous sampling process, r is the spatial position vector, is the spatial gradient tensor of the magnetic field, κ(·) is the temperature modulation coefficient function, and T(t, r) is the temperature distribution on the surface of the battery cell at position r and time t; S32. Perform time window smoothing processing on the magnetic anomaly response index to construct a sliding window feature curve; S33. Statistically process the magnetic anomaly response index window Φ (k) and calculate the average value of the magnetic disturbance intensity within the window and the maximum value max(Φ (k) . When the average value of the magnetic disturbance intensity is greater than the preset average threshold Φ th or the maximum value is greater than the preset peak threshold Φ max , mark the magnetic anomaly response index window Φ (k) as a magnetic field anomaly window S34. Stitch all magnetic anomaly windows in chronological order to generate the temperature anomaly distribution map T within the time period map .

5. A highly sensitive detection method for the safety of battery cells according to claim 1, characterized in that The generation of the thermal runaway warning signal specifically includes constructing a thermal risk index function and determining the thermal runaway warning signal; The construction of the thermal risk index function: Among them, H(t) is the heat risk index, λ1 is the weight coefficient of the temperature rise rate, and T map is the temperature anomaly distribution map, is the first-order time derivative of the temperature anomaly map, λ2 is the weight coefficient of the magnetic anomaly response, Φ(t) is the magnetic anomaly response index, λ3 is the temperature spatial gradient, is the spatial gradient of the temperature field, t represents the time variable in the continuous sampling process, r is the spatial position vector, λ4 is the weight coefficient of the historical temperature rise integral, δ is the length of the retrospective time window, τ is the historical time variable, and w(·) is the time decay weight function; The determined thermal runaway warning signal compares the thermal risk index H(t) with a preset threshold H th and when H(t)>H th is satisfied, a thermal runaway warning signal is generated at the corresponding time point t, marked as R thermal (t)=1, otherwise marked as R thermal (t)=0.

6. A highly sensitive detection method for the safety of battery cells according to claim 1, characterized in that, The magnetic distortion data map includes spatial position coordinates, magnetic response deviation intensity, and anomaly marking labels; the magnetic response deviation intensity represents the maximum difference between the actually measured magnetic response value and the reference magnetic response value at the corresponding position; the anomaly marking label performs anomaly determination on each position based on whether the magnetic response deviation exceeds the threshold.

7. A highly sensitive detection method for the safety of battery cells according to claim 1, characterized in that The specific content of S6 includes: S61. Based on the preliminary detection data set and in combination with the micro-short circuit risk identification signal, the temperature anomaly distribution map and the magnetic distortion data map, construct the multi-source detection data of the battery cell after aligning the time indexes of each item; S62. Input the multi-source detection data of the battery cell into the physics-informed neural network to construct a battery cell aging evolution model; S63. Construct a combined loss function: Among them, is the total loss function, t is the time index variable, is the predicted value of the battery cell aging state, S true (t) is the actually labeled battery cell aging state value, λ phys is the weight coefficient of the physical residual term, is the residual term of the physical equation for battery cell aging evolution; S64. Adopt a method combining forward propagation and backward propagation to perform joint minimization training on the parameter set θ of the physics-informed neural network; S65. During continuous multi-round training iterations, if both the data error and the physical residual in the joint loss function satisfy that the error change rate is lower than the threshold, the error value is lower than the target lower limit, or the training error no longer decreases, it is determined that the model training process is completed, and the final cell aging evolution model is output.

8. A highly sensitive detection method for the safety of battery cells according to claim 7, characterized in that, The specific content of S63 includes: S621. Construct the input layer of the neural network, and set the input vector as the multi-source detection data of the battery cell; S622. Based on the multi-source detection data of the battery cells, a feedforward main prediction branch neural network is constructed. The main neural network learns the mapping relationship between the input features and the aging state of the battery cells through multiple non-linear transformations, and outputs the predicted value of the aging state of the battery cells at the corresponding moment. The predicted value of the aging state of the battery cells includes the estimated value of the capacity retention rate, the growth value of the internal resistance, and the offset value of the voltage platform. S623. Construct a physical residual branch, add the physical model of the battery cell aging process as a constraint to the network training, and the physical residual branch fits the physical residual term of the battery cell degradation equation: Among them, is the residual term of the physical equation for the aging evolution of the battery cell, t is the time index variable, C Li is the lithium-ion concentration distribution function inside the battery cell, D is the lithium-ion diffusion coefficient, is the second-order gradient of lithium concentration in space, k is the lithium loss rate constant, γ is the defect modulation coefficient, f defect (r) is the defect perturbation function, and r is the spatial position vector; S624. Converge the outputs of the main branch and the residual branch, and respectively output the predicted values of the cell aging state and the residual term of the physical equation for cell aging evolution S625. Set the parameter set θ of the physics-informed neural network, including the parameters of the main branch and the physical residual branch; S626. The input layer, the main prediction branch and the physical residual branch jointly constitute the battery cell aging evolution model.

9. A highly sensitive detection method for the safety of battery cells according to claim 1, characterized in that The specific content of S7 includes: S71. Based on the final battery cell aging evolution model, input the multi-source detection data of the battery cell and output the predicted value of the battery cell aging state; S72. Extract the capacity retention rate estimate value, internal resistance growth value and voltage platform offset value in the predicted value of the battery cell aging state to form a state mapping vector; S73. Set an adjustable control action set A(t), including the maximum charging current, maximum discharging current, cooling air velocity level and temperature regulation power factor in the current time period; S74. Construct an optimization objective function for the charge and discharge control strategy: where, A(t) is a set of adjustable control actions, is to solve the optimal policy action, is the expected value of the policy return, t is the time index variable, R(t) is the policy return function at the current time step t, T is the total length of the policy optimization time domain, α1 is the weight coefficient of the capacity retention rate, ΔC ret (t) is the positive gain of the current action on the capacity retention rate, α2 is the weight coefficient of internal resistance suppression, ΔR in (t) is the increase in internal resistance caused by the current action, α3 is the weight coefficient of energy consumption suppression, and E(t) is the energy consumption per unit time of the current action; S75. Adopt a method based on deep reinforcement learning to train the optimization objective function of the charge and discharge control strategy. The policy network takes the state mapping vector as the input and the adjustable control action set as the output, uses the maximum expected return function as the optimization objective, and obtains the optimal control strategy when the expected value of the policy return no longer increases; S76. Based on the optimal control strategy, generate a charge and discharge strategy optimization instruction, and the charge and discharge strategy optimization instruction includes the optimal charging current setting value, optimal discharging current setting value, target cooling air velocity level and temperature regulation power level.

10. A highly sensitive detection method for the safety of battery cells according to claim 1, characterized in that The generation of the comprehensive response control instruction includes determining the control level label L(t) and constructing the comprehensive response control instruction C(t); The determination control level label L(t) generates a control level label based on the micro-short circuit risk identification signal and the thermal runaway warning signal. If R short (t) = 1, then the control level label L(t) = 2, which means short circuit handling takes priority. If R thermal (t) = 1 and R short (t) = 0, then the control level label L(t) = 1, which means thermal control takes priority. If R thermal (t) = 0 and R short (t) = 0, then the control level label L(t) = 0, which means the normal execution mode; The construction of the comprehensive response control instruction C(t) is constructed by combining the control level label L(t) and the current charge and discharge strategy optimization instruction. When L(t)=2, an emergency stop instruction or a power-off control quantity is output. When L(t)=1, the temperature control parameter is enhanced on the basis of the charge and discharge strategy optimization instruction. When L(t)=0, the charge and discharge and thermal management instructions set in the charge and discharge strategy optimization instruction are directly adopted.

Images