Fan fire causal reasoning early warning method based on multi-source evidence fusion

Abstract

The invention belongs to the technical field of wind power safety early warning, and discloses a fan fire causal reasoning early warning method based on multi-source evidence fusion, and the method comprises the steps: achieving the precise quantification of cross-modal evidence conflicts through a two-factor conflict quantification model in combination with a static conflict coefficient and a time sequence conflict attenuation coefficient calculated through a Jousseme distance; a differential self-adaptive fusion strategy is adopted for different conflict degrees, low-conflict scenes are weighted and averaged according to preset weights, weights of medium-high-conflict scenes are dynamically adjusted, relevance verification is carried out, and a time sequence sliding window and an effective evidence cache pool are matched, so that fusion deviation caused by single data fluctuation is avoided; the confidence coefficient of the finally output unified evidence set is not lower than 0.85, and the accuracy of hidden danger cause chain tracing is improved; a whole-process causal reasoning model is constructed, fire cases and practical experience are integrated, and a causal relationship knowledge base marked with time sequence dependence labels is established.

Description

Technical Field

[0001] This invention belongs to the field of wind power safety early warning technology, specifically a causal reasoning early warning method for wind turbine fires based on multi-source evidence fusion. Background Technology

[0002] With the large-scale development of the wind power industry, wind turbine fires are occurring frequently. The causes encompass multiple dimensions, including equipment aging, human error, and environmental interference, and most exhibit a latent, gradual evolutionary pattern, such as the evolutionary process of bearing insulation aging → partial discharge → arcing → fire. Existing early warning technologies mainly suffer from the following technical problems: On the one hand, existing multimodal early warning methods mostly focus on feature-level splicing and fusion, lacking the ability to quantify evidence-level conflicts. For example, wind turbine multi-source data such as visual images, partial discharge signals, and temperature time series data are heterogeneous and easily lead to evidence conflicts. For example, high temperature signals may originate from normal equipment operation or from hidden dangers. However, existing methods have not established conflict quantification models and only fuse data through fixed weights, resulting in distorted tracing of the hazard cause chain and an inability to effectively distinguish between direct causes and indirect related factors.

[0003] On the other hand, traditional causal reasoning models do not take into account the time-series characteristics of wind turbine operation and maintenance, and the weights of causal relationships are static. For example, the evolution process of hidden wind turbine hazards such as minor lubricating oil leakage → grease accumulation → high-temperature carbonization has significant time-series dependence. Existing models only rely on static data to construct causal graphs and cannot dynamically update the weights of the correlation between hazards and causes, resulting in a significant lag in early warning of progressive hazards and making it difficult to achieve early intervention.

[0004] In addition, the existing edge-cloud collaborative architecture focuses on data transmission and distribution, but does not achieve dynamic collaboration between model optimization and operating condition adaptation. It also lacks differentiated processing strategies for the evolution stages of wind turbine fire faults, making it difficult to balance model deployment efficiency and early warning accuracy. Summary of the Invention

[0005] The purpose of this invention is to provide a causal reasoning early warning method for wind turbine fires based on multi-source evidence fusion, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a causal reasoning early warning method for wind turbine fires based on multi-source evidence fusion, comprising a data acquisition stage, a preprocessing stage, a feature extraction stage, a conflict fusion stage, a temporal causal reasoning stage, a risk assessment and early warning stage, and an iterative optimization stage; Preferably, the data acquisition phase involves deploying a multi-dimensional sensing array based on the wind turbine nacelle scenario to acquire visual three-dimensional spatial evidence, combined electrochemical evidence, environmental evidence, and time-series operation and maintenance evidence. The visual 3D spatial evidence is acquired through a mobile platform equipped with a dual-light camera and a lidar fusion device. The mobile platform is specifically a rail-guided robot, adapted for deployment in the narrow spaces of a wind turbine nacelle, used to capture images of potential hazards such as abnormal smoke and grease buildup. Simultaneously, lidar point cloud modeling is used to locate the hazard in 3D space and monitor the flickering characteristics of smoke edges. The temporal characteristics of smoke are identified by the brightness and color change rate of adjacent frame pixels. The electrochemical joint evidence is acquired collaboratively by a wireless temperature sensor, a partial discharge monitoring module, and a nanoparticle sensor to achieve coordinated monitoring of electrical parameters and chemical properties. Dedicated guide rails are laid along the top and inner walls of the wind turbine nacelle. The total length of the guide rails is adapted to the length of the nacelle. The robot is equipped with two moving nodes, which cover the generator area at the front of the nacelle and the gearbox area at the rear, respectively. The moving speed between nodes is ≤0.5m / s, and the distance between the stopping monitoring points is ≤1.5m. A total of 6-8 wireless temperature sensors are deployed in key parts such as bearing end caps, cable connectors, and contactors, with at least one sensor installed in each key part. The installation position is ≤5cm away from the core heat-generating component. Two partial discharge monitoring modules are installed, one at the end of the generator stator winding and one inside the switch cabinet. Two nanoparticle sensors are deployed, one in the gearbox vent and one in the central ventilation area of ​​the engine compartment. Temperature and humidity sensors and combustible gas concentration sensors are installed in the middle of the cabin, while wind speed sensors are installed on the windward side of the cabin exterior, at a distance of ≥30cm from the cabin shell.

[0007] Environmental evidence includes temperature and humidity data, combustible gas concentration sensor data, and wind speed sensor data are acquired simultaneously to correct the impact of environmental interference on the monitoring signal; time-series operation and maintenance evidence includes the unit start-up and shutdown status recorded by the SCADA system, historical maintenance records, and is labeled with fault evolution stage tags based on the three-stage characteristics of grease thermal decomposition. The fault evolution stage label is based on the progressive evolution characteristics of wind turbine fire hazards, specifically including three stages: early, middle and late. The early stage is the budding stage of the hazard, such as minor grease leakage and slight insulation aging; the middle stage is the development stage of the hazard, such as grease accumulation and partial discharge; and the late stage is the critical stage of the hazard, such as high-temperature carbonization and electric arcing. The initial values ​​of the fault evolution stage correction coefficients for each stage are 0.8 for the early stage, 1.0 for the middle stage, and 1.2 for the late stage. These initial values ​​can be fine-tuned through the dynamic weight update mechanism of the time-series causal reasoning stage. GPS timing and lidar spatial calibration technology are used to achieve spatiotemporal synchronization of multi-source evidence, and all evidence is standardized into JSON format containing feature vectors, confidence labels, spatial coordinates and time series stages.

[0008] Preferably, the preprocessing stage employs a hierarchical preprocessing and dynamic screening strategy. Visual evidence uses 3×3 Gaussian filtering for noise reduction and adaptive brightness enhancement, combined with HSV color space screening rules to remove false smoke areas. Electrical evidence uses a 5-layer wavelet transform based on the db4 wavelet basis to filter electromagnetic interference and implements baseline drift correction for nanoparticle signals. Time-series evidence fills missing values ​​with linear interpolation and cleans up outliers using the DBSCAN-MAD algorithm. This algorithm combines density clustering (DBSCAN) and median absolute deviation (MAD) characteristics to accurately identify isolated outliers in time-series data, adapting to the continuity requirements of wind turbine operation and maintenance time-series data. A three-dimensional evaluation model was established with four input indicators: sensor health status, data consistency, temporal correlation, and evidence confidence. An adaptive dynamic threshold was used to screen the effective evidence set. The screening threshold was adjusted according to the real-time operating conditions of the wind turbine through the KS test to eliminate low-confidence and isolated temporal evidence.

[0009] Sensor health status indicators include sensor power supply stability (voltage fluctuation ≤ ±5% is acceptable), signal transmission packet loss rate (≤3% is acceptable), and historical fault frequency (≤2 times in the last 30 days is acceptable). Data consistency indicators include data deviation from the same type of sensor (≤10% is considered consistent) and data deviation from the historical average for the same period (≤15% is considered consistent). The temporal correlation index includes the similarity of evidence features between adjacent frames (≥0.7 indicates correlation) and the matching degree between data change trends and fault evolution patterns (≥0.6 indicates matching). The weights for sensor health status (0.3), data consistency (0.25), temporal correlation (0.25), and evidence confidence (0.2) are calculated by weighting the results. A score ≥ 0.7 is considered valid evidence.

[0010] Preferably, the feature extraction stage employs a dedicated extraction and lightweight adaptation algorithm for different types of valid evidence after preprocessing, and each piece of evidence is appended with a quintuple of hazard category, confidence level, evidence type, spatial coordinates, and temporal stage. Visual 3D features are extracted by improving Faster R-CNN to extract deep visual features. This algorithm uses ResNet101 as the backbone network. First, a data alignment module is used to achieve spatiotemporal alignment between visual images and LiDAR point clouds. The CBAM attention module and RoIAlign are integrated to improve the accuracy of small target hazard localization. At the same time, the geometric features of LiDAR point clouds are extracted, and deep fusion of visual and 3D features is achieved through a feature fusion matrix. A knowledge distillation strategy is used to optimize the model to adapt to edge deployment. Electrochemical features are extracted by Fourier transform to extract the frequency domain features of partial discharge signals in the 3MHz to 100MHz band. The mean, peak, and trend slope of temperature time series are extracted through statistical analysis. At the same time, the time series change rate of nanoparticle concentration is extracted and a feature vector relating electrical parameters and nanoparticle concentration is established. Time series operation and maintenance features are extracted by using a bidirectional LSTM network with 256 hidden layers to extract trend features. An attention gating mechanism is integrated to focus on key time series nodes of fault evolution. The model is optimized 100 times. At the same time, historical maintenance records are transformed into structured feature vectors and fused with time series operation features.

[0011] In the improved module connection relationship of Faster R-CNN, the input layer receives the preprocessed cabin visual image and LiDAR point cloud → through the data alignment module → ResNet101 backbone network extracts basic features → CBAM attention module → RoIAlign layer prunes target features → fully connected layer completes hazard classification and 3D coordinate regression → outputs visual 3D fusion feature vector. Data augmentation employs random flipping, brightness / contrast adjustment, Gaussian blur, and small target cropping and enlargement to adapt to the narrow space of the cabin and scenarios with potential hazards from small targets. IoU ≥ 0.5 is considered a positive sample, IoU < 0.3 is considered a negative sample, and IoU ≤ 0.3 < 0.5 is considered an ignored sample. During training, the initial learning rate was 0.001, the SGD optimizer was used, the momentum was 0.9, the weight decay was 0.0001, the backbone network head was frozen for the first 20 rounds, and the entire network was unfrozen and fine-tuned for the next 30 rounds with a learning rate of 0.0001. The optimal model was verified and saved every 5 rounds. Adding small-sized anchor frame sampling weights to the RoIAlign layer improves the recognition accuracy of small targets such as grease buildup and smoke.

[0012] In the module connection relationship of bidirectional LSTM, the input layer receives standardized time-series operation and maintenance data → the embedding layer transforms discrete features into a 256-dimensional vector → 2 layers of bidirectional LSTM hidden layers → the attention gating module focuses on the critical fault nodes → the fully connected layer reduces the dimension to 512 dimensions → outputs the time-series operation and maintenance feature vector. The window is divided into 30 time steps, the numerical features are normalized, and the missing values ​​are filled by linear interpolation and then smoothed by moving average. The Adam optimizer is used with an initial learning rate of 0.001, weight decay of 0.0005, dropout probability of 0.3, and activation function tanh. Training operation: 100 iterations, with the learning rate decaying by 50% every 10 iterations. An early stopping strategy is adopted, and the training stops if the validation set loss does not decrease for 5 consecutive iterations. The attention gating module incorporates a periodic weighting factor to enhance the feature weights during critical periods such as after maintenance and under high load.

[0013] In the improved Faster R-CNN parameters, the anchor box sizes were set to 16×16, 32×32, 64×64, and 128×128, with anchor box ratios of 1:1, 1:2, and 2:1; the training batch size was 8, and the initial learning rate was 0.001, which decayed to 0.0001 after 50 iterations. In the bidirectional LSTM network parameters, the hidden layer dimension is 256, the number of layers is 2, the dropout probability is 0.3, the activation function is tanh, and the optimizer is Adam. In the feature fusion matrix, visual features and LiDAR geometric features are fused to 512 dimensions using a 1×1 convolution kernel and then concatenated. Electrical parameters and nanoparticle concentration features are multiplied element-wise and then concatenated with temperature time-series features. The final fused feature dimension is 1024.

[0014] Preferably, the conflict fusion stage uses an improved Jousselme distance and temporal conflict attenuation dual-factor conflict quantification model based on the various evidence features output from the feature extraction stage. The conflict quantification model calculates the static conflict coefficient between evidence based on the Jousselme distance and designs a temporal conflict attenuation coefficient. The comprehensive conflict coefficient is obtained by weighting the two to achieve conflict quantification optimization. For evidence with a comprehensive conflict coefficient ≤ 0.3, a weighted average fusion is performed based on the weight ratios of visual 3D spatial evidence (0.4), electrochemical joint evidence (0.3), environmental evidence (0.15), and time-series maintenance evidence (0.15). For evidence with a conflict coefficient > 0.3, a dynamic weight adjustment mechanism is activated. If the confidence level of multiple consecutive frames exceeds a preset threshold, the weight is increased, and the consistency of trends of adjacent frame evidence is judged simultaneously. For high-conflict scenarios with C > 0.5, cross-evidence type correlation verification is triggered. The entire fusion process employs a batch data processing mechanism every 10ms, coupled with a cache pool that caches the most recent 50 frames of valid evidence and temporal sliding window analysis, ultimately outputting a unified evidence set with a confidence level ≥ 0.85 after fusion.

[0015] Preferably, the temporal causal reasoning stage is based on the unified evidence set output by the conflict fusion stage, and establishes a full-process model of causal temporal synchronization, hierarchical reasoning, and confidence calibration. First, fire cases and practical experience are integrated to build a knowledge base of causal relationships of wind turbine fires, and the temporal dependency tags of the causal chain and the cross-causal chain association rules are labeled. The knowledge base storage adopts a three-level structure of cause nodes, intermediate nodes, and potential problem nodes. Each node contains a unique identifier, descriptive text, a list of associated node IDs, and a time interval threshold range (corresponding to the time interval of each link in the cause chain). The time interval thresholds for each link in the causal chain are determined based on case statistics. For example, the time interval threshold for "lubricating oil leakage → grease accumulation" is 7 to 30 days, and for "grease accumulation → high-temperature carbonization" it is 3 to 15 days. The cross-causal chain association rules are expressed in the format of "triggering condition-association chain-influence weight". For example, when the concentration of combustible gas in environmental evidence is ≥0.5%LEL, the causal weight of the "partial discharge → arc ignition" causal chain is increased by 0.15.

[0016] Based on this, the FIRE-VLBERT model is improved by integrating a temporal attention module and a temporal causal synchronization module. The temporal attention module adopts a sliding window with a dynamically adjustable window size, while the causal synchronization module identifies lagging causal relationships based on the Granger causality test to achieve temporal alignment between causes and hazards. The model input integrates evidence features, fire terminology vectors, and temporal operation and maintenance features to dynamically update the causal relationship weights. The model training adopts a three-stage strategy: the first stage performs text, image, and temporal feature alignment training; the second stage performs full network fine-tuning; and the third stage implements causal temporal consistency constraint training. In the improved module connection relationship of FIRE-VLBERT, the input layer receives fused evidence features, fire terminology vectors, and time-series operation and maintenance features → feature alignment module → time-series attention module → time-series causal synchronization module → 6-layer Transformer encoder to mine causal relationships → confidence calibration module to correct results → output effective hidden danger cause chain, time interval and evolution stage; The first stage freezes the encoder and trains the feature alignment module, using contrastive loss to achieve feature clustering; the second stage fine-tunes the entire network, using a weighted sum of cross-entropy loss and MSE loss; the third stage adds temporal consistency constraints to enhance the accuracy of early-stage hidden danger inference. The fire terminology vector contains 180 fan-specific terms, and the embedding layer enhances the association weight between equipment and fault terms; the causal synchronization module presets the lag time range of common cause chains to improve inference efficiency.

[0017] In addition to triggered updates, the knowledge base undergoes a comprehensive update every three months. The updates include supplementing the causal chains of new fire cases, statistically correcting the time interval thresholds, and optimizing the cross-causal chain association rules. After the update, it must be verified through 10 sets of full-condition tests to ensure the adaptability of the knowledge base to the actual scenario. Only when the verification pass rate is ≥90% can it be put into use.

[0018] By mining causal relationships between evidence through a 6-layer Transformer encoder, and calibrating causal confidence by combining confusion matrix and causal time sequence error, the final output is an effective hidden danger cause chain with confidence ≥0.75, while marking the time interval and failure evolution stage of each link.

[0019] The model input features are fused with 1024 dimensions of evidence features, 256 dimensions of fire terminology vectors, and 512 dimensions of time-series operation and maintenance features, resulting in a total dimension of 1792 after concatenation. Causal weight updates are triggered when the matching degree between newly collected evidence and the causal chain in the knowledge base is ≥0.85; updates are triggered when the temporal interval deviation of the same causal chain exceeds 20% for three consecutive occurrences; updates are triggered when the fault evolution stage switches; the weight adjustment range for a single update is ±0.1 to avoid weight mutations affecting the stability of inference.

[0020] Preferably, the risk assessment and early warning stage is based on the effective hidden danger cause chain and related confidence data output from the time-series causal reasoning stage, and adopts a dynamic risk assessment and multi-channel push strategy; the dynamic risk calculation determines the weight of each indicator through fuzzy hierarchical analysis, incorporates the fault evolution stage correction coefficient, calculates the comprehensive risk value, and at the same time performs cluster analysis on historical risk data based on the DBSCAN-MAD algorithm to construct a dynamic risk threshold, and updates the four-level risk range according to historical fault data and real-time operating conditions; low risk is checked weekly, medium risk is rectified within 24 hours, higher risk is responded to within 4 hours, and high risk is dealt with immediately; The first-level indicators in the fuzzy hierarchical analysis method indicator system include hazard confidence, causal relationship strength, failure evolution stage, and environmental impact coefficient. The secondary indicators include the confidence level of potential hazards, which includes the confidence level of evidence fusion and the confidence level of the causal chain; the strength of causal relationship includes the weight of direct causation and the weight of indirect causation; and the environmental impact coefficient includes the wind speed correction coefficient and the temperature and humidity correction coefficient. The dynamic threshold is updated every 7 days based on the historical fault data of the past 30 days; it is updated immediately when a new high-risk fault case is added; the threshold adjustment range does not exceed 10% of the previous threshold to ensure threshold stability.

[0021] Differentiated early warning pushes are implemented through multiple channels, including cloud platforms, mobile apps, and cabin audible and visual alarms. For multi-fire source scenarios, the order of handling is prioritized according to risk value and fire source location. The push content includes the three-dimensional coordinates of the hazard, the prediction of the fault evolution stage, and the handling guidelines (including operation priorities). At the same time, the early warning data and on-site handling feedback will be transmitted to the iterative optimization stage.

[0022] Preferably, the iterative optimization stage establishes a three-level collaborative iterative mechanism of edge, cloud and professional verification based on the early warning data, non-early warning data and on-site handling feedback from the risk assessment and early warning stage; the edge regularly uploads non-early warning data and low-confidence difficult samples, the early warning data is uploaded in real time with on-site handling feedback, and the cloud filters samples according to preset standards to establish an incremental learning dataset containing working condition dimensions. Newly added hazard samples must be independently labeled by 3 professional technicians, and are considered valid if 2 or more people reach a consensus. The labeling content includes the hazard category, the stage of failure evolution, and the operating condition label. The operating condition label includes wind speed, load, and temperature and humidity range. Wind speed: ≤6m / s is low wind speed, 6~12m / s is medium wind speed, and >12m / s is high wind speed. Load: ≤40% is low load, 40%~80% is medium load, and >80% is high load. The incremental learning dataset consists of 70% training set, 20% validation set, and 10% test set. The dataset must cover all operating conditions, including 3 types of wind speed × 3 types of load × 3 types of temperature and humidity ranges, with each operating condition having a sample size of no less than 5% of the total sample size. Samples with a post-fusion confidence level of 0.6 to 0.85 that did not trigger an alert were marked as low-confidence, difficult samples.

[0023] The lightweight iteration of the model adopts a three-level network architecture: a cloud-based main network, edge slave networks, and a network adaptation and control unit. The cloud-based main network performs incremental learning and updates, while the edge slave networks inherit the core knowledge of the cloud-based main network through knowledge distillation. The network adaptation and control unit calculates the distillability and sparsity of the edge slave networks and dynamically adjusts the relevant parameters of the knowledge distillation intensity and sparsity pruning algorithm. The model update supports OTA breakpoint resumption, and the total update time is controlled within 8 minutes. If it exceeds 8 minutes, it will automatically be downgraded to a lightweight update package, which only transmits the core parameters. The distillability assessment index uses the consistency accuracy rate of teacher-student network predictions. A rate of ≥90% is considered high distillability, and a rate of ≤75% is considered low distillability. The sparsity evaluation index uses the network parameter redundancy rate. A redundancy rate ≥ 30% is considered to be highly sparsity-friendly. When high distillability and high sparsity are combined, the knowledge distillation intensity is set to 0.8~0.9, and the sparsity pruning ratio is set to 30%~40%; when high distillability and low sparsity are combined, the distillation intensity is set to 0.7~0.8, and the pruning ratio is set to 10%~20%; when low distillability is combined, the distillation intensity is set to 0.5~0.6, and the pruning ratio is set to 5%~10%. If network transmission is interrupted for more than 30 seconds, breakpoint resume will be initiated. After recovery, model updates will continue from the point of interruption. If the total update time exceeds 8 minutes, it will automatically downgrade to a lightweight update package, which only transmits core parameters.

[0024] The model accuracy is verified regularly through on-site measurements, including different wind speeds, temperatures, humidity, and unit load conditions. If the accuracy drops below a preset threshold (the preset threshold is 3%), a special fine-tuning is triggered, supplementing the sample under that condition for targeted training. The optimized model will then be applied back to the aforementioned stage.

[0025] The accuracy assessment uses accuracy, recall, F1 score, false negative rate, and false positive rate as core indicators. Among them, the passing standard is an early warning accuracy rate of ≥90%, recall rate of ≥85%, F1 score of ≥88%, false negative rate of ≤5%, and false positive rate of ≤8%. The full-condition testing scheme includes 27 operating conditions: wind speeds ≤6m / s, 6-12m / s, and >12m / s (3 groups each); temperature and humidity conditions -10~0℃ / 30-50%RH, 0-25℃ / 50-70%RH, and 25-40℃ / 70-90%RH (3 groups each); and unit load conditions ≤40%, 40-80%, and >80% (3 groups each). Each operating condition test sample size is ≥50, covering different stages of fault evolution.

[0026] The beneficial effects of this invention are as follows: 1. This invention achieves accurate quantification of cross-modal evidence conflict by using a two-factor conflict quantification model, combining the static conflict coefficient calculated by Jousselme distance with the temporal conflict attenuation coefficient. It employs a differentiated adaptive fusion strategy for different conflict levels: low-conflict scenarios are weighted averaged according to preset weights, while medium-to-high-conflict scenarios dynamically adjust weights and conduct correlation verification. Furthermore, it utilizes a temporal sliding window and an effective evidence cache pool to avoid fusion bias caused by single data fluctuations. The final output unified evidence set has a confidence level of no less than 0.85, improving the accuracy of tracing the causal chain of potential hazards.

[0027] 2. The full-process causal reasoning model constructed in this invention integrates fire cases and practical experience to establish a causal relationship knowledge base labeled with temporal dependency tags. The improved FIRE-VLBERT model integrates temporal attention and causal synchronization modules, identifies lagging associations through Granger causality tests, dynamically updates causal weights, and utilizes a 6-layer Transformer encoder to mine causal relationships between evidence. After error calibration, it outputs a valid hazard cause chain with a confidence level of not less than 0.75, and labels the temporal interval and evolution stage of each link. It can accurately capture the temporal evolution law of hidden hazards such as grease leakage and insulation aging, realize full-stage early warning, shorten response time, and reduce the probability of fire occurrence.

[0028] 3. This invention establishes a three-level collaborative iterative mechanism involving the edge, cloud, and professional verification ends. Various types of data are uploaded to the edge, while samples are selected in the cloud to construct an incremental learning dataset. Knowledge distillation enables lightweight deployment of the edge model, allowing for dynamic adjustment and optimization of parameters by the control unit. Model updates support OTA (Over-The-Air) resume capability. Simultaneously, regular full-condition field tests are conducted. If the accuracy drops by more than 3% under a certain condition, a special fine-tuning is triggered, supplementing corresponding samples for targeted training. This ensures efficient and flexible edge deployment, achieving full-condition model adaptation through continuous iteration. It maintains stable early warning accuracy under different wind speeds, temperatures, humidity levels, and unit load conditions, balancing deployment efficiency and early warning reliability, and adapting to the complex and ever-changing operating scenarios of wind turbines. Attached Figure Description

[0029] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a flowchart of the multi-source evidence collection and preprocessing process of the present invention; Figure 3 This is a flowchart of the multi-source evidence conflict fusion and causal reasoning process of the present invention; Figure 4 This is a flowchart of the risk assessment, early warning, and iterative optimization process of this invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] like Figures 1 to 4As shown, this embodiment of the invention provides a causal reasoning early warning method for wind turbine fires based on multi-source evidence fusion, including a data acquisition stage, a preprocessing stage, a feature extraction stage, a conflict fusion stage, a temporal causal reasoning stage, a risk assessment and early warning stage, and an iterative optimization stage. The specific implementation of each stage is as follows: The acquisition phase involves deploying a multi-dimensional sensing array based on the wind turbine nacelle scenario to collect visual three-dimensional spatial evidence, combined electrochemical evidence, environmental evidence, and time-series operation and maintenance evidence. The visual three-dimensional spatial evidence is acquired through a mobile platform equipped with a dual-light camera and a lidar fusion device. The mobile platform is specifically a rail-guided robot used to capture images of potential hazards such as abnormal smoke and grease buildup. The camera resolution is 2688×1520. Simultaneously, lidar point cloud modeling is used to achieve three-dimensional spatial positioning of the hazard with a spatial coordinate accuracy of ±2cm. It also performs smoke edge flicker feature monitoring and uses the brightness and color change rate of adjacent frame pixels to identify the temporal characteristics of the smoke. The electrochemical joint evidence is collected collaboratively by a wireless temperature sensor, a partial discharge monitoring module, and a nanoparticle sensor to achieve coordinated monitoring of electrical parameters and chemical properties. The infrared wavelength range of the dual-light camera is 8~14μm, the image resolution of both visible light and infrared is 2688×1520, the frame rate is 15~30fps, and the temperature measurement range is -20℃~550℃. The point cloud density of the lidar is ≥100 points / m. 2 The ranging range is 0.5~100m, the scanning frequency is 10~20Hz, and the angular resolution is ≤0.1°, ensuring that the three-dimensional positioning accuracy of potential hazards meets the standards.

[0032] Wireless temperature sensors are used to monitor the temperature of bearings and cables with an accuracy of ±0.5℃; partial discharge monitoring modules are used to monitor discharge pulse signals; and nanoparticle sensors are used to capture the concentration of characteristic nanoparticles generated by the thermal decomposition of lubricating oil. The wireless temperature sensor collects data at a 10-second interval; the partial discharge monitoring module collects data in real time at a sampling frequency of 1MHz; the nanoparticle sensor collects data at a 30-second interval; the temperature and humidity sensor and the combustible gas concentration sensor collect data at a 60-second interval; the wind speed sensor collects data at a 10-second interval; the dual-light camera and lidar mounted on the rail-guided robot collect data at a 500-ms interval, which is shortened to 200ms when switching mobile nodes.

[0033] The wireless temperature sensor has a measurement range of -40℃ to 125℃. The pulse signal measurement amplitude range of the partial discharge monitoring module is 100μV~10V; The nanoparticle sensor has a detection range of 0.01~10μm particle size and a concentration measurement range of 0~10mg / m³. The temperature and humidity sensor measures temperature from -20℃ to 60℃ and humidity from 0 to 100%RH. The combustible gas concentration sensor has a measurement range of 0~100%LEL and is designed for common combustible gases such as methane and propane. The wind speed sensor has a measurement range of 0~25m / s.

[0034] Environmental evidence includes temperature and humidity data, combustible gas concentration sensor data, and wind speed sensor data are acquired simultaneously to correct the impact of environmental interference on the monitoring signal; time-series maintenance evidence includes the unit start-up and shutdown status recorded by the SCADA system and historical maintenance records. The historical maintenance records include fields such as maintenance location and replaced parts, and are labeled with fault evolution stage tags based on the three-stage characteristics of grease thermal decomposition. GPS timing and LiDAR spatial calibration technology are used to achieve spatiotemporal synchronization of multi-source evidence, with a time error of ≤5ms and a spatial alignment error of ≤2cm. All evidence is standardized into JSON format, which includes feature vectors, confidence labels, spatial coordinates and temporal stages. For example, the visual 3D fusion evidence vector has a dimension of 1024, and data with a confidence level of ≥0.85 is considered valid.

[0035] The preprocessing stage employs a layered preprocessing and dynamic filtering strategy to improve data quality and relevance. Visual evidence is denoised using 3×3 Gaussian filtering and brightness adaptive enhancement, and false smoke areas are removed by combining HSV color space filtering rules. In the HSV screening threshold for smoke areas, the hue range is 0~360°, the saturation range is 0~60%, and the value range is 30~80%. Areas outside this range are judged as pseudo-smoke areas and are directly removed.

[0036] Electrical evidence uses a 5-layer wavelet transform based on the db4 wavelet basis to filter electromagnetic interference and performs baseline drift correction for nanoparticle signals; time-series evidence fills in missing values ​​through linear interpolation and cleans up outliers using the DBSCAN-MAD algorithm; this algorithm combines density clustering (DBSCAN) and median absolute deviation (MAD) characteristics to accurately identify isolated outliers in time-series data, adapting to the continuity requirements of wind turbine operation and maintenance time-series data; A three-dimensional evaluation model was established with four input indicators: sensor health status, data consistency, temporal correlation, and evidence confidence. An adaptive dynamic threshold was used to screen the effective evidence set. The screening threshold was adjusted by KS test according to the real-time operating conditions of the wind turbine, such as start-up, normal operation, and high load. The confidence threshold ranged from 0.6 to 0.75 to eliminate low-confidence and isolated temporal evidence, ensuring that the proportion of effective evidence after screening was ≥85%.

[0037] The confidence threshold for unit start-up and shutdown conditions is 0.7~0.75; the confidence threshold for normal operation conditions is 0.65~0.7; and the confidence threshold for high load conditions is 0.6~0.65. When adjusting, the operating condition switching signal is used as the trigger point, and the threshold update is completed within 3 seconds to ensure timely screening.

[0038] In the feature extraction stage, for different types of valid evidence after preprocessing, a dedicated extraction and lightweight adaptation algorithm is used to enhance the feature targeting and edge deployment feasibility. Each piece of evidence is attached with a five-tuple of hazard category, confidence level, evidence type, spatial coordinates and time series stage, which facilitates tracing the data source and causal chain association, and ensures the targeting, interpretability and edge deployment adaptability of feature extraction. Visual 3D features are extracted by improving Faster R-CNN to extract deep visual features. The algorithm uses ResNet101 as the backbone network. First, a data alignment module is used to achieve spatiotemporal alignment between visual images and LiDAR point clouds. The CBAM attention module and RoIAlign are integrated to improve the accuracy of small target hazard localization. At the same time, geometric features such as volume and surface area of ​​hazard areas in LiDAR point clouds are extracted, and deep fusion of visual and 3D features is achieved through feature fusion matrix. A knowledge distillation strategy is used to optimize the model to adapt to edge deployment. The teacher network is a full ResNet101, and the student network is a lightweight MobileNetV2. The module connections for knowledge distillation are as follows: Teacher network (complete improved Faster R-CNN) → Feature distillation module extracts key layer features of the backbone network → Student network (MobileNetV2) → Feature imitation module minimizes teacher-student feature error → Classification head distillation optimizes output → Obtaining a lightweight model at the edge.

[0039] The teacher's network was trained independently until convergence, and then the parameters were fixed. The student's network was pre-trained independently for 20 rounds. During distillation training, the learning rate was 0.0005, the AdamW optimizer was used, the distillation temperature was 4, and the loss function was a weighted average of the distillation loss (0.6) and the classification loss (0.4). After distillation, redundant channels are sparsely pruned, the number of parameters is compressed to 30% of the teacher network, and INT8 quantization is performed to ensure that the inference latency at the edge is ≤100ms / frame.

[0040] The knowledge distillation temperature was set to 3-5°C; the loss function was a weighted sum of distillation loss and classification loss, with weights of 0.6 and 0.4 respectively; during the distillation process, the teacher network parameters were frozen, and only the student network was trained to ensure that the edge model was lightweight and the accuracy loss was ≤5%.

[0041] Electrochemical characteristics were extracted by Fourier transform of the frequency domain features of partial discharge signals in the 3MHz to 100MHz band, and the mean, peak value, and trend slope of temperature time series were extracted by statistical analysis. At the same time, the time series change rate of nanoparticle concentration was extracted and a feature vector of correlation between electrical parameters and nanoparticle concentration was established. Three core features of partial discharge signals in the frequency domain are extracted: peak frequency, main frequency bandwidth, and total harmonic distortion. Two new types of features, variance and root mean square, have been added to the temperature time series features, which together with the original mean, peak value and trend slope constitute the temperature feature set. The temporal variation characteristics of nanoparticle concentration are supplemented with two other characteristics: the acceleration of change and the proportion of the steady segment, which strengthens the correlation with chemical properties.

[0042] The time-series operation and maintenance features are extracted using a bidirectional LSTM network with 256 hidden layers. An attention gating mechanism is integrated to focus on key time-series nodes of fault evolution. The model is optimized 100 times. At the same time, historical maintenance records are transformed into structured feature vectors and fused with time-series operation features.

[0043] The conflict fusion stage, based on the various evidence features output from the feature extraction stage, adopts an improved Jousselme distance and temporal conflict decay dual-factor conflict quantification model to achieve differentiated adaptive fusion, thus solving the problems of incomplete conflict quantification and static weights in existing methods. The conflict quantification model calculates the static conflict coefficient between evidence based on Jousselme distance, and simultaneously designs a temporal conflict attenuation coefficient. Attenuation weights are set according to the evidence collection time difference and the fault evolution stage. The attenuation coefficient for recent high-confidence evidence is ≤0.1, and the attenuation coefficient for historical evidence is ≥0.3. Conflict quantification is optimized through a comprehensive conflict coefficient. The comprehensive conflict coefficient C = α × static conflict coefficient + β × temporal conflict attenuation coefficient, where α and β are dynamic weights. The initial value range of α is 0.6~0.8, and the initial value range of β is 0.2~0.4. After offsetting, α is maintained at 0.5~0.9, and β is adjusted accordingly to ensure α+β=1. When the wind turbine is under high load or complex environment conditions, α shifts towards 0.8 to strengthen static conflict quantification. When the temporal correlation of evidence is strong, β shifts towards 0.4 to strengthen temporal conflict attenuation, ensuring the accuracy of conflict quantification under different operating conditions.

[0044] The timing conflict attenuation factor is set according to the following rules: When the time difference in evidence collection is ≤5 min, the attenuation coefficient is 0.05~0.1; when the time difference is ≤30 min and 5 min < time difference, the attenuation coefficient is 0.1~0.2; when the time difference is >30 min, the attenuation coefficient is 0.2~0.3. In the early stage of fault evolution, the attenuation coefficient is reduced by 0.02 within the corresponding time difference interval; in the middle stage, the baseline value is maintained; and in the late stage, it is increased by 0.03. When the fan is under high load (load > 80% of rated power) or in a complex environment (wind speed > 12m / s, humidity > 85%), α is taken as 0.7~0.8 and β is taken as 0.2~0.3; when the evidence time sequence correlation index (adjacent frame correlation degree ≥ 0.8) is used, β is taken as 0.3~0.4 and α is taken as 0.6~0.7.

[0045] For evidence with a comprehensive conflict coefficient ≤ 0.3, a weighted average fusion is performed based on the weight ratios of visual 3D spatial evidence (0.4), electrochemical joint evidence (0.3), environmental evidence (0.15), and time-series maintenance evidence (0.15). For evidence with a conflict coefficient > 0.3, a dynamic weight adjustment mechanism is activated. If the confidence level of a certain type of evidence is ≥ 0.95 for 5 consecutive frames, the weight is increased by 20%. Simultaneously, a trend consistency judgment is implemented for adjacent frames of evidence. If the consistency is ≥ 80%, the weight is further strengthened. For high-conflict scenarios with C > 0.5, cross-evidence type correlation verification is triggered, such as verification of high temperature signals combined with nanoparticle concentration or smoke characteristics. The default decision window is 5 consecutive frames, which can be adjusted to 3 to 8 frames according to the wind turbine conditions such as start-up / shutdown / normal operation / high load through cloud configuration; the default confidence threshold is 0.95, which can be lowered to 0.9 in low-stability scenarios; When the confidence condition of consecutive frames is met, the weight of the target evidence is increased by 20% relative to the initial weight, and the upper limit of the weight after a single adjustment is 1.5 times the initial weight; the consistency of the trend of evidence in adjacent frames is determined by the Pearson correlation coefficient, and when the correlation coefficient is ≥0.8, the weight is increased by an additional 5%. High-conflict scenarios require feature matching of at least two different types of source evidence. For example, high-temperature signals must match nanoparticle concentration ≥ 0.3 mg / m³ or smoke feature vector similarity ≥ 0.7. Only after verification can they participate in fusion.

[0046] The entire fusion process employs a batch data processing mechanism every 10ms, coupled with a cache pool that caches the most recent 50 frames of valid evidence and temporal sliding window analysis, to avoid fusion deviations caused by single data fluctuations. The final output is a unified evidence set with a confidence level ≥0.85, which improves the stability and temporal correlation of the fusion results.

[0047] The window length of the time-series sliding window is set to 10-20 frames, with a default of 15 frames; the window sliding step size is 5 frames, and the comprehensive conflict coefficient and fusion weight of the evidence within the window are recalculated after each slide; when the proportion of valid evidence within the window is less than 60%, the window length is automatically reduced to 10 frames to ensure the validity and relevance of the data within the window.

[0048] When the cache pool is not full, evidence is stored in the order of collection time; when the cache pool is full, evidence with a confidence level <0.7 is discarded first; if the confidence level of all cached evidence is ≥0.7, the earliest collected evidence is discarded to ensure that the cache pool always retains high-value evidence.

[0049] The temporal causal reasoning stage is based on the unified evidence set output by the conflict fusion stage. It establishes a full-process model of causal temporal synchronization, hierarchical reasoning, and confidence calibration to enhance temporal dependency and causal chain tracing accuracy. It first integrates more than 1,200 fire cases and the practical experience of more than 60 operation and maintenance technicians to build a causal relationship knowledge base for wind turbine fires, and marks the temporal dependency tags of causal chains, such as the temporal interval threshold of lubricating oil leakage → grease accumulation → high temperature carbonization, and also marks cross-causal chain association rules. Based on this, the FIRE-VLBERT model is improved by integrating a temporal attention module and a temporal causal synchronization module. The temporal attention module adopts a sliding window with a dynamically adjustable window size of 5 to 10 frames, which adapts to the speed of hazard evolution. The causal synchronization module identifies lagging causal relationships based on Granger causality test to achieve temporal alignment between causes and hazards. The model input integrates evidence features, fire terminology vectors, and temporal operation and maintenance features to dynamically update causal relationship weights. The model training adopts a three-stage strategy: the first stage performs text, image, and temporal feature alignment training; the second stage performs full network fine-tuning; and the third stage implements causal temporal consistency constraint training. The fire terminology vector is constructed based on wind turbine component terms, fault type terms, and fire-related substance terms: wind turbine component terms include generators, gearboxes, bearings, cables, etc.; fault type terms include leakage, aging, discharge, carbonization, etc.; fire-related substance terms include lubricating oil, combustible gas, nanoparticles, etc.

[0050] By mining causal relationships between evidence through a 6-layer Transformer encoder, and calibrating causal confidence by combining confusion matrix and causal time sequence error, the final output is an effective hidden danger cause chain with confidence ≥0.75, while marking the time interval and failure evolution stage of each link.

[0051] Error calibration is triggered when the deviation between the predicted time interval and the knowledge base threshold is greater than 30%. Error calibration is based on historical calibration data of similar cases, and the causal confidence is linearly adjusted according to the deviation ratio. The adjusted confidence should not be lower than 0.7 to avoid over-adjustment and distortion.

[0052] The risk assessment and early warning stage is based on the effective hazard causal chain and related confidence data output from the time-series causal reasoning stage. It employs dynamic risk assessment and multi-channel push strategies to improve the timeliness and targeted nature of early warnings. Dynamic risk calculation uses fuzzy hierarchical analysis to determine the weight of each indicator, incorporating correction coefficients for the fault evolution stage. The initial values ​​are 0.8 for the early stage, 1.0 for the middle stage, and 1.2 for the late stage. After dynamic fine-tuning in the time-series causal reasoning stage, the comprehensive risk value S is calculated as: S = Σ(hazard confidence score × 0.35 + causal correlation strength score × 0.3 + fault evolution stage score × 0.2 × correction coefficient + environmental impact coefficient score × 0.15). Simultaneously, the DBSCAN-MAD algorithm is used to cluster historical risk data, constructing dynamic risk thresholds and updating the four-level risk range based on historical fault data and real-time operating conditions. Low risk (S < 1.8) requires weekly inspection; medium risk (1.8 ≤ S < 2.8) requires 24-hour rectification; higher risk (2.8 ≤ S < 3.8) requires 4-hour response; and high risk (S ≥ 3.8) requires immediate action. The wind speed correction factor is set to 1.0 when the wind speed is ≤6m / s, 1.1 when the wind speed is 6m / s to 12m / s, and 1.2 when the wind speed is >12m / s. The temperature and humidity correction factor is 1.1 when the temperature is >30℃ and the humidity is <40%, 0.9 when the temperature is <0℃ and the humidity is >80%, and 1.0 for other operating conditions; the environmental impact factor is the product of the two correction factors.

[0053] Differentiated early warning pushes are implemented through multiple channels, including cloud platforms, mobile apps, and cabin audible and visual alarms. For multi-fire source scenarios, the order of handling is prioritized according to risk value and fire source location. The push content includes the three-dimensional coordinates of the hazard, the prediction of the fault evolution stage, and the priority guide for handling operations, ensuring that maintenance personnel can quickly locate and accurately handle hazards. At the same time, the early warning data and on-site handling feedback will be transmitted to the iterative optimization stage.

[0054] The handling guidelines are structured into five parts: hazard identification, risk assessment, operating procedures, safety precautions, and feedback requirements. Hazard identification clearly defines the three-dimensional coordinates corresponding to specific parts of the engine room, such as the generator front bearing, with coordinates X: 2.3m / Y: 1.5m / Z: 0.8m. The risk assessment briefly describes the evolution trend of the hazard and the possible consequences. The operating procedures list targeted handling actions step by step, such as for low-risk grease buildup: wipe it with a special cleaning agent after shutdown, and test the concentration of nanoparticles after cleaning. The safety precautions emphasize shutdown operations, fire prevention, and explosion prevention requirements. The feedback requirements clearly define the feedback time limit and feedback channels.

[0055] Fire sources located ≤1m from core components such as generators and gearboxes are classified as Level 1 (highest priority), 1~3m as Level 2, 3~5m as Level 3, and >5m as Level 4 (lowest priority). They are first sorted in descending order of risk value, and if the risk values ​​are the same, they are sorted in ascending order of location priority to ensure that potential hazards in critical areas are dealt with first.

[0056] The iterative optimization phase is based on various data such as early warning data, non-early warning data, and on-site handling feedback from the risk assessment and early warning phase. It establishes a three-level collaborative iterative mechanism of edge, cloud, and professional verification to achieve continuous model optimization and full-condition adaptation. The edge uploads non-early warning data and low-confidence difficult samples every hour, and early warning data is uploaded in real time with on-site handling feedback. The cloud selects samples according to the standard of "newly added hidden danger samples are verified by 3 professional technicians and labeled with working condition tags" to establish an incremental learning dataset containing working condition dimensions. The lightweight model iteration adopts a three-level network architecture consisting of a cloud-based main network, edge slave networks, and a network adaptation and control unit. The cloud-based main network performs incremental learning and updates, while the edge slave networks inherit the core knowledge of the cloud-based main network through knowledge distillation. The network adaptation and control unit calculates the distillability and sparsity of the edge slave networks and dynamically adjusts the relevant parameters of the knowledge distillation intensity and sparsity pruning algorithm. The model update supports OTA breakpoint resumption, and the total update time is controlled within 8 minutes. If it exceeds 8 minutes, it is automatically downgraded to a lightweight update package, which only transmits the core parameters, ensuring a balance between lightweight and accuracy of the edge model.

[0057] The cloud-based main network (a hybrid architecture of ResNet50 and Transformer) is responsible for incremental learning and model optimization; the edge network (MobileNetV2 and lightweight LSTM) is responsible for real-time data processing and early warning output; and the adaptation and control unit is responsible for calculating distillability and sparsity, dynamically adjusting the distillation intensity and pruning ratio, and connecting the cloud and the edge.

[0058] The cloud receives samples uploaded from the edge devices, and after professional annotation, it selects newly added hidden dangers, missing working conditions, and low-confidence correction samples to build an incremental dataset. The model was trained in the cloud and loaded with the best historical model. The feature fusion layer and inference layer were fine-tuned with a learning rate of 0.0001 and 20 iterations were performed. The model transfers core knowledge to the edge through knowledge distillation, enabling OTA breakpoint resumption and deployment time of ≤8 minutes; Edge verification is tested on 100 sets of samples. If the results meet the criteria, the update takes effect; otherwise, the feedback is rolled back.

[0059] The model's accuracy is verified monthly through on-site measurements, including different wind speeds, temperatures, humidity, and unit load conditions. If the accuracy drops by ≥3% under a certain condition, a special fine-tuning is triggered, and additional samples are added for that condition for targeted training to ensure the model's long-term stability and adaptability to all operating conditions. The optimized model will then be applied in reverse to preliminary stages such as cross-modal feature extraction, evidence conflict fusion, and temporal causal reasoning.

[0060] The preset threshold for model accuracy decline is uniformly set at 3%, with the average of accuracy, recall, and F1 score as the evaluation benchmark. If a single indicator (false negative rate > 5% or false positive rate > 8%) exceeds the threshold, a special fine-tuning will be triggered even if the average accuracy decline does not reach 3%, to ensure the reliability of the early warning.

[0061] The number of training rounds for specific fine-tuning is set to 30-50 rounds; the initial learning rate is 0.0005, which decays by 50% every 10 rounds; the batch size is set to 16; the training process adopts an early stopping strategy, and training is stopped if the accuracy of the validation set does not improve for 5 consecutive rounds, and the optimal model is retained.

[0062] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A causal reasoning early warning method for wind turbine fires based on multi-source evidence fusion, characterized in that, It includes the acquisition stage, preprocessing stage, feature extraction stage, conflict fusion stage, temporal causal reasoning stage, risk assessment and early warning stage, and iterative optimization stage; Data Acquisition Phase: Based on the wind turbine nacelle scenario, a multi-dimensional sensing array is deployed to collect visual three-dimensional spatial evidence, combined electrochemical evidence, environmental evidence, and time-series operation and maintenance evidence. Multi-source evidence is aligned through spatiotemporal synchronization technology, and the evidence is standardized into a unified format. Preprocessing stage: A hierarchical preprocessing and dynamic screening strategy is adopted to perform noise reduction, baseline drift correction and outlier cleaning on various types of evidence, and a three-dimensional evaluation model is used to adaptively screen valid evidence; Feature extraction stage: For preprocessed valid evidence, a dedicated extraction and lightweight adaptation algorithm is used to extract various features and perform deep fusion, and add multi-dimensional attribute information; Conflict fusion stage: Based on cross-modal features, a two-factor conflict quantification model is adopted, and a differentiated adaptive fusion strategy is executed. Through batch processing, time-series sliding window analysis and effective evidence caching pool, fusion bias is avoided in collaboration, and a unified evidence set with qualified confidence is output. The temporal causal reasoning stage: Based on a unified evidence set, a full-process causal reasoning model is established, a causal relationship knowledge base is constructed, the pre-trained model is improved and the causal weights are dynamically updated, and effective hidden danger cause chains and evolution information are output. Risk assessment and early warning stage: Based on effective hazard cause chain data, adopt dynamic risk assessment and multi-channel push strategy, classify risk levels and push out handling guidelines, and synchronously feed back early warning data to the iterative optimization stage; Iterative optimization phase: A three-level collaborative iterative mechanism is established based on early warning feedback data. The model is optimized through incremental learning and lightweight network architecture, and full-condition test verification and fine-tuning are performed.

2. The wind turbine fire causal reasoning early warning method based on multi-source evidence fusion according to claim 1, characterized in that, During the data collection phase, the deployment of the multi-dimensional sensing array and the evidence collection methods are as follows: visual three-dimensional spatial evidence is acquired through a mobile carrier equipped with a dual-light camera and lidar fusion device, simultaneously achieving hazard image capture, three-dimensional spatial positioning, and smoke temporal feature identification; electrochemical joint evidence is collected collaboratively by a wireless temperature sensor, a partial discharge monitoring module, and a nanoparticle sensor; environmental evidence includes temperature and humidity data, combustible gas concentration data, and wind speed data, with the wind speed data used to correct for environmental interference; and time-series operation and maintenance evidence includes unit start-up and shutdown status, historical maintenance records, and is labeled with fault evolution stage tags. The fault evolution stage label is divided based on the progressive evolution characteristics of wind turbine fire hazards, specifically including three stages: early, middle and late. Each stage corresponds to a preset initial value of the fault evolution stage correction coefficient, which is fine-tuned through the dynamic weight update mechanism of the time-series causal reasoning stage. The spatiotemporal synchronization technology uses GPS timing and LiDAR spatial calibration. The standardized unified format is JSON format, which includes feature vectors, confidence labels, spatial coordinates, and time series stages.

3. The wind turbine fire causal reasoning early warning method based on multi-source evidence fusion according to claim 2, characterized in that, In the preprocessing stage, a hierarchical preprocessing strategy is implemented differently for different types of evidence: visual evidence is denoised, brightened, and color space filtering rules are used to remove false target areas; electrical evidence is filtered for electromagnetic interference using wavelet transform and baseline drift correction is performed; and temporal evidence is filled with missing values ​​through interpolation and outlier cleanup is performed. The three-dimensional evaluation model uses sensor health status, data consistency, temporal correlation, and evidence confidence as input indicators, and adopts an adaptive dynamic threshold to screen valid evidence. The screening threshold is adjusted through statistical tests based on the real-time operating conditions of the wind turbine to eliminate low-confidence and isolated temporal evidence.

4. The wind turbine fire causal reasoning early warning method based on multi-source evidence fusion according to claim 3, characterized in that, In the feature extraction stage, each piece of evidence is supplemented with a five-tuple attribute information including hazard category, confidence level, evidence type, spatial coordinates, and temporal stage. Visual 3D features are extracted by improving the target detection algorithm to extract deep visual features, integrating an attention module to improve the accuracy of locating potential hazards of small targets, and simultaneously extracting and deeply fusing geometric features of LiDAR point clouds. A model optimization strategy is adopted to adapt to edge deployment. Electrochemical features are extracted to extract frequency domain features of partial discharge signals, time-series features of temperature, and time-series variation features of nanoparticle concentration, and establish a feature vector relating electrical parameters to nanoparticle concentration. Time-series operation and maintenance features are extracted using a recurrent neural network to extract trend features, integrating an attention mechanism to focus on key time-series nodes of fault evolution, and transforming historical maintenance records into structured features and fusing them with time-series operation features.

5. The wind turbine fire causal reasoning early warning method based on multi-source evidence fusion according to claim 4, characterized in that, In the conflict fusion stage, the two-factor conflict quantification model calculates the static conflict coefficient based on the Jousselme distance, and then calculates the comprehensive conflict coefficient by weighting it with the temporal conflict attenuation coefficient. The differentiated adaptive fusion strategy is implemented in stages based on the comprehensive conflict coefficient: low conflict scenarios are fused by weighted average according to preset weights; medium and high conflict scenarios activate a dynamic weight adjustment mechanism, which adjusts the weights by combining the confidence of consecutive frames and the consistency of evidence trends; high conflict scenarios trigger cross-evidence type correlation verification. The fusion process employs a data batch processing mechanism and temporal sliding window analysis, coupled with an effective evidence caching pool, to ultimately output a unified evidence set with a confidence level ≥ 0.85 after fusion.

6. The wind turbine fire causal reasoning early warning method based on multi-source evidence fusion according to claim 5, characterized in that, In the temporal causal reasoning stage, the full-process causal reasoning model includes causal temporal synchronization, hierarchical reasoning, and confidence calibration modules. The causal relationship knowledge base is built based on fire cases and practical experience, and labels the causal chain temporal dependency tags and cross-causal chain association rules. The improved pre-trained model integrates a temporal attention module and a temporal causal synchronization module. The temporal attention module adopts a dynamic window adjustment mechanism, while the causal synchronization module identifies lagging causal relationships through causal testing. The model input integrates evidence features, fire terminology vectors, and temporal operation and maintenance features to dynamically update the causal relationship weights, and is optimized through multi-stage training. By mining causal relationships between evidence through encoders and correcting causal confidence through error calibration mechanisms, the final output is an effective hidden danger cause chain with a confidence level ≥ 0.75, as well as the time interval and failure evolution stage of each link.

7. The wind turbine fire causal reasoning early warning method based on multi-source evidence fusion according to claim 6, characterized in that, In the risk assessment and early warning stage, dynamic risk assessment determines the indicator weights through fuzzy hierarchical analysis, incorporates the fault evolution stage correction coefficient to calculate the comprehensive risk value, establishes a dynamic risk threshold based on historical fault data and real-time operating conditions, divides the risk into four levels according to the risk threshold, and sets differentiated handling time limits. Low-risk areas are checked weekly, medium-risk areas require rectification within 24 hours, higher-risk areas require response within 4 hours, and high-risk areas require immediate action. Multi-channel early warning push is achieved through cloud platform, mobile APP, and cabin sound and light alarm. In multiple fire source scenarios, the priority of handling is sorted according to risk value and fire source location. The push content includes the three-dimensional coordinates of the hidden danger, the prediction of the failure evolution stage, and the handling guide. The early warning data and on-site handling feedback are transmitted synchronously to the iterative optimization stage.

8. The wind turbine fire causal reasoning early warning method based on multi-source evidence fusion according to claim 7, characterized in that, In the iterative optimization phase, the three-level collaborative iterative mechanism consists of the edge terminal, the cloud terminal, and the professional verification terminal; the edge terminal uploads non-early warning data, low-confidence difficult samples, and early warning data with feedback, while the cloud terminal filters samples to construct an incremental learning dataset; The lightweight model iteration adopts a three-level network architecture. The cloud-based main network performs incremental learning, while the edge slave network inherits core knowledge through knowledge distillation. The adaptive control unit dynamically adjusts the model optimization parameters and supports breakpoint resume for model updates. Regularly perform full-condition field tests and verifications. When the accuracy drops below a threshold, trigger a special fine-tuning. The optimized model is then applied in reverse to the aforementioned stages.

9. The wind turbine fire causal reasoning early warning method based on multi-source evidence fusion according to claim 8, characterized in that, The mobile carrier for collecting visual three-dimensional spatial evidence is a rail-guided robot.

10. The wind turbine fire causal reasoning early warning method based on multi-source evidence fusion according to claim 9, characterized in that, Outlier cleanup for time-series evidence uses the DBSCAN-MAD algorithm.

Images