Dynamic early warning method for specific fault of wind turbine generator and related device

Abstract

The invention discloses a dynamic early warning method for a specific fault of a wind turbine generator, and relates to the field of state monitoring of wind power generation equipment, and the method comprises the following steps: collecting real-time vibration data and real-time operation condition data of the wind turbine generator; extracting vibration depth features from the real-time vibration data by using a pre-trained depth feature extraction model; inputting the real-time operation condition data into a pre-trained health state prediction model to obtain a predicted value of the enhanced specificity index under the current operation condition, calculating a confidence interval according to the predicted value, and taking the confidence interval as a dynamic threshold range; when an early warning signal is triggered when the actual value of the enhanced specificity index exceeds a dynamic threshold range, the early warning signal is triggered; the enhanced specificity index comprises the vibration depth characteristic and a preset traditional specificity index. According to the invention, the problems of single feature representation and fixed early warning threshold in the prior art can be solved.

Description

Technical Field

[0001] This invention relates to the field of condition monitoring of wind power generation equipment, specifically to a dynamic early warning method and related device for specific faults of wind turbine units. Background Technology

[0002] Currently, in the field of condition monitoring and fault prediction for wind power equipment, the operational reliability monitoring of wind turbine units mainly relies on vibration-based condition monitoring systems. These systems typically collect vibration signals from the turbine, extract traditional characteristic indicators such as time and frequency domain parameters, and set fixed thresholds based on industry standards or experience to issue early warnings of anomalies, thereby achieving a preliminary assessment of the equipment's health status and preventing faults.

[0003] However, due to variations in turbine model, installation location, and operating conditions, wind turbines exhibit significant individual differences in their dynamic characteristics, making it difficult to universally apply fixed thresholds. This often leads to false alarms or missed alarms in early warning systems, resulting in insufficient reliability. Secondly, relying on limited traditional indicators such as RMS values, peak values, and spectrum analysis has weak characterization capabilities for early, subtle, and complex faults, making accurate fault location and type identification difficult. Furthermore, multi-source monitoring information such as vibration data, SCADA operating parameters, and acoustic signals is often stored and analyzed independently in practical applications, resulting in fragmented data that cannot achieve effective cross-validation and information complementarity. This creates information silos, limiting the depth and completeness of fault diagnosis. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic early warning method and related device for specific faults of wind turbine units, so as to solve the problems of single feature representation and fixed early warning threshold in the prior art.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a dynamic early warning method for specific faults of wind turbine units includes the following steps: Collect real-time vibration data and real-time operating condition data of wind turbine units; Vibration depth features are extracted from real-time vibration data using a pre-trained depth feature extraction model. Real-time operating condition data is input into a pre-trained health status prediction model to obtain predicted values ​​of enhanced specific indicators under the current operating conditions; confidence intervals are calculated based on the predicted values, and the confidence intervals are used as dynamic threshold ranges. When the actual value of the enhanced specificity indicator exceeds the dynamic threshold range, an early warning signal is triggered; the enhanced specificity indicator includes the vibration depth feature and the preset traditional specificity indicator.

[0006] In some implementations, after the warning signal is triggered, the following steps are also included: Real-time vibration data is input into a convolutional neural network for feature extraction to obtain a vibration feature vector. Real-time operating condition data is input into a long short-term memory network for feature extraction to obtain a SCADA feature vector. Based on an attention mechanism, the vibration feature vector and the SCADA feature vector are fused and classified to obtain the fault classification result and its confidence level, thus completing the fault diagnosis. By inputting the time series of enhanced specific indicators into a temporal convolutional network, the remaining service life of the wind turbine is predicted, thus completing the life diagnosis.

[0007] In some implementations, the pre-trained deep feature extraction model is trained through the following steps: Historical vibration data of wind turbines are acquired and used as the original input signal. The one-dimensional convolutional autoencoder model is trained unsupervised by minimizing the reconstruction loss. After the unsupervised training is completed, a pre-trained one-dimensional convolutional autoencoder model is obtained, and the encoder in the pre-trained one-dimensional convolutional autoencoder model is used as a pre-trained deep feature extraction model.

[0008] In some implementations, the pre-trained health status prediction model is trained through the following steps: Historical operating condition data of wind turbines are acquired. Using the historical operating condition data as input and the actual value of the enhanced specificity index as output, a long short-term memory network is trained in a supervised manner. After the supervised training is completed, the historical operating condition data is input into the long short-term memory network to obtain the predicted value of the enhanced specificity index. The residual between the predicted value and the actual value of the enhanced specificity index is calculated, and the confidence interval offset is determined according to the statistical distribution of the residual, thus completing the pre-training of the health status prediction model.

[0009] In some implementations, confidence intervals are calculated based on predicted values, specifically including: The confidence interval is obtained by adding and subtracting the predicted value from the confidence interval offset.

[0010] Secondly, a dynamic early warning system for wind turbine specific faults includes: The multi-source heterogeneous data acquisition module is used to collect real-time vibration data and real-time operating condition data of wind turbine units; The vibration depth feature extraction module is used to extract vibration depth features from real-time vibration data using a pre-trained depth feature extraction model. The adaptive threshold learning module is used to input real-time operating condition data into the pre-trained health status prediction model to obtain the predicted value of the enhanced specific index under the current operating condition, calculate the confidence interval based on the predicted value, and use the confidence interval as the dynamic threshold range. The dynamic early warning module is used to trigger an early warning signal when the actual value of the enhanced specific indicator exceeds the dynamic threshold range. The enhanced specific indicator includes the vibration depth feature and the preset traditional specific indicator.

[0011] In some implementations, it also includes: The fault diagnosis module is used to input real-time vibration data into a convolutional neural network for feature extraction to obtain vibration feature vectors, and input real-time operating condition data into a long short-term memory network for feature extraction to obtain SCADA feature vectors. Based on the attention mechanism, the vibration feature vectors and SCADA feature vectors are fused and classified to obtain fault classification results and their confidence levels, thus completing fault diagnosis. The remaining useful life prediction module is used to input the time series of enhanced specific indicators into a temporal convolutional network to predict the remaining useful life of the wind turbine and complete the life diagnosis.

[0012] Thirdly, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable in the processor, wherein the processor executes the computer program to implement a dynamic early warning method for wind turbine-specific faults.

[0013] Fourthly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements a dynamic early warning method for wind turbine-specific faults.

[0014] Fifthly, a computer program product, comprising a computer program, which, when executed by a processor, implements a dynamic early warning method for wind turbine-specific faults.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a dynamic early warning method for specific faults in wind turbine units. It extracts vibration depth features using a pre-trained deep feature extraction model and combines these features with preset traditional specific indicators to form enhanced specific indicators. This effectively integrates automatically learned deep features with expert experience, thereby enriching and enhancing the representational dimensions and capabilities of fault features, overcoming the shortcomings of traditional methods with single indicators. By inputting real-time operating condition data into a pre-trained health status prediction model, a dynamic threshold range matching the current operating condition is generated, and early warning judgments are made based on this range. This allows the early warning threshold to adapt to individual differences in the unit and changes in real-time operating status, solving the problem of poor adaptability of fixed early warning thresholds, which easily leads to false alarms or missed alarms, and achieving accurate and adaptive dynamic early warning.

[0016] Furthermore, after triggering the early warning, feature vectors from real-time vibration data and real-time operating condition data are extracted using convolutional neural networks and long short-term memory networks, respectively. These vectors are then fused and classified based on an attention mechanism. This allows for the comprehensive utilization of complementary information from real-time vibration data and real-time operating condition data, enabling intelligent and refined diagnosis of fault types and components, thus improving the accuracy and reliability of diagnostic results. Simultaneously, by inputting the time series data of enhanced specific indicators into a temporal convolutional network to predict remaining service life, the health degradation trend of components can be quantitatively assessed, providing a direct decision-making basis for implementing predictive maintenance. Attached Figure Description

[0017] Figure 1 A flowchart of a dynamic early warning method for specific faults of wind turbine units provided in an embodiment of the present invention; Figure 2 This is a structural diagram of a dynamic early warning system for specific faults of wind turbine generators provided in an embodiment of the present invention; Figure 3 This invention provides an overall architecture block diagram of a dynamic early warning system for specific faults in wind turbine units. Figure 4 This is a flowchart illustrating the model-specific index library and vibration depth feature extraction provided in this embodiment of the invention. Figure 5 A schematic diagram illustrating the adaptive threshold learning principle based on digital twins provided in an embodiment of the present invention; Figure 6 This is a structural diagram of the multi-source information fusion intelligent diagnostic model provided in an embodiment of the present invention; Detailed Implementation To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. The content is for explanation rather than limitation of the present invention.

[0018] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification and claims of this invention are intended to cover a 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, systems, products, or devices.

[0019] like Figure 1 and Figure 3 As shown, this embodiment provides a dynamic early warning method for specific faults in wind turbine generators, including the following steps: S1 collects real-time vibration data and real-time operating condition data of the wind turbine; Specifically, the system synchronously collects real-time vibration data, real-time operating condition data, and optional acoustic signals from the target wind turbine. Vibration data originates from multi-channel accelerometers installed on key components, such as the main bearing, gearbox input / output shafts, and generator bearings. The raw waveforms are acquired by the Condition Monitoring System (CMS) at a sampling rate of no less than 25.6 kHz. Operating condition data comes from the Supervisory Control and Data Acquisition (SCADA) system, including wind speed, generator speed, unit active power, gearbox oil temperature, and generator bearing temperature, typically acquired as 1-minute averages. At the edge or cloud platform, vibration signals undergo preprocessing such as denoising, resampling, and data alignment. SCADA data is cleaned and normalized to ensure that multi-source data are aligned in timestamps, laying the foundation for subsequent fusion analysis.

[0020] S2, using a pre-trained deep feature extraction model to extract vibration depth features from the real-time vibration data, the vibration depth features and the preset traditional specificity index together constitute the enhanced specificity index; Specifically, such as Figure 4 As shown, based on a pre-defined library of specific indicators for different transmission chain structures, such as doubly fed, semi-direct drive, direct drive, and different components, real-time vibration data is calculated to obtain a set of pre-defined traditional specific indicators. This indicator library includes time-domain indicators (RMS, peak value, kurtosis, waveform factor), frequency-domain indicators (1st, 2nd, and 3rd harmonic amplitudes at power frequency, gear meshing frequency and its sideband amplitudes, bearing fault characteristic frequency amplitudes), and time-frequency-domain indicators (such as wavelet packet decomposition energy entropy). Simultaneously, real-time vibration data segments from the same time period are input into the encoder part of a pre-trained one-dimensional convolutional autoencoder (1D-CAE). This encoder outputs a low-dimensional numerical vector, which is the vibration depth feature automatically learned from the original signal. This set of depth features, together with the calculated traditional specific indicators, constitutes an enhanced set of specific indicators.

[0021] The training process of this deep feature extraction model is as follows: In the offline phase before system deployment, long-term historical vibration data of the target wind turbine under known healthy conditions are collected. A one-dimensional convolutional autoencoder model is constructed, consisting of an encoder and a decoder. The encoder is composed of stacked one-dimensional convolutional and pooling layers to compress the input signal; the decoder is composed of stacked one-dimensional deconvolutional and upsampling layers to reconstruct the signal. During training, the historical vibration data under healthy conditions is segmented into multiple data fragments and fed into the model as the original input signal. The training objective is to make the reconstructed signal output by the decoder as close as possible to the original input signal, optimizing the model parameters by minimizing the reconstruction loss between the two, such as the mean squared error (MSE). Once the model training is complete, a pre-trained one-dimensional convolutional autoencoder model is obtained. In the subsequent online application phase, only the encoder part of the model is retained and used; this encoder serves as the pre-trained deep feature extraction model, used to encode new vibration data into low-dimensional vibration depth features in real time.

[0022] S3, input the real-time operating condition data into the pre-trained health status prediction model to obtain the predicted value of the enhanced specificity index under the current operating condition, calculate the confidence interval based on the predicted value, and use the confidence interval as the dynamic threshold range; the health status prediction model can be a time series prediction model built based on architectures such as Long Short-Term Memory Network (LSTM) or Transformer.

[0023] Specifically, such as Figure 5 As shown, the system reads real-time operating condition data and inputs it into a digital twin health status prediction model specifically built for the target unit, such as one trained on a Long Short-Term Memory (LSTM) network. During training, this model learns the complex mapping relationship between various enhanced specific indicators and operating conditions under a healthy state. When used online, the model outputs the predicted normal value for each indicator in the indicator set under the current operating conditions, based on real-time conditions. The system calculates a dynamic confidence interval centered on the predicted value, based on the confidence interval offset determined during model training (i.e., the 3σ standard of the prediction residuals from health data). This confidence interval serves as the dynamic threshold range for each indicator under the current operating conditions, thus enabling adaptive adjustment of the warning threshold according to the individual unit and real-time operating conditions.

[0024] The training process of this health status prediction model is as follows: First, historical operating condition data (SCADA data) and corresponding historical enhanced specific index actual values ​​are collected for the same health period as the training deep feature extraction model. These historical enhanced specific index actual values ​​are calculated using the vibration data for that period, applying the trained deep feature extraction model and a pre-set traditional index library. Then, using the historical operating condition data—a time-series vector composed of power, speed, and temperature—as input features and the corresponding historical enhanced specific index actual values ​​as the target output, a Long Short-Term Memory (LSTM) network is trained in a supervised manner. This allows the LSTM to learn the mapping relationship from operating conditions to normal index values. After training, the basic prediction model is obtained.

[0025] Next, the confidence interval offset is determined, and the trained basic prediction model is evaluated using historical health data: historical operating condition data is input into the model again to obtain a set of predicted values ​​for enhanced specific indicators. The residuals between these predicted values ​​and the corresponding actual values ​​of historical enhanced specific indicators are calculated, and the statistical distribution of the residuals of all training samples is analyzed, calculating their standard deviation σ. Based on a pre-set confidence level requirement, such as 99.73%, the confidence interval offset is determined, for example, taking 3 times the standard deviation of the residuals (3σ) as the offset. This completes the pre-training of the health status prediction model.

[0026] S4, compare the actual value of the enhanced specificity index with the dynamic threshold range, and trigger an early warning signal when the actual value of the enhanced specificity index exceeds the dynamic threshold range to achieve dynamic early warning.

[0027] Specifically, the actual values ​​of the enhanced specificity indicators, calculated and extracted in real time, are compared one by one with the corresponding dynamic threshold ranges generated by the health status prediction model. When the actual value of any indicator exceeds the upper or lower limit of its corresponding dynamic threshold range, the system determines that the indicator is abnormal and triggers an early warning signal of the corresponding level, thereby realizing adaptive dynamic early warning based on real-time operating conditions.

[0028] When the system is running online, the base prediction model (LSTM) in the health status prediction model outputs predicted values ​​of enhanced specific indicators based on real-time operating condition data. For each value in this predicted value sequence, the system reads the corresponding confidence interval offset (e.g., 3σ) determined during the model training phase. The upper limit of the dynamic threshold range is obtained by adding this offset to the predicted value, and the lower limit is obtained by subtracting this offset from the predicted value. Thus, for each enhanced specific indicator, a confidence interval centered on its predicted operating condition value and with a fixed offset is generated in real time. This interval serves as the dynamic threshold range for that indicator at that moment, used for comparison with the actual value.

[0029] Once the warning is triggered, the intelligent diagnostic process is automatically initiated, such as... Figure 6 As shown, firstly, real-time vibration data is extracted for a period of time before and after the warning time, such as 5 minutes before and after, and converted into a spectrogram, i.e., FFT spectrum or envelope demodulation spectrum. This spectrogram is input into a branch of a convolutional neural network (CNN), which automatically extracts fault features from the image and outputs a vibration feature vector. Simultaneously, the system extracts real-time operating condition data (SCADA data) for a period of time before and after the warning time, such as 30 minutes before and after, and inputs it into a branch of a long short-term memory network (LSTM), which extracts time-series operating condition features and outputs a SCADA feature vector. Subsequently, an attention fusion layer is used to concatenate the vibration feature vector and the SCADA feature vector, and the attention mechanism learns the weights of the two in the current diagnostic task to achieve feature fusion. The fused feature vector is finally passed through a fully connected layer and a softmax classifier to output the probability distribution of various faults (such as gear pitting, wear, bearing failure, etc.) and their participating components. The category with the highest probability and its probability value are the fault classification result and its confidence level.

[0030] The system automatically queries and extracts historical data of specific enhancement indicators that trigger warnings, such as the amplitude of gear meshing frequency sidebands, over the past few months, forming a time series. This time series is then input into a pre-trained Temporal Convolutional Network (TCN) model. The TCN model analyzes the degradation trend of this indicator and predicts its trajectory over a future period, such as the next 30 days. The system compares the time point corresponding to the predicted trajectory reaching a preset failure threshold with the current time point to calculate the remaining useful life (RUL) of the component, and outputs the prediction result in days or hours.

[0031] Based on the above method, this paper takes the early pitting fault warning of a doubly-fed wind turbine gearbox as an example to demonstrate the specific application of this method. The vibration acceleration signal of the input shaft bearing housing of the gearbox is collected in real time, along with data from SCADA showing the unit power (1500kW), speed (1200rpm), and gearbox oil temperature (65℃). After preprocessing, the data is calculated using a doubly-fed wind turbine gearbox-specific index library, and deep features are extracted using a 1D-CAE encoder, collectively forming an enhanced specific index set.

[0032] The current operating conditions—unit power 1500kW, speed 1200rpm, gearbox oil temperature 65℃—are input into the unit's dedicated LSTM health model. The model outputs the dynamic threshold ranges for each indicator. The system detects that the traditional indicator, the amplitude of the meshing frequency sideband in the envelope demodulation spectrum, and the third feature in the vibration depth feature vector exceed their dynamic threshold limits, thus triggering a level three warning. The warning message is labeled "Gearbox abnormal, pending diagnosis."

[0033] After the warning is triggered, the vibration spectrum and SCADA time-series data before and after the abnormal moment are input into a two-branch attention fusion network. One branch uses a CNN to process the spectrum, and the other uses an LSTM to process the SCADA data for diagnosis. The model outputs a diagnosis result of "suspected early pitting corrosion of the high-speed shaft gear of the gearbox," with a confidence level of 87%, and the warning level is upgraded to level two. The system then associates this high-confidence diagnosis result with the initial warning and automatically upgrades the warning level from level three (early warning) to level two (confirmed fault).

[0034] Simultaneously, the RUL prediction module retrieves historical data for this abnormal indicator from the past three months, inputs it into a Temporal Convolutional Network (TCN) model, predicts its future trend, estimates the remaining service life to be approximately 45 days, and generates a maintenance recommendation to "schedule an inspection within 30-40 days." The system automatically generates a work order containing the diagnostic results, RUL prediction, and maintenance recommendations, and pushes it to the operations and maintenance personnel. The maintenance recommendation will suggest combining wind speed forecasts to schedule shutdown inspections, achieving predictive maintenance scheduling with minimal power generation loss.

[0035] Throughout the process, the data quality monitoring module runs continuously in the background to ensure the continuity of the data stream used for analysis. Any interruptions are flagged on the dashboard, guaranteeing the reliability of decision-making. After on-site fault confirmation and repair by maintenance personnel, the results can be fed back to the system for continuous model optimization.

[0036] like Figure 2 As shown, this embodiment provides a dynamic early warning system for wind turbine specific faults, including: The multi-source heterogeneous data acquisition module is used to collect real-time vibration data and real-time operating condition data of wind turbine units; The vibration depth feature extraction module is used to extract vibration depth features from the real-time vibration data using a pre-trained depth feature extraction model. The vibration depth features and the preset traditional specificity index together constitute the enhanced specificity index. An adaptive threshold learning module is used to input the real-time operating condition data into a pre-trained health status prediction model to obtain the predicted value of the enhanced specificity index under the current operating condition, calculate the confidence interval based on the predicted value, and use the confidence interval as a dynamic threshold range. The dynamic early warning module is used to compare the actual value of the enhanced specificity indicator with the dynamic threshold range, and trigger an early warning signal when the actual value of the enhanced specificity indicator exceeds the dynamic threshold range, thereby realizing dynamic early warning. The fault diagnosis module is used to input the real-time vibration data into a convolutional neural network for feature extraction to obtain a vibration feature vector, and input the real-time operating condition data into a long short-term memory network for feature extraction to obtain a SCADA feature vector; based on an attention mechanism, the vibration feature vector and the SCADA feature vector are fused and classified to obtain a fault classification result and its confidence level. The remaining service life prediction module is used to input the time series of the enhanced specificity index into a temporal convolutional network to predict the remaining service life of the wind turbine. A high-performance parallel signal analysis engine, developed based on distributed computing frameworks (such as Spark) or utilizing the parallel computing capabilities of GPUs, can rapidly process vibration data from hundreds of units across the entire field in batches and in parallel, strictly adhering to relevant specifications. It supports simultaneous complex calculations such as spectrum analysis, envelope analysis, and order analysis, improving the efficiency of massive data analysis by tens of times and making routine health scans at the entire field possible. The engine integrates advanced signal processing algorithms, such as spectral kurtosis, wavelet packet transform, and empirical mode decomposition (EMD), supporting efficient spectrum analysis, envelope analysis, and order analysis of vibration data from hundreds of units across the entire field.

[0037] The automated data quality monitoring module continuously monitors the "heartbeat" of the CMS data stream for each unit. If the data stream is interrupted, it automatically records the start time, end time, and duration of the interruption and triggers an alarm. The system periodically generates data interruption statistics reports, displaying information such as data availability and interruption ranking for each unit through a visual dashboard. This ensures the integrity and reliability of the monitored data, forming a closed-loop meta-management system for the monitoring system itself.

[0038] The module division in this embodiment of the invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the invention can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0039] This embodiment also provides a computer device, which includes a processor and a memory. The memory is used to store a computer program (in this embodiment, the computer program includes a computing component and an iterative component, capable of model calculation and model updating). The computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the computer storage medium to realize the corresponding method flow or corresponding function. The processor of this embodiment can be used for the operation of a dynamic early warning method for wind turbine specific faults.

[0040] This embodiment also provides a storage medium, specifically a computer-readable storage medium (Memory). A computer-readable storage medium is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the dynamic early warning method for wind turbine specific faults in the above embodiment.

[0041] This embodiment also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the corresponding steps of the dynamic early warning method for wind turbine specific faults in the above embodiment.

[0042] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0043] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0044] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0045] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0046] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A dynamic early warning method for specific faults in wind turbine units, characterized in that, Includes the following steps: Collect real-time vibration data and real-time operating condition data of wind turbine units; Vibration depth features are extracted from the real-time vibration data using a pre-trained depth feature extraction model; The real-time operating condition data is input into a pre-trained health status prediction model to obtain the predicted value of the enhanced specificity index under the current operating condition. Calculate a confidence interval based on the predicted value, and use the confidence interval as a dynamic threshold range; When the actual value of the enhanced specificity indicator exceeds the dynamic threshold range, an early warning signal is triggered; The enhanced specificity index includes the vibration depth feature and the preset traditional specificity index.

2. The method for dynamic early warning of specific faults in wind turbine units according to claim 1, characterized in that, After the warning signal is triggered, the following steps are also included: The real-time vibration data is input into a convolutional neural network for feature extraction to obtain a vibration feature vector. The real-time operating condition data is input into a long short-term memory network for feature extraction to obtain a SCADA feature vector. The vibration feature vector and the SCADA feature vector are fused and classified based on an attention mechanism to obtain the fault classification result and its confidence level, thus completing the fault diagnosis. The time series of the enhanced specificity index is input into a temporal convolutional network to predict the remaining service life of the wind turbine, thus completing the life diagnosis.

3. The method for dynamic early warning of specific faults in wind turbine units according to claim 1, characterized in that, The pre-trained deep feature extraction model is trained through the following steps: Historical vibration data of wind turbines are acquired, and the historical vibration data is used as the original input signal. The one-dimensional convolutional autoencoder model is trained unsupervised by minimizing the reconstruction loss. After the unsupervised training is completed, a pre-trained one-dimensional convolutional autoencoder model is obtained, and the encoder in the pre-trained one-dimensional convolutional autoencoder model is used as a pre-trained deep feature extraction model.

4. The method for dynamic early warning of specific faults in wind turbine units according to claim 1, characterized in that, The pre-trained health status prediction model is trained through the following steps: Historical operating condition data of wind turbine units are acquired. Using the historical operating condition data as input and the actual value of the enhanced specificity index as output, a long short-term memory network is trained in a supervised manner. After the supervised training is completed, the historical operating condition data is input into the long short-term memory network to obtain the predicted value of the enhanced specificity index. The residual between the predicted value and the actual value of the enhanced specificity index is calculated, and the confidence interval offset is determined according to the statistical distribution of the residual, thus completing the pre-training of the health status prediction model.

5. The method for dynamic early warning of specific faults in wind turbine units according to claim 4, characterized in that, The confidence interval is calculated based on the predicted value, specifically including: The predicted value is added to and subtracted from the confidence interval offset to obtain the confidence interval.

6. A dynamic early warning system for specific faults in wind turbine generators, characterized in that, include: The multi-source heterogeneous data acquisition module is used to collect real-time vibration data and real-time operating condition data of wind turbine units; The vibration depth feature extraction module is used to extract vibration depth features from the real-time vibration data using a pre-trained depth feature extraction model. An adaptive threshold learning module is used to input the real-time operating condition data into a pre-trained health status prediction model to obtain the predicted value of the enhanced specificity index under the current operating condition. Calculate a confidence interval based on the predicted value, and use the confidence interval as a dynamic threshold range; The dynamic early warning module is used to trigger an early warning signal when the actual value of the enhanced specific indicator exceeds the dynamic threshold range; The enhanced specificity index includes the vibration depth feature and the preset traditional specificity index.

7. The dynamic early warning system for specific faults of wind turbine units according to claim 6, characterized in that, Also includes: The fault diagnosis module is used to input the real-time vibration data into a convolutional neural network for feature extraction to obtain a vibration feature vector, and input the real-time operating condition data into a long short-term memory network for feature extraction to obtain a SCADA feature vector; based on an attention mechanism, the vibration feature vector and the SCADA feature vector are fused and classified to obtain a fault classification result and its confidence level, thus completing the fault diagnosis. The remaining service life prediction module is used to input the time series of the enhanced specificity index into a temporal convolutional network to predict the remaining service life of the wind turbine and complete the service life diagnosis.

8. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable in the processor. When the processor executes the computer program, it implements the dynamic early warning method for specific faults of wind turbine units as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the dynamic early warning method for specific faults of wind turbine units as described in any one of claims 1 to 6.

10. A computer program product, the computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the dynamic early warning method for specific faults of wind turbine units as described in any one of claims 1 to 6.

Images