A wind turbine generator system failure early warning method and system

Abstract

The application discloses a kind of wind generating set fault early warning method and system, comprising: S1: collecting vibration waveform original data and temperature original data, and respectively pre-processing;S2: design isomorphic double-flow network respectively extracts high-frequency impact feature and low-frequency modulation feature;S3: according to high-frequency impact feature and low-frequency modulation feature, calculate vibration feature;S4: extract temperature rate of change sequence and temperature time series feature;S5: generate vibration on temperature attention weight matrix, temperature on vibration attention weight matrix, calculate fault sensitive mask, and generate fault sensitive feature;S6: to fault sensitive feature is compressed, again calculate three-stage early warning probability vector, unit fault early warning class, and issue alarm.The application can solve the problem of early warning lag caused by the dependence of traditional wind generating set fault early warning method on fixed threshold alarm.

Description

Technical Field

[0001] This invention relates to the field of wind power generation technology, and in particular to a method and system for early warning of wind turbine generator set faults. Background Technology

[0002] As a core component of the transmission system, the health of the wind turbine gearbox directly affects the unit's operating efficiency and lifespan. Currently, the industry commonly uses fixed threshold alarm mechanisms, such as temperature over-limit triggering warnings, as the primary monitoring method. While simple to implement, this method has significant limitations: Firstly, when the temperature parameter reaches the preset threshold, the gearbox often suffers irreversible mechanical damage such as severe pitting on the gear teeth or plastic deformation of the bearings, leading to a surge in maintenance costs. Secondly, traditional methods only monitor the effective value or peak value of vibration signals, failing to capture early, subtle fault characteristics or effectively explore the dynamic correlation between vibration and temperature parameters, resulting in a high false alarm rate and significant warning lag. Although recent research has attempted to use deep learning methods to analyze vibration spectra, these methods largely rely on cloud computing, involve large numbers of model parameters, and struggle to achieve joint analysis of multi-modal characteristics of temperature and vibration, making it difficult for existing technologies to provide early warnings in the early stages of faults. Summary of the Invention

[0003] In view of this, the present invention aims to provide a method and system for early warning of wind turbine generator faults, so as to solve the problem of early warning lag caused by the reliance on fixed threshold alarms in traditional methods.

[0004] A method for early warning of faults in wind turbine generator sets includes:

[0005] S1: Collect raw vibration waveform data and raw temperature data, and preprocess them respectively to obtain preprocessed vibration time series data and preprocessed temperature time series data;

[0006] S2: Based on the preprocessed vibration time-series data, short-time Fourier transform and continuous wavelet transform are combined to generate a time-spectrum map and a wavelet time-frequency map. Then, a heterogeneous dual-stream network is designed to extract high-frequency impact features and low-frequency modulation features respectively. The heterogeneous dual-stream network includes: sequentially performing 3×3 depthwise separable convolution, GELU activation, and a second 3×3 depthwise separable convolution on the time-spectrum map, and superimposing the output results of 1×1 depthwise separable convolution and 2×2 max pooling to obtain high-frequency impact features; sequentially performing 2×2 average pooling, ReLU activation, and 5×5 depthwise separable convolution on the wavelet time-frequency map, and performing a Hadamard product operation with the output of the 3×3 depthwise separable convolution activated by Sigmoid to obtain low-frequency modulation features.

[0007] S3: Based on the high-frequency impact characteristics and low-frequency modulation characteristics, calculate the vibration characteristics, including: performing attention calculation on the high-frequency impact characteristics, and then combining residual connections and layer normalization to generate high-frequency enhanced features; performing time and frequency dimension pooling on the low-frequency modulation characteristics respectively, and after fusion, generating a gated weight matrix through convolution and activation functions, and then combining dimension expansion to generate low-frequency enhanced features; finally, generating vibration characteristics through a dynamic weight gating mechanism.

[0008] S4: Extract the temperature change rate sequence and temperature time series features based on the preprocessed temperature time series data;

[0009] S5: Based on the temperature time series characteristics and vibration characteristics, generate the vibration attention weight matrix on temperature and the temperature attention weight matrix on vibration, calculate the fault-sensitive mask, and generate fault-sensitive features.

[0010] S6: Compress the fault-sensitive features to obtain the fault compressed features, then calculate the three-level early warning probability vector and the unit fault early warning category, and issue an alarm.

[0011] Furthermore, step S1 also includes:

[0012] S11: Based on the piezoelectric accelerometer installed at the gearbox measuring point, collect the raw vibration waveform data to obtain the vibration timing signal; the sampling rate of the raw vibration waveform data is not less than 20,000 times per second.

[0013] S12: Based on the temperature sensors installed on the gearbox bearing housing and the gear meshing surface, raw temperature data is collected to obtain a temperature timing signal; the sampling rate of the raw temperature data is once per second.

[0014] S13: For the vibration time series signal, the wavelet threshold denoising method is used to remove high-frequency noise to obtain the preprocessed vibration time series data;

[0015] S14: For the temperature time series signal, outliers are removed by the interquartile range method, and the dimensions are unified by the minimum-maximum normalization method to obtain the preprocessed temperature time series data.

[0016] Furthermore, step S2 also includes:

[0017] S21: Based on the preprocessed vibration time series data, generate the time spectrum X using short-time Fourier transform. stft And the wavelet time-frequency graph X is generated through continuous wavelet transform. cwt ;

[0018] S22: Based on the time-frequency spectrum and wavelet time-frequency diagram, a heterogeneous dual-stream network is designed to extract high-frequency impulse features and low-frequency modulation features, respectively. The calculation method is as follows:

[0019] Fhigh =DSConv 3×3 (GELU(DSConv 3×3 (X stft )))+MaxPool 2×2 (DSConv 1×1 (X stft ));

[0020] F low =DSConv 5×5 (ReLU(AvgPool 2×2 (X cwt )))⊙Sigmoid(DSConv 3×3 (X cwt ));

[0021] Among them, F high DSConv exhibits high-frequency impact characteristics. 3×3 () represents a depthwise separable convolution with a kernel size of 3×3, GELU() is the GELU activation function, and MaxPool 2×2 () represents max pooling with a pooling window size of 2×2, DSConv 1×1 () represents a depthwise separable convolution with a kernel size of 1×1, F low As a low-frequency modulation feature, DSConv 5×5 () represents a depthwise separable convolution with a kernel size of 5×5, ReLU() is the ReLU activation function, and AvgPool 2×2 () represents average pooling with a pooling window size of 2×2, ⊙ represents the Hadamard product, and Sigmoid() is the Sigmoid activation function.

[0022] Furthermore, in step S3, the dynamic weight gating mechanism specifically includes: performing global average pooling on the high-frequency enhancement features and the low-frequency enhancement features respectively, concatenating the pooling results, and then calculating the dynamic fusion weights through the Softmax function of the fully connected layer; and generating vibration features by weighted summation of the high-frequency enhancement features and the low-frequency enhancement features according to the dynamic fusion weights.

[0023] Furthermore, step S3 also includes:

[0024] S31: Based on the high-frequency impulse characteristics and low-frequency modulation characteristics, high-frequency enhancement characteristics and low-frequency enhancement characteristics are generated. The calculation method is as follows:

[0025]

[0026] F′ high =LayerNorm(F high +M high·Conv(F high ));

[0027] G=Sigmoid(Conv(TimeAvgPool(F low )+FreqMaxPool(F low )));

[0028] F′ low =F low ⊙Expand(G);

[0029] Among them, M high Here, is the high-frequency attention matrix, Softmax() is the softmax function, Conv() is the convolutional layer, and DSConv() is the depthwise separable convolution. F′ is the matrix transpose, d is the dimension of the high-frequency impact feature, and F′ is the matrix transpose. high For high-frequency enhancement features, LayerNorm() is used for layer normalization, G is the low-frequency gate weight matrix, TimeAvgPool() is the time-dimensional average pooling, FreqMaxPool() is the frequency-dimensional max pooling, and F′ is the frequency-dimensional max pooling. low For low-frequency enhancement features, Expand() is a vector dimension expansion operation;

[0030] S32: Based on the high-frequency enhancement features and low-frequency enhancement features, vibration features are generated through a dynamic weighted gating mechanism. The calculation method is as follows:

[0031] α=Softmax(FC(Concat(AvgPool(F′ high ),AvgPool(F′ low ))));

[0032] F fusion =α·F′ high +(1-α)·F′ low ;

[0033] Where α is the dynamic fusion weight, FC() is the fully connected layer, Concat() is the concatenation operation, AvgPool() is the average pooling, and F fusion This is a vibration characteristic.

[0034] Furthermore, step S4 also includes:

[0035] S41: Based on the preprocessed temperature time series data, the temperature change rate sequence is calculated using the sliding window difference method. Specifically, the preprocessed temperature time series data is truncated with a fixed window length. The difference between the temperature value at the end of the window and the temperature value at the beginning of the window is divided by the window length and then by the sampling interval time to obtain the temperature change rate sequence.

[0036] S42: Based on the preprocessed temperature time-series data and the temperature change rate sequence, temperature time-series features are generated using a dynamic convolution kernel. The calculation method is as follows:

[0037] K dynamic =Softplus(Conv1D(R′));

[0038] F temp =DepthwiseConv(R,K dynamic );

[0039] Among them, K dynamic Here are the parameters for the temperature-dynamic convolution kernel, Softplus() is the Softplus activation function, Conv1D() is a one-dimensional convolutional layer, R′ is the temperature change rate sequence, and F... temp R represents the temperature time series characteristics, and R represents the preprocessed temperature time series data.

[0040] Furthermore, the generation of fault-sensitive features in step S5 specifically includes: performing linear interpolation on the temperature time-series features to obtain aligned temperature features; concatenating the aligned temperature features with the vibration features and inputting them into a one-dimensional convolutional layer to generate attention weight matrices for vibration on temperature and temperature on vibration; multiplying the two attention weight matrices element-wise and generating a fault-sensitive mask through nonlinear activation; concatenating the vibration features with the aligned temperature features, performing layer normalization, and multiplying them element-wise with the fault-sensitive mask, and generating fault-sensitive features through a temporal convolutional network and random dropout operation.

[0041] Furthermore, step S5 also includes:

[0042] S51: Based on the temperature time series characteristics, aligned temperature characteristics are generated through linear interpolation. The calculation method is as follows:

[0043]

[0044] in, For the aligned temperature features, Linearlnterp() performs linear interpolation;

[0045] S52: Based on the aligned temperature and vibration characteristics, a bidirectional asymmetric attention mechanism is designed to generate the vibration-temperature attention weight matrix and the temperature-vibration attention weight matrix. The calculation method is as follows:

[0046]

[0047] Among them, A v2t Let A be the attention weighting matrix for vibration with respect to temperature. t2v This is the attention weighting matrix for the effect of temperature on vibration;

[0048] S53: Calculate the fault-sensitive mask based on the vibration-temperature attention weight matrix and the temperature-vibration attention weight matrix. The calculation method is as follows:

[0049] Mask risk =ReLU(A v2t ⊙A t2v );

[0050] Among them, Mask risk For fault-sensitive masking;

[0051] S54: Based on the fault-sensitive mask, vibration characteristics, and aligned temperature characteristics, fault-sensitive features are generated through a gated residual network. The calculation method is as follows:

[0052]

[0053] Where P is the cross-modal joint feature matrix, Q risk For fault-sensitive features, TCN() is a temporal convolutional network. For element-wise multiplication, Dropout() performs a random discard operation.

[0054] Furthermore, step S6 also includes:

[0055] S61: Based on the fault sensitivity characteristics, fault compression features are generated through time-averaged pooling. The calculation method is as follows:

[0056] Z = AvgPool(Conv1D(Q) risk ));

[0057] Where Z represents the fault compression feature;

[0058] S62: Based on the fault compression characteristics, a three-level early warning probability vector is generated through a fully connected network. The calculation method is as follows:

[0059] S warn =Softmax(FC(Z));

[0060] S warn =[s norm ,s early ,s emerg ];

[0061] Among them, S warn Let s be the probability vector for the third-level early warning system. norm ,s early ,s emerg These represent the probability values ​​for normal operation, early warning, and emergency shutdown, respectively.

[0062] S63: Select the category with the highest probability value from the three-level early warning probability vector as the unit fault early warning category and issue an alarm.

[0063] This invention also discloses a wind turbine generator fault early warning system, characterized in that it includes:

[0064] Data acquisition and preprocessing module: Acquires raw vibration waveform data and raw temperature data, and performs preprocessing on them respectively to obtain preprocessed vibration time series data and preprocessed temperature time series data;

[0065] High and low frequency feature extraction module: Based on the preprocessed vibration time series data, combined with short-time Fourier transform and continuous wavelet transform, time spectrum diagram and wavelet time frequency diagram are generated, and then a heterogeneous dual-stream network is designed to extract high-frequency impact features and low-frequency modulation features respectively.

[0066] Vibration feature extraction module: Calculates vibration features based on high-frequency impact features and low-frequency modulation features;

[0067] Temperature time series feature extraction module: Extracts the temperature change rate sequence and temperature time series features based on the preprocessed temperature time series data;

[0068] Fault-sensitive feature extraction module: Based on temperature time-series features and vibration features, it generates the vibration attention weight matrix on temperature and the temperature attention weight matrix on vibration, calculates the fault-sensitive mask, and generates fault-sensitive features.

[0069] The Level 3 Early Warning Module compresses fault-sensitive features to obtain fault compressed features, then calculates the Level 3 Early Warning Probability Vector and the Unit Fault Early Warning Category, and issues an alarm.

[0070] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0071] (1) This invention uses short-time Fourier transform to capture high-frequency impact features and combines wavelet transform to extract low-frequency modulation signals to achieve wideband signal characterization; it processes two types of features through a heterogeneous dual-stream network; finally, it designs a bidirectional asymmetric attention mechanism for the cross-modal correlation characteristics of temperature and vibration signals to achieve early warning of wind turbine generator failures.

[0072] (2) This invention innovatively proposes a heterogeneous dual-stream network to process vibration signals. The high-frequency path uses 3×3 depth-separable convolution with GELU activation and utilizes autocorrelation matrix to quantify the impact correlation intensity, effectively extracting transient impacts such as bearing pitting. The low-frequency path uses 5×5 convolution and Hadamard product operations, combined with spatiotemporal pooling to generate gating weights, accurately capturing slowly changing faults such as gear wear. At the same time, the dynamic weight gating mechanism automatically adjusts the fusion weights according to the feature energy distribution, improving the sensitivity of early fault detection and controlling the number of model parameters to meet the real-time requirements of edge computing.

[0073] (3) The bidirectional asymmetric attention mechanism designed in this invention effectively correlates vibration and temperature signals. The vibration-to-temperature path captures the impact hysteresis effect, and the temperature-to-vibration path models the vibration harmonic component offset law. By screening key features through fault-sensitive masks and combining them with gated residual networks, accurate detection is achieved, which improves the detection rate of faults with weak correlation between temperature rise and vibration impact. Attached Figure Description

[0074] Figure 1 A flowchart illustrating a wind turbine generator fault early warning method provided by the present invention;

[0075] Figure 2 A schematic diagram of the algorithm flow for extracting high-frequency impact features and low-frequency modulation features provided by the present invention;

[0076] Figure 3 A schematic diagram of the vibration feature extraction algorithm provided by the present invention;

[0077] Figure 4 A visualization example of the fault-sensitive features provided by this invention. Detailed Implementation

[0078] The present invention will be further described below with reference to the accompanying drawings, but this is not intended to limit the present invention in any way. Any modifications or substitutions made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0079] Example 1: A method for early warning of faults in wind turbine generator sets, such as... Figure 1 As shown, it includes the following steps:

[0080] S1: Collect raw vibration waveform data and raw temperature data, and preprocess them respectively to obtain preprocessed vibration time series data and preprocessed temperature time series data;

[0081] S11: Based on the piezoelectric accelerometer installed at the gearbox measuring point, collect the raw vibration waveform data to obtain the vibration timing signal; the sampling rate of the raw vibration waveform data is not less than 20,000 times per second.

[0082] S12: Based on the temperature sensors installed on the gearbox bearing housing and the gear meshing surface, raw temperature data is collected to obtain a temperature timing signal; the sampling rate of the raw temperature data is once per second.

[0083] S13: For the vibration time series signal, the wavelet threshold denoising method is used to remove high-frequency noise to obtain the preprocessed vibration time series data;

[0084] S14: For the temperature time series signal, outliers are removed by the interquartile range method, and the dimensions are unified by the minimum-maximum normalization method to obtain the preprocessed temperature time series data.

[0085] S2: Based on the preprocessed vibration time-series data, short-time Fourier transform and continuous wavelet transform are combined to generate a time-spectrum map and a wavelet time-frequency map. Then, a heterogeneous dual-stream network is designed to extract high-frequency impact features and low-frequency modulation features respectively. The heterogeneous dual-stream network includes: sequentially performing 3×3 depthwise separable convolution, GELU activation, and a second 3×3 depthwise separable convolution on the time-spectrum map, and superimposing the output results of 1×1 depthwise separable convolution and 2×2 max pooling to obtain high-frequency impact features; sequentially performing 2×2 average pooling, ReLU activation, and 5×5 depthwise separable convolution on the wavelet time-frequency map, and performing a Hadamard product operation with the output of the 3×3 depthwise separable convolution activated by Sigmoid to obtain low-frequency modulation features.

[0086] S21: Based on the preprocessed vibration time series data, generate the time spectrum X using short-time Fourier transform. stft And the wavelet time-frequency graph X is generated through continuous wavelet transform. cwt ;

[0087] S22: Based on the time-frequency spectrum and wavelet time-frequency diagram, a heterogeneous dual-stream network is designed to extract high-frequency impulse features and low-frequency modulation features, respectively. Figure 2 As shown, the calculation method is as follows:

[0088] F high =DSConv 3×3 (GELU(DSConv 3×3 (X stft )))+MaxPool 2×2 (DSConv 1×1 (X stft ));

[0089] F low =DSConv 5×5 (ReLU(AvgPool 2×2 (X cwt )))⊙Sigmoid(DSConv 3×3 (X cwt ));

[0090] Among them, F high DSConv exhibits high-frequency impact characteristics. 3×3 () represents a depthwise separable convolution with a kernel size of 3×3, GELU() is the GELU activation function, and MaxPool 2×2 () represents max pooling with a pooling window size of 2×2, DSConv 1×1 () represents a depthwise separable convolution with a kernel size of 1×1, F low As a low-frequency modulation feature, DSConv 5×5 () represents a depthwise separable convolution with a kernel size of 5×5, ReLU() is the ReLU activation function, and AvgPool 2×2 () represents average pooling with a pooling window size of 2×2, ⊙ represents the Hadamard product, and Sigmoid() is the Sigmoid activation function.

[0091] It should be further explained that, at the frequency domain analysis level, this invention constrains the time-frequency resolution of high-frequency impact components through a constant window function in the short-time Fourier transform, and captures the time-varying energy distribution of low-frequency modulated signals by combining the multi-scale characteristics of wavelet transform. This mathematically resolves the inherent contradictions of traditional single time-frequency analysis methods in broadband signal representation. Addressing the coexistence of high-frequency transient impacts and low-frequency slowly varying modulations in wind turbine vibration signals, the heterogeneous dual-stream network achieves feature decoupling through a combination of differentiated convolutional kernels and nonlinear activation functions: in the high-frequency path, small-scale depthwise separable convolution combined with max pooling is used, leveraging the smooth gradient characteristics of the GELU function to suppress high-frequency noise interference; in the low-frequency path, large-scale convolutional kernels are used for sensing... For long-term trends, the design utilizes Hadamard product dynamic weighting to highlight anomalous modulation frequency bands. Compared to traditional single-stream convolutional networks, this design separates the spectral sparsity of high-frequency components from the energy continuity of low-frequency components, avoiding modal confusion caused by mixed feature extraction. In practical applications, early-stage wind turbine bearing failures manifest as weak impact pulses (high-amplitude narrow-band peaks in the frequency domain), while gear wear exhibits low-frequency sideband modulation (wideband energy diffusion in the frequency domain). This invention uses dual-stream heterogeneous processing to ensure that the feature response domains of the two types of failures do not interfere with each other, overcoming the problem of increased misjudgment rate caused by frequency band overlap in traditional algorithms. At the same time, the number of parameters in depthwise separable convolution is only 1 / 8 of that in standard convolution, meeting the performance requirements for lightweight models.

[0092] S3: Based on the high-frequency impact features and low-frequency modulation features, vibration features are calculated, including: attention calculation on the high-frequency impact features, followed by residual connections and layer normalization to generate high-frequency enhancement features; time and frequency dimension pooling is performed on the low-frequency modulation features respectively, and after fusion, a gated weight matrix is ​​generated by convolution and activation functions, followed by dimension expansion to generate low-frequency enhancement features; finally, vibration features are generated through a dynamic weight gating mechanism; the dynamic weight gating mechanism specifically includes: global average pooling is performed on the high-frequency enhancement features and low-frequency enhancement features respectively, and the pooling results are concatenated, followed by calculation of dynamic fusion weights through the Softmax function of a fully connected layer; based on the dynamic fusion weights, the high-frequency enhancement features and low-frequency enhancement features are weighted and summed to generate vibration features;

[0093] S31: Based on the high-frequency impulse characteristics and low-frequency modulation characteristics, high-frequency enhancement characteristics and low-frequency enhancement characteristics are generated. The calculation method is as follows:

[0094]

[0095] F′ high =LayerNorm(F high +M high ·Conv(F high ));

[0096] G=Sigmoid(Conv(TimeAvgPool(F low )+FreqMaxPool(F low )));

[0097] F′ low =F low ⊙Expand(G);

[0098] Among them, M high Here, is the high-frequency attention matrix, Softmax() is the softmax function, Conv() is the convolutional layer, and DSConv() is the depthwise separable convolution. F′ is the matrix transpose, d is the dimension of the high-frequency impact feature, and F′ is the matrix transpose. high For high-frequency enhancement features, LayerNorm() is used for layer normalization, G is the low-frequency gate weight matrix, TimeAvgPool() is the time-dimensional average pooling, FreqMaxPool() is the frequency-dimensional max pooling, and F′ is the frequency-dimensional max pooling. low For low-frequency enhancement features, Expand() is a vector dimension expansion operation;

[0099] S32: Based on high-frequency and low-frequency enhancement features, vibration features are generated through a dynamic weighted gating mechanism, such as... Figure 3 As shown, the calculation method is as follows:

[0100] α=Softmax(FC(Concat(AvgPool(F′ high ),AvgPool(F′ low ))));

[0101] F fusion =α·F′ high +(1-α)·F′ low ;

[0102] Where α is the dynamic fusion weight, FC() is the fully connected layer, Concat() is the concatenation operation, AvgPool() is the average pooling, and F fusion This is a vibration characteristic.

[0103] It should be further explained that, at the mathematical modeling level, this invention employs an attention mechanism to enhance features by addressing the difference between the transient nature of high-frequency impact components and the continuity of low-frequency modulation components: the high-frequency path constructs attention weights through the autocorrelation matrix of convolutional features, quantifies the correlation strength of impact components at different time points using matrix dot product operations, and suppresses overfitting by combining normalization processing; the low-frequency path extracts long-term trends through time-dimensional average pooling, superimposes frequency-dimensional max pooling to capture abnormal frequency band energy peaks, and generates a low-frequency gating weight matrix through nonlinear mapping, thereby achieving targeted amplification of slowly changing fault features.

[0104] Furthermore, compared to traditional fixed-ratio fusion methods, the dynamic weight gating mechanism compresses feature dimensions through global average pooling and utilizes fully connected layers to learn the contribution ratio of high-frequency and low-frequency components, mathematically ensuring that the weight coefficients adaptively adjust with the energy distribution of the input features. This invention automatically allocates fusion weights in the time-frequency domain through a dynamic weight mechanism. When the energy of high-frequency impacts suddenly increases, its weight is increased to quickly respond to sudden faults, and when the low-frequency modulation is continuously abnormal, its contribution is enhanced to capture progressive damage, solving the problem of insufficient sensitivity of traditional algorithms to complex faults. At the same time, layer normalization and dimension expansion operations ensure the fusionability of features with different dimensions.

[0105] S4: Extract the temperature change rate sequence and temperature time series features based on the preprocessed temperature time series data;

[0106] S41: Based on the preprocessed temperature time series data, the temperature change rate sequence is calculated using the sliding window difference method. Specifically, the preprocessed temperature time series data is truncated with a fixed window length. The difference between the temperature value at the end of the window and the temperature value at the beginning of the window is divided by the window length and then by the sampling interval time to obtain the temperature change rate sequence.

[0107] S42: Based on the preprocessed temperature time-series data and the temperature change rate sequence, temperature time-series features are generated using a dynamic convolution kernel. The calculation method is as follows:

[0108] K dynamic =Softplus(Conv1D(R′));

[0109] F temp =DepthwiseConv(R,K dynamic );

[0110] Among them, K dynamic Here are the parameters for the temperature-dynamic convolution kernel, Softplus() is the Softplus activation function, Conv1D() is a one-dimensional convolutional layer, R′ is the temperature change rate sequence, and F... temp R represents the temperature time series characteristics, and R represents the preprocessed temperature time series data.

[0111] Furthermore, the calculation process for generating fault-sensitive features in step S5 includes: linearly interpolating the temperature time-series features to obtain aligned temperature features; concatenating the aligned temperature features with the vibration features and inputting them into a one-dimensional convolutional layer to generate attention weight matrices for vibration on temperature and temperature on vibration; multiplying the two attention weight matrices element-wise and generating a fault-sensitive mask through nonlinear activation; concatenating the vibration features with the aligned temperature features and performing layer normalization, and multiplying them element-wise with the fault-sensitive mask, generating fault-sensitive features through a temporal convolutional network and random dropout operation.

[0112] S5: Based on the temperature time series characteristics and vibration characteristics, generate the vibration attention weight matrix on temperature and the temperature attention weight matrix on vibration, calculate the fault-sensitive mask, and generate fault-sensitive features.

[0113] S51: Based on the temperature time series characteristics, aligned temperature characteristics are generated through linear interpolation. The calculation method is as follows:

[0114]

[0115] in, For the aligned temperature features, Linearlnterp() performs linear interpolation;

[0116] S52: Based on the aligned temperature and vibration characteristics, a bidirectional asymmetric attention mechanism is designed to generate the vibration-temperature attention weight matrix and the temperature-vibration attention weight matrix. The calculation method is as follows:

[0117]

[0118] Among them, A v2tLet A be the attention weighting matrix for vibration with respect to temperature. t2v This is the attention weighting matrix for the effect of temperature on vibration;

[0119] S53: Calculate the fault-sensitive mask based on the vibration-temperature attention weight matrix and the temperature-vibration attention weight matrix. The calculation method is as follows:

[0120] Mask risk =ReLU(A v2t ⊙A t2v );

[0121] Among them, Mask risk For fault-sensitive masking;

[0122] S54: Based on the fault-sensitive mask, vibration characteristics, and aligned temperature characteristics, fault-sensitive features are generated through a gated residual network. The calculation method is as follows:

[0123]

[0124] Where P is the cross-modal joint feature matrix, Q risk For fault-sensitive features, TCN() is a temporal convolutional network. For element-wise multiplication, Dropout() performs a random discard operation.

[0125] It should be further explained that this invention designs a bidirectional asymmetric attention mechanism for the cross-modal correlation characteristics of temperature and vibration signals: the vibration-to-temperature path captures the hysteresis effect of high-frequency mechanical shock on local temperature rise through independent convolutional layers, while the temperature-to-vibration path models the vibration harmonic component shift law caused by the decrease in material stiffness due to temperature rise. The Softmax function is used to probabilistically express the bidirectional correlation strength. Compared with traditional single-modal analysis or simple feature stitching methods, the fault-sensitive mask filters the overlapping region of bidirectional attention through Hadamard product, which is mathematically equivalent to constructing a convex hull of fault-related features in the joint time-frequency space. The ReLU function is used to remove negative interference components, forming a directional enhancement of composite fault features.

[0126] In practical applications, wind turbine gearbox failures often exhibit a weak correlation between gradual temperature rise and vibration shock. For example, in the early stages of lubricant failure, the temperature rise lags behind the vibration anomaly. Traditional methods often lead to misjudgments due to neglecting cross-modal temporal phase differences. This invention uses a gated residual network to perform residual learning on the masked weighted features and cross-modal joint features, utilizing the hole structure of the temporal convolutional network to cover the fault evolution mode. Compared to fixed-ratio multimodal fusion algorithms, this design mathematically achieves a dynamic balance between the temporal sensitivity of vibration signals and the cumulative effect of temperature signals, solving the problem of delayed response to early complex faults in traditional methods.

[0127] For example, the present invention in Figure 4 The fault-sensitive features were visualized, with a dimension of 24 hours × 64 feature channels, including two types of typical fault features:

[0128] (1) Channel 16 shows high activation (red area) with an amplitude of more than 2.5 during the period of 8-16 hours, which corresponds to the high frequency impact signal caused by bearing pitting;

[0129] (2) Channel 43 shows a sustained activation of about 1.8 (orange area) during the 12-20 hour period, reflecting temperature-related modulation caused by gear wear; the negative value area (blue) accounts for about 35% of the matrix, which is due to the random suppression of non-critical features by the Dropout layer.

[0130] S6: Compress the fault-sensitive features to obtain the fault compressed features, then calculate the three-level early warning probability vector and the unit fault early warning category, and issue an alarm.

[0131] S61: Based on the fault sensitivity characteristics, fault compression features are generated through time-averaged pooling. The calculation method is as follows:

[0132] Z = AvgPool(Conv1D(Q) risk ));

[0133] Where Z represents the fault compression feature;

[0134] S62: Based on the fault compression characteristics, a three-level early warning probability vector is generated through a fully connected network. The calculation method is as follows:

[0135] S warn =Softmax(FC(Z));

[0136] S warn =[s norm ,s early ,s emerg ];

[0137] Among them, S warn Let s be the probability vector for the third-level early warning system. norm ,s early ,s emerg These represent the probability values ​​for normal operation, early warning, and emergency shutdown, respectively.

[0138] S63: Select the category with the highest probability value from the three-level early warning probability vector as the unit fault early warning category and issue an alarm.

[0139] Specifically, for monitoring scenarios of wind turbine generators where intermittent impact faults are frequent, such as instantaneous bearing spalling or the initial stage of gear tooth breakage, characterized by alternating sudden high-amplitude pulses and normal low-energy signals, this invention also provides a gated weighted pooling calculation method to replace step S61. The calculation method is as follows:

[0140] Z = γ·MaxPool(Conv1D(Q) risk ))+(1-γ)·AvgPool(Conv1D(Q risk ));

[0141] γ=Sigmoid(Conv1D(ReLU(Conv1D(Q risk ))));

[0142] Where Z represents the fault compression feature, γ represents the intermittent impact gating weight, and MaxPool() represents the maximum pooling.

[0143] Example 2: The present invention also discloses a wind turbine generator fault early warning system, comprising:

[0144] Data acquisition and preprocessing module: Acquires raw vibration waveform data and raw temperature data, and performs preprocessing on them respectively to obtain preprocessed vibration time series data and preprocessed temperature time series data;

[0145] High and low frequency feature extraction module: Based on the preprocessed vibration time series data, combined with short-time Fourier transform and continuous wavelet transform, time spectrum diagram and wavelet time frequency diagram are generated, and then a heterogeneous dual-stream network is designed to extract high-frequency impact features and low-frequency modulation features respectively.

[0146] Vibration feature extraction module: Calculates vibration features based on high-frequency impact features and low-frequency modulation features;

[0147] Temperature time series feature extraction module: Extracts the temperature change rate sequence and temperature time series features based on the preprocessed temperature time series data;

[0148] Fault-sensitive feature extraction module: Based on temperature time-series features and vibration features, it generates the vibration attention weight matrix on temperature and the temperature attention weight matrix on vibration, calculates the fault-sensitive mask, and generates fault-sensitive features.

[0149] The Level 3 Early Warning Module compresses fault-sensitive features to obtain fault compressed features, then calculates the Level 3 Early Warning Probability Vector and the Unit Fault Early Warning Category, and issues an alarm.

[0150] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0151] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0152] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for early warning of faults in wind turbine generator sets, characterized in that, Includes the following steps: S1: Collect raw vibration waveform data and raw temperature data, and preprocess them respectively to obtain preprocessed vibration time series data and preprocessed temperature time series data; S2: Based on the preprocessed vibration time series data, combined with short-time Fourier transform and continuous wavelet transform, a time spectrum diagram and a wavelet time-frequency diagram are generated. Then, a heterogeneous dual-stream network is designed to extract high-frequency impact features and low-frequency modulation features respectively. The heterogeneous dual-stream network includes: sequentially performing 3×3 depthwise separable convolution, GELU activation, and a second 3×3 depthwise separable convolution on the time-frequency spectrum, and superimposing the output of 1×1 depthwise separable convolution and 2×2 max pooling to obtain high-frequency impulse features; sequentially performing 2×2 average pooling, ReLU activation, and 5×5 depthwise separable convolution on the wavelet time-frequency spectrum, and performing a Hadamard product operation with the output of 3×3 depthwise separable convolution activated by Sigmoid to obtain low-frequency modulation features; S3: Based on the high-frequency impact characteristics and low-frequency modulation characteristics, calculate the vibration characteristics, including: performing attention calculation on the high-frequency impact characteristics, and then combining residual connections and layer normalization to generate high-frequency enhanced features; performing time and frequency dimension pooling on the low-frequency modulation characteristics respectively, and after fusion, generating a gated weight matrix through convolution and activation functions, and then combining dimension expansion to generate low-frequency enhanced features; finally, generating vibration characteristics through a dynamic weight gating mechanism. S4: Extract the temperature change rate sequence and temperature time series features based on the preprocessed temperature time series data; S5: Based on the temporal and vibration characteristics, generate the attention weight matrix of vibration on temperature and the attention weight matrix of temperature on vibration, calculate the fault-sensitive mask, and generate fault-sensitive features. Specifically, this includes: performing linear interpolation on the temporal temperature characteristics to obtain aligned temperature characteristics; concatenating the aligned temperature characteristics with the vibration characteristics and inputting them into a one-dimensional convolutional layer to generate the attention weight matrices of vibration on temperature and temperature on vibration; performing element-wise multiplication of the two attention weight matrices and generating the fault-sensitive mask through nonlinear activation; concatenating the vibration characteristics with the aligned temperature characteristics, performing layer normalization, and multiplying them element-wise with the fault-sensitive mask, then generating the fault-sensitive features through a temporal convolutional network and random dropout operation. S6: Compress the fault-sensitive features to obtain the fault compressed features, then calculate the three-level early warning probability vector and the unit fault early warning category, and issue an alarm.

2. The wind turbine generator set fault early warning method according to claim 1, characterized in that, Step S1 includes: S11: Based on the piezoelectric accelerometer installed at the gearbox measuring point, collect the raw vibration waveform data to obtain the vibration timing signal; the sampling rate of the raw vibration waveform data is not less than 20,000 times per second. S12: Based on the temperature sensors installed on the gearbox bearing housing and the gear meshing surface, raw temperature data is collected to obtain a temperature timing signal; the sampling rate of the raw temperature data is once per second. S13: For the vibration time series signal, the wavelet threshold denoising method is used to remove high-frequency noise to obtain the preprocessed vibration time series data; S14: For the temperature time series signal, outliers are removed by the interquartile range method, and the dimensions are unified by the minimum-maximum normalization method to obtain the preprocessed temperature time series data.

3. The wind turbine generator set fault early warning method according to claim 1, characterized in that, Step S2 includes: S21: Generate a time spectrum diagram based on the preprocessed vibration time series data using short-time Fourier transform. Wavelet time-frequency diagrams are generated through continuous wavelet transform. ; S22: Based on the time-frequency spectrum and wavelet time-frequency diagram, a heterogeneous dual-stream network is designed to extract high-frequency impulse features and low-frequency modulation features, respectively. The calculation method is as follows: in, It is characterized by high-frequency impact. This is a depthwise separable convolution with a kernel size of 3×3. For GELU activation function, For max pooling with a pooling window size of 2×2, For depthwise separable convolutions with a kernel size of 1×1, It is a low-frequency modulation feature. For depthwise separable convolutions with a kernel size of 5×5, It is the ReLU activation function. For average pooling with a pooling window size of 2×2, For Hadama accumulation, This is the Sigmoid activation function.

4. The wind turbine generator set fault early warning method according to claim 3, characterized in that, In step S3, the dynamic weight gating mechanism specifically includes: performing global average pooling on the high-frequency enhancement features and the low-frequency enhancement features respectively, concatenating the pooling results, and then calculating the dynamic fusion weights through the Softmax function of the fully connected layer; and generating vibration features by weighted summation of the high-frequency enhancement features and the low-frequency enhancement features according to the dynamic fusion weights.

5. The wind turbine generator set fault early warning method according to claim 4, characterized in that, Step S3 includes: S31: Based on the high-frequency impulse characteristics and low-frequency modulation characteristics, high-frequency enhancement characteristics and low-frequency enhancement characteristics are generated. The calculation method is as follows: in, For high-frequency attention matrices, For the Softmax function, It is a convolutional layer. For depthwise separable convolution, For matrix transpose, For high-frequency impact characteristic dimensions, It is a high-frequency enhancement feature. For layer normalization, This is a low-frequency gated weight matrix. For time-dimensional average pooling, Max pooling is performed for the frequency dimension. It is a low-frequency enhancement feature. For vector dimension expansion operations; S32: Based on the high-frequency enhancement features and low-frequency enhancement features, vibration features are generated through a dynamic weighted gating mechanism. The calculation method is as follows: in, For dynamic weight fusion, It is a fully connected layer. For splicing operations, For average pooling, This is a vibration characteristic.

6. The wind turbine generator set fault early warning method according to claim 5, characterized in that, Step S4 includes: S41: Based on the preprocessed temperature time series data, the temperature change rate sequence is calculated using the sliding window difference method. Specifically, the preprocessed temperature time series data is truncated with a fixed window length. The difference between the temperature value at the end of the window and the temperature value at the beginning of the window is divided by the window length and then by the sampling interval time to obtain the temperature change rate sequence. S42: Based on the preprocessed temperature time-series data and the temperature change rate sequence, temperature time-series features are generated using a dynamic convolution kernel. The calculation method is as follows: in, For temperature-dynamic convolution kernel parameters, This is the Softplus activation function. It is a one-dimensional convolutional layer. It is a temperature change rate sequence. Temperature time series characteristics, For depthwise separable convolution, This is the preprocessed temperature time series data.

7. The wind turbine generator set fault early warning method according to claim 6, characterized in that, Step S5 includes: S51: Based on the temperature time series characteristics, aligned temperature characteristics are generated through linear interpolation. The calculation method is as follows: in, For the aligned temperature characteristics, Linear interpolation; S52: Based on the aligned temperature and vibration characteristics, a bidirectional asymmetric attention mechanism is designed to generate the vibration-temperature attention weight matrix and the temperature-vibration attention weight matrix. The calculation method is as follows: in, Let be the attention weight matrix for vibration with respect to temperature. This is the attention weighting matrix for the effect of temperature on vibration; S53: Calculate the fault-sensitive mask based on the vibration-temperature attention weight matrix and the temperature-vibration attention weight matrix. The calculation method is as follows: in, For fault-sensitive masking; S54: Based on the fault-sensitive mask, vibration characteristics, and aligned temperature characteristics, fault-sensitive features are generated through a gated residual network. The calculation method is as follows: in, For cross-modal joint feature matrix, It is a fault-sensitive feature. For temporal convolutional networks, For element-wise multiplication, This is a random discard operation.

8. The wind turbine generator set fault early warning method according to claim 7, characterized in that, Step S6 includes: S61: Based on the fault sensitivity characteristics, fault compression features are generated through time-averaged pooling. The calculation method is as follows: in, This is a fault compression feature; S62: Based on the fault compression characteristics, a three-level early warning probability vector is generated through a fully connected network. The calculation method is as follows: in, This is the probability vector for the third-level early warning system. These represent the probability values ​​for normal operation, early warning, and emergency shutdown, respectively. S63: Select the category with the highest probability value from the three-level early warning probability vector as the unit fault early warning category and issue an alarm.

9. A fault early warning system for wind turbine generator sets, characterized in that, include: Data acquisition and preprocessing module: Acquires raw vibration waveform data and raw temperature data, and performs preprocessing on them respectively to obtain preprocessed vibration time series data and preprocessed temperature time series data; High and low frequency feature extraction module: Based on the preprocessed vibration time series data, combined with short-time Fourier transform and continuous wavelet transform, time spectrum diagram and wavelet time frequency diagram are generated, and then a heterogeneous dual-stream network is designed to extract high-frequency impact features and low-frequency modulation features respectively. Vibration feature extraction module: Calculates vibration features based on high-frequency impact features and low-frequency modulation features; Temperature time series feature extraction module: Extracts the temperature change rate sequence and temperature time series features based on the preprocessed temperature time series data; Fault-sensitive feature extraction module: Based on temperature time-series features and vibration features, it generates the vibration attention weight matrix on temperature and the temperature attention weight matrix on vibration, calculates the fault-sensitive mask, and generates fault-sensitive features. The Level 3 Early Warning Module: Compresses fault-sensitive features to obtain fault compressed features, then calculates the Level 3 Early Warning Probability Vector and the Unit Fault Early Warning Category, and issues an alarm. To achieve the wind turbine generator set fault early warning method as described in any one of claims 1-8.

Images