Generator set fault diagnosis and detection method based on deep learning

Abstract

The invention relates to the technical field of motor detection, in particular to a generator set fault diagnosis and detection method based on deep learning, and the method comprises the steps: firstly, synchronously collecting three-phase voltage, current and rotating speed signals, and constructing a multi-dimensional time sequence matrix through data cleaning and sliding window segmentation; then, carrying out multi-scale decomposition on the matrix by adopting adaptive wavelet packet transformation, combining each frequency band reconstruction coefficient with an original signal channel, and constructing an enhanced feature tensor; then, a CNN-BiLSTM parallel network is constructed; and finally, dynamically fusing the features of the two branches through a self-adaptive weighted fusion strategy, and inputting a multi-layer full-connection classifier to output a fault type. According to the method, early weak fault features are effectively enhanced, bearing faults, rotor eccentricity, electrical imbalance and composite faults thereof can be accurately recognized, the intelligent level and accuracy of fault diagnosis of the generator set are remarkably improved, and the method can be widely applied to online monitoring and health management of power generation equipment of a power system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motor detection, and particularly relates to a generator set fault diagnosis and detection method based on deep learning. BACKGROUND

[0002] As the core equipment of the power system, the operation state of the generator set is directly related to the stability and safety of the power grid. The traditional generator set fault diagnosis method mainly relies on threshold comparison and simple spectrum analysis. For example, whether the bearing is abnormal is judged by monitoring whether the amplitude of the vibration signal is over-standard, or the degree of electrical imbalance is evaluated by analyzing the harmonic content of the voltage and current. However, these methods have obvious limitations:

[0003] The traditional method needs engineers to manually design features (such as the amplitude of a specific frequency band, kurtosis index, etc.) according to prior knowledge, which is not only inefficient, but also difficult to mine deep implicit information in the signal.

[0004] The electrical parameters of the generator present non-stationary characteristics under the conditions of start-stop, load change, etc. The traditional Fourier transform is difficult to simultaneously consider the local changes of time and frequency, resulting in a decrease in diagnostic accuracy.

[0005] The diagnosis model based on shallow machine learning (such as support vector machine) often has insufficient generalization ability when facing complex nonlinear fault coupling relationships, and is prone to false positives or false negatives.

[0006] With the development of artificial intelligence technology, although some research attempts to introduce deep learning, most of them are still at the theoretical simulation stage, and there is a lack of an end-to-end solution that can effectively fuse multi-source electrical parameters, automatically extract spatiotemporal features, and accurately identify complex faults. SUMMARY

[0007] The purpose of the present application is to provide a generator set fault diagnosis and detection method based on deep learning, which can automatically learn and identify deep features that can represent different fault types from the electrical parameters of the generator, thereby realizing high-precision and intelligent detection of production defects such as bearing faults, rotor eccentricity, and electrical imbalance.

[0008] To achieve the above purpose, the present application provides a generator set fault diagnosis and detection method based on deep learning, comprising the following steps:

[0009] Synchronously collecting the stator three-phase voltage, stator three-phase current and rotor speed analog signal of the generator set, performing data cleaning and sliding time window segmentation on the multi-channel digital signal sequence obtained after preprocessing, and obtaining a multi-dimensional time sequence matrix;

[0010] performing wavelet packet transform on each column signal in the multi-dimensional time sequence matrix, merging reconstruction coefficients of a plurality of frequency bands obtained by the wavelet packet transform with the multi-dimensional time sequence matrix in a channel dimension, and constructing an enhanced feature tensor;

[0011] inputting the enhanced feature tensor into two parallel feature extraction branches respectively to perform feature extraction, to obtain a local spatial feature vector and a global time sequence feature vector, wherein the two parallel feature extraction branches are a convolutional neural network branch and a bidirectional long short-term memory network branch respectively;

[0012] performing adaptive weighted fusion on the local spatial feature vector and the global time sequence feature vector, inputting a fusion feature vector obtained by the fusion into a multi-layer fully connected classifier, outputting a fault type probability distribution, and completing diagnosis and detection of the generator set fault;

[0013] The extraction process of the convolutional neural network branch specifically includes:

[0014] A multi-scale one-dimensional convolution layer is used to perform parallel convolution operation on the enhanced feature tensor using convolution kernels with sizes of 3x1, 7x1 and 15x1 respectively, and the outputs of the three convolutions are spliced in the channel dimension; then a plurality of one-dimensional convolution layers and maximum pooling layers are stacked for deep feature extraction; finally, a global maximum pooling layer is used to aggregate the feature maps into a local spatial feature vector;

[0015] The extraction process of the bidirectional long short-term memory network branch specifically includes:

[0016] The enhanced feature tensor is input into a bidirectional LSTM layer as a time step sequence to obtain a hidden state sequence containing forward and backward context information; an attention weight of each time step is calculated through a time attention mechanism, and the hidden state sequence is weighted and summed to be aggregated into a global time sequence feature vector;

[0017] The adaptive weighted fusion on the local spatial feature vector and the global time sequence feature vector includes:

[0018] The local spatial feature vector and the global time sequence feature vector are respectively input into independent single-layer fully connected networks to be mapped into two scalar importance scores; the two scores are spliced and then normalized by a Softmax function to obtain fusion weights of the two features; the two feature vectors are weighted and summed using the fusion weights to obtain a fusion feature vector.

[0019] The synchronous acquisition is implemented in the following manner:

[0020] A multi-channel synchronous data acquisition card is adopted, a built-in constant temperature crystal oscillator is used as a main clock, and IEEE 1588 precise time protocol or external hardware triggering is supported, so that the error of sampling time of each channel is less than 1 microsecond, and the time axis alignment of three-phase voltage, three-phase current and rotating speed signals is obtained.

[0021] The multi-channel digital signal sequence obtained after preprocessing is subjected to data cleaning and sliding time window segmentation, including:

[0022] The method of sliding window combined with median absolute deviation is adopted to detect and eliminate abnormal values in the digital signal sequence, and the median in the window is used to replace it.

[0023] The method of polynomial fitting or high-pass filtering is adopted to remove the set change trend item in the signal; and an adaptive notch filter is used to filter out fixed power frequency and harmonic interference.

[0024] The wavelet packet transform uses an adaptively selected wavelet basis function and decomposition layer number; the adaptive selection includes: constructing a candidate wavelet basis library, calculating the kurtosis and the weighted comprehensive evaluation index of energy retention rate of the reconstructed signal of each frequency band after each candidate wavelet basis decomposes the historical fault data, and selecting the wavelet basis corresponding to the maximum value of the index as the optimal wavelet basis; for the selected optimal wavelet basis, the number of decomposition layers is increased, and the information entropy of the energy distribution of all frequency bands of each layer is calculated, and the number of layers at which the first inflection point appears in the entropy value growth curve is selected as the best decomposition layer number.

[0025] The reconstructed coefficients of the plurality of frequency bands obtained are merged with the multi-dimensional time sequence matrix in the channel dimension to obtain a multi-dimensional time sequence matrix, including:

[0026] The 2^J frequency band reconstructed signals obtained after wavelet packet transform of each original channel signal are used as new channels, which are concatenated according to the original channel grouping, and are spliced with the original multi-dimensional time sequence matrix in the channel dimension to form an enhanced feature tensor with a dimension of L × (7 + 7×2^J), wherein L is the length of the time window, and J is the number of decomposition layers.

[0027] The multi-layer fully connected classifier specifically includes:

[0028] The first fully connected hidden layer, the second fully connected hidden layer and the output layer are sequentially connected; the first fully connected hidden layer and the second fully connected hidden layer both include a ReLU activation function and a Dropout operation immediately after it; the number of neurons of the output layer is equal to the total number of preset fault categories, and the output thereof is converted into the probability of each fault category through a Softmax function.

[0029] The method further includes:

[0030] A training sample set containing historical fault data is constructed, each sample is labeled with a corresponding fault type label, and the fault type includes normal state, bearing fault, rotor eccentricity, electrical imbalance and multiple composite faults, and the multiple composite faults refer to a fault mode containing two or more single fault types at the same time;

[0031] The training sample set is input into a deep learning network composed of two parallel feature extraction branches and a multi-layer fully connected classifier, cross-entropy loss function and Adam optimizer are used for back propagation training, and early stopping strategy is monitored by a validation set to prevent overfitting until the model converges.

[0032] The generator set fault diagnosis and detection method based on deep learning provided by the present application is characterized in that, in order to solve the problems of existing methods that feature extraction relies on manual work, it is difficult to process non-stationary signals and the recognition rate of composite faults is low, first, three-phase voltage, current and speed signals are synchronously collected, a multi-dimensional time sequence matrix is constructed through data cleaning and sliding window segmentation; then, the matrix is decomposed in multiple scales by using adaptive wavelet packet transform, the reconstruction coefficients of each frequency band are combined with the original signal channel to construct an enhanced feature tensor; then, a CNN-BiLSTM parallel network is constructed, the CNN branch extracts local spatial features, and the BiLSTM branch extracts global time sequence features combined with an attention mechanism; finally, the features of the two branches are dynamically fused through an adaptive weighted fusion strategy, and a multi-layer fully connected classifier is input to output the fault type. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced.

[0034] Figure 1 is a step schematic diagram of a generator set fault diagnosis and detection method based on deep learning according to the first embodiment of the present application.

[0035] Figure 2 is a flowchart of a generator set fault diagnosis and detection method based on deep learning provided by the present application.

[0036] Figure 3 is a flowchart of a generator set fault diagnosis and detection method based on deep learning provided by the present application, including model training. DETAILED DESCRIPTION

[0037] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The description herein relates to the drawings, in which the same numbers represent the same or similar elements, unless otherwise represented. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application.

[0038] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0039] It is to be understood that the singular forms "a", "an", and "the" include plural referents unless the context clearly dictates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0040] Reference will now be made to Figures 1-3 The present application provides a deep learning-based generator set fault diagnosis and detection method, comprising the following steps:

[0041] S101, synchronously collect the stator three-phase voltage, stator three-phase current and rotor speed analog signal of the generator set, perform data cleaning and sliding time window segmentation on the obtained multi-channel digital signal sequence after preprocessing, and obtain a multi-dimensional time sequence matrix.

[0042] Specifically, in order to realize the accurate alignment of voltage, current and speed signals on the time axis, the application adopts a data acquisition system based on high-precision clock synchronization. Three-phase voltage transformers and current transformers are installed on the stator side of the generator set to obtain three-phase stator voltages U a , U b , U c and three-phase stator currents I a , I b , I cA photoelectric encoder or magnetoelectric speed sensor is installed at the rotor shaft end to acquire the instantaneous rotor speed signal n. All analog signals output from the sensors are connected to a multi-channel synchronous data acquisition card. This card has a built-in high-stability temperature-controlled crystal oscillator as the master clock and supports the IEEE 1588 precise time protocol or external hardware triggering, ensuring that each channel samples the signal at the same time (with an error of less than 1 microsecond), fundamentally eliminating phase deviation and timing misalignment problems caused by asynchronous sampling. The sampling frequency is set according to the highest possible fault frequency component of the generator (such as the characteristic frequency of bearing faults), typically not lower than 10 kHz, to satisfy the Nyquist sampling theorem.

[0043] The raw analog signals output by sensors often have small amplitudes and are mixed with high-frequency noise, making direct analog-to-digital conversion impossible. Therefore, before entering the data acquisition card, they are pre-processed by a signal conditioning module, including impedance matching, differential amplification (amplifying the weak signal to the full-scale range of the ADC), and isolation protection (preventing high-voltage surges from damaging subsequent circuitry). The conditioned signal is then fed into an anti-aliasing filter, a low-pass filter with a cutoff frequency set to half the sampling frequency to filter out high-frequency noise above the Nyquist frequency and prevent spectral aliasing after sampling. The filtered analog signal is finally quantized into a discrete digital signal sequence by a high-precision analog-to-digital converter (ADC), yielding the raw sampled data points for each channel, denoted as U. a [t],U b [t],U c [t],I a [t],I b [t],I c [t], n[t], where t represents the discrete-time index.

[0044] Although out-of-band noise has been suppressed (i.e., denoising) through anti-aliasing filtering, digital signals may still contain outliers (outliers) and trend terms introduced by sensor transient interference, electromagnetic pulses, or data transmission errors. Data cleaning here is performed on digital signal sequences to remove these unwanted components and improve signal quality. The following methods are used specifically:

[0045] Outlier removal: For each channel's sampling sequence, a sliding window combined with the median absolute deviation method is used for detection. The median of the data is calculated within the window, and the standard deviation is estimated. If the amplitude of a sampling point deviates from the median by more than a preset multiple (e.g., 3 times the standard deviation), it is identified as an outlier and replaced with the median within the window to avoid interfering with subsequent analysis.

[0046] Trend removal: Due to sensor zero-point drift or ambient temperature changes, a slowly varying trend may be superimposed on the signal. A polynomial fitting or high-pass filtering method is used to remove the trend, making the signal fluctuate around zero mean, meeting the requirement of deep learning model for input data distribution stability.

[0047] Power frequency notch: For voltage and current signals, an adaptive notch filter is used to filter out the fixed 50 Hz (or 60 Hz) power frequency and its harmonic interference, while retaining the non-power frequency components related to faults.

[0048] After the above cleaning, a pure multi-channel digital signal sequence is obtained.

[0049] In order to construct samples suitable for deep learning model training, continuous multi-channel signals need to be segmented. A sliding time window method is used to cut long time series into multiple equal length short time segments. The selection of window length L needs to consider the time domain persistence and frequency resolution of the fault, usually covering several power frequency periods (such as 0.2 seconds, i.e. 10 cycles) to fully capture the periodic characteristics of the fault. The window sliding step S is generally less than the window length, so that there is an overlap between adjacent windows (the overlap rate is usually set to 50%), thereby increasing the number of samples and improving the robustness of model training. For each window starting point, all channel data with length L starting from the point are intercepted.

[0050] The data intercepted by each window contains the time series of all 7 channels (three-phase voltage, three-phase current, speed) in the window. It is organized in the form of a two-dimensional matrix, where the rows correspond to time sampling points (a total of L rows), and the columns correspond to different signal channels (a total of 7 columns). This matrix is called a multi-dimensional time series sample, denoted as X ∈ R L×7 Each row represents the instantaneous value of all sensor measurement points at a certain time, and each column represents the change trajectory of a measurement point in the entire time window. This structure completely retains the time correlation of the signal and the spatial correlation between channels, laying a foundation for subsequent deep learning models to extract spatiotemporal coupling features. All window-generated samples together constitute the original sample set, along with the corresponding operating condition labels (such as load level) and fault labels (annotated by subsequent steps) for model training.

[0051] S102, wavelet packet transform is performed on each column signal in the multi-dimensional time series matrix, and the reconstruction coefficients of the plurality of frequency bands obtained are merged with the multi-dimensional time series matrix in the channel dimension to construct an enhanced feature tensor.

[0052] Specifically, the wavelet packet transform (WPT) is an extension of the discrete wavelet transform. It not only decomposes the low-frequency part of the signal but also recursively decomposes the high-frequency part, thereby achieving uniform division of the full frequency band of the signal. For a discrete signal x(t) of length L, after J layers of wavelet packet decomposition, 2 J bandwidth frequency band subspaces can be obtained, each corresponding to a set of wavelet packet coefficients. By single-branch reconstruction of the frequency band coefficients (i.e., retaining the frequency band coefficients and setting other frequency band coefficients to zero, and then performing inverse transformation), a time-domain reconstructed signal of the same length as the original signal can be obtained, which represents the component of the original signal in a specific frequency band. This decomposition method can separate weak local features (such as high-frequency impacts caused by bearing faults and low-frequency fluctuations caused by rotor eccentricity) hidden in wideband signals into their respective narrow frequency bands, thereby effectively enhancing the recognizability of fault features.

[0053] The selection of wavelet basis functions and the number of decomposition layers directly affects the enhancement effect of wavelet packet transform on fault features. To match the signal decomposition with the fault characteristics of the generator set, the present application introduces an adaptive selection mechanism based on fault feature saliency.

[0054] Selection of wavelet basis functions: First, a candidate wavelet basis library is constructed, containing commonly used wavelet families with different compact support, vanishing moments, and symmetry, such as Daubechies (dbN), Symlets (symN), Coiflets (coifN), etc. For each candidate wavelet basis, typical samples (covering bearing faults, rotor eccentricity, electrical imbalance, etc.) in historical fault data are used for J-layer wavelet packet decomposition (J is temporarily set to a default value, such as 3 layers), and the fault impact index of each frequency band reconstructed signal is calculated, which is the sum of the absolute values of the kurtosis of each frequency band signal. Kurtosis is sensitive to transient impacts in the signal and can reflect local impact features produced by bearing faults, etc. The ratio of the energy of each frequency band to the total energy of the original signal is also calculated to evaluate the degree of preservation of fault information by decomposition. By combining the kurtosis and energy preservation rate of all frequency bands, a comprehensive evaluation index Q = α × K sum + β × E ratio is constructed, where α and β are weight coefficients optimized through grid search combined with the validation accuracy of the subsequent classification model. The wavelet basis with the maximum Q value is selected as the optimal wavelet basis for the current scenario. This process can be completed during offline model training, and once selected, it is fixed for online diagnosis.

[0055] Determination of decomposition layers: The number of decomposition layers J determines the fineness of adaptive frequency band division. Too few layers, it is difficult to separate different fault characteristics into different frequency bands; too many layers, each frequency band is too narrow, easy to introduce noise and increase the amount of calculation. The stability criterion based on frequency band energy entropy is used to determine the optimal decomposition layer number. For the selected optimal wavelet basis, start from J=1 and increase the decomposition layer number, and perform wavelet packet decomposition on the typical fault sample. The energy of the reconstructed signal of all frequency bands in each layer is calculated, and the information entropy of the energy distribution is calculated:

[0056]

[0057] Wherein, is the proportion of the energy of the i-th frequency band to the total energy. With the increase of J, the energy distribution is gradually refined, and the entropy value H(J) shows an upward trend; when J reaches a certain value, the frequency band division is too fine, which leads to energy dispersion, and the entropy value growth tends to be saturated or even fluctuates. Select the J at which the first inflection point (i.e. the first-order difference is less than the threshold value) appears in the entropy growth curve as the best decomposition layer number, which can ensure sufficient frequency band resolution and avoid instability caused by over-decomposition. Similarly, the selection process is completed offline.

[0058] For a multi-dimensional time series matrix X∈R L×7 , the column vectors correspond to three-phase voltages (U a , U b , U c ), three-phase currents (I a , I b , I c ) and speed (n) respectively, a total of 7 channels. For each column signal, independently perform J-layer wavelet packet decomposition with adaptive selected wavelet basis, and the specific operation is as follows:

[0059] Take the c-th column signal x c (t) (length L) as an example:

[0060] Perform J-layer wavelet packet decomposition on x c (t) to obtain 2 J wavelet packet coefficients of nodes.

[0061] For each node i (i=1, 2, …, 2J), perform single-branch reconstruction on the coefficients: keep the node coefficients, set the rest of the node coefficients to zero, and then perform wavelet packet inverse transform to obtain a time-domain reconstructed signal r c (t) with the same length as x c,i (t), which represents the component of the original signal in the corresponding frequency band.

[0062] Repeat the above process for all 7 channels to finally obtain 7×2 J frequency band reconstructed signals.

[0063] In order to simultaneously retain the original full-band information and the detailed information of each frequency band, the application merges the original signal and all the frequency band reconstruction signals in the channel dimension to form an enhanced feature tensor (which can also be referred to as an enhanced tensor). The specific merging method is as follows:

[0064] The original multi-dimensional time sequence matrix X itself serves as 7 basic channels.

[0065] For the 2 J frequency band reconstruction signals r c,i (t) generated by the c c th original channel (c = 1, 2,..., 7), regard them as new independent channels and arrange them in the channel order to form a frequency band channel matrix R c , that is, R c,1 = [r c,2 , r c,2 ,..., r

[0066] Serially concatenate all the new channels according to the original channels to finally obtain an enhanced feature tensor X J with a dimension of L x (7 + 7 x 2 J ), that is, X enhanced = [X, R1, R2,..., R7], R enhanced = [r c , r c,1 ,..., r c,2 c,2 ] is the frequency band channel matrix expanded by the c J th original channel, with a dimension of L x 2 J .

[0067] Each row of the enhanced feature tensor X enhanced still represents the observation value at the same time, but the observation dimension is expanded from the original 7 dimensions to 7 (1 + 2 J ) dimensions, containing the original signal and its projection on each fine frequency band. The fusion of such multi-scale information enables the subsequent deep learning model (CNN-BiLSTM) to simultaneously learn the deep features of faults from the wide frequency profile and the narrow frequency details, greatly improving the sensitivity and recognition ability of the model to weak faults and compound faults.

[0068] S103, input the enhanced feature tensor into two parallel feature extraction branches respectively for feature extraction to obtain a local spatial feature vector and a global time sequence feature vector, wherein the two parallel feature extraction branches are a convolutional neural network branch and a bidirectional long short-term memory network branch respectively.

[0069] Specifically, the enhanced feature tensor output by the above step is denoted as X enhanced ∈R L×Cwhere L is the length of the time window (number of samples), C = 7 x (1 + 2 J ) is the total number of channels (including original channels and all band- reconstructed channels). This tensor will serve as the input for both branches. Among them, the convolutional neural network branch focuses on mining the spatial morphological features of the signal within the local time window, i.e., the transient waveform distortion caused by the fault; the bidirectional long short-term memory network branch focuses on capturing the global temporal dependence of the signal over time, i.e., the dynamic change rule of the fault. The outputs of the two branches are fused through feature concatenation to form a comprehensive representation of the generator set operating state.

[0070] First, in the convolutional neural network branch (also the CNN path in the multi-scale convolution process), one-dimensional convolution operation is used to extract local correlation patterns within each channel and across channels, similar to identifying "local shapes" in the signal. The specific structure design is as follows: Figure 2

[0071] First, X enhanced is directly input into the first set of parallel one-dimensional convolution layers. In order to capture local features of different scales, a multi-scale convolution kernel design is adopted. Specifically, it contains three parallel convolution operations:

[0072] Small-scale convolution kernel: size 3 x 1 (i.e., convolution kernel length 3, acting on the time dimension, input channel number C, and output channel number set as F1). This convolution kernel is good at extracting subtle local fluctuations, such as high-frequency impact details caused by initial bearing fault.

[0073] Medium-scale convolution kernel: size 7 x 1, output channel number set as F2. Used to capture medium-length waveform patterns, such as periodic harmonic envelopes caused by electrical imbalance.

[0074] Large-scale convolution kernel: size 15 x 1, output channel number set as F3. Capable of covering trend features over a longer time range, such as slow drift caused by rotor eccentricity.

[0075] Each convolution operation uses a step size of 1 and a padding method of same (i.e., zero padding to keep the output time length L). After each convolution layer, a batch normalization layer is added to standardize the feature maps of each output channel, accelerate training convergence, and alleviate overfitting. Then, a ReLU activation function is introduced to introduce nonlinearity. Finally, the outputs of the three parallel convolutions are concatenated in the channel dimension to obtain a feature map that integrates multi-scale local features.

[0076] ​To further abstract the features and reduce the dimensionality, several (e.g., 2) one-dimensional convolutional layers are stacked after the concatenated feature map. The convolutional kernel size of these convolutional layers can be set to a fixed size (e.g., 5x1), and the number of output channels is gradually increased (e.g., from 256 to 512), and each convolutional layer is followed by batch normalization and ReLU activation. After passing through one or two convolutional layers, a one-dimensional max-pooling layer (pooling size of 2, stride of 2) is inserted to downsample the time dimension, reduce the subsequent computational load, and enhance the translation invariance of the features. After several times of pooling, the time dimension is reduced to 1 / 2 of the original k , denoted as L', and the final feature map is M cnn2 ∈R L′×512 .

[0077] To convert the variable-length spatio-temporal feature map into a fixed-length feature vector, a global max-pooling layer is used. This layer takes the maximum value of each channel in the time dimension, aggregates the most representative local features, and obtains a local spatial feature vector F cnn ∈R 512 . This vector condenses the most significant local waveform patterns in the entire time window, providing key detailed evidence for fault type identification.

[0078] In the bidirectional long short-term memory network branch (i.e., the BiLSTM path in the attached Figure 2 , the BiLSTM captures the dependency and evolution of the signal in a long time scale, utilizes the memory capability of BiLSTM for forward and backward information of time series, and combines the attention mechanism to focus on key time steps.

[0079] The enhanced feature tensor X enhanced ∈R L×C is regarded as a time series of length L, and each time step t corresponds to a C-dimensional feature vector x t . That is, the input of the branch is the sequence [x1, x2, …, x L ], and each x t ∈R C , R C indicates that the feature vector at each time step is a C-dimensional real vector, i.e., x t is a C-dimensional real vector.

[0080] This sequence is input into a bidirectional LSTM layer. The layer contains two independent LSTM sub-layers:

[0081] Forward LSTM: reads the sequence in time order from t=1 to L, and outputs the hidden state at each time step t, which integrates information from the past to the current time.

[0082] Backward LSTM: read the sequence in reverse order from t=L to 1, output the hidden state , which integrates the information from future to current time.

[0083] At each time step t, concatenate the forward and backward hidden states to get the complete hidden state of this step , which has a dimension of 2H (H is the hidden layer dimension of LSTM unit). The output of the entire BiLSTM layer is a hidden state sequence . Each h t in this sequence encodes the context information on the entire time period centered at time step t, which can effectively capture the timing evolution patterns of the fault signal, such as periodic repetition and trend change.

[0084] Not all time steps contribute equally to fault diagnosis. To highlight key timing segments (such as the moment of fault occurrence or the position of periodic peaks), a time attention mechanism is introduced to automatically learn the importance weight of each time step. The specific implementation is as follows:

[0085] Take the hidden state sequence H bilstm output by BiLSTM as input, and calculate the original score of each time step through a trainable attention scoring function. For example, use additive attention:

[0086] ;

[0087] Where W is a weight matrix, v is a vector, and b is a bias, all of which are learnable parameters.

[0088] Normalize the original score by Softmax to get the attention weight:

[0089] , which satisfies .

[0090] Use the attention weight to weight the sum of the hidden states to get the global timing feature vector:

[0091] ;

[0092] Its dimension is 2H. This vector is a weighted aggregation of the BiLSTM output sequence, focusing on time points that are more critical to diagnosis, thereby enhancing the discriminability of the features.

[0093] S104, adaptively weight the local spatial feature vector and the global timing feature vector, input the fusion feature vector obtained by fusion into a multi-layer fully connected classifier, output the fault type probability distribution, and complete the diagnosis and detection of the generator set fault.

[0094] Specifically, the two branches extract features that describe the generator's operating state from different perspectives: F cnn The former focuses on the local transient morphology of the signal, and is sensitive to sudden impacts and waveform distortion. bilstm The latter focuses on the global temporal dependence, and is sensitive to periodic changes and trend evolution. The importance of these two features may change dynamically for different fault types or different stages of the same fault. Therefore, simple feature concatenation or fixed weighting cannot fully exploit the complementary advantages of both. This invention introduces a lightweight attention mechanism to dynamically assign fusion weights to two features for each input sample, achieving adaptive fusion.

[0095] First, the local spatial feature vector F cnn ∈R 512 and the global temporal feature vector F bilstm ∈R 2H are input into two independent single-layer fully connected networks, respectively, each mapping to a scalar value:

[0096] ;

[0097] where Linear cnn and Linear bilstm are learnable linear transformations with output dimension 1. These two scalars s cnn and s bilstm represent the raw scores of the importance of the corresponding features to the current sample diagnosis.

[0098] Then, these two scores are concatenated into a two-dimensional vector [s cnn ,s bilstm ], and normalized by the Softmax function to obtain the fusion weights of the two features:

[0099] ;

[0100] The weights and satisfy , which adaptively adjusts the contribution ratio of the two types of features according to the content of the input sample.

[0101] The two feature vectors are weighted and summed using the above weights to obtain the final fusion feature vector F fused :

[0102] ;

[0103] The dimension of the fusion feature vector F fused is the same as F cnn and F bilstmThe same (the dimension must be kept consistent, so when designing, let 2H=512, or linearly transform one of them to align the dimension). Through this adaptive weighted fusion or adaptive fusion dynamic weighting process, the model can dynamically adjust the degree of dependence on local details and global timing information for different fault samples, for example, for early bearing failure, it may give F cnn a higher weight to capture the impact feature, while for rotor eccentricity, it may focus more on the periodic variation information in F bilstm , thereby improving the expression ability of the feature and the accuracy of the diagnosis.

[0104] The fusion feature vector F fused is obtained, which is then sent to a multi-layer fully connected classifier for fault type recognition. The multi-layer fully connected classifier is stacked by multiple fully connected layers, and the structure is as follows:

[0105] The fusion feature vector is first input to the first fully connected layer, which contains 256 neurons to realize further abstraction and dimension reduction of the feature. Its calculation method is:

[0106] h1=ReLU(W1×F fused +b1);

[0107] where W1 and b1 are learnable parameters, and the ReLU activation function introduces nonlinearity. Subsequently, to suppress overfitting, Dropout operation is applied to h1 to randomly discard part of the neuron output with a probability of p=0.5, enhancing the generalization ability of the model.

[0108] The h1 after Dropout is input to the second fully connected layer, which contains 128 neurons, also using ReLU activation function and followed by Dropout, further refining the feature.

[0109] Finally, the output of the second hidden layer is connected to the output layer, and the number of neurons in the output layer is equal to the total number of preset fault categories K (including normal state and each single fault and composite fault). The output layer uses linear activation (i.e., no activation function) to obtain the raw score (logits) of each category:

[0110] z=W3×h2+b3, z∈R K ;

[0111] where K represents the total number of fault categories, W3 represents the weight matrix of the output layer; h2 represents the output of the second hidden layer, which is a feature vector with a dimension of 128; b3 represents the bias vector of the output layer; z represents the raw score vector of the output layer, each element corresponds to the unnormalized score of a category; R K is a K-dimensional real vector space, indicating that z is a K-dimensional real vector, and each component corresponds to the score of a category.

[0112] To convert the scores into a probability distribution, apply the Softmax function to z:

[0113] k = 1, 2, …, K;

[0114] where, represents the probability that the model predicts the sample belongs to the k-th fault class.

[0115] The entire network (including the feature extraction network, the fusion module, and the multi-layer fully connected classifier) is trained in an end-to-end manner. During the training process, the cross-entropy loss function is used to measure the difference between the predicted probability distribution and the true label y (one-hot encoding):

[0116]

[0117] The Adam optimizer is used, with an initial learning rate of 0.001, and the mini-batch gradient descent method is used to update the network parameters. During the training process, the accuracy on the validation set is monitored, and an early stopping strategy is introduced to stop training when the performance on the validation set does not improve for multiple epochs, preventing overfitting. After training is complete, the optimal model is saved for online diagnosis.

[0118] Use historical operation data or fault simulation experimental data to construct a training sample set, each sample is labeled with the corresponding fault type label. Through the back propagation algorithm and the Adam optimizer, the cross-entropy loss function is used as the target to train the CNN-BiLSTM network in an end-to-end manner until the model converges. After training is complete, the model is solidified and deployed to the generator set online monitoring system. In actual operation, the system real-time collects electrical parameters and automatically executes the processing procedures of steps 1 to 4, finally outputs the current fault diagnosis result, realizes the intelligent detection of the health status of the generator set.

[0119] It should be noted that the fault diagnosis and detection method of the present application is divided into two stages of offline training and online diagnosis. As Figure 3 shown by the dashed box in the figure, in the offline training stage, first, a training sample set containing historical fault data is constructed, and the sample is input into the CNN branch, BiLSTM branch, and multi-layer fully connected classifier (FC) Figure 3 ​The deep learning network composed of the CNN and BiLSTM constitutes a full connection classifier in the end-to-end training process. The network is trained by the back propagation algorithm until the model converges and the trained optimal model is saved. In the online diagnosis stage, the real-time collected and preprocessed data are directly input into the trained model, and the diagnosis result, i.e., the probability distribution of the fault type, is finally output through the forward propagation calculation. The "end-to-end training" block in the figure illustrates the joint optimization process, and the "trained model" represents the solidified network parameters for online deployment.

[0120] The present application has the following beneficial effects and innovations:

[0121] End-to-end automatic feature extraction: Without the need for manual design of complex feature extraction rules, the CNN-BiLSTM network automatically learns and mines deep fault features from the original electrical parameters, eliminating the dependence on expert experience and improving the diagnosis efficiency.

[0122] Deep fusion of spatial and temporal features: A parallel dual-branch network structure is innovatively constructed. The CNN branch focuses on extracting the local spatial form features of the signal (such as specific waveform distortion caused by faults), and the BiLSTM branch combines the attention mechanism to focus on capturing the global time-dependent features of the signal (such as the evolution law of the fault over time). The deep fusion of the two makes the model more comprehensive and accurate in describing the nature of complex faults, especially composite faults.

[0123] Signal enhancement and noise resistance: Through wavelet packet transform, the input signal is preprocessed, and the time-domain signal is decomposed into different frequency bands, effectively enhancing the early weak fault features and suppressing noise interference, improving the robustness and early warning ability of the model under complex working conditions.

[0124] High-precision composite fault recognition: By fusing local details and global temporal information, the model can distinguish composite faults with coupled features, solving the problem of low recognition rate of traditional methods in the face of composite faults, and significantly improving the accuracy and reliability of the generator set fault diagnosis.

[0125] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. The specification and examples given herein are not intended to be exhaustive or limiting, and other adaptations and permutations of the application will be apparent to those skilled in the art in light thereof, the intention being to cover all modifications and alternatives within the scope of the application. Various modifications and changes can be made as would be obvious to a person skilled in the art having the benefit of this disclosure, without departing from the true spirit and scope of this application.

[0126] It should be understood that the application is not limited to the precise construction that has been described above and illustrated in the accompanying drawings, and that various modifications and changes can be made without departing from the scope thereof, and it is intended to cover in the appended claims all such modifications and changes as fall within the scope of the application.

Claims

1. A deep learning-based generator set fault diagnosis and detection method, characterized in that, The method comprises the following steps: Synchronously collecting stator three-phase voltage, stator three-phase current and rotor speed analog signals of the generator set, performing data cleaning and sliding time window segmentation on the obtained multi-channel digital signal sequence after preprocessing, and obtaining a multi-dimensional time sequence matrix; Performing wavelet packet transformation on each column signal in the multi-dimensional time sequence matrix, merging the obtained reconstruction coefficients of multiple frequency bands with the multi-dimensional time sequence matrix in the channel dimension, and constructing an enhanced feature tensor; Inputting the enhanced feature tensor into two parallel feature extraction branches respectively for feature extraction to obtain a local spatial feature vector and a global time sequence feature vector, wherein the two parallel feature extraction branches are a convolutional neural network branch and a bidirectional long short-term memory network branch respectively; Performing adaptive weighted fusion on the local spatial feature vector and the global time sequence feature vector, inputting the fused feature vector into a multi-layer fully connected classifier, outputting a fault type probability distribution, and completing diagnosis and detection of the generator set fault; The extraction process of the convolutional neural network branch specifically comprises: A multi-scale one-dimensional convolution layer is used, and convolution kernels with sizes of 3x1, 7x1 and 15x1 are used respectively to perform parallel convolution operations on the enhanced feature tensor, and the outputs of the three convolutions are spliced in the channel dimension; then a plurality of one-dimensional convolution layers and maximum pooling layers are stacked for deep feature extraction; finally, a global maximum pooling layer is used to aggregate the feature maps into a local spatial feature vector; The extraction process of the bidirectional long short-term memory network branch specifically comprises: The enhanced feature tensor is regarded as a time step sequence and input into a bidirectional LSTM layer to obtain a hidden state sequence containing forward and backward context information; the attention weight of each time step is calculated through a time attention mechanism, and the hidden state sequence is weighted and summed to be aggregated into a global time sequence feature vector; The adaptive weighted fusion of the local spatial feature vector and the global time sequence feature vector comprises: The local spatial feature vector and the global time sequence feature vector are respectively input into independent single-layer fully connected networks to be mapped into two scalar importance scores; the two scores are spliced and then normalized by a Softmax function to obtain the fusion weights of the two features; the two feature vectors are weighted and summed by using the fusion weights to obtain a fused feature vector.

2. The deep learning-based generator set fault diagnosis and detection method of claim 1, wherein, The implementation of synchronous acquisition is: A multi-channel synchronous data acquisition card is used, a constant temperature crystal oscillator is built-in as a master clock, and IEEE 1588 precise time protocol or external hardware triggering is supported, so that the sampling time error of each channel is less than 1 microsecond, thereby obtaining time axis aligned three-phase voltage, three-phase current and speed signals. 3.The deep learning-based generator set fault diagnosis and detection method of claim 1, wherein, The data cleaning and sliding time window segmentation of the obtained multi-channel digital signal sequence after preprocessing comprises: An abnormal value in the digital signal sequence is detected and removed by using a sliding window combined with a median absolute deviation, and the median in the window is used to replace it; A polynomial fitting or high-pass filtering method is used to remove the set trend item in the signal; an adaptive notch filter is used to filter out fixed power frequency and harmonic interference.

4. The deep learning-based generator set fault diagnosis and detection method of claim 1, wherein, The wavelet packet transform uses an adaptively selected wavelet basis function and decomposition level; the adaptive selection includes: constructing a candidate wavelet basis library, calculating the weighted comprehensive evaluation index of kurtosis and energy retention rate of each frequency band reconstructed signal after decomposition of each candidate wavelet basis on historical fault data, and selecting the wavelet basis corresponding to the maximum value of the index as the optimal wavelet basis; for the selected optimal wavelet basis, the number of decomposition levels is increased and the information entropy of the energy distribution of all frequency bands at each level is calculated, and the number of levels at which the first inflection point appears in the entropy value growth curve is selected as the best decomposition level.

5. The deep learning-based generator set fault diagnosis and detection method of claim 1, wherein, The reconstructed coefficients of the plurality of frequency bands obtained are merged with the multi-dimensional time series matrix in the channel dimension to obtain a multi-dimensional time series matrix, including: The 2^J frequency band reconstructed signals obtained after wavelet packet transform of each original channel signal are used as new channels, which are concatenated according to the original channel grouping and spliced with the original multi-dimensional time series matrix in the channel dimension to form an enhanced feature tensor with a dimension of L × (7 + 7×2^J), where L is the length of the time window and J is the number of decomposition levels.

6. The deep learning-based generator set fault diagnosis and detection method of claim 1, wherein, The multi-layer fully connected classifier specifically includes: a first fully connected hidden layer, a second fully connected hidden layer and an output layer connected in sequence; the first fully connected hidden layer and the second fully connected hidden layer both include a ReLU activation function and a Dropout operation immediately after the ReLU activation function; the number of neurons of the output layer is equal to the total number of preset fault categories, and the output thereof is converted into the probability of each fault type by a Softmax function.

7. The deep learning-based generator set fault diagnosis and detection method according to any one of claims 1-6, characterized in that, The method further includes: a training sample set containing historical fault data is constructed, each sample is labeled with a corresponding fault type label, and the fault type includes normal state, bearing fault, rotor eccentricity, electrical imbalance and multiple compound faults; the multiple compound faults refer to a fault mode containing two or more than two single fault types at the same time; the training sample set is input into a deep learning network composed of two parallel feature extraction branches and a multi-layer fully connected classifier, cross-entropy loss function and Adam optimizer are used for back propagation training, and early stopping strategy is monitored through a validation set to prevent overfitting until the model converges.

Images