Lithium ion battery early abnormality detection method and system based on dynamic security domain

Abstract

The application belongs to the field of lithium ion battery safety monitoring, and provides a lithium ion battery early abnormality detection method and system based on a dynamic safety domain. The method is based on the synchronous acquisition of the end voltage, current and cell surface temperature of the battery cell to form a package level context vector. The package level context vector is input into a pre-trained causal probability network to respectively output the temperature of the current step and the conditional distribution of the next step increment. The measured voltage increment and temperature increment data are used to perform affine calibration on the output conditional distribution of the next step increment. Based on the calibrated conditional distribution, a cumulative quality rule is used to construct a dynamic safety domain for each time step, and the instantaneous over-domain cost of the actual observation value relative to the dynamic safety domain is calculated. The cumulative hazard potential is obtained by first-order leaky integration accumulation. When the cumulative hazard potential exceeds the preset alarm threshold in consecutive multiple time steps, an early abnormality alarm is triggered. The false negative rate is reduced without increasing the number of sensors.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery safety monitoring, and particularly relates to a method and system for early anomaly detection of lithium-ion batteries based on dynamic safety domains. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] The safety of lithium-ion batteries in automotive and energy storage scenarios largely depends on the ability to issue effective early warnings before thermal runaway occurs. Existing battery management systems typically only collect three types of signals: terminal voltage, current, and surface temperature, and use these to determine the cell's condition. Current early warning and fault diagnosis solutions can be broadly categorized into three technical approaches; The first type is the threshold-based alarm scheme. This type of scheme sets fixed upper and lower limits for multiple monitored quantities, and triggers an alarm when any quantity exceeds the limit. This type of scheme requires additional sensors in terms of hardware, resulting in high deployment costs; the judgment logic is a hard threshold OR combination at the same time, which is sensitive to fluctuations in operating conditions, and the threshold cannot simultaneously take into account the false alarm rate and the missed alarm rate; moreover, it can only be triggered when the fault has developed to the stage of near thermal runaway, making it difficult to provide early warning in the early stages of the fault. The second category is the residual scheme. This type of scheme compares voltage consistency within a group or predicts the nominal response using a model, and then uses the magnitude of the residual as the basis for anomalies. Examples include voltage consistency diagnostic methods based on Shannon entropy or Z-score. This approach is sensitive to residual drift under operating condition modulation; the window length and data length significantly affect the judgment result. The residual only stabilizes and exceeds the threshold after the fault has developed to a certain extent, resulting in a relatively late warning time. Essentially, it still relies on the immediate determination of a hard threshold for the residual scalar.

[0004] The third category is end-to-end deep learning solutions. These solutions use neural networks to directly map operational data to a risk scalar, and then perform online inference after training on large sample data.

[0005] The above solutions share the following common drawbacks: the judgment criteria are all based on single-moment hard thresholds, and sensitivity and noise resistance are mutually constrained; the thresholds or residuals are sensitive to operating conditions and systems, and cross-system deployment requires readjustment; they rely on additional sensors or a large number of labeled samples, which increases costs and deployment difficulty; most solutions can only alarm when the fault develops to the stage close to thermal runaway, and do not have early warning capabilities. Summary of the Invention

[0006] To address at least one of the technical problems mentioned above, this invention provides a method and system for early anomaly detection of lithium-ion batteries based on dynamic safety domains. This method relies solely on the existing terminal voltage, current, and surface temperature signals of the battery management system to overcome the problems of condition sensitivity, false alarms and missed alarms, and delayed warning timing caused by the "single-moment-hard threshold" judgment in existing solutions.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A first aspect of the present invention provides a method for detecting early anomalies in lithium-ion batteries based on dynamic safety domains, comprising the following steps: Based on the synchronously acquired terminal voltage, current, and surface temperature of the battery cell, a packet-level context vector is formed. The packet-level context vector is input into a pre-trained causal probability network, which outputs the conditional distributions of the current step temperature, the next step voltage increment, and the next step temperature increment, respectively. Within the power-on warm-up window, the conditional distribution of the next voltage increment and the next temperature increment is affinely calibrated using the measured voltage increment and temperature increment data to obtain the calibrated conditional distribution. Based on the calibrated conditional distribution, a dynamic safety domain is constructed for each time step using the cumulative quality rule, and the instantaneous out-of-domain cost of the actual observation relative to the dynamic safety domain is calculated. Based on the instantaneous out-of-range costs of the voltage and temperature channels, an instantaneous danger rate is synthesized, and an accumulated danger potential is obtained by accumulating through a first-order leakage integral. When the accumulated danger potential exceeds a preset alarm threshold in multiple consecutive time steps, an early abnormal alarm for the battery cell is triggered.

[0008] Furthermore, the causal probability network uses a multi-layer Mamba selective state space block as its backbone and connects three hybrid density network heads in parallel at the end. Each hybrid density network head parameterizes the prediction density with K Gaussian components and outputs the conditional distributions of the current step temperature, the next step voltage increment, and the next step temperature increment, respectively.

[0009] Furthermore, during the training of the causal probabilistic network, the training objective is the joint negative log-likelihood over all normally functioning samples, with the addition of minimum variance regularization to suppress the collapse of mixed components.

[0010] Furthermore, affine calibration specifically includes: The bias and scaling of the battery cell are calculated by using the mean and variance of the measured voltage and temperature increments within the warm-up window, as well as the marginal mean and marginal variance predicted by the causal probability network in the same window. The bias and scaling of the battery cell are applied to the mixed component parameters of the causal probability network output, and the calibrated conditional distribution is obtained.

[0011] Furthermore, the dynamic security domain is represented as: , , in, Represents a dynamic security domain. Indicates the signals in each security domain. This represents the calibrated conditional density. This represents the density threshold determined by the cumulative mass rule. This represents the lower bound threshold of the probability density of the safe region. This is the lower limit of absolute density. The cumulative quality ratio means that only signals with probability densities higher than the absolute threshold and whose probability densities rank among the highest among all potential signals at any given moment are considered to be in the safe region.

[0012] Furthermore, the instantaneous cross-domain cost is expressed as: , in, Indicates the cost of instantaneous cross-domain access. Represents the actual observed value at time t. Represents a dynamic security domain. This represents the density threshold determined by the cumulative mass rule. This indicates the conditional density at the measured observation point x after calibration. t The value at that location, This indicates the lower limit of absolute density.

[0013] Furthermore, the accumulated danger... Due to voltage channel cost Instantaneous risk rate is synthesized by adding and coupling the temperature channel cost. Then, the result is obtained by first-order leakage integral.

[0014] A second aspect of the present invention provides an early anomaly detection system for lithium-ion batteries based on dynamic safety domains, comprising: The vector construction module is used to form a packet-level context vector based on the synchronously acquired terminal voltage, current and surface temperature of the battery cell; The conditional distribution output module is used to input the packet-level context vector into a pre-trained causal probability network and output the conditional distributions of the current step temperature, the next step voltage increment, and the next step temperature increment, respectively. The affine calibration module is used to perform affine calibration on the conditional distribution of the next voltage increment and the next temperature increment in the output during the power-on warm-up window, using the measured voltage increment and temperature increment data, and obtain the calibrated conditional distribution. The dynamic safety domain construction module is used to construct the dynamic safety domain for each time step based on the calibrated conditional distribution and the cumulative quality rule, and to calculate the instantaneous out-of-domain cost of the actual observation relative to the dynamic safety domain. The anomaly detection module is used to synthesize the instantaneous danger rate based on the instantaneous out-of-range cost of the voltage channel and temperature channel, and to obtain the cumulative danger potential through the first-order leakage integral. When the cumulative danger potential exceeds the preset alarm threshold in multiple consecutive time steps, the early anomaly alarm of the cell is triggered.

[0015] A third aspect of the present invention provides a computer-readable storage medium.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for early anomaly detection of lithium-ion batteries based on dynamic security domains.

[0017] A fourth aspect of the present invention provides a computer device.

[0018] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method for early anomaly detection of lithium-ion batteries based on dynamic safety domains.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This invention relies solely on the existing terminal voltage, current, and surface temperature signals of the battery management system. It employs a battery anomaly detection method that is adaptive to operating conditions, robust to transient interference, and capable of continuously accumulating evidence and providing stable alarms in the early stages of a fault. Instead of the "single moment + hard threshold" criterion, it uses "dynamic safety domain + continuous evidence accumulation" to reduce both false alarm rate and missed alarm rate without adding additional sensors.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a flowchart of the early anomaly detection method for lithium-ion batteries based on dynamic safety domains provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the overall framework of the early anomaly detection method for lithium-ion batteries based on dynamic safety domains provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the structure of the Causal Probability Network (CPN) provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the construction and time-varying nature of the Dynamic Security Domain (DSD) provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of leakage hazard potential integration and alarm triggering provided in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] Example 1 like Figures 1-2 As shown, this embodiment provides a method for early anomaly detection in lithium-ion batteries based on dynamic safety domains, including the following steps: Step 1: Based on the synchronously acquired terminal voltage, current, and surface temperature of the battery cell, construct a packet-level context vector; This embodiment uses an 8-cell blade lithium iron phosphate battery module as the carrier, with a sampling period of 1 second. The battery management system synchronously collects the terminal voltage, module current, and cell surface temperature of each cell within the module at 1Hz, forming a package-level context vector. .

[0027] Step 2: Input the packet-level context vector into the pre-trained causal probability network, and output the conditional distributions of the current temperature T, the next voltage increment ΔV, and the next temperature increment ΔT respectively; like Figure 3 The diagram shows a Causal Probabilistic Network (CPN), which uses a multi-layer Mamba selective state space block as its backbone and three hybrid density network heads connected in parallel at the end. These heads output the conditional distributions of the current step temperature T, the next step voltage increment ΔV, and the next step temperature increment ΔT, respectively.

[0028] Specifically, a single-layer Mamba selective state-space block outputs the output vector at time t, and multiple Mamba blocks are stacked layer by layer along the time dimension, with the previous block outputting the vector. As the next input, it continues to perform causal convolution, selective state update and gated residual; the output of the last Mamba selective state space block is then fed in parallel to the three mixed density network heads, which respectively parameterize the conditional distribution p(·|c_t) of the current step temperature T, the next step voltage increment ΔV and the next step temperature increment ΔT. Specifically, this includes the following: First, the packet-level context vector... Input projection : (1), in, Represents the projected features. Indicates the input weights. Indicates input bias; based on The backbone Mamba blocks perform causal convolution, selective state updates, and gated residuals in the time dimension. The expression for each block is: (2), (3), in, This represents the output vector at time t. This represents the hidden state of the network at time t. Represents the state decay coefficient. This represents the hidden state of the network at time t-1. This represents the projection parameters from the input features to the output. Represents a linear transformation function. This represents the swish activation function. The gating function is represented; each hybrid density network head is parameterized with K Gaussian components to predict the density, in voltage increments. Taking a channel as an example, its conditional density is expressed as: (4), in, This represents the conditional probability density of the next voltage increment ΔV under the packet-level context. , , These represent the weight, mean, and standard deviation of the k-th component, respectively. This indicates the number of Gaussian components in the mixed distribution; a value of 3 is recommended. Represents the Gaussian distribution function; Channels for temperature increment ΔT and voltage increment The channels have the same structure, and the mean and standard deviation of the temperature marginal distribution are spliced ​​into the two incremental channels as thermal context conditions, so that the incremental prediction adapts to thermal evolution.

[0029] As a further implementation, during the training of the causal probabilistic network, the training objective is the joint negative log-likelihood over all normally functioning samples, with the addition of a minimum variance regularization to suppress the collapse of the mixed components, expressed as: (5), in, The training objective (loss function) is the sum of the joint negative log-likelihood and minimum variance regularization for the voltage and temperature channels of all normally operating samples. This indicates the voltage increment channel's response to the measured voltage increment under specific context conditions. Conditional density, This represents the conditional density of the temperature increment channel with respect to the measured temperature increment ΔTt under context conditions, where λ is the coefficient (weight) of the minimum variance regularization term, used to suppress the collapse of the mixed components.

[0030] By freezing the parameters of the causal probability network after one-time offline training, it can be deployed to the BMS main control unit on the vehicle / energy storage side, and online inference only requires forward computation.

[0031] Step 3: In the power-on warm-up window, use the measured voltage and temperature increment data to perform affine calibration on the conditional distribution of the next voltage and temperature increment of the output, and obtain the calibrated conditional distribution. In this embodiment, the specific process of affine calibration includes: CPN learns the overall conditional distribution from normal data of the same type of battery, but cannot absorb the subtle biases and SOH differences between individual cells. Therefore, this embodiment sets a power-on warm-up window with a length of K samples (default K=300, i.e., 300 s), and the measured increments within this window are regarded as normal states.

[0032] For cell i, the mean of the measured increments ΔV and ΔT within the warm-up window is used. Measured variance Marginal mean predicted by CPN within the same window Marginal variance Solve the bias of the battery cell. With scaling , is represented as: (6), (7), in, Indicates the lower bound of the scaling factor (default is 0.3).

[0033] The upper bound of the scaling is 1.0, and the lower bound is... (Default 0.3) ensures that calibration only allows tightening of the security domain and does not allow excessive expansion.

[0034] Then, applying it to the mixed component parameters of the CPN output yields the calibrated conditional distribution, expressed as: (8), (9), in, This represents the mean of the k-th Gaussian component after calibration. This represents the mean of the k-th Gaussian component before calibration (i.e., directly output by CPN). This represents the standard deviation of the k-th Gaussian component after calibration. This represents the standard deviation of the k-th Gaussian component before calibration (i.e., direct output of CPN).

[0035] The calibration coefficients are fixed at the end of the warm-up window and will not be adjusted during the running period, eliminating the need for online retraining.

[0036] Step 4: Based on the calibrated conditional density, construct the dynamic safety region for each time step using the cumulative quality rule, and calculate the instantaneous out-of-domain cost of the actual observation relative to the dynamic safety region; In this embodiment, the calibrated conditional density is denoted as This method no longer uses static threshold upper and lower limits, but instead directly generates the variable safety region D from the conditional density using cumulative quality rules. t ,like Figure 4 As shown; Security Domain D t Defined jointly by equations (10) and (11): (10) (11), in, This indicates the possible values ​​of the state variable (voltage increment or temperature increment). Indicates the candidate density threshold. This represents the density threshold determined according to the cumulative mass rule, i.e., the maximum density level that satisfies the requirement that the cumulative probability mass of a high-density region is not less than τ. The actual density threshold of the safe region is represented by the larger of the dynamic threshold pτ(t) and the absolute lower limit β. This is the cumulative mass ratio (default 0.80). This is the lower limit of absolute density, used to prevent the safety region from degenerating into an excessively wide region when sudden changes in operating conditions lead to excessively wide conditional densities. Safety region D tThe density gradually tightens or loosens according to the conditions, thus naturally adapting to the working conditions.

[0037] The corresponding instantaneous out-of-domain cost is given by equation (12), with one portion taken from each of the voltage and temperature channels: (12), in, This indicates the calibrated conditional density at the measured observation point. The value at that point represents the predicted probability density corresponding to the measured state.

[0038] The safety domain is generated online by conditional density according to cumulative mass rules and can adaptively scale according to operating conditions; the warm-up affine calibration absorbs individual differences in one go, and no online training is required during the deployment period; The logarithmic form guarantees that: the cost is zero within the safe region; the cost is finite and does not diverge during accidental single-step out-of-domain outages; and the cost continues to increase during deep out-of-domain outages, thus preserving sensitivity to depth anomalies.

[0039] Step 5: Based on the instantaneous out-of-range costs of the voltage channel and temperature channel, synthesize the instantaneous danger rate and accumulate it through the first-order leakage integral to obtain the cumulative danger potential; when the cumulative danger potential exceeds the preset alarm threshold in multiple consecutive time steps, trigger the early abnormal alarm of the cell.

[0040] Voltage channel cost Temperature channel cost Instantaneous hazard rate is synthesized through addition and coupled product. Then, the cumulative danger potential is obtained by integrating the first-order leakage. And trigger an alarm based on a continuous threshold condition (see Figure 5 ).

[0041] (13) (14) (15) In formula (13) The coefficient is the channel coupling coefficient; the additive term retains the individual sensitivity of each channel, while the product term amplifies the coordinated deviation of the two channels, thus contributing only linearly to occasional spikes in a single channel, while generating superlinear growth for the common drift of both channels. In equation (14) The leakage coefficient is used to ensure that transient disturbances decay over time and do not accumulate indefinitely. In equation (15) M is the alarm threshold and M is the continuous step threshold. Together, they suppress false alarms caused by occasional spikes. When a cell satisfies equation (15), an early abnormal alarm is triggered for that cell, and the cell number can be directly located.

[0042] Dual-channel addition combined with coupled product balances single-channel sensitivity and cross-channel collaborative response; leakage integral and continuous over-threshold dual thresholds suppress transient disturbances from the time dimension. Example 2 This embodiment uses an 8-cell blade lithium iron phosphate battery module as the carrier, with a sampling period of 1 second. The CPN backbone consists of 4 layers of Mamba blocks with a hidden width of 128; each of the three hybrid density heads uses K=3 Gaussian components. The specific implementation steps are as follows: S1: The BMS synchronously collects the terminal voltage, module current, and cell surface temperature of each cell in the module at 1 Hz, forming a packet-level context vector c. t ; S2: According to formula (1)–(4), c t Input the pre-trained CPN, and output the marginal distribution of temperature T and the conditional distributions of ΔV and ΔT at each step. Before network deployment, train it on normal data of the same type of battery according to Equation (5), take the AdamW optimizer, and freeze the parameters after deployment.

[0043] S3: The first 300 seconds after power-on are considered a warm-up window. Calculate Δμ for each cell using equations (6)-(9). i s i And it is applied to the mixing component parameters, taking s min = 0.3, fixed during operation.

[0044] S4: For the calibrated ΔV and ΔT conditional densities, generate the safety region according to equations (10) and (11) with τ = 0.80 and β being 5% of the peak value of the conditional density. , Calculate the instantaneous cross-domain cost according to formula (9). , .

[0045] S5: Synthesize the instantaneous hazard rate h according to formula (13) t (Take ρ = 1.0), and use Equation (14) to perform leakage integration with γ = 0.05 to obtain H. t According to formula (15) with H th = 6.0, M = 3 triggers an alarm. When an alarm occurs, the cell number is reported and original data for 60 seconds before and after the trigger time is retained.

[0046] Example 3 Under the overall process of Example 2, this method is deployed on battery fault data covering four systems and morphologies: blade LFP, cylindrical 18650 LFP, blade ternary NCM, and square ternary NCM. CPN training and operation points (τ, β, ρ, H) are completed for each system. thThe selection of the (M) operation point maximizes the detection rate and minimizes the false alarm rate on the system's leave validation set. Fault injection types include poor contact short circuit (PCF) and internal short circuit (ISCF).

[0047] On a strictly reserved test set (32 runs, 256 cell-level judgments), this method achieved a detection rate of 0.906, a false alarm rate of 0.066, a precision of 0.659, and an F1 score of 0.770 (based on system macro-average). Compared with four representative baselines (temperature residual proxy, temperature-voltage residual proxy, intra-group voltage Z-score detection, and corrected sample entropy), this method has a detection rate 0.16 higher, a false alarm rate 0.09 lower, and an F1 score 0.25 higher, while also occupying the upper left of the detection rate-false alarm rate tradeoff.

[0048] Example 4 This embodiment further examines the method's ability to identify degraded battery cells after abuse. Six strings of pouch lithium cobalt oxide modules were used, and the C1 cell was first subjected to an offline 1C, 120% deep charge cycle overcharge to create a damaged state. After being reinstalled into the module, it was operated under UDDS conditions. The entire set of steps S1–S5 is consistent with that in Embodiment 2.

[0049] The results showed that the original voltage and temperature curves throughout the UDDS process could not distinguish C1 from the five normal cells; while the cumulative dangerous potential output by this method... The overcharge rate of this cell is significantly higher than that of the other five sibling cells, and it stably triggers an alarm in the middle of the UDDS (Underlying Defibrillation Discharge) stage. This example demonstrates that the method still has early detection capability under conditions of mild abuse, where internal degradation has already occurred but thermal runaway has not yet occurred.

[0050] Example 5 This embodiment provides a variant of vehicle-side embedded deployment: the trained CPN backbone is embedded into the BMS main control MCU using INT8 quantization. S3 is automatically executed for the first 5 minutes after each vehicle power-on or battery swap, while S4 and S5 run continuously online during vehicle operation. When equation (12) is true and triggers an alarm, the main control automatically saves the original V, I, and T sequences and corresponding values ​​for 60 seconds before and after the trigger time. The sequence is provided for cloud verification and offline analysis.

[0051] The cumulative evidence-based alarm of this method can be deployed in parallel with the existing hard threshold protection: the hard threshold is used for the forced interception of irreversible dangers in the tail segment, while this method is responsible for the early graded warning before the tail segment, and the two do not conflict with each other.

[0052] Example 6 This embodiment provides an early anomaly detection system for lithium-ion batteries based on dynamic safety domains, including: The vector construction module is used to form a packet-level context vector based on the synchronously acquired terminal voltage, current and surface temperature of the battery cell; The conditional distribution output module is used to input the packet-level context vector into a pre-trained causal probability network and output the conditional distributions of the current step temperature, the next step voltage increment, and the next step temperature increment, respectively. The affine calibration module is used to perform affine calibration on the conditional distribution of the next voltage increment and the next temperature increment in the output during the power-on warm-up window, using the measured voltage increment and temperature increment data, and obtain the calibrated conditional distribution. The dynamic safety domain construction module is used to construct the dynamic safety domain for each time step based on the calibrated conditional density and the cumulative quality rule, and to calculate the instantaneous cross-domain cost of the actual observation relative to the dynamic safety domain. The anomaly detection module is used to synthesize the instantaneous danger rate based on the instantaneous out-of-range cost of the voltage channel and temperature channel, and to obtain the cumulative danger potential through the first-order leakage integral. When the cumulative danger potential exceeds the preset alarm threshold in multiple consecutive time steps, the early anomaly alarm of the cell is triggered.

[0053] It should be noted that the specific implementation of the lithium-ion battery early anomaly detection system based on dynamic safety domain in this embodiment of the invention is similar to the specific implementation of the lithium-ion battery early anomaly detection method based on dynamic safety domain in this embodiment of the invention. For details, please refer to the description in the method section. To reduce redundancy, it will not be repeated here.

[0054] Example 7 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the above-described method for early anomaly detection of lithium-ion batteries based on dynamic security domains.

[0055] Example 8 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the above-described method for early anomaly detection of lithium-ion batteries based on dynamic security domains.

[0056] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0057] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0058] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0059] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0060] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for early anomaly detection in lithium-ion batteries based on dynamic safety domains, characterized in that, Includes the following steps: Based on the synchronously acquired terminal voltage, current, and surface temperature of the battery cell, a packet-level context vector is formed. The packet-level context vector is input into a pre-trained causal probability network, which outputs the conditional distributions of the current step temperature, the next step voltage increment, and the next step temperature increment, respectively. Within the power-on warm-up window, the conditional distribution of the next voltage increment and the next temperature increment is affinely calibrated using the measured voltage increment and temperature increment data to obtain the calibrated conditional distribution. Based on the calibrated conditional distribution, a dynamic safety domain is constructed for each time step using the cumulative quality rule, and the instantaneous out-of-domain cost of the actual observation relative to the dynamic safety domain is calculated. Based on the instantaneous out-of-range costs of the voltage and temperature channels, an instantaneous danger rate is synthesized, and an accumulated danger potential is obtained by accumulating through a first-order leakage integral. When the accumulated danger potential exceeds a preset alarm threshold in multiple consecutive time steps, an early abnormal alarm for the cell is triggered. The instantaneous cross-domain cost is expressed as: , in, Indicates the cost of instantaneous cross-domain access. Represents the actual observed value at time t. Represents a dynamic security domain. This represents the density threshold determined by the cumulative mass rule. This indicates the conditional density at the measured observation point after calibration. The value at that location, Indicates the lower limit of absolute density; Among them, the cumulative danger potential Due to voltage channel cost With temperature channel The instantaneous risk rate is synthesized by addition and coupling product of costs. Then, the result is obtained by first-order leakage integral.

2. The method for early anomaly detection of lithium-ion batteries based on dynamic safety domain as described in claim 1, characterized in that, The causal probability network uses a multi-layer Mamba selective state space block as its backbone and three hybrid density network heads connected in parallel at the end. Each hybrid density network head is parameterized with K Gaussian components to predict the density and outputs the conditional distributions of the current step temperature, the next step voltage increment, and the next step temperature increment, respectively.

3. The method for early anomaly detection of lithium-ion batteries based on dynamic safety domain as described in claim 1, characterized in that, When training a causal probabilistic network, the training objective is the joint negative log-likelihood over all normally functioning samples, with the addition of a minimum variance regularization to suppress the collapse of mixed components.

4. The method for early anomaly detection of lithium-ion batteries based on dynamic safety domain as described in claim 1, characterized in that, Affine calibration specifically includes: The bias and scaling of the battery cell are calculated by using the mean and variance of the measured voltage and temperature increments within the warm-up window, as well as the marginal mean and marginal variance predicted by the causal probability network in the same window. The bias and scaling of the battery cell are applied to the mixed component parameters of the causal probability network output, and the calibrated conditional distribution is obtained.

5. The method for early anomaly detection of lithium-ion batteries based on dynamic safety domain as described in claim 1, characterized in that, Dynamic security domains are represented as: ， ， in, Represents a dynamic security domain. This represents the values ​​that the state variable can take. This represents the calibrated conditional density. This represents the density threshold determined by the cumulative mass rule. Indicates the candidate density threshold. This is the lower limit of absolute density. This represents the cumulative mass ratio.

6. A lithium-ion battery early anomaly detection system based on dynamic safety domain, characterized in that, The method for detecting early anomalies in lithium-ion batteries based on dynamic safety domains as described in any one of claims 1-5 includes: The vector construction module is used to form a packet-level context vector based on the synchronously acquired terminal voltage, current and surface temperature of the battery cell; The conditional distribution output module is used to input the packet-level context vector into a pre-trained causal probability network and output the conditional distributions of the current step temperature, the next step voltage increment, and the next step temperature increment, respectively. The affine calibration module is used to perform affine calibration on the conditional distribution of the next voltage increment and the next temperature increment in the output during the power-on warm-up window, using the measured voltage increment and temperature increment data, and obtain the calibrated conditional distribution. The dynamic safety domain construction module is used to construct the dynamic safety domain for each time step based on the calibrated conditional distribution and the cumulative quality rule, and to calculate the instantaneous out-of-domain cost of the actual observation relative to the dynamic safety domain. The anomaly detection module is used to synthesize the instantaneous danger rate based on the instantaneous out-of-range cost of the voltage channel and temperature channel, and to obtain the cumulative danger potential through the first-order leakage integral. When the cumulative danger potential exceeds the preset alarm threshold in multiple consecutive time steps, the early anomaly alarm of the cell is triggered.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the method for early anomaly detection of lithium-ion batteries based on dynamic safety domains as described in any one of claims 1-5.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the method for early anomaly detection of lithium-ion batteries based on dynamic safety domains as described in any one of claims 1-5.

Images