A method and system for identifying fire in a charging cabin based on video monitoring

Abstract

The present application relates to the field of safety monitoring and early warning, and particularly relates to a charging cabin fire identification method and system based on video monitoring, which acquires multi-source monitoring data from the charging cabin in real time; performs visual feature analysis on the video image data to determine a visual feature vector; performs state feature analysis on the environmental sensor data to obtain a state feature vector; performs abnormal analysis on the visual feature vector and the state feature vector based on the extracted results, and judges whether the conditions for early warning are met according to the analysis results; if the result of the judgment is that the conditions for early warning are not met, time series analysis is performed through a risk prediction model based on the visual fire feature vector, the state fire feature vector, the equipment operation data and the historical time series data to generate fire risk trend prediction information; a comprehensive early warning level is generated according to the fire risk trend prediction information, and a device linkage control instruction matched with the comprehensive early warning level is generated. The problem of low efficiency caused by manual judgment and operation is overcome.

Description

Technical Field

[0001] This application relates to the field of safety monitoring and early warning technology, and in particular to a method and system for fire identification in a charging compartment based on video surveillance. Background Technology

[0002] Fire detection within charging compartments is a technical means of detecting and alarming potential fires using monitoring equipment installed inside and around the compartment. Current technologies for safety monitoring of specific spaces like charging compartments typically employ threshold alarm mechanisms based on single or a few types of sensors, combined with manual video surveillance. These traditional solutions generally suffer from problems such as limited monitoring dimensions, poor system coordination, delayed early warnings, and reliance on manual intervention in the response process. Specifically, they struggle to effectively capture early, subtle signs of fire; different subsystems, such as video and environmental sensing, operate in isolation, failing to form comprehensive judgments and mutual verification, leading to both false alarms and missed alarms; and the entire process lacks proactive analysis and predictive capabilities regarding risk development patterns, resulting in insufficient automation and intelligence in the closed-loop process from detection to response.

[0003] Secondly, with the development of intelligent technologies, some technical solutions have emerged that attempt to improve monitoring capabilities through data fusion. However, these solutions still face inherent technical bottlenecks when facing real and complex industrial application environments. Firstly, the reliability of multi-source heterogeneous data in collaborative analysis and fusion judgment is insufficient. When data is inconsistent, partially missing, or mismatched in time and space, it is difficult to make stable and accurate judgments. Secondly, in dynamic, multi-interference-source operating scenarios, the system's anti-interference capability and scenario adaptability are limited, making it difficult to accurately distinguish between interference generated by normal operating activities and real safety risk signs, thus affecting the system's practicality and reliability. Summary of the Invention

[0004] In order to solve one or more problems in the prior art, the main objective of this application is to provide a method and system for fire identification in charging compartments based on video surveillance.

[0005] To achieve the aforementioned objectives, this application proposes a method for fire detection in a charging compartment based on video surveillance, the method comprising: Real-time acquisition of multi-source monitoring data from the charging compartment, including video image data, equipment operating parameters, and environmental sensor data; Visual feature analysis is performed on the video image data to determine the visual feature vector; The environmental sensor data is analyzed for state characteristics to obtain a state feature vector; Based on the extracted results, anomaly analysis is performed on the visual feature vector and state feature vector, and the results of the analysis are used to determine whether the conditions for early warning are met. If the judgment result is that the conditions for early warning are not met, based on the visual fire feature vector, the status fire feature vector, the equipment operation data and historical time series data, a time series analysis is performed through a risk prediction model to generate fire risk trend prediction information. Based on the fire risk trend prediction information, a comprehensive early warning level is generated, and equipment linkage control commands are matched according to the comprehensive early warning level.

[0006] This application also provides a video surveillance-based fire detection system for charging compartments, including: The acquisition module is used to acquire multi-source monitoring data from the charging compartment in real time, including video image data, equipment operating parameters and environmental sensor data. The first analysis module is used to perform visual feature analysis on the video image data and determine the visual feature vector; The second analysis module is used to perform state feature analysis on the environmental sensor data to obtain a state feature vector; The judgment module is used to perform anomaly analysis on the visual feature vector and state feature vector based on the extracted results, and to determine whether the conditions for early warning are met based on the analysis results. The first generation module is used to generate fire risk trend prediction information by performing time-series analysis through a risk prediction model based on the visual fire feature vector, the status fire feature vector, the equipment operation data and historical time-series data, when the judgment result is that the conditions for early warning are not met. The second generation module is used to generate a comprehensive early warning level based on the fire risk trend prediction information, and to match the equipment linkage control commands according to the comprehensive early warning level.

[0007] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0008] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0009] The video surveillance-based fire detection method and system for charging compartments in this application, through simultaneous analysis of video images, equipment operating parameters, and environmental sensor data, can determine the presence of a fire from multiple perspectives, including visual appearance, equipment status, and the physical environment. For example, when a camera captures a suspected flame, it can check whether there is an abnormal temperature rise in the temperature sensor data of that area, thereby effectively filtering common interferences such as light reflections and significantly reducing the inherent risks of false alarms and missed alarms in traditional single-sensor solutions. Secondly, this method introduces risk prediction capabilities based on time-series data, realizing a shift from passive alarm to proactive early warning. When the conditions for immediate alarm are not met, it can comprehensively analyze current characteristics and historical data to determine the development trend of fire risk through a risk prediction model. This enables the system to identify potential precursors to thermal runaway and issue warnings before open flames or dense smoke appear, buying valuable time for early intervention and solving the problem of delayed early warning in traditional technologies. Finally, based on the severity of the prediction results, a graded early warning is generated, and matching equipment linkage control commands are automatically triggered, such as adjusting ventilation, cutting off power, or activating fire extinguishing devices. This shortens the emergency response time, improves the automation and intelligence level of overall safety management, and overcomes the inefficiency caused by the reliance on manual judgment and operation in traditional solutions. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a video surveillance-based fire detection method for charging compartments according to an embodiment of this application. Figure 2 This is a flowchart illustrating a video surveillance-based fire detection method for charging compartments according to an embodiment of this application. Figure 3 This is a schematic block diagram of a video surveillance-based fire detection system in a charging compartment according to an embodiment of this application. Figure 4 This is a schematic block diagram of the structure of a computer device according to an embodiment of this application.

[0011] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0013] Reference Figure 1 This application provides a method for fire detection in a charging compartment based on video surveillance, the method comprising: S1. Real-time acquisition of multi-source monitoring data from the charging compartment, including video image data, equipment operating parameters and environmental sensor data; S2. Perform visual feature analysis on the video image data to determine the visual feature vector; S3. Perform state feature analysis on the environmental sensor data to obtain a state feature vector; S4. Based on the extracted results, perform anomaly analysis on the visual feature vector and state feature vector, and determine whether the conditions for early warning are met based on the analysis results; S5. If the judgment result is that the conditions for early warning are not met, based on the visual fire feature vector, the state fire feature vector, the equipment operation data and historical time series data, a time series analysis is performed through the risk prediction model to generate fire risk trend prediction information. S6. Based on the fire risk trend prediction information, generate a comprehensive early warning level and match the equipment linkage control command according to the comprehensive early warning level.

[0014] As described in steps S1-S3 above, the principle of step one is to move away from relying on a single type of data source and instead simultaneously collect three types of key data: video image data, the charging equipment's own operating parameters, and independent environmental sensor data. For example, the system will simultaneously acquire real-time images of the cabin captured by the camera, current and voltage data output by the charging pile, and readings from dispersed temperature and gas sensors. This constructs a three-dimensional, complementary information matrix for subsequent analysis, jointly depicting the real-time safety status of the charging cabin from three dimensions: visual, equipment status, and physical environment. The principle of step two is to use computer vision technology to automatically process the video stream, transforming unstructured image information into structured, quantifiable feature descriptions. Specifically, through a pre-trained algorithm model, visual elements strongly related to fire in the image are identified and extracted, such as the specific color distribution and irregular jumping patterns of flames, or the translucent texture and diffusion direction of smoke. These extracted elements are combined into a data structure called a visual feature vector. This achieves automatic machine identification and quantification of visible fire signs, replacing inefficient and unreliable manual monitoring. Step three is similar in principle to step two but focuses on processing signals from various physical sensors. The principle involves cleaning, processing, and extracting features from this continuous, sometimes noisy, sensor data. For example, the system calculates the current temperature and its rate of change for a specific area from temperature sensor data, and obtains the concentration of a specific gas from gas sensor data. These values ​​are organized into another feature vector, namely the state feature vector. The effect of this step is to transform discrete sensor readings into quantitative indicators that characterize the evolution trend of the environmental state, providing precise physical parameter basis for judgment.

[0015] As described in steps S4-S6 above, step four is the first information fusion and decision point. The visual feature vectors and state feature vectors generated in the preceding steps are comprehensively compared and analyzed. Based on preset rules or models, these feature values ​​are evaluated to determine whether they deviate from the normal range. For example, when a suspected flame is identified in the visual feature vector, and the corresponding local temperature in the state feature vector also increases significantly, it is judged as highly abnormal. This enables rapid identification and alarm for open flames or smoke that have already shown significant characteristics, achieving real-time response to visible fires. Step five, assuming no immediate alarm is triggered at the current moment, correlates the current visual features, state features, and equipment operating data with historical data under similar operating conditions, and inputs this data into a specialized time series prediction model for calculation. This model can learn and capture the potential patterns of risk parameters changing over time, thereby predicting the development trajectory of fire risk over a future period, such as predicting whether the temperature of a certain battery cluster will accelerate to a dangerous level in the next few minutes. This achieves early risk insight and trend warning, significantly advancing the prevention and control point. Step six maps the severity of the trend prediction information output by the risk prediction model to different comprehensive early warning levels. Each level is pre-associated with a specific set of equipment linkage control commands. For example, when a slow temperature rise risk is predicted, a low-level early warning may be triggered, only recording and notification; when a rapid thermal runaway risk is predicted, a high-level early warning is triggered, automatically instructing the charging to stop and ventilation to start. This forms a complete closed loop of monitoring, early warning, and response, realizing intelligent linkage from risk prediction to automated intervention, and improving the speed and accuracy of emergency response.

[0016] As mentioned above, by simultaneously analyzing video images, equipment operating parameters, and environmental sensor data, the presence of a fire can be determined from multiple perspectives, including visual appearance, equipment status, and the physical environment. For example, when a camera captures a suspected flame, the temperature sensor data for that area can be checked for abnormal temperature rises, effectively filtering common interferences such as light reflections and significantly reducing the inherent risks of false alarms and missed alarms in traditional single-sensor solutions. Secondly, this method introduces risk prediction capabilities based on time-series data, realizing a shift from passive alarm to proactive early warning. Even when immediate alarm conditions are not met, the system can comprehensively analyze current characteristics and historical data to assess the development trend of fire risk through a risk prediction model. This enables the system to identify potential precursors to thermal runaway, issuing warnings before open flames or dense smoke appear, buying valuable time for early intervention and solving the problem of delayed warnings in traditional technologies. Finally, based on the severity of the prediction results, graded warnings are generated, and matching equipment linkage control commands are automatically triggered, such as adjusting ventilation, cutting off power, or activating fire extinguishing devices. This shortens emergency response time, improves the automation and intelligence level of overall safety management, and overcomes the inefficiency caused by reliance on manual judgment and operation in traditional solutions.

[0017] It should be added that this embodiment uses a risk prediction model to perform time-series analysis and generate fire risk trend prediction information. This embodiment constructs a dedicated time-series prediction model to assess the future trend of risk. Considering the need to process time-dependent sequence data and capture long-term dependencies, this embodiment uses a Long Short-Term Memory (LSTM) network as the basic model architecture. The model specifically includes the following layers: Input layer: receives multi-dimensional feature sequences with a fixed time step. LSTM hidden layer: consists of two stacked LSTM units. The first LSTM unit count is set to 128 for initial extraction of time-series features; the second LSTM unit count is set to 64 for further abstraction of advanced time-series patterns. A Dropout layer is introduced after each layer with a dropout rate of 0.3 to prevent overfitting during training. Fully connected output layer: transforms the LSTM output features of the last time step through a fully connected layer with 64 neurons, and finally outputs the final predicted value through a linearly activated neuron. The training process specifically involves collecting several months of historical data on the safe operation of the charging pod, including data from normal days, warning days, and a very small number of fire drills or real event days. Samples were extracted from historical data using a sliding window approach, and each sample was assigned a "future risk truth value" label. This label was determined by expert rules or post-event analysis based on whether an alarm or confirmed hazard would occur within the next 5 minutes, and the urgency of the hazard's development. For example, a label of 0 was assigned if nothing happened within the next 5 minutes, and 0.8 was assigned if a level 2 alarm occurred. The prepared dataset was divided into training, validation, and test sets in an 8:1:1 ratio. Mean squared error was used as the loss function. The Adam optimizer was used with an initial learning rate of 0.001 and a learning rate decay strategy. The batch size was set to 64. Training lasted for 200 epochs, and was terminated early when the validation set loss stopped decreasing after 10 consecutive epochs to preserve the optimal model parameters. To enhance the model's robustness in charging pod scenarios, noise samples such as transient sensor malfunctions and normal data fluctuations caused by routine maintenance were intentionally retained during data preprocessing, enabling the model to distinguish between abnormal risks and normal operational disturbances.

[0018] Reference Figure 2 In one embodiment, the step of performing anomaly analysis on the visual feature vector and state feature vector, and determining whether the conditions for early warning are met based on the analysis results, includes: S41. Compare each feature item in the visual feature vector with the corresponding preset mapping relationship, and generate a first quantitative score based on the comparison result of each feature item. S42. Compare each feature item in the state feature vector with its corresponding preset mapping relationship, and generate a second quantitative score based on the comparison result of each feature item. S43. According to the preset weighting strategy, the first quantitative score and the second quantitative score are fused and calculated to obtain the comprehensive risk value; S44. Compare the comprehensive risk value with the preset warning threshold. If the comprehensive risk value is greater than or equal to the warning threshold, it is determined that the warning condition is met. S45. If the comprehensive risk value is less than the warning threshold, it is determined that the warning conditions are not met.

[0019] As described above, Step One transforms abstract visual features into unified, comparable risk scores. For each item in the visual feature vector, such as the flickering frequency of a flame or the diffusion rate of smoke, a mapping relationship is pre-defined. This relationship defines the conversion rules from specific feature values ​​to a risk score component. For example, when a high value is detected in the flickering frequency of a flame, it is converted into a higher risk score by looking up the corresponding mapping relationship. The goal is to normalize visual features of different properties and dimensions, such as color, shape, and motion, into the same risk language. Step Two processes state features extracted from environmental sensor data. For example, for temperature features, a rule is pre-defined that maps temperature values ​​and temperature rise rates to risk scores. When the temperature exceeds a certain threshold or the temperature rises abnormally fast, the corresponding state feature item receives a higher risk score. This step also transforms physical sensor signals into standardized risk quantification indicators, allowing different types of state parameters, such as temperature and gas concentration, to be compared and used in calculations. Step Three involves weighted information fusion to arrive at a holistic, comprehensive risk assessment. A weight is preset for both visual feature scoring and state feature scoring. This weight can reflect prior knowledge or strategic emphasis on the reliability of different data sources. For example, in stable lighting conditions, video analysis may be given a higher weight; while in dusty conditions, the weight of sensor data may be appropriately increased. The scores of the two are added together according to their weights to obtain the comprehensive risk value. This overcomes the limitations of a single information source and achieves a more robust and comprehensive risk assessment through weighted balancing. Step four compares the calculated comprehensive risk value with a preset warning threshold. If the comprehensive risk value reaches or exceeds the threshold, the warning condition is met, triggering subsequent processes; otherwise, it is deemed not to meet the condition. This transforms a complex, multi-feature fusion analysis process into a clear, binary warning instruction, ensuring the clarity and executability of the decision.

[0020] This embodiment addresses the reliability of multi-source information fusion and its anti-interference capability in complex scenarios mentioned in the background technology. By setting an independent mapping relationship for each feature, the correlation between different symptoms and risk levels can be precisely characterized, rather than a simple "present / absent" judgment. More importantly, by introducing a weighting strategy for fusion calculation, the method acknowledges the differences in the reliability of different data sources in different scenarios, allowing the system to perform intelligent balancing. For example, when a video is briefly interfered with, reliable judgment can be made based on sensor data with higher weights, thereby avoiding overall misjudgment that may occur due to the failure of a single path in traditional solutions.

[0021] In one embodiment, the visual feature vector includes one or more of flame features and smoke features identified based on the video image data; the state feature vector includes one or more of local high-temperature area temperature features and preset gas concentration features identified based on the environmental sensing data.

[0022] As mentioned above, flame characteristics are based on the physical properties of flames in terms of color, shape, and dynamics. For example, flames typically have high red and orange channel values ​​in the RGB color space, and their shapes are irregular and flicker and jump violently over time. Flame characteristics are determined by identifying connected regions in the image that conform to these color models and have specific flicker frequencies. Smoke characteristics are based on the visual texture, translucency, and diffusion characteristics of smoke. For example, smoke typically appears as a grayish-white, unevenly textured, translucent area that diffuses and spreads in space over time. Smoke characteristics are determined by analyzing the motion vectors, transparency changes, and expansion trends of specific textured areas in the video sequence. By precisely focusing video analysis on the two types of visual phenomena most directly related to fire, the system can mimic the core judgment criteria used by humans when observing fires, efficiently extracting key risk clues from massive amounts of video information, and providing clear targets for subsequent quantitative scoring. Temperature characteristics of local high-temperature areas are monitored by tracking the abnormal distribution of the temperature field within the charging compartment. Using thermal imaging sensors or high-density point temperature sensors, abnormal hot spots deviating from the ambient background temperature can be identified, and their temperature values ​​and rates of increase over time can be tracked. This directly corresponds to the core exothermic phenomenon during battery thermal runaway and other malfunctions. Preset gas concentration characteristics: This monitors the concentration levels of specific gases. For lithium battery fire risks, the system focuses on monitoring the concentrations of characteristic gases such as hydrogen, carbon monoxide, or electrolyte volatiles. When an internal battery malfunctions, these gases may be released before open flames and smoke. It directly monitors the essential driving factors of fire at a physical level—abnormal energy release (manifested as high temperature) and abnormal material transformation (release of characteristic gases). This provides objective physical parameters independent of video evidence, especially suitable for early warning of fires that are internal or visually obstructed.

[0023] In one embodiment, the step of comparing each feature item in the visual feature vector with its corresponding preset mapping relationship and generating a first quantitative score based on the comparison results of each feature item includes: If flame features are identified in the visual feature vector, the duration of the flame and the rate of expansion of the flame area are analyzed. Based on the analysis results, and according to the first preset mapping rule, the corresponding flame feature scoring components are mapped. If smoke features are identified in the visual feature vector, the concentration of the smoke features and the diffusion coverage rate of the smoke are analyzed. Based on the analysis results, and according to the second preset mapping rule, the corresponding smoke feature score components are mapped. The flame feature score component and the smoke feature score component are weighted and summed according to preset weights to generate the first quantitative score.

[0024] As mentioned above, the principle behind generating flame feature scores is not only to identify the presence of a flame, but more importantly, to assess whether its dynamic behavior matches the development pattern of a real fire. This involves two analytical dimensions: First, analyzing the duration of the flame's existence. This is based on the principle that real flames are continuous, while many interfering sources, such as the instantaneous reflection of vehicle headlights or welding sparks, are often fleeting. The system calculates the duration for which the identified flame area appears continuously across multiple video frames. The longer the duration, the lower the likelihood of it being a transient interference. Second, analyzing the area expansion rate of the flame area. This is based on the principle that the flame area expands over time during fire growth. The system calculates the expansion rate by comparing the changes in the flame pixel area between consecutive frames. The faster the rate, the more rapidly the fire develops and the more imminent the risk. These two dynamic parameters are input into a preset mapping rule, which comprehensively evaluates the duration and expansion rate, transforming them into a standardized flame feature score component. By introducing the temporal and spatial growth dimensions, the accuracy of identifying real flames is greatly improved, effectively filtering out most static or transient fire-like interference. The scoring of smoke features follows a similar logic to flame analysis, focusing on the dynamic characteristics of smoke. The analysis also includes two aspects: first, analyzing the concentration of smoke features. This concentration is a relative value estimated using computer vision algorithms based on visual characteristics such as texture density and opacity of suspected smoke areas in video images. Higher concentration generally indicates denser smoke. Second, analyzing the rate of smoke diffusion and coverage. This involves observing the speed at which smoke spreads in space, measured by tracking the movement of smoke edges or calculating the growth rate of its coverage area. Faster diffusion indicates that smoke is rapidly filling the space, posing a higher risk. These two parameters are input into another independent pre-defined mapping rule, outputting a smoke feature score component. This allows the system to distinguish between slowly accumulating water vapor, fixed-location dust, and rapidly spreading fire smoke, enhancing the reliability of smoke identification. Regarding the weighted fusion of the score components, the obtained flame feature score component and smoke feature score component are finally weighted and summed according to a pre-defined weight to generate the first quantitative score. The principle is to acknowledge that different features may have different indicative values ​​in different scenarios. For example, in scenarios with low visibility but obvious heat sources, flame characteristics could perhaps be given higher weight. This achieves intelligent weighing and synthesis within visual evidence, forming a single quantitative indicator that can comprehensively reflect fire risk within the visible light range.

[0025] In one embodiment, the step of comparing each feature item in the state feature vector with its corresponding preset mapping relationship and generating a second quantitative score based on the comparison results of each feature item includes: Based on the state feature vector, analyze the highest temperature value and temperature rise rate of the local high temperature region. Based on the analysis results, and in accordance with the third preset mapping rule, the results are mapped to the corresponding temperature feature score components. Analyze the concentration value and duration of the preset gas based on the preset gas concentration characteristics; Based on the analysis results, and in accordance with the fourth preset mapping rule, the corresponding gas characteristic score components are mapped. The temperature feature score component and the gas feature score component are weighted and summed according to preset weights to generate the second quantitative score.

[0026] As mentioned above, the principle behind generating the temperature characteristic score component is to move beyond a single temperature threshold in judging localized high temperatures and instead comprehensively evaluate it from two dimensions: absolute intensity and trend of change. First, the highest temperature value is analyzed. This directly measures the energy release level of the thermal anomaly. The instantaneous temperature value of the current hottest point is located and read from sensor network data. This value directly reflects the core intensity of the heat source and is the fundamental physical basis for judging the severity of the risk. Second, the rate of temperature rise is analyzed. This captures the dynamic development process of the thermal anomaly. By calculating the rate of temperature increase in the high-temperature area per unit time, such as the degree Celsius increase per second, the urgency of thermal runaway is assessed. A point of slow temperature rise may be normal heat dissipation by the equipment, while a point of rapid temperature increase is highly likely to be a clear signal of malfunction or fire. These two key parameters are input into a specially designed third preset mapping rule. The rule comprehensively considers both the absolute temperature value and the rate of temperature rise, outputting a standardized temperature characteristic score component. It can effectively distinguish between the stable heat generated during normal equipment operation and the dangerous, dynamically developing thermal runaway of batteries within the charging compartment, improving the targeting and accuracy of temperature monitoring and early warning. The design logic for generating gas characteristic scoring components is similar to that of temperature analysis, aiming to provide a more reliable assessment of gas concentration. First, it analyzes the concentration values ​​of preset gases, directly monitoring the accumulation level of specific hazardous gases. It reads real-time concentration readings from target sensors such as hydrogen and carbon monoxide. The concentration value directly characterizes the scale of flammable or toxic gas leakage. Second, it analyzes the duration of this concentration value. The principle is to eliminate transient interference or sensor false alarms. A brief concentration spike may originate from environmental disturbances, while a sustained high concentration strongly points to a continuous gas release source. The system records the length of time the concentration exceeds the attention threshold. These two parameters are input into a fourth preset mapping rule to generate gas characteristic scoring components. This enhances the ability to identify early, minute but continuous gas leaks, while suppressing false alarms caused by occasional sensor fluctuations or brief environmental interference, making gas monitoring results more stable and reliable. The system then performs weighted fusion of the scoring components. Finally, the calculated temperature and gas characteristic score components are weighted and summed according to preset weights to generate a second quantitative score. The principle is based on the recognition that the indicative significance and reliability of temperature and gas evidence differ across different scenarios or failure modes. For example, temperature signals may be more sensitive in the early stages of some battery thermal runaway events; while characteristic gas signals may appear earlier in electrolyte leakage. Through weight adjustment, a more balanced and intelligent comprehensive evaluation of the environmental state can be achieved.

[0027] This embodiment enhances the reliability and early warning value of state feature analysis to address the aforementioned challenges of sensor susceptibility to interference and delayed warnings. Traditional solutions rely solely on single concentration or temperature exceedance thresholds, failing to distinguish between normal fluctuations and genuine emergencies, nor can they capture trends of accelerating risk deterioration. This method introduces a dynamic parameter (temperature rise rate, duration) representing the trend of change for each type of feature, upgrading the analysis from static "point" judgment to dynamic "line" tracking. This enables the system to keenly detect the trend of risk parameters "worsening," rather than merely their "already broken" state, thus achieving earlier warnings.

[0028] In one embodiment, prior to the step of performing anomaly analysis on the visual feature vector and the state feature vector, the method further includes: Obtain the spatial configuration information of the charging compartment, which includes the physical location coordinates of each point environmental sensor and the effective field of view of each video monitoring device in the spatial coordinate system of the charging compartment. During the generation of the state feature vector, if the value of a state feature item associated with a preset physical location is found to exceed a preset safety threshold, then based on the spatial configuration information, it is determined whether the preset physical location is within the effective field of view of any video surveillance device used for fire identification. If the judgment result is yes, then the standard fusion analysis strategy is executed to perform weighted fusion of the visual feature vector and the state feature vector to perform the anomaly analysis. If the judgment result is negative, the visual missing fusion strategy is executed, ignoring the fusion weight of the visual feature vector, and the anomaly analysis is performed based on the state feature vector and the associated device operating parameters.

[0029] As mentioned above, the first step is to establish a digital spatial knowledge base for the system. During implementation, this requires pre-entering or obtaining, through calibration measurements, the precise physical coordinates of all point-based environmental sensors within the charging compartment, as well as the spatial range clearly observable by each video surveillance device. Essentially, this involves building a virtual map within the computer that reflects the actual monitoring layout. This allows the system to not only know what data it has received, but also to clearly identify the specific location in space from which this data was collected, and whether that location is visible to the camera, providing crucial spatial context for subsequent intelligent judgment. The second step is to perform real-time, location-based data validity verification. When the system processes environmental sensor data to generate state feature vectors, if it finds that the sensor reading associated with a specific location exceeds a safety threshold, it immediately triggers a query. Using the spatial configuration information established in the first step, it determines whether the alarm sensor location falls within the effective field of view of any fire detection camera. It dynamically distinguishes between two different alarm scenarios: one where the camera can "see" the alarm point, providing visual confirmation; and another where the camera "cannot see" the alarm point, lacking visual information at that location. This judgment is the basis for subsequent differentiated processing. Step three involves dynamically selecting and executing the most reasonable analysis strategy based on the judgment result of the perceived state, demonstrating the flexibility of the decision-making mechanism. If the judgment result is yes, meaning the alarm point is within the camera's field of view, the standard fusion analysis strategy is executed. This means the system believes it can obtain visual information related to the alarm location, and therefore, according to the established process, performs weighted fusion analysis on the visual feature vector and the state feature vector, using video and environmental data for mutual verification to pursue the highest judgment accuracy. Step four involves switching to a visual omission fusion strategy if the judgment result is no, meaning the alarm point is outside the field of view of all cameras. Its core operation is to ignore the fusion weight of the visual feature vector, essentially temporarily shielding potentially irrelevant or invalid visual information input in the current analysis. At the same time, the system will rely heavily on the state feature vector that triggered the alarm itself, and combine it with the operating parameters of the equipment associated with the sensor location for a comprehensive risk assessment. For example, if a hydrogen sensor located in a corner alarms, although the camera cannot see that corner, the system can retrieve the current and voltage data of the charging pile charging the battery cluster in that corner to check for abnormal fluctuations, thereby assisting in determining the authenticity of the sensor alarm. Even in the absence of visual evidence, the system avoids underreporting due to rigid requirements for complete multi-source information. Instead, it can make conservative but timely risk assessments based on available reliable information, thereby improving the system's coverage and usability in complex spatial layouts.

[0030] This embodiment addresses the first deep-seated deficiency in in-depth analysis: the spatiotemporal consistency problem of multi-source data fusion. It specifically tackles two technical challenges: "sensor alarms, video shows no anomalies" and "spatiotemporal asynchrony." Under existing technology, when a temperature or gas sensor located in a camera blind spot alarms first, the system cannot obtain visual confirmation of that location. Due to rigid fusion rules, a single sensor alarm may be misjudged as interference, leading to missed detections of early thermal runaway or gas leaks. Similarly, when the sensor and camera fields of view are mismatched, forcibly fusing data from different spatial locations can distort risk assessment. This embodiment aims to overcome these challenges. For example, it first understands "what I can see and what I cannot see," and then flexibly adjusts its analysis logic based on this understanding, enabling effective risk assessment in both complete and incomplete information scenarios. In summary, this embodiment is based on an intelligent fusion routing mechanism that self-assesses perception capabilities. It upgrades data fusion from a fixed calculation formula that treats all inputs equally to an intelligent process with spatial cognition that dynamically selects paths based on the physical credibility of the information source. This invention solves the problem of fusion failure caused by the inability to ideally align multi-source data due to physical layout limitations in existing technologies. In particular, it ensures that even in extreme cases of "missing visual features," such as when a sensor alarms in a blind zone, the system will not delay judgment by waiting for visual evidence that will never appear, nor will it make misjudgments by forcibly fusing unrelated visual data. Instead, it can initiate a degraded but reliable backup analysis process to make decisions based on state features and device parameters.

[0031] In one embodiment, prior to performing anomaly analysis on the visual feature vector and state feature vector, the steps include: Acquire real-time operational status data of the charging compartment, including charging equipment start / stop and power information; Based on the real-time operation status data, the feature items in the visual feature vector and the status feature vector are analyzed and identified. If a flame feature is detected, it is determined whether the location of the flame feature coincides with a known fixed light source or reflective marker, and the corresponding flame feature score component is lowered based on the determination result. If smoke features are detected, it is determined whether the current operation is in a preset stage that is prone to generating dust or aerosols, and the corresponding smoke feature score is lowered according to the determination result. If a local high-temperature area temperature characteristic is identified, it is determined whether the high-temperature area matches the heat dissipation port of the charging device in high-power operation state, and the corresponding temperature characteristic score component is lowered according to the matching result. Anomaly analysis is performed based on the adjusted visual feature vector and the state feature vector.

[0032] As mentioned above, after the visual score is generated but before the final anomaly analysis, a feature credibility verification and correction process based on the real-time operational context is introduced. By understanding what the charging compartment is "doing," the system intelligently assesses what the monitored features "mean," thereby giving the system the ability to distinguish between appearance and essence. Step one provides the system with objective evidence to understand the current scenario. It continuously collects key operating condition information such as the start / stop status and real-time power of the charging equipment. For example, it knows whether the current charging is at full power, trickle charging, or in a shutdown state. A dynamic scene background is constructed, so that the system no longer views the perceived features in isolation but interprets them within the context of specific operational activities. Steps two to four involve designing dedicated logical rules based on operational status and spatial knowledge for each type of easily confused feature, and conducting a rationality review and dynamic correction of the initially extracted feature score components. For flame features, a position overlap judgment is performed. The principle is that many types of fire interference sources are fixed in spatial location. When video analysis identifies a flame feature, it immediately checks whether the center of the flame area overlaps with the pre-stored known fixed light source coordinates, high-reflectivity marker locations, or authorized temporary hot work areas in the system. If there is overlap, there is sufficient reason to believe that the current identification result is likely interference from these non-fire source light sources, and therefore the flame characteristic score component is significantly reduced according to the rules. It can effectively filter out false alarms caused by continuous or flickering light spots from cabin lighting, vehicle rearview mirrors, maintenance flashlights, etc. For smoke characteristics, the system performs operational phase correlation judgment. The principle is that the generation of certain types of smoke interference has a temporal regularity. It determines whether the current moment is within a preset operational phase that is prone to generating dust or aerosols, such as daily scheduled cleaning periods or equipment startup periods after maintenance. If smoke characteristics are identified and the current operational phase is such, the system will reasonably suspect that the smoke may originate from cleaning dust or equipment oil mist, thus reducing its smoke characteristic score component. During specific high-interference periods, it can automatically reduce the sensitivity to smoke characteristics, avoiding frequent false alarms caused by daily operational activities. For temperature characteristics of localized high-temperature areas, the system performs heat source matching judgment. Normal equipment heat dissipation has its fixed outlet and a power-related temperature rise pattern. When thermal imaging analysis identifies a localized high-temperature area, the system compares its location with the heat dissipation vent of a charging device operating at high power. If the locations match, and the device is indeed operating at high power, the high temperature is likely due to normal heat dissipation, thus lowering its temperature feature score component. This distinguishes between normal high temperatures at the charging pile's cooling fan outlet and abnormal overheating on the battery pack surface, preventing unnecessary thermal alarms from being triggered during normal device operation. Step five uses the visual and state feature vectors, which have undergone the aforementioned context verification and may have had their risk scores lowered, for subsequent anomaly analysis.

[0033] This embodiment addresses the problem of scene confusion. While existing technologies can accurately extract physical features such as flames, smoke, and high temperatures, in the dynamic industrial environment of a charging cabin, many fire-related symptoms overlap with those observed during normal operations. Without distinguishing between these two, any high-precision feature extraction will ultimately be drowned out by false alarms, rendering the system unusable. Therefore, this method does not simply draw conclusions after feature extraction but adds a knowledge-based rationality analysis step.

[0034] In one feasible embodiment, the warning sensitivity should not be the same when charging and when not charging, otherwise frequent activation will occur. To solve the above technical problem, the proposed solution further includes the following steps: The current operating status signal of the charging compartment is obtained, and the current operating status signal is used to indicate whether the charging operation is in progress or not. Based on the current operation status signal, a target warning parameter set matching the current operation status is selected from at least two preset warning parameter sets; wherein, the warning parameter set includes at least a preset safety threshold, a warning threshold, and a confirmation duration parameter. When the current operation status signal indicates that charging operation is in progress, the first early warning parameter set is used to perform the anomaly analysis, early warning condition judgment and early warning level confirmation. When the current operation status signal indicates that the charging operation is not in progress, the second early warning parameter set is used to perform the anomaly analysis, early warning condition judgment, and early warning level confirmation; wherein, the first early warning parameter set has a higher early warning sensitivity than the second early warning parameter set.

[0035] As described above, a dynamic adjustment mechanism is introduced to dynamically switch the scale and rhythm of the entire early warning judgment based on the most fundamental operational status of the charging compartment—whether it is performing the high-risk operation of charging—so that the system's alertness matches the current actual risk level. The first step is to perform a macro-level determination of the operational mode. By monitoring the total power status and power output of the charging pile, or receiving signals from charging management, it is determined whether the charging compartment is currently in operation or not. The second step is to implement and invoke pre-set strategies. At least two complete sets of early warning parameter packages are pre-stored, each containing key parameters required for anomaly analysis, threshold comparison, and level confirmation, such as preset safety thresholds for comparison, early warning thresholds for final judgment, and confirmation duration parameters for anti-vibration. Once the operational status is determined in Step 1, the corresponding parameter set is automatically selected. This directly links different operational statuses to different judgment criteria, preparing for different behavioral modes in different scenarios. The third step is the strategy execution phase, applying differentiated analysis strategies based on the operational status. When the system determines that charging is in progress, the first early warning parameter set is automatically activated. This parameter set is configured for higher warning sensitivity. Specifically, it may include lower temperature or gas concentration safety thresholds, allowing the system to detect minor anomalies earlier; it may also include lower overall risk warning thresholds, making alarm conditions easier to meet; and it may set shorter confirmation durations, allowing the system to respond faster to persistent risk signals. Conversely, when charging is not in progress, a second warning parameter set is activated. This parameter set has relatively lower warning sensitivity. Higher thresholds can be set for each parameter, requiring more obvious anomalies to trigger attention; a longer confirmation duration is also set, requiring the abnormal signal to persist more stably for a period of time before being confirmed as an alarm, filtering out more random environmental interference. During the relatively safe period when not charging, it acts like a robust patrolman, tending to "confirm before acting," thus avoiding frequent false alarms caused by routine factors such as equipment residual heat, personnel activity, and daily dust, protecting related equipment from unnecessary frequent start-ups and shutdowns.

[0036] Reference Figure 3 This application also provides a video surveillance-based fire detection system for charging compartments, comprising: Acquisition module 1 is used to acquire multi-source monitoring data from the charging compartment in real time, including video image data, equipment operating parameters and environmental sensor data; The first analysis module 2 is used to perform visual feature analysis on the video image data and determine the visual feature vector; The second analysis module 3 is used to perform state feature analysis on the environmental sensor data to obtain a state feature vector. The judgment module 4 is used to perform anomaly analysis on the visual feature vector and state feature vector based on the extracted results, and to determine whether the conditions for early warning are met based on the analysis results. The first generation module 5 is used to generate fire risk trend prediction information by performing time-series analysis through a risk prediction model based on the visual fire feature vector, the state fire feature vector, the equipment operation data and historical time-series data, when the judgment result is that the conditions for early warning are not met. The second generation module 6 is used to generate a comprehensive early warning level based on the fire risk trend prediction information, and to generate equipment linkage control commands matched according to the comprehensive early warning level.

[0037] As described above, it is understood that each component of the video surveillance-based fire identification system for charging compartments proposed in this application can achieve the function of any of the video surveillance-based fire identification methods for charging compartments described above, and the specific structure will not be repeated here.

[0038] Reference Figure 4 This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores monitoring data and other data. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a video surveillance-based fire detection method for charging compartments.

[0039] The processor described above executes the video surveillance-based fire identification method for charging compartments, comprising: acquiring multi-source monitoring data from the charging compartment in real time, the multi-source monitoring data including video image data, equipment operating parameters, and environmental sensor data; performing visual feature analysis on the video image data to determine visual feature vectors; performing state feature analysis on the environmental sensor data to obtain state feature vectors; performing anomaly analysis on the visual feature vectors and state feature vectors based on the extracted results, and determining whether the conditions for early warning are met based on the analysis results; if the determination result is that the conditions for early warning are not met, performing time-series analysis through a risk prediction model based on the visual fire feature vectors, the state fire feature vectors, the equipment operating data, and historical time-series data to generate fire risk trend prediction information; generating a comprehensive early warning level based on the fire risk trend prediction information, and matching equipment linkage control commands according to the comprehensive early warning level.

[0040] One embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for fire identification in a charging compartment based on video surveillance, comprising the following steps: real-time acquisition of multi-source monitoring data from the charging compartment, the multi-source monitoring data including video image data, equipment operating parameters, and environmental sensor data; performing visual feature analysis on the video image data to determine a visual feature vector; performing state feature analysis on the environmental sensor data to obtain a state feature vector; based on the extracted results, performing anomaly analysis on the visual feature vector and the state feature vector, and determining whether the conditions for an early warning are met based on the analysis results; if the determination result is that the conditions for an early warning are not met, performing time-series analysis through a risk prediction model based on the visual fire feature vector, the state fire feature vector, the equipment operating data, and historical time-series data to generate fire risk trend prediction information; generating a comprehensive early warning level based on the fire risk trend prediction information, and matching equipment linkage control commands according to the comprehensive early warning level.

[0041] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0042] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0043] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for fire detection in a charging compartment based on video surveillance, characterized in that, The method includes: Real-time acquisition of multi-source monitoring data from the charging compartment, including video image data, equipment operating parameters, and environmental sensor data; Visual feature analysis is performed on the video image data to determine the visual feature vector; The environmental sensor data is analyzed for state characteristics to obtain a state feature vector; Based on the extracted results, anomaly analysis is performed on the visual feature vector and state feature vector, and the results of the analysis are used to determine whether the conditions for early warning are met. If the judgment result is that the conditions for early warning are not met, based on the visual fire feature vector, the status fire feature vector, the equipment operation data and historical time series data, a time series analysis is performed through a risk prediction model to generate fire risk trend prediction information. Based on the fire risk trend prediction information, a comprehensive early warning level is generated, and equipment linkage control commands are matched according to the comprehensive early warning level.

2. The method for fire identification in a charging compartment based on video surveillance according to claim 1, characterized in that, The steps of performing anomaly analysis on the visual feature vector and state feature vector, and determining whether the conditions for early warning are met based on the analysis results, include: Each feature item in the visual feature vector is compared with its corresponding preset mapping relationship, and a first quantitative score is generated based on the comparison results of each feature item. Each feature item in the state feature vector is compared with its corresponding preset mapping relationship, and a second quantitative score is generated based on the comparison results of each feature item. According to the preset weighting strategy, the first quantitative score and the second quantitative score are fused and calculated to obtain the comprehensive risk value; The comprehensive risk value is compared with a preset warning threshold. If the comprehensive risk value is greater than or equal to the warning threshold, it is determined that the warning condition is met. If the overall risk value is less than the warning threshold, it is determined that the warning conditions are not met.

3. The method for fire identification in a charging compartment based on video surveillance according to claim 2, characterized in that, The visual feature vector includes one or more of the flame features and smoke features identified based on the video image data; the state feature vector includes one or more of the local high-temperature area temperature features and preset gas concentration features identified based on the environmental sensing data.

4. The method for fire identification in a charging compartment based on video surveillance according to claim 3, characterized in that, The step of comparing each feature item in the visual feature vector with its corresponding preset mapping relationship and generating a first quantitative score based on the comparison results of each feature item includes: If flame features are identified in the visual feature vector, the duration of the flame and the rate of expansion of the flame area are analyzed. Based on the analysis results, and according to the first preset mapping rule, the corresponding flame feature scoring components are mapped. If smoke features are identified in the visual feature vector, the concentration of the smoke features and the diffusion coverage rate of the smoke are analyzed. Based on the analysis results, and according to the second preset mapping rule, the corresponding smoke feature score components are mapped. The flame feature score component and the smoke feature score component are weighted and summed according to preset weights to generate the first quantitative score.

5. The method for fire identification in a charging compartment based on video surveillance according to claim 3, characterized in that, The step of comparing each feature item in the state feature vector with its corresponding preset mapping relationship and generating a second quantitative score based on the comparison results of each feature item includes: Based on the state feature vector, analyze the highest temperature value and temperature rise rate of the local high temperature region. Based on the analysis results, and in accordance with the third preset mapping rule, the results are mapped to the corresponding temperature feature score components. Analyze the concentration value and duration of the preset gas based on the preset gas concentration characteristics; Based on the analysis results, and in accordance with the fourth preset mapping rule, the corresponding gas characteristic score components are mapped. The temperature feature score component and the gas feature score component are weighted and summed according to preset weights to generate the second quantitative score.

6. The method for fire identification in a charging compartment based on video surveillance according to claim 2, characterized in that, Before the step of performing anomaly analysis on the visual feature vector and the state feature vector, the method further includes: Obtain the spatial configuration information of the charging compartment, which includes the physical location coordinates of each point environmental sensor and the effective field of view of each video monitoring device in the spatial coordinate system of the charging compartment. During the generation of the state feature vector, if the value of a state feature item associated with a preset physical location is found to exceed a preset safety threshold, then based on the spatial configuration information, it is determined whether the preset physical location is within the effective field of view of any video surveillance device used for fire identification. If the judgment result is yes, then the standard fusion analysis strategy is executed to perform weighted fusion of the visual feature vector and the state feature vector to perform the anomaly analysis. If the judgment result is negative, the visual missing fusion strategy is executed, ignoring the fusion weight of the visual feature vector, and the anomaly analysis is performed based on the state feature vector and the associated device operating parameters.

7. The method for fire identification in a charging compartment based on video surveillance according to claim 4, characterized in that, Prior to performing anomaly analysis on the visual feature vector and state feature vector, the steps include: Acquire real-time operational status data of the charging compartment, including charging equipment start / stop and power information; Based on the real-time operation status data, the feature items in the visual feature vector and the status feature vector are analyzed and identified. If a flame feature is detected, it is determined whether the location of the flame feature coincides with a known fixed light source or reflective marker, and the corresponding flame feature score component is lowered based on the determination result. If smoke features are detected, it is determined whether the current operation is in a preset stage that is prone to generating dust or aerosols, and the corresponding smoke feature score is lowered according to the determination result. If a local high-temperature area temperature characteristic is identified, it is determined whether the high-temperature area matches the heat dissipation port of the charging device in high-power operation state, and the corresponding temperature characteristic score component is lowered according to the matching result. Anomaly analysis is performed based on the adjusted visual feature vector and the state feature vector.

8. A fire detection system for charging compartments based on video surveillance, characterized in that, include: The acquisition module is used to acquire multi-source monitoring data from the charging compartment in real time, including video image data, equipment operating parameters and environmental sensor data. The first analysis module is used to perform visual feature analysis on the video image data and determine the visual feature vector; The second analysis module is used to perform state feature analysis on the environmental sensor data to obtain a state feature vector; The judgment module is used to perform anomaly analysis on the visual feature vector and state feature vector based on the extracted results, and to determine whether the conditions for early warning are met based on the analysis results. The first generation module is used to generate fire risk trend prediction information by performing time-series analysis through a risk prediction model based on the visual fire feature vector, the status fire feature vector, the equipment operation data and historical time-series data, when the judgment result is that the conditions for early warning are not met. The second generation module is used to generate a comprehensive early warning level based on the fire risk trend prediction information, and to match the equipment linkage control commands according to the comprehensive early warning level.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images