An electric vehicle fire tracing method, system, device, and medium
By using trapezoidal membership functions and Q-learning algorithms, and leveraging the fire feature classification and state transition constraints of electric vehicles, the subjectivity problem in tracing the source of electric vehicle fires is solved, enabling rapid and accurate tracing of fire causes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING VEHICLE TEST & RES INST CO LTD
- Filing Date
- 2025-11-12
- Publication Date
- 2026-06-19
AI Technical Summary
Current technologies for tracing the source of electric vehicle fires rely on expert experience, which is highly subjective, inefficient, and makes it difficult to guarantee objectivity and quickly pinpoint the cause of the fire.
Using a trapezoidal membership function and Q-learning algorithm, fire characteristics of electric vehicles are obtained, state categories and action spaces are classified, and an ε-greedy selection strategy and state transition constraints are used to iteratively learn Q-values. The largest relevant Q-value is selected as the cause of the fire.
It achieves comprehensiveness, objectivity, and accuracy in tracing the source of electric vehicle fires, and quickly locates the cause of the fire.
Smart Images
Figure CN121542835B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electric vehicle accident handling technology, and in particular to a method, system, equipment and medium for tracing the source of electric vehicle fires. Background Technology
[0002] In recent years, fires involving electric vehicles have been frequent. Therefore, when an electric vehicle catches fire, it is necessary to trace the source of the fire to find the root cause. This will allow for iterative design improvements to electric vehicles and reduce the occurrence of fires. However, current fire tracing efforts are hampered by issues such as missing or damaged accident-related data. Furthermore, accident tracing often relies on expert judgment, which is highly subjective and inefficient. The judgment of the cause of the fire is easily influenced by personal experience and cognitive biases, making it difficult to guarantee objectivity and quickly pinpoint the cause. This hinders the progress of efficient and accurate fire tracing. Summary of the Invention
[0003] To address the shortcomings of existing technologies, this invention provides a method, system, device, and medium for tracing the source of electric vehicle fires. It solves the problems of missing fire source tracing data and reliance on expert experience, which makes it difficult to guarantee objectivity and rapid location.
[0004] According to an embodiment of the present invention, a method for tracing the source of an electric vehicle fire includes:
[0005] Multiple fire characteristics and causes of electric vehicles are obtained, the fire characteristics are classified to obtain multiple state categories, the fire characteristics are combined into a state space, and the fire causes are combined into an action space.
[0006] The membership degree of each fire feature in the fuzzy linguistic variable range is determined by using a trapezoidal membership function.
[0007] State transition constraints are set according to the state category. Then, based on the state transition constraints and membership degree, the Q-learning algorithm is used to iteratively learn the Q value of any action in the action space for any state in the state space, thereby obtaining multiple related Q values.
[0008] The fire cause corresponding to the largest relevant Q value is selected as the fire cause for the current electric vehicle.
[0009] Preferably, the fuzzy language variable value thresholds include very small variable intervals, small variable intervals, medium variable intervals, large variable intervals, and very large variable intervals.
[0010] Preferably, the method for iterative learning using the Q-learning algorithm is as follows:
[0011] S1: Select any fire feature in the state space as the current state. Based on the Q value of the current state, use the ε-greedy selection strategy to select a fire cause from the action space as the action for the next moment.
[0012] S2: Calculate the state at the next time step based on the state transition constraints;
[0013] S3: Update the Q value of the current state based on the membership degree of the state and action at the next time step, and repeat steps S1-S3 until the Q value converges.
[0014] Preferably, the formula for calculating the Q value of the current state based on the Q value at the next moment is as follows:
[0015]
[0016] Where Q(s, a) is the Q value of the current state. Let r be the membership degree and r be the discount factor. For learning rate, For the state at the next moment, For the action in the next moment.
[0017] Preferably, the ε-greedy selection strategy is as follows:
[0018]
[0019] Where s is the current state, a is the current action, ε is called the exploration rate, and A(s) is the action space composed of the possible fire causes when the fire state is s.
[0020] Preferably, the state transition constraints are as follows:
[0021]
[0022] Where St is the current state, St+1 is the next state, and A, B, and C are different state categories.
[0023] On the other hand, according to embodiments of the present invention, an electric vehicle fire tracing system is also provided, which uses the above-described electric vehicle fire tracing method, including:
[0024] The data processing module is used to acquire multiple fire characteristics and causes of electric vehicles, classify the fire characteristics, and obtain multiple status categories.
[0025] A fuzzy control module is used to determine the membership degree of each fire feature on the fuzzy linguistic variable value threshold using a trapezoidal membership function.
[0026] The analysis module is used to set state transition constraints according to state categories, and then, based on the state transition constraints and membership degrees, it uses the Q-learning algorithm to iteratively learn the Q-value of any action in the action space for any state in the state space, obtain multiple relevant Q-values, and then select the fire cause corresponding to the largest relevant Q-value as the fire cause of the current electric vehicle.
[0027] On the other hand, according to an embodiment of the present invention, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the above-described method for tracing the source of an electric vehicle fire.
[0028] On the other hand, according to an embodiment of the present invention, a computer storage medium is also provided, which stores a computer program. When the computer program is executed by a processor, the processor performs the above-described method for tracing the source of an electric vehicle fire.
[0029] Compared with the prior art, the present invention has the following beneficial effects:
[0030] This invention utilizes a trapezoidal membership function to determine the nonlinear relationship between the fire characteristics and the cause of an electric vehicle fire. It solves the problem of state, action, and reward function modeling in Q-learning algorithms. Then, it uses the Q-learning algorithm to find the fire cause with the highest correlation to the fire characteristics. Through algorithm-driven approach, it achieves comprehensiveness, objectivity, and accuracy in the source tracing process. Attached Figure Description
[0031] Figure 1 This is a diagram illustrating the electric vehicle fire source tracing method according to an embodiment of the present invention.
[0032] Figure 2 This is a membership function curve diagram of an embodiment of the present invention. Detailed Implementation
[0033] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0034] like Figure 1 As shown in the figure, an embodiment of the present invention proposes a method for tracing the source of electric vehicle fires, including:
[0035] Multiple fire characteristics and causes of electric vehicles are obtained, the fire characteristics are classified to obtain multiple state categories, the fire characteristics are combined into a state space, and the fire causes are combined into an action space.
[0036] After an electric vehicle fire, some fire characteristics Si can be analyzed from the vehicle wreckage. These mainly include Si = {S1. Obvious tearing or damage hole at the bottom of the battery pack; S2. Obvious scratches at the bottom of the battery pack; S3. Obvious arcing marks on the top cover of the battery pack; S4. No damage to the battery pack exterior, only burn marks; S5. Obvious overcurrent melting beads on low-voltage electrical appliances or low-voltage wiring harnesses in the burning wreckage; S6. Milky white smoke before combustion; S7. Black smoke before combustion; S8. Violent flame ejection at the moment of ignition; S9. The fire started in the middle of the vehicle; S10. The fire started in the front engine compartment; S11. The fire started in the trunk; S12. The fire started inside the vehicle}. These fire characteristics can be directly combined to form a state space.
[0037] These fire characteristics correspond to different fire causes, mainly including aj={battery pack bottom impact a1, battery pack bottom scraping a2, battery pack individual cell short circuit a3, low voltage electrical system failure a4, fire of flammable items inside or outside the vehicle a5, charging system failure a6}. These fire causes can directly form the action space.
[0038] Based on the characteristics and data features of fires, they can be roughly divided into the following three categories:
[0039] The accident characteristic parameters related to battery fires are A={S1,S2,S3,S6,S8}, the causal characteristic parameters related to low-voltage system fires are B={S4,S5,S7,S9,S10,S11}, and the accident characteristic parameter set for fires caused by external ignition sources is C={S12}.
[0040] The membership degree of each fire feature on the threshold of fuzzy linguistic variable values is determined by using a trapezoidal membership function.
[0041] Based on real-world scenario databases and expert experience, when i is smaller or larger, the state si and action ai highlight the fire characteristics caused by the fire cause. Therefore, this invention uses membership degree to quantify the correlation between fire cause and fire characteristics. The larger the membership degree, the stronger the correlation, and vice versa.
[0042] The membership degree of this invention adopts a trapezoidal membership function. The fuzzy linguistic variable i of state (fire state) and action (fire cause) is described as small, medium and large. The fuzzy subset of state is: E(s)={L, M, H}, which respectively represent the degree of reflecting the fire state of the vehicle. The fuzzy subset of action is: E(a)={L, M, H}, where L, M and H represent the degree of relevance of the fire cause of the vehicle fire.
[0043] The threshold values for fuzzy linguistic variables are described as very small (very small variable interval), small (relatively small variable interval), medium (medium variable interval), large (large variable interval), and very large (very large variable interval), with E(r) = {VS, S, M, B, VB}. The membership function curve is shown below. Figure 2 The very small variable range is [0, 0.2), the small variable range is [0.2, 0.41), the medium variable range is [0.41, 0.59), the large variable range is [0.59, 0.81), and the very large variable range is [0.81, 1].
[0044] The input variables for the fuzzy control rule are state and action, and the output is the reward value. Based on the above analysis of how the output parameters change with the input parameters and the fire accident investigation database, fuzzy control rules that link the input and output variables were formulated, as shown in Table 1 below:
[0045] Table 1: Fuzzy Control Rule Table
[0046]
[0047] By utilizing the gradient membership function, the nonlinear relationship between the fire characteristics and the cause of an electric vehicle fire is determined, thus solving the problem of the state, action, and reward function model of the Q-learning algorithm.
[0048] State transition constraints are set according to the state category. Then, based on the state transition constraints and membership degree, the Q-learning algorithm is used to iteratively learn the Q value of any action in the action space for any state in the state space, thereby obtaining multiple related Q values.
[0049] Q-learning algorithms are globally optimal algorithms that transition to the next state after performing an action in the current state. However, in the process of tracing the source of electric vehicle fires, the next state is uncertain when the current state set performs an action, requiring a transition to the next state for global optimization. Since the next state set is large, this can lead to slow or non-convergent Q-value convergence. Therefore, this invention utilizes a database of past accident investigations to summarize and categorize data, defining state transitions. The state at the next moment is determined by the state transition factor. By eliminating completely irrelevant states, the speed of iterative updates and optimizations is accelerated.
[0050] In actual accident cause investigation, the probability that the current state and the next state are the same is 0, so identical states are directly eliminated. Furthermore, the state transition probability between different state datasets is also 0. The specific mathematical model is as follows:
[0051]
[0052] Therefore, when =0 indicates that the next state St+1 is unrelated to the previous state St, and state transition is eliminated. This can be understood as follows: if the next state and the previous state are of the same state category, the state transition probability can be considered to be 0, and the next state is eliminated.
[0053] In addition, the core of Q-learning is to continuously update the Q-values corresponding to state variables and action variables. This process is essentially a continuous optimization of the Q-table. The size of the Q-table directly affects the update speed of each iteration, and its size is determined by both state variables and action variables, as shown in Table 2. Therefore, when the dimension or size of the state variables decreases, the speed of the iteration loop will also increase accordingly.
[0054] Table 2: Q-table size corresponding to the number of state variables and action variables
[0055]
[0056] At the initial moment, the initial Q-value of Q-learning is calculated as follows:
[0057]
[0058] in, Discount factor and , In the state of And the action is The reward function at that time.
[0059] Then Q-learning iterations began:
[0060] S1: Select any fire feature in the state space as the current state. Based on the Q value of the current state, use the ε-greedy selection strategy to select a fire cause from the action space as the action for the next moment.
[0061] The ε-greedy selection strategy is used to explore all environments. ε is called the exploration rate, which is usually a small constant between 0 and 1, representing the probability of exploring the environment with ε.
[0062]
[0063] Where s is the current state, a is the current action, ε is called the exploration rate, and A(s) is the action space composed of the possible fire causes when the fire state is s.
[0064] S2: Calculate the state at the next time step based on the state transition constraints;
[0065] Based on the current state and the action selected in step S1, the Q-learning algorithm is used to calculate the state at the next moment. According to the state transition constraints, if the state at the next moment is S(t, ... If the current state is the same, return to step S1 and select an action again.
[0066] S3: Update the Q value of the current state based on the membership degree of the state and action at the next time step, and repeat steps S1-S3 until the Q value converges.
[0067] Update the Q value of the current state according to the following formula:
[0068]
[0069] Where Q(s, a) is the Q value of the current state. Let r be the membership degree and r be the discount factor. For learning rate, For the state at the next moment, For the next moment's action, This is the Q value of the next state.
[0070] Then repeat S1-S3, using the Q-learning algorithm to iteratively learn the Q-value of any action in the action space for any state in the state space, and obtain multiple related Q-values.
[0071] Finally, the fire cause corresponding to the largest relevant Q value is selected as the fire cause of the current electric vehicle. In this way, the Q-learning algorithm is used to find the fire cause with the greatest correlation to the fire characteristics. The algorithm drives the comprehensiveness, objectivity and accuracy of the source tracing process.
[0072] On the other hand, embodiments of the present invention also provide an electric vehicle fire source tracing system, which uses the above-described electric vehicle fire source tracing method, including:
[0073] The data processing module is used to acquire multiple fire characteristics and causes of electric vehicles, classify the fire characteristics, and obtain multiple status categories.
[0074] A fuzzy control module is used to determine the membership degree of each fire feature on the fuzzy linguistic variable value threshold using a trapezoidal membership function.
[0075] The analysis module is used to set state transition constraints according to state categories, and then, based on the state transition constraints and membership degrees, it uses the Q-learning algorithm to iteratively learn the Q-value of any action in the action space for any state in the state space, obtain multiple relevant Q-values, and then select the fire cause corresponding to the largest relevant Q-value as the fire cause of the current electric vehicle.
[0076] On the other hand, embodiments of the present invention also provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the above-described method for tracing the source of an electric vehicle fire.
[0077] On the other hand, embodiments of the present invention also provide a computer storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the above-described method for tracing the source of an electric vehicle fire.
[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for tracing the source of electric vehicle fires, characterized in that: include: Multiple fire characteristics and causes of electric vehicles are obtained, the fire characteristics are classified to obtain multiple state categories, the fire characteristics are combined into a state space, and the fire causes are combined into an action space. The membership degree of each fire feature in the fuzzy linguistic variable range is determined by using a trapezoidal membership function. State transition constraints are set according to the state category, and the state transition constraints are as follows: Where St is the current state, St+1 is the next state, and A, B, and C are different state categories; Based on state transition constraints and membership degrees, the Q-learning algorithm is used to iteratively learn the Q-value of any action in the action space for any state in the state space, resulting in multiple related Q-values, including: S1: Select any fire feature in the state space as the current state. Based on the Q value of the current state, use the ε-greedy selection strategy to select a fire cause from the action space as the action for the next moment. The ε-greedy selection strategy is as follows: Where s is the current state, a is the current action, ε is called the exploration rate, and A(s) is the action space composed of the possible fire causes when the fire state is s; S2: Calculate the state at the next time step based on the state transition constraints; S3: Update the Q-value of the current state based on the membership degree of the next state and the next action. The calculation formula is as follows: Where Q(s, a) is the Q value of the current state. Here, r is the discount factor, and r is the reward value. For learning rate, For the state at the next moment, For the action at the next moment, repeat steps S1-S3 until the Q value converges; The fire cause corresponding to the largest relevant Q value is selected as the fire cause for the current electric vehicle.
2. The method for tracing the source of electric vehicle fires as described in claim 1, characterized in that: The range of fuzzy language variables includes very small variable intervals, small variable intervals, medium variable intervals, large variable intervals, and very large variable intervals.
3. An electric vehicle fire traceability system, characterized by: The system uses a method for tracing the source of electric vehicle fires as described in any one of claims 1-2, comprising: The data processing module is used to acquire multiple fire characteristics and causes of electric vehicles, classify the fire characteristics, and obtain multiple status categories. A fuzzy control module is used to determine the membership degree of each fire feature in the fuzzy linguistic variable value domain using a trapezoidal membership function. The analysis module is used to set state transition constraints according to state categories, and then, based on the state transition constraints and membership degrees, it uses the Q-learning algorithm to iteratively learn the Q-value of any action in the action space for any state in the state space, obtain multiple relevant Q-values, and then select the fire cause corresponding to the largest relevant Q-value as the fire cause of the current electric vehicle.
4. A computer device, characterized by: The device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform a method for tracing the source of an electric vehicle fire as described in any one of claims 1 to 2.
5. A computer storage medium, characterized in that: The device contains a computer program that, when executed by a processor, causes the processor to perform a method for tracing the source of an electric vehicle fire as described in any one of claims 1 to 2.