Wind power crane fault prediction method and system based on lstm and multi-modal data fusion

Abstract

This invention discloses a fault prediction method and system for wind turbine cranes based on LSTM and multimodal data fusion, belonging to the field of industrial intelligent diagnostic technology. The method includes: collecting multimodal data such as vibration, temperature, current, and load signals; extracting time-frequency features from the vibration signal through wavelet packet decomposition, and forming fusion features by combining multi-source signal statistical features; calculating machine learning anomalies and statistical anomalies using isolated forest and Z-score respectively, and obtaining a weighted fusion composite risk score; dynamically updating the early warning threshold based on fault-free historical data, triggering an early warning if the threshold is exceeded; the system includes a multi-source sensor data acquisition module, a time-frequency-statistical feature fusion engine, an adaptive threshold decision unit, and an ONNX model export interface. This invention can accurately identify weak and complex faults, reduce false alarms and missed alarms, and the model is lightweight and easily deployed at the edge, meeting the real-time, reliable, and low-latency intelligent early warning requirements of wind turbine cranes.

Description

Technical Field

[0001] This invention relates to the field of intelligent fault diagnosis, industrial Internet of Things and multimodal data analysis, and in particular to a fault prediction method and system for wind power cranes based on LSTM and multimodal data fusion. Background Technology

[0002] Wind power cranes (such as wind turbine installation vessels and crawler cranes) are key heavy equipment supporting the construction of offshore and onshore wind farms, operating under complex and harsh conditions such as high wind loads, strong vibrations, and variable loads for extended periods. Their critical components (such as transmission systems, hydraulic systems, and electrical systems) are prone to progressive, sudden, or intermittent compound failures due to mechanical fatigue, environmental corrosion, and other factors, posing a serious threat to operational safety and project progress. Traditional fault early warning systems mainly rely on rule-based threshold judgments or periodic offline maintenance, which are ill-suited to the diversity of fault modes and the strong coupling of operational disturbances. With the development of Industrial Internet of Things (IIoT) and artificial intelligence technologies, integrating multi-source sensor data and utilizing algorithms such as deep learning to achieve intelligent prediction and early warning of equipment health status has become an important research direction for improving the reliability of heavy equipment operation and maintenance and realizing predictive maintenance.

[0003] Currently, typical technical solutions in this field mainly include the following categories:

[0004] (1) Alarm method based on fixed threshold: For example, Chinese invention patent application with publication number CN109857032A discloses a power distribution construction safety monitoring system, which monitors a single feature such as the root mean square value (RMS) of vibration signal and sets a static safety threshold to judge the fault; when the signal feature exceeds the preset threshold, an alarm is triggered.

[0005] (2) Analysis methods based on single-mode signals: Some existing technologies focus on using a single type of sensor signal (such as motor current) for fault diagnosis, and identify specific types of motor faults by analyzing features such as current harmonics.

[0006] (3) Diagnostic systems based on complex cloud models: In existing research, deep learning models such as Long Short-Term Memory Networks (LSTM) and Transformer are often used for fault prediction. However, such models usually have high computational complexity and rely on cloud servers for centralized training and inference. As pointed out in the relevant literature (IEEE Internet of Things Journal, 2023), it is difficult to deploy them efficiently on resource-constrained edge devices.

[0007] Although the above technologies have been applied to some extent, they still have significant limitations:

[0008] (1) The static threshold model has poor adaptability: the fixed threshold cannot reflect the impact of dynamic operating parameters such as wind speed and load on the normal operating status of the equipment. Under high load or high wind speed conditions, the normal vibration level is already increased, which can easily lead to false alarms; while under light load or low wind speed conditions, it may not be sensitive to early minor faults, leading to missed alarms; its mathematical model is too simplified and does not introduce a dynamic adjustment mechanism that couples operating conditions.

[0009] (2) Single-mode analysis is incomplete and has weak robustness: Relying on data from only a single type of sensor (such as vibration or current only) limits the information dimensions and makes it impossible to fully characterize the overall health status of complex mechanical systems. It is not capable of diagnosing faults that are not sensitive to the sensor (such as loose structural components or bearing wear), and is easily affected by sudden changes in specific operating conditions, resulting in poor system robustness.

[0010] (3) Rigid model deployment and insufficient real-time performance: The complex deep learning model that relies on cloud deployment has the disadvantages of high inference latency and strong network dependence, which makes it difficult to meet the urgent need for millisecond-level real-time early warning for mobile equipment such as wind power cranes in remote or network-free areas; at the same time, model updates and maintenance are inconvenient and the edge side lacks autonomous intelligence capabilities.

[0011] Therefore, there is an urgent need to develop a fault prediction system and method that can deeply integrate multimodal information, adapt to dynamic operating conditions, and support lightweight edge deployment, in order to overcome the shortcomings of existing technologies. Summary of the Invention

[0012] This invention addresses the technical challenges in wind turbine crane fault prediction, such as weak static threshold adaptability, one-sided single-modal information, and difficulty in deploying complex models at the edge. It proposes a fault prediction method and system for wind turbine cranes based on LSTM and multimodal data fusion. This method achieves adaptive perception of complex operating conditions through deep fusion of multimodal information and hybrid intelligent analysis, improving fault prediction accuracy and generalization ability. Simultaneously, it maintains lightweight models, enabling efficient deployment on edge devices and providing reliable technical support for the safe operation and intelligent maintenance of wind turbine cranes.

[0013] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0014] A fault prediction method for wind turbine cranes based on LSTM and multimodal data fusion includes the following steps:

[0015] Step 1: Collect real-time data from multi-modal sensors during the operation of the wind turbine crane. The data includes at least vibration signals, temperature signals, current signals, and load signals.

[0016] Step 2: Perform wavelet packet decomposition on the acquired vibration signal and extract the energy of each sub-band as time-frequency features; at the same time, calculate the statistical features of all types of sensor signals within a sliding window; fuse the time-frequency features and statistical features into a fused feature.

[0017] Step 3: Input the fused features into the isolated forest model and calculate the machine learning anomaly score; based on the real-time values ​​of each sensor signal, calculate their Z-score and take the average value to obtain the statistical anomaly score;

[0018] Step 4: Use weighted fusion to fuse the machine learning anomaly and the statistical anomaly to generate a composite risk score;

[0019] Step 5: Based on the historical composite risk score of the wind turbine crane under recent fault-free conditions, dynamically calculate and update the fault warning threshold to obtain the dynamic warning threshold;

[0020] Step 6: Compare the composite risk score calculated in real time with the dynamic early warning threshold. If the composite risk score exceeds the dynamic early warning threshold, a fault warning is triggered.

[0021] A further improvement to the technical solution of the present invention is that: in step 2, the wavelet packet decomposition specifically uses the Daubechies 4th order wavelet as the mother wavelet, and sets the number of decomposition levels. The vibration signal is divided into 16 sub-bands.

[0022] A further improvement to the technical solution of this invention lies in: in step 2, the sliding window length used to calculate the statistical features is... There are 1 sampling points, among which .

[0023] A further improvement of the technical solution of the present invention is that, in step 2, the statistical features include, but are not limited to, mean, variance, kurtosis and kurtosis.

[0024] A further improvement to the technical solution of the present invention is that: in step 3, the trained isolated forest model and feature calculation logic are exported as a standard model file through the ONNX framework to support deployment and inference on edge computing devices.

[0025] A further improvement to the technical solution of the present invention is that, in step 4, the calculation formula for the composite risk score is as follows:

[0026]

[0027] In the formula, For composite risk scores; For machine learning anomaly degree; For statistical outlier; is the weight, and is an adjustable parameter, with an initial value set to 0.6;

[0028] The formula for calculating statistical outlier is as follows:

[0029]

[0030] In the formula, are the mean and standard deviation of the m-th type sensor under normal operating conditions, respectively; For sensor signals; The number of sensors; This represents the number of sensor types.

[0031] The formula for calculating machine learning anomaly is as follows:

[0032]

[0033] In the formula, These are outlier scores in isolated forests; It is a sample Path length; Given a number of samples The average path length is used for normalization, making... The closer the anomaly score is to 1, the more abnormal the sample is.

[0034] A further improvement of the technical solution of the present invention is that the weight w can be dynamically adjusted according to the optimized performance indicators on the historical verification dataset.

[0035] A further improvement to the technical solution of the present invention is that, in step 5, the dynamic early warning threshold is calculated using the following formula:

[0036]

[0037] In the formula, The threshold is a dynamic early warning threshold; This is the mean of the composite risk scores; The standard deviation of the composite risk score; This is the sensitivity coefficient, default value. .

[0038] A further improvement to the technical solution of this invention is that, in step 6, the final early warning decision logic is as follows:

[0039] .

[0040] A fault prediction system for wind turbine cranes based on LSTM and multimodal data fusion includes:

[0041] A multi-source sensor data acquisition module is configured to collect multi-modal operating data of key components of a wind turbine crane in real time. The data includes at least vibration signals, temperature signals, current signals, and load signals.

[0042] The time-frequency-statistical feature fusion engine is communicatively connected to the multi-source sensor data acquisition module. It is configured to perform wavelet packet decomposition on the vibration signal to extract time-frequency features, calculate the statistical features of all sensor signals within a sliding window, and fuse the time-frequency features and statistical features into a fused feature.

[0043] An adaptive threshold decision unit, communicatively connected to the time-frequency-statistical feature fusion engine, is configured to receive the fused features and perform the following operations: inputting the fused features into a preset isolated forest model to calculate machine learning anomaly, while simultaneously calculating statistical anomaly based on real-time data from sensor signals; weighting and fusing the machine learning anomaly and statistical anomaly to generate a composite risk score; dynamically calculating and updating a dynamic early warning threshold based on historical fault-free data; and comparing the composite risk score with the dynamic early warning threshold, generating and outputting a fault early warning signal when the composite risk score exceeds the dynamic early warning threshold.

[0044] The ONNX model export interface, integrated with the adaptive threshold decision unit, is configured to export the isolated forest model and feature calculation logic as a standard ONNX format file to support cross-platform deployment and efficient inference on edge computing devices.

[0045] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows:

[0046] 1. This invention constructs a high-dimensional, complementary feature representation by fusing the time-frequency characteristics (wavelet packet energy) of vibration signals with the statistical characteristics of multi-source signals. This comprehensively and precisely characterizes the health status of equipment, significantly improving the detection sensitivity for early, subtle, and complex faults. Simultaneously, it innovatively introduces a weighted composite risk estimation mechanism based on statistical anomalies and machine learning anomalies, and combines this with adaptive dynamic early warning thresholds set according to operating parameters such as wind speed and load. This enables the system to effectively distinguish between signal fluctuations caused by changes in normal operating conditions and actual fault precursors, thereby significantly reducing false alarm and false negative rates in complex and variable operating environments, achieving highly robust and accurate early warning.

[0047] 2. This invention designs a lightweight workflow from feature extraction and model inference to decision output, and exports the core model through the ONNX standardized interface, enabling it to be deployed directly on resource-constrained edge computing devices (such as industrial gateways and embedded controllers) without complex cloud dependencies. This not only achieves millisecond-level local real-time fault prediction and alarm, meeting the stringent requirements of low latency and high reliability in harsh operating scenarios such as high-altitude, offshore, and mobile environments, but also reduces the system's dependence on network continuity, reducing data transmission and cloud computing costs. Furthermore, the modular system architecture and standardized model format facilitate algorithm iteration and rapid migration across different hardware platforms, greatly improving the engineering implementation capabilities and operational efficiency of the entire solution. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of a wind turbine crane fault prediction method based on LSTM and multimodal data fusion provided by the present invention.

[0050] Figure 2 This is a timing diagram of the multimodal sensor signal provided in an embodiment of the present invention;

[0051] Figure 3 This is a histogram of the composite risk score distribution provided in this embodiment of the invention;

[0052] Figure 4 This is a composite risk score and dynamic early warning threshold analysis diagram provided in the embodiments of the present invention;

[0053] Figure 5 This is a comparison chart of the actual fault state and the predicted result provided in the embodiments of the present invention;

[0054] Figure 6 This is a radar chart for multi-dimensional evaluation of model performance provided in this embodiment of the invention. Detailed Implementation

[0055] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0056] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0057] like Figure 1 As shown, a fault prediction method for wind turbine cranes based on LSTM and multimodal data fusion includes the following:

[0058] (I) Establishment and Derivation of a Fault Prediction Model for Wind Turbine Cranes Based on LSTM and Multimodal Data Fusion

[0059] 1.1 Multimodal data preprocessing and feature extraction, such as Figure 2 As shown.

[0060] Let the sampling frequency be The original vibration signal is .

[0061] (1) Wavelet Packet Decomposition (WPD): The Daubechies 4th order wavelet (db4) is selected as the mother wavelet. For the signal conduct Layer WPD, obtained Individual frequency band coefficient Where j is the decomposition level, This is the node index for this layer.

[0062] The frequency range corresponding to the kth sub-band is:

[0063]

[0064] (2) Frequency band energy calculation: Energy of each sub-band Defined as the sum of squares of its coefficients:

[0065]

[0066] in, It is the first The coefficient length of each sub-band. Energy vector.

[0067] It captures the energy distribution characteristics of signals in different frequency bands and is sensitive to local faults (such as bearing defects).

[0068] (3) Construction of statistical feature vectors: For all Sensor signals (M represents the known types and number of sensors) This refers to four categories: vibration (v), temperature (T), current (I), and load (L). In a length of (For example Within a sliding window of (number of sampling points), calculate the following four second-order statistics:

[0069]

[0070] In the formula, The mean; For variance; For kurtosis; For raucousness;

[0071] Statistical eigenvectors It describes the overall distribution characteristics of the signal.

[0072] 1.2 Multimodal Risk Quantification

[0073] (1) Statistical outlier measurement : at time step Calculate the current The absolute values ​​of the Z-scores of each sensor (referring only to categories here) relative to their historical normal states are averaged. Let... The mean and standard deviation of the m-th type of sensor under normal operating conditions (calculated offline and stored using historical data):

[0074]

[0075] (2) Machine learning anomaly measurement : The fused feature vector Input an Isolation Forest model. This model partitions the data space using a random hyperplane and calculates the samples. Path length Outlier scores in isolated forests Defined as:

[0076]

[0077] in, Given a number of samples The average path length is used for normalization, making... The closer the anomaly score is to 1, the more anomalous the sample is.

[0078] 1.3 Composite Risk Function and Dynamic Early Warning Threshold

[0079] (1) Composite risk function: The core output of this invention is a composite risk score. It is a weighted sum of the two anomalies mentioned above:

[0080]

[0081] Among them, weight The value can be dynamically adjusted based on the performance of historical validation sets, with the initial value set to... To balance the interpretability (statistical model) and generalization ability (ML model) of the model, such as Figure 3 As shown

[0082] (2) Dynamic early warning threshold mechanism: To overcome the limitations of fixed thresholds To address the limitations of [previous inventions], this invention introduces a dynamic early warning threshold. This dynamic early warning threshold is calculated based on recent (e.g., the past 24 hours) fault-free data. mean with standard deviation :

[0083]

[0084] in, Sensitivity coefficient (default) This mechanism ensures that the dynamic early warning threshold can adaptively drift with the equipment's operating status and environmental conditions, effectively suppressing "false anomalies" caused by changes in operating conditions.

[0085] (3) Early warning decision-making: The final early warning decision-making logic is as follows:

[0086]

[0087] (II) System Initialization and Data Acquisition

[0088] 2.1 Hardware and Sensor Deployment

[0089] (1) Deploy sensors on key components of wind turbine cranes:

[0090] Vibration sensors (such as accelerometers): mounted on the bearing housings of the main winch, slewing bearing, and luffing mechanism; sampling frequency... It is used to capture mechanical vibration signals.

[0091] Temperature sensors (such as PT100): installed in critical components such as the main motor, gearbox bearings, and hydraulic oil tank, with a sampling frequency of... It is used to monitor temperature rise.

[0092] Current sensor (such as Hall effect sensor): installed at the main motor inlet, sampling frequency... It is used to monitor motor load.

[0093] Load sensors (such as pressure sensors): installed at the main hook wire rope or boom hinge point, sampling frequency...

[0094] It is used to monitor the real-time load.

[0095] (2) Connect all sensors to the edge computing unit (such as NVIDIA Jetson series) via industrial bus (such as CAN, Profinet) or wireless means (such as LoRa).

[0096] 2.2 Historical Data Collection and Model Training

[0097] (1) Continuously collect equipment operating data under known health conditions. The recommended time span is 1-3 months to ensure coverage of different seasons and different operating conditions (wind speed, load).

[0098] (2) Preprocess the collected data, including noise reduction (such as low-pass filtering), normalization, and removal of outliers.

[0099] (3) Calculate the statistical characteristics (mean) of historical health data. and standard deviation And store it in a local database.

[0100] (4) Train the Isolation Forest model using the preprocessed health data, and set the parameters:

[0101]

[0102] (III) Real-time Fault Prediction Process

[0103] 3.1 Feature Extraction and Fusion

[0104] (1) Data acquisition: Edge computing units use The frequency of acquiring four types of sensor data at the current time t: .

[0105] (2) Time-frequency feature extraction: The latest vibration signal segment (e.g., 10240 points in the past 1 second) is decomposed into 4-level wavelet packet decomposition (db4 basis) and the energy of 16 sub-bands is calculated. The time-frequency feature vector is obtained. .

[0106] (3) Statistical feature extraction: For the four types of signals with 1024 sampling points (sliding window) including the current time, the mean, variance, kurtosis and kurtosis are calculated respectively to obtain the statistical feature vector. .

[0107] (4) Feature fusion: The time-frequency features and statistical features are concatenated to obtain the fused feature vector:

[0108]

[0109] 3.2 Anomaly Calculation

[0110] (1) Statistical abnormality Use formula calculate.

[0111] (2) Machine learning anomaly : Fuse feature vectors Input the trained Isolation Forest model to obtain anomaly scores. .

[0112] 3.3 Calculation of Composite Risk and Dynamic Threshold

[0113] (1) Composite risk Use formula Calculate the composite risk score at the current moment.

[0114] (2) Dynamic early warning threshold Maintain a storage system in a fault-free state for the last 24 hours. A sliding window of values. All values ​​within this window are calculated in real time. mean of values and standard deviation Using the formula Update dynamic warning thresholds, such as Figure 4 As shown.

[0115] 3.4 Fault Early Warning Decision and Output

[0116] (1) Decision-making: comparison and .

[0117] (2) Early warning output:

[0118] like If the signal is positive, a "high risk" warning signal will be triggered (output is 1), and the signal will be sent to the remote monitoring center via the HMI interface, audible and visual alarm, or network.

[0119] At the same time, the system can record , each component , and trigger threshold This is used for subsequent analysis.

[0120] like If the signal is not found, a "normal" signal will be output (output is 0).

[0121] (3) ONNX Model Export: After the system is running stably, the isolated forest model and feature calculation logic (implemented using scikit-learn and numpy in Python) are exported through the ONNX framework. Exporting the library as This file can be efficiently loaded and inferred on resource-constrained edge devices without requiring a Python environment.

[0122] (iv) Analysis of typical application scenarios

[0123] 4.1 Setting up the work scenario

[0124] Imagine a large wind turbine installation vessel anchored 50 kilometers offshore, its main crane performing tower installation. The wind speed is 10 m / s, and the crane is loaded with 80 tons. Under these harsh conditions of high winds and heavy loads, critical components such as the crane's slewing bearing and main hoisting mechanism are subjected to immense stress, making them highly susceptible to early fatigue damage. The goal is to provide early warning of potential mechanical failures while ensuring the continuity of the installation operation, thereby avoiding catastrophic accidents and costly unplanned downtime.

[0125] 4.2 Multimodal Data Fusion Analysis

[0126] Traditional methods may only monitor the total vibration. This invention, however, processes four signals simultaneously. For example, when a slewing bearing exhibits early wear, it is characterized by: vibration energy E in a specific frequency band (e.g., 200-400Hz). j A significant increase in temperature (caused by impact) occurs; simultaneously, due to increased friction, the local temperature ST slowly rises; the current SI waveform of the drive motor may exhibit slight distortion; and these changes may not be noticeable on the total load SL. By fusing these four types of information, the system can detect fault symptoms earlier and more accurately.

[0127] 4.3 Verification of Dynamic Threshold Adaptive Mechanism

[0128] Under calm sea conditions (wind speed 2 m / s, load 20 tons), the equipment's normal composite risk score. The average value is approximately 0.2. At this point, if a fixed threshold of 0.7 is used, the system is very robust. However, when entering operational conditions (wind speed 10 m / s, load 80 tons), even if the equipment is in perfect working order, The threshold value can also rise to around 0.5 due to increased normal vibration. Using a fixed threshold of 0.7 would lead to numerous false alarms and disrupt operations. The dynamic threshold mechanism of this invention automatically adjusts the threshold value. From 0.7 to approximately This effectively avoids interference caused by changes in operating conditions and maintains the specificity of the early warning system, such as... Figure 5 As shown.

[0129] 4.4 ONNX Model Deployment and Real-Time Performance

[0130] To enable the algorithm to run in real time on the shipboard industrial control computer, the model was exported using ONNX. The exported model has the following advantages, such as... Figure 6 As shown:

[0131] Input: A time window (e.g., 10 seconds) containing vibration, temperature, current, and load data.

[0132] Output: Composite risk score at the current moment And binary early warning signals.

[0133] Reasoning speed: Milliseconds (on NVIDIA Jetson AGX Orin).

[0134] Easy deployment: No Python environment required, usable Direct call.

[0135] This means that the system can truly achieve a closed loop of "perception-analysis-early warning", providing a solid guarantee for the safety of maritime operations.

[0136] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fault prediction method for wind turbine cranes based on LSTM and multimodal data fusion, characterized in that, Includes the following steps: Step 1: Collect real-time data from multi-modal sensors during the operation of the wind turbine crane. The data includes at least vibration signals, temperature signals, current signals, and load signals. Step 2: Perform wavelet packet decomposition on the acquired vibration signal and extract the energy of each sub-band as time-frequency features; at the same time, calculate the statistical features of all types of sensor signals within a sliding window; fuse the time-frequency features and statistical features into a fused feature. Step 3: Input the fused features into the isolated forest model and calculate the machine learning anomaly score; based on the real-time values ​​of each sensor signal, calculate their Z-score and take the average value to obtain the statistical anomaly score; Step 4: Use weighted fusion to fuse the machine learning anomaly and the statistical anomaly to generate a composite risk score; Step 5: Based on the historical composite risk score of the wind turbine crane under recent fault-free conditions, dynamically calculate and update the fault warning threshold to obtain the dynamic warning threshold; Step 6: Compare the composite risk score calculated in real time with the dynamic early warning threshold. If the composite risk score exceeds the dynamic early warning threshold, a fault warning is triggered.

2. The method according to claim 1, characterized in that, In step 2, the wavelet packet decomposition specifically uses the Daubechies 4th order wavelet as the mother wavelet, and sets the number of decomposition levels. The vibration signal is divided into 16 sub-bands.

3. The method according to claim 1, characterized in that, In step 2, the sliding window length used to calculate the statistical features is... There are 1 sampling points, among which .

4. The method according to claim 1, characterized in that, In step 2, the statistical features include, but are not limited to, mean, variance, kurtosis, and kurtosis.

5. The method according to claim 1, characterized in that, Step 3 also includes: exporting the trained isolated forest model and feature calculation logic as a standard model file through the ONNX framework to support deployment and inference on edge computing devices.

6. The method according to claim 1, characterized in that, In step 4, the formula for calculating the composite risk score is as follows: In the formula, For composite risk scores; For machine learning anomaly degree; For statistical outlier; is the weight, and is an adjustable parameter, with an initial value set to 0.6; The formula for calculating statistical outlier is as follows: In the formula, are the mean and standard deviation of the m-th type sensor under normal operating conditions, respectively; is the sensor signal; M is the number of sensors; m is the number of sensor types; The formula for calculating machine learning anomaly is as follows: In the formula, These are outlier scores in isolated forests; It is a sample Path length; Given a number of samples The average path length is used for normalization, making... The closer the anomaly score is to 1, the more abnormal the sample is.

7. The method according to claim 6, characterized in that, The weight w can be dynamically adjusted based on the performance metrics optimized on the historical validation dataset.

8. The method according to claim 1, characterized in that, In step 5, the dynamic early warning threshold is calculated using the following formula: In the formula, Dynamic early warning threshold; This is the mean of the composite risk scores; The standard deviation of the composite risk score; This is the sensitivity coefficient, default value. .

9. The method according to claim 8, characterized in that, In step 6, the final early warning decision logic is as follows: 。 10. A system for implementing the method according to any one of claims 1-9, characterized in that, include: A multi-source sensor data acquisition module is configured to collect multi-modal operating data of key components of a wind turbine crane in real time. The data includes at least vibration signals, temperature signals, current signals, and load signals. The time-frequency-statistical feature fusion engine is communicatively connected to the multi-source sensor data acquisition module. It is configured to perform wavelet packet decomposition on the vibration signal to extract time-frequency features, calculate the statistical features of all sensor signals within a sliding window, and fuse the time-frequency features and statistical features into a fused feature. An adaptive threshold decision unit, communicatively connected to the time-frequency-statistical feature fusion engine, is configured to receive the fused features and perform the following operations: inputting the fused features into a preset isolated forest model to calculate machine learning anomaly, while simultaneously calculating statistical anomaly based on real-time data from sensor signals; weighting and fusing the machine learning anomaly and statistical anomaly to generate a composite risk score; dynamically calculating and updating a dynamic early warning threshold based on historical fault-free data; and comparing the composite risk score with the dynamic early warning threshold, generating and outputting a fault early warning signal when the composite risk score exceeds the dynamic early warning threshold. The ONNX model export interface, integrated with the adaptive threshold decision unit, is configured to export the isolated forest model and feature calculation logic as a standard ONNX format file to support cross-platform deployment and efficient inference on edge computing devices.

Images