A battery thermal runaway early warning method fusing time series prediction and physical residual check

Abstract

The application belongs to the technical field of battery safety monitoring, and particularly relates to a battery thermal runaway early warning method fusing time sequence prediction and physical residual checking, which comprises the following steps: collecting battery voltage and temperature time sequence data, and performing sliding window slicing and standardization preprocessing; constructing a time convolution network prediction main chain, training only with normal working condition data, introducing domain embedding to fuse working condition information, predicting future voltage and temperature values, calculating prediction residual and abnormal score, and generating a formal trigger signal by using threefold gating of an abnormal score threshold, a duration threshold and a cumulative and statistical quantity threshold; constructing a physical residual checking branch, extracting abnormal features, and obtaining a comprehensive risk score by weighted fusion; fusing the formal trigger signal and the comprehensive risk score, mapping the fused signal into a multi-level risk grade by persistence confirmation, and outputting the signal; and offline calculating early warning advance with a thermal runaway reference point as an anchor point. The application realizes online early warning with high advance, low false alarm rate and strong interpretability.

Description

Technical Field

[0001] This invention relates to the field of battery safety monitoring technology, and in particular to a battery thermal runaway early warning method that integrates timing prediction and physical residual verification. Background Technology

[0002] Lithium-ion batteries are widely used in energy storage power stations, electric vehicles, and portable electronic devices due to their high energy density and long cycle life. However, under abnormal operating conditions such as overcharging, external heating, and internal short circuits, a series of chain exothermic reactions occur inside the battery, ultimately leading to thermal runaway. Typically, thermal runaway begins with the decomposition of the solid electrolyte interfacial film at 100 to 120 degrees Celsius, exposing the negative electrode active material and releasing heat. Subsequently, when the temperature rises to 130 to 160 degrees Celsius, the separator melts and shrinks, triggering an internal short circuit and generating a large amount of Joule heat. Further, in the range of 180 to 220 degrees Celsius, the positive electrode material decomposes, releasing oxygen and undergoing a violent exothermic reaction with the electrolyte, ultimately leading to fire or even explosion. Once thermal runaway enters the self-accelerating phase, traditional thermal management or power-off measures are difficult to effectively suppress it. Therefore, early warning before irreversible thermal runaway occurs is of great significance for improving battery system safety.

[0003] Existing lithium-ion battery thermal runaway early warning methods can be mainly divided into four categories. The first category is threshold-based methods, which trigger alarms by setting fixed temperature, voltage, or differential pressure thresholds. While simple to implement, these methods struggle to adapt to different battery types, operating conditions, and aging states. They lack sensitivity to subtle anomalies in the early, slow evolution of thermal runaway, typically only triggering alarms when the anomaly is already significant, resulting in limited lead time. The second category is purely data-driven methods, utilizing deep learning models such as Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs) to learn the temporal evolution of thermal runaway. Anomaly detection is achieved by predicting temperature and voltage. While these methods can capture early trend changes, real-world thermal runaway data is scarce, the models are heavily reliant on training data, have limited cross-condition generalization ability, and lack physical interpretability. Most methods only output a single anomaly result and lack multi-level risk assessment mechanisms. The third category is purely physical model methods, which establish thermal runaway models based on electrochemical-thermal coupling mechanisms. These methods offer high theoretical accuracy but suffer from complex parameter calibration, high computational load, and poor real-time performance, making them difficult to deploy directly in online monitoring systems. The fourth category is the data-physics joint method, which attempts to integrate data-driven models with physical feature analysis. However, most methods only focus on temperature anomalies and ignore the important indicative role of voltage signals in the early stages of thermal runaway. The physical feature extraction dimensions are limited, and key precursor features such as residual growth, temperature rise rate, temperature rise acceleration, and temperature-voltage coupling anomalies are not fully utilized. Furthermore, there is a lack of continuous confirmation mechanisms, which can easily lead to the misjudgment of instantaneous fluctuations as anomalies. Summary of the Invention

[0004] This invention proposes a battery thermal runaway early warning method that integrates timing prediction and physical residual verification, aiming to solve the technical problems of insufficient early warning time, high false alarm rate and poor physical interpretability in the existing technology.

[0005] This invention provides a battery thermal runaway early warning method that integrates time-series prediction and physical residual verification, comprising: Collect timing data of battery operation, including voltage and temperature; A time series prediction main chain is constructed. The main chain is trained with voltage and temperature data under normal operating conditions. It is used to predict future time series data based on current time series data and generate anomaly scores based on prediction residuals. The main chain generates formal trigger signals through multiple gating, including anomaly score threshold, duration threshold, and cumulative and statistical thresholds. A physical residual verification branch is constructed, which is used to extract physical features from the predicted residual. The physical features include at least temperature residual and temperature-voltage coupling anomaly features. The physical features are converted into sub-fractions through a piecewise mapping function based on statistical quantiles, and then weighted and fused to generate a comprehensive risk score. The temperature-voltage coupling anomaly features are determined to be valid when the voltage anomaly degree reaches a first threshold and the temperature rise rate reaches a second threshold. By integrating the formal trigger signal with the comprehensive risk score, a multi-level risk level is output. When the historical data is insufficient for the sliding window length, the system enters a preheating state and does not output a valid risk level.

[0006] Furthermore, the time-series prediction main chain adopts a temporal convolutional network, which extracts temporal features through causal convolution and dilated convolution; the temporal convolutional network introduces a domain embedding layer, which maps the working condition identifier into a continuous vector and concatenates it with the temporal features before inputting it into the network. The working condition identifier includes at least a dataset identifier, a charging / discharging stage identifier, and a discharge depth.

[0007] Furthermore, the outlier score is obtained by weighted fusion of the residual score and the cumulative sum statistic of the residuals. The residual score is the absolute value of the predicted residual divided by the baseline standard deviation of the residual under normal operating conditions. The cumulative sum statistic is used to detect the persistent deviation trend of the residuals. Its recursive method is to take the cumulative sum statistic at the current time as the larger value between the cumulative sum statistic at the previous time, the current residual minus the mean of the residual under normal operating conditions, and zero.

[0008] Furthermore, the thresholds for the multiple gating are: anomaly score threshold of 5.0, duration threshold of 20 seconds, and cumulative sum statistic threshold of 15.0. The formal trigger signal is enabled when all three conditions are met simultaneously.

[0009] Furthermore, the physical characteristics also include residual growth and temperature rise acceleration; the residual growth is the difference between the temperature residual at the current moment and the temperature residual at the previous moment divided by the sampling interval; the temperature rise rate is the difference between the temperature at the current moment and the temperature at the previous moment divided by the sampling interval; the temperature rise acceleration is the difference between the temperature rise rate at the current moment and the temperature rise rate at the previous moment divided by the sampling interval; and the voltage anomaly is the absolute value of the voltage residual divided by the baseline standard deviation of the voltage residual under normal operating conditions.

[0010] Furthermore, after converting the physical features into sub-scores using a piecewise mapping function based on statistical quantiles, a weighted fusion is performed to generate a comprehensive risk score, specifically including: Each sub-fraction Based on the corresponding physical characteristics through the mapping function get: ; In the formula, For corresponding physical characteristics The maximum value within the sliding window, i.e. ; Sub-fraction vector ; In the formula, The fractional value of the temperature residual; The residual growth fraction; The fraction of the rate of temperature rise; The fraction of temperature rise acceleration; For voltage anomaly degree sub-fraction; The voltage-temperature coupling anomaly sub-fraction; Comprehensive risk score ; In the formula, For the corresponding physical characteristic scores; These are the weight coefficients for the corresponding features, and they satisfy the following constraints: .

[0011] Furthermore, when generating a comprehensive risk score by weighted fusion of physical features, the weights of each physical feature score are configured as follows: temperature residual score weight 0.20, residual growth score weight 0.15, temperature rise rate score weight 0.25, temperature rise acceleration score weight 0.15, voltage anomaly score weight 0.15, and temperature-voltage coupling anomaly indication value weight 0.10.

[0012] Furthermore, the first threshold and the second threshold are determined based on the 99th percentile value of the corresponding feature under normal operating conditions.

[0013] Furthermore, the process of integrating the formal trigger signal and the comprehensive risk score to output a multi-level risk rating specifically includes: performing continuous confirmation, whereby the comprehensive risk score is continuously higher than a preset warning threshold. Furthermore, if the duration reaches a preset value, and at least two physical feature scores exceed their respective attention thresholds or a temperature-voltage coupling anomaly is established, the final risk level is determined based on the continuous confirmation results and the validity of the formal trigger signal.

[0014] Furthermore, the multi-level risk assessment includes: Level 0 represents normal, Level 1 represents attention, Level 2 represents warning, and Level 3 represents high-priority alarm; among them, Level 1 is defined as a comprehensive risk score reaching the attention threshold. Triggered, Level 2: The overall risk score reaches the warning threshold. Furthermore, the continuous confirmation of the trigger is required, and Level 3 is triggered by the validity of the formal trigger signal and the continuous confirmation of the trigger. It also includes an offline verification step: using the thermal runaway reference time as the evaluation anchor point, the time difference between the time point when the warning level is first reached and the evaluation anchor point is calculated as the warning advance.

[0015] The technical effects of this invention are: 1. A time-series prediction main chain based on a temporal convolutional network is constructed. This network is trained only on voltage and temperature time-series data under normal operating conditions. Temporal features are extracted through causal convolution and dilated convolution. A domain embedding mechanism is introduced to map the operating condition identifier into a continuous vector and concatenate it with the temporal features before inputting it into the network. The effects are: the model can accurately learn the temporal evolution patterns under normal operating conditions and generate sensitive prediction residual responses to minor abnormal changes during the incubation period of thermal runaway; simultaneously, the domain embedding layer automatically encodes the data distribution offsets of different charging and discharging stages and discharge depths through end-to-end training, significantly improving the model's cross-operating condition generalization ability. This solves the problems of poor adaptability of traditional threshold methods and weak cross-operating condition generalization ability of pure data-driven methods, thereby increasing the early warning lead time and reducing the false negative rate.

[0016] 2. A physical residual verification branch is constructed, independently extracting six-dimensional physical features from the predicted residuals: temperature residual, residual growth, temperature rise rate, temperature rise acceleration, voltage anomaly, and temperature-voltage coupling anomaly. These features are then weighted and fused to generate a comprehensive risk score. The effect is that the six-dimensional physical features characterize the precursory behavior of thermal runaway from multiple perspectives, including the degree of thermal deviation, deviation trend, temperature rise rate, temperature rise acceleration, internal short circuit representation, and thermoelectric coupling. This provides a complete physical basis for early warning, solving the problem of lack of physical interpretability in purely data-driven methods. It enables maintenance personnel to clearly understand the physical reasons triggering early warnings, enhancing the credibility of decision-making.

[0017] 3. A triple gating mechanism is employed to generate a formal trigger signal, consisting of an anomaly score threshold of 5.0, a duration threshold of 20 seconds, and a cumulative sum statistic threshold of 15.0. This is combined with a persistent confirmation mechanism (where the comprehensive risk score consistently exceeds the warning threshold and at least two physical features exhibit synchronous anomalies or temperature-voltage coupling anomalies are established) for level mapping. The effect is that the triple gating mechanism jointly constrains the signal from three dimensions: statistical significance, time duration, and cumulative effect. The persistent confirmation mechanism further requires synchronous anomalies in multiple physical features, effectively eliminating instantaneous anomalies caused by non-thermal runaway factors such as sensor noise and load fluctuations. This significantly reduces the false alarm rate and addresses the pain point of instantaneous fluctuations easily leading to false triggers in existing technologies.

[0018] Furthermore, this invention establishes a four-level risk level (normal, watchful, early warning, high-priority alarm) hierarchical mapping mechanism and an offline lead time verification method, ensuring engineering practicality and evaluability. Experimental data shows that this invention can achieve an early warning lead time of 506 seconds (approximately 8.4 minutes) in an overcharge thermal runaway scenario of a 280Ah lithium iron phosphate battery cell, while achieving zero false alarms on a normal benchmark dataset. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a battery thermal runaway early warning method that integrates time-series prediction and physical residual verification, as proposed in an embodiment of the present invention. Figure 2 This is a graph showing the evolution of voltage, temperature, and warning level over time during the early warning process of overcharge thermal runaway in lithium iron phosphate cells, as proposed in this embodiment of the invention. Detailed Implementation

[0020] 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.

[0021] To address the technical problems mentioned in the background art, such as insufficient lead time for battery safety warnings, high false alarm rates, and poor physical interpretability, this invention provides a battery thermal runaway early warning method that integrates timing prediction and physical residual verification. (Refer to...) Figure 1 As shown, the specific steps include: Step 1: Data Acquisition and Preprocessing. In this embodiment, the training data comes from the historical operating data of multiple different types of battery cells under normal charge-discharge cycles. Each original record includes a timestamp, voltage (V), temperature (T), dataset identifier (dataset_id), charge / discharge stage identifier (step_type), and depth of discharge (depthofdischarge, dod). The sampling frequency for voltage and temperature is 1 Hz, that is, one data point is collected per second.

[0022] First, a legality check is performed: (1) Check if the cell number exists in the preset valid list; (2) Check if the timestamp sequence is strictly monotonically increasing; (3) Check if the voltage value is within a reasonable range; (4) Check if the temperature value is within a reasonable range; (5) For data that passes the physical range check, the 3 sigma principle is further used to remove statistical outliers: Calculate the mean and standard deviation of each feature in its respective operating condition group. If a data point deviates from the mean by more than 3 times the standard deviation, it is judged as an outlier and removed.

[0023] After completing the inspection, a sliding window is used to construct time series samples. Let the window length be... Then at time The input window is: ; in to They represent from time 1 to 2. arrive Each sampling point contains a feature vector, and each feature vector contains voltage and temperature values. In this embodiment, the window length w=120, meaning that each input is a continuous 120 seconds of historical data.

[0024] Then, the voltage and temperature data within the window are Z-score standardized separately: ; in and These are the global mean and standard deviation of the feature in the training set, respectively. These two statistics are calculated and saved during the training phase and used directly during inference to avoid data leakage caused by online calculation.

[0025] If the current accumulated historical data is less than 120 points, the system enters a warm-up state. At this time, no valid risk level is output. Subsequent reasoning can only proceed after the window is filled.

[0026] Step 2: Construction and training of the TCN prediction main chain.

[0027] This step constructs a Temporal Convolutional Network (TCN) as the prediction master chain. The core components of the TCN network include causal convolution and dilated convolution.

[0028] (i) Network architecture: Causal convolution ensures that the output at time t depends only on the input at time t and previous times, satisfying the causal constraint of online inference, i.e., it does not use future information to predict the current value. Dilated convolution expands the receptive field by inserting holes between the convolution kernel elements, enabling the network to cover a longer time range without changing the number of parameters.

[0029] The specific structural parameters of the TCN network in this embodiment are as follows: convolutional kernel size K=3; 5 residual block layers, with the dilation factor of each layer increasing exponentially, d∈{1,2,4,8,16}; the number of channels in each residual block is uniformly set to 64; the Dropout random inactivation rate is set to 0.2; the model input window length is 120, and the output length is 60, i.e., predicting the voltage and temperature values ​​at 60 time points (60 seconds) in one go. Each dilated convolution layer is followed by a ReLU activation function, and residual connections are used to add the layer input and output to alleviate the gradient vanishing problem in deep network training.

[0030] (ii) Domain embedding mechanism: Since the voltage and temperature evolution of batteries vary significantly under different datasets, different charging and discharging stages and different discharge depths, this invention introduces a domain embedding mechanism to enhance the model's adaptability to operating conditions.

[0031] The specific implementation of the domain embedding layer is as follows: the dataset identifier, charge / discharge stage identifier, and discharge depth are converted into integer indices. Specifically, the discharge depth is first discretized from a continuous range of 0% to 100% into 20 equally spaced intervals, and then converted into corresponding interval indices. These integer indices are then mapped to low-dimensional continuous vectors using independent embedding matrices. In this embodiment, the dataset identifier embedding dimension is 8, the charge / discharge stage identifier embedding dimension is 4, and the discharge depth embedding dimension is 8. The three embedding vectors are concatenated to obtain the domain embedding vector. The total dimensions are 8+4+8=20.

[0032] The voltage value at the current time t and temperature value With domain embedding vector The features are concatenated along the feature dimension to obtain the final input features of the TCN network. : ; Here, Concat represents the vector concatenation operation. This embedding layer and the subsequent TCN network are jointly trained end-to-end through backpropagation, enabling the embedding vectors to automatically encode the data distribution shift under different operating conditions, thereby enhancing the model's adaptability to unknown operating conditions.

[0033] (III) Training Configuration: The preprocessed normal operating condition data is divided into training set:validation set:test set in a ratio of 7:1.5:1.5. The training configuration parameters for this embodiment are as follows: batch size is 128, and the initial learning rate is 10. 3 The minimum learning rate is 10. 6 The weight decay factor is 10. 5 AdamW optimizer parameters , 0.999, =10 8 The Huber loss smoothing threshold δ=1.0, the maximum number of training rounds is 300, and the early stopping patience value is 20.

[0034] The network training employs the Huber loss function, with a smoothing threshold parameter δ=1.0 in this embodiment. The Huber loss manifests as mean squared error for small errors and mean absolute error for large errors, reducing the impact of outliers on gradient updates. The optimizer uses AdamW with a cosine annealing learning rate scheduling strategy. After each training epoch, the model's loss value on the validation set is calculated. If the validation set loss does not decrease for 20 consecutive epochs, training stops, and the model parameters that minimize the validation set loss are restored as the final deployment model.

[0035] (iv) Model selection criteria. This invention selects a temporal convolutional network as the prediction backbone based on the following technical considerations: (1) TCN can obtain an exponentially increased receptive field without increasing the number of parameters through dilated convolution. The receptive field of 5 layers of dilated convolution can reach 48 time steps (combined with multi-layer accumulation, the actual coverage exceeds 120 time steps), which is suitable for capturing the slow change trend of thermal runaway incubation period of several minutes; (2) Causal convolution ensures the causal relationship in time sequence, which meets the real-time requirements of online early warning; (3) Compared with recurrent networks such as LSTM, TCN supports parallel computing and has lower inference latency, which is suitable for deployment in resource-constrained embedded systems; (4) The residual connection structure of TCN has more stable training performance under the same number of parameters.

[0036] Step 3: Online reasoning and anomaly score calculation.

[0037] The preprocessed data of the current window is input into the trained TCN network, which then outputs the voltage prediction values ​​for the next 60 time points at once. and temperature forecast For the current time t, take the predicted value and the actual observed value at the corresponding step. , Compare and calculate the predicted residuals: ; in It can be a voltage value or a temperature value. This is the corresponding predicted value. This formula is applied to both voltage and temperature to obtain the voltage residual. and temperature residual The dimensions of the residuals are consistent with those of the original physical quantities (voltage residuals are in volts, and temperature residuals are in degrees Celsius).

[0038] Further calculate the residual fractions. Taking the temperature residual as an example, its residual fraction is: ; in This is the baseline standard deviation of the temperature residuals under normal operating conditions, expressed in degrees Celsius. This baseline value was obtained during the training phase by statistically analyzing the temperature residuals of all normal charging data. Similarly, the voltage residual fractions are calculated using the corresponding voltage residual standard deviation baseline. The residual fraction is a dimensionless quantity that represents the multiple of the current residual relative to the normal fluctuation range.

[0039] Simultaneously, the cumulative sum statistic (CUSUM) of the residuals is calculated to detect persistent bias trends in the residuals. Its recursive formula is: ; in The mean of the temperature residuals under normal operating conditions (unit is the same as the residuals, in degrees Celsius), initial value. =0. The physical meaning of this statistic is that when the residuals are consistently higher than the normal mean, the cumulative sum increases continuously, reflecting the cumulative effect of deviation from the normal state; when the residuals fall back below the mean, the cumulative sum is kept from declining by taking the maximum value, avoiding the loss of cumulative abnormal information due to the mutual cancellation of positive and negative fluctuations.

[0040] Finally, the residual scores are weighted and fused with the cumulative sum statistic to obtain the outlier score. Since the residual scores are dimensionless while CUSUM has original dimensions, CUSUM needs to be divided by the baseline standard deviation of CUSUM under normal operating conditions before fusion. (Unit: Celsius) Dimensionless transformation is performed. The fusion formula is: ; in and The preset fusion weight coefficients are, where, For the predicted residual fractions Contribution ratio in anomaly assessment The contribution ratio of the cumulative statistic CUSUM in the anomaly assessment satisfies α + β = 1. In this embodiment, α = 0.7 and β = 0.3.

[0041] Step 4: Triple Gating to Generate the Formal Trigger Signal. This invention employs a triple gating mechanism to generate the formal trigger signal, consisting of an anomaly score threshold of 5.0, a duration threshold of 20 seconds, and a cumulative sum statistic threshold of 15.0. Specifically, the judgment logic requires that three conditions be met simultaneously: Condition 1: Current abnormal score ≥5.0 (dimensionless). This threshold ensures that the current deviation is statistically significant; a value of 5.0 means that the current outlier score has exceeded 5 standard deviations from the normal baseline.

[0042] Condition 2: The continuous duration of the abnormal state is ≥ 20 seconds. Where duration ranges from... The timer starts counting down from the moment the score reaches 5.0 for the first time. If the score drops below 5.0 at any point in the timeline, the timer resets to zero. This condition requires the anomaly to persist for a sufficiently long time, effectively eliminating sporadic triggering caused by transient noise spikes.

[0043] Condition 3: Cumulative Sum Statistic ≥15.0 (unit: degrees Celsius; the dimension of this threshold is consistent with that of the temperature residual, corresponding to the 99.9th percentile of the CUSUM statistical distribution under normal operating conditions).

[0044] When all three conditions are met simultaneously, the formal trigger signal is activated. Enabled: ; in This is an indicator function; it returns 1 if the condition is true, and 0 otherwise. When =1, it indicates that the TCN main chain has been confirmed to be in a high-risk state.

[0045] Step 5: Physical residual verification branch. This branch runs independently of the main chain and extracts multiple abnormal features from the perspective of physical mechanism.

[0046] (i) Temperature-related physical characteristics; temperature residuals are directly calculated using the main chain. ,Right now The unit is Celsius.

[0047] Residual growth The time-series variation trend of the temperature residual: ; in =1 second is the sampling time interval, and the unit of this feature is degrees Celsius per second.

[0048] The rate of temperature rise is calculated as the first derivative of temperature with respect to time, and the acceleration of temperature rise is calculated as the second derivative of temperature with respect to time.

[0049] (ii) Voltage anomalies and temperature-voltage coupling anomalies. The voltage anomaly degree is defined as the normalized value of the voltage residual relative to the normal fluctuation amplitude: ; in This is the baseline standard deviation of the voltage residual under normal operating conditions, with the unit being volts, consistent with voltage.

[0050] Temperature-voltage coupling anomaly indication value The judgment logic is as follows: ; in (dimensionless) and (Unit: degrees Celsius / second) These are the voltage anomaly threshold and the temperature rise rate threshold, respectively.

[0051] Among them, the voltage anomaly threshold Take the 99th percentile value of the voltage anomaly characteristic under normal operating conditions, i.e. =2.01 (corresponding to an actual voltage deviation of approximately 126.94 millivolts); Temperature rise rate threshold Take the 99th percentile value of the temperature rise rate characteristic under normal operating conditions, that is... =0.050 degrees Celsius / second (i.e., 3.0 degrees Celsius per minute).

[0052] (III) Calculation of the comprehensive risk score: The above six physical characteristics are converted into dimensionless scores by dividing each characteristic value by the baseline standard deviation of the corresponding characteristic under normal operating conditions. The scores are as follows: Temperature residual fraction ; ℃; Residual growth fraction ; ℃ / s; Fractional rate of temperature rise , =0.015℃ / s; Fractional acceleration due to temperature rise , ==0.002℃ / s²; Voltage anomaly score (This is already a dimensionless quantity); Temperature-voltage coupling anomaly indication value (Take 0 or 1).

[0053] Then, a weighted sum is performed to obtain the comprehensive risk score: ; in to These correspond to the six feature scores mentioned above. Among them, the temperature rise rate and temperature residual characteristics are most sensitive to the incubation stage of thermal runaway, and therefore are given higher weights; the temperature-voltage coupling anomaly is mainly used as a key confirming feature, and is assigned a lower but reinforcing weight. All weights satisfy... =1.

[0054] The sub-fraction vector is defined as: ; The physical verification level is segmented and mapped according to the threshold range of the comprehensive risk score. This embodiment sets... For the attention threshold, This is the warning threshold.

[0055] The preferred approach provides a detailed explanation of the sub-fraction generation process as follows: (1) Sub-score generation: In order to eliminate the impact of the difference in the dimensions of different physical features on the risk assessment results, the physical features extracted by S31 and S32 are uniformly scored.

[0056] To uniformly represent various physical characteristics, the following definition is provided: It is a characteristic of temperature residuals; This is a characteristic of residual growth; Characterized by the rate of temperature rise; It is a characteristic of temperature rise acceleration; Characteristics of voltage anomaly; This is a voltage-temperature coupling anomaly.

[0057] To reduce the impact of instantaneous fluctuations on the risk assessment results, the maximum value of each physical characteristic is taken within the most recent 60-second sliding time window:

[0058] In the formula, Let i represent the maximum observed value of the i-th physical feature within the sliding window, where i = 1, 2, 3, 4, 5, 6.

[0059] Based on normal operating condition sample data from the training or calibration set, calculate the statistical quantile thresholds corresponding to each physical feature:

[0060]

[0061]

[0062] In the formula, This represents the 95th percentile value of the i-th physical feature, meaning that approximately 95% of normal samples have feature values ​​that do not exceed this value. This represents the 99th percentile value of the i-th physical feature, meaning that approximately 99% of normal samples have feature values ​​that do not exceed this value. This represents the 99.9th percentile value of the i-th physical characteristic, meaning that approximately 99.9% of normal samples correspond to a characteristic value that does not exceed this value. The above statistical percentile values ​​are calculated based on normal operating condition samples from the training or calibration set, without relying on abnormal samples; therefore, they can characterize the fluctuation range under normal battery operating conditions.

[0063] , , It does not represent the final statistical threshold, but rather the statistical quantile function, which is used to calculate the statistical quantile values ​​of the corresponding physical characteristics at different probability locations based on normal operating condition sample data.

[0064] in, This represents a statistical function that performs α-quantile calculations on a dataset X. Its input is physical feature sample data, and its output is the feature value at the corresponding quantile position.

[0065] therefore, , , They are represented as: (1) : A statistical function for calculating the 95th percentile of physical characteristics under normal operating conditions; (2) : A statistical function for calculating the 99th percentile of physical characteristics under normal operating conditions; (3) : A statistical function for calculating the 99.9th percentile of physical characteristics under normal operating conditions.

[0066] Each sub-fraction is obtained from the corresponding physical characteristic through a mapping function: ; In the formula, the mapping function is defined as:

[0067] In the formula, This is a stabilizing term set up to prevent the denominator from being zero.

[0068] When a physical characteristic value is below the 95th percentile, the characteristic is considered to be within the normal fluctuation range, and the corresponding sub-score is 0. When the physical characteristic value is between the 95th and 99th percentiles, the sub-score increases linearly from 40 to 60, corresponding to mild anomaly. When the physical characteristic value is between the 99th and 99.9th percentiles, the sub-score increases linearly from 60 to 80, corresponding to moderate anomaly. When the physical characteristic value exceeds the 99.9th percentile, the sub-score continues to increase from 80, with a maximum limit of 100, corresponding to high anomaly.

[0069] (2) Definition of sub-fraction vector, finally obtaining the sub-fraction vector:

[0070] In the formula, The fractional value of the temperature residual; The residual growth fraction; The fraction of the rate of temperature rise; The fraction of temperature rise acceleration; For voltage anomaly degree sub-fraction; The voltage-temperature coupling anomaly sub-fraction.

[0071] (3) Calculation method of comprehensive risk score: The comprehensive risk score is defined as the weighted sum of sub-scores:

[0072] In the formula, For the corresponding physical characteristic scores; These are the weight coefficients for the corresponding features, and they satisfy the following constraints: .

[0073] (4) Physical verification level mapping The physical verification level is segmented and mapped based on the comprehensive risk score, and the calculation formula is as follows:

[0074] In the formula, , , These are the risk score thresholds, used to distinguish between risk levels 0, 1, 2, and 3, and they must meet the following conditions: This physical verification level is used internally to characterize the strength of the physical residual branches in supporting the final output risk level. Specifically: Used to distinguish between normal and attention status; Used to distinguish between a state of concern and a state of alert; Used to distinguish between early warning status and high-priority alarm status.

[0075] The specific mapping relationship is as follows: When the comprehensive risk score is lower than At that time, the corresponding level is 0 (normal); When the comprehensive risk score reaches But lower At that time, the corresponding level is 1 (attention); When the comprehensive risk score reaches But lower At that time, the corresponding level is 2 (warning). When the formal trigger signal When =1, it corresponds to level 3 (high priority alarm).

[0076] Step Six: Integrate Decision Making and Hierarchical Mapping; the integration module simultaneously receives the formal trigger signal from the main chain. Signal and physical branches and vector.

[0077] First, perform continuous verification (denoted as...). The judgment logic includes the following conditions: (1) Comprehensive risk score Continuously exceeding the preset warning threshold =60 and the duration reaches 10 seconds; (2) at least two physical feature scores in the sub-score vector exceed their respective attention thresholds, or the temperature-voltage coupling anomaly indicator value. =1.

[0078] When all of the above conditions are met Set the state to true.

[0079] Then, the four risk levels are mapped according to the following rules: Level 0 (Normal): and =0, the output suggestion is no action required; Level 1 (Follow): If this state lasts for more than 5 seconds, the recommended output is to enhance the recording and manually monitor the trend of this cell and neighboring cells. Level 2 (Warning): For true and The output recommendations are to review and prepare for action, enhance monitoring, increase sampling frequency, and conduct manual review; Level 3 (High-priority alarm candidate): =1 and If true, the output suggestion is to enter a more robust handling process; whether automatic linkage is implemented is determined by the station's, battery management system's, or energy management system's strategy. The timestamp of the first attainment of each level is also recorded.

[0080] Step 7: Tiered Output and Format Validation. The online output fields include: current risk level identifier (level_advisory, an integer from 0 to 3); level name (level_name, corresponding to normal, watch out, warning, and high-priority alarms); main triggering reason (main_reason, such as abnormal temperature-voltage coupling, excessively high temperature rise rate, etc., diagnostic labels); and action suggestion (action_suggestion).

[0081] In actual engineering deployment, the system performs format validation on the output results. The validation includes: (1) whether the risk level identifier is an integer in the range of 0 to 3; (2) whether the timestamp is a valid value that increases monotonically; (3) whether each risk score is a finite value, excluding infinite or non-numerical cases; (4) whether the output data structure contains all required fields and the field types are correct; (5) The value must be a Boolean or a 0 / 1 integer. If any check fails, the system logs an error and returns the last valid output.

[0082] An offline playback mechanism is adopted, in which historical data with marked thermal runaway times are input into the model in chronological order to simulate online real-time stream processing.

[0083] Define the thermal runaway reference point t_anchor as the critical reference moment when the battery experiences irreversible thermal runaway (such as the starting point of a sharp voltage drop or temperature increase). Record the time when the model first outputs level 2, alarm_t_s. If the same test sample simultaneously exhibits level 2 and level 3, the time of the first occurrence of level 2 is still taken.

[0084] The lead time for early warning is calculated as: lead_anchor_s = t_anchor alarm_t_s; If the calculation result is positive, it means that the model issued an early warning before thermal runaway occurred; if it is negative, it means that the warning lagged behind the thermal runaway reference point.

[0085] In this embodiment, the thermal runaway anchor point t_anchor = 837.0 seconds, and the time to first reach level 2 is 331.0 seconds, resulting in a calculated warning lead time of 837.0 seconds. 331.0 = 506.0 seconds (approximately 8.4 minutes). Meanwhile, when tested on a long-term normal benchmark dataset, the model did not produce any false positives of level 2 or higher.

[0086] The model's overall risk score exceeded the Level 2 warning threshold for the first time at 331.0 seconds. =60 and the persistence confirmation condition is met, the system outputs a Level 2 warning for the first time; at 744.0 seconds... When set to 1, the system outputs a Level 3 high-priority alarm. The entire level change trajectory is Level 0 → Level 1 → Level 2 → Level 3, showing a monotonically increasing trend as the risk increases.

[0087] refer to Figure 2 As shown, for reference Figure 2 As shown, in the early stage of thermal runaway (0-300 seconds), the cell voltage and temperature were within the normal fluctuation range, and no warning level was triggered. The system continued to output Level 0 normal status.

[0088] Approximately 300 seconds later, the temperature residual and the rate of temperature rise began to show a synchronous upward trend. At approximately 311.0 seconds, the comprehensive risk score exceeded the concern threshold θ1=40 for the first time, and the system output a Level 1 concern level, indicating that the cell had exhibited preliminary abnormal thermal-electric coupling behavior. At this time, the actual voltage deviation was approximately 126.94mV, and the rate of temperature rise was approximately 0.050℃ / s.

[0089] At 331.0 seconds, the comprehensive risk score further exceeded the warning threshold θ2=60, and the continuous confirmation condition was met, indicating an abnormal voltage-temperature coupling value. The setting is 1, and the system enters Level 2 alert status. From 331.0 seconds to 744.0 seconds, the cell temperature shows an accelerated upward trend, the overall risk score continues to climb, and the system remains in Level 2 alert status. During this stage, maintenance personnel can perform manual review and prepare for handling.

[0090] At 744.0 seconds, the TCN main chain anomaly score, duration, and CUSUM statistic simultaneously met the Z5N20H15 triple gating condition. When set to 1, the system triggers a Level 3 high-priority alarm, indicating that the battery cell has entered a high-risk state, and it is recommended to immediately initiate a stronger handling procedure.

[0091] The cell ultimately reached the thermal runaway reference anchor point (the starting point of the voltage drop) at 837.0 seconds. From the first Level 2 warning (331.0 seconds) to the thermal runaway anchor point (837.0 seconds), the warning lead time was 506 seconds (approximately 8.4 minutes), verifying the effectiveness of the method of this invention in early warning of thermal runaway.

[0092] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A battery thermal runaway early warning method integrating time series prediction and physical residual verification, characterized in that, include: Collect timing data of battery operation, including voltage and temperature; A time series prediction main chain is constructed. The main chain is trained with voltage and temperature data under normal operating conditions. It is used to predict future time series data based on current time series data and generate anomaly scores based on prediction residuals. The main chain generates formal trigger signals through multiple gating, including anomaly score threshold, duration threshold, and cumulative and statistical thresholds. A physical residual verification branch is constructed, which is used to extract physical features from the predicted residual. The physical features include at least temperature residual and temperature-voltage coupling anomaly features. The physical features are converted into sub-fractions through a piecewise mapping function based on statistical quantiles, and then weighted and fused to generate a comprehensive risk score. The temperature-voltage coupling anomaly features are determined to be valid when the voltage anomaly degree reaches a first threshold and the temperature rise rate reaches a second threshold. By integrating the formal trigger signal with the comprehensive risk score, a multi-level risk level is output. When the historical data is insufficient for the sliding window length, the system enters a preheating state and does not output a valid risk level.

2. The method according to claim 1, characterized in that, The time-series prediction main chain adopts a temporal convolutional network, which extracts temporal features through causal convolution and dilated convolution. The temporal convolutional network introduces a domain embedding layer to map the working condition identifier into a continuous vector and concatenate it with the temporal features before inputting it into the network. The working condition identifier includes at least a dataset identifier, a charging / discharging stage identifier, and a discharge depth.

3. The method according to claim 1, characterized in that, The outlier score is obtained by weighted fusion of the residual score and the cumulative sum statistic of the residuals. The residual score is the absolute value of the predicted residual divided by the baseline standard deviation of the residual under normal operating conditions. The cumulative sum statistic is used to detect the persistent deviation trend of the residuals. Its recursive method is to take the cumulative sum statistic of the current moment as the larger value between the cumulative sum statistic of the previous moment, the current residual minus the mean of the residual under normal operating conditions, and zero.

4. The method according to claim 1, characterized in that, The thresholds for the multiple gating are: anomaly score threshold of 5.0, duration threshold of 20 seconds, and cumulative sum statistic threshold of 15.

0. The formal trigger signal is set to valid when all three conditions are met.

5. The method according to claim 1, characterized in that, The physical characteristics also include residual growth and temperature rise acceleration; the residual growth is the difference between the temperature residual at the current moment and the temperature residual at the previous moment divided by the sampling interval; the temperature rise rate is the difference between the temperature at the current moment and the temperature at the previous moment divided by the sampling interval; the temperature rise acceleration is the difference between the temperature rise rate at the current moment and the temperature rise rate at the previous moment divided by the sampling interval; the voltage anomaly is the absolute value of the voltage residual divided by the baseline standard deviation of the voltage residual under normal operating conditions.

6. The method according to claim 1, characterized in that, After converting the physical features into sub-scores using a piecewise mapping function based on statistical quantiles, a weighted fusion is performed to generate a comprehensive risk score, specifically including: Each sub-fraction Based on the corresponding physical characteristics through the mapping function get: ； In the formula, For corresponding physical characteristics The maximum value within the sliding window, i.e. ； Sub-fraction vector ; In the formula, The fractional value of the temperature residual; The residual growth fraction; The fraction of the rate of temperature rise; The fraction of temperature rise acceleration; For voltage anomaly degree sub-fraction; The voltage-temperature coupling anomaly sub-fraction; Comprehensive risk score ; In the formula, For the corresponding physical characteristic scores; These are the weight coefficients for the corresponding features, and they satisfy the following constraints: .

7. The method according to claim 1, characterized in that, When generating a comprehensive risk score by weighted fusion of physical features, the weights of each physical feature score are configured as follows: temperature residual score weight 0.20, residual growth score weight 0.15, temperature rise rate score weight 0.25, temperature rise acceleration score weight 0.15, voltage anomaly score weight 0.15, and temperature-voltage coupling anomaly indication value weight 0.

10.

8. The method according to claim 1, characterized in that, The first threshold and the second threshold are determined based on the 99th percentile value of the corresponding feature under normal operating conditions.

9. The method according to claim 1, characterized in that, The process of integrating the formal trigger signal and the comprehensive risk score to output a multi-level risk level specifically includes: performing continuous confirmation, whereby the comprehensive risk score is continuously higher than a preset warning threshold. Furthermore, if the duration reaches a preset value, and at least two physical feature scores exceed their respective attention thresholds or a temperature-voltage coupling anomaly is established, the final risk level is determined based on the continuous confirmation results and the validity of the formal trigger signal.

10. The method according to claim 1, characterized in that, The multi-level risk assessment system includes: Level 0 represents normal, Level 1 represents attention, Level 2 represents warning, and Level 3 represents high-priority alarm; Level 1 is defined as a comprehensive risk score reaching the attention threshold. Triggered, Level 2: The overall risk score reaches the warning threshold. Furthermore, the continuous confirmation of the trigger is required, and Level 3 is triggered by the validity of the formal trigger signal and the continuous confirmation of the trigger. It also includes an offline verification step: using the thermal runaway reference time as the evaluation anchor point, the time difference between the time point when the warning level is first reached and the evaluation anchor point is calculated as the warning advance.

Images