Online sensing system for full-life-cycle safety situation of lithium battery box

Abstract

The invention, which relates to the technical field of lithium battery safety monitoring, discloses an on-line sensing system for a full life cycle safety situation of a lithium battery box, the on-line sensing system comprises a signal excitation and acquisition unit arranged in the battery box, and a data processing and control unit electrically connected with the signal excitation and acquisition unit. And the signal excitation and acquisition unit is configured to apply a micro-amplitude alternating current excitation signal to the single battery and acquire the voltage and current response of the single battery. According to the online sensing system for the full-life-cycle safety situation of the lithium battery box, by constructing a collaborative sensing architecture with dual-engine state estimation and dynamic credibility weighted fusion, the defect that a single model is insufficient in adaptability in the full life cycle of a battery is effectively overcome. By introducing a feedback verification closed-loop mechanism, online secondary verification and self-correction of an internal state estimation result are realized, so that the accuracy and reliability of early thermal runaway risk early warning are greatly improved, and occurrence of misinformation and missing report is remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of lithium battery safety monitoring technology, specifically to an online sensing system for the safety status of lithium battery boxes throughout their entire lifecycle. Background Technology

[0002] In the field of lithium battery box safety monitoring, existing technologies mainly rely on online monitoring of external electrical parameters such as voltage, current, and surface temperature, and provide safety warnings based on preset fixed thresholds. These solutions fall short when faced with the complex evolution of the battery throughout its entire lifecycle. Particularly noteworthy is that while the internal state of a battery undergoes subtle changes in the early stages of use, key internal parameters such as internal resistance and interface film thickness evolve nonlinearly with increasing cycle count. Existing systems generally lack real-time, adaptive sensing capabilities for these internal states, and their static models cannot accurately capture the minute characteristic changes in the early stages of battery aging, leading to highly unreliable early predictions of serious faults such as thermal runaway. Specifically, false alarms are frequent in the early stages of battery life, while true risks are easily missed in the later stages of aging. Although some improved solutions attempt to introduce more sensors or complex algorithms, they often require offline training with large amounts of historical data, or are computationally burdensome and difficult to continuously update online, failing to fundamentally solve the core challenge of accurately sensing early risks throughout the entire battery lifecycle, from its pristine state to the end of aging. Therefore, an online sensing system for the safety status of lithium battery boxes throughout their entire life cycle aims to solve the practical problem that existing technologies cannot achieve real-time and accurate prediction of early thermal runaway risks throughout the entire battery life cycle. Summary of the Invention

[0003] The purpose of this invention is to provide an online sensing system for the safety status of lithium battery boxes throughout their entire life cycle, in order to solve the problems mentioned in the background art.

[0004] To address the aforementioned technical problems, this invention provides the following technical solution: an online sensing system for the safety status of a lithium battery box throughout its entire lifecycle, comprising a signal excitation and acquisition unit arranged within the battery box, and a data processing and control unit electrically connected to the signal excitation and acquisition unit; the signal excitation and acquisition unit is configured to apply a micro-amplitude AC excitation signal to a single battery cell and acquire its voltage and current response; the data processing and control unit includes:

[0005] The dual-engine state estimation module has a parallel simplified physical engine and a data-driven engine. The simplified physical engine is configured to quickly extract the internal state parameters of the battery based on the voltage and current response data. The data-driven engine is configured to learn and map the internal health state of the battery based on the battery's historical and real-time external electrical parameter data.

[0006] The dynamic credibility weighted fusion unit, which is connected to the dual-engine state estimation module, is configured to receive the preliminary estimates of the same internal state parameter from the simplified physical engine and the data-driven engine, and assign appropriate weights to the preliminary estimates of the two engines according to the current operating condition and life stage of the battery, perform the first fusion calculation, and output the primary fusion state quantity.

[0007] The safety status scoring and prediction model, which is connected to the dynamic credibility weighted fusion unit, is configured to receive the primary fusion state quantity and comprehensively calculate the current safety status score of the battery, while making short-term predictions of the battery status at future times based on the current status.

[0008] The feedback verification closed-loop module, connected between the security situation scoring and prediction model and the dynamic credibility weighted fusion unit, is configured to continuously compare the short-term predicted value made by the security situation scoring and prediction model with the signal excitation and the actual measurement value subsequently collected by the acquisition unit, and send the generated prediction deviation information as a feedback signal to the dynamic credibility weighted fusion unit to drive the dynamic credibility weighted fusion unit to adjust its trust weights on the output values ​​of the two engines in the dual-engine state estimation module.

[0009] The system forms an enhanced closed loop of perception, fusion, prediction, verification and correction through the feedback verification closed loop module, and finally outputs a security situation score that tends to be stable after multiple iterations for early warning.

[0010] Furthermore, the simplified physics engine is a reduced-order electrochemical-thermal coupled model that focuses on the online extraction of key internal state parameters of the battery, such as internal resistance and lithium-ion diffusion coefficient.

[0011] Furthermore, the data-driven engine is a lightweight deep learning network model that takes as input the battery's voltage, current, and temperature data streams and outputs an estimate of the battery's internal health status.

[0012] Furthermore, the dynamic reliability weighted fusion accelerator allocates weights based on information about the battery's operating conditions, such as whether it is in a static, charging, or discharging state, as well as information about the early, middle, or late stages of battery life, based on the number of cycles.

[0013] Furthermore, the feedback verification closed-loop module feeds back the prediction deviation information to the dynamic credibility weighted fusion unit, thereby enabling online adaptive adjustment of the credibility of the output results of the dual-engine state estimation module, and thus completing the secondary verification and correction of the primary fusion state quantity.

[0014] Furthermore, the frequency components of the micro-amplitude AC excitation signal applied by the signal excitation and acquisition unit are adaptively selected according to the current health status and temperature conditions of the battery.

[0015] Furthermore, the safety situation score and the safety situation score calculated by the prediction model are continuous values ​​ranging from zero to one, which are associated with the thermal runaway risk level of the battery.

[0016] Furthermore, the data-driven engine can receive internal state parameter estimates provided by the simplified physics engine as its auxiliary input features to enhance the accuracy of its mapping relationships.

[0017] Furthermore, the final output of the system is a stable security situation score and internal state variables output by the security situation scoring and prediction model after multiple iterations of the enhanced closed loop.

[0018] Furthermore, the system also includes a cloud-based data platform that communicates with the data processing and control unit. This platform is used to perform statistical analysis on the operating modes and failure characteristics of the battery group throughout its entire life cycle, and to synchronize the analysis results downlink to the local safety status scoring and prediction model of individual batteries, so as to achieve continuous evolution of the model.

[0019] This invention provides an online sensing system for the safety status of lithium battery boxes throughout their entire lifecycle. It offers the following advantages:

[0020] This online perception system for the safety status of lithium battery boxes throughout their entire lifecycle effectively overcomes the shortcomings of a single model in terms of adaptability throughout the battery's lifecycle by constructing a collaborative perception architecture that integrates dual-engine state estimation and dynamic reliability weighted fusion. By introducing a feedback verification closed-loop mechanism, it achieves online secondary verification and self-correction of internal state estimation results, thereby significantly improving the accuracy and reliability of early thermal runaway risk warnings and significantly reducing false alarms and missed alarms.

[0021] This online sensing system for the entire lifecycle safety status of lithium battery boxes leverages a cloud-based data platform to enable group knowledge sharing and online evolution of individual models. This allows the local sensing system to continuously absorb the operational experience of the battery group and constantly optimize its parameters and judgment thresholds. This not only strengthens the system's ability to track the nonlinear aging process of batteries but also achieves a virtuous cycle where the accuracy of safety status perception continuously improves over time, providing long-term assurance for the comprehensive safety monitoring of lithium battery boxes. Attached Figure Description

[0022] Figure 1 This is a data flow diagram of an online sensing system for the safety status of a lithium battery box throughout its entire lifecycle, as described in this invention.

[0023] Figure 2 This is a weight adaptive adjustment logic diagram of an online sensing system for the safety status of a lithium battery box throughout its entire life cycle, as described in this invention.

[0024] Figure 3 This is a flowchart illustrating the safety scoring calculation process of an online sensing system for the full lifecycle safety status of lithium battery boxes, as described in this invention. Detailed Implementation

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

[0026] Please see Figures 1 to 3 This invention provides a technical solution: an online sensing system for the safety status of a lithium battery box throughout its entire lifecycle, comprising a signal excitation and acquisition unit arranged inside the battery box, and a data processing and control unit electrically connected to the signal excitation and acquisition unit; the signal excitation and acquisition unit is configured to apply a micro-amplitude AC excitation signal to each battery cell and acquire its voltage and current response; the data processing and control unit includes:

[0027] The dual-engine state estimation module has a parallel simplified physics engine and a data-driven engine. The simplified physics engine is constructed to quickly extract the internal state parameters of the battery based on voltage and current response data, while the data-driven engine is constructed to learn and map the internal health state of the battery based on the battery's historical operation and real-time external electrical parameter data.

[0028] The dynamic credibility weighted fusion unit, connected to the dual-engine state estimation module, is configured to receive preliminary estimates of the same internal state parameter from the simplified physical engine and the data-driven engine, and assign appropriate weights to the preliminary estimates of the two engines according to the current operating condition and life stage of the battery, perform the first fusion calculation, and output the primary fusion state quantity.

[0029] The safety status scoring and prediction model, which is connected to the dynamic credibility weighted fusion unit, is configured to receive the primary fusion state quantity and comprehensively calculate the current safety status score of the battery, while making short-term predictions of the battery status at future times based on the current state.

[0030] The feedback verification closed-loop module, which is connected between the security situation scoring and prediction model and the dynamic credibility weighted fusion unit, is configured to continuously compare the short-term predicted values ​​and signal excitations made by the security situation scoring and prediction model with the actual measurement values ​​subsequently collected by the acquisition unit, and send the generated prediction deviation information as a feedback signal to the dynamic credibility weighted fusion unit to drive the dynamic credibility weighted fusion unit to adjust its trust weights on the output values ​​of the two engines in the dual-engine state estimation module.

[0031] The system forms an enhanced closed loop of perception, fusion, prediction, verification and correction through a feedback verification closed loop module, and finally outputs a security situation score that tends to be stable after multiple iterations for early warning.

[0032] Further explanation is needed regarding an online sensing system for the safety status of lithium battery boxes throughout their entire lifecycle. The system implements the following: A specific frequency micro-amplitude AC excitation signal is applied to each battery cell via a signal excitation and acquisition unit, which simultaneously acquires voltage and current response data. This unit utilizes a programmable waveform generator and a high-precision acquisition circuit. The frequency range of the excitation signal is dynamically adjusted according to the actual operating state of the battery to obtain effective impedance spectrum information. The data processing and control unit constructs a dual-engine state estimation module. The simplified physics engine employs a reduced-order model based on equivalent circuits and heat transfer principles, directly calculating internal state parameters such as internal resistance and diffusion coefficient by real-time analysis of the excitation response data. The data-driven engine uses a trained convolutional neural network model, inferring the battery's health status by inputting historical voltage, current, and temperature sequences. The outputs of both engines are simultaneously fed into a dynamic reliability weighted fusion unit. This fusion unit assigns dynamic weights to each engine based on the battery's current operating condition and lifespan stage. The weight calculation is based on the estimation error statistics of each engine under similar historical scenarios, thus completing the first fusion and outputting the initial fusion state quantity. The current battery operating condition includes charge / discharge states, and the lifespan stage is based on the number of cycles.

[0033] This state variable is then received by the safety situation scoring and prediction model, which uses a hidden Markov chain structure to synthesize the state variable to calculate a safety score from zero to one and predict changes in battery parameters in the near future. The key to the system is the feedback verification closed-loop module, which continuously compares the predicted value with the subsequent actual measurement value. The resulting prediction deviation is not only used to evaluate the accuracy of the model, but also serves as a feedback signal to adjust the weight allocation in the dynamic credibility weighted fusion unit in real time, forming an enhanced closed loop of "perception-fusion-prediction-verification-recorrection". Through multiple iterative cycles, the system finally outputs a stable and reliable safety situation score and internal state variable, realizing accurate perception of early thermal runaway risk of the battery from the initial stage to the aging process. The entire solution effectively overcomes the problem that static models cannot adapt to the nonlinear aging of batteries through multi-model adaptive fusion and closed-loop self-verification mechanism.

[0034] The simplified physics engine is a reduced-order electrochemical-thermal coupled model that focuses on the online extraction of key internal state parameters of battery internal resistance and lithium-ion diffusion coefficient.

[0035] It should be further explained that the simplified physics engine adopts an electrochemical-thermal coupled model that has undergone rigorous order reduction. This model abandons the complex calculation path of solving full-order partial differential equations in the traditional way. It deeply integrates and simplifies the pseudo-two-dimensional model describing the internal electrochemical dynamics of lithium-ion batteries and the thermal model describing the temperature field distribution through model order reduction. Specifically, the electrochemical part adopts a single-particle model approximation and introduces methods such as polynomial approximation or Pad approximation to simplify the expression of lithium-ion concentration distribution on the surface of solid particles. Thus, the core dynamics of the electrode process are condensed into an algebraic relationship or low-order differential equation about the lithium-ion concentration on the surface of electrode particles and the electrolyte potential. The thermal model is simplified to a lumped parameter model focused on several key temperature measurement points in the battery body.

[0036] This reduced-order model receives micro-amplitude AC excitation signals and their voltage and current response data from the signal excitation and acquisition unit in real time, identifies and quickly outputs key state parameters inside the battery online. The internal resistance parameter is tracked by analyzing the variation trend of the real part of the impedance spectrum at different frequency points, while the lithium-ion diffusion coefficient is dynamically estimated by fitting the characteristic frequency dependence of the Warburg impedance segment related to the diffusion process in the impedance spectrum. Through this targeted model reduction and parameter focusing, this simplified physics engine achieves rapid, online extraction of battery internal state parameters under the limited computing resources of the embedded system, providing a reliable state estimation based on physical mechanisms for subsequent multi-engine fusion.

[0037] It should be further explained that the simplification of the physics engine is achieved through a combination of polynomial approximation and Padre approximation. Specifically, the simplification of the lithium-ion concentration distribution on the electrode particle surface is expressed using a third-order Legendre polynomial approximation, with the expression: c s (r,t)=a0(t)+a1(t)P1(r / R)+a2(t)P2(r / R)+a3(t)P3(r / R); where, c s Let r be the solid-phase lithium-ion concentration, r be the radial coordinate of the particle, R be the particle radius, and P be the solid-phase lithium-ion concentration. n Let a0-a3 be an nth-order Legendre polynomial, where a0-a3 are coefficients that vary with time. The simplified electrolyte potential distribution uses the first-order Pad approximation, with the approximate formula as follows:

[0038] ; where φ e Let x be the electrolyte potential, x be the coordinate along the electrode thickness, and b0, b1, and c1 be dynamic coefficients.

[0039] The initial values ​​of the core parameters of the equivalent circuit corresponding to this engine are set as follows: internal resistance R0 = 50 ± 5 mΩ, charge transfer resistance R... ct =200±20mΩ, double-layer capacitance C dl =50±5μF, the time constant τ corresponding to the diffusion impedance d =10±2s. The above initial value can be finely adjusted proportionally according to the battery's nominal capacity, such as 100Ah or 200Ah. The fine-tuning coefficient is... 150Ah is the base capacity.

[0040] The data-driven engine is a lightweight deep learning network model that takes battery voltage, current, and temperature data streams as input and outputs an estimate of the battery's internal health state. Specifically, the data-driven engine is implemented by constructing a lightweight deep learning network model using a hybrid architecture combining one-dimensional convolutional layers and gated recurrent units (ROUs). The one-dimensional convolutional layers extract local temporal features from the input battery voltage, current, and temperature data streams, with a kernel size of 3, a stride of 1, and two layers. The number of feature maps is controlled at 16 and 32 respectively to limit computational complexity. The subsequent gated recurrent unit layer captures long-term dependencies in the battery state, with 64 hidden units. The network's input is a normalized data sequence containing the most recent 256 sampling points, and the output layer uses a fully connected layer and a sigmoid activation function, ultimately outputting an estimate of the battery's internal health state ranging from zero to one.

[0041] This lightweight network introduces knowledge distillation technology during the model training phase and uses soft labels generated by a complex teacher network as supervision signals. This allows the lightweight student network to maintain high inference accuracy while strictly controlling the number of model parameters to below 500,000, thus ensuring its real-time operation on embedded processors. The data-driven engine effectively complements the shortcomings of the simplified physics engine in modeling uncertainty by directly learning the complex nonlinear mapping between external electrical parameters and internal health states.

[0042] It should be further explained that in the lightweight deep learning network of the data-driven engine, the connection between the one-dimensional convolutional layer and the gated recurrent unit (GRU) is as follows: the feature map output by the first convolutional layer is activated by ReLU, then reduced in dimensionality by a max pooling layer with a stride of 2, and then input into the second convolutional layer. The output of the second convolutional layer is activated by ReLU and flattened into a one-dimensional vector, which is directly connected to the GRU layer. The output of the GRU layer is passed through a Dropout layer with a dropout probability of 0.2, then connected to a fully connected layer, and finally outputs the health status estimate through the output layer. Among them, the first convolutional layer contains 16 3×1 convolutional kernels, the second convolutional layer contains 32 3×1 convolutional kernels, the GRU layer contains 64 hidden units, the fully connected layer contains 32 neurons with ReLU activation, and the output layer contains 1 neuron with Sigmoid activation.

[0043] The network is trained using the Adam optimizer with an initial learning rate of 0.001, decreasing by a factor of 0.8 every 20 epochs. The loss function is the mean squared error (MSE), expressed as follows: In the formula, N is the batch sample size, and y i This represents the true health status value. This is the predicted value; the number of training iterations is set to 150 rounds, and the result is obtained when the loss on the validation set decreases by less than 10 for 10 consecutive rounds. -5 Training is terminated early, and the training data is standardized using the Z-score method, i.e. μ is the mean of the training set, and σ is the standard deviation of the training set. For voltage data, μ = 3.6V and σ = 0.3V; for current data, μ = 0A and σ = 50A; and for temperature data, μ = 25℃ and σ = 8℃.

[0044] The dynamic reliability weighted fusion accelerator assigns weights based on information about the battery's operating conditions, including whether it is in a static, charging, or discharging state, as well as information about the early, middle, or late stages of battery life, categorized by the number of cycles.

[0045] It should be further explained that the specific implementation of the weight allocation of the dynamic credibility weighted fusion engine is based on a multi-factor decision-making mechanism: for the determination of battery operating conditions, the system distinguishes between resting, charging, or discharging states in real time by monitoring the rate of change of battery terminal voltage and the direction and magnitude of current. When the battery is in a high-current charging condition, due to the significant electrochemical polarization effect, the system will assign a relatively high weight to the data-driven engine, because it can better learn nonlinear polarization characteristics from historical data; when the battery is in a resting or low-current condition, the electrochemical process is closer to a quasi-equilibrium state, and the estimation of the simplified physics engine based on first principles is more reliable, so it is given a higher weight.

[0046] For lifespan segmentation, the system divides the battery's lifespan into three stages: early, middle, and late, based on the battery's cumulative equivalent full cycle count. In the early lifespan, the data-driven engine has a low initial weight due to insufficient training data, and the system primarily relies on the simplified physics engine's output. However, as operational data accumulates, the data-driven engine's weight is gradually increased through an online learning module. In the middle lifespan, the weights of the two engines tend to balance. In the late lifespan, due to accelerated battery aging and increased internal state nonlinearity, the system dynamically increases the data-driven engine's weight by incorporating deviation information from the feedback verification loop to better capture abnormal degradation patterns. This weight allocation mechanism is optimized in real-time using a historical error statistics table based on a sliding window, ensuring the fusion strategy's adaptability to the dynamic evolution of the battery throughout its entire lifespan.

[0047] It should be further explained that the credibility metric of the dynamic credibility-weighted fusion engine is achieved through historical error statistics, specifically using a sliding window with a window size W = 50 perception cycles. The root mean square error (RMSE) of the two engines is calculated, and the simplified RMSE expression for the physics engine is as follows:

[0048] ;x phy,k x is the estimated value of the physics engine in the k-th cycle. true,k For the calibration value in the k-th period, the RMSE expression of the data-driven engine is:

[0049] ;

[0050] The formula for assigning credibility weights is:

[0051] ω data =1-ω phy ω phy For the physics engine weights, ω data Weights for the data-driven engine. The threshold range for weight adjustment is set as follows: when |RMSE phy -RMSE data |>0.05x ref x ref For the parameter reference value, such as an internal resistance reference value of 200mΩ, a rapid weight adjustment is triggered, with an adjustment step size of Δω=0.1; when |RMSE phy -RMSE data |≤0.05x ref The weights are maintained or fine-tuned with a step size Δω = 0.02. The initial weight reference values ​​for different operating conditions are: ω for static operating condition. phy0 =0.7、ω data0 =0.3, charging condition, when current I≤0.5C, simplify the initial weight ω of the physics engine. phy0=0.4, Initial weight ω of the data-driven engine data0 =0.6, under discharge conditions, when the current I ≥ -0.5C, the initial weight ω of the simplified physics engine is set to 0.6. phy0 =0.5, Initial weight ω of the data-driven engine data0 =0.5, where C is the current value corresponding to the nominal capacity of the battery.

[0052] The feedback verification closed-loop module feeds back prediction deviation information to the dynamic credibility weighted fusion unit, enabling online adaptive adjustment of the credibility of the dual-engine state estimation module's output results, thereby completing the secondary verification and correction of the primary fused state quantities. Further explanation is needed: the implementation of the feedback verification closed-loop module involves constructing an online adaptive correction mechanism based on the deviation between model predictions and actual measurements. The module continuously receives the safety situation score and the prediction sequence of battery parameters within a specific future time window from the prediction model, while simultaneously receiving the actual measurement sequence acquired by the signal excitation and acquisition unit after that time window. Battery parameters include voltage or temperature. The root mean square error between the predicted and measured sequences is calculated as a quantification deviation index. When this deviation exceeds a preset threshold, a weight adjustment process is triggered. This process first analyzes the temporal distribution characteristics of the deviation. If the deviation is concentrated in a dynamically changing operating condition, it is determined that the transient response capability of the data-driven engine is insufficient, and its weight coefficient in the dynamic credibility weighted fusion unit is reduced accordingly.

[0053] Conversely, if the deviation persists under steady-state conditions, it indicates that the model parameters of the simplified physics engine may drift due to battery aging. In this case, its weight coefficient will be reduced. This directional weight adjustment signal based on deviation characteristics is fed back to the dynamic credibility weighted fusion unit in real time, enabling it to dynamically reconstruct the fusion strategy according to the latest system performance. This achieves continuous secondary verification and online correction of the primary fusion state variables, ultimately forming a perception closed loop that can self-diagnose and self-optimize.

[0054] The frequency components of the micro-amplitude AC excitation signal applied by the signal excitation and acquisition unit are adaptively selected according to the current health status and temperature conditions of the battery.

[0055] It should be further explained that the adaptive selection mechanism of the frequency components of the micro-amplitude AC excitation signal in the signal excitation and acquisition unit is implemented as follows: The system has a built-in reference frequency set covering from low frequency to high frequency, and the initial excitation uses multiple discrete frequency points in this set for scanning measurement; by analyzing the initially acquired impedance spectrum characteristics, especially by observing the integrity of the charge transfer impedance loop and the slope characteristics of the diffusion impedance segment in the Nyquist plot, the system dynamically determines the core frequency point set required for subsequent monitoring; when the battery health status indicates capacity decay or internal resistance increase, the system automatically increases the excitation proportion of low-frequency components to more accurately characterize the diffusion process of lithium ions in the bulk phase of the electrode material; when the temperature sensor detects an increase in battery temperature, the system will correspondingly increase the excitation frequency density in the mid-to-high frequency region to more sensitively capture changes in the charge transfer process at the electrode interface; this mechanism of dynamically optimizing the frequency components of the excitation signal based on the real-time state parameters of the battery ensures that the most representative impedance characteristic information can be obtained under different aging stages and operating conditions, providing a high-quality data foundation for subsequent state estimation.

[0056] The safety situation score calculated by the prediction model is a continuous value ranging from zero to one, which is associated with the battery's thermal runaway risk level.

[0057] It should be further explained that the safety situation scoring and prediction model generates a continuous safety situation score through a probabilistic evaluation framework with multiple feature inputs. The safety situation scoring and prediction model takes the primary fusion state quantity output by the front-end fusion unit as the core input, including but not limited to internal resistance, diffusion coefficient, capacity decay rate and temperature gradient, while also introducing the real-time operating load rate of the battery as an environmental factor.

[0058] The scoring calculation adopts a multivariate anomaly detection algorithm based on Mahalanobis distance. By establishing a joint distribution model of various parameters of the battery under normal conditions, the degree of deviation of the current state from the normal distribution is calculated, and the deviation is mapped to a continuous score value between zero and one. The correlation between the score value and the thermal runaway risk level is realized through a preset dynamic threshold range. When the score value is in a lower range, it corresponds to a normal and safe state. When the score value rises and crosses the first threshold, a warning state is triggered. When the score value rises further and crosses the second threshold, it corresponds to a high-risk alarm state.

[0059] These threshold ranges are not fixed, but are adaptively adjusted according to the battery's lifespan stage. A more lenient threshold is used in the early stages of lifespan, and the threshold standard is gradually tightened as aging progresses.

[0060] It should be further explained that the calculation steps for the Mahalanobis distance in the security situation scoring model are as follows: First, construct the parameter sample set X={x1,x2,...,x} under normal conditions.m}, m≤5000, the sample comes from normal operation data at the beginning of the battery life, each sample x i =[R int D L ,ΔT,R rate ], R int For internal resistance, D L R is the lithium-ion diffusion coefficient, ΔT is the temperature change rate, and R is the lithium-ion diffusion coefficient. rate Calculate the load factor; calculate the covariance matrix σ and mean vector μ of the sample set. X Current state x curr The distance to Maharanobis is: The distance d is mapped to a security situation score using the Sigmoid function, and the mapping formula is as follows: In the formula, α=0.8 is the adjustment coefficient, d0=3 is the baseline distance, and the corresponding score is 0.5.

[0061] The joint distribution of normal state parameters was constructed using a Gaussian Mixture Model (GMM), with 3 components. It was trained iteratively using the EM algorithm for 50 iterations. The dynamic threshold was calculated as follows: in the early stages of the lifespan, when the number of iterations n is less than 500, the warning threshold S... th1 Set to 0.6, high-risk threshold S th2 Set to 0.8, during the mid-life, when the number of cycles n is greater than or equal to 500 and less than 1500, the warning threshold S is... th1 Set to 0.5, high-risk threshold S th2 Set to 0.7, in the later stages of the lifespan, when the number of cycles n is greater than or equal to 1500, the warning threshold S is... th1 Set to 0.4, high-risk threshold S th2 Set to 0.6, where the number of cycles n is calculated based on the equivalent number of full cycles, and the partial cycles are converted based on "charging capacity / nominal capacity".

[0062] This scoring model achieves refined and full-cycle quantitative perception of battery thermal runaway risk through this dynamic threshold mechanism and multi-parameter joint evaluation.

[0063] The data-driven engine receives internal state parameter estimates from the simplified physics engine as auxiliary input features to enhance the accuracy of its mapping relationships. Further explanation is needed: the data-driven engine constructs a cross-engine data fusion input architecture. It not only receives external electrical parameter data streams composed of voltage, current, and temperature, but also incorporates internal state parameter estimates from the simplified physics engine in real-time, including internal resistance and lithium-ion diffusion coefficient, as auxiliary input features. In implementation, these feature vectors from different mechanistic models are first standardized to eliminate dimensional differences before being concatenated into a hybrid input vector. This hybrid vector is then input into a lightweight deep learning network, whose structure adds a fully connected feature layer before the aforementioned convolutional layers, specifically designed to learn the deep cross-correlation between external electrical parameters and internal state parameters.

[0064] Through this architecture, the data-driven engine can utilize both external performance and internal mechanism information. When the simplified physics engine exhibits estimation bias under certain extreme conditions due to model simplification, the data-driven engine can compensate and correct for this by leveraging patterns learned from massive amounts of data. This two-way information interaction mechanism enables the data-driven engine to detect early signs of battery degradation that are difficult to detect using external parameters alone, significantly enhancing the robustness and accuracy of state estimation throughout the entire system's lifecycle.

[0065] The final output of the system is a stable security situation score and internal state variables output by the security situation scoring and prediction model after multiple iterations of enhanced closed loop.

[0066] It should be further explained that the safety situation score and internal state variables are not calculated all at once. Instead, they are confirmed only after multiple iterations of calculation, through continuous adjustment of the output weights of the dynamic reliability weighted fusion unit by the feedback verification closed-loop module within each perception cycle. Specifically, the system sets an iteration termination condition based on the rate of state change. Only when the change in the safety situation score obtained from two consecutive iterations is less than a preset convergence threshold, and the fluctuations of key internal state variables also tend to stabilize, is the system marked as a valid output for that cycle. This multi-iteration convergence process ensures that when faced with sudden changes in battery operating conditions or transient interference from sensor data, the system can filter noise through its internal self-correction mechanism and output reliable and stable evaluation results. The final output safety situation score and internal state variables, as conclusions deeply verified by the system, provide a high-confidence decision-making basis for early warning of battery thermal runaway risks, thereby achieving accurate perception of the battery's safety status throughout its entire life cycle.

[0067] The system also includes a cloud-based data platform that communicates with the data processing and control unit. This platform is used to perform statistical analysis on the operating modes and failure characteristics of the battery group throughout its entire life cycle, and to synchronize the analysis results down to the local safety status score and prediction model of individual batteries, so as to achieve continuous evolution of the model.

[0068] It should be further explained that the cloud data platform constructs a centralized data analysis and knowledge mining system covering multiple battery pack groups. The cloud data platform periodically collects anonymized operational data from various distributed online sensing systems through encrypted communication channels, including historical safety status scores, internal state parameter trajectories, early warning records, and corresponding operating contexts. In the cloud, a distributed computing framework is used to process massive amounts of battery pack data in parallel, clustering analysis methods are used to identify common aging patterns and failure paths, and association rule mining technology is used to establish a quantitative relationship between different operating conditions and battery performance degradation rates.

[0069] The acquired collective knowledge, such as early characteristic patterns of internal resistance mutations under specific operating conditions or critical warning thresholds of diffusion coefficients, is encapsulated as model parameter increments or feature weight adjustment instructions. This incremental information is synchronized to the sensing systems of each local battery pack via a safety downlink and is integrated into their local safety situation scoring and prediction models and data-driven engines in a non-intrusive manner, enabling online updates of local model parameters and fine-tuning of discrimination rules. This continuous evolution mechanism, which combines collective experience with individual practice, allows each independent sensing system to learn from the operational wisdom of the entire battery pack and continuously improve its prediction accuracy and adaptability for early risks throughout its entire life cycle.

[0070] This system effectively overcomes the shortcomings of a single model in terms of adaptability throughout the entire battery lifecycle by constructing a collaborative perception architecture that integrates dual-engine state estimation and dynamic credibility weighted fusion. By introducing a feedback verification closed-loop mechanism, online secondary verification and self-correction of internal state estimation results are achieved, thereby significantly improving the accuracy and reliability of early thermal runaway risk warning and significantly reducing false alarms and missed alarms.

[0071] This system further leverages the collective knowledge sharing and online evolution capabilities of the cloud-based data platform, enabling the local sensing system to continuously absorb the operational experience of the battery group and constantly optimize its own parameters and judgment thresholds. This not only strengthens the system's ability to track the nonlinear aging process of batteries but also achieves a virtuous cycle where the accuracy of safety situational awareness continuously improves over time, providing long-term protection for the full-process safety monitoring of lithium battery boxes.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An online sensing system for the safety status of lithium battery boxes throughout their entire lifecycle, characterized in that, It includes a signal excitation and acquisition unit arranged inside the battery box, and a data processing and control unit electrically connected to the signal excitation and acquisition unit; the signal excitation and acquisition unit is configured to apply a small-amplitude AC excitation signal to a battery cell and acquire its voltage and current response; The data processing and control unit includes: The dual-engine state estimation module has a parallel simplified physical engine and a data-driven engine. The simplified physical engine is configured to extract the internal state parameters of the battery based on the voltage and current response data, and the data-driven engine is configured to learn and map the internal health state of the battery based on the battery's historical and real-time external electrical parameter data. The dynamic credibility weighted fusion unit, which is connected to the dual-engine state estimation module, is configured to receive the preliminary estimates of the same internal state parameter from the simplified physical engine and the data-driven engine, and assign appropriate weights to the preliminary estimates of the two engines according to the current operating condition and life stage of the battery, perform the first fusion calculation, and output the primary fusion state quantity. The safety status scoring and prediction model, which is connected to the dynamic credibility weighted fusion unit, is configured to receive the primary fusion state quantity and comprehensively calculate the current safety status score of the battery, while making short-term predictions of the battery status at future times based on the current status. The feedback verification closed-loop module, connected between the security situation scoring and prediction model and the dynamic credibility weighted fusion unit, is configured to continuously compare the short-term predicted value made by the security situation scoring and prediction model with the signal excitation and the actual measurement value subsequently collected by the acquisition unit, and send the generated prediction deviation information as a feedback signal to the dynamic credibility weighted fusion unit to drive the dynamic credibility weighted fusion unit to adjust its trust weights on the output values ​​of the two engines in the dual-engine state estimation module. The system forms an enhanced closed loop of perception, fusion, prediction, verification and correction through the feedback verification closed loop module, and finally outputs a security situation score that tends to be stable after multiple iterations for early warning.

2. The online sensing system for the safety status of a lithium battery box throughout its entire lifecycle, as described in claim 1, is characterized in that: The simplified physics engine is a reduced-order electrochemical-thermal coupled model that focuses on the online extraction of key internal state parameters of the battery, such as internal resistance and lithium-ion diffusion coefficient.

3. The online sensing system for the safety status of a lithium battery box throughout its entire lifecycle, as described in claim 1, is characterized in that: The data-driven engine is a lightweight deep learning network model that takes as input the battery's voltage, current, and temperature data streams and outputs an estimate of the battery's internal health status.

4. The online sensing system for the safety status of a lithium battery box throughout its entire lifecycle, as described in claim 1, is characterized in that: The dynamic reliability weighted fusion accelerator assigns weights based on information about the battery's operating conditions, including whether it is in a static, charging, or discharging state, as well as information about the early, middle, or late stages of battery life, categorized by the number of cycles.

5. The online sensing system for the safety status of a lithium battery box throughout its entire lifecycle, as described in claim 1, is characterized in that: The feedback verification closed-loop module feeds back the prediction deviation information to the dynamic credibility weighted fusion unit, thereby enabling online adaptive adjustment of the credibility of the output results of the dual-engine state estimation module, and thus completing the secondary verification and correction of the primary fusion state quantity.

6. The online sensing system for the safety status of a lithium battery box throughout its entire lifecycle, as described in claim 1, is characterized in that: The frequency components of the micro-amplitude AC excitation signal applied by the signal excitation and acquisition unit are adaptively selected according to the current health status and temperature conditions of the battery.

7. The online sensing system for the safety status of a lithium battery box throughout its entire lifecycle, as described in claim 1, is characterized in that: The safety situation score and the safety situation score calculated by the prediction model are continuous values ​​ranging from zero to one, which are associated with the thermal runaway risk level of the battery.

8. The online sensing system for the safety status of a lithium battery box throughout its entire lifecycle, as described in claim 1, is characterized in that: The data-driven engine can receive internal state parameter estimates provided by the simplified physics engine as its auxiliary input features. Used to enhance the accuracy of its mapping relationship.

9. The online sensing system for the safety status of a lithium battery box throughout its entire lifecycle, as described in claim 1, is characterized in that: The final output of the system is a stable security situation score and internal state variables output by the security situation scoring and prediction model after multiple iterations of the enhanced closed loop.

10. An online sensing system for the safety status of a lithium battery box throughout its entire lifecycle, according to any one of claims 1 to 9, characterized in that: The system also includes a cloud-based data platform that communicates with the data processing and control unit. This platform is used to perform statistical analysis on the operating modes and failure characteristics of the battery group throughout its entire life cycle, and to synchronize the analysis results down to the local safety status score and prediction model of individual batteries, so as to achieve continuous evolution of the model.

Images