Airplane cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage

Abstract

The invention relates to the technical field of aviation safety monitoring, in particular to an aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage, and the system comprises a visible light image collection module which is used for obtaining a real-time visible light image in an aircraft cargo hold; the infrared thermal imaging acquisition module is used for acquiring multi-region and multi-channel thermal imaging images in the cargo hold and constructing a thermal distribution diagram; the smoke and gas sensing device is used for monitoring the concentration change of hazardous gas in the air and outputting continuous detection data; the multi-modal AI identification processing module is used for carrying out fusion analysis on the multi-modal data; the pod linkage verification module is used for performing external verification sampling on the target area when the multi-modal analysis result is low in confidence but has risks, and forming a final judgment result in an auxiliary manner; and the alarm output module is used for converting the final judgment result into an alarm signal, sending the alarm signal to a cockpit HUD or a main control system through a flight control system, and automatically switching a prompt mode according to the fire behavior grade.

Description

Technical Field

[0001] The embodiments of this application relate to the field of aviation safety monitoring technology, and in particular to an aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage. Background Technology

[0002] Aircraft cargo holds, being enclosed, limited spaces that typically lack manual inspection, pose a significant threat to flight safety should a fire occur. Currently, most aircraft cargo holds are equipped only with smoke detectors or a single temperature sensor. These detection mechanisms are simplistic, have slow response times, and are prone to false alarms and missed alarms due to cargo obstruction, interference from abnormal heat sources, or uneven airflow.

[0003] With the diversification of flight missions and the increasing complexity of cargo types, traditional single-modal fire detection methods are no longer sufficient to meet the early fire identification requirements of aircraft cargo holds, which demand high confidence, low latency, and multi-source fusion. In recent years, infrared thermal imaging and artificial intelligence technologies have been widely used in security and industrial monitoring, providing feasibility for building multimodal intelligent fire identification systems. However, currently, there is no AI-infrared thermal image collaborative analysis mechanism designed for special environments such as aircraft cargo holds, especially lacking integrated solutions for non-fire source heat body identification, confidence fusion judgment, and external pod linkage verification. Summary of the Invention

[0004] To address the aforementioned technical issues, embodiments of this application propose an aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage. The system aims to improve the accuracy of aircraft cargo hold fire identification and response efficiency, and enhance the intelligent safety protection capabilities of the aircraft cargo hold through multi-source data fusion, deep learning model analysis, dynamic threshold control, and pod linkage verification.

[0005] To achieve the above objectives, embodiments of this application propose an aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage, applicable to aircraft. The system includes: a visible light image acquisition module for acquiring real-time visible light images of the aircraft cargo hold interior and providing image frames for smoke detection and visual feature extraction; an infrared thermal imaging acquisition module for acquiring multi-area, multi-channel thermal imaging images of the cargo hold interior and constructing a heat distribution map based on the thermal imaging images; and a smoke and gas sensing device, including... sensor, Sensors and particulate matter sensors are used to monitor changes in the concentration of hazardous gases in the air and output continuous detection data. A multimodal AI recognition and processing module is used to fuse and analyze detection data from visible light images, thermal imaging images, heat distribution maps, and smoke and gas sensors. This multimodal AI recognition and processing module consists of a thermal image detection and prediction unit, a thermal image adaptive learning and threshold adjustment unit, a visible light smoke recognition unit, a smoke and gas analysis unit, a heat source liveness detection unit, and a multimodal fusion decision unit. The thermal image detection and prediction unit detects temperature rise anomalies and predicts heat diffusion trends based on time-series heat distribution maps. The thermal image adaptive learning and threshold adjustment unit learns the baseline of the heat distribution map based on the cargo hold structure and cargo loading status and dynamically adjusts the temperature rise anomaly recognition threshold. The visible light smoke recognition unit identifies suspected smoke based on multi-scale texture and morphological features of visible light images. The system comprises the following components: a smoke and gas analysis unit, which extracts gas concentration abrupt change characteristics based on the detection data output by smoke and gas sensors and analyzes the fire trend based on these characteristics; a heat source liveness detection unit, which combines thermal imaging images and target contours to identify non-fire source heat sources; a multimodal fusion decision unit, which fuses the identification results of the above units and outputs the fire confidence level; a pod-linked verification module, which, when the multimodal analysis result is low confidence but there is a risk, calls the aircraft's external infrared pod or downward-looking thermal imaging system to perform external verification sampling of the target area and compares whether the abnormal area has a continuous temperature rise, an expansion trend, or infrared high brightness characteristics to assist in forming the final judgment result; and an alarm output module, which converts the final judgment result into an alarm signal, sends it to the cockpit HUD or the main control system display interface through the flight control system, and automatically switches the prompt mode according to the fire level.

[0006] To achieve the above objectives, embodiments of this application also propose a method for detecting fire alarms in an aircraft cargo hold based on the linkage of AI anomaly recognition and thermal imaging. This method is implemented based on the aforementioned aircraft cargo hold fire alarm detection system based on the linkage of AI anomaly recognition and thermal imaging. The method includes: acquiring multimodal data within the cargo hold using a visible light camera, an infrared thermal imaging camera, and a smoke and gas sensor, and preprocessing these data to obtain a heat map sequence, a visible light image sequence, a gas concentration curve, and smoke particle data; inputting the heat map sequence into a thermal temporal prediction network based on spatiotemporal sequence modeling to extract spatial convolutional features and temporal recursive features, predicting the temperature rise trend in future multiple frames, and outputting an abnormal hot zone mask and anomaly probability values; calculating the Euclidean distance between the current heat map features and the centers of historical clustering models, and dynamically adjusting the abnormal temperature rise threshold based on the calculated Euclidean distance; inputting the visible light image sequence into a SmokeNet model based on a multi-scale convolutional neural network to identify and segment suspected smoke areas, outputting a binary mask and smoke area percentage; and inputting the gas concentration curve into a CUSUM mutation detection model to identify gas concentration mutation points and... The slope of the change is combined with smoke particle data for joint trend analysis to obtain smoke and gas analysis results. The contour and temperature fluctuation characteristics of the thermal anomaly area are input into a random forest classifier to determine whether the thermal anomaly area is a non-ignition source heat body. The temperature rise trend, the mask and probability value of the anomalous thermal area, the proportion of suspected smoke area and smoke area, the smoke and gas analysis results, and the non-ignition source heat body identification results are input into the XGBoost fusion model to calculate the confidence level of the fire and classify the confidence level. At the same time, an anomaly tracing report is generated, which includes the contribution weight of multimodal data to the final decision and the triggering reason. When the multimodal analysis result is low confidence but there is a risk, the external infrared pod or downward thermal imaging system of the aircraft is called to perform external verification sampling of the target area and compare whether the anomaly area has continuous temperature rise, expansion trend or infrared high brightness characteristics to assist in forming the final judgment result. The final judgment result is converted into an alarm signal and sent to the display interface of the cockpit HUD or main control system through the flight control system. The prompt mode is automatically switched according to the fire level, and visual evidence and risk level description are attached.

[0007] Optionally, the heat map detection and prediction unit incorporates a thermal time-series prediction network based on spatiotemporal sequence modeling. The network structure of the thermal time-series prediction network is ConvLSTM or spatiotemporal Transformer. The input to the thermal time-series prediction network is the heat map sequence output by the infrared thermal imaging acquisition module, which is a multi-channel temperature matrix of tensors formed by stacking time series data. The thermal time-series prediction network first extracts spatial features through a two-dimensional convolutional layer, that is, extracts the hot spot distribution features of a single frame, and uses normalization and position encoding to retain the spatial position information of the cargo hold structure. Then, it extracts temporal features, that is, captures the temperature change trend, and finally outputs temperature prediction maps and abnormal hot zone probability maps for several future frames. The predicted temperature rise curve is compared with the safety threshold to generate an abnormal heat diffusion trend score. The thermal time-series prediction network supports dual redundancy operation during the inference stage, that is, the same input data can be used for both inference and inference. The temperature rise is processed by a ConvLSTM model and a spatiotemporal Transformer model, respectively. The ConvLSTM model focuses on the time-dependent modeling of temperature changes, while the spatiotemporal Transformer model focuses on self-attention focusing in high-temperature regions. The outputs of the two models are weighted and fused according to their respective prediction confidence. When the difference between the two predictions exceeds a set deviation threshold, a consistency judgment process is triggered to correct the fusion weights, thereby reducing the risk of misjudgment by a single model and improving prediction stability. The temperature rise anomaly identification threshold in the aircraft cargo hold fire alarm detection system is adaptively adjusted according to the Euclidean distance distribution of historical heat map sequences. The system periodically calculates the mean and standard deviation of this distribution, forming a dynamic safety range with the mean ± kσ, and corrects it according to environmental parameters such as cabin humidity, air pressure, and background temperature, realizing self-learning and environmental adaptive control of the temperature rise threshold.

[0008] Optionally, the input to the thermal image adaptive learning and threshold control unit is the historical thermal image sequence acquired by the infrared thermal imaging acquisition module during the historical operation phase and the currently acquired real-time thermal image frame. The thermal image adaptive learning and threshold control unit performs principal component analysis on the historical thermal image sequence to reduce dimensionality, uses K-means clustering in the principal component space to obtain the historical cluster model centers, forms thermal baseline models under different cargo states and cabin types, and forms a thermal distribution pattern library. The currently acquired real-time thermal image frame is mapped to the principal component space to obtain the current thermal image features. The Euclidean distance between the current thermal image features and the historical cluster model centers is calculated. If the calculated Euclidean distance deviates from the baseline by more than a preset dynamic safety range, the temperature rise anomaly identification threshold is lowered. If the calculated Euclidean distance deviates from the baseline by no more than the preset dynamic safety range, the temperature rise anomaly identification threshold is raised to reduce false alarms. When dynamically adjusting the temperature rise anomaly identification threshold, the output of the environmental sensor is introduced as a correction factor. Multiple regression or neural network models are used to perform environmental compensation on the temperature rise anomaly identification threshold to avoid false alarms caused by seasonality or changes in flight altitude. The output of the environmental sensor includes cabin humidity, air pressure, and background temperature.

[0009] Optionally, the input to the visible light smoke recognition unit is a visible light image acquired by the visible light image acquisition module; the visible light smoke recognition unit performs multi-scale pyramid sampling on the input visible light image, extracts texture features at different spatial resolutions, and calculates local binary patterns and edge gradient features to capture the blurred edge characteristics of smoke; the texture features at different spatial resolutions and the blurred edge characteristics of smoke are input into the SmokeNet model based on a multi-scale convolutional neural network to obtain its output binary smoke mask and smoke density score, and calculates the smoke area ratio in the image.

[0010] Optionally, the input to the smoke and gas analysis unit is timestamped. Concentration curve, Concentration curves and particulate matter sensor data streams; the smoke and gas analysis unit will... Concentration curve, The concentration curve is converted into a gas concentration sequence and denoised by moving average filtering. The denoised gas concentration sequence is then subjected to CUSUM mutation detection and sliding window regression to locate the mutation time point and calculate the rate of change before and after the mutation to determine the gas concentration change characteristics. The smoke and gas analysis unit determines the particulate matter concentration change characteristics based on the sensor data stream. The gas concentration change characteristics and particulate matter concentration change characteristics are time-aligned. When both show an upward trend and are highly correlated, the fire probability score is improved.

[0011] Optionally, the input to the heat source liveness detection unit is the thermal imaging image of the thermal anomaly region and its corresponding visible light contour. The heat source liveness detection unit extracts thermal dynamic features based on the thermal imaging image of the thermal anomaly region, including the standard deviation of inter-frame temperature change and respiratory frequency fluctuations. It also extracts geometric features based on the corresponding visible light contour, including aspect ratio, contour smoothness, and area fluctuations. The thermal dynamic features and geometric features are then input into a random forest classifier to determine whether the thermal anomaly region is a non-fire source heat body. The output of the heat source liveness detection unit is the confidence level that the thermal anomaly region is a non-fire source heat body, and the corresponding category label for the non-fire source heat body.

[0012] Optionally, the input to the multimodal fusion decision unit is the output of the heat map detection and prediction unit, the heat map adaptive learning and threshold control unit, the visible light smoke recognition unit, the smoke and gas analysis unit, and the heat source liveness detection unit. The multimodal fusion decision unit uses XGBoost or LightGBM to weight and fuse the outputs of each unit, automatically assigning weights to the outputs of each unit, outputting the confidence level and trend prediction results of the fire situation, and generating an anomaly tracing report. The anomaly tracing report includes the contribution weight of multimodal data to the final decision and the triggering reasons, which are used by the unit for rapid accident judgment and post-accident review.

[0013] Optionally, when the multimodal analysis result is low confidence but poses a risk, the pod linkage verification module, based on the pixel coordinates of the abnormal area in the cargo hold's internal coordinate system, combined with the attitude angles provided by the flight control system and the calibration matrix of the cargo hold camera obtained through calibration, realizes the conversion between cargo hold coordinates and external space coordinates. This allows for the calculation of the pod gimbal's pitch and yaw angles, and the control of the external infrared lens for verification imaging. The persistence and spread of the anomaly are verified by pixel-level difference comparison with the original heatmap; that is, pixel-level difference is performed between the external and internal heatmaps to calculate the anomaly. Regional overlap rate and continuous temperature rise trend; attitude angles include pitch angle, roll angle and heading angle; when the pod linkage verification module fails to work, the aircraft cargo hold fire alarm detection system will automatically determine whether to enter temporary lock observation mode based on historical data and the latest heat map evolution trend, delaying the alarm trigger to avoid false alarms or missed alarms; in temporary lock observation mode, the system continuously tracks the temperature fluctuation and smoke concentration changes of the target area. If the anomaly continues to expand, the alarm will be triggered with a delay. If the anomaly subsides on its own, the lock will be released, thereby effectively avoiding false alarms or missed alarms.

[0014] Optionally, the multimodal AI recognition and processing module supports model updates based on online transfer learning mechanisms and introduces a model drift detection algorithm to monitor the differences between the distribution of input data and the distribution of historical training data in real time. When the difference value based on the Kullback-Leibler divergence metric exceeds the set drift threshold, the system automatically triggers the online incremental training process of the model, and updates the parameters by combining historical data and new data to maintain the recognition accuracy and stability of the model under different routes, climate conditions and cargo types.

[0015] The embodiments of this application propose an aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage, which has the following advantages compared with traditional aircraft cargo hold fire alarm detection systems.

[0016] First, multimodal intelligent recognition improves the accuracy of fire alarm detection. This application utilizes multi-source data such as infrared thermal images, visible light images, and gas concentrations, and through collaborative analysis using deep learning algorithms, effectively overcomes the limitations of single-modal recognition, significantly improving the accuracy and response speed for early fire detection.

[0017] Second, it features multi-dynamic threshold control to adapt to various cargo hold conditions. This application can automatically adjust the thermal anomaly identification threshold according to the aircraft cabin type and load status, avoiding misjudgment problems caused by differences in cargo thermal capacity or obstruction.

[0018] Third, it enables the identification of multiple living targets, reducing the false alarm rate. For non-fire-source heat objects such as cats and dogs, this application uses infrared temperature fluctuations and target outline morphology to effectively eliminate interference from intrusive biological factors.

[0019] Fourth, multi-pod linkage verification to improve the confidence level of judgment. When the system judgment is uncertain or the recognition result is marginally blurred, this application can schedule the infrared pod for real-time review and observation, enhancing the overall safety tolerance.

[0020] Fifth, multi-full-process automated alarm output to ensure flight safety. The alarm output module supports hierarchical response output for different fire alarm levels and can be synchronized to the HUD, ground station, and main control interface, providing clear and intuitive risk alerts for pilots and dispatchers.

[0021] In summary, this application provides an intelligent fire detection system for aircraft cargo holds that integrates AI recognition, thermal imaging analysis, and pod verification mechanisms. It features high precision, high self-adaptability, and good practicality, meeting the integration requirements of the safety protection system for new-generation intelligent aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] To more clearly illustrate the embodiments of this application or the technical solutions in the related art, the following will briefly introduce the drawings required for use in the description of the embodiments of this application or the related art. Obviously, the following drawings are only some embodiments of this application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings. The drawings described herein are only used to explain this application and are not used to limit this application.

[0023] Figure 1 is a schematic structural diagram of an aircraft cargo hold fire detection system based on AI anomaly recognition and thermal imaging linkage provided in an embodiment of this application; Figure 2 is a flowchart of a method for detecting aircraft cargo hold fires based on AI anomaly recognition and thermal imaging linkage provided in another embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the following will elaborate on each embodiment of this application in detail with reference to the drawings. Those of ordinary skill in the art can understand that in each embodiment of this application, many technical details are provided for the better understanding of the readers. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions claimed in this application can still be achieved. The following division of the embodiments is for convenience of description and should not constitute any limitation on the specific implementation of this application. The following embodiments can be combined and cross-referenced with each other without conflict.

[0025] One embodiment of this application proposes an aircraft cargo hold fire alarm detection system based on the linkage of AI anomaly recognition and thermal imaging, applicable to aircraft. The implementation details of the aircraft cargo hold fire alarm detection system based on the linkage of AI anomaly recognition and thermal imaging proposed in this embodiment are described in detail below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.

[0026] The specific structure of the aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage proposed in this embodiment can be as follows: Figure 1 As shown, it includes: The visible light image acquisition module 10 is used to acquire real-time visible light images of the interior of the aircraft cargo hold and provide image frames for smoke detection and visual feature extraction.

[0027] The infrared thermal imaging acquisition module 20 is used to acquire thermal imaging images of multiple areas and multiple channels inside the cargo hold, and to construct a thermal distribution map based on the thermal imaging images.

[0028] Smoke and gas sensing device 30, including sensor, Sensors and particulate matter sensors are used to monitor changes in the concentration of hazardous gases in the air and output continuous detection data.

[0029] The multimodal AI recognition and processing module 40 is used to fuse and analyze visible light images, thermal imaging images, thermal distribution maps, and detection data output by smoke and gas sensing devices.

[0030] The multimodal AI recognition and processing module 40 consists of a heat map detection and prediction unit 41, a heat map adaptive learning and threshold control unit 42, a visible light smoke recognition unit 43, a smoke and gas analysis unit 44, a heat source liveness detection unit 45, and a multimodal fusion decision unit 46.

[0031] The thermal image detection and prediction unit 41 is used to detect temperature rise anomalies and predict heat diffusion trends based on time-series thermal distribution maps. The thermal image adaptive learning and threshold adjustment unit 42 is used to learn the baseline of the thermal distribution map based on the cargo hold structure and cargo status and dynamically adjust the temperature rise anomaly identification threshold. The visible light smoke recognition unit 43 is used to identify suspected smoke areas based on the multi-scale texture and morphological features of visible light images. The smoke and gas analysis unit 44 is used to extract gas concentration mutation features based on the detection data output by the smoke and gas sensing device, and analyze the fire trend based on the gas concentration mutation features. The heat source liveness detection unit 45 is used to determine non-fire source heat bodies by combining thermal imaging images and target contours. The multimodal fusion decision unit 46 is used to fuse the identification results (output) of all the above units and output the confidence level of the fire.

[0032] The pod linkage verification module 50 is used to call the aircraft's external infrared pod or downward thermal imaging system to perform external verification sampling on the target area when the multimodal analysis result is low confidence but there is a risk. It also compares whether the abnormal area has a continuous temperature rise, an expansion trend or infrared high brightness characteristics to assist in forming the final judgment result.

[0033] The alarm output module 60 is used to convert the final judgment result into an alarm signal, which is then sent to the cockpit HUD or the main control system display interface via the flight control system, and the alert mode is automatically switched according to the fire level.

[0034] The functions and details of each component of the aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage proposed in this embodiment will be described in detail below.

[0035] (1) Sensing layer device.

[0036] The visible light image acquisition module is equipped with a visible light camera, and the infrared thermal imaging acquisition module is equipped with an infrared thermal imager. (Visible light camera, infrared thermal imager) sensor, Sensors and particulate matter sensors together form a multimodal sensor network for the aircraft cargo hold fire alarm detection system to acquire multimodal data.

[0037] Visible light camera, infrared thermal imager, sensor, Both the sensors and the particulate matter sensors are deployed inside the aircraft's cargo hold.

[0038] The visible light images acquired by the visible light image acquisition module may be described differently below due to differences in their uses and storage methods, but they are essentially all visible light images. Similarly, the thermal images acquired by the infrared thermal imaging acquisition module may be described differently below due to differences in their uses and storage methods, but they are essentially all thermal images. The data output by the sensor is Concentration curve, The data output by the sensor is The concentration curves may be described differently in the following text, but they are essentially all about gas concentration.

[0039] (2) Heat map detection and prediction unit.

[0040] The heat map detection and prediction unit has a built-in thermal time series prediction network based on spatiotemporal sequence modeling. The network structure of the thermal time series prediction network is ConvLSTM or spatiotemporal Transformer.

[0041] The input to the thermal time series prediction network is the heat map sequence output by the infrared thermal imaging acquisition module, which is a multi-channel temperature matrix of tensor formed by stacking time series. The matrix has the shape T×H×W×C, where T represents time, H represents height, W represents width, and C represents the number of channels.

[0042] The thermal time series prediction network first extracts spatial features through a two-dimensional convolutional layer (Conv2D), that is, it extracts the hot spot distribution features of a single frame (such as the location of the maximum temperature and the direction of the temperature gradient). It then uses normalization and location encoding to retain the spatial location information of the cargo hold structure. Next, it extracts temporal features, that is, it captures the temperature change trend. Finally, it outputs a temperature prediction map (temperature rise prediction value matrix) and an abnormal hot zone probability map (probability range from 0 to 1) for several future frames. The predicted temperature rise curve is compared with the safety threshold to generate an abnormal heat diffusion trend score.

[0043] The thermal time series prediction network supports dual redundancy during the inference phase. That is, the same input data is processed by a ConvLSTM model and a spatiotemporal Transformer model respectively. The ConvLSTM model focuses on the time dependence modeling of temperature changes, while the spatiotemporal Transformer model focuses on self-attention focusing in high-temperature regions. The output results of the two models are weighted and fused according to their respective prediction confidence. When the difference between the two prediction results exceeds a set deviation threshold, a consistency judgment process is triggered to correct the fusion weight, thereby reducing the risk of misjudgment by a single model and improving prediction stability. The temperature rise anomaly identification threshold in the aircraft cargo hold fire alarm detection system is adaptively adjusted based on the Euclidean distance distribution of historical heat map sequences. The system periodically calculates the mean and standard deviation of this distribution, and forms a dynamic safety range with the mean ± kσ. It is then corrected based on environmental parameters such as cabin humidity, air pressure, and background temperature to achieve self-learning and environmental adaptive control of the temperature rise threshold.

[0044] (3) Heatmap adaptive learning and threshold control unit.

[0045] The input to the heatmap adaptive learning and threshold control unit is the historical heatmap sequence acquired by the infrared thermal imaging acquisition module during the historical operation phase and the currently acquired real-time heatmap frame.

[0046] The heatmap adaptive learning and threshold control unit performs principal component analysis (PCA) on historical heatmap sequences to reduce dimensionality (to 2 to 3 dimensions). K-means clustering is used in the principal component space to obtain historical cluster model centers, forming thermal baseline models under different cargo states and cabin types, and forming a thermal distribution pattern library, with each class representing a thermal baseline for a combination of cargo states / cabin types.

[0047] The currently acquired heatmap frame is mapped to the principal component space to obtain the current heatmap features. The Euclidean distance between the current heatmap features and the historical cluster model centers is calculated. If the calculated Euclidean distance deviates from the baseline by more than a preset dynamic safety interval, the temperature rise anomaly detection threshold is lowered; if the calculated Euclidean distance deviates from the baseline by no more than the preset dynamic safety interval, the temperature rise anomaly detection threshold is raised to reduce false alarms. The real-time temperature rise anomaly detection threshold is then output for use by other units or modules.

[0048] When dynamically adjusting the temperature rise anomaly identification threshold, the output of environmental sensors (including but not limited to cabin humidity, air pressure, and background temperature) is introduced as a correction factor. Multiple regression or neural network models are used to compensate for the temperature rise anomaly identification threshold, avoiding false alarms caused by seasonality or changes in flight altitude.

[0049] (4) Visible light smoke recognition unit.

[0050] The input to the visible light smoke recognition unit is a visible light image acquired by the visible light image acquisition module. The color space of the visible light image is RGB, and the size is 640px × 480px.

[0051] The visible light smoke recognition unit performs multi-scale pyramid sampling on the input visible light image, extracts texture features at different spatial resolutions, and calculates local binary pattern (LBP) and edge gradient features to capture the blurred edge characteristics of smoke.

[0052] The visible light smoke recognition unit inputs the texture features and blurred edge characteristics of smoke at different spatial resolutions into the SmokeNet model based on a multi-scale convolutional neural network to obtain its output binary smoke mask (1 represents smoke and 0 represents background) and smoke density score (the score ranges from 0 to 1), and calculates the proportion of smoke area in the image.

[0053] The SmokeNet model mainly consists of convolution, BatchNorm, ReLU, and upsampling decoding blocks.

[0054] (5) Smoke and gas analysis unit.

[0055] The input to the smoke and gas analysis unit is timestamped. Concentration curve, Concentration curves and particulate matter sensor data streams.

[0056] The smoke and gas analysis unit will Concentration curve, The concentration curve is converted into a gas concentration sequence and then subjected to moving average filtering for noise reduction. The denoised gas concentration sequence is then input into the CUSUM mutation detection model for CUSUM mutation detection and moving window regression to locate the mutation time point and calculate the rate of change (ppm / s) before and after the mutation to determine the gas concentration change characteristics.

[0057] Next, the smoke and gas analysis unit determines the particulate matter concentration change characteristics based on the sensor data stream, and aligns the gas concentration change characteristics with the particulate matter concentration change characteristics over time. When both show an upward trend and are highly correlated (correlation coefficient greater than 0.8), the fire probability score is improved.

[0058] (6) Heat source live body discrimination unit.

[0059] The input to the heat source liveness detection unit is the thermal imaging image of the thermal anomaly region and the corresponding visible light profile.

[0060] The heat source liveness detection unit extracts thermal dynamic features based on thermal imaging images of thermal anomaly areas, including the standard deviation of inter-frame temperature change and respiratory frequency fluctuations. It also extracts geometric features based on the corresponding visible light contours, including aspect ratio, contour smoothness, and area fluctuations. The thermal dynamic features and geometric features are then input into a random forest classifier to determine whether the thermal anomaly area is a non-fire source heat body.

[0061] The output of the heat source liveness detection unit is the confidence level (0 to 1) that the thermal anomaly area is a non-fire source heat body, and the corresponding category label of the non-fire source heat body (such as cat, dog, or stray person).

[0062] The training samples used in the training process of the random forest classifier include non-fire source heat body samples such as cats, dogs, and people, as well as real fire source samples.

[0063] (7) Multimodal fusion decision unit.

[0064] The inputs to the multimodal fusion decision unit are the outputs of the heat map detection and prediction unit, the heat map adaptive learning and threshold control unit, the visible light smoke recognition unit, the smoke and gas analysis unit, and the heat source liveness detection unit. These outputs include the heat map temperature rise gradient, the smoke mask area ratio, the gas change rate, the liveness confidence, and the flight state parameters (altitude and speed).

[0065] The multimodal fusion decision unit uses XGBoost or LightGBM to weight and fuse the outputs of each unit, automatically assigning weights to the outputs of each unit, and outputting the confidence level (low, medium, high) and trend prediction results (low, medium, high) of the fire situation. At the same time, it generates an anomaly tracing report, which includes the contribution weight of multimodal data to the final decision and the triggering reasons, for the unit to quickly interpret the accident and conduct post-accident review.

[0066] (8) Pod linkage verification module.

[0067] The trigger condition of the pod linkage verification module is that the multi-modal analysis result is in the low confidence interval but the risk score is greater than the preset medium threshold. That is, when the multi-modal analysis result is of low confidence but there is a risk, according to the pixel coordinates of the abnormal area in the cargo hold internal coordinate system, combined with the attitude angles (pitch angle, roll angle, heading angle) provided by the flight control system and the calibration matrix of the cargo hold camera obtained through calibration, the conversion between the cargo hold coordinates and the external space coordinates is realized, so as to calculate the pitch angle and yaw angle of the pod pan-tilt, and control the external infrared lens to perform a review shot. The persistence and diffusion of the abnormality are verified by pixel-level differential comparison with the original heat map, that is, the external heat map and the internal heat map are pixel-level differenced, and the overlap rate of the abnormal area and the continuous temperature rise trend are calculated.

[0068] When the pod linkage verification module fails to work, the aircraft cargo hold fire detection system will automatically judge whether to enter the temporary locking observation mode based on historical data and the latest heat map evolution trend, and delay triggering the alarm to avoid false alarms or missed alarms. That is, if the pod linkage verification cannot complete the verification operation due to equipment failure, signal occlusion, etc., the system will enable the abnormal evolution analysis mechanism of the historical heat map trajectory and real-time trend, and enter the "temporary locking observation mode" within a certain period of time. In the temporary locking observation mode, the system continuously tracks the temperature fluctuation and smoke concentration change of the target area. If the abnormality continues to expand, the alarm will be triggered with a delay. If the abnormality self-relieves, the lock will be released, thus effectively avoiding false alarms or missed alarms.

[0069] (9) Alarm output module.

[0070] The alarm output module supports hierarchical output according to the fire alarm confidence level, which is divided into three categories: early warning, primary alarm and emergency alarm. Different levels correspond to different visual cues, audio alerts and linkage responses. When the pod linkage verification module fails to return the review result within the specified time, the system calculates the risk score based on the latest heat map trend and the multi-modal fusion confidence level. If the score exceeds the safety threshold, the corresponding level of alarm will be directly triggered to ensure aviation safety; if the score is in the critical range, it will enter the short-term delay observation mode and continue sampling verification to avoid false alarms caused by temporary interference. The alarm output module supports converting the final judgment result and key evidence (heat map, smoke mask, gas curve) into alarm signals, and sending them to the cockpit HUD, ground station or main control display interface through the flight control system, and switching different sound / visual cue modes according to the fire situation level.

[0071] (10) Multi-modal AI recognition and processing module.

[0072] The multi-modal AI recognition and processing module supports model updates based on an online transfer learning mechanism and introduces a model drift detection algorithm (MDD) to continuously monitor the differences between the input data distribution and the historical training data distribution in real time. When the difference value based on Kullback-Leibler divergence or other distribution distance metrics exceeds the set drift threshold, the system automatically triggers the online incremental training process of the model, combines historical data with new data for parameter updates, so as to maintain the recognition accuracy and stability of the model under different flight routes, climate conditions, and cargo types.

[0073] In addition, the aircraft cargo hold fire detection system supports the dynamic switching of different parameters in the early training stage and the late flight stage, adapting to environmental factors such as temperature background, airflow disturbance, and external interference in different scenarios such as night / day, takeoff / cruise, etc., to improve the detection robustness and practicality. The multi-modal fusion decision-making unit, based on an adjustable weight mechanism, supports users or algorithms to automatically adjust the weight coefficients of each channel, enabling the aircraft cargo hold fire detection system to have an adaptive ability under different flight tasks (such as freight routes / special cargo / special climate). Through the combination of multi-modal perception and AI recognition decision-making algorithms, the aircraft cargo hold fire detection system significantly improves the intelligent level of aircraft cargo hold fire detection, taking into account false alarm control and information credibility while ensuring the response speed, providing key technical support for new aviation scenarios such as future unmanned cargo aircraft and urban air traffic.

[0074] An aircraft cargo hold fire detection system based on AI anomaly recognition and thermal imaging linkage proposed in this embodiment has the following advantages compared with traditional aircraft cargo hold fire detection systems.

[0075] First, multi-modal intelligent recognition improves the accuracy of fire detection. This embodiment uses multi-source data such as infrared thermal images, visible light images, and gas concentrations, and through collaborative analysis by deep learning algorithms, effectively overcomes the limitations of single-modal recognition, and significantly improves the recognition accuracy and response speed of early fire situations.

[0076] Second, multi-dynamic threshold regulation adapts to various cargo hold states. This embodiment can automatically adjust the thermal anomaly recognition threshold according to the aircraft cabin type and load state, avoiding misjudgment problems caused by differences in cargo heat capacity or occlusion.

[0077] Third, multi-living target discrimination reduces the false alarm rate. For non-fire heat sources such as cats and dogs, this embodiment judges through infrared temperature fluctuations and target contour shapes, effectively excluding the interference factors of accidentally entering organisms. <关于猫狗等非火源热体,本实施例通过红外温度波动与目标轮廓形态判断,有效排除误入生物干扰因素。

[0078] Fourth, multi-pod linkage verification improves the judgment confidence level. When the system's judgment is uncertain or the recognition result is marginally ambiguous, this embodiment can dispatch the infrared pod for real-time recheck observation, enhancing the overall safety fault tolerance ability.

[0079] Fifth, multi-stage automated alarm output ensures flight safety. The alarm output module supports tiered response output for different fire alarm levels and can be synchronized to the HUD, ground station, and main control interface, providing pilots and dispatchers with clear and intuitive risk alerts.

[0080] In summary, this embodiment provides an intelligent fire alarm detection system for aviation cargo holds that integrates AI recognition, thermal imaging analysis, and pod verification mechanisms. It features high precision, high adaptability, and good practicality, and is suitable for the integration requirements of safety protection systems for next-generation intelligent aircraft.

[0081] It is worth mentioning that all modules and units involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units do not exist in this embodiment.

[0082] One embodiment of this application proposes a method for detecting fire alarms in an aircraft cargo hold based on the linkage of AI anomaly recognition and thermal imaging, implemented based on an aircraft cargo hold fire alarm detection system based on the linkage of AI anomaly recognition and thermal imaging as described in the above system embodiment. The implementation details of the aircraft cargo hold fire alarm detection method based on the linkage of AI anomaly recognition and thermal imaging proposed in this embodiment are described below. The following implementation details are provided for ease of understanding and are not essential for implementing this solution.

[0083] The specific process of the aircraft cargo hold fire detection method based on the linkage of AI anomaly recognition and thermal imaging proposed in this embodiment can be described as follows: Figure 1 As shown, it includes: Step 71: Acquire multimodal data of the cargo hold using a visible light camera, an infrared thermal imaging camera, and a smoke gas sensor, and preprocess them to obtain a thermal image sequence, a visible light image sequence, a gas concentration curve, and smoke particle data.

[0084] Step 72: Input the heat map sequence into the thermal time series prediction network based on spatiotemporal sequence modeling, extract spatial convolutional features and temporal recursive features, predict the temperature rise trend of future multiple frames, and output the abnormal hot area mask and abnormal probability value.

[0085] Step 73: Calculate the Euclidean distance between the current heat map features and the historical cluster model centers, and dynamically adjust the abnormal temperature rise threshold based on the calculated Euclidean distance.

[0086] Step 74: Input the visible light image sequence into the SmokeNet model based on a multi-scale convolutional neural network to identify and segment suspected smoke regions, and output a binary mask and the proportion of smoke area.

[0087] Step 75: Input the gas concentration curve into the CUSUM mutation detection model to identify the gas concentration mutation points and slopes, and perform joint trend analysis in conjunction with smoke particle data to obtain the smoke and gas analysis results.

[0088] Step 76: Input the outline and temperature fluctuation features of the thermal anomaly region into the random forest classifier to determine whether the thermal anomaly region is a non-fire source heat body.

[0089] Step 77: Input the temperature rise trend, abnormal hot zone mask and abnormal probability value, suspected smoke area and smoke area ratio, smoke and gas analysis results, and non-fire source heat body identification results into the XGBoost fusion model to calculate the confidence level of the fire and classify the confidence level. At the same time, generate an anomaly tracing report. The anomaly tracing report includes the contribution weight of multimodal data to the final decision and the triggering reason.

[0090] Step 78: If the multimodal analysis result is low confidence but there is a risk, call the aircraft's external infrared pod or downward thermal imaging system to perform external verification sampling on the target area, and compare whether the abnormal area has a continuous temperature rise, expansion trend or infrared high brightness characteristics to help form the final judgment result.

[0091] Step 79: The final judgment result is converted into an alarm signal and sent to the cockpit HUD or the main control system display interface through the flight control system. The prompt mode is automatically switched according to the fire level, and visual evidence and risk level description are attached.

[0092] The steps described above are for clarity only. In implementation, they can be combined into one step, or some steps can be broken down into multiple steps, as long as they involve the same logical relationship, they are all within the scope of protection of this application. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, without changing the core design of the algorithm and process, are also within the scope of protection of this application.

[0093] It is not difficult to see that this embodiment is a method embodiment corresponding to the above system embodiment, and this embodiment can be implemented in conjunction with the above system embodiment. The relevant technical details and technical effects mentioned in the above system embodiment are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above system embodiment.

[0094] Those skilled in the art will understand that the above embodiments are specific implementations of this application, and in practical applications, various changes can be made in form and detail without departing from the spirit and scope of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A fire alarm detection system for aircraft cargo hold based on AI anomaly recognition and thermal imaging linkage, applicable to aircraft, characterized in that, The system includes: The visible light image acquisition module is used to acquire real-time visible light images of the interior of the aircraft cargo hold and provide image frames for smoke detection and visual feature extraction. The infrared thermal imaging acquisition module is used to acquire thermal imaging images of multiple areas and multiple channels inside the cargo hold, and to construct a thermal distribution map based on the thermal imaging images. Smoke and gas sensing devices, including sensor, Sensors and particulate matter sensors are used to monitor changes in the concentration of hazardous gases in the air and output continuous detection data; The multimodal AI recognition and processing module is used to fuse and analyze visible light images, thermal imaging images, heat distribution maps, and detection data output by smoke and gas sensing devices. The multimodal AI recognition and processing module consists of a thermal image detection and prediction unit, a thermal image adaptive learning and threshold adjustment unit, a visible light smoke recognition unit, a smoke and gas analysis unit, a heat source liveness detection unit, and a multimodal fusion decision unit. The thermal image detection and prediction unit detects temperature rise anomalies and predicts heat diffusion trends based on time-series thermal distribution maps. The thermal image adaptive learning and threshold adjustment unit learns the baseline of the thermal distribution map based on the cargo hold structure and cargo status and dynamically adjusts the temperature rise anomaly recognition threshold. The visible light smoke recognition unit identifies suspected smoke areas based on multi-scale texture and morphological features of visible light images. The smoke and gas analysis unit extracts gas concentration mutation features from the detection data output by smoke and gas sensors and analyzes fire trends based on these features. The heat source liveness detection unit combines thermal imaging images and target contours to identify non-fire source heat bodies. The multimodal fusion decision unit fuses the recognition results from the above units and outputs the fire confidence level. The pod linkage verification module is used to call the aircraft's external infrared pod or downward thermal imaging system to perform external verification sampling on the target area when the multimodal analysis result is low confidence but there is a risk. It also compares whether the abnormal area has a continuous temperature rise, an expansion trend or infrared high brightness characteristics to help form the final judgment result. The alarm output module is used to convert the final judgment result into an alarm signal, which is then sent to the cockpit HUD or the main control system display interface via the flight control system, and the alert mode is automatically switched according to the fire level.

2. The aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage as described in claim 1, characterized in that, The heat map detection and prediction unit has a built-in thermal time series prediction network based on spatiotemporal sequence modeling. The network structure of the thermal time series prediction network is ConvLSTM or spatiotemporal Transformer. The input to the thermal time series prediction network is the heat map sequence output by the infrared thermal imaging acquisition module, which is a multi-channel temperature matrix of tensor formed by stacking time series. The thermal time series prediction network first extracts spatial features through a two-dimensional convolutional layer, that is, extracts the hot spot distribution features of a single frame. It then uses normalization and position encoding to retain the spatial position information of the cargo hold structure. Finally, it extracts temporal features, that is, captures the temperature change trend. Finally, it outputs the temperature prediction map and abnormal hot zone probability map for several future frames, and compares the predicted temperature rise curve with the safety threshold to generate an abnormal heat diffusion trend score. The thermal time series prediction network supports dual redundancy during the inference phase. That is, the same input data is processed by a ConvLSTM model and a spatiotemporal Transformer model respectively. The ConvLSTM model focuses on the time dependence modeling of temperature changes, while the spatiotemporal Transformer model focuses on self-attention focusing in high-temperature regions. The output results of the two models are weighted and fused according to their respective prediction confidence. When the difference between the two prediction results exceeds a set deviation threshold, a consistency judgment process is triggered to correct the fusion weight, thereby reducing the risk of misjudgment by a single model and improving prediction stability. The temperature rise anomaly identification threshold in the aircraft cargo hold fire alarm detection system is adaptively adjusted based on the Euclidean distance distribution of historical heat map sequences. The system periodically calculates the mean and standard deviation of this distribution, and forms a dynamic safety range with the mean ± kσ. It is then corrected based on environmental parameters such as cabin humidity, air pressure, and background temperature to achieve self-learning and environmental adaptive control of the temperature rise threshold.

3. The aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage as described in claim 1, characterized in that, The input to the heatmap adaptive learning and threshold control unit is the historical heatmap sequence acquired by the infrared thermal imaging acquisition module during the historical operation phase and the currently acquired real-time heatmap frame; The heatmap adaptive learning and threshold control unit performs principal component analysis on historical heatmap sequences to reduce dimensionality, uses K-means clustering in the principal component space to obtain historical cluster model centers, forms thermal baseline models under different cargo states and cabin types, and constitutes a thermal distribution pattern library. The current real-time acquired heat map frame is mapped to the principal component space to obtain the current heat map features. The Euclidean distance between the current heat map features and the historical cluster model center is calculated. If the calculated Euclidean distance deviates from the baseline by more than the preset dynamic safety range, the temperature rise anomaly identification threshold is reduced. If the calculated Euclidean distance deviates from the baseline by no more than the preset dynamic safety range, the temperature rise anomaly identification threshold is increased to reduce false alarms. When dynamically adjusting the temperature rise anomaly identification threshold, the output of environmental sensors is introduced as a correction factor. Multiple regression or neural network models are used to compensate for the temperature rise anomaly identification threshold to avoid false alarms caused by seasonality or changes in flight altitude. The output of the environmental sensors includes cabin humidity, air pressure and background temperature.

4. The aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage as described in claim 1, characterized in that, The input to the visible light smoke recognition unit is the visible light image acquired by the visible light image acquisition module; The visible light smoke recognition unit performs multi-scale pyramid sampling on the input visible light image, extracts texture features at different spatial resolutions, and calculates local binary patterns and edge gradient features to capture the blurred edge characteristics of smoke. Texture features at different spatial resolutions and the blurred edge characteristics of smoke are input into the SmokeNet model based on a multi-scale convolutional neural network to obtain its output binary smoke mask and smoke density score, and to calculate the proportion of smoke area in the image.

5. The aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage as described in claim 1, characterized in that smoke... The input to the gas analysis unit is timestamped. Concentration curve, Concentration curves and particulate matter sensor data streams; The smoke and gas analysis unit will Concentration curve, The concentration curve is converted into a gas concentration sequence and then subjected to moving average filtering for noise reduction. The denoised gas concentration sequence is then subjected to CUSUM mutation detection and moving window regression to locate the mutation time point and calculate the rate of change before and after the mutation to determine the gas concentration change characteristics. The smoke and gas analysis unit determines the characteristics of particulate matter concentration changes based on sensor data streams, aligns the characteristics of gas concentration changes with the characteristics of particulate matter concentration changes over time, and improves the fire probability score when both are showing an upward trend and are highly correlated.

6. The aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage as described in claim 1, characterized in that, The input to the heat source liveness detection unit is the thermal imaging image of the thermal anomaly region and the corresponding visible light contour; The heat source liveness detection unit extracts thermal dynamic features based on thermal imaging images of thermal anomaly areas, including the standard deviation of inter-frame temperature change and respiratory frequency fluctuations. It also extracts geometric features based on the corresponding visible light contours, including aspect ratio, contour smoothness, and area fluctuations. The thermal dynamic features and geometric features are then input into a random forest classifier to determine whether the thermal anomaly area is a non-fire source heat body. The output of the heat source liveness detection unit is the confidence level that the thermal anomaly region is a non-fire source heat body, and the corresponding category label of the non-fire source heat body.

7. The aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage as described in claim 1, characterized in that, The inputs of the multimodal fusion decision unit are the outputs of the heat map detection and prediction unit, the heat map adaptive learning and threshold control unit, the visible light smoke recognition unit, the smoke and gas analysis unit, and the heat source liveness detection unit. The multimodal fusion decision unit uses XGBoost or LightGBM to weight and fuse the outputs of each unit, automatically assigning weights to the outputs of each unit, outputting the confidence level and trend prediction results of the fire situation, and generating an anomaly tracing report. The anomaly tracing report includes the contribution weight of multimodal data to the final decision and the triggering reasons, which is used by the unit for rapid accident judgment and post-accident review.

8. The aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage as described in claim 1, characterized in that, When the multimodal analysis result is low confidence but poses a risk, the pod linkage verification module, based on the pixel coordinates of the abnormal area in the cargo hold's internal coordinate system, combined with the attitude angles provided by the flight control system and the calibration matrix of the cargo hold camera obtained through calibration, realizes the conversion between cargo hold coordinates and external space coordinates. This allows the calculation of the pod gimbal's pitch and yaw angles, and controls the external infrared lens to perform verification shooting. The persistence and spread of the anomaly are verified by pixel-level difference comparison with the original heat map. That is, pixel-level difference is performed between the external heat map and the internal heat map to calculate the overlap rate of the abnormal area and the continuous temperature rise trend. The attitude angles include pitch, roll, and yaw angles. When the pod linkage verification module fails to work, the aircraft cargo hold fire alarm detection system will automatically determine whether to enter the temporary lock observation mode based on historical data and the latest heat map evolution trend, and delay the triggering of the alarm to avoid false alarms or missed alarms. In the temporary lockout observation mode, the system continuously tracks temperature fluctuations and smoke concentration changes in the target area. If the anomaly continues to escalate, the alarm will be triggered with a delay. If the anomaly resolves itself, the lockout will be lifted, thus effectively avoiding false alarms or missed alarms.

9. The aircraft cargo hold fire alarm detection system based on AI anomaly recognition and thermal imaging linkage as described in claim 1, characterized in that, The multimodal AI recognition and processing module supports model updates based on online transfer learning mechanisms and introduces a model drift detection algorithm to monitor the differences between the distribution of input data and the distribution of historical training data in real time. When the difference value based on the Kullback-Leibler divergence metric exceeds the set drift threshold, the system automatically triggers the online incremental training process of the model, and updates the parameters by combining historical data and new data to maintain the recognition accuracy and stability of the model under different routes, climate conditions and cargo types.

10. A method for detecting fires in an aircraft cargo hold based on the linkage of AI anomaly recognition and thermal imaging, implemented based on the aircraft cargo hold fire detection system based on the linkage of AI anomaly recognition and thermal imaging as described in any one of claims 1 to 9, the method comprising: Multimodal data inside the cargo hold were acquired using a visible light camera, an infrared thermal imaging camera, and a smoke gas sensor, and preprocessed to obtain thermal image sequences, visible light image sequences, gas concentration curves, and smoke particle data. The heat map sequence is input into a thermal time series prediction network based on spatiotemporal sequence modeling. Spatial convolutional features and temporal recursive features are extracted to predict the temperature rise trend in future multiple frames and output abnormal hot area masks and abnormal probability values. The Euclidean distance between the current heat map features and the historical cluster model centers is calculated, and the abnormal temperature rise threshold is dynamically adjusted based on the calculated Euclidean distance. The visible light image sequence is input into the SmokeNet model based on a multi-scale convolutional neural network to identify and segment suspected smoke regions, and output a binary mask and the proportion of smoke area. The gas concentration curve is input into the CUSUM mutation detection model to identify the gas concentration mutation points and slopes, and combined with smoke particle data for joint trend analysis to obtain the smoke and gas analysis results. The outline and temperature fluctuation features of the thermal anomaly region are input into a random forest classifier to determine whether the thermal anomaly region is a non-fire source heat body. The temperature rise trend, abnormal hot zone mask and abnormal probability value, suspected smoke area and smoke area ratio, smoke and gas analysis results, and non-fire source heat body identification results are input into the XGBoost fusion model to calculate the confidence level of the fire and classify the confidence level. At the same time, an anomaly tracing report is generated, which includes the contribution weight of multimodal data to the final decision and the triggering reason. When the multimodal analysis results are of low confidence but there is a risk, the external infrared pod or downward thermal imaging system of the aircraft is called to perform external verification sampling on the target area, and the abnormal area is compared to see if it has a continuous temperature rise, an expansion trend or infrared brightness characteristics, to help form the final judgment result. The final judgment result is converted into an alarm signal and sent to the cockpit HUD or main control system display interface through the flight control system. The alarm signal automatically switches the alert mode according to the fire level and is accompanied by visual evidence and risk level description.

Images