Wind power cabin fire early warning device and method based on multispectral fusion

Abstract

The invention discloses a wind power cabin fire early warning device and a wind power cabin fire early warning method based on multispectral fusion, and aims at the defects of high false alarm rate, insufficient motion artifact compensation and transient fire capture lag of single-spectrum detection in the prior art to construct a multispectral collaborative analysis architecture. A thermal deformation dynamic compensation model is constructed in combination with a main shaft rotating speed signal, and sub-pixel-level motion artifact elimination is achieved; an improved three-dimensional convolutional neural network is adopted to extract smoke diffusion trajectory, temperature conduction trend and arc pulse features, and multi-source feature adaptive fusion is realized through a cross-modal attention mechanism and a graph neural network; and establishing an oil stain reflection spectrum fingerprint database for false alarm filtering, and finally triggering third-level early warning according to an environment self-adaptive threshold value. According to the method, the electrical short circuit detection time can be greatly shortened, the false alarm rate is effectively reduced, the smoldering fire is warned in advance, and the industrial problem of accurate recognition of early fire in a complex cabin environment is solved.

Description

Technical Field

[0001] This invention belongs to the field of wind power equipment safety monitoring technology, and in particular relates to an early warning device, apparatus and electronic equipment for wind turbine nacelle fire based on multispectral fusion. Background Technology

[0002] There are three coupled fire hazards inside the wind turbine nacelle: electrical fires: short circuits in converters and cable joints generate electric arcs, mainly characterized by ultraviolet light; mechanical fires: overheating of gearboxes / bearings ignites lubricating oil, mainly characterized by infrared light; and composite material combustion: smoldering of epoxy resin shells when heated, mainly characterized by visible light smoke. In the complex nacelle environment of strong vibration, oil pollution, and electromagnetic interference, how to achieve high-precision early fire identification and rapid response (with delay) to avoid missed or false alarms caused by environmental interference from traditional single sensors is a technical problem that urgently needs to be solved in this field.

[0003] Existing technologies generally employ single-spectrum detection, dual-spectrum fusion, or motion compensation methods for processing. Single-spectrum detection typically uses infrared thermal imaging threshold alarms, while dual-spectrum fusion generally uses a weighted average decision method combining visible light and infrared light, followed by motion compensation to eliminate artifacts. However, existing technologies have the following drawbacks: 1. Feature fusion is relatively coarse. Existing technologies only perform feature fusion by adding weights, which cannot adapt to environmental changes. In high-temperature environments, excessive infrared weights lead to false alarms due to oil stain reflections, while insufficient ultraviolet weights under strong electromagnetic interference result in missed arc detections. 2. Motion artifacts are a serious problem. Traditional background subtraction methods ignore the problem of thermal deformation offset, resulting in large residual errors after compensation, which leads to false alarms. 3. Temporal correlation is missing. Existing single-frame analysis cannot capture the evolution of fire, such as the direction of smoke diffusion or temperature conduction trends, leading to misjudgment of smoke in the early stage of smoldering due to airflow disturbances and the failure to effectively capture transient features of electrical sparks. Therefore, it is necessary to design an early warning device and method for wind turbine nacelle fires based on multispectral fusion to solve the above problems. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an early warning device and method for wind turbine nacelle fire based on multispectral fusion. It aims to solve the problems of the prior art by constructing a multispectral-temporal deep analysis architecture, developing a physical-data dual-driven compensation model, and establishing an oil pollution spectral fingerprint database.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: An early warning device for wind turbine nacelle fires based on multispectral fusion, comprising: The multispectral sensing module is deployed in a key area inside the wind turbine nacelle and includes at least one visible light sensor, one infrared thermal imaging sensor and one ultraviolet flame sensor. The data fusion processing unit is connected to the multispectral sensing module and is configured to perform spatiotemporal registration and feature-level fusion of multi-source spectral data. The fire feature analysis module includes: The multispectral temporal analysis submodule based on the improved 3D-CNN takes a registered visible-infrared-ultraviolet image sequence as input. The dynamic background modeling submodule combines the fan spindle speed signal to eliminate motion artifacts of rotating parts. The false alarm suppression submodule distinguishes false fire sources by comparing them with an oil spill reflectance spectrum database. The graded early warning module outputs three levels of alarm signals based on the confidence level of the fire.

[0006] Preferably, the key area inside the wind turbine nacelle where the multispectral sensing module is deployed in step S1 is specifically: The spectral sensing module adopts a distributed layout, including: The sensor group in the spindle bearing area focuses on the connection between the gearbox and the generator. The hydraulic system area sensor group covers the pitch system and braking unit; Sensor array in electrical cabinet area, monitoring converters and cable joints.

[0007] Preferably, the improved 3D-CNN multispectral temporal analysis submodule includes: Parallel branch feature extractors process visible light texture features, infrared temperature distribution features, and ultraviolet radiation intensity features separately. A cross-modal attention fusion layer generates fused feature maps through a spatial-channel dual attention mechanism; The timing anomaly detection head outputs a fire probability change curve for N consecutive frames.

[0008] Preferably, the dynamic background modeling submodule performs: Real-time spindle speed ω is obtained based on the fan SCADA system; Construct the equations of motion for the rotating component: ； In the formula, x and y represent the horizontal and vertical coordinates, respectively; t represents the image acquisition timestamp; A represents the amplitude of the rotating component in the x-axis direction, and B represents the amplitude of the rotating component in the y-axis direction. Indicates the real-time angular velocity of the spindle. Indicates the initial phase angle of the x-axis motion. Indicates the initial phase angle of the y-axis motion; Periodic motion artifacts in image sequences are eliminated by using a motion compensation matrix.

[0009] Preferably, the false alarm suppression submodule includes: An oil spill reflectance spectral feature library stores reflectance curves of different oils in predetermined wavelength bands; The real-time spectral matching unit calculates the cosine similarity between the current UV-Vis joint spectrum and the feature library; False alarm filtering is triggered when the similarity exceeds a threshold.

[0010] Preferably, the early warning method for a wind turbine nacelle fire early warning device based on multispectral fusion includes the following steps: S1 synchronously collects visible light, infrared and ultraviolet spectral data streams inside the cabin through a multispectral sensing module; S2 is executed in the data fusion processing unit: S201, Motion artifact elimination: Based on the real-time angular velocity ω of the principal axis, the multispectral image sequence is aligned by the motion compensation matrix; S202, Spatiotemporal Registration: Map visible light, infrared, and ultraviolet data to a unified coordinate system; S3 performs multispectral time-series analysis through the fire feature analysis module: S301, processed using an improved 3D-CNN: Input the registered multispectral image sequence into the parallel branch feature extractor; A joint feature map is generated through a cross-modal attention fusion layer; S302, Dynamic Background Modeling: Based on the compensated image sequence, construct a background template for the rotating component; Foreground anomaly regions are extracted by background subtraction; S303, False Alarm Suppression Handling: Compare the spectra of the abnormal areas with the oil spill reflectance spectrum database; S4, in the graded early warning module: S401, calculate the fire confidence level based on the joint feature map and false alarm filtering results; S402, when the confidence level exceeds the dynamic threshold, a graded early warning signal is triggered.

[0011] Preferably, the specific method for motion artifact elimination in step S201 is as follows: Motion offset calculation: The instantaneous offset of the rotating component is calculated based on the motion equation described in claim 3: ; Injection of heat deformation compensation: ; In the formula, Represents the coefficient of thermal deformation along the x-axis. Indicates the thermal deformation coefficient along the y-axis; This represents the temperature ratio, where T is taken from the current cabin temperature and the calibration reference temperature, respectively. Motion compensation matrix construction: ; In the formula, For dynamic rotation compensation angle, Δ represents the time interval between consecutive frames. ; Image geometric correction: For each pixel in the k-th frame image Perform an affine transformation: ; The transformed system must meet the following constraints:

[0012] In the formula, These are the corrected pixel coordinates.

[0013] Preferably, the improved 3D-CNN processing in step S301 specifically includes: Parallel feature extraction, including: Visible light branch: Extracting spatiotemporal diffusion features of smoke using 3D convolution kernels: ; Infrared branch: Employing dilated convolution to capture thermal conduction features. ; Ultraviolet branch: Identifying radiation abrupt changes using a pulse detection layer. b); Cross-modal attention fusion includes: Calculate the correlation weights between visible light and infrared features: ; Weighted fusion UV characteristics: ; In the formula, This represents the registered visible light image sequence, where Cv represents the number of visible light channels. This represents the 3D convolution kernel parameters, used to extract the spatiotemporal features of smoke; This represents the registered infrared image sequence; This represents the parameters of the hollow convolution kernel, used to expand the receptive field to capture heat conduction; This represents the registered ultraviolet image sequence. PulseDetect indicates the pulse detection operation, specifically: ; In the formula, Δt represents the pulse detection time window. Indicates the current time point; This represents the similarity matrix between visible light and infrared features; This represents the Sigmoid activation function. This represents the intermodal attention weight matrix; Represents tensor multiplication. Indicates feature splicing, This represents feature map addition under a unified dimension.

[0014] Preferably, the linkage between steps S302 and S303 includes: The foreground mask M(x,y,t) output by dynamic background modeling is used as the input region for false alarm suppression. Spectral database alignment is performed only for pixel regions where M(x,y,t)=1; Logic for determining oil contamination: ; In the formula, This represents the combined ultraviolet-visible light spectrum of the current pixel. The CosineSimilarity is calculated as follows, representing the sample spectra in the oil spill reflectance spectral database: ; The threshold of 0.85 is the optimal classification threshold determined by the ROC curve.

[0015] Preferably, the fire confidence level calculation in step S401 adopts the following method:

[0016] Where GAP represents the importance of global average pooling output 3D-CNN features; M' represents the foreground mask after false positive suppression; and λ and μ represent the weight coefficients trained from historical fire data, respectively.

[0017] Furthermore, the overall architecture of the improved 3D-CNN network includes an input layer, a parallel branch structure, a cross-modal attention fusion layer, a temporal analysis layer, and an output layer.

[0018] The parallel branching structure includes visible light branch, infrared branch and ultraviolet branch. The operation sequences are Conv3D(5×5×3)-BN-ReLU → MaxPool3D, DilatedConv3D(r=2)-LayerNorm-GELU and PulseDetect → 1×1×1 Conv3D, respectively, which are used to extract smoke motion trajectory, capture heat conduction spatial pattern and enhance radiation pulse features.

[0019] The beneficial effects of this invention are as follows: 1. This invention solves the problem of coarse feature fusion and achieves multimodal adaptive decision-making. Addressing the high false alarm rate of existing fixed-weight fusion methods under complex conditions, this invention dynamically adjusts the weight allocation of different spectral features through a cross-modal attention mechanism and graph neural network fusion technology. In high-temperature environments, it automatically reduces infrared weights to suppress false alarms due to oil reflection, and increases ultraviolet weights to enhance arc detection capabilities under strong electromagnetic interference.

[0020] 2. This invention addresses the problem of residual errors caused by neglecting thermal deformation in traditional motion compensation methods. It innovatively introduces a thermal deformation coefficient and a temperature scaling factor to construct a collaborative compensation model for mechanical vibration and thermal expansion. Vibration amplitude parameters are obtained through laser calibration, and the offset is dynamically corrected using real-time temperature data, ultimately generating a high-precision affine transformation matrix.

[0021] 3. To address the problem that single-frame analysis cannot capture the evolution pattern of fires, this invention employs an improved 3D-CNN to extract smoke diffusion trajectories, heat conduction trends, and arc pulse characteristics, combined with a GRU time-series analysis unit to establish a fire evolution model. This achieves three major breakthroughs: early warning is triggered by determining the angle between the smoke diffusion direction and the airflow field during the early smoldering stage; transient arcs are captured with extremely high accuracy through a 100ms pulse detection window; and the GRU unit continuously tracks the temperature conduction trend for 5 minutes, completely solving the problems of missed detection of slowly changing fires and delayed capture of transient fires. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the device logic block diagram in an embodiment of the present invention; Figure 2 This is a flowchart of the system of the present invention; Figure 3 This is a schematic diagram comparing the effects of the method in this embodiment of the invention with existing technical methods. Detailed Implementation

[0023] Example 1: like Figure 1 As shown, an early warning device for wind turbine nacelle fires based on multispectral fusion includes: The multispectral sensing module is deployed in a key area inside the wind turbine nacelle and includes at least one visible light sensor, one infrared thermal imaging sensor and one ultraviolet flame sensor. The data fusion processing unit is connected to the multispectral sensing module and is configured to perform spatiotemporal registration and feature-level fusion of multi-source spectral data. The fire feature analysis module includes: The multispectral temporal analysis submodule based on the improved 3D-CNN takes a registered visible-infrared-ultraviolet image sequence as input. The dynamic background modeling submodule combines the fan spindle speed signal to eliminate motion artifacts of rotating parts. The false alarm suppression submodule distinguishes false fire sources by comparing them with an oil spill reflectance spectrum database. The graded early warning module outputs three levels of alarm signals based on the confidence level of the fire.

[0024] Preferably, the key area inside the wind turbine nacelle where the multispectral sensing module is deployed in step S1 is specifically: The spectral sensing module adopts a distributed layout, including: The sensor group in the spindle bearing area focuses on the connection between the gearbox and the generator. The hydraulic system area sensor group covers the pitch system and braking unit; Sensor array in electrical cabinet area, monitoring converters and cable joints.

[0025] Preferably, the improved 3D-CNN multispectral temporal analysis submodule includes: Parallel branch feature extractors process visible light texture features, infrared temperature distribution features, and ultraviolet radiation intensity features separately. A cross-modal attention fusion layer generates fused feature maps through a spatial-channel dual attention mechanism; The timing anomaly detection head outputs a fire probability change curve for N consecutive frames.

[0026] Preferably, the dynamic background modeling submodule performs: Real-time spindle speed ω is obtained based on the fan SCADA system; Construct the equations of motion for the rotating component: ； In the formula, x and y represent the horizontal and vertical coordinates, respectively; t represents the image acquisition timestamp; A represents the amplitude of the rotating component in the x-axis direction, and B represents the amplitude of the rotating component in the y-axis direction. Indicates the real-time angular velocity of the spindle. Indicates the initial phase angle of the x-axis motion. Indicates the initial phase angle of the y-axis motion; Periodic motion artifacts in image sequences are eliminated by using a motion compensation matrix.

[0027] Preferably, the false alarm suppression submodule includes: An oil spill reflectance spectral feature library stores reflectance curves of different oils in predetermined wavelength bands; The real-time spectral matching unit calculates the cosine similarity between the current UV-Vis joint spectrum and the feature library; False alarm filtering is triggered when the similarity exceeds a threshold.

[0028] like Figure 2 As shown, an early warning method for a wind turbine nacelle fire early warning device based on multispectral fusion includes the following steps: S1 synchronously collects visible light, infrared and ultraviolet spectral data streams inside the cabin through a multispectral sensing module; S2 is executed in the data fusion processing unit: S201, Motion artifact elimination: Based on the real-time angular velocity ω of the principal axis, the multispectral image sequence is aligned by the motion compensation matrix; S202, Spatiotemporal Registration: Map visible light, infrared, and ultraviolet data to a unified coordinate system; S3 performs multispectral time-series analysis through the fire feature analysis module: S301, processed using an improved 3D-CNN: Input the registered multispectral image sequence into the parallel branch feature extractor; A joint feature map is generated through a cross-modal attention fusion layer; S302, Dynamic Background Modeling: Based on the compensated image sequence, construct a background template for the rotating component; Foreground anomaly regions are extracted by background subtraction; S303, False Alarm Suppression Handling: Compare the spectra of the abnormal areas with the oil spill reflectance spectrum database; S4, in the graded early warning module: S401, calculate the fire confidence level based on the joint feature map and false alarm filtering results; S402, when the confidence level exceeds the dynamic threshold, a graded early warning signal is triggered.

[0029] Preferably, the specific method for motion artifact elimination in step S201 is as follows: Motion offset calculation: The instantaneous offset of the rotating component is calculated based on the motion equation described in claim 3: ; Injection of heat deformation compensation: ; In the formula, Represents the coefficient of thermal deformation along the x-axis. Indicates the thermal deformation coefficient along the y-axis; This represents the temperature ratio, where T is taken from the current cabin temperature and the calibration reference temperature, respectively. Motion compensation matrix construction: ; In the formula, For dynamic rotation compensation angle, Δ represents the time interval between consecutive frames. ; Image geometric correction: For each pixel in the k-th frame image Perform an affine transformation: ; The transformed system must meet the following constraints:

[0030] In the formula, These are the corrected pixel coordinates.

[0031] Preferably, the improved 3D-CNN processing in step S301 specifically includes: Parallel feature extraction, including: Visible light branch: Extracting spatiotemporal diffusion features of smoke using 3D convolution kernels: ; Infrared branch: Employing dilated convolution to capture thermal conduction features. ; Ultraviolet branch: Identifying radiation abrupt changes using a pulse detection layer. b); Cross-modal attention fusion includes: Calculate the correlation weights between visible light and infrared features: ; Weighted fusion UV characteristics: ; In the formula, This represents the registered visible light image sequence, where Cv represents the number of visible light channels. This represents the 3D convolution kernel parameters, used to extract the spatiotemporal features of smoke; This represents the registered infrared image sequence; This represents the parameters of the hollow convolution kernel, used to expand the receptive field to capture heat conduction; This represents the registered ultraviolet image sequence. PulseDetect indicates the pulse detection operation, specifically: ; In the formula, Δt represents the pulse detection time window. Indicates the current time point; This represents the similarity matrix between visible light and infrared features; This represents the Sigmoid activation function. This represents the intermodal attention weight matrix; Represents tensor multiplication. Indicates feature splicing, This represents feature map addition under a unified dimension.

[0032] Preferably, the linkage between steps S302 and S303 includes: The foreground mask M(x,y,t) output by dynamic background modeling is used as the input region for false alarm suppression. Spectral database alignment is performed only for pixel regions where M(x,y,t)=1; Logic for determining oil contamination: ; In the formula, This represents the combined ultraviolet-visible light spectrum of the current pixel. The CosineSimilarity is calculated as follows, representing the sample spectra in the oil spill reflectance spectral database: ; The threshold of 0.85 is the optimal classification threshold determined by the ROC curve.

[0033] Preferably, the fire confidence level calculation in step S401 adopts the following method:

[0034] Where GAP represents the importance of global average pooling output 3D-CNN features; M' represents the foreground mask after false positive suppression; and λ and μ represent the weight coefficients trained from historical fire data, respectively.

[0035] Furthermore, the overall architecture of the improved 3D-CNN network includes an input layer, a parallel branch structure, a cross-modal attention fusion layer, a temporal analysis layer, and an output layer.

[0036] The parallel branching structure includes visible light branch, infrared branch and ultraviolet branch. The operation sequences are Conv3D(5×5×3)-BN-ReLU → MaxPool3D, DilatedConv3D(r=2)-LayerNorm-GELU and PulseDetect → 1×1×1 Conv3D, respectively, which are used to extract smoke motion trajectory, capture heat conduction spatial pattern and enhance radiation pulse features.

[0037] Example 2: This embodiment provides an application of a multispectral fusion-based early warning device and method for wind turbine nacelle fires in a specific wind turbine unit. The design provided by this patent was used in a wind farm deployed in an autonomous prefecture in the southwestern region of a central province. The specific process and effects are as follows: Test environment: Wind turbine: 2.5MW double-fed induction generator, nacelle temperature -25℃~65℃; test duration 6 months, covering high temperature, high humidity and strong electromagnetic interference conditions.

[0038] Comparison of options: Existing Solution 1: Single infrared threshold alarm; Existing Solution 2: Visible light + infrared weighted fusion; The present invention provides a multispectral fusion early warning system.

[0039] The performance comparison is shown in Table 1 below: Table 1: Performance indicators compared with existing technologies;

[0040] From Table 1 above and Figure 3 It is evident that this solution achieved breakthroughs in response speed during actual testing on 30 wind turbine units in a wind farm: electrical short-circuit detection took only 0.15 seconds, compared to >0.9 seconds for traditional solutions. This is due to the synergistic effect of ultraviolet pulse detection and 3D-CNN temporal analysis, resulting in a speedup of 5 times compared to national standards. Reliability also saw a significant improvement: the false alarm rate was 0.1 times / thousand-hour, far less than the industry average of >3 times, thanks to the cross-modal attention mechanism that accurately suppresses oil stain reflection. Economic efficiency was also remarkable: maintenance costs were reduced to 12,000 yuan / unit / year, a decrease of 85%, due to a 96.8% elimination rate of motion artifacts, reducing unnecessary downtime. The key innovation lies in the multispectral spatiotemporal fusion architecture, which solves the three major pain points of traditional solutions: poor environmental adaptability, lack of thermal deformation compensation, and lag in transient capture, achieving a technological leap from "post-event alarm" to "early prevention" in fire early warning.

Claims

1. An early warning device for wind turbine nacelle fires based on multispectral fusion, characterized in that, include: The multispectral sensing module is deployed in a key area inside the wind turbine nacelle and includes at least one visible light sensor, one infrared thermal imaging sensor and one ultraviolet flame sensor. The data fusion processing unit is connected to the multispectral sensing module and is configured to perform spatiotemporal registration and feature-level fusion of multi-source spectral data. The fire feature analysis module includes: The multispectral temporal analysis submodule based on the improved 3D-CNN takes a registered visible-infrared-ultraviolet image sequence as input. The dynamic background modeling submodule combines the fan spindle speed signal to eliminate motion artifacts of rotating parts. The false alarm suppression submodule distinguishes false fire sources by comparing them with an oil spill reflectance spectrum database. The graded early warning module outputs three levels of alarm signals based on the confidence level of the fire.

2. The early warning device for wind turbine nacelle fire based on multispectral fusion according to claim 1, characterized in that, The key areas inside the wind turbine nacelle where the multispectral sensing module is deployed in step S1 are specifically: The spectral sensing module adopts a distributed layout, including: The sensor group in the spindle bearing area focuses on the connection between the gearbox and the generator. The hydraulic system area sensor group covers the pitch system and braking unit; Sensor array in electrical cabinet area, monitoring converters and cable joints.

3. The early warning device for wind turbine nacelle fire based on multispectral fusion according to claim 1, characterized in that, The improved 3D-CNN multispectral temporal analysis submodule includes: Parallel branch feature extractors process visible light texture features, infrared temperature distribution features, and ultraviolet radiation intensity features separately. A cross-modal attention fusion layer generates fused feature maps through a spatial-channel dual attention mechanism; The timing anomaly detection head outputs a fire probability change curve for N consecutive frames.

4. The early warning device for wind turbine nacelle fire based on multispectral fusion according to claim 1, characterized in that, The dynamic background modeling submodule executes: Real-time spindle speed ω is obtained based on the fan SCADA system; Construct the equations of motion for the rotating component: ； In the formula, x and y represent the horizontal and vertical coordinates, respectively; t represents the image acquisition timestamp; A represents the amplitude of the rotating component in the x-axis direction, and B represents the amplitude of the rotating component in the y-axis direction. Indicates the real-time angular velocity of the spindle. Indicates the initial phase angle of the x-axis motion. Indicates the initial phase angle of the y-axis motion; Periodic motion artifacts in image sequences are eliminated by using a motion compensation matrix.

5. The early warning device for wind turbine nacelle fire based on multispectral fusion according to claim 1, characterized in that, The false alarm suppression submodule includes: An oil spill reflectance spectral feature library stores reflectance curves of different oils in predetermined wavelength bands; The real-time spectral matching unit calculates the cosine similarity between the current UV-Vis joint spectrum and the feature library; False alarm filtering is triggered when the similarity exceeds a threshold.

6. A method for early warning of wind turbine nacelle fire based on multispectral fusion according to any one of claims 1-5, characterized in that, Includes the following steps: S1 synchronously collects visible light, infrared and ultraviolet spectral data streams inside the cabin through a multispectral sensing module; S2 is executed in the data fusion processing unit: S201, Motion artifact elimination: Based on the real-time angular velocity ω of the principal axis, the multispectral image sequence is aligned by the motion compensation matrix; S202, Spatiotemporal Registration: Map visible light, infrared, and ultraviolet data to a unified coordinate system; S3 performs multispectral time-series analysis through the fire feature analysis module: S301, processed using an improved 3D-CNN: Input the registered multispectral image sequence into the parallel branch feature extractor; A joint feature map is generated through a cross-modal attention fusion layer; S302, Dynamic Background Modeling: Based on the compensated image sequence, construct a background template for the rotating component; Foreground anomaly regions are extracted by background subtraction; S303, False Alarm Suppression Handling: Compare the spectra of the abnormal areas with the oil spill reflectance spectrum database; S4, in the graded early warning module: S401, calculate the fire confidence level based on the joint feature map and false alarm filtering results; S402, when the confidence level exceeds the dynamic threshold, a graded early warning signal is triggered.

7. The early warning method for a wind turbine nacelle fire early warning device based on multispectral fusion according to claim 6, characterized in that, The specific method for motion artifact elimination in step S201 is as follows: Motion offset calculation: The instantaneous offset of the rotating component is calculated based on the motion equation described in claim 3: ； Injection of heat deformation compensation: ； In the formula, Represents the coefficient of thermal deformation along the x-axis. Indicates the thermal deformation coefficient along the y-axis; This represents the temperature ratio, where T is taken from the current cabin temperature and the calibration reference temperature, respectively. Motion compensation matrix construction: ； In the formula, For dynamic rotation compensation angle, Δ represents the time interval between consecutive frames. ; Image geometric correction: For each pixel in the k-th frame image Perform an affine transformation: ； The transformed system must meet the following constraints: In the formula, These are the corrected pixel coordinates.

8. The early warning method for a wind turbine nacelle fire early warning device based on multispectral fusion according to claim 7, characterized in that, The improved 3D-CNN processing in step S301 specifically includes: Parallel feature extraction, including: Visible light branch: Extracting spatiotemporal diffusion features of smoke using 3D convolution kernels: ； Infrared branch: Employing dilated convolution to capture thermal conduction features. ； Ultraviolet branch: Identifying radiation abrupt changes using a pulse detection layer. b)； Cross-modal attention fusion includes: Calculate the correlation weights between visible light and infrared features: ； Weighted fusion UV characteristics: ； In the formula, This represents the registered visible light image sequence, where Cv represents the number of visible light channels. This represents the 3D convolution kernel parameters, used to extract the spatiotemporal features of smoke; This represents the registered infrared image sequence; This represents the parameters of the hollow convolution kernel, used to expand the receptive field to capture heat conduction; This represents the registered ultraviolet image sequence. PulseDetect indicates the pulse detection operation, specifically: ； In the formula, Δt represents the pulse detection time window. Indicates the current time point; This represents the similarity matrix between visible light and infrared features; This represents the Sigmoid activation function. This represents the intermodal attention weight matrix; Represents tensor multiplication. Indicates feature splicing, This represents feature map addition under a unified dimension.

9. The early warning method for a wind turbine nacelle fire early warning device based on multispectral fusion according to claim 8, characterized in that, The linkage between steps S302 and S303 includes: The foreground mask M(x,y,t) output by dynamic background modeling is used as the input region for false alarm suppression. Spectral database alignment is performed only for pixel regions where M(x,y,t)=1; Logic for determining oil contamination: ； In the formula, This represents the combined ultraviolet-visible light spectrum of the current pixel. The CosineSimilarity is calculated as follows, representing the sample spectra in the oil spill reflectance spectral database: ； The threshold of 0.85 is the optimal classification threshold determined by the ROC curve.

10. The early warning method for a wind turbine nacelle fire early warning device based on multispectral fusion according to claim 9, characterized in that, The fire confidence level calculation in step S401 uses the following method: Where GAP represents the importance of global average pooling output 3D-CNN features; M' represents the foreground mask after false positive suppression; and λ and μ represent the weight coefficients trained from historical fire data, respectively.

Images