A data-driven new energy vehicle fire accident cause determination method

Abstract

The present application relates to a kind of data-driven new energy vehicle fire accident cause determination method, belong to new energy vehicle safety technical field.Establish the fire trace database including fire accident scene and real vehicle fire experiment, build and train new energy vehicle fire trace identification model, output new energy vehicle burn loss feature vector;Establish new energy vehicle burn loss feature dataset, build up fire site research and judge network and carry out training, realize the intelligent research and judge of fire site;Establish the background data set including background data matrix and accident cause coding, build and train background data intelligent analysis network, realize the intelligent analysis of fire cause;Collect the trace image and background data matrix of the new energy vehicle fire scene to be investigated, input trained fire trace identification model and background data intelligent analysis network, output fire site and fire cause, realize the intelligent determination of new energy vehicle fire accident cause.

Description

Technical Field

[0001] This invention relates to a data-driven method for determining the cause of fire accidents in new energy vehicles, belonging to the field of new energy vehicle safety technology. Background Technology

[0002] Developing new energy vehicles is an essential path for my country to transform from a major automobile producer to a leading automobile power. In recent years, my country's sales and ownership of new energy vehicles have experienced explosive growth, making it a core engine for the global new energy vehicle industry. However, with the rapid increase in ownership, fire accidents involving new energy vehicles have become frequent and are showing an upward trend year by year. Untimely and unclear investigations into the causes of these accidents can easily escalate conflicts between the parties involved and hinder industry upgrading. For fire investigation agencies, the high value of each new energy vehicle and the resulting slow and inaccurate investigations can easily lead to economic and even legal disputes. For vehicle manufacturers, unclear tracing of the causes of fire accidents makes it difficult to provide targeted improvement plans for upgrading the safety of new energy vehicles. Therefore, conducting timely and effective investigations into fire accidents involving new energy vehicles is of great significance for reducing social disputes and upgrading safety technologies.

[0003] Currently, investigations into the causes of fires involving new energy vehicles primarily rely on frontline fire investigators from fire and rescue teams. The investigation process typically involves determining the location of the fire based on on-site traces, and then classifying the case based on the identification results of evidence collected from the fire site. Such investigations are generally lengthy, and the identification of the fire location heavily depends on the experience of the fire investigators; the accuracy of the investigation conclusions is directly related to the investigators' experience level. However, the traces in new energy vehicle fires are highly homogeneous, making it difficult to identify the spread patterns. Furthermore, frontline fire investigators generally face shortages of manpower and weak infrastructure, resulting in the timeliness and accuracy of investigations into new energy vehicle fires failing to meet societal expectations. Therefore, there is an urgent need to develop methods for determining the causes of new energy vehicle fires to achieve efficient and accurate investigations, improving investigation efficiency while reducing reliance on the experience of investigators.

[0004] Currently, there is no mature intelligent analysis method for the causes of fires in new energy vehicles, and existing methods have many shortcomings in various aspects. For example, the invention patent with authorization announcement number CN 114358663 B, "Comprehensive Judgment Method for Electric Vehicle Fire Accidents Based on Artificial Intelligence," requires manual export and analysis of BMS stored data, necessitating investigators with strong technical expertise. Furthermore, if the BMS is damaged, the data support from the power battery is lost. The analysis of the spread of fire on the vehicle's exterior mainly relies on changes in paint color, making it difficult to effectively identify the initial fire location for vehicles with completely burned paint. The invention patent with application number CN202310410415.5, "An Intelligent Recognition Method for Fire Traces in Electric Vehicles," can preliminarily determine the location of the fire source through intelligent recognition of multi-regional trace features, but it lacks an effective means of determining the cause of the fire.

[0005] Therefore, it is necessary to conduct research on data-driven methods for determining the causes of fires in new energy vehicles, and to use deep learning methods to achieve intelligent determination of the location and cause of the fire, thereby reducing the dependence of investigation conclusions on the experience of investigators and enabling rapid and accurate determination of the causes of fires in new energy vehicles. Summary of the Invention

[0006] Given that current investigations into the causes of fires involving new energy vehicles are primarily conducted by frontline fire investigators from fire and rescue teams, the investigation process suffers from low timeliness and accuracy heavily relies on the experience of the personnel. Furthermore, there is a lack of rapid, intelligent, and objective methods for determining the cause of fires. This invention provides a data-driven method for determining the causes of fires involving new energy vehicles. It constructs a multi-angle trace intelligent recognition network for new energy vehicles, outputting intelligent recognition results of the degree of burns at different angles and locations. It also constructs a burn feature vector for new energy vehicles, using a CNN model to intelligently analyze burn features and intelligently output the location of the fire. Finally, it constructs an intelligent recognition model for backend data of new energy vehicles; by inputting a backend data matrix, it can intelligently output the most probable cause of the fire. Combining the intelligently identified fire location and cause, intelligent determination of the cause of fires involving new energy vehicles can be achieved.

[0007] The technical solution adopted in this invention is: a data-driven method for determining the cause of fire accidents in new energy vehicles, the steps of which are as follows: Step 1: Establishment of a fire trace database for new energy vehicles; establish a fire trace database containing on-site investigation cases of fire accidents in new energy vehicles and simulated fire experiments of actual vehicles, and define the degree of burn damage of vehicles from different angles in the fire trace database. Step 2: Construction of a fire trace recognition model for new energy vehicles; Establish a DeepLab V3+ image segmentation model and train it using a fire trace database to achieve intelligent segmentation of the areas around the exterior of the vehicle and the power battery with different degrees of burns; Establish a CNN image recognition model and train it using a fire trace database to achieve intelligent recognition of the degree of burns in the driver's compartment, the rear trunk, and the engine compartment. Step 3: Construction of a network for judging the fire location of new energy vehicles; using the new energy vehicle fire trace database to obtain the burn feature vector and fire location code of new energy vehicles, and establish a new energy vehicle burn feature dataset; build a one-dimensional CNN model and train it using the feature dataset to realize intelligent judgment of the fire location; Step 4: Construction of a smart analysis network for new energy vehicle back-end data; collect back-end data on new energy vehicle fire accidents, establish a back-end dataset containing a back-end data matrix and accident cause codes; construct a smart analysis network for back-end data and train it using the back-end dataset, input the back-end data matrix into the smart analysis network for back-end data to achieve intelligent analysis of the cause of the fire; Step 5: Intelligent Determination of the Cause of Fires in New Energy Vehicles; Input the fire trace images collected at the scene of the fire in the new energy vehicle to be investigated into the fire trace recognition model to obtain the burn feature vector of the new energy vehicle. Input the feature vector into the fire location analysis network to determine the fire location; Extract the background data of the fire accident to construct a background data matrix and input it into the intelligent analysis network of the background data of the new energy vehicle to output the accident cause code to determine the cause of the fire; Combine the output fire location and cause of the fire to determine the cause of the fire in the new energy vehicle.

[0008] The method for establishing the new energy vehicle fire trace database in step one is as follows: Based on a large number of on-site investigation cases of new energy vehicle fire accidents and vehicle fire simulation experiments, a new energy vehicle fire trace database is established. Each vehicle fire case in the database is required to have investigation conclusions including the location of the fire and the cause of the fire. The database includes image data from five parts: the exterior of the vehicle, the passenger compartment, the trunk, the engine compartment, and the power battery. The exterior of the vehicle includes four angles: the front, rear, left, and right sides. The passenger compartment includes two areas: the front and rear seats. Images from each angle of the exterior and the power battery are categorized as "undamaged" or "slightly damaged." The degree of burn damage is divided into multiple areas, such as "moderate burn damage" and "severe burn damage". For the vehicle's exterior, if the paint has not changed color, it is defined as no burn damage; if the paint has been burned and discolored, it is defined as slight burn damage; if the paint is missing and the metal body is exposed, it is defined as moderate burn damage; and if the metal body is rusted, discolored, or oxidized and blackened, it is defined as severe burn damage. The power battery is divided into multiple areas based on the burn damage to the casing: areas where the casing has not changed color are defined as no burn damage; areas where the casing has been partially heated and discolored are defined as slight burn damage; areas where the casing has partially melted and burned through are defined as moderate burn damage; and areas where the casing has melted extensively and the interior is exposed are defined as severe burn damage. The driver's compartment, rear trunk, and engine compartment are categorized into "no burn damage", "slight burn damage", "moderate burn damage", and "severe burn damage" according to the overall degree of burn damage.

[0009] The method for constructing the new energy vehicle fire trace recognition model in step two is as follows: A DeepLab V3+ image segmentation model is established, trained using image data of the vehicle's exterior and power battery from the new energy vehicle fire trace database, to achieve intelligent segmentation and recognition of the degree of burn damage to the vehicle's exterior and power battery; a CNN image recognition model is established, trained using image data of the front and rear seats of the driver's compartment and the rear trunk and engine compartment from the new energy vehicle fire trace database, to achieve analysis of the degree of burn damage to each part.

[0010] The method for constructing the fire location analysis network for new energy vehicles described in step three is as follows: (1) Construction of the burn feature dataset; the segmented images of the front, rear, and power battery burn extent in the new energy vehicle fire trace database are divided into four equal regions: upper left, lower left, upper right, and lower right; the segmented images of the left and right sides of the vehicle burn extent in the new energy vehicle fire trace database are divided into four equal regions from front to back: front, front-middle, middle-rear, and rear; the upper left, lower left, upper right, and lower right regions of the front of the vehicle, and the upper left, lower left, and upper right regions of the rear of the vehicle are calculated respectively. There are 20 regions in total: the lower right region, the upper left, lower left, upper right, and lower right regions of the power battery, the front left, front middle, middle rear, and rear regions of the vehicle, and the front right, front middle, middle rear, and rear regions of the vehicle. Each region has a burn characteristic vector [a, b, c, d], where a is the percentage of "unburned" area, b is the percentage of "slightly burned" area, c is the percentage of "moderately burned" area, and d is the percentage of "severely burned" area. The front and rear of the vehicle, and the power battery are pieced together in the order of upper left, lower left, upper right, and lower right to form three 16-bit arrays. A 16-dimensional vector is used to represent the burn characteristics of the front, rear, and power battery of the vehicle, respectively. The left and right sides of the vehicle are concatenated into two 16-dimensional vectors in the order of front, front-middle, middle-rear, and rear, representing the burn characteristics of the left and right sides of the vehicle, respectively. Burn severity feature codes are established for the driver's compartment, trunk, and engine compartment. Unburned front seats in the driver's compartment are coded as [1,0,0,0], slightly burned as [0,1,0,0], moderately burned as [0,0,1,0], and severely burned as [0,0,0,1]. Unburned rear seats in the driver's compartment are coded as [1,0,0,0]... [0,0,0], Slight burn damage is recorded as [0,1,0,0], Moderate burn damage is recorded as [0,0,1,0], Severe burn damage is recorded as [0,0,0,1]; The rear trunk is not burned and is recorded as [1,0,0,0], Slight burn damage is recorded as [0,1,0,0], Moderate burn damage is recorded as [0,0,1,0], Severe burn damage is recorded as [0,0,0,1]; The engine compartment of the car is not burned and is recorded as [1,0,0,0], Slight burn damage is recorded as [0,1,0,0], Moderate burn damage is recorded as [0,0,1,0], Severe burn damage is recorded as [0,0,0,1]. For the example of a new energy vehicle with a front-heavy burn damage pattern, the engine compartment is severely burned (characteristic denoted as [0,0,0,1]), the front seats of the driver's compartment are moderately burned (characteristic denoted as [0,0,1,0]), the rear seats of the driver's compartment are slightly burned (characteristic denoted as [0,1,0,0]), and the trunk is unburned (characteristic denoted as [1,0,0,0]). The burn damage feature vectors of each part are merged and concatenated into a 96-dimensional vector in the following order: front of the vehicle, rear of the vehicle, left side of the vehicle, right side of the vehicle, power battery, engine compartment, front seats of the driver's compartment, rear seats of the driver's compartment, and trunk. This vector is the burn damage feature vector of the new energy vehicle. A coding system for the fire ignition points of new energy vehicles is established. The fire ignition point in the power battery is denoted as [1,0,0,0,0,0,0,0,0], in the engine compartment as [0,1,0,0,0,0,0,0,0], in the front row of the driver's cab as [0,0,1,0,0,0,0,0,0], in the rear row of the driver's cab as [0,0,0,1,0,0,0,0,0], in the trunk as [0,0,0,0,1,0,0,0,0], and in the exterior front, rear, left, and right sides of the vehicle as [0,0,0,0,0,1,0,0,0], [0,0,0,0,0,0,0,1,0], [0,0,0,0,0,0,0,1,0], [0,0,0,0,0,0,0,0,1], and [0,0,0,0,0,0,0,0,1], respectively. We collected and mapped the burn feature vectors and one-hot codes of the fire locations of multiple new energy vehicles to establish a new energy vehicle fire burn feature dataset. (2) Construction of fire location judgment network: A CNN model is built as the fire location judgment network. The fire location judgment network is trained using the established new energy vehicle fire burn feature dataset, so that it can ultimately input the burn feature vector of new energy vehicle and output the one-hot code of the fire location, thereby judging the fire location.

[0011] The method for constructing the intelligent analysis network for new energy vehicle background data in step four is as follows: Based on a large number of survey cases, collect a large amount of background data on new energy vehicle fire accidents, extract the insulation resistance, single cell voltage, power battery temperature, ambient temperature, and fault codes of each new energy vehicle that caught fire, and extract the valid data of the 5000 frames before the data is disconnected to establish an n×5000 matrix, where n is the number of features of the new energy vehicle background data; let the background data contain n1 single cell voltage data, n2 battery temperature data, n3 ambient temperature data, and n4 fault code data, then n = n1 + n2 + n3 + n4 +1; Establish an accident cause code for each new energy vehicle fire case, where power battery failure is [1,0,0], electrical circuit failure is [0,1,0], and external factors are [0,0,1]. The n×5000 matrix for each fire case corresponds to the accident cause code; Establish a CNN model. Since the number of battery cells and temperature collection points are different for each model of new energy vehicle, n is also different. Therefore, adaptive pooling is used in the pooling layer of the CNN model to dynamically match the size of the vehicle background data matrix; Train the CNN model by inputting the collected background data and the cause code, so that the input background data can output the cause of the fire accident.

[0012] The method for intelligently determining the cause of a new energy vehicle fire, as described in step five, is as follows: Multi-angle images of the new energy vehicle fire accident to be investigated are collected. Images of the vehicle's exterior and the power battery are input into a trained DeepLab V3+ image segmentation model, which outputs 16-dimensional burn feature vectors for the front, rear, and power battery of the vehicle. Images of the driver's compartment, rear trunk, and engine compartment are input into a CNN image recognition model, which outputs burn feature vectors for the engine compartment, front and rear seats of the driver's compartment, and rear trunk. These burn feature vectors are combined to form a 96-dimensional new energy vehicle burn feature vector. This burn feature vector is then input into a fire location analysis network, which outputs a one-hot code for the fire location of the new energy vehicle, intelligently determining the fire location. Background data uploaded by the new energy vehicle is collected, and an n×5000 background data matrix is ​​established. This matrix is ​​input into a background data analysis network to obtain the accident cause code, thereby determining the cause of the vehicle fire. Finally, the cause of the new energy vehicle fire accident is comprehensively determined by combining the intelligently output fire location and cause.

[0013] The technical effects of this invention are as follows: By employing a method driven by image trace data and background data, combined with numerous actual investigation cases and physical fire simulation experiments, this invention establishes an intelligent identification network for fire traces in new energy vehicles and an intelligent background data analysis network, which can produce the following technical effects: (1) The fire trace recognition model for new energy vehicles can intelligently divide the exterior of the vehicle and the burned area of ​​the power battery using deep learning and computer vision technology. It can effectively determine the degree of burn in the driver's compartment, the rear trunk, and the engine compartment. It can avoid the traditional "experience-based" fire investigation and effectively quantify the burn characteristics of different areas of new energy vehicles, and obtain the burn feature vector of new energy vehicles.

[0014] (2) The constructed fire location analysis network for new energy vehicles can effectively extract the feature information in the burn feature vector of new energy vehicles, output the fire location code, realize the intelligent judgment of the fire location of new energy vehicles, and effectively avoid misjudgment of the fire location due to insufficient experience of investigators.

[0015] (3) The intelligent analysis network for the background data of new energy vehicles can intelligently mine the deep features in the background data and analyze the cause of fire of new energy vehicles through feature integration. This can effectively avoid misjudgment of the cause of fire due to insufficient data analysis ability of grassroots investigators.

[0016] (4) This method can effectively determine the location and cause of fire in new energy vehicles. It is timely and objective, and is of great significance for reducing social disputes and upgrading the safety technology of new energy vehicles. Attached Figure Description

[0017] Figure 1Image of burn marks on the front of a car; Figure 2 Image of burn marks on the left side of the car; Figure 3 Images of burn marks on a power battery; Figure 4 Image segmentation to determine the extent of burn damage to the front of the car; Figure 5 Image segmentation to determine the extent of burn damage on the left side of the car; Figure 6 Image segmentation to determine the degree of battery burnout; Figure 7 The image shows the extent of the burn damage to the front of the car divided into four equal parts. Figure 8 The image shows the extent of the burn damage on the left side of the car, divided into four equal parts. Figure 9 The image shows the degree of burn damage to the power battery divided into four equal parts. Detailed Implementation

[0018] Example 1 takes the investigation of a fire in a new energy vehicle caused by an electrical circuit fault in the engine compartment as an example; (1) Establishment of a database of fire traces for new energy vehicles: Based on a large number of on-site investigation cases of fire accidents involving new energy vehicles and simulated fire experiments on actual vehicles, a database of fire traces for new energy vehicles is established. Each fire case in the database is required to have an investigation conclusion including the location of the fire and the cause of the fire. The database includes image data of five parts: the exterior of the vehicle, the driver's compartment, the trunk, the engine compartment, and the power battery. The exterior of the vehicle includes four angles: the front, rear, left, and right sides of the vehicle. The driver's compartment includes two areas: the front and rear seats of the driver's compartment. The images of each angle of the exterior of the vehicle and the power battery are categorized as "undamaged", "slightly damaged", "moderately damaged", and "medium damaged". The degree of burn damage is divided into multiple areas, including "severe burn damage" and "minor burn damage". For the vehicle's exterior, if the paint has not changed color, it is defined as "no burn damage"; if the paint has been burned and discolored, it is "minor burn damage"; if the paint is missing and the metal body is exposed, it is "moderate burn damage"; and if the metal body is rusted, discolored, or oxidized and blackened, it is "severe burn damage". The power battery is also divided into multiple areas based on the extent of burn damage to the casing: areas where the casing has not changed color are defined as "no burn damage"; areas where the casing has been partially heated and discolored are defined as "minor burn damage"; areas where the casing has partially melted and burned through are defined as "moderate burn damage"; and areas where the casing has melted extensively and the interior is exposed are defined as "severe burn damage". The cab, trunk, and engine compartment are categorized into "no burn damage", "minor burn damage", "moderate burn damage", and "severe burn damage" based on the overall degree of burn damage.

[0019] (2) Construction of fire trace recognition model for new energy vehicles: DeepLab V3+ image segmentation model is established and trained using image data of the exterior of the vehicle and the power battery in the fire trace database of new energy vehicles to realize intelligent segmentation and recognition of the degree of burns of the exterior of the vehicle and the power battery; CNN image recognition model is established and trained using image data of the front row, rear row and rear trunk and engine compartment in the driver's cab in the fire trace database of new energy vehicles to realize the analysis of the degree of burns of each part.

[0020] (3) Construction of the fire location judgment network: Construct a new energy vehicle burn damage feature dataset, divide the segmented images of the front, rear, and power battery burn damage in the new energy vehicle fire trace database into four equal regions: upper left, lower left, upper right, and lower right. Divide the segmented images of the left and right sides of the vehicle burn damage in the new energy vehicle fire trace database into four equal regions: front, front-middle, middle-rear, and rear. Calculate the upper left, lower left, upper right, and lower right regions of the front of the vehicle, the upper left, lower left, upper right, and lower right regions of the rear of the vehicle, the upper left, lower left, upper right, and lower right regions of the power battery, and the front, front-middle, middle-rear, and rear regions of the left side of the vehicle. The burn feature vectors for each of the 20 regions—front right, front middle, middle rear, and rear—are [a, b, c, d]. Here, a represents the percentage of the region that is "not burned," b represents the percentage of the region that is "slightly burned," c represents the percentage of the region that is "moderately burned," and d represents the percentage of the region that is "severely burned." The front of the car, the rear of the car, and the power battery are concatenated into three 16-dimensional vectors in the order of top left, bottom left, top right, and bottom right, representing the burn features of the front of the car, the rear of the car, and the power battery, respectively. The left side of the car and the right side of the car are concatenated into two 16-dimensional vectors in the order of front, front middle, middle rear, and rear, representing the burn features of the left side of the car and the right side of the car, respectively. Establish characteristic codes for the degree of burn damage in the cab, rear cargo box, and engine compartment. For the front row of the cab, no burn damage is recorded as [1,0,0,0], slight burn damage as [0,1,0,0], moderate burn damage as [0,0,1,0], and severe burn damage as [0,0,0,1]. For the rear row of the cab, no burn damage is recorded as [1,0,0,0], slight burn damage as [0,1,0,0], moderate burn damage as [0,0,1,0], and severe burn damage as [0,0,1,0]. Burn damage is denoted as [0,0,0,1]; no burn damage in the rear trunk is denoted as [1,0,0,0], slight burn damage as [0,1,0,0], moderate burn damage as [0,0,1,0], and severe burn damage as [0,0,0,1]; no burn damage in the engine compartment is denoted as [1,0,0,0], slight burn damage as [0,1,0,0], moderate burn damage as [0,0,1,0], and severe burn damage as [0,0,0,1]. The burn damage feature vectors of the vehicle's exterior, power battery, engine compartment, front and rear seats of the driver's cab, and rear trunk are merged and concatenated into a 96-dimensional vector. A coding system for the fire ignition points of new energy vehicles is established. The fire ignition point in the power battery is denoted as [1,0,0,0,0,0,0,0,0], in the engine compartment as [0,1,0,0,0,0,0,0,0], in the front row of the driver's cab as [0,0,1,0,0,0,0,0,0], in the rear row of the driver's cab as [0,0,0,1,0,0,0,0,0], in the trunk as [0,0,0,0,1,0,0,0,0], and in the exterior front, rear, left, and right sides of the vehicle as [0,0,0,0,0,1,0,0,0], [0,0,0,0,0,0,0,1,0], [0,0,0,0,0,0,0,1,0], [0,0,0,0,0,0,0,0,1], and [0,0,0,0,0,0,0,0,1], respectively. We collected and mapped the burn feature vectors and one-hot codes of the fire locations of multiple new energy vehicles to establish a new energy vehicle fire burn feature dataset.

[0021] (4) Construction of intelligent analysis network for background data of new energy vehicles: Based on a large number of investigation cases, a large amount of background data of new energy vehicle fire accidents was collected. The insulation resistance, single cell voltage, power battery temperature, ambient temperature and fault code of each new energy vehicle fire were extracted. The valid data of the first 5000 frames before the data was disconnected was extracted and an n×5000 matrix was established, where n is the number of features of the background data of new energy vehicles. Let the background data contain n1 single cell voltage data, n2 battery temperature data, n3 ambient temperature data and n4 fault code data, then n=n1+n2+n3+n4+1. Establish an accident cause code for each new energy vehicle fire case, where power battery failure is [1,0,0], electrical circuit failure is [0,1,0], and external factors are [0,0,1]. The n×5000 matrix of each fire case corresponds to the accident cause code. A CNN model is built. Since the number of battery cells and temperature collection points varies for each model of new energy vehicle, n is also different. Therefore, adaptive pooling is used in the pooling layer of the CNN model to dynamically match the size of the vehicle background data matrix. The collected background data and cause codes are input to train the CNN model, so that the input background data can output the cause of the fire accident.

[0022] (5) Investigation into a fire in a new energy vehicle caused by an electrical wiring fault in the engine compartment: ① Upon arriving at the scene of the new energy vehicle fire, fire investigators, in accordance with the requirements of the new energy vehicle fire trace database, collected photographs of five parts: the exterior of the vehicle (including the front, rear, left, and right sides), the driver's cabin (including the front and rear seats), the trunk, the engine compartment, and the power battery. Images of the burn traces on the front, left side, and power battery are shown below. Figure 1 , Figure 2 , Figure 3 As shown, Figure 1 Image of burn marks on the front of the car. Figure 2 Image of burn marks on the left side of the car. Figure 3 Images of burn marks on a power battery; ② Input the collected photos of the car's exterior and the power battery into the DeepLab V3+ image segmentation model, and output segmented images of the burn extent of the car's exterior and the power battery; input the photos of the driver's compartment, trunk, and engine compartment into the CNN image recognition model, and output feature vectors of the burn extent of the driver's compartment, trunk, and engine compartment: The engine compartment suffered severe burn damage, coded as [0,0,0,1]; the front row of the cab suffered moderate burn damage, coded as [0,0,1,0]; the rear row of the cab suffered slight burn damage, coded as [0,1,0,0]; the rear cargo box was not burned, coded as [1,0,0,0]; Segmented images of the front of the car, the left side of the car, and the extent of battery burn damage, as shown below. Figure 4 , Figure 5 , Figure 6 As shown, Figure 4 Image segmentation to determine the extent of burn damage to the front of the car. Figure 5 Image segmentation to determine the extent of burn damage on the left side of the car. Figure 6 The image is segmented to represent the degree of burn damage to the power battery; in the image, the green area represents "no burn damage", the yellow area represents "slight burn damage", the brown area represents "moderate burn damage", and the red area represents "severe burn damage".

[0023] Calculate the feature vector of the burn damage to the front of the car, and divide the image of the burn damage to the front of the car into four equal parts: upper left, upper right, lower left, and lower right. See [link / description]. Figure 7 For the upper left region ①, the area of ​​"moderate burn damage" accounts for 95%, and the area of ​​"severe burn damage" accounts for 5%. For the upper right region ②, the area of ​​"moderate burn damage" accounts for 93%, and the area of ​​"severe burn damage" accounts for 7%. For the lower left region ③, the area of ​​"moderate burn damage" accounts for 11%, and the area of ​​"severe burn damage" accounts for 89%. For the lower right region ④, the area of ​​"moderate burn damage" accounts for 8%, and the area of ​​"severe burn damage" accounts for 92%. Therefore, the burn damage feature vector of the front of the car is [0,0,0.95,0.05,0,0,0.93,0.07,0,0,0.11,0.89,0,0,0.08,0.92]. Calculate the burn feature vector of the rear of the car. Since the rear of the car is not burned, its burn feature vector is [1,0,0,0,1,0,0,0,1,0,0,0,1,0,0,0]; Calculate the feature vector of the burn damage on the left side of the car, and divide the image of the burn damage on the left side of the car into four equal parts: front, front-middle, middle-rear, and rear. See [link / details]. Figure 8 For the front region ①, the area of ​​"unburned" region accounts for 39%, the area of ​​"moderately burned" region accounts for 35%, and the area of ​​"severely burned" region accounts for 26%. For the front-middle region ②, the area of ​​"unburned" region accounts for 85%, the area of ​​"slightly burned" region accounts for 11%, and the area of ​​"moderately burned" region accounts for 4%. For the middle-rear region ③, the area of ​​"unburned" region accounts for 100%. For the rear region ④, the area of ​​"unburned" region accounts for 100%. Therefore, the burn feature vector of the left side of the car is [0.39,0,0.35,0.26,0.85,0.11,0.04,0,1,0,0,0,1,0,0,0]. The calculation method for the burn feature vector on the right side of the car is the same as that on the left side. The calculated burn feature vector on the right side of the car is [0.37,0,0.37,0.26,0.84,0.12,0.04,0,1,0,0,0,1,0,0,0]; Calculate the feature vector of power battery burnout, and divide the power battery burnout level segmentation image into four equal parts: upper left, upper right, lower left, and lower right. See [link / reference]. Figure 9For the upper left region ①, the area of ​​"unburned" region accounts for 53% and the area of ​​"slightly burned" region accounts for 47%. For the upper right region ②, the area of ​​"unburned" region accounts for 45% and the area of ​​"slightly burned" region accounts for 55%. For the lower left region ③, the area of ​​"unburned" region accounts for 100%. For the lower right region ④, the area of ​​"unburned" region accounts for 100%. Therefore, the burn feature vector of the power battery is [0.53,0.47,0,0,0.55,0.45,0,0,1,0,0,0,1,0,0,0].

[0024] ③ After obtaining the burn feature codes of each part, the burn feature codes are merged and spliced ​​into a 96-dimensional vector in the order of front, rear, left, right, power battery, power compartment, front row of driver's cab, rear row of driver's cab, and rear trunk, which is the burn feature vector of the new energy vehicle. ④ Input the burn feature vector of the new energy vehicle into the fire location analysis network of the new energy vehicle. The network outputs [0,1,0,0,0,0,0,0,0], that is, the fire location analysis network of the new energy vehicle predicts that the fire location of the new energy vehicle is inside the engine compartment. ⑤ Collect background data of the new energy vehicle 7 days before the fire, extract the insulation resistance, single cell voltage, power battery temperature, ambient temperature and fault code information of the new energy vehicle background data, construct a background data matrix with a size of 150×5000, input the background data matrix into the new energy vehicle background data intelligent analysis network, the new energy vehicle background data intelligent analysis network outputs [0,1,0], that is, determine that the cause of the fire of the new energy vehicle is electrical circuit fault; ⑥ Based on the combined results of the new energy vehicle fire location analysis network and the new energy vehicle background data intelligent analysis network, this new energy vehicle fire was caused by an electrical circuit fault in the engine compartment. The engine compartment contains a large number of normally energized lines, posing a potential electrical fault. Therefore, the cause of the fire output by the network is reasonable.

Claims

1. A data-driven method for determining the cause of fire accidents in new energy vehicles, characterized in that, The steps are as follows: Step 1: Establishment of a fire trace database for new energy vehicles; Establish a fire trace database containing on-site investigation cases of fire accidents involving new energy vehicles and simulated fire experiments of actual vehicles, and define the degree of burn damage of vehicles from different angles in the fire trace database; Step 2: Construction of a fire trace recognition model for new energy vehicles; Establish a DeepLab V3+ image segmentation model and train it using a fire trace database to achieve intelligent segmentation of the areas around the exterior of the vehicle and the power battery with different degrees of burns; Establish a CNN image recognition model and train it using a fire trace database to achieve intelligent recognition of the degree of burns in the driver's compartment, the rear trunk, and the engine compartment. Step 3: Construction of a network for judging the fire location of new energy vehicles; using the new energy vehicle fire trace database to obtain the burn feature vector and fire location code of new energy vehicles, and establish a new energy vehicle burn feature dataset; build a one-dimensional CNN model and train it using the feature dataset to realize intelligent judgment of the fire location; Step 4: Construction of a smart analysis network for new energy vehicle back-end data; collect back-end data on new energy vehicle fire accidents, establish a back-end dataset containing a back-end data matrix and accident cause codes; construct a smart analysis network for back-end data and train it using the back-end dataset, input the back-end data matrix into the smart analysis network for back-end data to achieve intelligent analysis of the cause of the fire; Step 5: Intelligent Determination of the Cause of Fires in New Energy Vehicles; Input the fire trace images collected at the scene of the fire in the new energy vehicle to be investigated into the fire trace recognition model to obtain the burn feature vector of the new energy vehicle. Input the feature vector into the fire location analysis network to determine the fire location; Extract the background data of the fire accident to construct a background data matrix and input it into the intelligent analysis network of the background data of the new energy vehicle to output the accident cause code to determine the cause of the fire; Combine the output fire location and cause of the fire to determine the cause of the fire in the new energy vehicle.

2. The data-driven method for determining the cause of fire accidents in new energy vehicles according to claim 1, characterized in that, The method for establishing the new energy vehicle fire trace database in step one is as follows: Based on a large number of on-site investigation cases of new energy vehicle fire accidents and vehicle fire simulation experiments, a new energy vehicle fire trace database is established. Each vehicle fire case in the database is required to have investigation conclusions including the location of the fire and the cause of the fire. The database includes image data from five parts: the exterior of the vehicle, the driver's compartment, the trunk, the engine compartment, and the power battery. The exterior of the vehicle includes four angles: the front, rear, left, and right sides. The driver's compartment includes two areas: the front and rear seats. Images from each angle of the exterior and the power battery are categorized as "undamaged" or "slightly damaged." The degree of burn damage is divided into multiple areas, namely "moderate burn damage" and "severe burn damage". For the marks around the vehicle's exterior, the degree of burn damage is divided into multiple areas: no burn damage is defined as no paint discoloration, light burn damage is defined as paint discoloration, moderate burn damage is defined as paint loss exposing the metal body, and severe burn damage is defined as rust discoloration or oxidation blackening of the metal body. The power battery is divided into multiple areas according to the burn damage of the outer casing: no burn damage is defined as an area where the outer casing does not discoloration, light burn damage is defined as an area where the outer casing is partially discolored due to heat, moderate burn damage is defined as an area where the outer casing is partially melted and burned through, and severe burn damage is defined as an area where the outer casing is largely melted and the interior is exposed. The driver's compartment, rear trunk, and engine compartment are divided into "no burn damage", "light burn damage", "moderate burn damage", and "severe burn damage" according to the overall degree of burn damage.

3. The data-driven method for determining the cause of fire accidents in new energy vehicles according to claim 1, characterized in that, The method for constructing the new energy vehicle fire trace recognition model in step two is as follows: A DeepLab V3+ image segmentation model is established, trained using image data of the vehicle's exterior and power battery from the new energy vehicle fire trace database, to achieve intelligent segmentation and recognition of the degree of burn damage to the vehicle's exterior and power battery; a CNN image recognition model is established, trained using image data of the front and rear seats of the driver's compartment and the rear trunk and engine compartment from the new energy vehicle fire trace database, to achieve analysis of the degree of burn damage to each part.

4. The data-driven method for determining the cause of fire accidents in new energy vehicles according to claim 1, characterized in that, The method for constructing the fire location judgment network of new energy vehicles in step three is as follows: (1) Construction of burn feature dataset; the segmented images of the front, rear and power battery burn degree in the new energy vehicle fire trace database are divided into four equal regions: upper left, lower left, upper right and lower right; the segmented images of the left and right sides of the vehicle burn degree in the new energy vehicle fire trace database are divided into four equal regions: front, front middle, middle rear and rear; the upper left, lower left, upper right and lower right regions of the front of the vehicle, the upper left, lower left, upper right and lower right regions of the rear of the vehicle, the upper left, lower left, upper right and lower right regions of the power battery, the front, front middle, middle rear and rear regions of the left side of the vehicle and the front right side of the vehicle are calculated respectively. The system is divided into 20 regions: front-middle, middle-rear, and rear. Each region has a burn characteristic vector [a, b, c, d], where a represents the percentage of "unburned" area, b represents the percentage of "slightly burned" area, c represents the percentage of "moderately burned" area, and d represents the percentage of "severely burned" area. The front of the vehicle, rear of the vehicle, and power battery are concatenated into three 16-dimensional vectors in the order of top left, bottom left, top right, and bottom right, representing the burn characteristics of the front, rear, and power battery, respectively. The left and right sides of the vehicle are concatenated into two 16-dimensional vectors in the order of front, front-middle, middle-rear, and rear, representing the burn characteristics of the left and right sides of the vehicle, respectively. Burn characteristics are also established for the cab, trunk, and engine compartment. The degree of damage is coded as follows: Front row of the driver's cab: no burn marks [1,0,0,0], slight burn marks [0,1,0,0], moderate burn marks [0,0,1,0], severe burn marks [0,0,0,1]; Rear row of the driver's cab: no burn marks [1,0,0,0], slight burn marks [0,1,0,0], moderate burn marks [0,0,1,0], severe burn marks [0,0,0,1]; Trunk: no burn marks [1,0,0,0], slight burn marks [0,1,0,0], moderate burn marks [0,0,1,0], severe burn marks [0,0,0,1]; Engine compartment: no burn marks [1,0,0,0], slight burn marks [0,0,0,1], ... [0,0,1,0] represents moderate burn damage, and [0,0,0,1] represents severe burn damage. For the example of a new energy vehicle with a front-heavy burn damage and a rear-light burn damage, the engine compartment is severely burned and the feature is recorded as [0,0,0,1], the front row of the driver's cab is moderately burned and the feature is recorded as [0,0,1,0], the rear row of the driver's cab is lightly burned and the feature is recorded as [0,1,0,0], and the trunk is not burned and the feature is recorded as [1,0,0,0]. The burn damage feature vectors of each part are merged and spliced ​​into a 96-dimensional vector in the order of front of the car, rear of the car, left side of the car, right side of the car, power battery, engine compartment, front row of the driver's cab, rear row of the driver's cab, and trunk. This vector is the burn damage feature vector of the new energy vehicle. A coding system for the fire ignition points of new energy vehicles is established. The fire ignition point in the power battery is denoted as [1,0,0,0,0,0,0,0,0], in the engine compartment as [0,1,0,0,0,0,0,0,0], in the front row of the driver's cab as [0,0,1,0,0,0,0,0,0], in the rear row of the driver's cab as [0,0,0,1,0,0,0,0,0], in the trunk as [0,0,0,0,1,0,0,0,0], and in the exterior front, rear, left, and right sides of the vehicle as [0,0,0,0,0,1,0,0,0], [0,0,0,0,0,0,0,1,0], [0,0,0,0,0,0,0,1,0], [0,0,0,0,0,0,0,0,1], and [0,0,0,0,0,0,0,0,1], respectively. We collected and mapped the burn feature vectors and one-hot codes of the fire locations of multiple new energy vehicles to establish a new energy vehicle fire burn feature dataset. (2) Construction of fire location judgment network: A CNN model is built as the fire location judgment network. The fire location judgment network is trained using the established new energy vehicle fire burn feature dataset, so that it can ultimately input the burn feature vector of new energy vehicle and output the one-hot code of the fire location, thereby judging the fire location.

5. The data-driven method for determining the cause of fire accidents in new energy vehicles according to claim 1, characterized in that, The method for constructing the intelligent analysis network for new energy vehicle background data in step four is as follows: Based on a large number of survey cases, collect a large amount of background data on new energy vehicle fire accidents, extract the insulation resistance, single cell voltage, power battery temperature, ambient temperature, and fault codes of each new energy vehicle that caught fire, and extract the valid data of the 5000 frames before the data is disconnected to establish an n×5000 matrix, where n is the number of features of the new energy vehicle background data; let the background data contain n1 single cell voltage data, n2 battery temperature data, n3 ambient temperature data, and n4 fault code data, then n = n1 + n2 + n3 + n4 +1; Establish an accident cause code for each new energy vehicle fire case, where power battery failure is [1,0,0], electrical circuit failure is [0,1,0], and external factors are [0,0,1]. The n×5000 matrix for each fire case corresponds to the accident cause code; Establish a CNN model. Since the number of battery cells and temperature collection points are different for each model of new energy vehicle, n is also different. Therefore, adaptive pooling is used in the pooling layer of the CNN model to dynamically match the size of the vehicle background data matrix; Train the CNN model by inputting the collected background data and the cause code, so that the input background data can output the cause of the fire accident.

6. The data-driven method for determining the cause of fire accidents in new energy vehicles according to claim 1, characterized in that, The method for intelligently determining the cause of a new energy vehicle fire, as described in step five, is as follows: Multi-angle images of the new energy vehicle fire accident to be investigated are collected. Images of the vehicle's exterior and the power battery are input into a trained DeepLab V3+ image segmentation model, which outputs 16-dimensional burn feature vectors for the front, rear, and power battery of the vehicle. Images of the driver's compartment, rear trunk, and engine compartment are input into a CNN image recognition model, which outputs burn feature vectors for the engine compartment, front and rear seats of the driver's compartment, and rear trunk. These burn feature vectors are combined to form a 96-dimensional new energy vehicle burn feature vector. This burn feature vector is then input into a fire location analysis network, which outputs a one-hot code for the fire location of the new energy vehicle, intelligently determining the fire location. Background data uploaded by the new energy vehicle is collected, and an n×5000 background data matrix is ​​established. This matrix is ​​input into a background data analysis network to obtain the accident cause code, thereby determining the cause of the vehicle fire. Finally, the cause of the new energy vehicle fire accident is comprehensively determined by combining the intelligently output fire location and cause.

Images