Fire behavior intelligent detection and quick response method based on image recognition

Abstract

The invention discloses an intelligent fire detection and quick response method based on image recognition. The method comprises the following steps: S1, synchronously acquiring and preprocessing multi-source image data; s2, carrying out multi-modal feature fusion and early fire identification; s3, intelligent grading response based on deep reinforcement learning: constructing a fire behavior grade dynamic evaluation network, inputting fused multi-modal features, outputting fire behavior grade probability distribution and a crisis index, judging a fire behavior grade according to a dynamic threshold value, and triggering a predefined grading response strategy; s4, performing cross-scene adaptive transfer learning; and S5, carrying out system self-diagnosis and dynamic optimization. According to the invention, through time-space synchronous acquisition and deep fusion of visible light, infrared and smoke multi-mode data, vision, thermal radiation and smoke characteristics of flames are integrated, and common interferences such as real fire behavior and lamplight, light reflection, a moving heat source, water mist and dust are effectively distinguished by using an improved lightweight CNN, a self-adaptive background temperature model and an environment airflow model.

Description

Technical Field

[0001] This application belongs to the field of fire detection technology, specifically involving a method for intelligent fire detection and rapid response based on image recognition. Background Technology

[0002] Fire is one of the major disasters that seriously threaten life and property. In recent years, with the reduction in cost and widespread use of hardware equipment such as surveillance cameras and infrared sensors, vision-based automatic fire detection technology has become a research hotspot in the field of fire safety. However, existing technical solutions still face many challenges and limitations in large-scale deployment and application.

[0003] At the perception level, most systems rely on single-type sensor data (such as using only visible light cameras), which lacks robustness to scenarios involving changes in ambient light, smoke obstruction, and complex background interference (such as lights, solar reflection, and moving heat sources). Traditional color segmentation and temperature judgment methods based on fixed thresholds, as well as early machine learning models based on handcrafted features (such as LBP and HOG), struggle to accurately distinguish between real fires and similar interference sources, resulting in high false alarm and false negative rates, especially in the early stages of fire ignition or smoldering, where detection is particularly difficult.

[0004] At the information fusion and decision-making level, existing technologies often employ simple rule-based fusion or post-fusion strategies, failing to fully utilize the complementarity of multimodal data such as visible light, infrared, and smoke in spatial, temporal, and spectral dimensions. There is a lack of a dynamic weight allocation mechanism that can adapt to environmental changes (such as day / night, season, and weather), and a lack of intelligent models for quantitatively assessing and predicting fire development trends, resulting in delayed early warnings or rigid response strategies.

[0005] At the system deployment and operation level, well-trained models are usually optimized for specific scenarios. When deployed to a new environment (such as from a chemical plant to a library), their performance will drop significantly, and the cost of re-labeling and retraining is high. At the same time, the system lacks the ability to actively monitor and self-calibrate the status of front-end sensors (such as lens contamination, defocus, and offset) during long-term operation, as well as the adaptive optimization mechanism for model performance degradation as the environment changes slowly, which increases the operation and maintenance burden and the risk of failure.

[0006] Therefore, there is an urgent need for an integrated fire detection and response solution that can achieve high-precision early identification, low false alarms, strong generalization, adaptive optimization, and drive intelligent hierarchical rapid response. Summary of the Invention

[0007] This application provides a method for intelligent fire detection and rapid response based on image recognition, which aims to solve the problems of existing technologies that make it difficult to accurately distinguish between real fires and similar interference sources, resulting in delayed early warning or rigid response strategies.

[0008] A method for intelligent fire detection and rapid response based on image recognition, the method comprising:

[0009] S1. Synchronous acquisition and preprocessing of multi-source image data: The video stream of the monitored area is synchronously acquired through the time-stamped visible light camera, infrared thermal imaging camera and smoke-sensitive camera to form spatiotemporally aligned multimodal raw data, and the images of each modality are enhanced and corrected respectively.

[0010] S2. Multimodal Feature Fusion and Early Fire Identification: Based on the multimodal image data preprocessed in S1, an improved lightweight convolutional neural network is used to extract the flame color, shape, and dynamic flicker frequency features in the visible light image. An adaptive background temperature model is used to identify abnormal temperature rise areas in the infrared image. Smoke features are extracted based on the fusion of ultraviolet and visible light images. The features of each modality are fused through a feature fusion module with dynamically adjustable weights to achieve early fire identification.

[0011] S3. Intelligent hierarchical response based on deep reinforcement learning: Based on the multimodal features fused in S2, a dynamic fire level assessment network is constructed. The multimodal features are input, the probability distribution of fire level and crisis index are output, and the fire level is determined according to the dynamic threshold to trigger the predefined hierarchical response strategy.

[0012] S4, Cross-Scene Adaptive Transfer Learning: Based on cross-domain multi-scenario fire data, a domain-invariant feature representation space is constructed through self-supervised contrastive learning and adversarial training, and a generative adversarial network and incremental learning framework are combined to achieve rapid adaptation to new scenarios in S1.

[0013] S5. System self-diagnosis and dynamic optimization: Real-time monitoring of image quality, occlusion, and image stability of each camera in S1. Based on the diagnostic results, software correction, active self-calibration, or maintenance alarms are triggered in stages. The model in S2 to S4 is continuously optimized through periodic online fine-tuning and federated learning.

[0014] Optionally, in S1, the enhancement and correction processing of each modal image includes: adaptive brightness equalization and flame feature color gamut enhancement processing of the visible light image;

[0015] Non-uniformity correction and temperature calibration are performed on the infrared image to generate an absolute temperature distribution matrix;

[0016] Fog-sensitive images are enhanced with fog penetration and dynamic range compression.

[0017] Optionally, in S2, the improved lightweight convolutional neural network includes a feature extraction backbone constructed from depthwise separable convolutions, an embedded temporal attention module, and a cascaded frame difference analysis branch and a group of one-dimensional convolutional layers for extracting the dynamic flicker frequency of the flame.

[0018] Optionally, in step S2, the adaptive background temperature model uses a Gaussian mixture model or kernel density estimation method to establish a temperature background probability distribution for each pixel, and updates the background model through a dynamic adaptive learning rate strategy, combined with spatial alignment verification to improve the reliability of fire identification.

[0019] Optionally, in step S2, the feature fusion module with dynamically adjustable weights performs weighted fusion of the confidence levels of visible light, infrared and smoke features based on the initial basic weights of each modality output by the environmental context perceiver, and uses an improved early warning rule based on evidence theory to make early fire judgment.

[0020] Optionally, in step S3, the fire level dynamic assessment network includes a long short-term memory network to capture the temporal evolution of the fire state and output the fire level probability distribution, crisis index, and state prediction value for the next time step; the crisis index is calculated by a nonlinear combination of normalized flame area, temperature rise rate, smoke concentration, diffusion rate, and comprehensive confidence level.

[0021] Optionally, in step S3, the dynamic threshold is adaptively adjusted by an online learning algorithm based on sliding window quantile estimation and smoothly updated in combination with scene context features, in order to determine the fire level and trigger the corresponding graded response strategy.

[0022] Optionally, in step S4, the domain-invariant feature representation space is obtained by training a self-supervised contrastive learning model and an adversarial domain classifier, so that fire samples from different scenarios are clustered in the feature space, thereby suppressing scenario-related features.

[0023] Optionally, in S4, the rapid adaptation to new scenarios employs a recurrent consistent generative adversarial network for domain adaptation training, combined with an incremental learning framework of elastic weight consolidation and dynamic memory replay, to mitigate catastrophic forgetting of old knowledge when adapting to new scenarios.

[0024] Optionally, in S5, the system self-diagnosis includes real-time detection of image blur, occlusion, and image shift, and triggers local software correction, PTZ self-calibration, or remote maintenance alarms according to the severity level; the dynamic optimization includes periodic online fine-tuning based on edge data buffers and multi-terminal collaborative model updates based on federated learning.

[0025] Compared with the prior art, this application has at least the following beneficial effects:

[0026] This application integrates the visual, thermal radiation, and smoke characteristics of flames by simultaneously acquiring and deeply fusing multimodal data in the time and space of visible light, infrared, and smoke. By utilizing an improved lightweight CNN, an adaptive background temperature model, and an environmental airflow model, it effectively distinguishes real fires from common interferences such as lights, reflections, moving heat sources, water mist, and dust, thereby improving the identification capability of early fires (especially the smoldering stage) and reducing false alarms.

[0027] This application utilizes a dynamic fire severity assessment network integrating LSTM and attention mechanisms to quantitatively analyze the temporal evolution of key parameters such as flame area, temperature rise rate, smoke concentration, and diffusion speed, outputting a crisis index and severity probability. This enables real-time assessment of fire development trends and early warning of potential risks, providing a scientific basis for graded response.

[0028] This application presents a hierarchical response strategy library built upon deep reinforcement learning. Based on dynamic evaluation results and adaptive thresholds, it can automatically trigger response actions of different levels, ranging from local audible and visual alarms and remote notifications to power cut-off and activation of fire extinguishing systems. This achieves an optimal balance between response speed, resource consumption, and risk control.

[0029] This application employs self-supervised contrastive learning and adversarial training to construct domain-invariant feature representations. It combines CycleGAN domain adaptation and an incremental learning framework based on elastic weight consolidation, enabling the model to quickly adapt to new scenarios. This greatly reduces the dependence on a large amount of labeled data in the target scenario and improves the system's generalization ability and deployment efficiency. Attached Figure Description

[0030] Figure 1 A flowchart of an image recognition-based intelligent fire detection and rapid response method provided in one embodiment of this application. Detailed Implementation

[0031] 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.

[0032] The image recognition-based intelligent fire detection and rapid response method provided in this application includes the following steps:

[0033] S1: Synchronous acquisition and preprocessing of multi-source image data. Through time-stamped visible light cameras, infrared thermal imaging cameras and smoke-sensitive cameras, video streams of the monitored area are acquired simultaneously to form spatiotemporally aligned multimodal raw data.

[0034] Adaptive brightness equalization and flame feature color gamut enhancement processing are applied to visible light images;

[0035] Specifically, for visible light images, adaptive brightness equalization based on Retinex theory is first performed to eliminate the effects of uneven scene lighting, backlighting, or local overexposure. Subsequently, joint color correction is performed in the RGB and HSV color spaces, and enhancement is achieved through a preset flame characteristic color gamut filter (mainly targeting the saturation and brightness channels of red, orange, and yellow tones) to suppress interference sources similar to flame colors (such as lights and reflective objects).

[0036] Non-uniformity correction and temperature calibration are performed on the infrared image to generate an absolute temperature distribution matrix;

[0037] Specifically, for infrared images, non-uniformity correction is first performed to eliminate fixed-pattern noise caused by differences in the response of individual detector pixels. The correction process employs a blackbody calibration method based on two or more temperature references to generate a table of gain and offset correction coefficients for each pixel. Subsequently, based on the detector's radiation response model and atmospheric transmittance parameters, the original radiation count values ​​are converted into absolute temperature values ​​or relative temperature rise values, generating a temperature distribution matrix. Furthermore, spatial temperature gradient operators (such as Sobel and Prewitt operators) are applied to calculate the rate of change of the temperature field, highlighting locally overheated regions.

[0038] Fog penetration enhancement and dynamic range compression are applied to smoke-sensitive images;

[0039] Specifically, for smoke-sensitive images, fog penetration enhancement processing based on dark channel priors or physical atmospheric scattering models is first performed to improve image contrast under low visibility conditions such as fog and haze. Subsequently, dynamic range compression is performed to map the original high dynamic range data to an 8-bit or 16-bit display range while preserving texture details and motion information in the smoke area.

[0040] S2: Multimodal feature fusion and early fire identification, based on an improved lightweight convolutional neural network to extract flame color, shape and dynamic flicker frequency features from visible light images;

[0041] The improved lightweight convolutional neural network takes a preprocessed visible light image sequence as input. Its main structure uses depthwise separable convolution to construct the feature extraction backbone and embeds a temporal attention module designed specifically for the dynamic characteristics of flames.

[0042] The network outputs two types of features in parallel: spatial features and dynamic features. Spatial features are obtained by spatial pyramid pooling from the deep feature maps of the network and are used to encode the color distribution of the flame (mainly the R / G ratio in RGB space and the statistical histogram of the H component in HSV space) and shape contour (extracted by extracting its irregularity, area, and aspect ratio through high-response regions of the activation map). Dynamic features are extracted by introducing a cascaded frame difference analysis branch into the network. This branch performs element-wise difference operations on the feature maps of consecutive frames, and then processes them through a one-dimensional temporal convolutional layer to quantize and extract the periodic flicker frequency (typically 1-10Hz) of the pixel intensity of the flame region and the pulsation intensity of local regions.

[0043] To further analyze the fire's development trend, a temporal convolutional network (TCN) is introduced. The input to this TCN is the multi-frame high-level feature vectors continuously output by the aforementioned lightweight CNN. It captures dependencies at different time scales through dilated causal convolutional layers. The TCN's output is a set of trend indicators, including but not limited to: the inter-frame growth rate of the flame area, the movement vector of the flame centroid, and the slope of the change in the overall average brightness of the region. These indicators together constitute a quantitative description of the fire's spread trend.

[0044] The specific improvement schemes for lightweight convolutional neural networks (CNNs) are as follows:

[0045] This module takes the output feature map from the preceding layer as input. First, the input feature map is processed through a depthwise separable convolutional layer, which consists of a cascaded depthwise convolution and a pointwise convolution (i.e., a 1×1 convolution) to efficiently extract spatial features and fuse channel information. Then, the resulting feature map is input into an improved channel attention submodule.

[0046] The workflow of the channel attention submodule is as follows: First, a global average pooling layer compresses the spatial dimensions (height H and width W) of the feature map, generating a 1×1×C feature vector describing the global statistics of each channel. Next, this feature vector passes through a "bottleneck" structure consisting of two fully connected layers. The first fully connected layer reduces the number of channels to C / r (r is the reduction ratio, usually set to 16) and uses the ReLU activation function; the second fully connected layer restores the number of channels to C and uses the Sigmoid activation function, thus generating a weight coefficient between 0 and 1 for each channel. Finally, the generated channel weight coefficients are multiplied and weighted channel-by-channel with the feature map output from the original depthwise separable convolution, achieving feature recalibration, enhancing the feature response of key channels in flame detection, and suppressing irrelevant or noisy channels.

[0047] The temporal attention module is described as follows: to effectively capture the dynamic evolution of flames within a short timeframe and their long-range spatial dependencies, a lightweight "spatiotemporal attention module" is introduced at a high level of the feature extraction network. The input to this module is a sequence of N consecutive frames (e.g., N=5) of feature maps from the output of the last convolutional layer of the CNN backbone network, with dimensions [N, C, H, W].

[0048] This module first uses a 1×1 convolutional layer to reduce the dimensionality of the feature map for each frame, reducing subsequent computation. Then, it employs a simplified attention mechanism based on the non-local concept for embedded deployment. Specifically, for the dimensionality-reduced feature map sequence, three parallel 1×1 convolutional layers generate three feature maps: Query, Key, and Value. Attention weights are calculated along the joint temporal-spatial dimension: the Query and Key are matrix multiplied and then normalized using Softmax to obtain a spatiotemporal attention map that encodes the correlation strength between any position in the current frame and all positions in all frames. Finally, this attention map is weighted and summed with the Value feature map to generate an output feature that enhances the spatiotemporal context information. This module enables the network to adaptively focus on the spatiotemporal context information most relevant to the dynamic changes (such as flickering and diffusion) of the current flame region.

[0049] The design of a one-dimensional convolutional layer for flicker frequency extraction is based on the flicker frequency of a flame, which is a key dynamic optical feature. To quantify and extract this feature from image sequences, a dedicated flicker frequency analysis branch is designed. This branch takes as input the temporal feature vector (e.g., the average brightness of the region or a sequence of intensity values ​​for a specific color channel within T consecutive frames) extracted from the aforementioned base network for the suspected flame region.

[0050] The core of this analysis consists of a set of parallel one-dimensional convolutional layers, each acting as a bandpass filter. The filter size (kernel size) is designed based on the target flicker frequency range (typically 1-10Hz) and the video sampling frequency (frame rate, e.g., 30fps). According to the Nyquist sampling theorem, the highest frequency that can be analyzed is 15Hz. To cover the main flicker frequency band of 1-10Hz, multiple one-dimensional convolutional kernels of different sizes are designed. For example, smaller kernels (corresponding to a time window of 3-5 frames) are more sensitive to higher frequencies (e.g., 6-10Hz), while larger kernels (corresponding to a time window of 7-9 frames) are more sensitive to lower frequencies (e.g., 1-4Hz). The number of convolutional kernels can be set according to the number of fine frequency bands to be divided as needed. For example, four kernels of different sizes can be set to roughly distinguish low-frequency, mid-frequency, mid-high-frequency, and high-frequency components.

[0051] Each one-dimensional convolutional layer is followed by absolute value activation and average pooling, converting the temporal signal into scalar features characterizing the energy intensity of that frequency band. The outputs of all parallel branches are concatenated into a multi-dimensional feature vector, which encodes the scintillation energy distribution of the input region across multiple feature frequency bands. This vector serves as an important component of the flame dynamics features and is input into the subsequent classifier for decision-making. This design provides a learnable end-to-end solution for directly extracting physically meaningful frequency features from pixel-level temporal fluctuations.

[0052] An adaptive background temperature model is constructed, based on a sequence of infrared images over a continuous period (e.g., 30 seconds to 5 minutes). Using a Gaussian mixture model or kernel density estimation method, a probability distribution of temperature values ​​is established for each pixel location in the scene. In real-time processing, the measured temperature value of each pixel in the current frame is compared with the background model, and its standardized residual is calculated.

[0053] Based on this, the two-dimensional temperature gradient field is calculated in real time, and the central difference method is used to calculate the temperature gradient components of each pixel in the x and y directions. , This process yields the gradient magnitude and direction. Abnormal temperature rise regions are identified through a dual-threshold segmentation process: First, pixels with standardized residuals exceeding a first threshold (e.g., 3-5 times the standard deviation) are marked as candidate hotspots; second, morphological closure operations are performed on these candidate regions to connect adjacent regions, and the geometric center, bounding box, average temperature, and maximum temperature of each connected region are calculated; finally, only pixels with an average temperature exceeding a second absolute threshold (e.g., ambient temperature + 100°C) are retained. , Areas that can be configured to be 20-50℃ according to the scenario and whose area is larger than the minimum effective area are identified as effective overheating hotspots.

[0054] To improve reliability, the system performs spatial alignment verification. Using a pre-calibrated homography transformation matrix or a feature-based image registration algorithm, the overheated hotspot regions identified in the infrared image are mapped to the visible light image coordinate system. Then, it is checked whether the corresponding region in the mapped visible light image is simultaneously identified as a flame candidate region by the flame feature extraction submodule. Only hotspot-flame region pairs with a spatial overlap exceeding a preset ratio (e.g., 50%) are confirmed as high-confidence fire sources.

[0055] To adapt to the varying rates of change in ambient temperature under different monitoring scenarios and to balance model update speed and stability, the background temperature model adopts a dynamic adaptive learning rate strategy. The core of this strategy is to evaluate the overall deviation between the current temperature field and the background model in real time and adjust the learning rate accordingly.

[0056] Specifically, it continuously calculates the measured temperature value of each pixel (or block region) in the current infrared frame. Compared with the background model predictions absolute residuals between Subsequently, the mean value of the residual R of the entire image frame was calculated. and standard deviation The scene stability index SS is defined as follows:

[0057] in This is an empirical adjustment coefficient.

[0058] The value range of this indicator S is (0,1]. When the ratio of the residual mean to the standard deviation is small (i.e. the scenario is stable and the changes are mild and consistent), S approaches 1; when there are drastic or disordered changes, S decreases significantly.

[0059] Finally, the pixel-level learning rate used for updating the background model in the current frame. Based on the base learning rate It is determined together with the scene stability metric S, and a local weight based on pixel residuals is introduced. :

[0060] Among them, local weights By employing a Gaussian kernel, pixels with excessive residuals (potentially foreground or anomalies) receive extremely low learning rates, thus effectively suppressing them during updates. Through this mechanism, the model updates normally to adapt to natural changes (such as day-night temperature drift) when the environment is stable. However, when large-area, high-intensity anomalies occur (such as suspected fires or strong heat source interference), the model update rate automatically slows down to prevent the background from being "contaminated" by the foreground.

[0061] For the pollution detection and treatment process of temporary high-temperature moving objects, in order to effectively distinguish between continuous fire heat sources and short-lived high-temperature objects (such as vehicles, pedestrians, and mechanical equipment), the system implements a multi-stage pollution suppression process.

[0062] First, motion consistency detection is performed. The system runs in parallel a lightweight moving target detection algorithm (such as background subtraction or optical flow) based on visible light or infrared images. When an infrared mode identifies a candidate overheated area, it immediately checks whether the area is marked as a moving object (with a coherent movement trajectory) in the corresponding visible light image sequence. If the area is confirmed as a moving object and its temperature change pattern is highly correlated with the movement trajectory (e.g., the temperature rises when the object enters the field of view and the temperature of the original area recovers rapidly after it leaves), the system marks the area as a "temporary moving heat source" and applies an extremely low learning rate or a complete freeze update to the area and its neighboring areas along the movement path during background updates.

[0063] Secondly, time-duration verification and thermal persistence management are introduced. For all candidate overheated regions not excluded by motion detection, the system maintains a historical queue of thermal events. Each thermal event records its first occurrence time, duration, highest temperature, and spatial location. A minimum duration threshold is defined. (For example, 3-10 seconds). If the duration of a thermal event is less than... If the event disappears and does not exhibit temperature growth and spatial expansion characteristics consistent with the flame diffusion pattern during its existence, the system classifies it as a transient thermal disturbance (such as thermal reflection or instantaneous thermal eruption). During background model updates, the update learning rate of the pixel region covered by the event identified as a transient thermal disturbance will be additionally suppressed for a period of time after the event ends, in order to accelerate the fading of the "thermal persistence" artifact from the background model.

[0064] Finally, a background restoration and model reset strategy is implemented. The system periodically (e.g., every 24 hours) or when a significant structural change in the scene is detected (which can be determined through global feature matching of visible light images), it initiates local or global restoration of the background model. For areas marked as contaminated and where the source of interference has been confirmed to have left, neighborhood interpolation or restoration algorithms based on historical clean background snapshots can be used to accelerate the recovery of accurate background temperature estimates. In extreme cases, such as irreversible changes in the monitoring perspective, the system can trigger a complete reset and reinitialization of the background model.

[0065] Smoke analysis is based on the pixel-level fusion results of ultraviolet and visible light images. First, the registered ultraviolet and visible light images are decomposed and fused at multiple scales based on Laplacian pyramid or wavelet transform to enhance the texture contrast and edge information of the smoke region in the fused image.

[0066] From the fused image sequence, static and dynamic features of smoke are extracted. Static texture features are obtained by calculating parameters such as the local binary mode variance, contrast, and homogeneity of the gray-level co-occurrence matrix of suspected smoke regions in the fused image. Diffusion morphology features are described by analyzing the contour changes, area growth rate, and centroid movement trajectory of the smoke region in consecutive frames. Motion trajectory is calculated using an improved dense optical flow algorithm (such as Farneback optical flow) to determine the velocity field of the smoke region and to statistically analyze its dominant flow direction and diffusion velocity.

[0067] To reduce false alarms, the system integrates a simplified environmental airflow model. This model is preset based on the physical layout of the monitored scene (such as the location of doors and windows, and vents) or estimates the main airflow direction and velocity obtained in real time from sensors. When the identified smoke trajectory deviates significantly from the dominant diffusion direction predicted by the current airflow model (e.g., the angle exceeds a threshold), and its texture features match more closely with typical feature libraries of water mist (which presents a more uniform granular feel) or dust (whose trajectory is usually more significantly affected by gravity), the area is determined to be an interference source and suppressed.

[0068] In addition, a feature fusion module with dynamically adjustable weights is designed. This module receives real-time feature vectors and confidence scores from the three sub-modules mentioned above. The weight allocation strategy is based on a predefined environmental context perceiver, which outputs the initial basic weights for each modality based on information such as time (day / night), weather (sunny, rainy, foggy), and scene visibility. , , For example, in night mode. Significantly improved, especially in dense fog. improve.

[0069] The final decision adopts an improved early warning rule based on evidence theory. The triggering condition for early fire warning is defined as follows: at least one modality has an abnormal confidence level exceeding its independent alarm threshold, and among the remaining modalities, at least one modality has an abnormal confidence level exceeding its auxiliary verification threshold (which is lower than the independent alarm threshold). For example, when the infrared modality identifies a rapidly developing overheated hot spot with high confidence, and the visible light modality, although not meeting the flame alarm standard, has a "suspected bright spot" feature confidence level exceeding the auxiliary verification threshold, the system determines it to be an early fire (possibly smoldering or in its initial stage) and triggers an early warning.

[0070] S3: Based on deep reinforcement learning, an intelligent hierarchical response is constructed to build a dynamic assessment network for fire level. The input is the fused multimodal features, and the output is the probability distribution of fire level and the crisis index. The assessment network includes a long short-term memory network to capture temporal evolution.

[0071] The core of the dynamic fire severity assessment network is a multi-head neural network structure with an attention mechanism. The network input is the multimodal state vector obtained in S2, after time alignment and feature fusion. The state vector at each time step t contains quantified data for the following key evaluation factors: the projected area of ​​the flame (number of pixels and physical space estimate), the instantaneous value of the average temperature of the flame region, and its past values. The rate of change over time (temperature rise rate), the smoke concentration index in the core area (calculated based on the transmittance attenuation model), the movement vector of the flame centroid and the dominant smoke diffusion direction vector, and the confidence level of the early warning;

[0072] The network first encodes and reduces the dimensionality of the input state through a fully connected layer. Then, the encoded features are fed into a Long Short-Term Memory (LSTM) layer to capture the temporal evolution pattern of the fire situation, outputting a hidden state containing historical context information. Finally, a fully connected layer with three output heads processes this hidden state. The first output head uses a Softmax function to generate a probability distribution of the fire severity level. The second output header outputs a scalar value representing the Crisis Index (CI) of the current state, used for ranking the severity of fires of the same level. The third output header outputs the predicted values ​​of each assessment factor for the next time step, which will be used for model building in subsequent reinforcement learning.

[0073] Specifically, the Crisis Index (CI) is a continuous scalar value that comprehensively quantifies the instantaneous danger and potential development momentum of a fire. By introducing nonlinear coupling and a time derivative term, it more accurately characterizes the "crisis" level of a fire. The CI is calculated once every assessment cycle (e.g., 1 second), and its specific calculation function is defined as follows:

[0074] Let the current time be t, the key evaluation factors include:

[0075] A(t): Normalized effective flame area (relative to the maximum measurable area of ​​the monitoring field of view);

[0076] : Normalized core region temperature rise rate (temperature change per unit time).

[0077] Normalized smoke concentration index;

[0078] Normalized flame front spatial diffusion rate (based on centroid movement and profile changes);

[0079] G(t): The overall confidence level after multimodal fusion (range [0,1]);

[0080] First, each factor is normalized to map its value range to the [0,1] interval. The normalization radix (denominator) is not a fixed value, but is dynamically set based on the historical maximum value or physical upper limit of the specific monitoring scenario. For example: This represents the maximum possible area of ​​fire within the camera's field of view. The maximum value is set based on the material's flammability (e.g., 100°C / s).

[0081] The core calculation function of the crisis index CI(t) adopts the form of a log-linear combination and interaction terms:

[0082]

[0083] in:

[0084] w1, w2, w3, w4, and w5 are adjustable weight coefficients that satisfy Σwi=1 (i=1 to 4, w5 is the independent weight of the interaction term). Their initial values ​​are determined based on expert knowledge or historical data during offline training and can be fine-tuned in subsequent online learning.

[0085] , This is a scaling factor used to adjust the sensitivity of area and temperature rise rate in the function, ensuring that they reach a suitable numerical range before logarithmic operations.

[0086] The introduction of this technology makes CI sufficiently sensitive to initial small flames and slow temperature rises (i.e., the germination period), while avoiding situations where A(t) or The infinite expansion of CI when the value is very large is consistent with the diminishing marginal effect of crisis perception;

[0087] Interactive items Used to capture the extremely high-risk mode of "large-area rapid heating", its effect is to generate an additional nonlinear increment to CI when both the area and the rate of temperature rise are high, which is more in line with the physical laws of fire dynamics.

[0088] The overall confidence level G(t) serves as a global multiplier. When multi-source information is contradictory or of low quality, it directly reduces the output value of CI, reflecting the uncertainty of the system's own judgment.

[0089] The function outputs CI(t), a dimensionless continuous value, theoretically ranging from... However, through parameter design and normalization, it is usually constrained to a meaningful range (such as 0~10) in practical applications, which facilitates intuitive understanding and threshold setting.

[0090] Dynamic threshold , , To address the varying risk tolerance levels across different application scenarios (such as chemical plants and libraries), as well as baseline drift caused by environmental and equipment aging within the same scenario, this system employs an online learning algorithm based on sliding window quantile estimation to dynamically adjust the trigger thresholds for each level of response. , , ;

[0091] The core steps of the algorithm are as follows:

[0092] Data pool maintenance: The system maintains three first-in-first-out (FIFO) queue data pools for each monitoring point, corresponding to the CI sequences recorded when historical CI values ​​were determined to be at the "emergence," "development," or "spread" levels. Only after a complete fire event (from warning to cancellation) has ended, and the event has been manually reviewed or the system has confirmed the correct classification with high confidence, will the CI sequences generated during the process be classified and stored in the corresponding data pools.

[0093] Quantile Regression Model: For each level k (k=1,2,3), a lightweight conditional quantile regression model is trained online using historical CI data from its corresponding data pool. This model aims to estimate the conditional distribution of CI values ​​for that level given a certain "scene context feature vector X" (which may include time, season, weather, device status, etc.). We pay particular attention to its upper tail quantiles, for example, setting them as quantiles. This means that CI values ​​exceeding this quantile are considered to belong to the more dangerous portion of that level that "requires a response";

[0094] Dynamic threshold calculation involves the system acquiring the current scene context feature vector in real time during runtime. Inputting this data into the conditional quantile regression model corresponding to the three levels yields the estimated CI conditional quantile values ​​for the three levels under the current conditions. This estimated value is used as the base value for the dynamic threshold:

[0095] To prevent the threshold from fluctuating drastically due to a single abnormal event, an exponential smoothing filter is applied to the base threshold:

[0096] in For example, a smoothing factor (e.g., 0.9);

[0097] Security boundaries and manual intervention: the system allows administrators to set absolute upper and lower thresholds for each level. ), dynamic threshold This will be constrained within this scope to ensure security. At the same time, a management interface will be provided, allowing experts to directly adjust based on a review of recent events. The value or the weight of specific historical events in the data pool can be adjusted to indirectly and smoothly affect the threshold.

[0098] This algorithm enables the response threshold to automatically adapt to different time periods (such as high population density during the day and low population density at night), different seasons (dry summer and humid winter), and changes in equipment sensitivity, thereby achieving adaptive adjustment of risk sensitivity.

[0099] The fire severity level is determined using a dynamic threshold method. When the probability of the highest severity level exceeds a preset threshold (e.g., 0.7), and its crisis index (CI) exceeds the dynamic threshold corresponding to that level (this threshold can be adaptively adjusted based on historical fire data), the fire is determined to have entered that level. This design allows the system to distinguish between situations of "early development" and "rapid development" that are of the same level but have different degrees of danger.

[0100] A predefined, extensible, tiered response strategy library is provided, which links specific response actions to fire severity levels and crisis indices. Each response action is defined as a tuple (trigger condition, execution instruction set, expected effect, rollback mechanism), for example:

[0101] Level 1 Response (Emerging Stage): The triggering condition is " The execution instruction set includes: activating the high-decibel audible and visual alarms at the monitoring points; sending structured alarm information containing location, snapshots, and evaluation data to the preset monitoring center duty station; and automatically recording and saving 30-second video clips before and after the recording.

[0102] Level 2 Response (Development Phase): Triggering condition is " The execution instruction set, in addition to the first-level actions, adds: sending instructions to the smart distribution box in the fire area via the Internet of Things to cut off non-critical power circuits; shutting down the ventilation or air conditioning system in the area to slow airflow; activating emergency lighting; and unlocking preset safety escape route access control.

[0103] Level 3 Response (Proliferation Phase): Triggering condition is " The execution instruction set, in addition to including secondary actions, adds the following core features: triggering automatic fire extinguishing devices (such as gas or fine water mist) within the area; packaging and pushing real-time video streams, temperature heat maps, building floor plans, and personnel heat maps (if available) to the city fire command system via a dedicated data link; and playing preset evacuation voice instructions via the broadcast system.

[0104] The execution status of all commands (success, failure, timeout) is monitored and reported in real time. The system is designed with an action rollback mechanism. For example, once the fire is confirmed to be extinguished, reverse operations such as restoring power supply and ventilation can be performed automatically or after confirmation.

[0105] To train the agent to make optimal sequence decisions in complex and uncertain real-world environments, this system employs a proximate policy optimization algorithm for training in a simulated fire environment, as detailed below:

[0106] The simulation environment is constructed using a simplified model based on real physical laws. The state transition function incorporates the prediction head of the fire assessment network and introduces random perturbations to simulate uncertainty. The reward function R is carefully designed as a multi-objective weighted sum. Among them, fire loss estimates are dynamically calculated based on flame area, temperature, and simulated asset value maps; response costs include quantitative indicators such as energy interruption and equipment wear and tear.

[0107] Agent training, where the input to the agent (Actor-Critic architecture) is the environment state. (i.e., the state vector of the fire assessment network), the output is the probability distribution of each possible response action in the current state, and the Critic network evaluates the value of the state. During training, the agent learns the policy through a large number of episodes. The goal is to balance response speed, action effectiveness and cost control under the premise of maximizing long-term cumulative reward. The trained policy network can be deployed in the online system. Based on the real-time fire situation, it outputs the probability of the optimal response action. The system selects the action with the highest probability or performs it by sampling according to probability to increase the exploratory nature.

[0108] A human-machine collaborative optimization closed loop is established. After each real alarm and response process, the system automatically generates a handling report. Authorized senior managers or experts can evaluate the system's assessment level and the sequence of response actions taken through the feedback interface (such as "appropriate assessment", "insufficient response", "over-response"). They can also manually correct or supplement the response actions. These "status-action-result" data pairs with scores and correction labels will be stored in a priority experience playback pool.

[0109] The system periodically (e.g., weekly) initiates an offline fine-tuning process, using accumulated human feedback data to fine-tune the reinforcement learning policy network. Fine-tuning employs a conservative policy iteration approach, ensuring that new policies gradually align with human expert preferences while inheriting existing good decision-making patterns, and avoiding catastrophic forgetting of existing policies. This mechanism enables the system to adapt to new risk profiles, equipment upgrades, or changes in operational strategies.

[0110] S4: Cross-scenario adaptive transfer learning, based on cross-domain multi-scenario fire data, constructs a domain-invariant feature representation space through self-supervised comparative learning and adversarial training;

[0111] Specifically, to achieve the transferability of the model across different physical scenarios, a hierarchical cross-domain fire feature library is constructed. The construction of this feature library adopts a semi-automated method based on self-supervised pre-training and expert knowledge guidance.

[0112] First, normal and fire-related images and video sequences covering the target domain (such as industrial workshops, forests, ship cabins, and high-rise buildings) were collected from public datasets, collaborating institutions, and simulation platforms to form the original multi-domain dataset. This dataset only requires scene type labels and does not need detailed fire annotations. A convolutional neural network pre-trained on a large general-purpose image dataset was then used as the initial feature extractor to extract deep features from all images.

[0113] Secondly, a domain-invariant feature learning network is designed. This network takes a triple (anchor sample, positive sample, negative sample) as input and is trained using an improved contrastive loss to learn a unified feature representation space. Specifically, anchor samples and positive samples come from different fire instances of the same scene type, while negative samples come from fire or non-fire images of other scene types. The training objective is to minimize the distance between anchor samples and positive samples in the feature space, while maximizing the distance with negative samples. Furthermore, a domain classifier is introduced as an adversarial component to attempt to distinguish scene types from features, while a feature extractor is trained to confuse this classifier, thereby forcing the network to strip away scene-specific features and retain domain-invariant features relevant to the essence of the fire.

[0114] Ultimately, the trained domain-invariant feature learning network can map any input image to a unified feature representation space. In this space, fire samples from different scenarios cluster in similar regions, while scene-related background features are suppressed. The high-level feature vectors output by this network constitute the core representation of the cross-domain fire feature library, providing high-quality initial features for subsequent domain-adaptive training.

[0115] When the back-end of the South is deployed to a new target scene, in order to quickly achieve high accuracy with limited new scene labeled data, a fine-tuning strategy that combines generative adversarial networks and minimizing domain differences is adopted.

[0116] We maintain a source domain pre-trained model, which is a complete fire identification and assessment model trained on existing multi-source data (feature library). For new target scenarios, we collect a small amount of labeled data (labeled data in the target domain) and a large amount of unlabeled scene images (unlabeled data in the target domain).

[0117] The core of domain adaptive training is a recurrent consistent generative adversarial network (RCA). This network contains two generators (…). , ) and two discriminators ( , Generator Learning to convert the style of a source domain image to the style of a target domain image, while Then perform the opposite transformation, discriminator. and They are used to distinguish between real target domain images and images made from... The generated image, and how to distinguish it from the real source domain image. The generated image;

[0118] The training process includes three losses: adversarial loss, cycle consistency loss, and task-specific loss. Adversarial loss ensures that the generated images are stylistically indistinguishable from the target (or source) domain. Cycle consistency loss requires that an image transformed from the source domain to the target domain and back should be as consistent as possible with the original image to preserve semantic content (e.g., fire area). Task-specific loss involves inputting the style-transformed source domain image (which looks like the target domain image) into a pre-trained source domain model to calculate the loss for the fire recognition task (e.g., classification cross-entropy). Simultaneously, labeled target domain data is input into a model sharing a backbone network but with the task header initialized to the source domain model parameters, and the supervised loss is calculated.

[0119] By jointly optimizing these losses, the model's backbone network gradually learns to extract domain-invariant features that are discriminative to both the source and target domains, while the task head adapts to the data distribution in the target domain. This method effectively utilizes the knowledge of the source domain model and a large amount of unlabeled target domain data, reducing the dependence on a large amount of labeled target domain data.

[0120] To enable the system to continuously adapt to new scenarios or slow changes within the same scenario (such as equipment updates or seasonal changes) without forgetting the knowledge it has learned, this system designs an incremental learning framework based on elastic weight consolidation and dynamic memory replay.

[0121] When fine-tuning training for a new scenario or a new batch of data, the system calculates the importance of each parameter in the source domain pre-trained model for the previous task (usually approximated by the diagonal value of the Fisher information matrix of the parameter on the old data). In the target loss function of fine-tuning, a regularization term is added to penalize large modifications to important parameters, so that while the model adapts to new data, the key parameters do not deviate drastically from their optimal values ​​learned on the old task, effectively mitigating catastrophic forgetting.

[0122] A fixed-size sample memory is maintained to store a small number of typical samples (including images and their feature representations) representing various past scenes. The update strategy of the memory adopts a diversity preservation algorithm based on categories and scenes. When new data arrives, it is determined whether to replace the old samples based on their feature similarity with the samples in the memory and the scene they belong to. During incremental learning training, each iteration not only uses data from new scenes, but also uniformly samples old scene data from the memory to participate in the training, so that the model can continuously "review" old knowledge.

[0123] For multiple learned scenarios, the system learns a lightweight "scenario descriptor" vector for each scenario. When the system is deployed in a known scenario, it can quickly call the corresponding optimized model parameters by matching the scenario descriptor; if it is a completely new scenario, the above domain adaptation and incremental learning process is initiated, and a new descriptor and model branch are generated for the new scenario.

[0124] S5: System self-diagnosis and dynamic optimization, real-time monitoring of image quality, occlusion and image stability of each camera, and triggering software correction, active self-calibration or maintenance alarm according to the diagnostic results.

[0125] To ensure the reliability of the front-end sensing data, the system incorporates a real-time, online health status diagnostic pipeline for each imaging unit. This pipeline analyzes the image quality, integrity, and geometric stability in parallel and triggers corresponding maintenance actions.

[0126] The image quality analysis submodule periodically extracts frames from the video stream for analysis. For blur detection, it calculates the local gradient variance or Laplacian arithmetic response. If the sharpness index of the entire frame or a specific region consistently falls below a dynamic threshold (based on historical sharp frames), the lens is suspected of being contaminated or out of focus. For occlusion detection, it combines background modeling with foreground analysis. Long-term static foreground areas with abnormally uniform textures (e.g., color blocks) and regular edges may be identified as man-made occlusions (e.g., stickers) or insect nests; large areas with abrupt and persistent color and texture changes may indicate lens smudges. The system calculates the area percentage of the occluded region; if it exceeds a set threshold, an alarm is triggered.

[0127] The image stability monitoring submodule continuously tracks a set of stable, high-contrast natural feature points (such as building corners and fixed landmarks) in the image. After system initialization or each self-calibration, the reference positions of these feature points are recorded. During operation, these points are tracked in real time using optical flow or feature matching. If a large number of feature points undergo consistent translation or rotation exceeding the allowable range, it is determined that the camera has physically shifted or become loose. The system calculates the rotation angle and translation vector of the shift.

[0128] Based on the diagnostic results, the system executes a graded response and self-calibration trigger:

[0129] Level 1 Alarm (Performance Degradation): If slight blurring or less than 5% local occlusion is detected, the system marks the sensor data confidence level as reduced in the background and provides a maintenance prompt in the management interface;

[0130] Level 2 Alarm (Function Impairment): If moderate to severe blurring, large-area occlusion, or slight shift is detected, the system will automatically attempt software correction (such as applying an image deblurring filter or masking the occluded area). Simultaneously, a local audible and visual alarm will be triggered, and a high-level alarm will be sent to the monitoring center, clearly specifying the fault type and location.

[0131] For physical offsets, the system can control units supporting gimbals or motorized lenses to automatically adjust in the opposite direction based on the calculated offset vector, attempting to restore the original viewpoint. For out-of-focus issues, the system can control the motorized lens to perform a focusing scan until the sharpness index is restored. All self-calibration operations have strict safety boundaries and limited attempts; failures escalate to a request for manual intervention.

[0132] To cope with slow environmental changes (such as seasonal changes, renovations, and equipment aging) and to improve generalization capabilities, the system is designed with a dual-track model dynamic optimization mechanism.

[0133] The periodic online fine-tuning system maintains a fixed-size circular data buffer locally on the edge computing unit, continuously storing recently collected, high-quality labeled (either automatically labeled with high confidence by the system and filtered by simple rules) "normal" and "fire" samples. Every fixed period (e.g., 24 hours), or when a statistically significant decrease in the model's average confidence on recent data is detected, a fine-tuning task is automatically initiated. Fine-tuning uses the data in the buffer and employs a flexible weight consolidation strategy. When calculating new loss, it imposes constraints on the model's original key parameters to prevent overfitting to new data and forgetting old knowledge. The fine-tuning process is entirely local, and the updated model takes effect immediately, forming a closed-loop optimization.

[0134] Multi-terminal collaborative optimization based on federated learning aims to improve the overall performance of the model by leveraging collective intelligence while protecting the data privacy of each site. The system adopts a cross-terminal federated learning framework.

[0135] Each edge node participating in federated learning acts as a client. After completing the local fine-tuning cycle, it encrypts the update amount of the model parameters (i.e. the difference between the old and new model parameters) and attaches a summary of the amount of data and category distribution used for this fine-tuning (excluding the original data) for uploading.

[0136] The central server periodically gathers eligible clients (e.g., those with idle network access and sufficient battery power). Upon receiving model updates from each client, it aggregates them using a secure weighted average algorithm. Weights are dynamically allocated based on the client's data volume and data quality (confidence level). The aggregation produces a global model update.

[0137] Before uploading updates to the client, calibrated Gaussian noise can be added to the update vector to ensure that no information about any individual training sample can be deduced from the update, thus meeting strict privacy protection requirements.

[0138] The central server distributes the aggregated global updates to all clients. Instead of directly replacing their local models, clients intelligently integrate the global updates with their local data distribution (e.g., using model interpolation or selective integration for specific layers). This allows them to gain global knowledge gains while retaining their specific capabilities tailored to their own scenarios. This mechanism enables systems deployed in different environments (such as frigid outdoor environments in the north and humid, hot factory buildings in the south) to learn from each other and evolve together without sharing sensitive on-site video data.

[0139] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for intelligent fire detection and rapid response based on image recognition, characterized in that, The method includes: S1. Synchronous acquisition and preprocessing of multi-source image data: The video stream of the monitored area is synchronously acquired through the time-stamped visible light camera, infrared thermal imaging camera and smoke-sensitive camera to form spatiotemporally aligned multimodal raw data, and the images of each modality are enhanced and corrected respectively. S2. Multimodal Feature Fusion and Early Fire Identification: Based on the multimodal image data preprocessed in S1, an improved lightweight convolutional neural network is used to extract the flame color, shape, and dynamic flicker frequency features in the visible light image. An adaptive background temperature model is used to identify abnormal temperature rise areas in the infrared image. Smoke features are extracted based on the fusion of ultraviolet and visible light images. The features of each modality are fused through a feature fusion module with dynamically adjustable weights to achieve early fire identification. S3. Intelligent hierarchical response based on deep reinforcement learning: Based on the multimodal features fused in S2, a dynamic fire level assessment network is constructed. The multimodal features are input, the probability distribution of fire level and crisis index are output, and the fire level is determined according to the dynamic threshold to trigger the predefined hierarchical response strategy. S4, Cross-Scene Adaptive Transfer Learning: Based on cross-domain multi-scenario fire data, a domain-invariant feature representation space is constructed through self-supervised contrastive learning and adversarial training, and a generative adversarial network and incremental learning framework are combined to achieve rapid adaptation to new scenarios in S1. S5. System self-diagnosis and dynamic optimization: Real-time monitoring of image quality, occlusion, and image stability of each camera in S1. Based on the diagnostic results, software correction, active self-calibration, or maintenance alarms are triggered in stages. The model in S2 to S4 is continuously optimized through periodic online fine-tuning and federated learning.

2. The intelligent fire detection and rapid response method based on image recognition according to claim 1, characterized in that, In S1, the enhancement and correction processing of each modal image includes: adaptive brightness equalization and flame feature color gamut enhancement processing of the visible light image; Non-uniformity correction and temperature calibration are performed on the infrared image to generate an absolute temperature distribution matrix; Fog-sensitive images are enhanced with fog penetration and dynamic range compression.

3. The intelligent fire detection and rapid response method based on image recognition according to claim 1, characterized in that, In S2, the improved lightweight convolutional neural network includes a feature extraction backbone constructed from deep separable convolutions, an embedded temporal attention module, and a cascaded frame difference analysis branch and a one-dimensional convolutional layer group for extracting the dynamic flicker frequency of the flame.

4. The intelligent fire detection and rapid response method based on image recognition according to claim 1, characterized in that, In step S2, the adaptive background temperature model uses a Gaussian mixture model or kernel density estimation method to establish a temperature background probability distribution for each pixel, and updates the background model through a dynamic adaptive learning rate strategy, combined with spatial alignment verification to improve the reliability of fire identification.

5. The intelligent fire detection and rapid response method based on image recognition according to claim 1, characterized in that, In S2, the feature fusion module with dynamically adjustable weights performs weighted fusion of the confidence levels of visible light, infrared and smoke features based on the initial basic weights of each modality output by the environmental context sensor, and uses an improved early warning rule based on evidence theory to make early fire judgment.

6. The intelligent fire detection and rapid response method based on image recognition according to claim 1, characterized in that, In S3, the fire level dynamic assessment network includes a long short-term memory network to capture the temporal evolution of the fire state and output the fire level probability distribution, crisis index and the state prediction value for the next time step; the crisis index is calculated by a nonlinear combination of normalized flame area, temperature rise rate, smoke concentration, diffusion rate and comprehensive confidence.

7. The intelligent fire detection and rapid response method based on image recognition according to claim 1, characterized in that, In step S3, the dynamic threshold is adaptively adjusted by an online learning algorithm based on sliding window quantile estimation and smoothly updated in combination with scene context features, which is used to determine the fire level and trigger the corresponding graded response strategy.

8. The intelligent fire detection and rapid response method based on image recognition according to claim 1, characterized in that, In step S4, the domain-invariant feature representation space is obtained by training a self-supervised contrastive learning model and an adversarial domain classifier, which enables fire samples from different scenarios to cluster in the feature space and suppress scene-related features.

9. The intelligent fire detection and rapid response method based on image recognition according to claim 1, characterized in that, In step S4, the rapid adaptation to new scenarios employs a recurrent consistent generative adversarial network for domain adaptation training, combined with an incremental learning framework of elastic weight consolidation and dynamic memory replay, to mitigate catastrophic forgetting of old knowledge when adapting to new scenarios.

10. The intelligent fire detection and rapid response method based on image recognition according to claim 1, characterized in that, In S5, the system self-diagnosis includes real-time detection of image blur, occlusion and screen offset, and triggers local software correction, PTZ self-calibration or remote maintenance alarm according to the severity level; the dynamic optimization includes periodic online fine-tuning based on edge data buffer and multi-terminal collaborative model update based on federated learning.

Images