Fault early warning method and system for wind generating set

Abstract

The invention discloses a wind generating set fault early warning method and system, and the method comprises the steps: S1, collecting vibration waveform original data and temperature original data, and carrying out the preprocessing of the data; s2, designing a heterogeneous double-current network to extract high-frequency impact features and low-frequency modulation features respectively; s3, calculating vibration characteristics according to the high-frequency impact characteristics and the low-frequency modulation characteristics; s4, extracting a temperature change rate sequence and temperature time sequence characteristics; s5, generating an attention weight matrix of vibration to temperature and an attention weight matrix of temperature to vibration, calculating a fault sensitive mask, and generating fault sensitive features; and S6, compressing the fault sensitive features, calculating a three-level early warning probability vector and a unit fault early warning category, and giving an alarm. According to the method, the problem of early warning lag caused by dependence on fixed threshold value alarm of a traditional wind generating set fault early warning method can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation, and in particular to a wind generator set fault early warning method and system. Background Art

[0002] As a core component of the transmission system, the health of a wind turbine gearbox directly impacts the unit's operating efficiency and lifespan. Currently, the industry generally uses fixed-threshold alarm mechanisms, such as temperature overshoot triggering early warnings, as the primary monitoring method. While this method is simple to implement, it has significant limitations. On the one hand, by the time the temperature parameter reaches the preset threshold, the gearbox has often already suffered irreversible mechanical damage such as severe pitting on the tooth surface or plastic deformation of the bearing, leading to a surge in repair costs. On the other hand, traditional methods only perform simple effective value or peak value monitoring of vibration signals, failing to capture early, subtle fault characteristics or effectively explore the dynamic correlation between vibration and temperature parameters. This can easily lead to a high system false alarm rate and severe early warning lags. Although recent studies have attempted to use deep learning methods to analyze vibration spectra, these methods often rely on cloud computing, have large model parameters, and struggle to achieve joint analysis of temperature and vibration multimodal characteristics, making it difficult for existing technologies to provide early warnings. Summary of the Invention

[0003] In view of this, the present invention aims to provide a wind turbine generator set fault early warning method and system to solve the problem of early warning lag caused by traditional methods relying on fixed threshold alarms.

[0004] A wind turbine generator set fault early warning method, comprising:

[0005] S1: collecting vibration waveform raw data and temperature raw data, and preprocessing 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, a time-frequency spectrum and a wavelet time-frequency spectrum are generated by combining short-time Fourier transform and continuous wavelet transform, and 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: performing 3×3 depth-wise separable convolution, GELU activation, and secondary 3×3 depth-wise separable convolution on the time-frequency spectrum, and superimposing the output results of 1×1 depth-wise separable convolution and 2×2 maximum pooling to obtain high-frequency impact features; performing 2×2 average pooling, ReLU activation, and 5×5 depth-wise separable convolution on the wavelet time-frequency spectrum, and performing a Hadamard product operation on the result with the output of the 3×3 depth-wise separable convolution after Sigmoid activation to obtain low-frequency modulation features;

[0007] S3: Calculate vibration features based on high-frequency impact features and low-frequency modulation features, including: performing attention calculation on high-frequency impact features, and then combining residual connections and layer normalization to generate high-frequency enhancement features; pooling low-frequency modulation features in time and frequency dimensions, fusing them through convolution and activation functions to generate a gated weight matrix, and then combining dimensional expansion to generate low-frequency enhancement features; finally, generating vibration features through a dynamic weight gating mechanism;

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

[0009] S5: Generate the attention weight matrix of vibration to temperature and the attention weight matrix of temperature to vibration based on the temperature time series characteristics and vibration characteristics, calculate the fault sensitive mask, and generate the fault sensitive features;

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

[0011] Furthermore, the step S1 further includes:

[0012] S11: Collecting raw vibration waveform data from a piezoelectric acceleration sensor installed at a measuring point on the gearbox to obtain a vibration time series signal; the sampling rate of the raw vibration waveform data is not less than 20,000 times per second;

[0013] S12: collecting raw temperature data from temperature sensors provided on the gearbox bearing seat and the gear meshing surface to obtain a temperature time series signal; the sampling rate of the raw temperature data is once per second;

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

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

[0016] Furthermore, the step S2 further includes:

[0017] S21: Generate a time-frequency spectrum X through short-time Fourier transform based on the pre-processed vibration time series data stft , and generate the wavelet time-frequency graph X by continuous wavelet transform cwt ;

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

[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 is the high-frequency impact feature, DSConv 3×3 () is a depth-separable convolution with a kernel size of 3×3, GELU() is a GELU activation function, and MaxPool 2×2 () is the maximum pooling with a pooling window size of 2×2, DSConv 1×1 () is a depth-wise separable convolution with a kernel size of 1×1, F low is the low-frequency modulation feature, DSConv 5×5 () is a depth-separable convolution with a kernel size of 5×5, ReLU() is the ReLU activation function, and AvgPool 2×2 () is the average pooling with a pooling window size of 2×2, ⊙ is the Hadamard product, and Sigmoid() is the Sigmoid activation function.

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

[0023] Furthermore, the step S3 further includes:

[0024] S31: Generate high-frequency enhancement features and low-frequency enhancement features based on high-frequency impact features and low-frequency modulation features. The calculation method is:

[0025]

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

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

[0028] F′ low =F low ⊙Expand(G);

[0029] Among them, M high is the high-frequency attention matrix, Softmax() is the Softmax function, Conv() is the convolution layer, DSConv() is the depth-wise separable convolution, is the matrix transpose, d is the high-frequency impact feature dimension, F′ high is the high-frequency enhancement feature, LayerNorm() is the layer normalization, G is the low-frequency gating weight matrix, TimeAvgPool() is the time dimension average pooling, FreqMaxPool() is the frequency dimension maximum pooling, F′ low is the low-frequency enhancement feature, and Expand() is the vector dimension expansion operation;

[0030] S32: Generate vibration features based on the high-frequency enhancement features and the low-frequency enhancement features through a dynamic weight gating mechanism. The calculation method is:

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

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

[0033] Among them, α is the dynamic fusion weight, FC() is the fully connected layer, Concat() is the splicing operation, AvgPool() is the average pooling, F fusion Vibration characteristics.

[0034] Furthermore, the step S4 further includes:

[0035] S41: Calculate the temperature change rate sequence based on the preprocessed temperature time series data using a sliding window difference method. Specifically, the preprocessed temperature time series data is intercepted with a fixed window length, and 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 divided by the sampling interval to obtain the temperature change rate sequence.

[0036] S42: Generate temperature time series features using a dynamic convolution kernel based on the preprocessed temperature time series data and the temperature change rate sequence. The calculation method is:

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

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

[0039] Among them, K dynamic is the temperature dynamic convolution kernel parameter, Softplus() is the Softplus activation function, Conv1D() is the one-dimensional convolution layer, R′ is the temperature change rate sequence, F temp is the temperature time series feature, and R is the preprocessed temperature time series data.

[0040] Furthermore, the fault-sensitive features are generated in step S5, specifically including: linear interpolation of the temperature time series features to obtain aligned temperature features; concatenating the aligned temperature features with the vibration features and inputting them into the one-dimensional convolution layer respectively to generate attention weight matrices of vibration to temperature and temperature to vibration; element-wise multiplication of the two attention weight matrices and generating a fault-sensitive mask through nonlinear activation; concatenating the vibration features with the aligned temperature features, performing layer normalization, and performing element-wise multiplication with the fault-sensitive mask, and generating the fault-sensitive features through a temporal convolution network and random drop operation.

[0041] Furthermore, the step S5 further includes:

[0042] S51: Generate aligned temperature features through linear interpolation based on the temperature time series features. The calculation method is:

[0043]

[0044] in, is the aligned temperature feature, Linearlnterp() is the linear interpolation;

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

[0046]

[0047] Among them, A v2t is the attention weight matrix of vibration to temperature, A t2v is the attention weight matrix of temperature to vibration;

[0048] S53: Calculate the fault sensitive mask according to the attention weight matrix of vibration to temperature and the attention weight matrix of temperature to vibration. The calculation method is:

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

[0050] Among them, Mask risk is a fault sensitive mask;

[0051] S54: Generate fault-sensitive features through a gated residual network based on the fault-sensitive mask, vibration features, and aligned temperature features. The calculation method is:

[0052]

[0053] Among them, P is the cross-modal joint feature matrix, Q risk is the fault-sensitive feature, TCN() is the temporal convolutional network, is element-by-element multiplication, and Dropout() is a random drop operation.

[0054] Furthermore, the step S6 further includes:

[0055] S61: Generate fault compression features through time average pooling based on fault-sensitive features. The calculation method is:

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

[0057] Where Z is the fault compression feature;

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

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

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

[0061] Among them, S warn is the three-level warning probability vector, s norm ,s early ,s emerg are the probability values ​​of normal, early warning, and emergency shutdown respectively;

[0062] S63: The category with the largest probability value in the three-level warning probability vector is selected as the unit fault warning category, and an alarm is issued.

[0063] The present invention also discloses a wind turbine generator set fault warning system, which is characterized by comprising:

[0064] Data acquisition and preprocessing module: collects the original vibration waveform data and the original temperature data, and preprocesses them respectively to obtain the 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, short-time Fourier transform and continuous wavelet transform are combined to generate time-frequency spectrum and wavelet time-frequency spectrum. 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: Generates the attention weight matrix of vibration to temperature and the attention weight matrix of temperature to vibration based on the temperature time series features and vibration features, calculates the fault-sensitive mask, and generates fault-sensitive features;

[0069] Three-level warning module: compresses fault-sensitive features to obtain fault compression features, then calculates the three-level warning probability vector and unit fault warning category, and issues an alarm.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] (1) The present invention uses short-time Fourier transform to capture high-frequency impact features and combines it with wavelet transform to extract low-frequency modulation signals to achieve wide-band signal representation; it processes the two types of features through a heterogeneous dual-stream network. Finally, based on the cross-modal correlation characteristics of temperature and vibration signals, a bidirectional asymmetric attention mechanism is designed to achieve early warning of wind turbine faults.

[0072] (2) The present 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 uses the autocorrelation matrix to quantify the impact correlation strength, 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 slow-changing faults such as gear wear. At the same time, the dynamic weight gating mechanism automatically adjusts the fusion weight according to the characteristic energy distribution, thereby 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 the present invention effectively associates vibration and temperature signals. The vibration-to-temperature path captures the shock hysteresis effect, and the temperature-to-vibration path models the offset law of the vibration harmonic components. By screening key features through fault-sensitive masks and combining with gated residual networks to achieve accurate detection, the detection rate of faults with weak correlation between slow temperature rise and vibration shock is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 A schematic flow chart of a wind turbine generator set fault 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 algorithm flow for vibration feature extraction provided by the present invention;

[0077] Figure 4 This is a visualization example diagram of the fault-sensitive characteristics provided by the present invention. DETAILED DESCRIPTION

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

[0079] Example 1: A wind turbine generator set fault warning method, such as Figure 1 As shown, the following steps are included:

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

[0081] S11: Collecting raw vibration waveform data from a piezoelectric acceleration sensor installed at a measuring point on the gearbox to obtain a vibration time series signal; the sampling rate of the raw vibration waveform data is not less than 20,000 times per second;

[0082] S12: collecting raw temperature data from temperature sensors provided on the gearbox bearing seat and the gear meshing surface to obtain a temperature time series signal; the sampling rate of the raw temperature data is once per second;

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

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

[0085] S2: Based on the preprocessed vibration time series data, a time-frequency spectrum and a wavelet time-frequency spectrum are generated by combining short-time Fourier transform and continuous wavelet transform, and 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: performing 3×3 depth-wise separable convolution, GELU activation, and secondary 3×3 depth-wise separable convolution on the time-frequency spectrum, and superimposing the output results of 1×1 depth-wise separable convolution and 2×2 maximum pooling to obtain high-frequency impact features; performing 2×2 average pooling, ReLU activation, and 5×5 depth-wise separable convolution on the wavelet time-frequency spectrum, and performing a Hadamard product operation on the result with the output of the 3×3 depth-wise separable convolution after Sigmoid activation to obtain low-frequency modulation features;

[0086] S21: Generate a time-frequency spectrum X through short-time Fourier transform based on the pre-processed vibration time series data stft , and generate the wavelet time-frequency graph X by continuous wavelet transform cwt ;

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

[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 is the high-frequency impact feature, DSConv 3×3 () is a depth-separable convolution with a kernel size of 3×3, GELU() is a GELU activation function, and MaxPool 2×2 () is the maximum pooling with a pooling window size of 2×2, DSConv 1×1 () is a depth-wise separable convolution with a kernel size of 1×1, F low is the low-frequency modulation feature, DSConv 5×5 () is a depth-separable convolution with a kernel size of 5×5, ReLU() is the ReLU activation function, and AvgPool 2×2 () is the average pooling with a pooling window size of 2×2, ⊙ is the Hadamard product, and Sigmoid() is the Sigmoid activation function.

[0091] It should be further explained that, at the frequency domain analysis level, the present invention constrains the time-frequency resolution of high-frequency impact components through the constant window function in the short-time Fourier transform, and captures the time-varying energy distribution of low-frequency modulated signals in combination with the multi-scale characteristics of the wavelet transform, thereby solving the inherent contradiction of the traditional single time-frequency analysis method in the characterization of wide-band signals from a mathematical principle; in view of the coexistence of high-frequency transient impacts and low-frequency slowly varying modulations in the vibration signals of wind turbines, the heterogeneous dual-stream network realizes feature decoupling through the combination of differentiated convolution kernels and nonlinear activation functions: in the high-frequency path, small-scale depth-separable convolution is used in combination with maximum pooling, and the smooth gradient characteristics of the GELU function are used to suppress high-frequency noise interference; in the low-frequency path, large-scale convolution kernel perception is used Long-term trend, using Hadamard product dynamic weighting to highlight abnormal modulation frequency bands; compared with the traditional single-stream convolutional network, this design separates and models the spectral sparsity of high-frequency components and the energy continuity of low-frequency components, avoiding the modal confusion problem caused by mixed feature extraction; in actual application scenarios, the early manifestation of wind turbine bearing failure is a weak impact pulse (high-amplitude narrowband peak in the frequency domain), while gear wear presents low-frequency sideband modulation (wideband energy diffusion in the frequency domain). The present invention uses dual-stream heterogeneous processing to make the characteristic response domains of the two types of faults non-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 of the deep separable convolution is only 1 / 8 of that of the standard convolution, meeting the performance requirements of lightweight models.

[0092] S3: Calculate vibration features based on high-frequency impact features and low-frequency modulation features, including: performing attention calculation on high-frequency impact features, and then combining residual connections and layer normalization to generate high-frequency enhancement features; pooling low-frequency modulation features in time and frequency dimensions respectively, generating a gating weight matrix after fusion through convolution and activation functions, and then combining dimensional expansion to generate low-frequency enhancement features; finally, generating vibration features through a dynamic weight gating mechanism; the dynamic weight gating mechanism specifically includes: performing global average pooling on high-frequency enhancement features and low-frequency enhancement features respectively, and splicing the pooling results, and then calculating dynamic fusion weights through the fully connected layer Softmax function; based on the dynamic fusion weights, performing weighted summation on high-frequency enhancement features and low-frequency enhancement features to generate vibration features;

[0093] S31: Generate high-frequency enhancement features and low-frequency enhancement features based on high-frequency impact features and low-frequency modulation features. The calculation method is:

[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 is the high-frequency attention matrix, Softmax() is the Softmax function, Conv() is the convolution layer, DSConv() is the depth-wise separable convolution, is the matrix transpose, d is the high-frequency impact feature dimension, F′ high is the high-frequency enhancement feature, LayerNorm() is the layer normalization, G is the low-frequency gating weight matrix, TimeAvgPool() is the time dimension average pooling, FreqMaxPool() is the frequency dimension maximum pooling, F′ low is the low-frequency enhancement feature, and Expand() is the vector dimension expansion operation;

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

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

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

[0102] Among them, α is the dynamic fusion weight, FC() is the fully connected layer, Concat() is the splicing operation, AvgPool() is the average pooling, F fusion Vibration characteristics.

[0103] It should be further explained that, at the mathematical modeling level, the present invention adopts an attention mechanism to achieve feature enhancement based on the difference in 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 the convolution features, uses matrix dot product operations to quantify the correlation strength of impact components at different time points, and combines normalization processing to suppress overfitting; the low-frequency path extracts long-term trends through average pooling in the time dimension, superimposes maximum pooling in the frequency dimension to capture the energy peak of the abnormal frequency band, and generates a low-frequency gating weight matrix through nonlinear mapping, thereby realizing directional amplification of slowly varying fault features.

[0104] In addition, compared with the traditional fixed-ratio fusion method, the dynamic weight gating mechanism compresses the feature dimension through global average pooling, uses the fully connected layer to learn the contribution ratio of high-frequency and low-frequency components, and mathematically ensures that the weight coefficient is adaptively adjusted with the input feature energy distribution; the present invention automatically allocates fusion weights in the time-frequency domain space through a dynamic weight mechanism. When the high-frequency impact energy suddenly increases, its weight is increased to quickly respond to sudden faults. When the low-frequency modulation continues to be abnormal, its contribution is enhanced to capture progressive damage, which solves the problem of insufficient sensitivity of traditional algorithms to complex faults; at the same time, layer normalization and dimensionality expansion operations ensure the fusion of features of different dimensions.

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

[0106] S41: Calculate the temperature change rate sequence based on the preprocessed temperature time series data using a sliding window difference method. Specifically, the preprocessed temperature time series data is intercepted with a fixed window length, and 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 divided by the sampling interval to obtain the temperature change rate sequence.

[0107] S42: Generate temperature time series features using a dynamic convolution kernel based on the preprocessed temperature time series data and the temperature change rate sequence. The calculation method is:

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

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

[0110] Among them, K dynamic is the temperature dynamic convolution kernel parameter, Softplus() is the Softplus activation function, Conv1D() is the one-dimensional convolution layer, R′ is the temperature change rate sequence, F temp is the temperature time series feature, and R is the preprocessed temperature time series data.

[0111] Furthermore, the calculation process of generating fault-sensitive features in step S5 includes: performing linear interpolation on the temperature time series features to obtain aligned temperature features; splicing the aligned temperature features with the vibration features and inputting them into the one-dimensional convolution layer respectively to generate attention weight matrices of vibration to temperature and temperature to vibration; performing element-by-element multiplication of the two attention weight matrices and generating a fault-sensitive mask through nonlinear activation; splicing the vibration features with the aligned temperature features, performing layer normalization, and performing element-by-element multiplication with the fault-sensitive mask, and generating fault-sensitive features through a temporal convolution network and a random dropout operation.

[0112] S5: Generate the attention weight matrix of vibration to temperature and the attention weight matrix of temperature to vibration based on the temperature time series characteristics and vibration characteristics, calculate the fault sensitive mask, and generate the fault sensitive features;

[0113] S51: Generate aligned temperature features through linear interpolation based on the temperature time series features. The calculation method is:

[0114]

[0115] in, is the aligned temperature feature, Linearlnterp() is the linear interpolation;

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

[0117]

[0118] Among them, A v2tis the attention weight matrix of vibration to temperature, A t2v is the attention weight matrix of temperature to vibration;

[0119] S53: Calculate the fault sensitive mask according to the attention weight matrix of vibration to temperature and the attention weight matrix of temperature to vibration. The calculation method is:

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

[0121] Among them, Mask risk is a fault sensitive mask;

[0122] S54: Generate fault-sensitive features through a gated residual network based on the fault-sensitive mask, vibration features, and aligned temperature features. The calculation method is:

[0123]

[0124] Among them, P is the cross-modal joint feature matrix, Q risk is the fault-sensitive feature, TCN() is the temporal convolutional network, is element-by-element multiplication, and Dropout() is a random drop operation.

[0125] It should be further explained that the present invention designs a bidirectional asymmetric attention mechanism based on 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 an independent convolutional layer, and the temperature-to-vibration path models the offset law of vibration harmonic components caused by the decrease in material stiffness due to temperature rise, and uses the Softmax function to probabilistically express the bidirectional correlation strength; compared with traditional single-modal analysis or simple feature splicing methods, the fault-sensitive mask uses the Hadamard product to filter the overlapping areas of bidirectional attention, which is mathematically equivalent to constructing a convex envelope of fault-related features in the joint time-frequency space, and eliminating negative interference components through the ReLU function, thereby forming a directional enhancement of the composite fault features.

[0126] In actual application scenarios, wind turbine gearbox failures often manifest as a weak correlation between a slow temperature rise and vibration shock. For example, the initial temperature rise of lubricating oil failure lags behind the vibration anomaly. Traditional methods ignore the cross-modal temporal phase difference and lead to misjudgment. The present invention uses a gated residual network to perform residual learning on the mask-weighted features and the cross-modal joint features, and uses the void structure of the time convolutional network to cover the fault evolution pattern. Compared with the fixed-ratio multimodal fusion algorithm, this design mathematically achieves a dynamic balance between the time domain sensitivity of the vibration signal and the cumulative effect of the temperature signal, solving the problem of the traditional method's delayed response to early composite faults.

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

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

[0129] (2) Channel 43 shows a sustained activation of about 1.8 in the 12-20 hour period (orange area), reflecting the temperature-related modulation caused by gear wear; the negative area (blue) in the matrix accounts for about 35%, 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 compression features, then calculate the three-level warning probability vector and the unit fault warning category, and issue an alarm;

[0131] S61: Generate fault compression features through time average pooling based on fault-sensitive features. The calculation method is:

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

[0133] Where Z is the fault compression feature;

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

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

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

[0137] Among them, S warn is the three-level warning probability vector, s norm ,s early ,s emerg are the probability values ​​of normal, early warning, and emergency shutdown respectively;

[0138] S63: The category with the largest probability value in the three-level warning probability vector is selected as the unit fault warning category, and an alarm is issued.

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

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

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

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

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

[0144] Data acquisition and preprocessing module: collects the original vibration waveform data and the original temperature data, and preprocesses them respectively to obtain the 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, short-time Fourier transform and continuous wavelet transform are combined to generate time-frequency spectrum and wavelet time-frequency spectrum. 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: Generates the attention weight matrix of vibration to temperature and the attention weight matrix of temperature to vibration based on the temperature time series features and vibration features, calculates the fault-sensitive mask, and generates fault-sensitive features;

[0149] Three-level warning module: compresses fault-sensitive features to obtain fault compression features, then calculates the three-level warning probability vector and unit fault warning category, and issues an alarm.

[0150] It should be noted that the serial numbers of the above-mentioned embodiments of the present invention are for descriptive purposes only and do not represent the advantages or disadvantages of the embodiments. In addition, the terms "including", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method comprising the element.

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

[0152] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A wind turbine generator set fault warning method, characterized in that: The following steps are involved: S1: collecting vibration waveform raw data and temperature raw data, and preprocessing them respectively to obtain preprocessed vibration time series data and preprocessed temperature time series data; S2: Based on the preprocessed vibration time series data, short-time Fourier transform and continuous wavelet transform are combined to generate time-frequency spectrum and wavelet time-frequency spectrum. Then, a heterogeneous dual-stream network is designed to extract high-frequency impact features and low-frequency modulation features respectively. The heterogeneous two-stream network includes: performing 3×3 depthwise separable convolution, GELU activation, and secondary 3×3 depthwise separable convolution on the time-frequency spectrum, and superimposing the output results of 1×1 depthwise separable convolution and 2×2 maximum pooling to obtain high-frequency impact features; 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 on the result and the output of the 3×3 depthwise separable convolution after Sigmoid activation to obtain low-frequency modulation features; S3: Calculate vibration features based on high-frequency impact features and low-frequency modulation features, including: performing attention calculation on high-frequency impact features, and then combining residual connections and layer normalization to generate high-frequency enhancement features; pooling low-frequency modulation features in time and frequency dimensions, fusing them through convolution and activation functions to generate a gated weight matrix, and then combining dimensional expansion to generate low-frequency enhancement features; finally, generating vibration features through a dynamic weight gating mechanism; S4: extracting the temperature change rate sequence and temperature time series features based on the preprocessed temperature time series data; S5: Generate the attention weight matrix of vibration to temperature and the attention weight matrix of temperature to vibration based on the temperature time series characteristics and vibration characteristics, calculate the fault sensitive mask, and generate the fault sensitive features; S6: Compress the fault-sensitive features to obtain the fault compression features, then calculate the three-level warning probability vector and the unit fault warning category, and issue an alarm.

2. The wind turbine generator set fault early warning method according to claim 1, characterized in that: The S1 step includes: S11: Collecting raw vibration waveform data from a piezoelectric acceleration sensor installed at a measuring point on the gearbox to obtain a vibration time series signal; the sampling rate of the raw vibration waveform data is not less than 20,000 times per second; S12: collecting raw temperature data from temperature sensors provided on the gearbox bearing seat and the gear meshing surface to obtain a temperature time series signal; the sampling rate of the raw temperature data is once per second; S13: For the vibration time series signal, a wavelet threshold noise reduction method is used to remove high-frequency noise to obtain preprocessed vibration time series data; S14: For the temperature time series signal, outliers are eliminated by the interquartile range method, and the minimum-maximum normalization method is used to unify the dimensions 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: The S2 step includes: S21: Generate a time-frequency spectrum X through short-time Fourier transform based on the pre-processed vibration time series data stft , and generate the wavelet time-frequency graph X by continuous wavelet transform cwt ; S22: Based on the time-frequency spectrum and wavelet time-frequency spectrum, a heterogeneous dual-stream network is designed to extract high-frequency impact features and low-frequency modulation features respectively. The calculation method is: F high =DSConv 3×3 (GELU(DSConv 3×3 (X stft )))+MaxPool 2×2 (DSConv 1×1 (X stft )); F low =DSConv 5×5 (ReLU(AvgPool 2×2 (X cwt )))⊙Sigmoid(DSConv 3×3 (X cwt )); Among them, F high is the high-frequency impact feature, DSConv 3×3 () is a depth-separable convolution with a kernel size of 3×3, GELU() is a GELU activation function, and MaxPool 2×2 () is the maximum pooling with a pooling window size of 2×2, DSConv 1×1 () is a depth-wise separable convolution with a kernel size of 1×1, F low is the low-frequency modulation feature, DSConv 5×5 () is a depth-separable convolution with a kernel size of 5×5, ReLU() is the ReLU activation function, and AvgPool 2×2 () is the average pooling with a pooling window size of 2×2, ⊙ is the Hadamard product, and Sigmoid() is the Sigmoid activation function.

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

5. The wind turbine generator set fault warning method according to claim 4, characterized in that: The S3 step includes: S31: Generate high-frequency enhancement features and low-frequency enhancement features based on high-frequency impact features and low-frequency modulation features. The calculation method is: F′ high =LayerNorm(F high +M high ·Conv(F high )); G=Sigmoid(Conv(TimeAvgPool(F low )+FreqMaxPool(F low ))); F′ low =F low ⊙Expand(G); Among them, M high is the high-frequency attention matrix, Softmax() is the Softmax function, Conv() is the convolution layer, DSConv() is the depth-wise separable convolution, is the matrix transpose, d is the high-frequency impact feature dimension, F′ high is the high-frequency enhancement feature, LayerNorm() is the layer normalization, G is the low-frequency gating weight matrix, TimeAvgPool() is the time dimension average pooling, FreqMaxPool() is the frequency dimension maximum pooling, F′ low is the low-frequency enhancement feature, and Expand() is the vector dimension expansion operation; S32: Generate vibration features based on the high-frequency enhancement features and the low-frequency enhancement features through a dynamic weight gating mechanism. The calculation method is: α=Softmax(FC(Concat(AvgPool(F′ high ),AvgPool(F′ low )))); F fusion =α·F′ high +(1-α)·F′ low ; Among them, α is the dynamic fusion weight, FC() is the fully connected layer, Concat() is the splicing operation, AvgPool() is the average pooling, F fusion Vibration characteristics.

6. The wind turbine generator set fault early warning method according to claim 5, characterized in that: The S4 step comprises: S41: Calculate the temperature change rate sequence based on the preprocessed temperature time series data using a sliding window difference method. Specifically, the preprocessed temperature time series data is intercepted with a fixed window length, and 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 divided by the sampling interval to obtain the temperature change rate sequence. S42: Generate temperature time series features using a dynamic convolution kernel based on the preprocessed temperature time series data and the temperature change rate sequence. The calculation method is: K dynamic =Softplus(Conv1D(R′)); F temp =DepthwiseConv(R,K dynamic ); Among them, K dynamic is the temperature dynamic convolution kernel parameter, Softplus() is the Softplus activation function, Conv1D() is the one-dimensional convolution layer, R′ is the temperature change rate sequence, F temp is the temperature time series feature, and R is the preprocessed temperature time series data.

7. The wind turbine generator set fault warning method according to claim 6, characterized in that: Generating fault-sensitive features in the 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 convolution layer respectively to generate attention weight matrices of vibration to temperature and temperature to vibration; performing element-by-element multiplication of the two attention weight matrices and generating a fault-sensitive mask through nonlinear activation; concatenating the vibration features with the aligned temperature features, performing layer normalization, and performing element-by-element multiplication with the fault-sensitive mask, and generating the fault-sensitive features through a temporal convolution network and a random dropout operation.

8. The wind turbine generator set fault warning method according to claim 7, characterized in that: The step S5 comprises: S51: Generate aligned temperature features through linear interpolation based on the temperature time series features. The calculation method is: in, is the aligned temperature feature, Linearlnterp() is the linear interpolation; S52: Based on the aligned temperature and vibration features, a bidirectional asymmetric attention mechanism is designed to generate the attention weight matrix of vibration to temperature and the attention weight matrix of temperature to vibration. The calculation method is: Among them, A v2t is the attention weight matrix of vibration to temperature, A t2v is the attention weight matrix of temperature to vibration; S53: Calculate the fault sensitive mask according to the attention weight matrix of vibration to temperature and the attention weight matrix of temperature to vibration. The calculation method is: Maske risk =ReLU(A v2t ⊙A t2v ); Among them, Mask risk is a fault sensitive mask; S54: Generate fault-sensitive features through a gated residual network based on the fault-sensitive mask, vibration features, and aligned temperature features. The calculation method is: Among them, P is the cross-modal joint feature matrix, Q risk is the fault-sensitive feature, TCN() is the temporal convolutional network, is element-by-element multiplication, and Dropout() is a random drop operation.

9. The wind turbine generator set fault early warning method according to claim 8, characterized in that: The step S6 comprises: S61: Generate fault compression features through time average pooling based on fault-sensitive features. The calculation method is: Z=AvgPool(Conv1D(Q risk )); Where Z is the fault compression feature; S62: Based on the fault compression characteristics, a three-level warning probability vector is generated through a fully connected network. The calculation method is: With warn =Softmax(FC(Z)); S warn =[s norm ,s early ,s emerg ]; Among them, S warn is the three-level warning probability vector, s norm ,s early ,s emerg are the probability values ​​of normal, early warning, and emergency shutdown respectively; S63: The category with the largest probability value in the three-level warning probability vector is selected as the unit fault warning category, and an alarm is issued.

10. A wind turbine generator fault warning system, characterized in that: include: Data acquisition and preprocessing module: collects the original vibration waveform data and the original temperature data, and preprocesses them respectively to obtain the 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, short-time Fourier transform and continuous wavelet transform are combined to generate time-frequency spectrum and wavelet time-frequency spectrum. 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: Generates the attention weight matrix of vibration to temperature and the attention weight matrix of temperature to vibration based on the temperature time series features and vibration features, calculates the fault-sensitive mask, and generates fault-sensitive features; Level 3 warning module: compresses fault-sensitive features to obtain fault compression features, calculates level 3 warning probability vectors, unit fault warning categories, and issues alarms. To realize a wind turbine generator set fault early warning method as described in any one of claims 1-9.

Images