A method and system for identifying the operating state of a hydroelectric generating unit based on the gram angle field

Abstract

The present application relates to the field of hydroelectric generating set state monitoring, in particular to a hydroelectric generating set operation state recognition method and system based on gram angle field; state parameters under different operation states of the hydroelectric generating set are acquired; the state parameters under different operation states are preprocessed to obtain one-dimensional fusion time domain signals; the one-dimensional fusion time domain signals are image coded based on GAF to obtain GADF images and GASF images; the GADF images and the GASF images under the same operation state are input into a pre-constructed DCLCNN model to obtain the operation state recognition result of the hydroelectric generating set. The present application realizes effectively improving the feature extraction capability of the DCLCNN model in the frequent adjustment process of the hydroelectric generating set and under the complex operation conditions of the generating set, improving the state recognition precision of the DCLCNN model, and reducing the calculation complexity of the DCLCNN model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of hydroelectric unit state monitoring, and particularly relates to a hydroelectric unit operation state recognition method and system based on Gramian angular field. BACKGROUND

[0002] With the promotion of clean energy structure transformation, hydroelectric power generation has become the first choice for maintaining power grid balance in high proportion renewable energy grid connection scenarios due to its flexible and rapid load regulation capability. However, frequent peak shaving and frequency modulation tasks make the hydroelectric unit long-term operate in non-steady state, which causes the deterioration of the hydroelectric unit to intensify and the fault risk to rise significantly. Therefore, the research on hydroelectric unit state monitoring and fault diagnosis has important engineering value.

[0003] The prior art generally analyzes the vibration, temperature, acoustics and other multi-modal signals of the equipment, extracts the key features in the time domain, frequency domain and time-frequency domain, and realizes the state recognition of the hydroelectric unit by combining machine learning algorithms such as random forest or support vector machine. Although the above method can recognize typical fault modes, it performs poorly when facing non-linear and non-stationary signals under complex working conditions. Although time-frequency analysis can reveal the dynamic characteristics of the signal, the analysis process is complicated and needs to rely on professional knowledge to select appropriate window functions or scale parameters and the like, which is difficult to adapt to the complex and changeable operation conditions and fault modes of the hydroelectric unit, resulting in insufficient state recognition accuracy.

[0004] Gramian Angular Field (GAF) as a new time series signal processing method has been widely applied in the state monitoring and fault diagnosis of various large and complex mechanical and electrical equipment, and therefore it has great development potential in the field of hydroelectric units. GAF can map the vibration signal to the polar coordinate system and further generate Gramian Angular Difference Field (GADF) image and Gramian Angular Summation Field (GASF) image using Gramian matrix, which can retain the time sequence dependence and global features of the signal. This method is simple and fast to calculate, does not need complex parameter adjustment, has wide adaptability, and has the ability to represent the operation state and fault state of complex mechanical equipment.

[0005] In order to realize the state recognition of hydroelectric generating set by GADF or GASF image, various image recognition models need to be introduced. Among them, the Convolutional Neural Network (CNN) model has certain advantages in such tasks. The CNN can automatically extract deep signal features through multi-layer convolution and pooling operations, greatly reduces the dependence on artificial feature engineering, and shows strong adaptability to complex environment. However, the conventional CNN model is difficult to effectively capture the signal dynamic characteristics under complex working conditions of equipment, and the model recognition accuracy is poor. At the same time, the high parameter quantity and calculation cost of CNN make it difficult to adapt to the industrial environment with limited resources, low energy consumption and high real-time requirements. Most importantly, the CNN usually only targets a single image, so it can only recognize GADF or GASF image independently, cannot fully utilize the information of the two kinds of images, and is difficult to fully represent the signal characteristics of hydroelectric generating set under complex working conditions. SUMMARY

[0006] In view of the problems in the prior art, the present application provides a hydroelectric generating set operating state recognition method and system based on Gram angle field, which realizes the joint recognition application of GADF and GASF images by constructing a new DCLCNN model, effectively improves the feature extraction capability of the DCLCNN model in the process of frequent adjustment of the hydroelectric generating set and under complex operating conditions of the generating set, and realizes the deep fusion application of GADF and GASF images in the field of hydroelectric generating set monitoring and diagnosis.

[0007] In order to achieve the above purpose, the present application adopts the following technical scheme:

[0008] The hydroelectric generating set operating state recognition method based on Gram angle field comprises the following steps:

[0009] S1, obtaining state parameters under different operating states of the hydroelectric generating set;

[0010] S2, pre-processing the state parameters under different operating states to obtain a one-dimensional fusion time domain signal;

[0011] S3, image encoding the one-dimensional fusion time domain signal based on GAF to obtain GADF image and GASF image;

[0012] S4, input the GADF image and the GASF image under the same operating state into a pre-constructed DCLCNN model, the DCLCNN model comprising two parallel feature extraction branches, a feature fusion layer and a classification layer; one of the feature extraction branches extracts features of the GADF image to obtain first multi-scale features, and the other feature extraction branch extracts features of the GASF image to obtain second multi-scale features; the feature fusion layer fuses the first multi-scale features and the second multi-scale features to obtain fused features; and the classification layer maps and outputs the fused features to obtain an operating state recognition result of the hydroelectric generating unit.

[0013] As a further improvement of the application, the state parameters in S1 include a swing signal and a vibration signal.

[0014] The swing signal includes X-direction swing of the upper guide bearing, Y-direction swing of the upper guide bearing, X-direction swing of the lower guide bearing, Y-direction swing of the lower guide bearing, X-direction swing of the water guide bearing and Y-direction swing of the water guide bearing.

[0015] The vibration signal includes X-direction vibration of the upper bracket, Y-direction vibration of the upper bracket, X-direction vibration of the lower bracket, Y-direction vibration of the lower bracket, X-direction vibration of the generator stator and Y-direction vibration of the generator stator.

[0016] The different operating states include a stable operating condition, a transition operating condition and an abnormal operating condition.

[0017] The stable operating condition is divided into the following conditions according to the percentage of the current operating load to the rated load:

[0018] The condition of 0% to below 20% of the current operating load is divided into a low-load operating condition.

[0019] The condition of 20% and above to below 40% of the current operating load is divided into a low-to-medium load operating condition.

[0020] The condition of 40% and above to below 60% of the current operating load is divided into a medium load operating condition.

[0021] The condition of 60% and above to below 80% of the current operating load is divided into a medium-to-high load operating condition.

[0022] The condition of 80% and above to 100% of the current operating load is divided into a high load operating condition.

[0023] The transition operating condition includes a start-up process, a shutdown process, a speed increasing / decreasing process and a load increasing / decreasing process.

[0024] The abnormal operating condition includes hydraulic vibration abnormality, mechanical vibration abnormality and electromagnetic vibration abnormality.

[0025] As a further improvement of the application, the preprocessing in S2 includes signal fusion of state parameters in different operating states by PCA to obtain a one-dimensional fused time domain signal.

[0026] As a further improvement of the application, the process of S3 includes:

[0027] The one-dimensional fused time domain signal is smoothed and compressed to obtain a compressed signal.

[0028] The compressed signal is normalized to obtain a normalized signal.

[0029] The normalized signal is processed by polar coordinate conversion to obtain polar coordinates, including polar angle and polar radius.

[0030] The polar coordinates are processed by Gram angle and difference field coding to obtain GADF images by calculating the sine value of the angle difference between data points in the polar coordinates, and GASF images by calculating the cosine value of the angle sum between data points in the polar coordinates.

[0031] As a further improvement of the application, the two feature extraction branch structures in S4 are the same, each including an initial feature extraction layer, a deep feature extraction layer, and a feature recombination layer.

[0032] The initial feature extraction layer performs 3x3 standard convolution on the input GADF image or GASF image, uses a ReLU activation function, and has a step size of 2 to obtain preliminary features.

[0033] The deep feature extraction layer includes four layers of inverted residual structures, which contain MLCA attention mechanisms in the inverted residual structures.

[0034] The inverted residual structure sequentially performs 1x1 point convolution and 3x3 deep convolution on the preliminary features to obtain deep convolution output.

[0035] The MLCA attention mechanism is used to calibrate the deep convolution output to obtain calibrated features.

[0036] The calibrated features are subjected to 1x1 point convolution to obtain deep features.

[0037] The feature recombination layer performs 1x1 standard convolution on the deep features, uses a h-swish activation function, and has a step size of 1 to obtain first multi-scale features or second multi-scale features.

[0038] As a further improvement of the application, the feature fusion layer in S4 sequentially performs global average pooling, dimension merging, and flattening on the first multi-scale features and the second multi-scale features to obtain fused features.

[0039] As a further improvement of the application, the classification layer in S4 comprises two fully connected layers, the first fully connected layer is used to upgrade the fusion features, and an h-swish activation function and a Dropout layer are introduced to obtain high-dimensional features;

[0040] The second fully connected layer is used to reduce the dimension of the high-dimensional features and map the classification, and the operation state recognition result of the hydroelectric generating set is output.

[0041] The application provides a hydroelectric generating set operation state recognition system based on a Gram angle field, which comprises:

[0042] A data acquisition module is configured to acquire state parameters under different operation states of the hydroelectric generating set.

[0043] A data preprocessing module is configured to preprocess the state parameters under different operation states to obtain one-dimensional fusion time domain signals.

[0044] An image encoding module is configured to encode the one-dimensional fusion time domain signals based on GAF to obtain GADF images and GASF images.

[0045] A result output module is configured to input the GADF images and the GASF images under the same operation state into a pre-constructed DCLCNN model, wherein the DCLCNN model comprises two parallel feature extraction branches, a feature fusion layer and a classification layer; one of the feature extraction branches is configured to extract features of the GADF images to obtain first multi-scale features, and the other feature extraction branch is configured to extract features of the GASF images to obtain second multi-scale features; the feature fusion layer is configured to fuse the first multi-scale features and the second multi-scale features to obtain fusion features; and the classification layer is configured to map and output the fusion features to obtain the operation state recognition result of the hydroelectric generating set.

[0046] The application provides a hydroelectric generating set operation state recognition device based on a Gram angle field, which comprises a processor and a memory, wherein the processor executes a computer program stored in the memory to realize the hydroelectric generating set operation state recognition method based on the Gram angle field.

[0047] The application provides a computer readable storage medium for storing a computer program, wherein the computer program is executed by a processor to realize the hydroelectric generating set operation state recognition method based on the Gram angle field.

[0048] Compared with the prior art, the application has the following technical effects:

[0049] The GAF is used for image coding of the vibration signal and the swing signal of the hydroelectric generating set, time series are converted into two-dimensional images retaining time dependence and global features, the GAF coding does not need a window function, parameter setting is simple, operation efficiency is high, information loss and high calculation burden of time-frequency conversion are avoided, and high-quality image input is provided for the subsequent DCLCNN model.

[0050] The GADF and GASF images generated by the GAF are simultaneously used, complementary advantages of the two in symmetry and dynamic feature extraction are fully given, limitations of insufficient feature expression of single image coding are overcome, a DCLCNN model is built through the application, feature extraction and recognition of double images are realized, information integrity is significantly enhanced by double image input, recognition accuracy of the model under complex operating conditions is greatly improved, and the reliability of state monitoring is effectively improved.

[0051] The inverted residual structure is combined in the DCLCNN model, the parameter quantity and the calculation complexity of the DCLCNN model are significantly reduced, high discrimination ability and low power consumption demand can be considered, real-time response and resource-limited deployment requirements of an industrial field are met, the MLCA attention mechanism is introduced into the DCLCNN model, the multi-scale feature interaction strategy is combined, the perception ability of the DCLCNN model to the slight difference between complex operating states of the hydroelectric generating set is improved, and the recognition accuracy and the generalization performance of the DCLCNN model are significantly enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 It is a whole flowchart of the application;

[0053] Figure 2 It is a GAF image coding process diagram of the vibration signal and the swing signal of the hydroelectric generating set;

[0054] Figure 3 It is a DCLCNN model diagram of the application;

[0055] Figure 4 It is a MLCA attention mechanism diagram in the DCLCNN model of the application;

[0056] Figure 5 It is an inverted residual structure diagram in the DCLCNN model of the application;

[0057] Figure 6 It is a state recognition accuracy comparison diagram of different double-channel models in the embodiment;

[0058] Figure 7 It is a system diagram of the application. DETAILED DESCRIPTION

[0059] The application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the application, but not to limit the application. In addition, it should be noted that only the parts related to the application are shown in the drawings for the convenience of description, and not all the structures. Referring to Figure 1 The application provides a hydroelectric generator set operating state recognition method based on a Gram angle field, comprising the following steps:

[0060] S1, acquiring state parameters of the hydroelectric generator set under different operating states;

[0061] S2, preprocessing the state parameters under different operating states to obtain a one-dimensional fusion time domain signal;

[0062] S3, image coding the one-dimensional fusion time domain signal based on GAF to obtain a GADF image and a GASF image;

[0063] S4, inputting the GADF image and the GASF image under the same operating state into a pre-constructed DCLCNN (Dual-Channel Lightweight Convolution Neutral Network) model, the DCLCNN model comprising two parallel feature extraction branches, a feature fusion layer, and a classification layer; one of the feature extraction branches extracts features of the GADF image to obtain first multi-scale features, and the other feature extraction branch extracts features of the GASF image to obtain second multi-scale features; the feature fusion layer fuses the first multi-scale features and the second multi-scale features to obtain fusion features; and the classification layer maps and outputs the fusion features to obtain the operating state recognition result of the hydroelectric generator set.

[0064] The application will be further described in detail below with reference to the accompanying drawings and embodiments:

[0065] Step one, the acceleration sensor is used to measure the swing signal of the large shaft of the hydroelectric generator set and the vibration signal of the frame in this embodiment, and the sampling frequency is 1024 Hz; the swing signal includes the X-direction swing of the upper guide bearing, the Y-direction swing of the upper guide bearing, the X-direction swing of the lower guide bearing, the Y-direction swing of the lower guide bearing, the X-direction swing of the water guide bearing, and the Y-direction swing of the water guide bearing, including 6-channel swing signals.

[0066] The vibration signal includes the X-direction vibration of the upper frame, the Y-direction vibration of the upper frame, the X-direction vibration of the lower frame, the Y-direction vibration of the lower frame, the X-direction vibration of the generator stator, and the Y-direction vibration of the generator stator, including 6-channel vibration signals, a total of 12-channel signals as data, but not limited to this. In actual application, more vibration signals and swing signals can be selected according to the conditions of different hydroelectric generator sets.

[0067] In the embodiment, the swing signals and the vibration signals in different operating states need to be selected, including stable operating conditions, transition operating conditions and abnormal operating conditions.

[0068] Specifically, the stable operating conditions are divided according to the percentage of the current operating load to the rated load: 0% or more to less than 20% of the current operating load is divided into a low load operating condition; 20% or more to less than 40% of the current operating load is divided into a medium-low load operating condition; 40% or more to less than 60% of the current operating load is divided into a medium load operating condition; 60% or more to less than 80% of the current operating load is divided into a medium-high load operating condition; and 80% or more to 100% of the current operating load is divided into a high load operating condition.

[0069] The transition operating conditions include a starting process, a shutdown process, a speed increasing and decreasing process and a load increasing and decreasing process.

[0070] The abnormal operating conditions include hydraulic vibration abnormalities caused by factors such as hydraulic imbalance, tail water pipe eccentric vortex band and Karman vortex train; mechanical vibration abnormalities caused by factors such as shaft misalignment, rotor mass imbalance and dynamic and static part rubbing; and electromagnetic vibration abnormalities caused by factors such as magnetic pull imbalance, generator stator core loosening and generator stator-rotor air gap asymmetry.

[0071] In the embodiment, the hydroelectric generating set is divided into six operating states of high load operating condition, medium load operating condition, low load operating condition, speed increasing and decreasing process, shutdown process and starting process by combining historical operating data of the hydroelectric generating set, and each operating state includes vibration signals and swing signals of the above-mentioned 12 channels.

[0072] In actual operation, the duration of different states of the hydroelectric generating set is different, and the data of each state is of different lengths, as shown in Table 1, and there is an obvious sample imbalance problem.

[0073] Table 1: Experimental data for state recognition of hydroelectric generating set

[0074]

[0075] Step three, the 12 channels of vibration signals and swing signals in different operating states are fused into one-dimensional fusion time domain signals by PCA (Principal Component Analysis), which can retain the main time sequence characteristics of the signals and compress the dimension of the data, providing data input for the subsequent process.

[0076] It should be noted that in the subsequent process, only the one-dimensional fused time-domain signal under the same operating state is processed to determine the operating state of the hydropower unit, and data from different operating states cannot be used interchangeably.

[0077] Step 4: In this embodiment, image encoding of the one-dimensional fused time-domain signal is performed using GAF. See [link to relevant documentation]. Figure 2 The process is as follows:

[0078] Given a time series Given n data points, use GAF to generate two images of size n. GADF and GASF images;

[0079] First, a piecewise aggregation approximation is used to smooth and compress the time series. For a time series with n data points... Using a new time series with m data points It is represented as follows, in which The new time series Y is represented as a compressed signal, and the calculation of the new time series Y is shown in equation (1):

[0080] (1)

[0081] In the formula: And rounding up indicates that the time series X is divided into k segments; Represents the first in the new time series Y i One element; x j Represents the first time series X j One element; y i Represents the first time series Y i One element; Represents the time series number 1 i The last point of the segment.

[0082] New time series Normalizing to the interval [-1, 1] will yield... As a normalized signal; the calculation is shown in equation (2):

[0083] (2)

[0084] In the formula: Represents the first in the new time series Y i One element; Representing normalized time series The i One element; This represents the maximum value in the new time series Y; denotes the minimum value in the new time series Y.

[0085] The normalized signal is mapped to polar coordinates, the calculation of which is shown in equation (3):

[0086] (3)

[0087] In the equation: is the polar angle after the inverse cosine transformation; is the polar radius; is the time stamp; M is a constant factor of the normalized polar coordinate system.

[0088] It should be noted that the above mapping is a bidirectional mapping that does not lose the original data feature information. For a given arbitrary time series, there is only one result after mapping to polar coordinates, and its inverse mapping is also unique, which can be used to accurately restore the data. At the same time, the polar coordinate system retains the time information of the original time series, and the corresponding time value can be determined through the polar radius value.

[0089] Finally, the Gram angle and difference field encoding is performed, the data points of the polar coordinates are encoded into the corresponding Gram angle field, the sine value of the angle difference between different polar coordinate data points is used to construct the GADF image, and the cosine value of the angle sum between different polar coordinate data points is used to construct the GASF image, the calculation of which is shown in equations (4) and (5):

[0090] (4)

[0091] (5)

[0092] In the equation: GADF denotes the Gram angle difference field; GASF denotes the Gram angle sum field; denotes the sine; denotes the cosine; denotes the angle of the data at the 1st, 2nd, and mth time points in the polar coordinate system, respectively.

[0093] Step five, construct a DCLCNN (Dual-Channel Lightweight Convolution Neutral Network) model, train the constructed DCLCNN model to obtain a trained DCLCNN model, and the training process is as follows:

[0094] The GADF image and the GASF image obtained in step four are formed into a data set, and the data set is randomly divided into a training set, a validation set and a test set according to a ratio of 7:2:1; the training set is subjected to data enhancement, and similar training samples are generated and saved by means of random flipping, random brightness change, random Gaussian blur and the like, thereby forming a data-enhanced training set, a data-enhanced validation set and a data-enhanced test set; and the data-enhanced training set, the data-enhanced validation set and the data-enhanced test set are input into the DCLCNN model for training, thereby obtaining the trained DCLCNN model.

[0095] Referring to Figure 3 In the embodiment, the DCLCNN model is composed of two parallel feature extraction branch layers, a feature fusion layer and a classification layer, and the feature extraction branch among them is used for feature extraction of the GADF image and the other feature extraction branch is used for feature extraction of the GASF image.

[0096] In the embodiment, the two feature extraction branches have the same structure but different processing objects; the feature extraction branch includes an initial feature extraction layer, a deep feature extraction layer and a feature recombination layer.

[0097] In the embodiment, the initial feature extraction layer adopts 3x3 standard convolution, adopts a ReLU (Rectified Linear Unit) activation function and has a stride of 2, thereby obtaining preliminary features;

[0098] The deep feature extraction layer includes four layers of inverted residual structures, wherein the inverted residual structure contains an MLCA (Mixed Local Channel Attention) attention mechanism; each layer of the inverted residual structure sequentially performs 1x1 point-by-point convolution and 3x3 deep convolution on the preliminary features, while adopting a ReLU activation function for linear activation, the first layer and the third layer of the inverted residual structure have a stride of 2, and the second layer and the fourth layer of the inverted residual structure have a stride of 1, thereby obtaining deep convolution output.

[0099] Referring to Figure 4 In the embodiment, the MLCA attention mechanism adopts local average pooling and global average pooling to extract multi-scale information from the deep features, wherein the local average pooling captures spatial local detail features, and the global average pooling extracts global context information, thereby obtaining pooled local features and global features.

[0100] The pooled local features and the global features are respectively subjected to channel compression and information conversion by using one-dimensional convolution, the local and global features after one-dimensional convolution are respectively restored to the initial local pooling dimension by using inverse average pooling and reconstruction operation, and local and global information fusion is completed by point-by-point addition operation.

[0101] The fused local and global features are restored to the original input dimension through reverse average pooling, combined with the deep convolution output through point-by-point multiplication to obtain calibrated features to realize the fusion of channel and spatial features, local and global features, thereby enhancing the model feature expression capability.

[0102] Finally, the calibrated features are subjected to 1x1 point-by-point convolution to obtain deep features.

[0103] Referring to Figure 5 The inverted residual structure of the embodiment adopts a channel management strategy of dimension reduction-convolution-dimension increase, reduces the number of parameters through 1x1 point-by-point convolution for dimension reduction, performs 3x3 deep convolution for feature extraction, and finally recovers the number of feature channels at the output end using 1x1 point-by-point convolution and introduces a skip connection to allow direct cross-layer transmission of feature information in the network, thereby improving the training difficulty of the DCLCNN model while maintaining the complexity of the DCLCNN model and ensuring the performance of the deep network.

[0104] In the embodiment, the feature reorganization layer performs channel reorganization on the deep features subjected to 1x1 standard convolution, adopts an h-swish activation function with a step size of 1 to obtain first multi-scale features or second multi-scale features, wherein the output of the GADF image is the first multi-scale features and the output of the GASF image is the second multi-scale features.

[0105] The feature fusion layer in the embodiment is used to fuse the first multi-scale features and the second multi-scale features to obtain fused features, wherein the feature fusion layer includes a pooling layer, a concatenation layer and a flattening layer.

[0106] The pooling layer adopts global average pooling with a step size of 1 to extract global statistical information of each channel of the first multi-scale features and the second multi-scale features to obtain first feature vectors and second feature vectors.

[0107] The concatenation layer uses a dimension merging operation to concatenate the first feature vectors and the second feature vectors in the channel dimension to obtain concatenated vectors.

[0108] The flattening layer flattens the concatenated vectors into one-dimensional vectors to obtain fused features.

[0109] The classification layer in the embodiment is composed of two fully connected layers. First, the first fully connected layer performs a dimension increasing operation on the fused features and introduces an h-swish nonlinear activation function to improve the complex feature expression capability of the DCLCNN model. Meanwhile, a Dropout layer is added to randomly shield part of the neurons to reduce the risk of overfitting of the DCLCNN model to obtain high-dimensional features.

[0110] Finally, the last fully connected layer reduces the dimension of the high-dimensional features and maps them to the classification output space, outputting the operation state recognition result of the hydroelectric generating set.

[0111] The embodiment provides verification results of the method, and compares performances of a single-channel model processing a GADF image, the single-channel model processing a GASF image, and a DCLCNN model processing a GADF+GASF image in terms of accuracy, precision, recall, and F1 score.

[0112] The DCLCNN model processes the GADF+GASF image.

[0113] One of the two feature extraction branches in the DCLCNN model is removed, the GADF image and the GASF image are processed by the single feature extraction branch respectively, the splicing layer and the flattening layer in the feature fusion layer are removed, and the original classification layer is retained, so as to construct a single-channel model.

[0114] The parameters of the DCLCNN model and the single-channel model are uniformly set, the training round is 50, and the batch size is 16. AdamW (Adaptive Moment Estimation with Weight Decay Fix) and cross-entropy are used as the optimizer and the loss function respectively.

[0115] The cosine annealing learning rate decay algorithm is used in the training stage of the DCLCNN model and the single-channel model, and the learning rate is dynamically adjusted to accelerate the convergence of the model. The initial learning rate is 0.005.

[0116] The results are shown in Table 2, wherein the DCLCNN model performs best, with an accuracy of 99.51%, a precision of 99.55%, a recall of 98.82%, and an F1 score of 99.17%, indicating that the DCLCNN model can effectively enhance the feature expression ability of the model and exhibit stronger robustness and higher recognition accuracy in the state recognition task.

[0117] Table 2. Comparison of state recognition accuracy of single-channel and double-channel

[0118]

[0119] The embodiment further verifies the performance of the DCLCNN model in the method.

[0120] Replace the inverted residual structure in the feature extraction layer of the DCLCNN model with a standard convolution to obtain model 1; add an MLCA attention mechanism to model 1 to obtain model 2; replace the inverted residual structure in the feature extraction layer of the DCLCNN model with a module in ShuffleNetV2 (ShuffleNet version 2) to obtain model 3; replace the inverted residual structure in the feature extraction layer of the DCLCNN model with a module in Xception (extreme convolutional network) to obtain model 4.

[0121] The identification accuracy of each model is analyzed using accuracy, precision, recall, F1 score and other indicators, as shown in Table 2. Figure 6 As shown in Table 2, the accuracy of the DCLCNN model is 99.51%, the precision is 99.55%, the recall is 98.82%, and the F1 score is 99.17%. The DCLCNN model performs excellently in all evaluation indicators and far exceeds other models.

[0122] The identification efficiency of different models is shown in Table 3. Considering the decomposition operation of the depth separable convolution in the DCLCNN model and the multi-scale feature fusion of the MLCA, the memory consumption will increase to some extent when performing double-channel image feature extraction, and the network inference speed will be slightly affected. The inference time of the constructed model is not more than 3ms, and the parameter quantity and computational complexity of the DCLCNN model are at a low level. At the same time, the DCLCNN model is superior to other models in identification accuracy and can meet the real-time state identification requirements of hydroelectric generating units.

[0123] Table 3 Comparison of identification efficiency of different double-channel models

[0124]

[0125] Based on the same inventive concept, the embodiments of the present application also provide a hydroelectric generating unit operating state identification system based on a Gram angle field. Since the principle of solving problems of the hydroelectric generating unit operating state identification system based on the Gram angle field is similar to the aforementioned hydroelectric generating unit operating state identification method based on the Gram angle field, the implementation of the hydroelectric generating unit operating state identification system based on the Gram angle field can be referred to the implementation of the hydroelectric generating unit operating state identification method based on the Gram angle field, and the repeated parts will not be described again.

[0126] In specific implementation, the hydroelectric generating unit operating state identification system based on the Gram angle field provided by the embodiments of the present application specifically comprises:

[0127] A data acquisition module is configured to acquire state parameters under different operating states of a hydroelectric generating unit.

[0128] A data preprocessing module is configured to preprocess the state parameters under different operating states to obtain a one-dimensional fused time domain signal.

[0129] an image encoding module, configured to perform image encoding on the one-dimensional fused time-domain signal based on the GAF to obtain a GADF image and a GASF image;

[0130] a result output module, configured to input the GADF image and the GASF image under the same operation state into a pre-constructed DCLCNN model, the DCLCNN model comprising two parallel feature extraction branches, a feature fusion layer and a classification layer; one of the feature extraction branches extracts features of the GADF image to obtain first multi-scale features, and the other of the feature extraction branches extracts features of the GASF image to obtain second multi-scale features; the feature fusion layer fuses the first multi-scale features and the second multi-scale features to obtain fused features; and the classification layer maps and outputs the fused features to obtain an operation state recognition result of the hydroelectric generating set.

[0131] Correspondingly, the embodiment of the present application further provides a hydroelectric generating set operation state recognition device based on Gram angle field, comprising a processor and a memory, wherein the processor executes a computer program saved in the memory to realize the hydroelectric generating set operation state recognition method based on Gram angle field provided by the embodiment of the present application.

[0132] The more specific process of the above method can refer to the corresponding content disclosed in the foregoing embodiments, which will not be repeated here.

[0133] Correspondingly, the embodiment of the present application further provides a computer readable storage medium for storing a computer program, wherein the computer program is executed by a processor to realize the above hydroelectric generating set operation state recognition method based on Gram angle field provided by the embodiment of the present application.

[0134] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts of each embodiment can be referred to each other. For the system, device and storage medium disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0135] Those skilled in the art will further appreciate that the elements and algorithms described in connection with the examples disclosed herein can be embodied in electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various examples have been described herein in terms of exemplary components and processes that can include various examples of hardware and / or software. Whether such functionality is implemented in hardware or software depends on the particular application and design constraints imposed on the overall system. Skilled artisans can implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.

[0136] The steps of a method or algorithm described in connection with the examples disclosed herein can be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module can reside in random access memory (RAM), flash memory, read-only memory (ROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0137] Finally, it should be noted that the terms "first", "second", and the like, herein do not denote any order, quantity, combination, or importance, but rather are used to distinguish one element from another, and are more especially used for the purpose of distinction from other elements in the specification. Also, the terms "include", "comprise", or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements does not include only those elements recited, but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. The terms "exemplary" and "embodiment" do not necessarily refer to the only or the best example, but instead indicate that the example is one of a possible set of examples.

[0138] The above describes in detail the method, system, device and storage medium provided by the application for identifying the operation state of a hydroelectric generator set based on a Gram angle field. The principles and implementation modes of the application are described using specific examples. The above description of the examples is only intended to help understand the method of the application and its core idea. For those skilled in the art, the specific implementation modes and application scope can be changed according to the idea of the application. In summary, the content of the specification should not be interpreted as limiting the application.

Claims

1. A method for identifying the operating state of a hydroelectric generating unit based on the Gram angle field, characterized by, The method comprises the following steps: S1, obtaining state parameters under different operating states of a hydroelectric generating set; S2, preprocessing the state parameters under different operating states to obtain a one-dimensional fusion time domain signal; S3, image encoding the one-dimensional fusion time domain signal based on GAF to obtain a GADF image and a GASF image; S4, inputting the GADF image and the GASF image under the same operating state into a pre-constructed DCLCNN model, the DCLCNN model comprising two parallel feature extraction branches, a feature fusion layer and a classification layer; The two feature extraction branches are structurally identical, each comprising an initial feature extraction layer, a deep feature extraction layer and a feature recombination layer; The initial feature extraction layer performs 3×3 standard convolution on the input GADF image or GASF image, adopts a ReLU activation function and has a stride of 2 to obtain preliminary features; The deep feature extraction layer comprises four inverted residual structures, each of which comprises an MLCA attention mechanism; The inverted residual structure sequentially performs 1×1 point convolution and 3×3 deep convolution on the preliminary features to obtain deep convolution output, calibrates the deep convolution output through the MLCA attention mechanism to obtain calibrated features, and performs 1×1 point convolution on the calibrated features to obtain deep features; The feature recombination layer performs 1×1 standard convolution on the deep features, adopts an h-swish activation function and has a stride of 1; One of the feature extraction branches extracts GADF image features to obtain first multi-scale features, and the other feature extraction branch extracts GASF image features to obtain second multi-scale features; The feature fusion layer sequentially performs global average pooling, dimension merging and flattening on the first multi-scale features and the second multi-scale features to obtain fusion features; The classification layer maps the fusion features to output, and comprises two fully connected layers, the first fully connected layer increases the dimension of the fusion features and introduces an h-swish activation function and a Dropout layer to obtain high-dimensional features, and the second fully connected layer reduces the dimension of the high-dimensional features and maps the classification to obtain the operating state recognition result of the hydroelectric generating set.

2. The method for identifying the operating state of a hydroelectric generating unit based on the Gram angular field according to claim 1, characterized in that, The state parameters in S1 include swing signals and vibration signals; The swing signals include X-direction swing of the upper guide bearing, Y-direction swing of the upper guide bearing, X-direction swing of the lower guide bearing, Y-direction swing of the lower guide bearing, X-direction swing of the water guide bearing and Y-direction swing of the water guide bearing; The vibration signals include X-direction vibration of the upper bracket, Y-direction vibration of the upper bracket, X-direction vibration of the lower bracket, Y-direction vibration of the lower bracket, X-direction vibration of the generator stator and Y-direction vibration of the generator stator; The different operating states include stable operating conditions, transition operating conditions and abnormal operating conditions; The stable operating conditions are divided into low-load operating conditions, medium-low-load operating conditions, medium-load operating conditions and high-load operating conditions according to the percentage of the current operating load relative to the rated load; 0% or more to less than 20% of the current operating load is classified as a low-load operating condition; 20% or more to less than 40% of the current operating load is classified as a medium-low-load operating condition; 40% or more to less than 60% of the current operating load is classified as a medium-load operating condition; dividing 60% and above of current running load to 80% below into middle-high load running condition; dividing 80% and above of current running load to 100% into high load running condition; the transition running condition includes start-up process, shutdown process, speed increasing and decreasing process, and load increasing and decreasing process; the abnormal running condition includes hydraulic vibration abnormality, mechanical vibration abnormality, and electromagnetic vibration abnormality.

3. The method for identifying the operating state of a hydroelectric generating unit based on the Gram angular field according to claim 1, characterized in that, the preprocessing in S2 includes adopting PCA to perform signal fusion on the state parameters under different running states to obtain one-dimensional fused time domain signal.

4. The method for identifying the operating state of a hydroelectric generating unit based on the Gram angular field according to claim 1, characterized in that, the process of S3 includes: performing smoothing compression on the one-dimensional fused time domain signal to obtain compressed signal; performing normalization processing on the compressed signal to obtain normalized signal; performing polar coordinate conversion processing on the normalized signal to obtain polar coordinates, the polar coordinates including polar angle and polar radius; performing gram angle and difference field coding processing on the polar coordinates, obtaining GADF image by calculating the sine value of the angle difference between the data points in the polar coordinates, and obtaining GASF image by calculating the cosine value of the angle sum between the data points in the polar coordinates.

5. A system for identifying the operating state of a hydroelectric generating unit based on the Gram angle field, applied to the method for identifying the operating state of a hydroelectric generating unit based on the Gram angle field according to any one of claims 1 to 4, characterized in that, comprise: a data acquisition module configured to acquire state parameters of a hydroelectric generating set under different running states; a data preprocessing module configured to preprocess the state parameters under different running states to obtain one-dimensional fused time domain signal; an image coding module configured to perform image coding on the one-dimensional fused time domain signal based on GAF to obtain GADF image and GASF image; a result output module configured to input the GADF image and the GASF image under the same running state into a pre-constructed DCLCNN model, the DCLCNN model comprising two parallel feature extraction branches, a feature fusion layer, and a classification layer; one of the feature extraction branches extracts features of the GADF image to obtain first multi-scale features, and the other of the feature extraction branches extracts features of the GASF image to obtain second multi-scale features; the feature fusion layer fuses the first multi-scale features and the second multi-scale features to obtain fused features; and the classification layer maps and outputs the fused features to obtain a running state recognition result of the hydroelectric generating set.

6. A device for identifying the operating state of a hydroelectric unit based on the Gram angle field, characterized by comprise a processor and a memory, wherein the processor implements the gram angle field-based hydroelectric generating set running state recognition method according to any one of claims 1 to 4 when executing a computer program stored in the memory.

7. A computer readable storage medium characterized in that, a computer program for storing, wherein the computer program is executed by a processor to implement the gram angle field-based hydroelectric generating set running state recognition method according to any one of claims 1 to 4.

Images