Thermal runaway risk assessment method and device and storage medium

Abstract

The invention relates to a thermal runaway risk assessment method and device and a storage medium, relates to the technical field of vehicles, and at least solves the technical problem of how to carry out early warning on a battery before thermal runaway of the battery in related technologies. The method comprises the following steps: acquiring time sequence state data of a battery, wherein the time sequence state data is used for indicating change conditions of battery parameters of the battery at different time points; the time sequence state data is input into a trained thermal runaway early warning model, a risk assessment result is obtained, the thermal runaway early warning model is used for detecting whether the battery has a thermal runaway risk or not, and the thermal runaway early warning model comprises a long and short-term memory network module and a soft decision tree module, the long-short-term memory network module is used for extracting characteristic data related to the thermal runaway risk, and the soft decision tree module is used for determining whether thermal runaway occurs in the battery or not.

Description

Technical Field

[0001] The present application relates to the technical field of vehicles, in particular to the technical field of batteries, and specifically relates to a method, device and storage medium for evaluating the risk of thermal runaway. Background Art

[0002] With the rapid development of vehicle technology, people's requirements for vehicle safety are getting higher and higher. As a power source of a vehicle, if the internal temperature of a single battery cell of the battery rises abnormally and thermal runaway occurs, it will cause heat radiation or conduction through conductive substances to adjacent battery cells, resulting in the melting of the battery separator, violent reaction of the electrolyte and release of a large amount of combustible gas, leading to fire or even explosion.

[0003] Therefore, it is necessary to monitor the battery and give an early warning to the battery before thermal runaway occurs to ensure the safety of the vehicle. And how to give an early warning to the battery before thermal runaway has become a technical problem to be solved urgently. Summary of the Invention

[0004] The present application provides a method, device and storage medium for evaluating the risk of thermal runaway to at least solve the technical problem of how to give an early warning to the battery before thermal runaway in the related art. The technical solution of the present application is as follows: According to a first aspect of the present application, there is provided a method for evaluating the risk of thermal runaway, including: obtaining time series state data of a battery, where the time series state data is used to indicate the change of battery parameters of the battery at different time points. Inputting the time series state data into a trained thermal runaway early warning model to obtain a risk assessment result, where the thermal runaway early warning model is used to detect whether there is a risk of thermal runaway in the battery, and the thermal runaway early warning model includes: a long short-term memory network module and a soft decision tree module, where the long short-term memory network module is used to extract feature data related to the risk of thermal runaway, and the soft decision tree module is used to determine whether thermal runaway occurs in the battery.

[0005] In a possible implementation manner, inputting the time series state data into a trained thermal runaway early warning model to obtain a risk assessment result includes: inputting the time series state data into the long short-term memory network module to obtain a risk feature vector of the battery, where the risk feature vector is used to indicate feature data related to the risk of thermal runaway. Inputting the risk feature vector into the soft decision tree module to obtain a risk assessment result.

[0006] In a possible implementation, the soft decision tree module includes: multiple preset evaluation paths. Inputting the risk feature vector into the soft decision tree module to obtain a risk assessment result, including: for each preset evaluation path, based on the risk feature vector and the preset evaluation path, obtaining the initial probability evaluation information corresponding to the preset evaluation path, so as to obtain the initial probability evaluation information corresponding to multiple preset evaluation paths. The initial probability evaluation information is used to indicate the probability of the battery having a thermal runaway risk, and one preset evaluation path corresponds to one initial probability evaluation information. Based on the multiple initial probability evaluation information, the risk assessment result is obtained.

[0007] In a possible implementation, the preset evaluation path includes multiple evaluation factors and multiple probability weights. The evaluation factors are used to indicate the conditions for the battery to have a thermal runaway risk, and the probability weights are used to indicate the influence degree of the evaluation factors on the battery having a thermal runaway risk. One evaluation factor corresponds to multiple probability weights. Based on the risk feature vector and the preset evaluation path, obtaining the initial probability evaluation information corresponding to the preset evaluation path includes: based on the risk feature vector, multiple evaluation factors and multiple probability weights, obtaining the initial probability evaluation information corresponding to the preset evaluation path.

[0008] In a possible implementation, the long short-term memory network module includes: a fully connected layer, and the fully connected layer is used to input the risk feature vector into the soft decision tree module. Among them, the dimension of the fully connected layer of the long short-term memory network module is the same as the dimension of the input layer of the soft decision tree module.

[0009] In a possible implementation, the trained thermal runaway warning model is obtained through the following method: obtaining a training set, a first model loss, a second model loss, and an initial thermal runaway warning model. The first model loss is used to adjust the accuracy of the soft decision tree module, and the second model loss is used to adjust the accuracy of the long short-term memory network module. Training the initial thermal runaway warning model based on the training set, the first model loss, and the second model loss to obtain the trained thermal runaway warning model.

[0010] In a possible implementation, the initial thermal runaway warning model is a pruned thermal runaway warning model. Training the initial thermal runaway warning model based on the training set, the first model loss, and the second model loss to obtain the trained thermal runaway warning model includes: training the initial thermal runaway warning model based on the training set, the first model loss, and the second model loss to obtain the first model parameters. Sending the training set and the first model parameters to the server. The server deploys an unpruned thermal runaway warning model, and the server is used to optimize the first model parameters based on the training set and the unpruned thermal runaway warning model. Receiving the second model parameters. The second model parameters are the optimized first model parameters. Based on the second model parameters, the trained thermal runaway warning model is obtained.

[0011] In a possible implementation manner, the method further includes: obtaining an initial soft decision forest model, where the initial soft decision forest model includes a plurality of initial evaluation paths, and the initial evaluation paths include: a branch structure. Based on the plurality of branch structures, determine the similarity of the plurality of branch structures. Based on the similarity of the plurality of branch structures, determine multiple groups of similar structure groups, where a group of similar structure groups includes at least two branch structures whose similarity is less than a preset similarity threshold. Fuse the branch structure parameters in each group of similar structure groups to obtain a soft decision forest model.

[0012] In a possible implementation manner, the branch structure includes at least one of the following: a branch node and a decision edge. The branch node is used to indicate an evaluation factor, and the decision edge is used to indicate a probability weight. The evaluation factor is used to indicate the conditions for the battery to have a thermal runaway risk, and the probability weight is used to indicate the influence degree of the evaluation factor on the thermal runaway risk of the battery. One evaluation factor corresponds to multiple probability weights.

[0013] According to the second aspect provided by the present application, there is provided an apparatus for evaluating thermal runaway risk, where the apparatus includes an acquisition module and a processing module.

[0014] The acquisition module is used to acquire time series state data of the battery, and the time series state data is used to indicate the change of the battery parameters of the battery at different time points. The processing module is used to input the time series state data into the trained thermal runaway early warning model to obtain a risk assessment result. The thermal runaway early warning model is used to detect whether the battery has a thermal runaway risk, and the thermal runaway early warning model includes: a long short-term memory network module and a soft decision tree module. The long short-term memory network module is used to extract feature data related to the thermal runaway risk, and the soft decision tree module is used to determine whether the battery has a thermal runaway.

[0015] In a possible implementation manner, the processing module is used to input the time series state data into the long short-term memory network module to obtain a risk feature vector of the battery, and the risk feature vector is used to indicate feature data related to the thermal runaway risk. The processing module is further used to input the risk feature vector into the soft decision tree module to obtain a risk assessment result.

[0016] In a possible implementation manner, the soft decision tree module includes: a plurality of preset evaluation paths. The processing module is used for each preset evaluation path, based on the risk feature vector and the preset evaluation path, to obtain initial probability evaluation information corresponding to the preset evaluation path, so as to obtain initial probability evaluation information corresponding to the plurality of preset evaluation paths. The initial probability evaluation information is used to indicate the probability of the battery having a thermal runaway risk, and one preset evaluation path corresponds to one initial probability evaluation information. The processing module is further used to obtain a risk assessment result based on the plurality of initial probability evaluation information.

[0017] In a possible implementation manner, the preset evaluation path includes multiple evaluation factors and multiple probability weights. The evaluation factors are used to indicate the conditions for the battery to have a thermal runaway risk, and the probability weights are used to indicate the influence degree of the evaluation factors on the thermal runaway risk of the battery. One evaluation factor corresponds to multiple probability weights. The processing module is used to obtain the initial probability evaluation information corresponding to the preset evaluation path based on the risk feature vector, multiple evaluation factors, and multiple probability weights.

[0018] In a possible implementation manner, the long short-term memory network module includes: a fully connected layer, and the fully connected layer is used to input the risk feature vector into the soft decision tree module. Among them, the dimension of the fully connected layer of the long short-term memory network module is the same as the dimension of the input layer of the soft decision tree module.

[0019] In a possible implementation manner, the acquisition module is used to acquire the training set, the first model loss, the second model loss, and the initial thermal runaway warning model. The first model loss is used to adjust the accuracy of the soft decision tree module, and the second model loss is used to adjust the accuracy of the long short-term memory network module. The processing module is used to train the initial thermal runaway warning model based on the training set, the first model loss, and the second model loss to obtain the trained thermal runaway warning model.

[0020] In a possible implementation manner, the initial thermal runaway warning model is a pruned thermal runaway warning model. The device further includes a sending module and a receiving module. The processing module is used to train the initial thermal runaway warning model based on the training set, the first model loss, and the second model loss to obtain the first model parameters. The sending module is used to send the training set and the first model parameters to the server. The server deploys an unpruned thermal runaway warning model, and the server is used to optimize the first model parameters based on the training set and the unpruned thermal runaway warning model. The receiving module is used to receive the second model parameters. The second model parameters are the optimized first model parameters. The processing module is further used to obtain the trained thermal runaway warning model based on the second model parameters.

[0021] In a possible implementation manner, the acquisition module is used to acquire the initial soft decision forest model. The initial soft decision forest model includes multiple initial evaluation paths, and the initial evaluation path includes: a branch structure. The processing module is used to determine the similarity of multiple branch structures based on the multiple branch structures. The processing module is further used to determine multiple groups of similar structure groups based on the similarity of the multiple branch structures. A group of similar structure groups includes at least two branch structures whose similarity is less than the preset similarity threshold. The processing module is further used to fuse the branch structure parameters in each group of similar structure groups to obtain the soft decision forest model.

[0022] In a possible implementation, the branch structure includes at least one of the following: a branch node and a decision edge. The branch node is used to indicate an evaluation factor, and the decision edge is used to indicate a probability weight. The evaluation factor is used to indicate the conditions for the battery to have a thermal runaway risk, and the probability weight is used to indicate the degree of influence of the evaluation factor on the thermal runaway risk of the battery. One evaluation factor corresponds to multiple probability weights.

[0023] According to a third aspect provided by the present application, there is provided an apparatus for evaluating thermal runaway risk, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to execute the instructions to implement the method according to the first aspect and any of its possible implementations.

[0024] According to a fourth aspect provided by the present application, there is provided a computer-readable storage medium. When the instructions in the computer-readable storage medium are executed by the processor of the apparatus for evaluating thermal runaway risk, the apparatus for evaluating thermal runaway risk can execute the method according to the first aspect and any of its possible implementations.

[0025] Advantages of the present invention: (1) By obtaining the time series status data of the battery and inputting the time series status data into the trained thermal runaway warning model, a risk assessment result can be obtained. The thermal runaway warning model includes: a long short-term memory network module and a soft decision tree module. Since the long short-term memory network module can capture the long-term trends and periodic changes in the time series status data, it can better determine the characteristic data related to the thermal runaway risk in the time series status data. The soft decision tree module can smooth the decision boundary and reduce the sensitivity to noise, so as to more accurately determine the probability of the battery having a thermal runaway risk, and further more accurately determine whether the battery has a thermal runaway. Moreover, since the data processing process of the soft decision tree module can perform interpretable hierarchical reasoning, the interpretability of the thermal runaway warning model can be improved.

[0026] (2) By combining the long short-term memory network model with the soft decision tree, while capturing higher-level abstract features, the performance and interpretability of the thermal runaway warning model can be balanced, making the risk assessment result more accurate and interpretable.

[0027] (3) By obtaining the initial probability assessment information of each preset evaluation path, the probability results of multiple preset evaluation paths of the soft decision tree model can be obtained. That is to say, the probability distribution of the battery having a thermal runaway risk can be obtained. In this way, the risk assessment result can be obtained more accurately according to the probability distribution of the battery having a thermal runaway risk.

[0028] (4) Since the preset evaluation paths of the soft decision tree model include multiple evaluation factors and multiple probability weights, each preset evaluation path corresponds to a clear feature combination logic. In this way, the meaning of each preset evaluation path is clear and interpretable, and when the risk assessment result is used to indicate that the battery has a thermal runaway, the cause of the thermal runaway risk of the battery can be determined.

[0029] (5) It can enable the long short-term memory network module and the soft decision tree module to be directly connected and trained as a whole, thus simplifying the model architecture of the thermal runaway warning model and improving the computational efficiency of the thermal runaway warning model.

[0030] (6) By obtaining a training set, a first model loss, a second model loss, and an initial thermal runaway warning model, and training the initial thermal runaway warning model based on the training set, the first model loss, and the second model loss, a trained thermal runaway warning model is obtained. The first model loss is used to adjust the accuracy of the soft decision tree module, and the second model loss is used to adjust the accuracy of the long short-term memory network module. In this way, the training direction of the thermal runaway warning model can be effectively guided, preventing the thermal runaway warning model from overfitting on the training data, improving the generalization ability of the thermal runaway warning model, enabling the thermal runaway warning model to have better robustness when facing different types of input data, and not causing the performance of the thermal runaway warning model to decline due to the overuse of some paths.

[0031] (7) By deploying the pruned thermal runaway warning model in the thermal runaway risk assessment device, the thermal runaway risk assessment device can quickly predict whether the battery has a thermal runaway. By deploying the unpruned thermal runaway warning model on the server and tuning the first model parameters through the server, while improving the speed of predicting whether the battery has a thermal runaway, the prediction result of the thermal runaway warning model can be made more accurate.

[0032] (8) By fusing the branch structures in each group of similar structure groups to obtain multiple fused branch structures, pruning of the thermal runaway warning model can be achieved. Moreover, since a group of similar structure groups has at least two branch structures with a similarity less than a preset similarity threshold, and the branch structure includes branch nodes and decision edges, while lightweight processing is performed on the branch nodes and decision edges, the accuracy of the thermal runaway warning model can be maintained.

[0033] It should be noted that the technical effects brought by any implementation manner in the second aspect to the fourth aspect can refer to the technical effects brought by the corresponding implementation manner in the first aspect, which will not be elaborated here.

[0034] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit this application. Brief Description of the Drawings

[0035] The accompanying drawings here are incorporated into the specification and constitute a part of this specification, showing embodiments in line with this application, and are used together with the specification to explain the principles of this application, and do not constitute an improper limitation to this application.

[0036] Figure 1 is a schematic diagram of an evaluation system for thermal runaway risk shown according to an exemplary embodiment; Figure 2 is a schematic flowchart of an evaluation method for thermal runaway risk shown according to an exemplary embodiment; Figure 3 is a schematic flowchart of another evaluation method for thermal runaway risk shown according to an exemplary embodiment; Figure 4 is an example schematic diagram of a system of vehicle end and server shown according to an exemplary embodiment; Figure 5 is an example schematic diagram of an evaluation method for thermal runaway risk shown according to an exemplary embodiment; Figure 6 is a schematic structural diagram of an evaluation device for thermal runaway risk shown according to an exemplary embodiment; Figure 7 is a schematic structural diagram of another evaluation device for thermal runaway risk shown according to an exemplary embodiment. Detailed implementation manners

[0037] In order to enable those of ordinary skill in the art to better understand the technical solutions of this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the accompanying drawings.

[0038] It should be noted that the terms "first", "second", etc. in the specification, claims and above-mentioned accompanying drawings of this application are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that such used data can be interchanged under appropriate circumstances so that the embodiments of this application described here can be implemented in an order other than those illustrated or described here. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. On the contrary, they are merely examples of devices and methods consistent with some aspects of this application as detailed in the appended claims.

[0039] Before introducing in detail the evaluation method for thermal runaway risk in the embodiments of this application, the implementation environment and application scenarios of the embodiments of this application will be introduced first.

[0040] An embodiment of the present application provides a method for evaluating the risk of thermal runaway, including: obtaining time-series state data of a battery and inputting the time-series state data into a trained thermal runaway warning model to obtain a risk assessment result. The thermal runaway warning model includes: a long short-term memory network module and a soft decision tree module. Since the long short-term memory network module can capture long-term trends and periodic changes in the time-series state data, it can better determine the feature data related to the thermal runaway risk in the time-series state data. The soft decision tree module can smooth the decision boundary and reduce the sensitivity to noise, thereby more accurately determining the probability of the battery having a thermal runaway risk and further more accurately determining whether the battery has a thermal runaway. Moreover, since the data processing process of the soft decision tree module can perform interpretable hierarchical reasoning, it can improve the interpretability of the thermal runaway warning model.

[0041] It should be noted that the execution subject of the method for evaluating the risk of thermal runaway provided in the present application can be an evaluation device for the risk of thermal runaway, and this device can be a battery controller or a vehicle. At the same time, this device can also be the central processing unit (CPU) of the vehicle, or the module for evaluating the risk of thermal runaway in this device, or the in-vehicle device in the vehicle. The present application does not limit this. In the embodiment of the present application, taking the vehicle as the execution subject of the method for evaluating the risk of thermal runaway as an example, the method for evaluating the risk of thermal runaway provided in the embodiment of the present application is described.

[0042] It should be noted that the vehicle can be, but is not limited to, a pure electric vehicle (Pure Electric Vehicle / Battery Electric Vehicle, PEV / BEV), a hybrid electric vehicle (Hybrid Electric Vehicle, HEV), a range extended electric vehicle (Range Extended Electric Vehicle, REEV), a plug-in hybrid electric vehicle (Plug-in Hybrid Electric Vehicle, PHEV), etc.

[0043] The implementation environment of the embodiment of the present application is introduced below.

[0044] Figure 1 As shown in the schematic diagram of an evaluation system for the risk of thermal runaway according to an exemplary embodiment, Figure 1 the evaluation system for the risk of thermal runaway includes: a time-series data acquisition and preprocessing module, a model construction module, a model joint training and parameter optimization module, and a model lightweight deployment and alarm module.

[0045] Among them, the time series data acquisition and preprocessing module includes: a data acquisition module and a data preprocessing module. The data acquisition module can obtain the initial battery state data, and the data preprocessing module is used to perform data preprocessing on the initial battery state data.

[0046] Extract the battery state data of a preset time period from the initial battery state data. Then, the battery state data of the preset time period can be subjected to data cleaning processing to obtain the first state data. Then, the first state data can be subjected to data normalization processing to obtain the second state data. Then, based on the sliding time window algorithm, the second state data can be processed to obtain the training set.

[0047] Optionally, the data acquisition module includes a voltage sensor, a current sensor, a temperature sensor, a State Of Charge (SOC) and State Of Health (SOH) estimation unit.

[0048] It should be noted that the signals directly collected by the sensors will pass through a conditioning circuit to amplify the signals input by the sensors, remove noise within a certain frequency range through a filter, and indirectly calculate the SOC and SOH.

[0049] It should be noted that this application does not limit the method of data preprocessing. For example, the method of data preprocessing can include: data cleaning, data normalization, and sliding time window algorithm.

[0050] The model construction module is used to construct a long short-term memory network model and a soft decision tree model, and based on the long short-term memory network model and the soft decision tree model, obtain an initial thermal runaway warning model.

[0051] The model joint training and parameter optimization module includes: a model training module and a parameter optimization module. The model training module is used to train the initial thermal runaway warning model. The parameter optimization module is used to optimize the parameters of the initial thermal runaway warning model.

[0052] The model lightweight deployment and alarm module includes: a model lightweight deployment module and an alarm module.

[0053] Optionally, the model lightweight deployment module is used to perform pruning processing on the long short-term memory network model and the soft decision tree model to obtain a pruned thermal runaway warning model. Moreover, the model lightweight deployment module is also used to deploy the pruned thermal runaway warning model to the in-vehicle terminal embedded system. Moreover, the model lightweight deployment module is also used to send the vehicle-side collected data and training parameters to the cloud and obtain the global model parameters from the cloud.

[0054] The alarm module is used to determine the target alarm level when the risk assessment result triggers the alarm condition, and issue an alarm based on the target alarm level.

[0055] Optionally, when the target alarm level is the first alarm level, alarm information can be sent through the instrument, and an alarm signal can be sent to the cloud platform monitoring center. When the target alarm level is the second alarm level, alarm information can be sent through the instrument, the high voltage of the power battery can be disconnected, and an alarm signal can be sent to the cloud platform monitoring center.

[0056] For ease of understanding, the following specifically introduces the method for evaluating the thermal runaway risk provided by this application in combination with the accompanying drawings.

[0057] As Figure 2 shown, the method includes the following steps: S201. Obtain the time series status data of the battery.

[0058] Among them, the time series status data is used to indicate the change of the battery parameters of the battery at different time points.

[0059] Optionally, the time series status data includes: voltage change data within the target time period, current change data within the target time period, temperature change data within the target time period, state of charge change data within the target time period, and health state change data within the target time period. The target time period includes: multiple time points.

[0060] Specifically, the battery includes: multiple single cells. The voltage change data includes at least one of the following: the maximum voltage among the voltages of the multiple single cells, the minimum voltage among the voltages of the multiple single cells, and the total battery voltage. The current change data includes at least one of the following: the highest temperature among the temperatures of the multiple single cells, the lowest temperature among the temperatures of the multiple single cells, and the temperature gradient. The state of charge change data includes at least one of the following: the true SOC and the displayed SOC. The health state change data includes: SOH. The total battery voltage is the sum of the voltages presented after all single cells in the battery are combined in series, parallel, or series-parallel. The temperature gradient is used to indicate the temperature difference between different positions inside the battery.

[0061] S202. Input the time series status data into the trained thermal runaway early warning model to obtain the risk assessment result.

[0062] Among them, the thermal runaway early warning model is used to detect whether there is a risk of thermal runaway in the battery. The thermal runaway early warning model includes: a long short-term memory (LSTM) module and a soft decision tree (SDT) module. The long short-term memory network module is used to extract feature data related to the thermal runaway risk, and the soft decision tree module is used to determine whether the battery has a thermal runaway.

[0063] In a possible implementation, the time series state data can be input into the long short-term memory network module to obtain the risk feature vector of the battery. The risk feature vector is used to indicate the feature data related to the thermal runaway risk. Then, the risk feature vector can be input into the soft decision tree module to obtain the risk assessment result.

[0064] It can be understood that by combining the long short-term memory network model with the soft decision tree, while capturing higher-level abstract features, the performance and interpretability of the thermal runaway early warning model can be balanced, making the risk assessment result more accurate and interpretable.

[0065] In the embodiment of the present application, the long short-term memory network module includes: an input layer, a hidden layer (i.e., the long short-term memory network layer), and a fully connected layer. The fully connected layer is used to input the risk feature vector into the soft decision tree module. Among them, the dimension of the fully connected layer of the long short-term memory network module is the same as the dimension of the input layer of the soft decision tree module.

[0066] It should be understood that the long short-term memory network module is obtained based on the long short-term memory network model, and the soft decision tree module is obtained based on the soft decision tree model.

[0067] In this way, the long short-term memory network module and the soft decision tree module can be directly connected and can be trained as a whole, thereby simplifying the model architecture of the thermal runaway early warning model and improving the calculation efficiency of the thermal runaway early warning model.

[0068] In a possible design, the time series state data can be data-converted through the input layer to obtain the target sequence state data. The data format of the target sequence state data is the data format that the long short-term memory network module can process. Then, the target sequence state data can be data-processed through the hidden layer to obtain the hidden state feature vector. Then, the hidden state feature vector can be data-mapped through the fully connected layer to obtain the risk feature vector of the battery.

[0069] In the embodiments of the present application, the hidden state feature vector includes: a plurality of hidden state sub-vectors, and the plurality of hidden state sub-vectors include at least one of the following: voltage feature vector, current feature vector, temperature feature vector, SOC feature vector, and SOH feature vector. The voltage feature vector includes at least one of the following: voltage mean vector, voltage variance vector, voltage range vector, voltage change rate vector, voltage maximum vector, and voltage minimum vector. The current feature vector includes at least one of the following: current mean vector, current variance vector, current range vector, current change rate vector, current maximum vector, and current minimum vector. The temperature feature vector includes at least one of the following: temperature mean vector, temperature variance vector, temperature range vector, temperature change rate vector, temperature maximum vector, and temperature minimum vector.

[0070] It should be noted that the types of the plurality of hidden state sub-vectors are different.

[0071] Optionally, the hidden layer includes: a forgetting gate, an input gate, and an output gate. The forgetting gate is used to retain or discard the hidden state sub-vector (i.e., cell state) of the previous moment at the current moment. The input gate is used to determine the hidden state sub-vector to be stored, and the output gate is used to determine the hidden state sub-vector to be output.

[0072] Specifically, the target sequence state data can be processed through the forgetting gate to obtain the output data of the forgetting gate. The target sequence state data can be processed through the input gate to obtain the candidate state vector to be updated. Then, the output data and the candidate state vector to be updated can be processed through the forgetting gate and the input gate to obtain the updated state vector. Then, the updated state vector and the target sequence state data can be processed through the output gate to obtain the hidden state feature vector.

[0073] Exemplarily, the calculation formula of the forgetting gate satisfies Formula 1.

[0074] Formula 1.

[0075] Where, f t is the output data of the forgetting gate, σ represents the sigmoid activation function, W f is the weight matrix of the forgetting gate, h t-1 is the hidden state sub-vector of the previous moment (i.e., the output result of the neuron), U f is the weight matrix of the forgetting gate, x t is the time series state data, b f is the bias term of the forgetting gate.

[0076] It should be noted that f t is a constant greater than or equal to 0 and less than or equal to 1.

[0077] Exemplarily, the input gate calculation formula satisfies Formula 2 and Formula 3.

[0078] Formula 2.

[0079] Formula 3.

[0080] where, i t is the update parameter of the target sequence state data, W i is the weight matrix of the input gate signal, b i is the bias term of the input gate, is the state candidate vector to be updated (i.e., the candidate cell state to be updated), tanh() is the hyperbolic tangent function, W c is the weight matrix of the state candidate vector, b c is the bias term of the state candidate vector.

[0081] Optionally, the update parameter is used to control whether the target sequence state data is updated.

[0082] It should be noted that if the update parameter is 1, it is determined to update and control the target sequence state data; if the update parameter is 0, it is determined not to update and control the target sequence state data.

[0083] Exemplarily, the calculation formula of the updated state candidate vector satisfies Formula 4.

[0084] Formula 4.

[0085] where, C t is the updated state vector (i.e., the updated candidate cell state), C t-1 is the state vector of the previous moment.

[0086] Exemplarily, the calculation formula of the output gate satisfies Formula 5 and Formula 6.

[0087] Formula 5.

[0088] Formula 6.

[0089] where, o t is the output gate signal, W o is the weight matrix of the output gate, b o is the bias term of the output gate, h t is the hidden state sub-vector (i.e., the battery feature output value at the current moment).

[0090] It should be noted that, o t is an output value greater than or equal to 0 and less than or equal to 1.

[0091] It should be noted that the hidden state feature vector is a comprehensive representation of the current input and all previous inputs, and is the feature vector output by the hidden layer of the long short-term memory network.

[0092] Exemplarily, the hidden state feature vector satisfies Formula Seven.

[0093] Formula Seven.

[0094] Wherein, H is the hidden state feature vector, h1 is the hidden state sub-vector at the first moment, h2 is the hidden state sub-vector at the second moment, and h t is the hidden state sub-vector at the t-th moment.

[0095] Optionally, the long short-term memory network module further includes: an attention mechanism layer.

[0096] In another possible design, the hidden state feature vector can be processed through the attention mechanism layer to obtain an attention vector set. Subsequently, the attention vector set can be mapped through a fully connected layer to obtain the risk feature vector of the battery.

[0097] It should be noted that the risk feature vector of the battery is a feature vector with the same dimension as the input layer of the soft decision tree module, that is, a low-dimensional feature vector.

[0098] In the embodiment of the present application, the attention vector set includes: a plurality of attention sub-vectors.

[0099] Optionally, the attention weight parameter of the hidden state feature vector can be determined through the attention mechanism layer. Subsequently, for each hidden state sub-vector among the plurality of hidden state sub-vectors, each hidden state sub-vector can be weighted and aggregated based on the attention weight parameter to obtain an attention sub-vector, so as to obtain the attention vector set.

[0100] Exemplarily, the attention vector satisfies Formula Eight.

[0101] Formula Eight.

[0102] Wherein, is any one of the plurality of attention sub-vectors, is the attention weight parameter, and h i is any one of the plurality of hidden state sub-vectors.

[0103] It should be noted that the attention vector set can be obtained according to different types of time series data.

[0104] Exemplarily, the attention vector set satisfies Formula Nine.

[0105] Formula Nine.

[0106] Among them, is the attention vector set, is the attention sub-vector at the first moment, is the attention sub-vector at the second moment, is the attention sub-vector at the t-th moment.

[0107] Exemplarily, the attention weight parameters satisfy Formula Ten and Formula Eleven.

[0108] Formula Ten.

[0109] Formula Eleven.

[0110] Among them, e i is the similarity score.

[0111] Exemplarily, the similarity score satisfies Formula Twelve.

[0112] Formula Twelve.

[0113] Among them, W h is the weight matrix of the similarity score, and b h is the bias term of the similarity score.

[0114] It should be noted that the soft decision tree is a deep neural network with a binary tree structure logically. The soft decision tree module is composed of multiple differentiable soft decision trees, and each tree can implement probabilistic path selection through the Sigmoid activation function. The number of soft decision trees in the soft decision tree module in this application is not limited. For example, the number of soft decision trees can be 5, 10, 50, 100, or 101.

[0115] It should be understood that the number of soft decision trees in the soft decision tree module needs to balance the accuracy and computational cost of the model.

[0116] When the number of soft decision trees is 100 (i.e., a medium number), it can ensure a certain degree of diversity without being too time-consuming. The maximum depth of a single tree is set to 10 to balance the complexity and generalization ability of the model.

[0117] In the embodiments of this application, the soft decision tree module includes: a plurality of preset evaluation paths and a plurality of initial evaluation paths. One preset evaluation path corresponds to one initial evaluation path. The preset evaluation path is used to indicate the probability distribution of the battery having a thermal runaway risk, and the initial evaluation path is used to indicate the static hierarchical structure of the soft decision tree module.

[0118] In another possible design, for each preset evaluation path, initial probability evaluation information corresponding to the preset evaluation path can be obtained based on the risk feature vector and the preset evaluation path, so as to obtain the initial probability evaluation information corresponding to multiple preset evaluation paths. The initial probability evaluation information is used to indicate the probability of the battery having a thermal runaway risk, and one preset evaluation path corresponds to one initial probability evaluation information. Then, based on the multiple initial probability evaluation information, a risk evaluation result can be obtained.

[0119] Among them, the preset evaluation path includes multiple evaluation factors and multiple probability weights, and the initial evaluation path includes multiple branch nodes and multiple decision edges. The evaluation factor is used to indicate the condition for the battery to have a thermal runaway risk, the probability weight is used to indicate the influence degree of the evaluation factor on the battery having a thermal runaway risk, the branch node is used to indicate the evaluation factor, and the decision edge is used to indicate the probability weight. One evaluation factor corresponds to multiple probability weights, one evaluation factor corresponds to one branch node, and one probability weight corresponds to one decision edge.

[0120] Optionally, the initial probability evaluation information corresponding to the preset evaluation path can be obtained based on the risk feature vector, multiple evaluation factors, and multiple probability weights.

[0121] It should be noted that the present application does not limit the acquisition method of the probability weight. For example, the probability weight corresponding to each evaluation factor can be set in advance. Another example is that the probability weight corresponding to each evaluation factor can be obtained based on the weight generator of the branch node.

[0122] It can be understood that since the preset evaluation path of the soft decision tree model includes multiple evaluation factors and multiple probability weights, each preset evaluation path corresponds to a clear feature combination logic. In this way, the meaning of each preset evaluation path is clear and interpretable, and when the risk evaluation result is used to indicate that the battery has a thermal runaway, the reason for the battery to have a thermal runaway risk can be determined.

[0123] Optionally, the risk evaluation result is used to indicate whether the battery has a thermal runaway.

[0124] Optionally, for each preset evaluation path, the risk feature vector can be input into the sub-classifier of the branch node, and multiple evaluation probabilities can be obtained through the sub-classifier and the evaluation factors. The risk feature vector can be input into the weight generator of the branch node to obtain multiple probability weights corresponding to each evaluation factor, and one evaluation probability corresponds to one probability weight. Then, based on the multiple evaluation probabilities and multiple probability weights, the initial probability evaluation information corresponding to the preset evaluation path can be obtained to obtain the initial probability evaluation information corresponding to multiple preset evaluation paths. Then, based on the multiple initial probability evaluation information, the initial probability evaluation information with the highest probability of the battery having a thermal runaway risk can be determined as the risk evaluation result.

[0125] Alternatively, if there is an initial probability assessment information with a probability of thermal runaway risk greater than the preset probability threshold among multiple initial probability assessment information, it can be determined that the risk assessment result is that the battery has a thermal runaway; if there is no initial probability assessment information with a probability of thermal runaway risk greater than the preset probability threshold among multiple initial probability assessment information, it can be determined that the risk assessment result is that the battery has no thermal runaway.

[0126] In this way, by obtaining the initial probability assessment information of each preset evaluation path, the probability results of multiple preset evaluation paths of the soft decision tree model can be obtained. That is to say, the probability distribution of the thermal runaway risk of the battery can be obtained. In this way, the risk assessment result can be obtained more accurately according to the probability distribution of the thermal runaway risk of the battery.

[0127] It can be understood that the long short-term memory network model can capture the time dependence of battery parameters, the time attention mechanism can focus on key time steps (such as the stage of sudden temperature rise or sudden voltage rise), and the fully connected layer can adjust the feature dimension to adapt to the input of the soft decision tree model. The soft decision tree model adjusts the contribution degree of each tree through weight parameters, so that the final prediction result is a weighted integrated output, that is, an interpretable hierarchical decision result is generated based on the power battery feature vector output by the long short-term memory network.

[0128] Based on the above technical solution, by obtaining the time series state data of the battery and inputting the time series state data into the trained thermal runaway warning model, the risk assessment result can be obtained. The thermal runaway warning model includes: a long short-term memory network module and a soft decision tree module. Since the long short-term memory network module can capture the long-term trend and periodic changes in the time series state data, it can better determine the feature data related to the thermal runaway risk in the time series state data. The soft decision tree module can smooth the decision boundary and reduce the sensitivity to noise, so as to more accurately determine the probability of the battery having a thermal runaway risk, and then more accurately determine whether the battery has a thermal runaway. And, since the data processing process of the soft decision tree module can perform interpretable hierarchical reasoning, the interpretability of the thermal runaway warning model can be improved.

[0129] As Figure 3 shown, combined with Figure 2 , the method includes the following steps: S301. Obtain a training set, a first model loss, a second model loss, and an initial thermal runaway warning model.

[0130] Among them, the first model loss is used to adjust the accuracy of the soft decision tree module, and the second model loss is used to adjust the accuracy of the long short-term memory network module.

[0131] In a possible implementation, initial battery state data can be obtained. After that, battery state data for a preset period can be extracted from the initial battery state data. After that, the battery state data for the preset period can be subjected to data cleaning processing to obtain first state data. After that, the first state data can be subjected to data normalization processing to obtain second state data. After that, based on the sliding time window algorithm, the second state data can be processed to obtain a training set.

[0132] In the embodiments of the present application, the thermal runaway risk assessment device includes: an operation unit and a time series database. The operation unit is used to estimate the SOC value and the SOH value of the battery, and the time series database is used to store data.

[0133] Optionally, data can be collected through sensors such as a voltage sensor, a current sensor, and a temperature sensor to obtain raw battery state data. After that, the raw state data can be transmitted to the operation unit to obtain initial battery state data. After that, the initial battery state data can be uploaded to the time series database for storage.

[0134] It should be noted that the present application does not limit the specific processing methods of data cleaning and data normalization. For example, data cleaning includes: handling missing values and handling outliers. Data normalization includes: the maximum-minimum normalization algorithm. Among them, the method for handling missing values includes: the linear interpolation algorithm, and the method for handling outliers includes: the 3- sigma principle.

[0135] Optionally, the battery state data for the preset period can be processed based on the linear interpolation algorithm to obtain the interpolated battery state data. After that, based on the 3-sigma principle, the interpolated initial battery state data can be processed to obtain the first state data.

[0136] It should be noted that it can be assumed that there is a linear relationship between adjacent data points to estimate the missing value, and it is assumed that in the battery state data for the preset period, the missing value is located between the first valid value and the second valid value.

[0137] Among them, the interpolated battery state data includes: multiple battery parameters.

[0138] Specifically, the missing points can be determined based on the battery status data of a preset time period. After that, based on the missing points and the battery status data of the preset time period, the first valid value and the second valid value can be determined. After that, based on the first valid value and the second valid value, the missing value corresponding to the missing point can be obtained to obtain the interpolated battery status data. After that, based on the interpolated battery status data, the mean value (i.e., μ) of multiple battery parameters and the standard deviation (i.e., σ) of multiple battery parameters can be obtained. After that, based on the mean value of multiple battery parameters and the standard deviation of multiple battery parameters, the upper bound value (i.e., μ + 3σ) and the lower bound value (i.e., μ - 3σ) of multiple battery parameters can be determined. After that, based on multiple battery parameters, the upper bound value of multiple battery parameters, and the lower bound value of multiple battery parameters, the target outliers can be determined, and the target outliers are the battery parameters among multiple battery parameters that are greater than the upper bound value or less than the lower bound value. After that, the target outliers can be removed to obtain the first status data.

[0139] Exemplarily, among multiple battery parameters, battery parameter 1 is 25, battery parameter 2 is 28, battery parameter 3 is 22, battery parameter 4 is 26, battery parameter 5 is 27, battery parameter 6 is 23, battery parameter 7 is 24, battery parameter 8 is 29, battery parameter 9 is 30, and battery parameter 10 is 50. After that, calculations can be performed on multiple battery parameters to obtain the mean value of multiple battery parameters as 27.4 and the standard deviation of multiple battery parameters as 5.3. Then, the upper bound value of multiple battery parameters is 43.3, and the lower bound value of multiple battery parameters is 11.5. Since battery parameter 10 is 50, which is greater than the upper bound value 43.3, battery parameter 10 is the target outlier.

[0140] Exemplarily, the missing value satisfies Equation XIII.

[0141] Equation XIII.

[0142] Wherein, x i is the missing value, x i-1 is the first valid value, x i+1 is the second valid value, t i is the timestamp of the missing value, t i-1 is the timestamp of the first valid value, t i+1 is the timestamp of the second valid value.

[0143] Exemplarily, the mean value satisfies Equation XIV.

[0144] Equation XIV.

[0145] Wherein, x j is any one of the multiple battery parameters of the interpolated battery status data, and N is the number of the multiple battery parameters in the interpolated battery status data.

[0146] Exemplarily, the standard deviation satisfies Formula XV.

[0147] Formula XV.

[0148] It should be noted that the preset time period is M consecutive unit time periods. The present application does not limit the specific value of M. For example, M can be a positive integer greater than or equal to 3. The present application does not limit the unit time period. For example, the unit time period can be 10 seconds, 30 seconds, 1 minute, 5 minutes, 30 minutes, 1 hour, 4 hours, or 12 hours.

[0149] Optionally, the first state data can be linearly transformed based on the maximum-minimum normalization algorithm to obtain the second state data.

[0150] Specifically, the first state data can be mapped to the interval [0, 1] for normalization processing to obtain the second state data.

[0151] In this way, the adverse effects caused by singular sample data can be eliminated.

[0152] Exemplarily, the second state data satisfies Formula XVI.

[0153] Formula XVI.

[0154] Wherein, is the second state data, x is any one of the multiple battery parameters of the first state data, x min is the smallest battery parameter among the multiple battery parameters of the first state data, x max is the largest battery parameter among the multiple battery parameters of the first state data.

[0155] Optionally, the window size and time step of the time window can be determined. Then, based on the second state data, time window, and time step, an input sample can be obtained. Then, the input sample can be divided according to the division ratio to obtain a training set, a validation set, and a test set.

[0156] Wherein, the input sample is a matrix (N, d), where N is the window size and d is the number of features.

[0157] Specifically, the second state data within the first preset time can be intercepted through a time window to obtain the first intercepted data. After an interval of the second preset time, the battery time series data within the first preset time is intercepted again through a sliding window to obtain the second intercepted data. This is iteratively performed until the sliding window slides from the first time step to the last time step of the second state data to obtain the target intercepted data. Subsequently, the intercepted data of X consecutive unit time periods in sequence in time can be extracted from the target intercepted data as the model input items. Subsequently, the intercepted data of Y consecutive unit time periods in sequence in time, which is located after the X unit time periods, can be extracted from the target intercepted data as the model output items. Subsequently, the model input items and the model output items can be used as input samples.

[0158] It should be noted that this application places no restrictions on the first preset time and the second preset time. For example, the first preset time can be 5 minutes (min), 10 min, 15 min, 20 min, or 25 min, and the second preset time can be 1 min, 2 min, 3 min, 4 min, or 5 min.

[0159] It should be noted that both X and Y are positive integers.

[0160] It should be noted that this application places no restrictions on the division ratio. For example, the division ratio can be 14:3:3, and 70% of the time series data can be used as the training set, 15% of the time series data can be used as the validation set, and 15% of the time series data can be used as the test set.

[0161] In a possible implementation, an initial long short-term memory network model and an initial soft decision tree model can be obtained. Subsequently, based on the initial long short-term memory network model and the initial soft decision tree model, an initial thermal runaway warning model can be obtained.

[0162] Optionally, the initial thermal runaway warning model is a pruned thermal runaway warning model.

[0163] Optionally, the pruning threshold of the initial long short-term memory network module can be obtained. Subsequently, structural pruning processing can be performed on the initial long short-term memory network model based on the pruning threshold to obtain the pruned long short-term memory network model. Subsequently, pruning processing can be performed on the soft decision tree model to obtain the pruned soft decision tree model. Subsequently, based on the pruned long short-term memory network model and the pruned soft decision tree model, a pruned thermal runaway warning model can be obtained.

[0164] In a possible design, the pruned long short-term memory network model can be initially trained to obtain the long short-term memory network model after the initial training. The long short-term memory network model after the initial training includes: long short-term memory network model parameters. Subsequently, multiple floating-point numbers in the long short-term memory network model parameters can be converted into low-precision integers through quantization-aware training to obtain the optimized long short-term memory network model.

[0165] Exemplarily, the low-precision integer satisfies Formula XVII.

[0166] Formula XVII.

[0167] Wherein, is the low-precision integer, is the quantization step, is any value among the multiple floating-point numbers in the long short-term memory network model parameters, is the minimum value among the multiple floating-point numbers in the long short-term memory network model parameters.

[0168] Exemplarily, the quantization step satisfies Formula XVIII.

[0169] Formula XVIII.

[0170] Wherein, is the maximum value of the floating-point numbers, and b is the quantization bit number.

[0171] In some embodiments, the optimized long short-term memory network model can be evaluated.

[0172] In this way, the generalization performance and performance stability of the pruned long short-term memory network model can be ensured.

[0173] It should be understood that in combination with S202, the soft decision tree model includes: multiple initial evaluation paths, and the initial evaluation path includes: a branch structure.

[0174] In another possible design, based on multiple branch structures, the similarity of the multiple branch structures can be determined (i.e., ). Subsequently, based on the similarity of the multiple branch structures, multiple groups of similar structure groups can be determined. A group of similar structure groups includes at least two branch structures with a similarity less than a preset similarity threshold. Subsequently, the branch structures in each group of similar structure groups can be fused to obtain multiple fused branch structures. Subsequently, based on the multiple fused branch structures, the pruned thermal runaway warning model can be obtained.

[0175] It should be understood that in combination with S202, the branch structure includes at least one of the following: a branch node and a decision edge. The branch node is used to indicate an evaluation factor, and the decision edge is used to indicate a probability weight.

[0176] It should be noted that this application does not limit the preset similarity threshold. For example, the preset similarity threshold can be 1.

[0177] Exemplarily, if the similarity between branch node 1 and branch node 2 is less than 0.1, then branch node 1 and branch node 2 are fused to obtain a fused branch node.

[0178] In this way, by fusing the branch structures in each group of similar structure groups to obtain multiple fused branch structures, pruning of the thermal runaway warning model can be achieved. And, since there are at least two branch structures in a group of similar structure groups with a similarity less than the preset similarity threshold, and the branch structure includes branch nodes and decision edges, while lightweight processing is performed on the branch nodes and decision edges, the accuracy of the thermal runaway warning model can be maintained.

[0179] It can be understood that by performing structured pruning on the long short-term memory network model, neurons and connections with little impact on thermal runaway prediction can be removed. By fusing the branch structures in each group of similar structure groups to obtain multiple fused branch structures, repeated calculations can be reduced. Further, by lightweighting the long short-term memory network model and the soft decision tree model, while ensuring the prediction accuracy of the model, the network architecture can be streamlined, the number of parameters and the amount of calculation can be reduced, so that the trained lightweight model can be adapted to the in-vehicle terminal embedded system.

[0180] S302. Train the initial thermal runaway warning model based on the training set, the first model loss, and the second model loss to obtain a trained thermal runaway warning model.

[0181] In a possible implementation, the parameters of the initial thermal runaway warning model can be initialized, and an optimizer for the initial thermal runaway warning model can be defined. Then, based on the first model loss and the second model loss, an objective loss function can be obtained. Then, the training set can be input into the initial thermal runaway warning model so that the initial thermal runaway warning model can be continuously optimized through the training set and the objective loss function to obtain a trained thermal runaway warning model.

[0182] Optionally, the objective loss function satisfies Equation XIX.

[0183] Equation XIX.

[0184] Wherein, is the first weight parameter, is the second weight parameter, and are used to balance the training of the long short-term memory network module and the soft decision tree module. L task is the cross-entropy loss (i.e., the first model loss), Ltree is the tree structure regularization (i.e., the second model loss).

[0185] Exemplarily, the first model loss satisfies Formula 20.

[0186] Formula 20.

[0187] Where n is the number of samples, y i is the real data, y i ' is the predicted value of the long short-term memory network module.

[0188] Exemplarily, the second model loss satisfies Formula 21.

[0189] Formula 21.

[0190] in, is the hyperparameter of the soft decision tree model. left is the probability of selecting the first node, p right is the probability of selecting the second node.

[0191] It should be noted that the soft decision tree module includes multiple soft decision tree levels, and the first node and the second node are nodes located in the same soft decision tree level.

[0192] Optionally, the training set can be subjected to feature extraction through a long short-term memory network module to obtain an extracted feature vector. The extracted feature vector can then be input into a soft decision tree module, and the initial output result can be obtained through the differentiable hierarchical decision path of the soft decision tree module. The model loss of the initial thermal runaway warning model can then be calculated, and the gradient of each layer of parameters in the initial thermal runaway warning model can be reversely calculated using the chain rule. The model parameters can then be updated based on the gradient and the optimizer to obtain the initial thermal runaway warning model after initial training. The initial thermal runaway warning model after initial training can then be trained, and the training can be iteratively repeated until the initial thermal runaway warning model converges, thereby obtaining a trained thermal runaway warning model.

[0193] It should be noted that this application does not limit the optimizer. For example, the optimizer may be an Adaptive Moment Estimation (Adam) optimizer.

[0194] It should be noted that this application does not limit the initial learning rate and number of iterations during the training process of the thermal runaway warning model. For example, the initial learning rate can be 0.001 and the number of iterations can be 100.

[0195] It should be noted that when the long short-term memory network model is used to extract features from the training set to obtain the extracted feature vectors, the extracted feature vectors should be closely related to thermal runaway and representative, such as battery temperature, voltage, current, SOC, and SOH.

[0196] It should be noted that this application does not limit the number of long short-term memory network layers of the long short-term memory network model, the number of hidden units of the long short-term memory network model, the dimension of the extracted feature vectors, or the data sampling frequency. For example, the number of long short-term memory network layers can be 2, 3, or 4, the dimension of the feature vectors can be 9, 10, or 11, the number of hidden units can be 32, 64, or 128. The sensor sampling data frequency can be 1 Hertz (HZ), 10HZ, or 100HZ.

[0197] It should be noted that to ensure model performance while reducing computational tasks, the long short-term memory network layer can be set to 2 layers. To capture complex non-linear relationships and avoid model overfitting, the number of hidden units can be 64.

[0198] It should be noted that the data sampling frequency needs to be set in combination with the timeliness requirements of thermal runaway warning. It is necessary to ensure that sufficient time-dependent relationships can be captured while avoiding introducing too much noise. Taking the sensor sampling data frequency of 1HZ as an example, by selecting 300 time steps, the critical stages of sudden temperature rise (such as temperature rise greater than 1 degree Celsius per second (℃ / s)) and sudden voltage drop (such as voltage drop greater than 50 millivolts per second (mV / s)) can be captured.

[0199] It can be understood that the Adam optimizer can better optimize the parameters, and the trained thermal runaway warning model can better learn how to map the data to the correct category, improving the accuracy of the thermal runaway warning model prediction.

[0200] In some embodiments, a test set can be obtained, and the trained thermal runaway warning model can be evaluated based on the test set.

[0201] In another possible implementation, the initial thermal runaway warning model can be trained based on the training set, the first model loss, and the second model loss to obtain the first model parameters. Then, the training set and the first model parameters can be sent to the server, where an unpruned thermal runaway warning model is deployed. The server is used to optimize the first model parameters based on the training set and the unpruned thermal runaway warning model. Then, the second model parameters can be received. The second model parameters are the optimized first model parameters. Then, based on the second model parameters, the trained thermal runaway warning model can be obtained.

[0202] It should be noted that this application does not limit the number of vehicles. For example, the number of vehicles can be 1 or multiple.

[0203] Exemplarily, as Figure 4 shown, the unpruned thermal runaway warning model is deployed in the cloud, and the vehicle side (i.e., the evaluation device for thermal runaway risk) includes multiple vehicles, and a lightweight model (i.e., the pruned thermal runaway warning model) is deployed in the embedded system of each vehicle.

[0204] It should be noted that the multiple vehicles in this application can be the same or different, and this application does not limit the types of multiple vehicles.

[0205] The vehicle can perform data collection to obtain a real-time data set, and input the real-time data set into the lightweight model for model training. The cloud (i.e., the server) can obtain the initial model parameters. Then, the cloud can send the initial model parameters to all vehicles. Each vehicle can train the lightweight model based on the initial model parameters and the real-time data set to obtain updated model parameters (i.e., the first model parameters). Then, each vehicle can send the real-time data set and the updated model parameters to the cloud.

[0206] It should be noted that the content of the real-time data set is not shared among the vehicles.

[0207] Then, the cloud can obtain the global model parameters based on the updated model parameters sent by all vehicles; the cloud can obtain the global model training set based on the real-time data sets sent by all vehicles. Then, the cloud can train the unpruned thermal runaway warning model based on the global model parameters and the global model training set to obtain updated global model parameters. Then, the updated global model parameters can be sent. Then, the cloud sends the updated global model parameters (i.e., the second model parameters) to all vehicles.

[0208] Exemplarily, the global model parameters satisfy Formula XXII.

[0209] XXII.

[0210] Wherein, is the global model parameter, is the updated model parameter.

[0211] For example, Vehicle 1 can perform data collection to obtain Real-time Data Set 1. Then, Real-time Data Set 1 can be input into Lightweight Model 1 for training to obtain updated model parameters . Then, Vehicle 1 can send Real-time Data Set 1 and the updated model parameters to the cloud ; Vehicle 2 can perform data collection to obtain Real-time Data Set 2. Then, Real-time Data Set 2 can be input into Lightweight Model 2 for training to obtain updated model parameters After that, vehicle 2 can send the real-time dataset 2 and the updated model parameters to the cloud ; Vehicle N can collect data to obtain the real-time dataset N. After that, the real-time dataset N can be input into the lightweight model N for training to obtain the updated model parameters After that, vehicle N can send the real-time dataset N and the updated model parameters to the cloud After that, the cloud can be based on the real-time dataset 1, the updated model parameters , the real-time dataset 2, the updated model parameters , the real-time dataset N and the updated model parameters Train the unpruned thermal runaway warning model to obtain the updated global model parameters After that, the updated global model parameters can be sent to all vehicles .

[0212] In this way, by deploying the optimized lightweight model in the in-vehicle terminal embedded system, fast thermal runaway prediction can be achieved. By uploading the real-time collected data to the cloud and the cloud performing high-precision dynamic prediction, vehicle-cloud data synchronization can be achieved, and the balance between fast response and high-precision warning can be achieved.

[0213] In this way, by deploying the pruned thermal runaway warning model in the thermal runaway risk assessment device, the thermal runaway risk assessment device can quickly predict whether the battery has a thermal runaway. By deploying the unpruned thermal runaway warning model on the server and tuning the first model parameters through the server, while improving the speed of predicting whether the battery has a thermal runaway, the prediction result of the thermal runaway warning model can be made more accurate.

[0214] Based on the above technical solution, by obtaining the training set, the first model loss, the second model loss, and the initial thermal runaway warning model, and training the initial thermal runaway warning model based on the training set, the first model loss, and the second model loss, the trained thermal runaway warning model is obtained. The first model loss is used to adjust the accuracy of the soft decision tree module, and the second model loss is used to adjust the accuracy of the long short-term memory network module. In this way, the training direction of the thermal runaway warning model can be effectively guided, preventing the thermal runaway warning model from overfitting on the training data, improving the generalization ability of the thermal runaway warning model, so that the thermal runaway warning model can have good robustness when facing different types of input data, and the performance of the thermal runaway warning model will not decline due to the overuse of some paths.

[0215] Next, a specific example is used to introduce the thermal runaway risk assessment method provided by the embodiments of the present application. As Figure 5As shown in the figure, the thermal runaway warning model includes a soft decision tree module, an input layer, a hidden layer, an attention mechanism layer, and a fully connected layer. Among them, the time series status data can be obtained through the input layer. The time series status data includes: voltage change data within the target time period, current change data within the target time period, temperature change data within the target time period, state of charge change data within the target time period, and health status change data within the target time period. Then, the input layer can perform data conversion on the time series status data to obtain target sequence status data, and the data format of the target sequence status data is the data format that can be processed by the long short-term memory network module. Then, the hidden layer can perform data processing on the target sequence status data to obtain a hidden state feature vector. Then, the attention mechanism layer and the fully connected layer can perform data processing on the hidden state feature vector to obtain a risk feature vector of the battery. Then, for each preset evaluation path among the multiple preset evaluation paths of the soft decision tree module, based on the risk feature vector and the preset evaluation path, the initial probability evaluation information corresponding to the preset evaluation path can be obtained to obtain the initial probability evaluation information corresponding to the multiple preset evaluation paths. For example, the initial probability evaluation information corresponding to the multiple preset evaluation paths can be initial probability evaluation information 1 (i.e., P1), initial probability evaluation information 2 (i.e., P2), ……, initial probability evaluation information m - 1 (i.e., P m-1 ), and initial probability evaluation information m (i.e., P m ). Then, based on the multiple initial probability evaluation information, a risk assessment result can be obtained.

[0216] The above mainly introduces the solution provided by the embodiments of the present application from the perspective of the method. To implement the above functions, the thermal runaway risk assessment device includes the corresponding hardware structure and / or software module for executing each function. Those skilled in the art should easily realize that, combined with the units and algorithm steps of each example described in the embodiments disclosed in this article, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a certain function is executed in the way of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0217] The embodiments of the present application can divide the functional modules of the thermal runaway risk assessment device according to the above method. For example, the thermal runaway risk assessment device can include each functional module corresponding to each functional division, or two or more functions can be integrated into one processing module. The above integrated module can be implemented in the form of hardware or in the form of a software functional module. It should be noted that the division of modules in the embodiments of the present application is illustrative, only a logical functional division, and there can be other division methods in actual implementation.

[0218] Figure 6 is a schematic structural diagram of an apparatus for assessing the risk of thermal runaway shown according to an exemplary embodiment. Referring to Figure 6 , the apparatus for assessing the risk of thermal runaway includes an acquisition module 601 and a processing module 602.

[0219] The acquisition module 601 is configured to acquire time-series state data of the battery, and the time-series state data is used to indicate the change of the battery parameters of the battery at different time points. The processing module 602 is configured to input the time-series state data into the trained thermal runaway early warning model to obtain a risk assessment result. The thermal runaway early warning model is used to detect whether there is a thermal runaway risk in the battery. The thermal runaway early warning model includes: a long short-term memory network module and a soft decision tree module. The long short-term memory network module is used to extract feature data related to the thermal runaway risk, and the soft decision tree module is used to determine whether the battery has a thermal runaway.

[0220] Regarding the apparatus in the above embodiments, the specific manners in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated herein.

[0221] Figure 7 is a schematic structural diagram of another apparatus for assessing the risk of thermal runaway shown according to an exemplary embodiment. As Figure 7 shown, the apparatus for assessing the risk of thermal runaway includes, but is not limited to: a processor 701 and a memory 702.

[0222] Among them, the above-mentioned memory 702 is used to store the executable instructions of the above-mentioned processor 701. It can be understood that the above-mentioned processor 701 is configured to execute instructions to implement the method for assessing the risk of thermal runaway in the above embodiments.

[0223] It should be noted that those skilled in the art can understand that Figure 7 the structure of the apparatus for assessing the risk of thermal runaway shown in Figure 7 does not constitute a limitation on the apparatus for assessing the risk of thermal runaway. The apparatus for assessing the risk of thermal runaway can include more or fewer components than shown, or combine certain components, or arrange different components.

[0224] The processor 701 is the control center of the thermal runaway risk assessment device, connecting various parts of the entire thermal runaway risk assessment device through various interfaces and lines. By running or executing software programs and / or modules stored in the memory 702, and calling data stored in the memory 702, it executes various functions of the thermal runaway risk assessment device and processes data, thereby monitoring the thermal runaway risk assessment device as a whole. The processor 701 may include one or more processing units. Optionally, the processor 701 may integrate an application processor and a modem processor. Among them, the application processor mainly processes the operating system, user interface, application programs, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor may not be integrated into the processor 701 either.

[0225] The memory 702 can be used to store software programs and various data. The memory 702 mainly includes a program storage area and a data storage area. Among them, the program storage area can store the operating system, application programs required by at least one functional module (such as the determination unit, processing unit, etc.). In addition, the memory 702 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage devices.

[0226] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided, such as the memory 702 including instructions. The above instructions can be executed by the processor 701 of the thermal runaway risk assessment device to implement the method in the above embodiment.

[0227] In actual implementation, Figure 6 the functions of the acquisition module 601 and the processing module 602 in Figure 7 can both be implemented by the processor 701 in

[0228] calling the computer program stored in the memory 702. The specific execution process can refer to the description of the method part in the above embodiment, which will not be elaborated here.

[0229] In an exemplary embodiment, the embodiment of the present application further provides a computer program product including one or more instructions, and the one or more instructions can be executed by a processor 701 of an apparatus for evaluating the risk of thermal runaway to complete the method in the above embodiment.

[0230] It should be noted that when the instructions in the above computer-readable storage medium or the one or more instructions in the computer program product are executed by the processor of the apparatus for evaluating the risk of thermal runaway, each process of the above method embodiment is implemented, and the same technical effects as the above method can be achieved. To avoid repetition, it will not be elaborated here.

[0231] Through the description of the above embodiments, those skilled in the art can clearly understand that for the convenience and brevity of description, only the above division of each functional module is used as an example. In actual applications, the above functions can be allocated to different functional modules according to needs, that is, the internal structure of the apparatus is divided into different functional modules to complete all or part of the functions described above.

[0232] In several embodiments provided by the present application, it should be understood that the disclosed apparatus and method can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of modules or units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another apparatus, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces. The indirect coupling or communication connection of the apparatus or unit can be in an electrical, mechanical or other form.

[0233] The unit described as a separated component may or may not be physically separated, and the component displayed as a unit may be a physical unit or multiple physical units, that is, it may be located in one place, or may be distributed to multiple different places. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0234] In addition, each functional unit in various embodiments of the present application can be integrated in one processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0235] When the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solution of the embodiments of the present application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions for causing a device (which can be a single-chip microcomputer, a chip, etc.) or a processor to execute all or part of the steps of the methods of the various embodiments of the present application. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, ROM, RAM, magnetic disks, or optical discs that can store program codes.

[0236] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for assessing the risk of thermal runaway, characterized in that, The method includes: Obtaining time-series status data of the battery, where the time-series status data is used to indicate the changes in battery parameters of the battery at different time points; Inputting the time-series status data into a trained thermal runaway warning model to obtain a risk assessment result. The thermal runaway warning model is used to detect whether there is a thermal runaway risk for the battery. The thermal runaway warning model includes: a long short-term memory network module and a soft decision tree module. The long short-term memory network module is used to extract feature data related to the thermal runaway risk, and the soft decision tree module is used to determine whether the battery has a thermal runaway.

2. The method according to claim 1, wherein The inputting the time-series status data into the trained thermal runaway warning model to obtain a risk assessment result includes: Inputting the time-series status data into the long short-term memory network module to obtain a risk feature vector of the battery, where the risk feature vector is used to indicate feature data related to the thermal runaway risk; Inputting the risk feature vector into the soft decision tree module to obtain the risk assessment result.

3. The method according to claim 2, wherein The soft decision tree module includes: a plurality of preset evaluation paths; the inputting the risk feature vector into the soft decision tree module to obtain the risk assessment result includes: For each of the preset evaluation paths, based on the risk feature vector and the preset evaluation path, obtaining initial probability evaluation information corresponding to the preset evaluation path to obtain the initial probability evaluation information corresponding to the plurality of preset evaluation paths. The initial probability evaluation information is used to indicate the probability of the battery having a thermal runaway risk, and one preset evaluation path corresponds to one initial probability evaluation information; Based on the plurality of initial probability evaluation information, obtaining the risk assessment result.

4. The method according to claim 3, characterized in that, The preset evaluation path includes a plurality of evaluation factors and a plurality of probability weights. The evaluation factors are used to indicate the conditions for the battery to have a thermal runaway risk, and the probability weights are used to indicate the influence degree of the evaluation factors on the battery having a thermal runaway risk. One evaluation factor corresponds to a plurality of probability weights; The obtaining the initial probability evaluation information corresponding to the preset evaluation path based on the risk feature vector and the preset evaluation path includes: Based on the risk feature vector, the plurality of evaluation factors, and the plurality of probability weights, obtaining the initial probability evaluation information corresponding to the preset evaluation path.

5. The method according to any one of claims 2-4, characterized in that, The long short-term memory network module includes: a fully connected layer, and the fully connected layer is used to input the risk feature vector into the soft decision tree module; wherein, the dimension of the fully connected layer is the same as the dimension of the input layer of the soft decision tree module.

6. The method according to any one of claims 1-4, characterized in that The trained thermal runaway warning model is obtained by the following method: Obtaining a training set, a first model loss, a second model loss, and an initial thermal runaway warning model. The first model loss is used to adjust the accuracy of the soft decision tree module, and the second model loss is used to adjust the accuracy of the long short-term memory network module; Training the initial thermal runaway warning model based on the training set, the first model loss, and the second model loss to obtain the trained thermal runaway warning model.

7. The method according to claim 6, wherein The initial thermal runaway warning model is the pruned thermal runaway warning model; training the initial thermal runaway warning model based on the training set, the first model loss, and the second model loss to obtain the trained thermal runaway warning model includes: Training the initial thermal runaway warning model based on the training set, the first model loss, and the second model loss to obtain first model parameters; Sending the training set and the first model parameters to a server, where an unpruned thermal runaway warning model is deployed on the server, and the server is configured to optimize the first model parameters based on the training set and the unpruned thermal runaway warning model; Receiving second model parameters; the second model parameters are the optimized first model parameters; Obtaining the trained thermal runaway warning model based on the second model parameters.

8. The method according to any one of claims 1 to 4, characterized in that, The method further includes: Obtaining an initial soft decision forest model, where the initial soft decision forest model includes a plurality of initial evaluation paths, and the initial evaluation paths include: branch structures; Determining the similarity of the plurality of branch structures based on the plurality of branch structures; Determining multiple groups of similar structure groups based on the similarity of the plurality of branch structures, where one group of similar structure groups includes at least two branch structures whose similarity is less than a preset similarity threshold; Fusing the branch structure parameters in each group of similar structure groups to obtain the soft decision forest model.

9. The method according to claim 8, wherein The branch structure includes at least one of the following: a branch node and a decision edge, the branch node is used to indicate an evaluation factor, the decision edge is used to indicate a probability weight, the evaluation factor is used to indicate the condition for the battery to have a thermal runaway risk, the probability weight is used to indicate the influence degree of the evaluation factor on the battery having a thermal runaway risk, and one evaluation factor corresponds to multiple probability weights.

10. An evaluation device for thermal runaway risk, characterized in that, The apparatus includes: An acquisition module, configured to acquire time-series state data of a battery, where the time-series state data is used to indicate the change of battery parameters of the battery at different time points; A processing module, configured to input the time-series state data into a trained thermal runaway warning model to obtain a risk assessment result, where the thermal runaway warning model is used to detect whether the battery has a thermal runaway risk, and the thermal runaway warning model includes: a long short-term memory network module and a soft decision tree module, the long short-term memory network module is used to extract feature data related to the thermal runaway risk, and the soft decision tree module is used to determine whether the battery has a thermal runaway.

11. An evaluation device for thermal runaway risk, characterized in that, Includes: A processor; A memory for storing executable instructions of the processor; wherein, the processor is configured to execute the instructions to implement the method according to any one of claims 1-9.

12. A computer-readable storage medium, characterized in that, When the computer-executable instructions stored in the computer-readable storage medium are executed by the processor of the thermal runaway risk assessment apparatus, the thermal runaway risk assessment apparatus can execute the method according to any one of claims 1-9.

Images