Hydroelectric generating set vibration information fusion fault diagnosis method based on deep learning
By integrating multi-channel vibration signals from hydropower units using deep learning technology and employing 2D convolutional neural networks for end-to-end fault diagnosis, the complexity and single-signal dependence of fault diagnosis in existing technologies for hydropower units are solved, thereby improving the accuracy and adaptability of the diagnosis.
Patent Information
- Application Number
- CN202410978525.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-11-12
AI Technical Summary
Existing fault diagnosis technologies for hydropower units rely on manual analysis, which makes it difficult to effectively process complex monitoring signals. Furthermore, multi-signal diagnosis relies on a single signal, resulting in limited diagnostic effectiveness and making it easy to overlook weak faults.
A deep learning-based fault diagnosis method for hydropower unit vibration information fusion is adopted. It uses a 2D convolutional neural network for end-to-end processing, integrates multi-channel vibration signals, performs multi-scale feature extraction and fusion, and combines multiple softmax outputs for fault identification and diagnosis.
It improved the accuracy and reliability of fault diagnosis, reduced noise interference, enhanced adaptive diagnostic performance, and achieved safe and stable operation of hydropower units.
Smart Images

Figure CN118940213A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of hydropower unit fault diagnosis, and in particular to a hydropower unit vibration information fusion fault diagnosis method based on deep learning. Background Art
[0002] As the core equipment of hydropower stations, the safe and stable operation of hydropower units is crucial to the stability of the entire power system. However, the safety management and fault diagnosis of hydropower units face a series of challenges, especially in the new energy, the randomness and volatility of wind and photovoltaic power generation pose a challenge to the stability of the power system. As a mature clean energy, hydropower has the characteristics of flexible regulation and rapid action, and plays a key role in the complementary development of new energy. Therefore, the fault diagnosis technology of hydropower units has become particularly important.
[0003] The original monitoring signal of the hydropower unit is very complex and has a high signal dimension, making it difficult to identify faults manually. The main implementation process of the existing hydropower unit fault diagnosis technology includes signal processing, feature construction, feature classification, and diagnosis prediction. Starting from the original unit monitoring signal, data preprocessing, denoising and reconstruction analysis are required, and then various signal processing technologies are used to transform the signal into a feature representation with mathematical and physical meanings. These features need to be able to fully reflect the key information of various fault state signals of the unit and have sufficient distinguishability. However, obtaining certain features through mathematical and physical analysis by humans requires high theoretical requirements and practical experimental experience, and it cannot be guaranteed that the constructed features have good diagnostic effects. Moreover, the fault signals of hydropower units are usually weak, easily ignored, and may even be discarded as noise. The hydropower unit state monitoring system has multiple monitoring signal acquisition channels, but most hydropower unit fault diagnosis only uses one signal as the diagnosis basis, so the diagnostic effect is limited.
[0004] Therefore, developing a vibration information fusion fault diagnosis method for hydropower units based on deep learning can not only improve the safety management level of hydropower units, but also use deep learning technology to process complex monitoring signals, extract key features, improve the accuracy of fault diagnosis, and ensure the stable operation of hydropower stations. Summary of the invention
[0005] The present invention aims to address the technical defects of the prior art and provide a hydropower unit vibration information fusion fault diagnosis method based on deep learning, which uses deep learning technology to form end-to-end processing from signal input to diagnosis result, thereby improving the accuracy, reliability and adaptive diagnosis performance of fault diagnosis.
[0006] The present invention provides the following technical solution: a hydropower unit vibration information fusion fault diagnosis method based on deep learning, the method comprising the following steps:
[0007] Step 1: Data collection and preprocessing;
[0008] Acquire monitoring data of vibration signals at the guide bearing of the main shaft of the hydropower unit as raw data and construct a data set, wherein the vibration signals include orthogonal multi-channel vibration signals of spatial vibration information;
[0009] The vibration signal is divided into several sample segments, the length of the sample segment is 1024, each sample segment contains several vibration cycles, each sample segment is normalized, and then features are extracted from the sample segments. The extracted features represent the characteristics of the fault signal, and the fault sample is determined according to the fault signal and a corresponding fault label is added;
[0010] Step 2: Construction of deep diagnostic model;
[0011] 2D convolution is used for feature extraction. By integrating independent and complementary fault information in the monitoring data of multiple vibration signals, an end-to-end multi-input fault diagnosis model is established. The multi-input fault diagnosis model simultaneously inputs the vibration signals obtained in orthogonal directions and performs multi-scale feature extraction and fusion. For each input channel, a length of n is first 2 The original input signal is converted into a two-dimensional data graph, and then several features in the original signal are obtained through a multi-scale feature extraction structure. The two input branches adopt the same symmetrical structure, and a dual-channel parallel processing framework is established to perform channel-related feature extraction and channel-level fusion. Then, based on the fusion features, a deep residual learning structure is constructed to perform deep feature extraction. The high-level features are activated by softmax through neuron combination to obtain the fault probability prediction distribution vector, and the diagnosis prediction result is obtained;
[0012] Step 3: Diagnosis based on multiple softmax outputs;
[0013] Based on the model established in step 2, a bypass classification structure directly connected to the single-branch vibration feature is added, the intermediate-level features are used for fault identification, the convolution extraction structure is used to reduce the dimension of the features, and the softmax is used to map the high-level features into the probability vector of the corresponding fault;
[0014] Step 4: Diagnostic model training and optimization process;
[0015] The diagnostic model constructed according to steps 2-3 includes two vibration input branches and three softmax probability vector outputs. The model is trained according to the three softmax probability vector outputs, a cross entropy loss is established for each output probability vector, and model parameters are adjusted through forward propagation calculation and reverse gradient propagation training. The multi-scale feature extraction output loss weight in step 2 is 1, and the output loss weight of the two connected to the single vibration input feature in step 3 is 0.3. After multiple iterative trainings, a training model is obtained;
[0016] Step 5: Fault output probability joint decision diagnosis;
[0017] According to the latest monitoring data of the vibration signal at the guide bearing of the main shaft of the hydropower unit, the training model is input. The forward calculation process of the training model is used to gradually extract features and identify faults, and three softmax probability vectors are output to predict the fault probability. The corrected fault probability is obtained by summing up the output probability vectors with three groups of dimensions as fault categories corresponding to the fault categories. The comprehensive fault identification results take the category with the largest joint probability as the output fault.
[0018] Preferably, in step 2, the original data is directly received as the model input, and a 2D data conversion relationship is constructed. The vibration signal of each input channel,
[0019] X k =x k (i),k=1,2,i=1,2,3,…,N.
[0020] Where k represents the input vibration signal number, represents different input sequences, N represents the length of the input signal time series, and the length is N = n 2 The original data is converted into a two-dimensional data graph of n×n, and the calculation formula is as follows:
[0021] X=[x1,x2,…,x N ]
[0022]
[0023] X=x i , i=1,2,3,…,N is the original time series signal, X * is the two-dimensional data graph obtained after transformation. The corresponding relationship between the vector sequence signal and the matrix graph elements is N=n 2 .
[0024] Preferably, the step 1 uses a two-dimensional convolution layer to extract features and construct the model structure of each part. The two-dimensional convolution calculation formula is as follows:
[0025]
[0026] The above formula is a kernel matrix K with a shape of n×n and an input two-dimensional matrix M. Through the convolution kernel and the n in the input image 2 The corresponding products of the elements are summed up to calculate an output feature point O(i, j). The input image is moved horizontally and vertically, and different spans of element points are selected. The same kernel is used to analyze the data features of different image space positions. The data is fully analyzed to obtain the convolution feature results.
[0027] Preferably, in the multi-scale feature extraction in step 2, the data graph is input, convolution layers with different kernel sizes are first constructed, and features are extracted by stacking multiple layers of convolution to form various convolution kernel combinations to fully extract fault feature information in the data. The convolution process of the multi-scale feature extraction includes the first part, and the convolution kernel size of the multi-scale convolution structure and the parameters of the multi-parallel convolution structure are:
[0028] (1)1×1conv
[0029] (2) 3×3conv, 5×5conv
[0030] (3)7×7conv,11×11conv
[0031] (4) 3×3maxpool, 1×1conv
[0032] Among them, Conv represents a convolutional layer with multiple kernel numbers, and maxpool represents a maximum pooling layer;
[0033] In the second part, the multi-scale convolution structure uses a combination of smaller kernels to achieve the effect of a large convolution kernel. The parameters of the convolution kernel size and the multi-parallel convolution structure are:
[0034] (1) 1×3conv, 3×1conv, 1×3conv, 3×1conv
[0035] (2) 3×3avgpool, 1×1conv
[0036] (3) 1×3conv, 3×1conv
[0037] (4)1×1conv
[0038] Each part of the multi-scale convolution contains 4 branches. After the convolution calculation, the Relu activation function is used for nonlinear processing. The formula is as follows:
[0039] Preferably, each of the four branches of the multi-scale convolution can extract an output convolution feature set, and the output features y1, y2, y3, and y4 of the four branches are channel-concatenated to obtain multi-scale features, and the formula is as follows:
[0040] T = concept (y1, y2, y3, y4).
[0041] Wherein Y is a multi-scale feature extracted from the output of multiple branch convolution features, and the convolution processing branches of the two input signals adopt the same symmetrical structure to establish a dual-channel parallel processing framework.
[0042] Preferably, the independent and complementary fault information in the monitoring data of the multiple vibration signals in step 2 is based on the convolution features F1=[x1, x2, ..., xc] and F2=[y1, y2, ..., yc] output from a single vibration signal processing branch, where x i ,y i , i=1,...,c is a two-dimensional feature matrix. First, the associated spatial vibration feature is constructed. The formula is as follows:
[0043] R i =avgpool(x i *y i ),i=1,2,3,…,c.
[0044] Avgpool is an average pooling operation, which calculates the element sum average within the receptive field of the input image data corresponding to the pooling kernel, and takes the common distribution of multiple points in the neighborhood as the local smoothing feature. The formula is as follows:
[0045] R=Avgpool(F1*F2)
[0046] O = concat(F1, R, F2)
[0047] Among them, F1 and F2 are elements between channel maps, Avgpool is a multi-convolution channel feature processing operation, and avgpool is calculated for each channel of the input feature. R represents the similarity of independent features of different signals, and O is the channel fusion feature output.
[0048] Preferably, the cross entropy loss in step 4 is calculated as follows:
[0049]
[0050] Where y is the one-hot encoding of the true label of the sample, It is the sample failure probability vector of the predicted output under the current parameter combination of the model.
[0051] Preferably, the failure probability is predicted in step 4, and the formula is as follows:
[0052] P i =[p i1 ,p i2 ,p i3 ,…,p ic ] T ,i=1,2,3
[0053] P mod =(w1*P1+w2*P2+w3*P3)
[0054] Where P i is the probability vector of each diagnosis prediction classification output, 0 <p ij <1, c is the number of fault types, w i is the correction weight.
[0055] Preferably, the softmax is an output activation function for multi-category output mapping, and its formula is as follows:
[0056] o=XW T +b
[0057]
[0058] Where X is the output feature vector of the previous layer, W is the weight, b is the bias vector, o is the linear combination output value, n is the number of fault categories, represents the probability vector of different faults corresponding to the activation output, and
[0059] Preferably, the fault labels are set to 0, 1, 2, 3, 4, 5 according to the fault categories, and the vibration signal waveform samples and the corresponding fault labels are shuffled together, and the training set and test set data are divided into 4:1.
[0060] The present invention discloses a hydropower unit vibration information fusion fault diagnosis method based on deep learning. For the fault diagnosis of hydropower unit vibration information, multi-channel vibration signals are fused and deep learning technology is used to perform feature extraction and fault identification.
[0061] Data collection and preprocessing: First, the vibration signals at the guide bearing of the main shaft of the hydropower unit are collected and divided into sample segments with a length of 1024. Normalization processing and feature extraction are performed to prepare data for subsequent model training;
[0062] Deep diagnostic model construction:
[0063] Use 2D convolutional neural network (CNN) to extract features, integrate fault information in orthogonal multi-channel vibration signals, establish an end-to-end multi-input fault diagnosis model, perform multi-scale feature extraction and fusion, adopt a dual-channel parallel processing framework, perform channel-related feature extraction and fusion, build a deep residual learning structure, perform deep feature extraction, and use the softmax function to map high-level features into fault probability prediction distribution vectors;
[0064] Diagnosis based on multiple softmax outputs: Add a bypass classification structure, use intermediate features for fault identification, use a convolutional extraction structure for feature dimensionality reduction, and map features into fault probability vectors through a softmax function;
[0065] Diagnostic model training and optimization: According to the model structure, establish the cross entropy loss function, perform model training, adjust the loss weight, optimize the model parameters, and perform multiple iterative training through forward propagation and reverse gradient propagation;
[0066] Fault output probability joint decision diagnosis: Input the latest monitoring data, use the training model to extract features and identify faults, output three softmax probability vectors, predict the fault probability, and determine the fault category through joint probability decision. It has the following advantages:
[0067] (1) The multi-angle spatial data of the hydropower unit shaft vibration monitoring is fully utilized to expand the fault information capacity of the hydropower unit from the information source. Through deep feature extraction and fusion, the problem of diagnostic performance degradation caused by insufficient fault information based on a single input data is effectively avoided;
[0068] (2) For multiple input shaft vibration signals containing rich fault information of hydropower units, firstly, sufficient multi-type feature mining is performed to construct a basic feature set, and then channel-related fault information mining at the intermediate feature level is performed to extract fault information existing in the spatial vibration mode; this effectively improves the comprehensiveness of feature mining, and multi-step feature extraction can effectively suppress noise interference;
[0069] (3) Expand the output channels of the deep model to form an end-to-end diagnostic processing model with multiple inputs and multiple outputs. Multi-output training can provide regularization and reduce model overfitting. The joint output diagnostic decision method can stably and effectively improve the model diagnostic performance.
[0070] In general, this deep learning-based hydropower unit vibration information fusion fault diagnosis method, through advanced data processing and model building technology, eliminates the need for complex manual participation in the hydropower unit diagnosis process to the greatest extent, improves the automation and intelligence level of fault diagnosis, and provides a strong guarantee for the safe and stable operation of the hydropower unit. BRIEF DESCRIPTION OF THE DRAWINGS
[0071] Figure 1 It is a flow chart of fault diagnosis of hydropower unit of the present invention;
[0072] Figure 2 Visualization results of features extracted by the model of the present invention;
[0073] Figure 3 This is a network structure diagram of the model proposed in the present invention;
[0074] Figure 4 This is a structural diagram of the adaptive multi-scale feature extraction module proposed in the present invention;
[0075] Figure 5 This is the normalized channel attention structure diagram used in the present invention;
[0076] Figure 6 This is a structural diagram of the residual module used in the present invention;
[0077] Figure 7 A high-level feature representation diagram extracted by the model of the present invention;
[0078] Figure 8 It is the accuracy of multi-classifier output and decision fusion. DETAILED DESCRIPTION
[0079] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. All other embodiments obtained by ordinary technicians in the field without making creative work based on the embodiments of the present invention shall fall within the scope of protection of the present invention.
[0080] like Figures 1 to 8 As shown, it shows a specific embodiment of the present invention: Figures 1 to 8 As shown, the present invention discloses a hydropower unit vibration information fusion fault diagnosis method based on deep learning, the method comprising the following steps:
[0081] Step 1: Data collection and preprocessing;
[0082] Acquire monitoring data of vibration signals at the guide bearing of the main shaft of the hydropower unit as raw data and construct a data set, wherein the vibration signals include orthogonal multi-channel vibration signals of spatial vibration information;
[0083] The vibration signal is divided into several sample segments, the length of the sample segment is 1024, each sample segment contains several vibration cycles, each sample segment is normalized, and then features are extracted from the sample segments. The extracted features represent the characteristics of the fault signal, and the fault sample is determined according to the fault signal and a corresponding fault label is added;
[0084] Step 2: Deep diagnostic model construction
[0085] 2D convolution is used for feature extraction. By integrating independent and complementary fault information in the monitoring data of multiple vibration signals, an end-to-end multi-input fault diagnosis model is established. The multi-input fault diagnosis model simultaneously inputs the vibration signals obtained in orthogonal directions and performs multi-scale feature extraction and fusion. For each input channel, a length of n is first 2 The original input signal is converted into a two-dimensional data graph, and then several features in the original signal are obtained through a multi-scale feature extraction structure. The two input branches adopt the same symmetrical structure, and a dual-channel parallel processing framework is established to perform channel-related feature extraction and channel-level fusion. Then, based on the fusion features, a deep residual learning structure is constructed to perform deep feature extraction. The high-level features are activated by softmax through neuron combination to obtain the fault probability prediction distribution vector, and the diagnosis prediction result is obtained;
[0086] Step 3: Diagnosis based on multiple softmax outputs
[0087] Based on the model established in step 2, a bypass classification structure directly connected to the single-branch vibration feature is added, the intermediate-level features are used for fault identification, the convolution extraction structure is used to reduce the dimension of the features, and the softmax is used to map the high-level features into the probability vector of the corresponding fault;
[0088] Step 4: Diagnose model training and optimization process
[0089] The diagnostic model constructed according to steps 2-3 includes two vibration input branches and three softmax probability vector outputs. The model is trained according to the three softmax probability vector outputs, a cross entropy loss is established for each output probability vector, and model parameters are adjusted through forward propagation calculation and reverse gradient propagation training. The multi-scale feature extraction output loss weight in step 2 is 1, and the output loss weight of the two connected to the single vibration input feature in step 3 is 0.3. After multiple iterative trainings, a training model is obtained;
[0090] Step 5: Fault output probability joint decision diagnosis
[0091] According to the latest monitoring data of the vibration signal at the guide bearing of the main shaft of the hydropower unit, the training model is input. The forward calculation process of the training model is used to gradually extract features and identify faults, and three softmax probability vectors are output to predict the fault probability. The corrected fault probability is obtained by summing up the output probability vectors with three groups of dimensions as fault categories corresponding to the fault categories. The comprehensive fault identification results take the category with the largest joint probability as the output fault.
[0092] like Figure 1The fault diagnosis process of the hydropower unit shown in the figure first collects vibration monitoring data, obtains the original vibration time series signal data, constructs a data set, divides the vibration signal into multiple sample segments without overlap, and generally selects a data length of 1024. Each sample signal must contain several vibration cycles to fully express the fault information. Then, each sample is normalized to have a mean of 0 and a unit variance, so as to focus on the fault components in the signal. Monitoring data samples of 6 health states, including health and fault, are collected for processing, and corresponding fault labels are added to the fault samples. The labels are set to 0, 1, 2, 3, 4, and 5 according to the fault type. After the vibration signal waveform samples and the corresponding fault labels are shuffled together, they are divided into training set and test set data in a ratio of 4:1.
[0093] According to step 2, a deep diagnosis model is constructed, and 2D convolution is used as the main structure for feature extraction. The rich independent and complementary fault information in the multi-vibration monitoring signal is utilized. The model simultaneously inputs the vibration signals obtained in orthogonal directions, and multi-scale feature extraction and fusion are performed in parallel. For each input channel, the original input signal with a length of 1024 is first converted into a 32×32 two-dimensional data map, and then a large number of features in the original signal are obtained through the multi-scale feature extraction structure. The same symmetric structure is used for the two input branches, and a dual-channel parallel processing framework is established to perform channel-related feature extraction and channel-level fusion. Then, based on the fused features, a deep residual learning structure is constructed for deep feature extraction. The high-level features are activated by neurons through softmax combination to obtain the fault probability prediction distribution vector, and the diagnosis prediction result is obtained. According to step 3, on the basis of the established model, a bypass classification structure directly connected to the single-branch vibration feature is added, and the intermediate-level features are used for fault identification, providing additional diagnosis, and improving the degree of fault information mining and utilization in the signal. The convolution extraction structure with a relatively simple structure is used to reduce the dimension of the features, and the high-level features are mapped to the probability vector of the corresponding fault using softmax. The newly added convolution classification structure performs classification diagnosis based on the characteristics of a single input vibration signal, which can fully mine the fault differentiation information of each independent vibration signal and provide additional diagnostic output.
[0094] The bypass classification structure,the classifier is a feature dimension reduction classifier with a relatively simple structure,,and its main components, the network layer, are shown in Table 1:
[0095] Table 1:
[0096]
[0097] Train the model according to step 4. The diagnostic model constructed through the above process has two vibration input branches and 3 softmax diagnostic probability outputs. During training, a cross entropy loss is established for the probability vector of each output. The overall training performance measurement target of the model is set to the weighted sum of the cross entropy losses of the three outputs. Through forward propagation calculation, the Nadam algorithm is used for reverse gradient propagation training to adjust the model parameters. The channel fusion deep feature extraction output loss weight is set to 1, and the output loss weight of the two connected to the single vibration input feature is set to 0.3. The model learning rate is set to 0.001, the number of iterations is set to 50 times, and the training sample batch size is 32. After multiple iterative training processes, the model loss is continuously reduced and the diagnostic accuracy is improved.
[0098] After the model training is completed, fault diagnosis is applied. The application method is to use the latest monitored vibration signals of multiple main shafts of the hydropower unit, and input the measured vibration sample signals into the model according to the model structure. The forward calculation process of the model is used to perform step-by-step feature extraction and fault identification, and output the diagnosis probability and fault category. For the three-output diagnosis results obtained by the model diagnosis, the predicted fault probability obtained by the multi-output structure of the collection model is used for joint decision diagnosis according to step 5. The three groups of output probability vectors with the dimension of the number of fault categories are summed up corresponding to the fault category to obtain the corrected fault probability to represent the fault identification result of comprehensive multi-faceted information. Finally, the category with the largest joint probability is used as the output fault, so as to obtain a more accurate and effective fault result. The accuracy results of the comprehensive diagnosis are shown in Table 2, compared with the single-input diagnosis model:
[0099] Table 2:
[0100]
[0101] The comparison results with the multi-input diagnosis model are shown in Table 3:
[0102] Table 3:
[0103]
[0104] It can be seen that the proposed method has higher and more stable diagnostic performance than various models that extract features from a single signal. Compared with other deep models that also extract features from multiple signals, it is found that the proposed deep fusion diagnosis model also has obvious performance advantages. The high-level feature visualization results extracted by the model are shown in Figure 2. Figure 2 As shown, it can be seen that the original data is difficult to distinguish and the fault type cannot be identified. After model feature extraction, different faults can be accurately distinguished.
[0105] Preferably, in step 2, the original data is directly received as the model input, and a 2D data conversion relationship is constructed. The vibration signal of each input channel,
[0106] X k =x k (i),k=1,2,i=1,2,3,…,N.
[0107] Where k represents the input vibration signal number, represents different input sequences, N represents the length of the input signal time series, and the length is N = n 2 The original data is converted into a two-dimensional data graph of n×n, and the calculation formula is as follows:
[0108] X=[x1,x2,…,x N ]
[0109]
[0110] X=x i , i=1,2,3,…,N is the original time series signal, X * is the two-dimensional data graph obtained after transformation. The corresponding relationship between the vector sequence signal and the matrix graph elements is N=n 2 .
[0111] like Figure 3 As shown, in step 2, the original one-dimensional time series vibration signal is converted into a two-dimensional data graph. This conversion provides a richer information structure for the input of the deep learning model and has the following advantages:
[0112] 1. Enhanced feature representation capabilities: By converting a one-dimensional signal into a two-dimensional graph, the time-frequency characteristics of the signal can be better captured. The two-dimensional data graph can show the vibration pattern of the signal in different time periods, providing a richer feature space for the deep learning model.
[0113] 2. Improve the generalization ability of the model: Two-dimensional data graphs can contain more contextual information, which helps the model understand the characteristics of the signal from different perspectives, thereby improving the model's generalization ability for unseen data.
[0114] 3. Utilize spatial correlation: In a two-dimensional data graph, there may be spatial correlation between adjacent time points. Deep learning models can use this spatial correlation to extract more meaningful features and improve the accuracy of fault diagnosis.
[0115] 4. Simplify the model structure: By directly converting the original signal into a two-dimensional data graph, the input structure of the model can be simplified, the preprocessing steps can be reduced, and the model can learn features more directly from the original data.
[0116] 5. Adaptability: This conversion method can be applied to different types of vibration signals and has good adaptability. Whether it is a periodic signal or a non-periodic signal, it can be effectively processed by this conversion method.
[0117] 6. Reduce computational complexity: Compared with directly extracting features from one-dimensional signals, two-dimensional data graphs may allow the use of simpler convolution kernels for feature extraction, thereby reducing the computational complexity of the model to a certain extent.
[0118] 7. Improved diagnostic speed: Since the model can learn features directly from the two-dimensional data graph, it may reduce the need for feature engineering, thereby speeding up fault diagnosis.
[0119] 8. Easy to visualize: Two-dimensional data graphs are easier to visualize than one-dimensional signals, which helps to understand the model.
[0120] 9. Multi-scale analysis: Based on the two-dimensional data graph, multi-scale analysis can be more easily implemented, such as capturing features of different scales through convolution kernels of different sizes, which helps the model better understand the local and global characteristics of the signal.
[0121] In summary, converting the original vibration signal into a two-dimensional data graph and using it as the input of the deep learning model can make full use of the advantages of deep learning in feature extraction and pattern recognition to improve the accuracy and efficiency of fault diagnosis of hydropower units.
[0122] Preferably, the step 1 uses a two-dimensional convolution layer to extract features and construct the model structure of each part. The two-dimensional convolution calculation formula is as follows:
[0123]
[0124] The above formula is a kernel matrix K with a shape of n×n and an input two-dimensional matrix M. Through the convolution kernel and the n in the input image 2 The corresponding products of the elements are summed up to calculate an output feature point O(i, j). The input image is moved horizontally and vertically, and different spans of element points are selected. The same kernel is used to analyze the data features of different image space positions. The data is fully analyzed to obtain the convolution feature results.
[0125] Two-dimensional convolution extracts features through convolution kernels. Common kernel sizes are 1×1, 3×3, 5×5, etc. The two-dimensional convolution formula uses multiple convolution kernels of scale to calculate features at the same time to increase the richness of various features and output feature results corresponding to different kernel parameters. The two-dimensional convolution kernel is only connected to the local area of the input data, which enables the model to capture local features, such as local frequency changes in vibration signals. Compared with the fully connected layer, the convolution layer greatly reduces the amount of calculation and the required storage space due to its local connection characteristics, which helps to improve the accuracy and efficiency of fault diagnosis.
[0126] Preferably, in the multi-scale feature extraction in step 2, the data graph is input, convolution layers with different kernel sizes are first constructed, and features are extracted by stacking multiple layers of convolution to form various convolution kernel combinations to fully extract fault feature information in the data. The convolution process of the multi-scale feature extraction includes the first part, and the convolution kernel size of the multi-scale convolution structure and the parameters of the multi-parallel convolution structure are:
[0127] (1)1×1conv
[0128] (2) 3×3conv, 5×5conv
[0129] (3)7×7conv,11×11conv
[0130] (4) 3×3maxpool, 1×1conv
[0131] Among them, Conv represents a convolutional layer with multiple kernel numbers, and maxpool represents a maximum pooling layer;
[0132] In the second part, the multi-scale convolution structure uses a combination of smaller kernels to achieve the effect of a large convolution kernel. The parameters of the convolution kernel size and the multi-parallel convolution structure are:
[0133] (1) 1×3conv, 3×1conv, 1×3conv, 3×1conv
[0134] (2) 3×3avgpool, 1×1conv
[0135] (3) 1×3conv, 3×1conv
[0136] (4)1×1conv
[0137] Each part of the multi-scale convolution contains 4 branches. After the convolution calculation, the Relu activation function is used for nonlinear processing. The formula is as follows:
[0138] The vibration signals in orthogonal directions are input at the same time, and then a large number of features in the original signal are obtained through the multi-scale feature extraction structure. The multi-scale feature extraction structure is constructed based on two-dimensional convolution. The method is to first construct convolution layers with different kernel sizes for the input data graph, and use multi-layer convolution stacking to extract features with various convolution kernel combinations to fully extract the fault feature information in the data.
[0139] Preferably, each of the four branches of the multi-scale convolution can extract an output convolution feature set, and the output features y1, y2, y3, and y4 of the four branches are channel-concatenated to obtain multi-scale features, and the formula is as follows:
[0140] Y=concept(y1,y2,y3,y4).
[0141] Wherein Y is a multi-scale feature extracted from the output of multiple branch convolution features, and the convolution processing branches of the two input signals adopt the same symmetrical structure to establish a dual-channel parallel processing framework.
[0142] The four branches in the multi-scale convolution structure extract features respectively, and then these features are cascaded at the channel level to form a comprehensive multi-scale feature set; the dual-channel parallel processing framework allows two input signals to be processed simultaneously, which can significantly improve computational efficiency, especially in real-time or near real-time fault diagnosis systems. The convolution processing branches of the two input signals adopt the same symmetric structure, which helps to ensure that the model processes the two input signals consistently, thereby improving the fairness and generalization ability of the model.
[0143] Preferably, the independent and complementary fault information in the monitoring data of the multiple vibration signals in step 2 is based on the convolution features F1=[x1, x2, ..., xc] and F2=[y1, y2, ..., yc] output from a single vibration signal processing branch, where x i ,y i , i=1,...,c is a two-dimensional feature matrix. First, the associated spatial vibration feature is constructed. The formula is as follows:
[0144] R i =avgpool(x i *y i ),i=1,2,3,…,c.
[0145] Avgpool is an average pooling operation, which calculates the element sum average within the receptive field of the input image data corresponding to the pooling kernel, and takes the common distribution of multiple points in the neighborhood as the local smoothing feature. The formula is as follows:
[0146] R=Avgpool(F1*F2)
[0147] O = concat(F1, R, F2)
[0148] Among them, F1 and F2 are elements between channel maps, Avgpool is a multi-convolution channel feature processing operation, and avgpool is calculated for each channel of the input feature. R represents the similarity of independent features of different signals, and O is the channel fusion feature output.
[0149] In step 2, the process of constructing the associated spatial vibration features based on the convolution features output by a single vibration signal processing branch involves the integration and fusion of features, which improves the model's ability to capture fault information and provides an effective technical means for fault diagnosis of hydropower units.
[0150] Preferably, the cross entropy loss in step 4 is calculated as follows:
[0151]
[0152] Where y is the one-hot encoding of the true label of the sample, It is the sample failure probability vector of the predicted output under the current parameter combination of the model.
[0153] The cross entropy loss measures the difference between the probability distribution predicted by the model and the probability distribution of the true label. The smaller the value, the more accurate the model's prediction. By minimizing the cross entropy loss, the model learns to adjust its parameters to make the predicted probability as close as possible to the probability of the true label. The cross entropy loss is used in conjunction with the softmax activation function. The softmax function can convert the output of the model into a probability distribution, which is then directly used to calculate the cross entropy loss. The cross entropy loss can be used to train the model so that it can accurately predict the probabilities of different fault types, thereby achieving efficient and accurate fault identification.
[0154] Preferably, the failure probability is predicted in step 4, and the formula is as follows:
[0155] P i =[p i1 ,p i2 ,p i3 ,…,p ic ] T ,i=1,2,3
[0156] P mod =(w1*P1+w2*P2+w3*P3)
[0157] Where P i is the probability vector of each diagnosis prediction classification output, 0 <p ij <1, c is the number of fault types, w i is the correction weight.
[0158] By introducing the modified weight w i , the predicted probabilities of different fault categories can be adjusted to reflect their importance or frequency of occurrence in actual situations, helping the model to more accurately predict the probability of fault occurrence, thereby achieving more effective fault prevention and maintenance strategies.
[0159] Preferably, the softmax is an output activation function for multi-category output mapping, and its formula is as follows:
[0160] o=XW T +b
[0161]
[0162] Where X is the output feature vector of the previous layer, W is the weight, b is the bias vector, o is the linear combination output value, n is the number of fault categories, represents the probability vector of different faults corresponding to the activation output, and
[0163] The output of the softmax function can be directly interpreted as probability, which makes the result easy to understand and explain. Each element of the output vector is between 0 and 1, and the sum of all elements is 1, forming a valid probability distribution. The softmax function can output a probability for each category, which helps the model make more robust predictions in the face of uncertainty. Using the softmax function can help the model accurately predict the probability of faults belonging to different categories, thereby providing strong support for fault diagnosis.
[0164] Preferably, the fault labels are set to 0, 1, 2, 3, 4, 5 according to the fault categories, and the vibration signal waveform samples and the corresponding fault labels are shuffled together, and the training set and test set data are divided into 4:1.
[0165] Fault label setting: First, according to different types of faults, a unique label is set for each category, such as 0, 1, 2, 3, 4, 5. Each vibration signal waveform sample is assigned a label according to its corresponding fault type.
[0166] Data shuffling: Before dividing the training set and the test set, the order of the set of vibration signal waveform samples and their corresponding fault labels is shuffled. This step is to ensure the randomness of the data and avoid deviations in the model training process.
[0167] Dataset division: The shuffled data set is divided into training set and test set according to a certain ratio. The ratio is 4:1, about 80% of the data will be used for training, and 20% of the data will be used for testing. In the fault diagnosis of vibration signals of hydropower units, this method can help develop and evaluate a model that can accurately identify different fault types. The model first learns the relationship between vibration signals and fault types on the training set, and then verifies its diagnostic ability on the test set.
[0168] In summary, dividing the dataset into training and test sets and using these sets during model development are key steps to ensure model effectiveness and reliability.
[0169] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in the field without departing from the purpose of the present invention. These changes involve related technologies well known to those skilled in the art, which all fall within the scope of protection of the patent of the present invention.
[0170] Many other changes and modifications may be made without departing from the concept and scope of the present invention.It should be understood that the present invention is not limited to the specific embodiments, and the scope of the present invention is defined by the appended claims.
Claims
1. A hydropower unit vibration information fusion fault diagnosis method based on deep learning, It is characterized in that The method comprises the following steps: Step 1: Data collection and preprocessing; Acquire monitoring data of vibration signals at the guide bearing of the main shaft of the hydropower unit as raw data and construct a data set, wherein the vibration signals include orthogonal multi-channel vibration signals of spatial vibration information; The vibration signal is divided into several sample segments, the length of the sample segment is 1024, each sample segment contains several vibration cycles, each sample segment is normalized, and then features are extracted from the sample segments. The extracted features represent the characteristics of the fault signal, and the fault sample is determined according to the fault signal and a corresponding fault label is added; Step 2: Construction of deep diagnostic model; 2D convolution is used for feature extraction. By integrating independent and complementary fault information in the monitoring data of multiple vibration signals, an end-to-end multi-input fault diagnosis model is established. The multi-input fault diagnosis model simultaneously inputs the vibration signals obtained in orthogonal directions and performs multi-scale feature extraction and fusion. For each input channel, a length of n is first 2 The original input signal is converted into a two-dimensional data graph, and then several features in the original signal are obtained through a multi-scale feature extraction structure. The two input branches adopt the same symmetrical structure, and a dual-channel parallel processing framework is established to perform channel-related feature extraction and channel-level fusion. Then, based on the fusion features, a deep residual learning structure is constructed to perform deep feature extraction. The high-level features are activated by softmax through neuron combination to obtain the fault probability prediction distribution vector, and the diagnosis prediction result is obtained; Step 3: Diagnosis based on multiple softmax outputs; Based on the model established in step 2, a bypass classification structure directly connected to the single-branch vibration feature is added, the intermediate-level features are used for fault identification, the convolution extraction structure is used to reduce the dimension of the features, and the softmax is used to map the high-level features into the probability vector of the corresponding fault; Step 4: Diagnostic model training and optimization process; The diagnostic model constructed according to steps 2-3 includes two vibration input branches and three softmax probability vector outputs. The model is trained according to the three softmax probability vector outputs, a cross entropy loss is established for each output probability vector, and model parameters are adjusted through forward propagation calculation and reverse gradient propagation training. The multi-scale feature extraction output loss weight in step 2 is 1, and the output loss weight of the two connected to the single vibration input feature in step 3 is 0.
3. After multiple iterative trainings, a training model is obtained; Step 5: Fault output probability joint decision diagnosis; According to the latest monitoring data of the vibration signal at the guide bearing of the main shaft of the hydropower unit, the training model is input. The forward calculation process of the training model is used to gradually extract features and identify faults, and three softmax probability vectors are output to predict the fault probability. The corrected fault probability is obtained by summing up the output probability vectors with three groups of dimensions as fault categories corresponding to the fault categories. The comprehensive fault identification results take the category with the largest joint probability as the output fault.
2. According to the method for fusion fault diagnosis of hydropower unit vibration information based on deep learning in claim 1, It is characterized in that In step 2, the original data is directly received as the model input, and a 2D data conversion relationship is constructed. The vibration signal of each input channel, X k =x k (t),k=1,2,i=1,2,3,...,N. Where k represents the input vibration signal number, represents different input sequences, N represents the length of the input signal time series, and the length is N = n 2 The original data is converted into a two-dimensional data graph of n×n, and the calculation formula is as follows: X=[x 1 ,x 2 ,…,x N ] X=x i , i=1,2,3,…,N is the original time series signal, X * is the two-dimensional data graph obtained after transformation. The corresponding relationship between the vector sequence signal and the matrix graph elements is 3. According to the method for hydropower unit vibration information fusion fault diagnosis based on deep learning in claim 2, It is characterized in that The step 1 uses a two-dimensional convolution layer to extract features and construct the model structure of each part. The two-dimensional convolution calculation formula is as follows: The above formula is an n×n kernel matrix K and an input two-dimensional matrix M. Through the convolution kernel and the n in the input image 2 The corresponding products of the elements are summed up to calculate an output feature point O(i, j). The input image is moved horizontally and vertically, and different spans of element points are selected. The same kernel is used to analyze the data features of different image space positions. The data is fully analyzed to obtain the convolution feature results.
4. A hydropower unit vibration information fusion fault diagnosis method based on deep learning according to any one of claims 1 to 3, It is characterized in that In the step 2, the multi-scale feature extraction inputs the data graph, first constructs convolution layers with different kernel sizes, extracts features by stacking multiple layers of convolution, and forms various convolution kernel combinations to fully extract fault feature information in the data. The convolution process of the multi-scale feature extraction includes the first part, and the convolution kernel size of the multi-scale convolution structure and the parameters of the multi-parallel convolution structure are: (1)1×1conv (2) 3×3conv, 5×5conv (3) 7×7conv, 11×11conv (4) 3×3maxpool, 1×1conv Among them, Conv represents a convolutional layer with multiple kernel numbers, and maxpool represents a maximum pooling layer; In the second part, the multi-scale convolution structure uses a combination of smaller kernels to achieve the effect of a large convolution kernel. The parameters of the convolution kernel size and the multi-parallel convolution structure are: (1) 1×3conv, 3×1conv, 1×3conv, 3×1conv (2) 3×3avgpool, 1×1conv (3) 1×3conv, 3×1conv (4)1×1conv Each part of the multi-scale convolution contains 4 branches. After the convolution calculation, the Relu activation function is used for nonlinear processing. The formula is as follows:
5. According to the method for hydropower unit vibration information fusion fault diagnosis based on deep learning in claim 4, It is characterized in that Each of the four branches of the multi-scale convolution can extract an output convolution feature set, and the output features y of the four branches are 1 ,y 2 ,y 3 ,y 4 Channel cascading is performed to obtain multi-scale features, and the formula is as follows: Y=concept(y1,y2,y3,y4). Wherein Y is a multi-scale feature extracted from the output of multiple branch convolution features, and the convolution processing branches of the two input signals adopt the same symmetrical structure to establish a dual-channel parallel processing framework.
6. According to the method of vibration information fusion fault diagnosis of hydropower unit based on deep learning in claim 5, It is characterized in that The independent and complementary fault information in the monitoring data of the multiple vibration signals in step 2 is based on the convolution features F1=[x1, x2, ..., xc] and F2=[y1, y2, ..., yc] output from the single vibration signal processing branch, where x i ,y i , i=1,..., c is a two-dimensional feature matrix. First, the associated spatial vibration feature is constructed, and the formula is as follows: R i =avgpool(x i *y i ),i=1,2,3,...,c. Avgpool is an average pooling operation, which calculates the average value of the elements in the input image data receptive field corresponding to the pooling kernel, and takes the common distribution of multiple points in the neighborhood as the local smoothing feature. The formula is as follows: R=Avgpool(F 1 *F 2 ) O=concat(F 1 ,R,F 2 ) Among them, F 1 、F 2 It is an element between channel graphs. Avgpool is a feature processing operation of multiple convolution channels. Avgpool is calculated for each channel of the input feature. R represents the similarity of independent features of different signals. O is the channel fusion feature output.
7. A hydropower unit vibration information fusion fault diagnosis method based on deep learning according to claim 6, It is characterized in that The cross entropy loss in step 4 is as follows: Where y is the one-hot encoding of the true label of the sample, It is the sample failure probability vector of the predicted output under the current parameter combination of the model.
8. According to the method for hydropower unit vibration information fusion fault diagnosis based on deep learning in claim 1, It is characterized in that The failure probability is predicted in step 4, and the formula is as follows: P i =[p i1 ,p i2 ,p i3 ,...,p ic ] T ,i=1,2,3 P mod =(w 1 *P 1 +w 2 *P 2 +w 3 *P 3 ) Where P i is the probability vector of each diagnosis prediction classification output, 0 <p ij <1, c is the number of fault types, w i is the correction weight.
9. According to the method for hydropower unit vibration information fusion fault diagnosis based on deep learning in claim 1, It is characterized in that The softmax is an output activation function for multi-category output mapping, and its formula is as follows: o=XW T +b Where X is the output feature vector of the previous layer, W is the weight, b is the bias vector, o is the linear combination output value, n is the number of fault categories, represents the probability vector of different faults corresponding to the activation output, and 10. According to the method of vibration information fusion fault diagnosis of hydropower unit based on deep learning in claim 1, It is characterized in that The fault labels are set to 0, 1, 2, 3, 4, and 5 according to the fault categories. After the vibration signal waveform samples and the corresponding fault labels are shuffled, the training set and test set data are divided into 4:1.
Citation Information
Cited By
Off-road forklift transmission system fault diagnosis method and system based on artificial intelligence
CN119179999A
Diesel generating set fault prediction method, model and storage medium
CN119538102A
Special vehicle key component health assessment method based on orthogonal characteristic dynamic coupling
CN119862394A
High-reliability fault diagnosis method based on physical-data joint optimization driving and related device
CN120046003A
Bearing fault variable working condition diagnosis method based on multi-scale convolutional network and MAML
CN120822090A
