SCNN-based early mechanical looseness diagnosis method for wind turbine variable pitch system

Abstract

The application provides a wind turbine variable pitch system early mechanical loosening diagnosis method based on SCNN, belongs to the technical field of mechanical loosening diagnosis, and comprises the following steps: a high-precision vibration sensor is installed on a wind turbine variable pitch system to collect signals; after denoising and segmentation, intrinsic mode functions are obtained by using variational mode decomposition, and multi-scale sample entropy is calculated as a feature; the feature is coded into a pulse sequence by using a hybrid pulse coding strategy, and a convolution pulse neural network model is constructed; after pre-training and training with labeled data, a signal to be diagnosed is input, and the output result is statistically analyzed. The wind turbine variable pitch system early mechanical loosening diagnosis method based on SCNN solves the problems that the traditional wind turbine variable pitch system mechanical loosening diagnosis method is not sensitive to early weak loosening signals, depends on artificial threshold, and has poor model generalization ability.

Description

Technical Field

[0001] This invention relates to the field of mechanical loosening diagnosis technology, and in particular to an early mechanical loosening diagnosis method for wind turbine pitch systems based on SCNN. Background Technology

[0002] As a core component for adjusting blade angle and optimizing wind energy capture, the stability of the pitch system in a wind turbine directly affects power generation efficiency and equipment lifespan. However, under complex operating conditions (such as vibration, temperature differences, and corrosion), the bearings, gearboxes, or connecting parts of the pitch system are prone to mechanical loosening due to fatigue, wear, or assembly errors. Such faults initially manifest as subtle vibration anomalies; if not detected promptly, they can lead to serious accidents such as blade runaway and gear breakage, resulting in millions of dollars in economic losses. Traditional diagnostic methods rely on manual inspections or threshold-based vibration analysis, which struggle to capture early signs of loosening, leading to delayed fault detection and failing to meet the "preventive maintenance" requirements of wind farms.

[0003] Currently, the diagnosis of mechanical loosening in wind turbines mainly employs two types of technologies: spectral analysis-based methods, which extract frequency domain features (such as peak values ​​and harmonics) from vibration signals to determine faults, but require preset fixed thresholds, are insensitive to early nonlinear loosening, and are easily affected by noise interference; and machine learning classification methods, which can automatically extract features, but rely on a large amount of labeled data, and the model's generalization ability is greatly affected by changes in operating conditions. Furthermore, traditional neural networks do not fully consider the temporal characteristics of vibration signals, resulting in insufficient early characterization capabilities for loosening faults. While spiking neural networks (SNNs) can simulate the temporal pulse mechanism of biological neurons, they lack effective feature encoding and training strategies, making them difficult to directly apply to complex industrial scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide an early mechanical loosening diagnosis method for wind turbine pitch systems based on SCNN, which solves the problems of traditional wind turbine pitch system mechanical loosening diagnosis methods being insensitive to early weak loosening signals, relying on manual thresholds, and having poor model generalization ability, and achieves high-precision, adaptive early fault detection.

[0005] To achieve the above objectives, this invention provides an early mechanical loosening diagnosis method for wind turbine pitch systems based on SCNN, comprising:

[0006] High-precision vibration sensors are installed on the pitch system of wind turbines to collect vibration signals. High-frequency noise and low-frequency interference in the original vibration signals are removed, and the denoised vibration signals are segmented.

[0007] Variational mode decomposition is used to decompose each vibration signal segment to obtain the intrinsic mode function, and the multi-scale sample entropy of each intrinsic mode function component is calculated as the feature vector.

[0008] A hybrid pulse coding strategy is used to encode the extracted feature vectors into pulse sequences. The pulse triggering threshold is set according to the magnitude of the feature values ​​of the feature vectors to construct a convolutional spiking neural network model.

[0009] The convolutional spiking neural network model was pre-trained using a wind turbine vibration dataset, and then trained using labeled training data to adjust the synaptic weights.

[0010] The vibration signal of the wind turbine pitch system to be diagnosed is processed and input into the convolutional pulse neural network model. The output diagnostic results are judged and statistically analyzed.

[0011] Preferably, the objective function expression for variational mode decomposition is:

[0012] ;

[0013] in, For the first One intrinsic mode function component For the first The center frequency of an intrinsic mode For balance function, It is a unit impulse function. For convolution operations, For the time partial derivative, The imaginary unit, It is a norm.

[0014] The preferred expression for calculating multi-scale sample entropy is:

[0015] ;

[0016] in, It is a time series. For embedded dimensions, For similarity tolerance, As a scale factor, For time series in Similarity counts.

[0017] Preferably, the hybrid pulse coding strategy specifically involves wavelet packet decomposition of the vibration signal, first pulse time coding for the high-frequency subband, and pulse frequency coding for the low-frequency subband. An adaptive threshold adjustment algorithm is introduced during the coding process to dynamically optimize the pulse trigger threshold based on historical feature vectors.

[0018] Preferably, the convolutional spiking neural network is constructed by including spiking convolutional layers, spiking pooling layers, and fully connected layers. The spiking convolutional layers use variable pulse timing kernels, and the synaptic weights are adjusted by the pulse time-dependent plasticity STDP rule.

[0019] The expression for adjusting synaptic weights according to the STDP rule is:

[0020] ;

[0021] in, The time difference between the pulses of the preceding and following neurons. A positive learning rate A negative learning rate is the time constant.

[0022] Preferably, the high-frequency sub-band uses first pulse time coding to record the first pulse exceeding the threshold. Pulse moment Generate pulse sequence The low-frequency subband uses pulse frequency coding to count the number of pulses per unit time. Generate pulse density sequence ;

[0023] The adaptive threshold adjustment algorithm is expressed as follows:

[0024] ;

[0025] in, As the initial threshold, To adjust the coefficient, This represents the variance of the most recent 100 eigenvectors.

[0026] Preferably, the output diagnostic results include and The expressions for the accuracy and false alarm rate of the diagnostic results are as follows:

[0027] ;

[0028] ;

[0029] in, It is a true positive. True negative, It was a false positive. It is a false negative.

[0030] Therefore, the present invention employs the aforementioned SCNN-based early mechanical loosening diagnosis method for wind turbine pitch systems, and the technical effects are as follows:

[0031] 1. Multi-scale feature fusion: Combining variational mode decomposition (VMD) with multi-scale sample entropy, the time-frequency-nonlinear composite features of vibration signals are extracted, breaking through the linear limitations of traditional spectrum analysis.

[0032] 2. Hybrid pulse coding strategy: The high-frequency subband uses first pulse time coding and the low-frequency subband uses pulse frequency coding, and an adaptive threshold adjustment is introduced to enhance the sensitivity to weak loose signals.

[0033] 3. STDP-driven SCNN model: Unsupervised pre-training is achieved through the impulse time-dependent plasticity (STDP) rule, and supervised learning is combined to fine-tune synaptic weights, solving the model convergence problem in small sample scenarios. Attached Figure Description

[0034] Figure 1 This is a flowchart of the early mechanical loosening diagnosis method for wind turbine pitch system based on SCNN of the present invention. Detailed Implementation

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0036] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0037] Example 1

[0038] like Figure 1 As shown, this invention provides an early mechanical loosening diagnosis method for wind turbine pitch systems based on SCNN, including:

[0039] High-precision vibration sensors are installed on the pitch system of wind turbines to collect vibration signals. High-frequency noise and low-frequency interference in the original vibration signals are removed, and the denoised vibration signals are segmented.

[0040] High-precision, high-sensitivity vibration sensors are selected to ensure the capture of even weak vibration signals. The sensors are installed securely in critical areas of the wind turbine's pitch system, such as bearing housings and gearbox housings, minimizing external interference. A data acquisition card or dedicated data acquisition system is used, with appropriate sampling rates and times set, to continuously or periodically collect vibration signals. The raw vibration signals are stored in a computer or embedded system for subsequent processing and analysis.

[0041] Digital filtering techniques, such as Butterworth or Chebyshev filters, are applied to remove high-frequency noise and low-frequency interference from the original vibration signal. Appropriate filter parameters and cutoff frequencies are selected based on signal characteristics to ensure effective noise removal while preserving the useful signal. Depending on the signal length and diagnostic requirements, the denoised vibration signal is divided into several segments, each of moderate length to maintain signal continuity. Fixed-length segmentation or segmentation based on signal characteristics can be used to ensure that each signal segment contains a complete vibration cycle.

[0042] Variational mode decomposition is used to decompose each vibration signal segment to obtain the intrinsic mode function, and the multi-scale sample entropy of each intrinsic mode function component is calculated as the feature vector.

[0043] For each vibration signal segment, VMD decomposition is performed to obtain a series of intrinsic mode functions (IMFs). Appropriate decomposition levels and penalty factors are set to ensure that the obtained IMF components effectively reflect the time-frequency characteristics of the signal.

[0044] The objective function expression for variational mode decomposition is:

[0045] ;

[0046] in, For the first One intrinsic mode function component For the first The center frequency of an intrinsic mode For balance function, It is a unit impulse function. For convolution operations, For the time partial derivative, The imaginary unit, It is a norm.

[0047] For each IMF component, calculate the multi-scale sample entropy as a feature vector. Set appropriate embedding dimensions, similarity tolerances, and scale factors to ensure that the calculated multi-scale sample entropy can effectively reflect the complexity and nonlinear characteristics of the signal.

[0048] The expression for calculating multi-scale sample entropy is:

[0049] ;

[0050] in, It is a time series. For embedded dimensions, For similarity tolerance, As a scale factor, For time series in Similarity counts.

[0051] A hybrid pulse coding strategy is used to encode the extracted feature vectors into pulse sequences. The pulse trigger threshold is set according to the magnitude of the feature values ​​of the feature vectors to construct a convolutional spiking neural network model.

[0052] Wavelet packet decomposition is performed on the vibration signal to obtain high-frequency and low-frequency sub-bands. Appropriate wavelet basis functions and decomposition levels are selected to ensure that the obtained high-frequency and low-frequency sub-bands effectively reflect the time-frequency characteristics of the signal. A first-pulse time coding method is used to record the moment of the first pulse exceeding a set threshold, generating a pulse sequence. Appropriate thresholds and sampling window sizes are set to ensure accurate capture of pulse events in the high-frequency sub-band. A pulse frequency coding method is used to count the number of pulses per unit time, generating a pulse density sequence. Appropriate time window sizes and counting methods are set to ensure accurate reflection of the pulse frequency characteristics in the low-frequency sub-band.

[0053] The high-frequency subband uses first pulse time coding to record the first pulse exceeding the threshold. Pulse moment Generate pulse sequence The low-frequency subband uses pulse frequency coding to count the number of pulses per unit time. Generate pulse density sequence ;

[0054] The adaptive threshold adjustment algorithm is expressed as follows:

[0055] ;

[0056] in, As the initial threshold, To adjust the coefficient, This represents the variance of the most recent 100 eigenvectors.

[0057] An SCNN model consisting of pulsed convolutional layers, pulsed pooling layers, and fully connected layers is constructed. The pulsed convolutional layers employ variable pulse temporal kernels, and the synaptic weights are adjusted using the pulse temporal dependent plasticity (STDP) rule.

[0058] The expression for adjusting synaptic weights according to the STDP rule is:

[0059] ;

[0060] in, The time difference between the pulses of the preceding and following neurons. A positive learning rate A negative learning rate is the time constant.

[0061] The convolutional spiking neural network model was pre-trained using a wind turbine vibration dataset, and then trained using labeled training data to adjust the synaptic weights.

[0062] The SCNN model was pre-trained using a wind turbine vibration dataset to initialize network weights. Supervised learning training was then performed on labeled training data, adjusting synaptic weights to minimize classification error. Cross-validation and grid search were used to optimize the network structure and hyperparameters, improving the model's generalization ability and classification accuracy.

[0063] The vibration signal of the wind turbine pitch system to be diagnosed is processed and input into the convolutional pulse neural network model. The output diagnostic results are judged and statistically analyzed.

[0064] The vibration signal of the wind turbine pitch system to be diagnosed is processed through the same procedure (denoising, segmentation, feature extraction, and hybrid pulse coding) and then input into the trained SCNN model. The SCNN model outputs diagnostic results, including categories such as normal and loose. Based on the diagnostic results, corresponding maintenance measures are taken, such as immediate shutdown for inspection and planned maintenance.

[0065] Statistical analysis is performed on the accuracy and false alarm rate of the diagnostic results to evaluate model performance. The output diagnostic results include... and The expressions for the accuracy and false alarm rate of the diagnostic results are as follows:

[0066] ;

[0067] ;

[0068] in, It is a true positive. True negative, It was a false positive. It is a false negative.

[0069] Ten 2MW doubly-fed induction generator (DFIG) wind turbines from a wind farm were selected as the experimental subjects. High-precision piezoelectric vibration sensors were installed at two key locations on the pitch system of each turbine: the bearing housing and the gearbox housing. The sensor sampling rate was set to 5120Hz, and vibration signals were continuously collected for three months. Simultaneously, three groups of early mechanical loosening faults of different degrees were artificially set: slight loosening, moderate loosening, and severe loosening, corresponding to a 15% reduction, a 30% reduction, and a 50% reduction in bolt preload, respectively. Fault labels were assigned to the corresponding signals. Finally, a dataset containing the normal state and the three loosening states was constructed, totaling 1200 samples (300 samples for each state, with each sample containing 10240 sampling points).

[0070] The original vibration signal was collected and a fourth-order Butterworth bandpass filter was used to remove high-frequency noise and low-frequency interference. The denoised signal was then segmented into segments of 1024 sampling points each, with each segment corresponding to one analysis sample.

[0071] Variational mode decomposition (VMD) is performed on each signal segment with a decomposition layer of 6 and a penalty factor of 2000, resulting in 6 intrinsic mode function (IMF) components. Multiscale sample entropy is calculated for each IMF component, and finally a feature vector with a dimension of 30 is generated.

[0072] Wavelet packet decomposition (using db4 wavelet basis, decomposition level 3) is performed on the vibration signal corresponding to the feature vector to obtain high-frequency subband and low-frequency subband. The high-frequency subband uses first pulse time coding and the low-frequency subband uses pulse frequency coding. The threshold is dynamically optimized by an adaptive threshold adjustment algorithm to generate pulse sequence.

[0073] An SCNN model (containing 2 pulsed convolutional layers, 1 pulsed pooling layer, and 2 fully connected layers) is constructed. First, it is pre-trained using the STDP rule with an unlabeled wind turbine vibration dataset, and then fine-tuned with labeled samples under supervision. The signal to be diagnosed is processed according to the above process and then input into the model to output the diagnostic results.

[0074] To verify the superiority of the method of this invention, two traditional diagnostic methods were selected as control groups: Control group 1: threshold diagnosis method based on spectrum analysis; Control group 2: diagnosis method based on BP neural network. The diagnostic accuracy and false alarm rate of the three methods under different degrees of loosening were statistically analyzed, and the results are shown in Table 1 below.

[0075]

[0076] In scenarios involving slight loosening, the diagnostic accuracy of the SCNN method in this invention reaches 98.5%, far exceeding the 62.3% of the spectral thresholding method and the 81.7% of the BP neural network method. This is because the traditional spectral thresholding method relies on a fixed frequency domain threshold, making it difficult to identify the nonlinear vibration characteristics of weak loosening; the time domain features of the BP neural network are insufficient for representing early faults. In contrast, this invention combines the composite features of VMD multi-scale sample entropy with hybrid pulse coding, which can accurately capture the temporal differences of weak signals.

[0077] The method of this invention has an average accuracy of 99.05% and an average false alarm rate of only 0.85%, which is significantly better than the control group. The spectrum thresholding method is greatly affected by noise and operating condition fluctuations, and has the highest false alarm rate (average 9.62%). Although the BP neural network is better than the spectrum method, its generalization ability is insufficient and it is prone to feature failure in cross-operating condition samples. SCNN, through unsupervised pre-training by STDP, improves the model's adaptability to complex operating conditions.

[0078] Therefore, this invention adopts the aforementioned SCNN-based early mechanical loosening diagnosis method for wind turbine pitch systems. By installing high-precision vibration sensors at key parts of the wind turbine pitch system to collect signals, and after denoising and segmented preprocessing, the inherent mode functions are obtained through variational mode decomposition, and time-frequency-nonlinear composite features are extracted by calculating multi-scale sample entropy. Then, a hybrid pulse coding strategy using high-frequency subband first pulse time coding, low-frequency subband pulse frequency coding, and adaptive threshold adjustment is employed to transform the features into pulse sequences. An SCNN model containing pulse convolutional layers, pooling layers, and fully connected layers is constructed, and synaptic weights are adjusted based on STDP rules. The model is first pre-trained using a wind turbine vibration dataset, then fine-tuned using labeled data, and finally, the input signal to be diagnosed is used to complete the diagnosis and result statistics. This achieves accurate capture of early mechanical loosening signals in wind turbine pitch systems, solving the problems of traditional methods being insensitive to weak early loosening signals, relying on manual thresholds, and having poor model generalization ability. It achieves high-precision, adaptive early fault detection, providing reliable technical support for preventative maintenance of wind farms.

[0079] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for early mechanical loosening diagnosis of wind turbine pitch system based on SCNN, characterized in that, include: High-precision vibration sensors are installed on the pitch system of wind turbines to collect vibration signals. High-frequency noise and low-frequency interference in the original vibration signals are removed, and the denoised vibration signals are segmented. Variational mode decomposition is used to decompose each vibration signal segment to obtain the intrinsic mode function, and the multi-scale sample entropy of each intrinsic mode function component is calculated as the feature vector. A hybrid pulse coding strategy is used to encode the vibration signal corresponding to the extracted feature vector into a pulse sequence. The pulse trigger threshold is set according to the magnitude of the feature value of the feature vector, and a convolutional pulse neural network model is constructed. The convolutional spiking neural network model was pre-trained using a wind turbine vibration dataset, and then trained using labeled training data to adjust the synaptic weights. The vibration signal of the wind turbine pitch system to be diagnosed is processed and input into the convolutional pulse neural network model. The output diagnostic results are judged and statistically analyzed. The hybrid pulse coding strategy specifically involves wavelet packet decomposition of the vibration signal corresponding to the feature vector. The high-frequency sub-band uses first pulse time coding, while the low-frequency sub-band uses pulse frequency coding. An adaptive threshold adjustment algorithm is introduced during the coding process to dynamically optimize the pulse trigger threshold based on historical feature vectors.

2. The method for early mechanical loosening diagnosis of wind turbine pitch system based on SCNN according to claim 1, characterized in that, The objective function expression for variational mode decomposition is: ； in, For the first One intrinsic mode function component For the first The center frequency of an intrinsic mode For balance function, It is a unit impulse function. For convolution operations, For the time partial derivative, The imaginary unit, It is a norm.

3. The method for early mechanical loosening diagnosis of wind turbine pitch system based on SCNN according to claim 1, characterized in that, The expression for calculating multi-scale sample entropy is: ； in, It is a time series. For embedded dimensions, For similarity tolerance, As a scale factor, For time series in Similarity counts.

4. The method for early mechanical loosening diagnosis of wind turbine pitch system based on SCNN according to claim 1, characterized in that, The construction of the convolutional spiking neural network includes spiking convolutional layers, spiking pooling layers, and fully connected layers. The spiking convolutional layers use variable pulse temporal kernels and adjust synaptic weights through the pulse time-dependent plasticity STDP rule. The expression for adjusting synaptic weights according to the STDP rule is: ； in, The time difference between the pulses of the preceding and following neurons. A positive learning rate A negative learning rate is the time constant.

5. The method for early mechanical loosening diagnosis of wind turbine pitch system based on SCNN according to claim 1, characterized in that, The high-frequency subband uses first pulse time coding to record the first pulse exceeding the threshold. Pulse moment Generate pulse sequence ; The low-frequency subband uses pulse frequency coding to count the number of pulses per unit time. Generate pulse density sequence ; The adaptive threshold adjustment algorithm is expressed as follows: ； in, As the initial threshold, To adjust the coefficient, This represents the variance of the most recent 100 eigenvectors.

6. The method for early mechanical loosening diagnosis of wind turbine pitch system based on SCNN according to claim 1, characterized in that, The output diagnostic results include and The expressions for the accuracy and false alarm rate of the diagnostic results are as follows: ； ； in, It is a true positive. True negative, It was a false positive. It is a false negative.

Images