False alarm prevention ai fire review device and method

Abstract

The application provides an anti-misreporting AI fire review device and method, which realizes in-situ intelligent review of air suction type smoke warning, accurate filtering of salt spray interference, and high-precision positioning of fire by constructing a marine special fire multi-modal data set, improving a lightweight detection model, deploying a local quantization of an edge NPU, a multi-dimensional progressive time sequence review logic, and a homography matrix spatial positioning algorithm, solving the technical problems of high-frequency misreporting, network disconnection failure, and inaccurate positioning of the closed electrical core cabin of the marine booster station, and adopting a pure edge local inference, which is reliable and efficient in operation, accurate in small target identification, fast in response speed, accurate in distortion adaptation and positioning during operation, and efficient in emergency disposal, thereby effectively reducing the operation and maintenance cost of the marine booster station.

Description

Technical Field

[0001] This invention relates to the field of fire prevention, and in particular to an AI-based fire verification device and method for preventing false alarms. Background Technology

[0002] Offshore booster stations are exposed to harsh marine environments characterized by high salt spray, high humidity, strong temperature differences, and strong electromagnetic interference. They generally adopt an "unmanned or minimally staffed" operation and maintenance mode. Due to their long distance from the shore, high cost of offshore maintenance, and difficulty in on-site emergency response, core electrical areas such as cable mezzanines, GIS rooms, high-voltage electrical cabinets, and main transformer compartments rely mainly on aspirating early smoke detectors for fire early warning. However, these detectors rely solely on a single smoke concentration parameter for alarm determination and are highly susceptible to interference from salt spray aerosols, water vapor condensation, marine dust, equipment arc flashes, and lens beads, resulting in high-frequency false alarms.

[0003] Currently, the industry is gradually adopting video-linked verification methods to reduce false alarms, but existing technologies have several fundamental flaws and cannot be adapted to the specific operating conditions of offshore booster stations: 1. Poor model versatility and weak anti-interference capability: Existing AI verification solutions mostly use YOLO and SSD models trained on general datasets such as COCO and VOC. They are not trained for specific scene features such as hazy salt fog at sea, lens condensation, electric arc flash, and thin smoldering smoke. They cannot distinguish between white salt fog aerosols and real electric smoldering smoke, and generally have the problem of false judgment of "reporting as soon as white is seen".

[0004] 2. Dependence on backend architecture and insufficient reliability: Traditional video AI verification is mostly based on the architecture of "front-end acquisition + cloud / backend server inference", which is highly dependent on submarine optical cables and remote servers. Once the offshore network is interrupted or the server crashes, the verification function will be completely invalid, which cannot meet the high reliability operation requirements of unattended booster stations.

[0005] 3. Simple identification logic and low anti-interference ability: Existing solutions mostly use instantaneous judgment of single-frame images, without a continuous temporal correlation verification mechanism, which cannot filter short-term interference such as salt fog clouds that drift with the wind and instantaneous reflections, resulting in extremely poor false alarm filtering effect.

[0006] 4. Unclear fire location and low response efficiency: Traditional verification can only realize zone alarms and cannot output the precise physical coordinates of the fire point. Maintenance personnel cannot quickly locate the hidden danger and delay emergency response.

[0007] Therefore, existing technologies cannot simultaneously meet the core requirements of offshore substations, such as high interference environment, reliable offline operation, low false alarms, and accurate positioning. There is an urgent need to develop an AI-based fire prevention and false alarm verification method and dedicated device that is adapted to offshore substations, adopts edge-localized deployment, and multi-dimensional progressive verification. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an AI-based fire verification method and device for preventing false alarms in offshore substations. By constructing a multimodal dataset of fires specific to the sea, improving the lightweight detection model, deploying local quantization of the edge NPU, implementing multi-dimensional progressive temporal verification logic, and using a homography matrix spatial positioning algorithm, this invention achieves on-site intelligent verification of aspirating smoke warnings, accurate filtering of salt spray interference, and high-precision fire location. It solves the technical problems of high-frequency false alarms, network disconnection failures, and inaccurate positioning of marine detectors. This invention is specifically applied to the very early fire verification of the enclosed electrical core compartments (cable interlayer, GIS room, high-voltage electrical cabinet, main transformer compartment) of offshore substations, excluding non-target scenarios such as wind turbine nacelles, offshore towers, and onshore substations, and precisely focusing on the subdivided technical gaps in unmanned offshore substations.

[0009] In a first aspect, the present invention provides an AI fire verification device to prevent false alarms, comprising an early smoke detection unit (101), a salt spray camera unit (102), an integrated main control unit (103), a dual-link communication unit (104), an effective fire confirmation and fire point coordinate output unit (105), and an AR visualization label (106), wherein: The very early smoke detection unit (101) has a built-in adjustable speed air pump, anti-salt spray filter, LED detection chamber and airflow sensor. It dynamically adjusts the air intake flow according to the environmental load, and is equipped with a backflush air path module to realize automatic self-cleaning of the pipeline network. It also transmits the signal to the integrated main control unit (103). The integrated main control unit (103) further includes an edge AI main control module (1031), a false alarm prevention composite model (1032) and a salt spray distortion positioning module (1033). The anti-salt spray camera unit (102) is directly connected to the integrated main control unit (103) via an RJ45 interface. When the very early smoke detection unit (101) outputs a smoke exceeding the limit warning signal, it collects real-time high-definition video stream images of the monitoring cabin. The dual-link communication unit (104) exchanges signals bidirectionally with the integrated main control unit (103); The effective fire confirmation and fire point coordinate output unit (105) receives the AI ​​reasoning result output by the integrated main control unit (103) and completes the fire determination according to the three-level progressive timing review logic; The AR visualization identifier (106) receives the physical coordinates of the fire point output by the valid fire confirmation and fire point coordinate output unit (105) and the signal from the integrated main control unit (103), and completes the coordinate registration and mapping by combining the imaging parameters of the on-site camera and the indoor CAD plan base map data of the booster station.

[0010] Secondly, this invention provides a method for preventing false alarms in AI fire verification, comprising the following steps: Step 201: Inhalation-type smoke concentration exceeding the limit warning. When the detected smoke index exceeds the preset first-level warning threshold, an exceeding the limit warning trigger signal is generated. Step 202: The anti-salt spray camera acquires video streams, and the anti-salt spray camera unit acquires real-time high-definition video image sequences of the monitoring area and outputs continuous frame images; Step 203: Image dehazing and water droplet blur preprocessing. Receive the original video frames output in step 202. Use the dark channel prior dehazing algorithm to eliminate image grayscale distortion caused by sea salt fog. Use the guided filtering algorithm to repair local blurred pixels caused by lens beads. Normalize and scale the image to generate a standardized image that matches the NPU model input specifications. Step 204: Local offline inference at the edge: The standardized image preprocessed in step 203 is sent to the onboard embedded NPU, and the offshore substation-specific fire identification model quantized by INT8 is loaded. The inference outputs the multi-class detection results of each frame of the image, including the target category, the target pixel box position, and the confidence score of each category. Step 205: Timing stability determination, receive the multi-class detection result of the single frame image output in step 204, and execute the first-level progressive verification logic; Step 206: Fireworks association determination, receive the timing stability determination result output in step 205, and execute the second-level progressive verification logic; Step 207: Multi-class confusion suppression determination, receive the fireworks association determination result output in step 206, and execute the third-level progressive review logic; Step 208: Valid fire confirmation, receive the multi-class confusion suppression judgment result output from step 207, output the final verification result of three-level progressive review, and complete the fire situation determination; Step 209: False alarm attribution analysis. Receive the ruling result from step 208 and perform attribution identification analysis on the reviewed events. Step 210: Homography matrix coordinate mapping. Receive the target pixel box data transmitted from step 209, and convert the pixel coordinates of the target box points in the image pixel coordinate system into the physical coordinates of the substation site plane through inverse matrix transformation. The coordinate mapping calculation result is transmitted to step 211. Step 211: Output the physical coordinates of the fire point, receive the valid fire situation adjudication command output in step 208, retrieve the physical coordinate values ​​calculated in step 210, and output the standardized fire information message. Step 212: Overlay the video image with the base map, receive the branch signal output from Step 208, read the output physical coordinates of the fire point, overlay and render the AR fire point highlight mark onto the real-time monitoring video image and the indoor CAD base map of the substation, generate a visualized video stream with AR markers, and push it to the onshore maintenance terminal; at the same time, receive the short-term interference event from Step 205, overlay the interference location mark onto the image, save the interference event recording and base map snapshot as maintenance records, and after completing the interference event recording, return to Step 205 to continue performing timing monitoring on subsequent video frames.

[0011] Furthermore, the following offline deployment preparation steps are included before step 201: Step S1: Construct a multimodal training dataset specifically for offshore booster stations; Step S2: Construct and optimize a lightweight fire detection basic model; Step S3: Model transfer training, pruning, and edge quantization deployment; Step S4: Edge-adaptive image preprocessing and progressive AI verification inference; Step S5: Precise fire location using monocular vision homography matrix.

[0012] Furthermore, the preset first-level warning threshold is 0.015dB / m.

[0013] Furthermore, the output target category includes any one or a combination of smoke, flame, salt spray interference, lens water droplets, and electric arc interference.

[0014] The technical solution of this invention has the following beneficial effects: 1. Scenario-specific adaptation, excellent anti-interference performance, and significant incremental value: Utilizing a dedicated positive and negative sample dataset for offshore booster stations and a GAN adversarial sample training mechanism, this invention specifically distinguishes between salt spray, water droplets, electric arc interference, and real smoke and flame characteristics. Combined with a three-level progressive temporal verification logic with priority constraints, after six months of real-world testing in marine areas, the false alarm filtering rate of this invention reaches over 99.5%, significantly better than the 98% false alarm filtering rate of mainstream industry solutions. The average number of false alarms per day is ≤0.1, compared to more than 15 false alarms per day for traditional air-breathing detection + general AI verification solutions. This solves the technical problems of poor scenario adaptability of general models and high-frequency false alarms at sea. Furthermore, offshore operations and maintenance have extremely specific scenarios, with booster stations generally located more than 20 kilometers from shore. The cost of invalid offshore operations and maintenance triggered by a single false alarm exceeds 50,000 yuan. This invention can reduce the frequency of invalid offshore operations and maintenance by 95% and the overall operation and maintenance cost by more than 70%.

[0015] 2. Pure edge local inference, offline high-reliability operation: The model is completely embedded in the onboard NPU of the device. Image preprocessing, AI inference, fire situation determination and positioning calculation are all completed locally. It does not rely on the cloud, backend server and external network. It can still work independently under extreme network outage and device downtime conditions. It is fully adapted to the high-reliability operation requirements of unattended and unmanned operation and maintenance at sea, which is fundamentally different from the existing cloud remote verification technology.

[0016] 3. Accurate small target identification, fast response speed, and strong model adaptability: For the small target characteristics of smoldering, thin smoke from offshore electrical fires, a customized optimization scheme combining the NWD normalized Wasserstein distance loss function and CBAM attention mechanism is adopted. This is a dedicated optimization strategy adapted to the weak features and strong interference scenarios of offshore substations. After INT8 quantization, the model's accuracy loss is ≤0.9%, the single-frame inference time is stable at 22–28ms, and the overall fire identification latency is less than 3 seconds, meeting the requirements for very early warning of electrical fires. Standardized comparative experiments verify that under the same offshore salt spray interference conditions and the same hardware equipment, the general COCO pre-trained YOLOv8 model has a misclassification rate of 47.6% and a false negative rate of 31.2% in salt spray scenarios. The optimized model of this invention has a misclassification rate as low as 0.4% and a false negative rate of <1%, fully demonstrating the necessity of the dedicated dataset and customized model optimization.

[0017] 4. Precise positioning with distortion adaptation and efficient emergency response: Based on the homography matrix coordinate mapping algorithm with optimized salt spray distortion compensation, high-precision fire positioning within 0.5 meters can be achieved without binocular cameras. The AR visualization overlay of the fire point location provides accurate coordinate basis for remote fire fighting and equipment power failure interlocking under unattended operation at sea, solving the problems of fuzzy positioning and inability to accurately trace the source in traditional positioning.

[0018] 5. Dedicated hardware for offshore engineering, resulting in significant cost reduction and efficiency improvement: The whole machine is designed with corrosion protection, dehumidification, self-cleaning, and redundant power supply and communication. It is fully adapted to the harsh environment of high salt spray and high interference at sea, enabling long-term maintenance-free operation. It can reduce the cost of offshore maintenance by more than 70%, reduce the frequency of ineffective offshore maintenance by 95%, and eliminate equipment downtime and waste of emergency resources caused by false alarms. Attached Figure Description

[0020] Figure 1 : A schematic diagram of the overall hardware architecture of the AI ​​fire verification device for preventing false alarms of the present invention; Figure 2 : Schematic diagram of the AI ​​fire verification method for preventing false alarms in this invention; Figure 3 : Flowchart of model training, pruning quantization, and edge deployment in this invention; Figure 4 This invention provides a schematic diagram of the fire spatial positioning principle based on homography transformation. Detailed Implementation

[0022] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0023] Please refer to Figure 1 The schematic diagram of the overall hardware architecture of the AI ​​fire verification device for preventing false alarms of the present invention includes an early smoke detection unit 101, an anti-salt spray camera unit 102, an integrated main control unit 103, a dual-link communication unit 104, an effective fire confirmation and fire point coordinate output unit 105, and an AR visualization label 106, wherein: The very early smoke detection unit 101 has a built-in adjustable speed air pump, anti-salt spray filter, LED detection chamber and airflow sensor. It can dynamically adjust the air intake flow according to the environmental load and achieve automatic self-cleaning of the pipeline network when combined with the backflush air path module.

[0024] The salt spray-resistant camera unit 102 is an industrial-grade video acquisition device adapted to the high salt spray and high humidity environment of offshore substations. It features lens self-cleaning and anti-condensation heating functions. Hardware-wise, it is directly connected to the integrated main control unit via an RJ45 interface and supports PoE power supply. When the early smoke detection unit outputs a smoke exceedance warning signal, this unit is triggered and starts to acquire real-time high-definition video stream images of the monitoring cabin. Under normal operating conditions, it can also continuously output video data as needed. The acquired raw image data is transmitted to the integrated main control unit 103 for subsequent image preprocessing and edge AI fire inference recognition, providing raw visual input for smoke and flame target detection. Simultaneously, the lens self-cleaning and heating anti-condensation mechanisms reduce image blurring and distortion caused by salt spray adhesion and water droplet buildup, ensuring imaging quality in harsh marine environments.

[0025] The integrated main control unit 103 is solidified within the housing, featuring an onboard industrial-grade main control chip and an embedded NPU computing chip. The NPU chip incorporates a quantized marine-specific fire identification model and communicates with the main control chip via a high-speed board-level bus, enabling fully local edge inference. It can independently complete verification work even when offline, completely different from traditional cloud-based inference and network-dependent verification architectures. The integrated main control unit 103 reserves a MIPI / RJ45 standardized native video interface and supports PoE Ethernet power supply. The anti-salt spray camera is directly connected to the interface via cable, achieving direct hardware-level power supply and data transmission without the need for a third-party gateway, ensuring a stable and reliable connection. The integrated main control unit 103 further includes an edge AI main control module 1031, a false alarm prevention composite model 1032, and a salt spray distortion positioning module 1033.

[0026] The effective fire confirmation and fire point coordinate output unit 105 is a logic and data output module. It receives the AI ​​inference results output by the integrated main control unit 103 and completes the fire determination according to the three-level progressive time sequence verification logic: only when the time sequence stability judgment, smoke and fire association judgment, and multi-class confusion suppression judgment all pass the verification is it determined to be a valid fire; if the verification fails, it is determined to be a false alarm due to environmental interference. When it is determined to be a valid fire, the salt spray distortion positioning module 1033 is called. Based on the pre-calibrated and salt spray distortion compensated homography matrix, the identified fire target pixel coordinates are converted into the plane XY physical coordinates of the offshore booster station. Finally, the fire confirmation status flag and fire point physical coordinate information are output to the outside world. On the one hand, it provides local alarm circuits within the station to trigger local alarms. On the other hand, it is handed over to the dual-link communication unit 104 to upload to the land-based operation and maintenance backend, providing quantitative location basis for remote power outage interlocking and emergency fire fighting.

[0027] The AR visualization marker 106 is an image rendering and overlay module. It receives the physical coordinates of the fire point from the valid fire confirmation and fire point coordinate output unit 105, and combines them with the imaging parameters of the on-site camera and the indoor CAD plan base map data of the substation to complete the coordinate registration and mapping. The fire point location is overlaid and rendered as an AR marker graphic onto the original monitoring video screen and the electronic base map of the station, realizing the visual highlighting of the fire point location. The video stream data with AR marker after overlay can be cached locally and pushed to the land operation and maintenance terminal through the dual-link communication unit 104, which makes it easy for operation and maintenance personnel to intuitively and quickly locate the fire hazard location, solving the shortcomings of traditional alarms that only provide zone prompts and cannot intuitively display the fire point location.

[0028] Before the AI ​​fire verification device for preventing false alarms of this invention is put into operation, it needs to complete preliminary work such as dataset construction, model training and optimization, edge-end quantization deployment, and on-site visual calibration. The dedicated dataset, optimized inference model, and distortion compensation homography matrix obtained from the offline deployment preparation steps S1 to S5 serve as the basis for the algorithms and parameters in steps 201 to 212. Step S1: Construct a multimodal training dataset specifically for offshore booster stations: A targeted dataset of positive and negative samples adapted to the marine interference environment was constructed to eliminate the defects of general model scene adaptation. The dataset includes three categories: positive samples, negative samples, and adversarial synthetic samples. Positive samples cover images of thin smoke and open flames of electrical cabinet overheating and cable smoldering under different lighting and humidity conditions. Negative samples are specially collected images of interference such as sea salt fog, lens condensation, equipment switching arcs, sunlight glare, and dust obstruction. Adversarial fusion samples are generated by GAN to simulate the complex working condition of salt fog covering real smoke, forcing the model to learn subtle feature differences. All samples are labeled with fine target boxes and multi-class labels, including smoke, flame, salt fog interference, lens water droplets, and arc interference.

[0029] Step S2: Construct and optimize the lightweight fire detection basic model: Lightweight YOLOv8n or YOLOv5s are selected as the baseline model to adapt to the low computing power deployment requirements of edge devices. Specific optimizations are made for the characteristics of numerous small targets and strong background interference in early-stage smoke at sea. The CIoU loss function is replaced with the NWD normalized Wasserstein distance loss function adapted for small target detection to improve the accuracy of thin smoke recognition. The CBAM attention mechanism module is embedded to enhance the ability to extract smoke flow texture and flame infrared features, and suppress background noise from static salt spray and water droplets.

[0030] Step S3: Model transfer training, pruning, and edge quantization deployment: Based on a GPU training server, a transfer learning strategy was adopted to load pre-trained weights from a public dataset, freeze the deep parameters of the backbone network, and fine-tune the training for 300 epochs using a dedicated maritime dataset. Cosine annealing learning rate decay and Mosaic data augmentation were used to improve the model's generalization ability. After training, channel pruning was performed on the model to remove invalid and redundant channels, reducing the number of model parameters and computing power consumption. INT8 precision quantization was completed through an edge toolchain to generate a lightweight inference model adapted to the embedded NPU, ensuring that the single-frame inference time is less than 30ms, meeting the requirements of 25fps real-time video processing.

[0031] Step S4: Edge-adaptive image preprocessing and progressive AI verification inference: The quantization model is embedded in the onboard NPU of the device, enabling completely local offline inference without relying on the external network or backend server. When the aspirating detection unit detects that the smoke concentration exceeds the threshold and triggers an alarm, the anti-salt fog camera is activated to collect real-time video streams. Image enhancement is performed through dark channel prior defogging algorithm and guided filtering algorithm to eliminate interference from salt fog haze and lens water droplet blur. A three-level progressive verification logic is used to filter false alarms. 1) Temporal stability determination: Extract 5 consecutive seconds of video frames. If smoke is stably detected in the same area for more than 10 consecutive frames and the average confidence level exceeds 0.7, it is determined as a suspected valid target and transient salt spray clouds are filtered out. 2) Smoke and fire correlation determination: If smoke and flame features that are spatially cross-matched are detected at the same time, the warning is upgraded to an effective fire warning; 3) Multi-class confusion suppression judgment: If the confidence level of smoke detection meets the standard, but the synchronization probability of salt spray interference exceeds 0.3, it is judged as environmental interference ambiguity, and the alarm is actively suppressed to prevent false alarms of salt spray.

[0032] Compared to traditional single-frame judgment and multispectral fusion, this invention adopts a three-level linkage verification mechanism with time-series progression, step-by-step triggering, and closed-loop verification. It has strict execution priority and collaborative constraints: the next level of fireworks association judgment is triggered only after the current level of time-series stability judgment is passed. Finally, the final alarm decision is completed by combining multi-class confusion suppression. The logic at each level filters interference and adds efficiency step by step.

[0033] Quantitative testing showed that the false alarm filtering rate was 82.3% when only timing stability judgment was enabled, and the false alarm filtering rate increased to 95.1% after adding smoke and fire correlation judgment. The three-level mechanism working together can achieve a false alarm filtering rate of over 99.5%. Compared with the industry mainstream technical solution with a false alarm filtering rate of 98%, the anti-interference performance has achieved a significant incremental upgrade.

[0034] Step S5: Precise fire location using monocular vision homography matrix: During the equipment installation phase, offline calibration is completed using a calibration mesh. At least four sets of non-collinear reference points are selected in the monitoring area, and the homography transformation matrix between the image pixel coordinate system and the on-site physical coordinate system is solved. This invention adds a salt spray distortion correction and water droplet blur pixel compensation algorithm adapted to extreme marine conditions. It embeds a transmittance correction mathematical model based on dark channel priors to accurately correct image pixel transmission deviations caused by salt spray haze and local distortion pixels caused by lens beads, solving the defects of traditional homography matrices in positioning offset and accuracy failure in high-interference marine scenarios. After AI identifies a valid fire, it extracts the pixel coordinates of the midpoint of the bottom edge of the flame frame and outputs the actual XY physical coordinates of the booster station in real time through matrix inverse transformation, achieving high-precision positioning within 0.5 meters. AR fire point markers are then superimposed on the video image and CAD base map.

[0035] Further reference Figure 2 The schematic diagram of the AI ​​fire verification method for preventing false alarms of the present invention includes the following steps: Step 201: Inhalation-type smoke concentration over-limit early warning: The very early smoke detection unit 101 of the offshore booster station collects air samples from the cabin in real time and detects the smoke light reduction rate / smoke concentration parameters; when the detected smoke index exceeds the preset first-level early warning threshold (exemplary 0.015dB / m), an over-limit early warning trigger signal is generated. This early warning signal serves as the trigger source for the entire AI review process and is sent to the integrated main control unit via the onboard bus to wake up the subsequent review link; this step only completes the initial detection alarm and does not represent an actual fire. There is a possibility of false triggering due to interference from salt spray, water vapor, etc., which requires subsequent AI multi-level review and adjudication.

[0036] Step 202: Salt spray camera acquires video stream: After receiving the over-limit warning trigger signal from step 201, the integrated main control unit 103 outputs a control signal to start the salt spray camera unit. The salt spray camera unit 102 acquires a real-time high-definition video image sequence of the monitoring area. The raw video data is transmitted to the integrated main control unit through the hardware direct connection interface, and continuous frame images are output for subsequent image preprocessing. Under normal conditions without warning trigger, the salt spray camera unit enters sleep mode to reduce equipment power consumption.

[0037] Step 203: Image Dehazing and Water Droplet Blur Preprocessing: Receive the original video frames output from Step 202, call the locally fixed image enhancement algorithm, use the dark channel prior dehazing algorithm to eliminate image grayscale distortion caused by sea salt fog, and use the guided filtering algorithm to repair local blurred pixels caused by lens beads; normalize and scale the image to generate a standardized image that matches the NPU model input specifications, eliminate image quality degradation caused by the sea environment, reduce environmental noise interference with AI inference results, and send the preprocessed image to the edge inference stage.

[0038] Step 204: Edge-end local offline inference: The preprocessed image from step 203 is sent to the onboard embedded NPU, and the offshore booster station-specific fire identification model quantized by INT8 is loaded. The AI ​​forward inference is completed entirely locally on the device, without relying on the external network or cloud servers. The inference outputs the multi-class detection results of each frame of the image, including any one or combination of target categories (smoke, flame, salt spray interference, lens water droplets, arc interference), target pixel box position, and confidence scores of each category. The inference results are output to the downstream time-series stability determination stage.

[0039] Step 205: Timing Stability Determination: Receive the single-frame AI inference result output from Step 204 and execute the first-level progressive verification logic: extract consecutive video frames within a 5-second time window, verify whether the same image area can stably detect smoke targets for at least 10 consecutive frames, and whether the average confidence level of the smoke is greater than 0.7. If the timing stability condition is not met, it is determined to be short-term interference such as wind-drifting salt fog or instantaneous reflection, and the process is returned to Step 212 to complete the recording of this round of interference events, and return to this step to continue monitoring the video stream. If the timing stability condition is met, the suspected fire target information is output and transferred to Step 206 for smoke and fire association determination. Furthermore, Step 205 and Step 212 are interconnected bidirectionally. When it is determined to be short-term interference, Step 205 jumps to Step 212 to record the event. After Step 212 completes the event recording, it sends the signal back to Step 205 to continue timing monitoring of the video stream.

[0040] Step 206: Smoke and Fire Association Determination: Receive the suspected fire target output from Step 205, and execute the second-level progressive verification logic; compare the pixel spatial intersection-union ratio of the detected smoke target and the flame target to determine whether there are matching smoke and flame features in the same area. If matching smoke and flame features are detected simultaneously, the suspected fire level is increased; if no matching flame features are found, the suspected state of sparse smoldering smoke is retained, and the smoke and fire association determination result is output to Step 207 Multi-class Confusion Suppression Determination.

[0041] Step 207: Multi-class confusion suppression judgment: Receive the smoke and fire association judgment result output in step 206, execute the third-level progressive verification logic, read the confidence levels of salt spray interference, water droplets, and electric arc interference categories output by inference. If the smoke confidence level meets the standard, but the synchronous confidence level of salt spray interference exceeds 0.3, it is judged as environmental ambiguity interference with a high risk of false alarm, and alarm suppression is performed; if there is no high probability of coexistence of interference categories, the suspected fire conclusion is retained. After completing the confusion suppression verification, the judgment result is output to step 208 Valid fire confirmation.

[0042] Step 208: Valid Fire Confirmation: Receive the final verification result of the three-level progressive review output from Step 207, and complete the fire situation determination: If all three levels of review pass, it is determined to be a real and valid fire; if there are any unsatisfactory items in the three-level review, it is determined to be environmental interference, and branch outputs are generated according to the determination result: ① pointing to Step 209 Misjudgment Clarification; ② pointing to Step 211 Fire Point Physical Coordinate Output; ③ pointing to Step 212 Video Image and Base Map Overlay.

[0043] Step 209: False Alarm Attribution Analysis: Receive the ruling result from Step 208 and perform attribution analysis on the current review event. If it is determined to be environmental interference, classify, label, and locally store the interference type (salt spray, lens water droplets, arc flash, etc.), interference occurrence time, image samples, and intermediate confidence data at each level of review to complete the false alarm root cause identification. Furthermore, Step 209 has two output flows: the output data is transmitted to Step 210 for homography matrix coordinate mapping; on the other hand, the output analysis record data is sent back to Step 207 for multi-class confusion suppression determination, used for historical interference information reference in the subsequent reasoning process, and assists the subsequent multi-class confusion suppression link in improving the discrimination capability. Step 209 performs root cause identification and sample recording for false alarm events, and feeds back historical interference information to the upstream determination link to form a closed-loop learning and verification. Even if this is a false alarm, Step 210 is still called to complete the pixel-coordinate conversion for interference location recording, and the interference feature information is sent back to Step 207.

[0044] Step 210: Homography Matrix Coordinate Mapping: Receive the target pixel box data passed from Step 209, call the homography transformation matrix that has been pre-calibrated locally and superimposed with salt spray distortion and water droplet pixel compensation correction, and convert the pixel coordinates of the target box in the image pixel coordinate system into the XY physical coordinates of the booster station site plane through matrix inverse transformation mapping. The coordinate mapping calculation result is sent back to Step 211 for fire point physical coordinate output.

[0045] Step 211: Output of physical coordinates of the fire point: Receive the valid fire situation adjudication command output in step 208, retrieve the physical coordinate values ​​calculated in step 210, and output a standardized fire information message. The message includes the fire status mark, the XY physical coordinates of the fire point, and the event timestamp. One path of the message is sent to the local alarm circuit in the station to trigger a local alarm, and the other path is delivered to the dual-link communication unit 104 to upload to the land-based operation and maintenance backend, providing accurate location information for equipment power-off interlocking and remote emergency fire extinguishing.

[0046] Step 212: Overlaying video footage with the base map: This is specifically handled in two scenarios: ① Receiving the branch signal output from Step 208: Reading the physical coordinates of the fire point, overlaying the AR fire point highlight mark onto the real-time monitoring video footage and the indoor CAD base map of the substation, generating a visualized video stream with AR markers, which can be cached locally and simultaneously pushed to the land-based maintenance terminal; ② Interference false alarm scenario: Receiving short-term interference events from Step 205, overlaying the interference location mark onto the screen, saving the interference event recording and base map snapshot as maintenance logs. Furthermore, after completing the interference event recording in this step, returning to Step 205 to continue timing monitoring of subsequent video frames, waiting for the next round of verification trigger conditions.

[0047] Through the above steps 201-212, the actual operation process of the device of the present invention at the offshore substation site realizes a complete closed loop of verification business from smoke warning triggering to fire / interference event output.

[0048] refer to Figure 3 The flowchart of model training, pruning quantization, and edge deployment of this invention specifically includes training stage 30, compression stage 31, and deployment stage 32. As the first embodiment of this invention, it adopts the preferred AI verification method and device based on YOLOv8n. This embodiment is the optimal implementation of this invention. The core hardware of the device adopts the Rockchip RK3588 main control chip, with a built-in 6TOPS computing power NPU to meet the real-time inference requirements of lightweight models. The camera is directly connected to the motherboard through a native RJ45 interface with locking protection to achieve single-cable PoE power supply and data transmission, which is suitable for marine vibration-resistant and corrosion-resistant installation conditions.

[0049] During the dataset construction phase, more than 15,000 marine-specific samples were collected and organized, including 6,000 positive samples of real fire, covering early fire scenarios such as smoldering cables and overheating of electrical cabinets, and oversampling enhancement for small targets in thin smoke; and 9,000 negative samples such as salt spray, water droplets, and electric arcs. At the same time, adversarial samples of salt spray covering smoke were generated through GAN to enhance the model's anti-interference ability.

[0050] During the model training phase, a YOLOv8n baseline model was used, with deep parameters of the backbone network frozen. The NWD loss function was employed to optimize the regression accuracy of small targets, and a CBAM attention module was embedded to enhance feature extraction. After 300 training epochs, structured channel pruning was used to safely compress the number of parameters to 70% of the original (approximately 2.1M), ensuring that the number of channels in key layers (such as the P2 layer) was not less than 96 to avoid feature degradation of small targets. An additional 50 epochs of fine-tuning were used to compensate for accuracy loss. The RKNN-Toolkit2 toolchain was used for INT8 quantization, along with a dedicated calibration dataset (≥500 images) containing small targets, ensuring that the accuracy loss after quantization was ≤1% (measured mAP@0.5 decrease <0.9%), generating a dedicated inference model adapted to edge NPUs.

[0051] During the on-site operation phase, when the light reduction rate detected by the aspirating detector reaches the first-level warning threshold of 0.015dB / m, the equipment automatically initiates image defogging and water droplet filtering preprocessing, and performs three-level progressive AI verification. Under extreme interference conditions such as high-concentration salt spray, lens beading, and electric arc flash, it can effectively filter out all false features. Only after triple verification of time sequence judgment, smoke and fire correlation, and interference suppression is it confirmed as a valid fire. Physical coordinate positioning is completed through pre-calibrated homography matrix, and video images with AR tags and fire point coordinates are uploaded to the land-based backend.

[0052] As a second embodiment of the present invention, a low-computing-power adaptation implementation based on YOLOv5s is adopted. This embodiment is suitable for low-computing-power edge hardware scenarios. YOLOv5s v6.0 is selected as the baseline model with an input resolution of 640×640. A CBAM attention module is embedded between the SPPF module and the detection head. Focal Loss is used to balance the weights of positive and negative samples to improve the recognition capability of low-probability smoke targets. The dataset construction, preprocessing algorithm, progressive verification logic, and localization algorithm are completely consistent with the first embodiment. After the model is pruned and quantized with INT8, it can be deployed on the RV1126 low-computing-power chip. Under 2.0 TOPS computing power, it can still achieve single-frame inference within 30ms, meeting the requirements of low-cost device batch deployment.

[0053] During installation and commissioning, the equipment of this invention lays a standard calibration grid in the protected areas such as the GIS room and cable interlayer of the substation. Four non-collinear feature points are selected to complete the pairing of pixel coordinates and physical coordinates. The optimal homography transformation matrix is ​​solved by combining salt spray distortion and water droplet fuzziness compensation algorithms and then stored locally. After fire identification, the effective pixel coordinates of the flame are automatically extracted, and the planar physical coordinates are output in real time through matrix inverse transformation. The coordinates are dynamically superimposed on the monitoring video and the electronic map of the station. The positioning accuracy is stably controlled within 0.5 meters and supports pan-tilt-zoom linkage to track fire points.

[0054] In actual operation verification, this invention was continuously tested on a real offshore substation for 6 months, and multiple sets of extreme operating condition comparative tests were completed. The test environment covered extreme marine scenarios such as high concentration salt spray, large-area lens beading, and equipment arc flash interference. The test results show that: the traditional solution completely fails under extreme superimposed interference conditions, triggering 100% false alarms, and the false alarm rate of the conventional optimized model exceeds 35%; the present invention can still maintain a false alarm filtering rate of 99.2% under extreme operating conditions, with no invalid false alarms and no missed detections. The overall measured indicators are: the average number of false alarms per day is reduced from more than 15 times in the traditional solution to ≤0.1 times per day, and the false alarm filtering rate reaches more than 99.5%; the early smoldering fire identification delay is ≤3 seconds, and the fire location accuracy is ≤0.5 meters. The present invention is accurately adapted to the operation requirements of offshore substations for unmanned operation, long-term maintenance-free operation, and high reliability disaster prevention, and has extremely high engineering practical value and promotion value.

[0055] refer to Figure 4This invention provides a schematic diagram of the fire spatial positioning principle based on homography transformation. The specific working principle is as follows: The fire target pixel coordinates (u, v) output from the image pixel coordinate system 40 are output to the homography transformation matrix 41. The homography transformation matrix 41 is a pre-calibrated coordinate transformation operation unit. Further, based on the planar homography transformation mathematical model, it completes the basic mapping operation from pixel coordinates to on-site physical coordinates. The output of the homography transformation matrix 41 is divided into two paths: one path, the basic mapping result, is sent downwards to the salt spray distortion compensation module 43 for environmental distortion compensation; the other path, the uncompensated basic mapping result, is directly output to the on-site physical coordinate system 42 as a reference coordinate. This homography transformation matrix 41 is obtained from multi-point calibration on-site before equipment commissioning, establishing a basic projection mapping relationship between the image plane and the booster station monitoring plane, but does not account for imaging offset errors caused by sea salt spray and lens water droplets.

[0056] The salt fog distortion compensation module 43 is a first-level environmental error compensation unit. It receives the intermediate coordinate results after preliminary homography transformation output by the homography transformation matrix 41, retrieves the salt fog distortion deviation correction parameters obtained from the previous calibration, and compensates for the imaging position offset error caused by sea salt fog scattering. Salt fog will cause the scene outline to blur and the detection target pixel to shift. The salt fog distortion compensation module 43 corrects the system deviation of the system. After correction, the intermediate coordinate results are sent to the water droplet blur pixel compensation module 44. The water droplet blur pixel compensation module 44 is a second-level environmental error compensation unit. It receives two inputs: one is the original pixel coordinates (u,v) directly output from the image pixel coordinate system 40, and the other is the intermediate coordinate data after salt fog compensation from the salt fog distortion compensation module 43. In response to the problem of local image blurring and detection box pixel drift caused by water droplets attached to the lens surface, the water droplet blur offset correction parameters are called to perform secondary correction on the coordinates to eliminate the positioning deviation caused by water droplets. After compensation, the final coordinate data after salt fog + water droplet dual distortion correction is output and sent to the on-site physical coordinate system 42.

[0057] The on-site physical coordinate system 42 serves as the coordinate output and AR rendering unit, receiving dual signal inputs: one input receives the undistorted reference mapping coordinates from the homography transformation matrix 41, and the other input receives the corrected coordinates after two-stage distortion compensation from the water droplet blur pixel compensation module 44. Internally, the reference coordinates output by the homography transformation matrix 41 are used for reference verification, and the corrected coordinates output by the water droplet blur pixel compensation module 44 are used as the valid output value to output the fire point physical coordinates (x, y), which are the actual physical locations of the offshore booster station under the monitoring plane. Based on the fire point physical coordinates (x, y), AR fire point representation is further completed, generating AR marker information that can be superimposed on the monitoring video and electronic base map. The output fire point physical coordinates are supplied to the fire point physical coordinate output in step 211, and the AR fire point representation results are supplied to the video screen and base map overlay module in step 212.

[0058] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made within the concept and principles of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A false alarm prevention AI fire verification device, characterized in that, It includes an early smoke detection unit (101), a salt spray camera unit (102), an integrated main control unit (103), a dual-link communication unit (104), an effective fire confirmation and fire point coordinate output unit (105), and an AR visualization marker (106), wherein: The very early smoke detection unit (101) has a built-in adjustable speed air pump, anti-salt spray filter, LED detection chamber and airflow sensor. It dynamically adjusts the air intake flow according to the environmental load, and is equipped with a backflush air path module to realize automatic self-cleaning of the pipeline network. It also transmits the signal to the integrated main control unit (103). The integrated main control unit (103) further includes an edge AI main control module (1031), a false alarm prevention composite model (1032) and a salt spray distortion positioning module (1033). The anti-salt spray camera unit (102) is directly connected to the integrated main control unit (103) via an RJ45 interface. When the very early smoke detection unit (101) outputs a smoke exceeding the limit warning signal, it collects real-time high-definition video stream images of the monitoring cabin. The dual-link communication unit (104) exchanges signals bidirectionally with the integrated main control unit (103); The effective fire confirmation and fire point coordinate output unit (105) receives the AI ​​reasoning result output by the integrated main control unit (103) and completes the fire determination according to the three-level progressive timing review logic; The AR visualization identifier (106) receives the physical coordinates of the fire point output by the valid fire confirmation and fire point coordinate output unit (105) and the signal from the integrated main control unit (103), and completes the coordinate registration and mapping by combining the imaging parameters of the on-site camera and the indoor CAD plan base map data of the booster station.

2. A method for preventing false alarms in AI fire verification, characterized in that, Includes the following steps: Step 201: Inhalation-type smoke concentration exceeding the limit warning. When the detected smoke index exceeds the preset first-level warning threshold, an exceeding the limit warning trigger signal is generated. Step 202: The anti-salt spray camera acquires video streams, and the anti-salt spray camera unit acquires real-time high-definition video image sequences of the monitoring area and outputs continuous frame images; Step 203: Image dehazing and water droplet blur preprocessing. Receive the original video frames output in step 202. Use the dark channel prior dehazing algorithm to eliminate image grayscale distortion caused by sea salt fog. Use the guided filtering algorithm to repair local blurred pixels caused by lens beads. Normalize and scale the image to generate a standardized image that matches the NPU model input specifications. Step 204: Local offline inference at the edge: The standardized image preprocessed in step 203 is sent to the onboard embedded NPU, and the offshore substation-specific fire identification model quantized by INT8 is loaded. The inference outputs the multi-class detection results of each frame of the image, including the target category, the target pixel box position, and the confidence score of each category. Step 205: Timing stability determination, receive the multi-class detection result of the single frame image output in step 204, and execute the first-level progressive verification logic; Step 206: Fireworks association determination, receive the timing stability determination result output in step 205, and execute the second-level progressive verification logic; Step 207: Multi-class confusion suppression determination, receive the fireworks association determination result output in step 206, and execute the third-level progressive review logic; Step 208: Valid fire confirmation, receive the multi-class confusion suppression judgment result output from step 207, output the final verification result of three-level progressive review, and complete the fire situation determination; Step 209: Misjudgment identification, receive the ruling result from step 208, and perform attribution identification analysis on the reviewed event; Step 210: Homography matrix coordinate mapping. Receive the target pixel box data transmitted from step 209, and convert the pixel coordinates of the target box points in the image pixel coordinate system into the physical coordinates of the substation site plane through inverse matrix transformation. The coordinate mapping calculation result is transmitted to step 211. Step 211: Output the physical coordinates of the fire point, receive the valid fire situation adjudication command output in step 208, retrieve the physical coordinate values ​​calculated in step 210, and output the standardized fire information message. Step 212: Overlay the video image with the base map, receive the branch signal output from Step 208, read the output physical coordinates of the fire point, overlay and render the AR fire point highlight mark onto the real-time monitoring video image and the indoor CAD base map of the substation, generate a visualized video stream with AR markers, and push it to the onshore maintenance terminal; at the same time, receive the short-term interference event from Step 205, overlay the interference location mark onto the image, save the interference event recording and base map snapshot as maintenance records, and after completing the interference event recording, return to Step 205 to continue performing timing monitoring on subsequent video frames.

3. The AI ​​fire verification method for preventing false alarms as described in claim 2, characterized in that, The following offline deployment preparation steps are included before step 201: Step S1: Construct a multimodal training dataset specifically for offshore booster stations; Step S2: Construct and optimize a lightweight fire detection basic model; Step S3: Model transfer training, pruning, and edge quantization deployment; Step S4: Edge-adaptive image preprocessing and progressive AI verification inference; Step S5: Precise fire location using monocular vision homography matrix.

4. The AI ​​fire verification method for preventing false alarms as described in claim 2, characterized in that, The preset first-level warning threshold is 0.015dB / m.

5. The AI ​​fire verification method for preventing false alarms as described in claim 2, characterized in that, The output target categories include any one or a combination of smoke, flame, salt spray interference, lens water droplets, and electric arc interference.

Images