Rotor type fault diagnosis method and device for rotating equipment

Abstract

The invention provides a rotor type fault diagnosis method and device for rotating equipment, and the method comprises the steps: obtaining original vibration data of to-be-diagnosed equipment, and carrying out the data processing and feature fusion of the original vibration data, so as to obtain fusion data; inputting the fusion data into a pre-trained fault diagnosis model to obtain a diagnosis result output by the fault diagnosis model; wherein the fault diagnosis model is obtained by training a pre-constructed VGG-SVM model by using sample data and a corresponding diagnosis result label, and the sample data is fused data obtained by fusing multi-source original data; the sample data is multi-source original data obtained at a plurality of target sampling points of a rotating equipment sample. The problems of lagging and poor accuracy of rotor fault diagnosis in the prior art are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent diagnosis of high-end equipment, and particularly to a rotor fault diagnosis method and device for rotating equipment. Background Art

[0002] Turbomachinery represented by high-end grinders, steam turbine generator sets, pump sets, compressors, and engines, as common types of mechanical equipment, is widely used in industrial fields such as petrochemical, aerospace, and metallurgical manufacturing. With the construction and development of the equipment manufacturing industry and the industrial Internet, mechanical equipment is becoming increasingly large-scale, safe, and intelligent. Based on the equipment self-diagnosis principle in the artificial self-healing theory, the condition monitoring and fault diagnosis of rotating equipment are the solid foundation and key to realizing equipment intelligence. For rotating machinery, the rotor is one of the most critical and fault-prone components, with the greatest impact on vibration during operation. Common faults include rotor imbalance, rotor misalignment, and looseness. Different faults have different vibration principles and induced hazards, and must be accurately identified to prescribe the right remedy, otherwise it will only be counterproductive and achieve half the result with twice the effort. Therefore, realizing the intelligent, fast, and accurate diagnosis of typical rotor faults in rotating equipment is crucial for system design and equipment maintenance, and has engineering practical significance.

[0003] In the initial stage, equipment operation and maintenance were basically carried out by experienced operators or professional maintenance personnel, who judged operation and maintenance through signals such as sound and temperature. In this case, the equipment was often severely damaged, and due to differences in the operating characteristics and structural parameters of different equipment, there were significant differences in the diagnostic results, and the experience passed down by word of mouth was not suitable for wide dissemination and application. With the development of computers, some machine learning and signal recognition began to be applied to some equipment with simple structures, but often required manual selection of features and multi-faceted processing of vibration signals, requiring operators to have a high level of professional quality, and there were difficulties in large-scale application. With the diversification of production and life, equipment is also developing towards large-scale and complex. Basic time-domain waveforms and spectrograms can no longer meet the requirements of equipment diagnosis. With the development of industrial big data and equipment monitoring technology, existing technologies have integrated deep learning technology with mechanical equipment monitoring and diagnosis.

[0004] However, there are many problems with existing artificial intelligence diagnosis algorithms. For example, algorithms such as ANN and SVM are prone to gradient explosion and data overfitting; algorithms such as deep learning require a large amount of sample data and take a long time to train. In short, a single algorithm model diagnosis process is difficult to meet diverse and complex requirements. For rotor components, their faults such as imbalance, misalignment, and looseness are mostly a type of fault dominated by power frequency. There are certain similarities in their spectrum characteristics. For example, the fault characteristics in normal and unbalanced states are mainly single-frequency, and misalignment and looseness states both contain single-frequency and double-frequency. Therefore, it is difficult to accurately identify specific faults by relying on a single spectrum analysis and a simple diagnosis model; at the same time, the vibration signal of the rotor during movement will also make it more difficult to identify due to the influence of flexural deformation. More importantly, during the operation of the equipment, the acquisition signal is insensitive due to reasons such as the measurement point location, vibration transmission path, and multi-position vibration coupling. The influence of transmission structure and noise will also cause vibration attenuation or change, making the frequency component unclear. Spectral analysis cannot accurately capture fault information, and ultimately cannot effectively reflect fault characteristics and accurately identify them.

[0005] In summary, the rotor fault diagnosis of rotating equipment in the prior art has the problems of hysteresis and poor fault identification accuracy. Summary of the invention

[0006] The present invention provides a rotor fault diagnosis method and device for rotating equipment, which are used to solve the defects of delayed and poor accuracy of rotor fault diagnosis in the prior art, so as to realize timely and accurate identification of rotor faults of rotating equipment.

[0007] The present invention provides a rotor fault diagnosis method for a rotating device, the method comprising:

[0008] Acquire original vibration data of the device to be diagnosed, and perform data processing and feature fusion on the original vibration data to obtain fused data;

[0009] Inputting the fused data into a pre-trained fault diagnosis model to obtain a diagnosis result output by the fault diagnosis model;

[0010] Among them, the fault diagnosis model is obtained by training a pre-built VGG-SVM model using sample data and corresponding diagnosis result labels, the sample data is fused data obtained after fusing multi-source original data, and the sample data is multi-source original data obtained at multiple target sampling points of rotating equipment samples.

[0011] In some embodiments, data processing and feature fusion are performed on the raw vibration data to obtain fused data, specifically including:

[0012] Perform data partitioning, data cleaning, normalization, and data augmentation on the original vibration data in sequence to obtain fused data of different fault types with consistent data sizes.

[0013] In some embodiments, use the fused sample data and the corresponding diagnostic result labels to train a pre-constructed VGG-SVM model to obtain the fault diagnosis model, which specifically includes:

[0014] Multi-source original data obtained at multiple target sampling points of rotating equipment samples, and the sample data obtained after fusing the multi-source original data;

[0015] Construct a sample data set using paired sample data and the corresponding diagnostic result labels, and divide the sample data set into a training set and a validation set;

[0016] Use the sample data in the training set to train a pre-constructed VGG network to obtain an initial model;

[0017] Use the feature data obtained after extracting features by the convolutional blocks in the initial model to perform secondary training on a pre-constructed SVM network to obtain an integrated fault diagnosis model.

[0018] In some embodiments, the network architecture of the VGG network specifically includes 5 convolutional blocks. Among them, the first two convolutional blocks contain 2 convolutional layers and 1 pooling layer, and the last three convolutional blocks contain 3 convolutional layers and 1 pooling layer, and finally are connected through 3 fully connected layers.

[0019] In some embodiments, during the training process, the VGG network adopts the stochastic gradient descent algorithm, with an initial learning rate of 0.01, a momentum of 0.9, a weight decay coefficient of 1e-4, and the number of training batches is 50.

[0020] In some embodiments, during the secondary training of the SVM network, feature data is obtained after extracting features by the convolutional blocks of the VGG network, input into the SVM network for a new round of feature extraction and training, and by adjusting the penalty coefficient C of the SVM, the optimal solution of the support vector is found to obtain an integrated fault diagnosis model.

[0021] In some embodiments, after obtaining the integrated fault diagnosis model, it further includes:

[0022] Randomly extract multiple groups of data from the multi-source original data as a test set;

[0023] Input the test set into the trained fault diagnosis model, and test the fault diagnosis model through the test results.

[0024] The present invention also provides a rotor fault diagnosis device for a rotating device, the device comprising:

[0025] A data acquisition unit, configured to acquire the original vibration data of the device to be diagnosed, and perform data processing and feature fusion on the original vibration data to obtain fusion data;

[0026] A result generation unit, configured to input the fusion data into a pre-trained fault diagnosis model to obtain a diagnosis result output by the fault diagnosis model;

[0027] Wherein, the fault diagnosis model is obtained by training a pre-constructed VGG-SVM model using sample data and corresponding diagnosis result labels, the sample data is fusion data obtained by fusing multi-source original data, and the sample data is multi-source original data obtained at multiple target sampling points of a rotating device sample.

[0028] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, where when the processor executes the program, it implements the rotor fault diagnosis method for a rotating device as described in any one of the above.

[0029] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the rotor fault diagnosis method for a rotating device as described in any one of the above.

[0030] The rotor fault diagnosis method and device provided by the present invention, by acquiring the original vibration data of the device to be diagnosed, and performing data processing and feature fusion on the original vibration data to obtain fusion data; inputting the fusion data into a pre-trained fault diagnosis model, the diagnosis result output by the fault diagnosis model can be obtained; wherein, the fault diagnosis model is based on an integrated model algorithm of VGG and SVM, combining the algorithm advantages, using the excellent feature extraction ability of the VGG network and the powerful non-linear classification ability of SVM, after two feature extractions, the risk of gradient explosion and overfitting can be effectively prevented during the model construction and testing process, the fault features are more obvious, and the diagnosis accuracy is higher, thus solving the defects of lagging and poor accuracy in rotor fault diagnosis in the prior art, and realizing the timely and accurate identification of rotor faults of rotating devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0032] Figure 1 One of the flow schematic diagrams of the rotor fault diagnosis method for rotating equipment provided by the present invention;

[0033] Figure 2 Schematic diagram of the construction principle of the VGG_16 model provided by the present invention;

[0034] Figure 3 Schematic diagram of the construction principle of the VGG-SVM integrated model provided by the present invention;

[0035] Figure 4 Another flow schematic diagram of the rotor fault diagnosis method for rotating equipment provided by the present invention;

[0036] Figure 5 For Figure 2 Schematic diagram of the convolutional block in the VGG_16 model shown;

[0037] Figure 6 Schematic diagram of the SVM classification principle provided by the present invention;

[0038] Figure 7 Schematic diagram of the construction result of the 2100rpm model of the double-span three-support rotor test bench provided by the embodiment of the present invention;

[0039] Figure 8 For Figure 7 Test result display diagram in the shown embodiment;

[0040] Figure 9 For Figure 7 Fault diagnosis test result display diagram in the shown embodiment;

[0041] Figure 10 Structural block diagram of the rotor fault diagnosis device for rotating equipment provided by the present invention;

[0042] Figure 11 Schematic diagram of the structure of the electronic device provided by the present invention. Detailed implementation manners

[0043] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without making creative efforts shall fall within the protection scope of the present invention.

[0044] To solve the problems existing in the prior art, the method provided by the present invention is for the intelligent diagnosis of typical faults of high-end turbomachinery rotors. Through the method of multi-source data fusion, data fusion, feature fusion, and decision fusion are carried out, and an integrated model is used to make intelligent early warnings for common unbalances, misalignments, and loosenesses of rotating equipment. Diagnostic algorithms are used to ensure the stability and reliability of equipment operation, maintain equipment safety, and increase the service life of equipment.

[0045] In a specific embodiment, as Figure 1 shown, the rotor fault diagnosis method for rotating equipment provided by the present invention includes the following steps:

[0046] S110: Obtain the original vibration data of the equipment to be diagnosed, and perform data processing and feature fusion on the original vibration data to obtain fusion data; during the diagnosis process, use sensors set at the acquisition points to collect multiple sets of original vibration data at multiple locations. The acquisition points can be determined according to usage requirements. For example, multiple data acquisition points are set at key parts such as bearings, motors, and couplings.

[0047] S120: Input the fusion data into a pre-trained fault diagnosis model to obtain the diagnosis result output by the fault diagnosis model; specifically, the output form of the diagnosis result can be in the form of a table or a visual form such as a bar chart.

[0048] Among them, the fault diagnosis model is obtained by training a pre-constructed VGG-SVM model using sample data and corresponding diagnosis result labels. The sample data is the fusion data obtained by fusing multi-source original data, and the sample data is the multi-source original data obtained at multiple target sampling points of the rotating equipment sample.

[0049] In one embodiment, in the VGG-SVM integrated model, the VGG part uses the performance-better VGG_16 model. The VGG_16 model is the part that has been trained. After extracting features through 5 convolutional blocks, the obtained m×512×1×1 (data volume × number of channels × height × width) data size is as Figure 2As shown in the figure, it is then input into the SVM model for a new round of feature extraction and training. After removing the fully connected layer from the VGG_16 model, the data size is m×512. This data volume is input into the SVM (stacked autoencoder), and by adjusting the penalty coefficient C of the SVM, the optimal solution of the support vector is found, so as to achieve the output result that includes both feature extraction and final classification, as Figure 3 shown. By using the training method of the integrated model, the fault diagnosis accuracy can be effectively improved. After the data extracted by the VGG_16 feature extraction enters the SVM, the influence of noise and outliers can be effectively reduced. Based on the previously extracted features, further combination and abstraction are carried out to obtain more advanced data information, making the SVM classification and feature extraction effects more obvious. At the same time, after two rounds of feature extraction, the data features are more obvious, and even in the face of complex, multi-dimensional and mutually fused data, good classification and extraction can be achieved. Compared with a single deep learning network, the integrated model has stronger generalization ability and higher fault diagnosis accuracy.

[0050] In step S110, data processing and feature fusion are performed on the original vibration data to obtain fused data, which specifically includes:

[0051] The original vibration data is successively subjected to data division, data cleaning, normalization, and data augmentation to obtain fused data of different fault types with consistent data sizes.

[0052] Since the directly collected data cannot be directly used for intelligent algorithm fault diagnosis, a series of data preprocessing is required. Among them, data cleaning is used to remove the error data caused by electromagnetic interference and environmental noise; data normalization is used to solve the problem of too large differences in the data ranges of different physical quantities. In addition, in the actual use process, faults often occur in an instant, so the fault data acquisition time is less during the experiment. To ensure that the data volume is consistent with that in the normal state, in this embodiment, the stacked autoencoder (SAE) is used to perform data augmentation on the fault information. After preprocessing, the data volumes of different faults are basically the same. At this time, the fault signals in four states (unbalance, misalignment, looseness, and normal) are labeled and then feature fusion is performed.

[0053] In some embodiments, as Figure 4 shown, the pre-constructed VGG-SVM model is trained using the fused sample data and the corresponding diagnostic result labels to obtain the fault diagnosis model, which specifically includes the following steps:

[0054] S410: The multi-source raw data obtained at multiple target sampling points of the rotating equipment sample, and the sample data obtained after fusing the multi-source raw data; In a specific usage scenario, data is collected from a rotating equipment to be fault-diagnosed (i.e., the equipment to be diagnosed). After installing multiple sensors at different key parts and conducting different fault experiments, the size of the data collected in each state is A×B×16384 (A is the number of sensors, B is the amount of data acquisition waveforms, and 16384 is the number of data points in a single waveform). After fusing the original vibration data and randomly selecting, the size of the fused sample data is (N×16384). After data preprocessing such as data partitioning, data cleaning, normalization, and data augmentation, different fault type data with basically the same size of sample data is obtained. After setting labels for different fault types, they are fused and input, and finally the decision fusion output model verification result is obtained through the confusion matrix method.

[0055] During data acquisition and fusion, since the method provided by the present invention is applied to rotating equipment of different rotor types, the first problem to be solved is fault diagnosis data acquisition. Since the fault diagnosis algorithm proposed by the present invention targets complex fault types, including rotor imbalance, rotor misalignment, and looseness faults, these faults are all synchronous faults mainly with power frequency. The signal characteristics are slightly different but also contain each other, making it difficult to identify. Therefore, during data acquisition, in addition to collecting data on fault signals, in order to verify the fault identification effect, a simulation experiment under normal conditions is designed and carried out simultaneously as a control group to participate in model training and testing. During the data acquisition process, eddy current displacement sensors are selected, and multiple sensor measuring points are arranged at key parts such as bearings, motors, and couplings. One of the multi-source data fusions proposed by the present invention is to fuse multi-source data from different measuring points, making the data source more universal;

[0056] Furthermore, the directly collected data cannot be directly used for intelligent algorithm fault diagnosis and requires a series of data preprocessing, including using data cleaning to remove error data caused by electromagnetic interference and environmental noise; using data normalization to solve the problem of too large differences in the data ranges of different physical quantities. In addition, in actual production and life, faults often occur in an instant, so the fault data acquisition time is less during the experiment. To ensure the same amount of data as in the normal state, a stacked autoencoder (SAE) is used to perform data augmentation on the fault information. After preprocessing, the amounts of different fault data are basically the same. At this time, the fault signals in the four states are set with labels and then feature fusion is performed.

[0057] S420: Construct a sample data set using paired sample data and corresponding diagnostic result labels, and divide the sample data set into a training set and a validation set; during feature extraction and decision fusion, since the extracted fault features need to be able to truly reflect the device fault state, time-domain feature analysis is selected for vibration signals. The integrated model needs to first select a suitable basic diagnostic algorithm, analyze and evaluate the algorithms of the single models VGG and SVM, prevent the extracted features from being redundant or irrelevant, increase the complexity for subsequent calculations, and avoid overfitting. After selecting the basic diagnostic algorithm, select the results of model training and validation. In this embodiment, it is expressed using a confusion matrix (4×4), which can simultaneously identify and verify the vibration states under four different conditions.

[0058] S430: Use the sample data in the training set to train a pre-constructed VGG network to obtain an initial model;

[0059] S440: Use the feature data obtained after extracting features by the convolutional blocks in the initial model to perform secondary training on a pre-constructed SVM network to obtain an integrated fault diagnosis model.

[0060] During the training process, combine VGG_16 and SVM, integrate the advantages of the two algorithms, and make the model construction effect better through two feature extractions. First, train the data through the VGG_16 model, then remove the fully connected layer of the trained part, input it into the SVM algorithm, find the optimal solution of the support vector by adjusting the penalty coefficient C of the SVM, and finally output the trained integrated algorithm model.

[0061] In some embodiments, the network architecture of the VGG network specifically includes 5 convolutional blocks. Among them, the first two convolutional blocks contain 2 convolutional layers and 1 pooling layer, the last three convolutional blocks contain 3 convolutional layers and 1 pooling layer, and finally they are connected through 3 fully connected layers. During the training process, the VGG network uses the stochastic gradient descent algorithm, with an initial learning rate of 0.01, a momentum of 0.9, a weight decay coefficient of 1e-4, and the number of training batches is 50.

[0062] Specifically, the VGG network is preferably the VGG_16 convolutional neural network. In the field of deep learning, convolutional neural networks have become one of the core technologies for data processing. As a highly representative one, VGG has excellent performance and far-reaching influence. In this embodiment, the VGG_16 network is selected, that is, the number of layers of trainable parameters in the network is 16 (13 convolutional layers and 3 fully connected layers). The VGG_16 network as a whole includes 5 convolutional blocks, and each convolutional block is composed of consecutive convolutional layers and pooling layers. Specifically, the first two convolutional blocks contain 2 convolutional layers, the middle two convolutional blocks each contain 3 convolutional layers, and the last convolutional block contains 3 convolutional layers, asFigure 5 as shown

[0063] The size of the convolution kernel used in each convolutional layer is 3×3, the stride is usually set to 1, and the same padding method is adopted during the convolution operation, so that the size of the feature map remains relatively consistent in the spatial dimension after convolution; after each convolutional block in VGG_16, there is a pooling layer immediately following. In this embodiment, the maximum pooling method is adopted in VGG_16, the size of the pooling kernel is 2×2, and the stride is 2. Through the pooling layer, the data extracted by the convolutional layer can be downsampled. On the one hand, it can reduce the amount of data and the complexity of subsequent calculations. On the other hand, it can maintain the main information of the features to a certain extent, enhance the robustness of the network to features and the adaptability to different sizes. After completing the feature extraction and downsampling of the five convolutional blocks, the network flattens the feature data output by the last pooling layer into a one-dimensional vector, and then connects to 3 fully connected layers to integrate and classify the features extracted previously.

[0064] The deep characteristics of VGG_16 make it have stronger expressive power compared with shallower network structures. More layers mean that more complex and more discriminative feature representations can be learned. However, the increase in depth will also bring risks such as gradient disappearance and overfitting. To solve such problems, the method provided in this embodiment uses the ReLU activation function to alleviate gradient disappearance during the use of VGG_16, so that the gradient can be relatively smooth during backpropagation, ensuring the effective training of the network; on the other hand, SAE data augmentation is used before training to reduce overfitting and improve the generalization ability of the model.

[0065] In the above embodiment, during the secondary training of the SVM network, feature data is obtained by extracting features through the convolutional blocks of the VGG network, and then input into the SVM network for a new round of feature extraction and training. By adjusting the penalty coefficient C of the SVM, the optimal solution of the support vector is found to obtain an integrated fault diagnosis model.

[0066] SVM (Support Vector Machine), as a common machine learning algorithm, has been widely used in many classification and regression tasks. According to different data dimensions and linear separability, there are roughly two situations. In the case of linear separability, for the training data set D={(x 1 ,y 1 ),(x 2 ,y 2 ),...,(x n ,y n )}, a hyperplane is found in its space to completely separate the two types of data points. The hyperplane in the two-dimensional space is a straight line, as Figure 6 as shown

[0067] A hyperplane in three-dimensional space is a plane, while a hyperplane in n-dimensional space is an n-1 subspace that satisfies a specific linear equation. The common linear plane equation is ω T x + b = 0, where ω is the normal vector of the hyperplane, which determines the direction of the hyperplane, and b is the bias term, which determines the position of the hyperplane in space. In practical applications, many data are often not linearly separable, that is, different types of data cannot be accurately separated by a simple linear hyperplane. Therefore, the concept of kernel function is introduced. Through a non-linear mapping maps the data point x in the original space to a higher-dimensional feature space H, that is, the Hilbert space, so that the data becomes linearly separable in the high-dimensional space.

[0068] For actual data, due to the influence of noise or outliers, if strictly following the hard margin requirement of maximizing the margin, the hyperplane may be too sensitive to noise and outliers, resulting in poor generalization ability of the model. Therefore, the concept of soft margin is proposed in the optimization process, allowing some data points to appear inside the margin or even be misclassified. By introducing the penalty coefficient C, the balance between maximizing the margin and allowing data points to violate the margin constraint is weighed. In summary, SVM constructs a powerful and flexible machine learning model with the help of kernel functions and soft margins, and is widely used in classification and regression tasks in many fields.

[0069] Furthermore, to verify the advantages of the method provided by the present invention, the same data is input into a single algorithm to train a single algorithm model, and the accuracy of the integrated model algorithm is determined by comparing the verification results. Specifically, after obtaining the integrated fault diagnosis model, it further includes:

[0070] Randomly extract multiple groups of data from the multi-source original data as the test set;

[0071] Input the test set into the trained fault diagnosis model, and test the fault diagnosis model through the test results.

[0072] That is to say, the data used for training and verification is removed, and the remaining part is used as the test set to test the integrated model. By comparing the model prediction results with the actual results, the test accuracy of the algorithm is verified to ensure the accuracy and stability of the intelligent algorithm.

[0073] In the above specific embodiments, the rotor fault diagnosis method for rotating equipment provided by the present invention obtains the original vibration data of the equipment to be diagnosed, processes the original vibration data and performs feature fusion to obtain fusion data; inputs the fusion data into a pre-trained fault diagnosis model, and the diagnosis result output by the fault diagnosis model can be obtained; wherein, the fault diagnosis model is based on an integrated model algorithm of VGG and SVM, combines the algorithm advantages, utilizes the excellent feature extraction ability of the VGG network and the powerful non-linear classification ability of SVM, and after two feature extractions, the risks of gradient explosion and overfitting can be effectively prevented during the model construction and testing processes, the fault features are more obvious, and the diagnosis accuracy is higher, thus solving the defects of lagging rotor fault diagnosis and poor accuracy in the prior art, and realizing the timely and accurate identification of rotor faults in rotating equipment.

[0074] Furthermore, in the method provided by the present invention, actual data collection is carried out by conducting different fault simulation experiments under different working conditions, and the signals collected by sensors at different measuring points are combined through original data fusion; after preprocessing the data, the vibration information of different fault types is subjected to feature fusion; for the basic diagnosis algorithm, it is proposed to use decision fusion to display the model construction results. By modifying and improving a single algorithm, it is proposed to integrate different intelligent diagnosis algorithms, and through two feature extractions, the intelligent diagnosis of typical rotor faults in rotating equipment is realized. After data testing, this method has high robustness, stability and accuracy.

[0075] For the sake of easy understanding, the implementation process and technical effects of the method provided by the present invention are briefly described below by taking a specific usage scenario as an example.

[0076] This embodiment takes a double-span three-support rotor system equipment as an example, selects the common working condition of 2100 rpm, arranges multiple sensors at key parts such as motors, bearings, and couplings. For example, two eddy current displacement sensors (1×45° and 1×135°) are installed diagonally at the bearing part. After setting up a faulty part to conduct a fault experiment for data collection, the original data is obtained. The data collected by all sensors under each fault type are fused and shuffled to obtain the data after data fusion. After data fusion, D groups of data are randomly selected and divided into samples of size F×1024 (32×32) for subsequent processing.

[0077] By observing the time-domain waveform of the sample data, the "bad points" collected are subjected to data cleaning, and a series of displacement vibration signals X 1 , X 2 , X 3 ,... X i ,..., X n are obtained, and they are normalized. Then, data augmentation is performed through SAE (Stacked Autoencoder); after data preprocessing, different labels are set for different fault types, unbalance (0), misalignment (1), looseness (2), normal (3), and then all types of fault signals are feature-fused to obtain an overall sample data set.

[0078] The sample data set is classified into a training set and a validation set in a ratio of 8:2 for model construction; the model is constructed using the Pytorch framework in Pycharm, and a single model algorithm and an ensemble model algorithm are built respectively.

[0079] Among them, SVM uses the scikit-learn 1.0.1 version. Since vibration coupling mostly shows non-linearity, the radial basis function (RBF) is selected as the kernel function, which has good fitting for non-linear continuous random functions; for actual data, due to the influence of noise or outliers, if the hard margin requirement of maximizing the margin is strictly followed, the classification hyperplane may be too sensitive to noise and outliers, resulting in poor generalization ability of the model. Therefore, the concept of soft margin is proposed in the optimization process, allowing some data points to appear inside the margin or even be misclassified; by introducing the penalty coefficient C, the balance between maximizing the margin and allowing data points to violate the margin constraint is weighed; during the training process, the penalty coefficient C is verified by setting different series of values to obtain the optimal value.

[0080] Among them, for the VGG part, VGG_16 is selected. VGG_16 contains 5 convolutional blocks. The first two convolutional blocks contain 2 convolutional layers and 1 pooling layer, and the last three convolutional blocks contain 3 convolutional layers and 1 pooling layer. Finally, it is connected through 3 fully connected layers; the size of the convolutional kernel used in each convolutional layer is 3×3, the stride is usually set to 1, and the same padding method is adopted during the convolution operation, so that the size of the feature map after convolution can remain relatively consistent in the spatial dimension; the pooling layer selects the maximum pooling method, the pooling kernel size is 2×2, and the stride is 2. Through the pooling layer, the data extracted by the convolutional layer can be downsampled. On the one hand, it can reduce the data volume and the complexity of subsequent calculations. On the other hand, it can maintain the main information of the features to a certain extent, enhance the robustness of the network to features and the adaptability to different sizes.

[0081] After completing the feature extraction and downsampling of the five convolutional blocks, the VGG network flattens the feature data output by the last pooling layer into a one-dimensional vector, and then connects to three fully connected layers to integrate and classify the previously extracted features; the VGG_16 optimizer uses stochastic gradient descent (SGD), the initial learning rate is 0.01, the momentum is set to 0.9 to prevent gradient discreteness, and the weight decay coefficient is introduced to prevent the model from overfitting. The size is set to 1e-4, and the number of training batches is 50. The learning rate adjuster mode selects max, and patience selects 4 times; in the process of building the integrated model, VGG_16 is first trained, and the trained model is exported in .pth format. The size of the trained data set is m×512×1×1 (data volume×number of channels×height×width). After removing the fully connected layer, the data size is m×512; the data after feature extraction is used as input to the SVM for secondary training, and the optimal solution is obtained by continuously adjusting the size of the penalty coefficient C. The obtained model training results are expressed in the form of a confusion matrix through decision fusion, as shown below: Figure 7 As shown in the figure, the final trained integrated model is exported in .pth format.

[0082] In order to further verify the fault diagnosis accuracy of the intelligent diagnosis algorithm proposed in this patent and for the actual application in subsequent projects, a model testing stage is deliberately added in the subsequent process of model construction. The data set used for the test comes from the fused data after removing the sample data. The remaining data is divided into 30 test sets for each fault type, and a total of 120 test sets are used as the test database.

[0083] 40 test sets were randomly selected from the database to test the diagnostic algorithm. The test type results are as follows: Figure 8 As shown, it is displayed in the form of a histogram, and the display content includes fault diagnosis rate, fault type, etc.; after testing 40 test sets, the model prediction results are compared with the actual results after inputting the model. If they are not correct, they are judged as errors if they are inconsistent. The test data is recorded as follows Figure 9 The test table shown shows that among the 40 test sets, 39 were correct and 1 was wrong, proving that the accuracy of the intelligent diagnosis algorithm results is as high as over 97%. Figure 8 Fault_tname represents the result of the algorithm diagnosis fault test set; the x-axis Fault_type represents the fault type (Unbalance: unbalance; Misalignment: misalignment; Looseness: looseness; Normal: normal); the y-axis Fault_taccuracy represents the fault diagnosis accuracy.

[0084] From the above test results, it can be seen that in this embodiment, for the fault diagnosis method provided by the present invention, by actually collecting data on the built test bench and verifying after multi-source fusion, it can be concluded that the integrated model algorithm has high stability, robustness, and accuracy, and can be effectively applied. Moreover, this method can effectively identify typical rotor faults of rotating equipment, not limited to the unbalanced state. It can effectively identify rotor misalignment, bolt loosening, etc., and can input multiple types of data at one time. This multi-fault identification ability can reduce misdiagnosis. Considering the mutual influence of multiple faults, the diagnosis result will be more accurate. The ability to identify multiple faults can ensure precise maintenance of the equipment, improve equipment availability, and provide a reliable basis for the optimal design of the equipment.

[0085] Furthermore, the relatively high diagnostic accuracy of the integrated model can effectively improve the reliability and availability of the equipment, reduce unexpected downtime, and extend the service life of the equipment. It can accurately detect problems at the early stage of the fault, even when the fault has not yet had an obvious impact on the operation of the equipment, facilitating timely measures to be taken. At the same time, this method can monitor the operating status of each production equipment, optimize the production process, ensure product quality. In addition, it can achieve precise maintenance based on the actual status of the equipment and reduce spare parts inventory.

[0086] In addition to the above method, the present invention also provides a rotor fault diagnosis device for rotating equipment, as Figure 10 shown. The device includes:

[0087] A data acquisition unit 1010, configured to obtain the original vibration data of the equipment to be diagnosed, and perform data processing and feature fusion on the original vibration data to obtain fusion data;

[0088] A result generation unit 1020, configured to input the fusion data into a pre-trained fault diagnosis model to obtain a diagnosis result output by the fault diagnosis model;

[0089] Wherein, the fault diagnosis model is obtained by training a pre-constructed VGG-SVM model using sample data and corresponding diagnosis result labels. The sample data is fusion data obtained by fusing multi-source original data, and the sample data is multi-source original data obtained at multiple target sampling points of rotating equipment samples.

[0090] In some embodiments, performing data processing and feature fusion on the original vibration data to obtain fusion data specifically includes:

[0091] Performing data partitioning, data cleaning, normalization, and data augmentation on the original vibration data in sequence to obtain fusion data of different fault types with consistent data sizes.

[0092] In some embodiments, the pre-constructed VGG-SVM model is trained using the fused sample data and the corresponding diagnostic result labels to obtain the fault diagnosis model, which specifically includes:

[0093] Multi-source raw data obtained at multiple target sampling points of the rotating equipment samples, and the sample data obtained after fusing the multi-source raw data;

[0094] A sample data set is constructed using paired sample data and the corresponding diagnostic result labels, and the sample data set is divided into a training set and a validation set;

[0095] The pre-constructed VGG network is trained using the sample data in the training set to obtain an initial model;

[0096] The pre-constructed SVM network is retrained using the feature data obtained after extracting features by the convolutional blocks in the initial model to obtain an integrated fault diagnosis model.

[0097] In some embodiments, the network architecture of the VGG network specifically includes 5 convolutional blocks. Among them, the first two convolutional blocks contain 2 convolutional layers and 1 pooling layer, the last three convolutional blocks contain 3 convolutional layers and 1 pooling layer, and finally they are connected through 3 fully connected layers.

[0098] In some embodiments, during the training process, the VGG network adopts the stochastic gradient descent algorithm, with an initial learning rate of 0.01, a momentum of 0.9, a weight decay coefficient of 1e-4, and the number of training batches is 50.

[0099] In some embodiments, during the retraining of the SVM network, the feature data obtained after extracting features by the convolutional blocks of the VGG network is input into the SVM network for a new round of feature extraction and training. By adjusting the penalty coefficient C of the SVM, the optimal solution of the support vector is found to obtain an integrated fault diagnosis model.

[0100] In some embodiments, after obtaining the integrated fault diagnosis model, it further includes:

[0101] Randomly extract multiple groups of data from the multi-source raw data as the test set;

[0102] The test set is input into the trained fault diagnosis model, and the fault diagnosis model is tested through the test results.

[0103] In the above specific embodiments, the rotor fault diagnosis device for a rotating device provided by the present invention obtains the original vibration data of the device to be diagnosed, processes the original vibration data and performs feature fusion to obtain fusion data; inputs the fusion data into a pre-trained fault diagnosis model, and the diagnosis result output by the fault diagnosis model can be obtained; wherein, the fault diagnosis model is based on an integrated model algorithm of VGG and SVM, combines the algorithm advantages, utilizes the excellent feature extraction ability of the VGG network and the powerful non-linear classification ability of SVM, and after two feature extractions, the risks of gradient explosion and overfitting can be effectively prevented during the model construction and testing processes, the fault features are more obvious, and the diagnosis accuracy is higher, thereby solving the defects of lagging and poor accuracy in rotor fault diagnosis in the prior art, and realizing the timely and accurate identification of rotor faults in rotating devices.

[0104] Figure 11 An example of a schematic physical structure diagram of an electronic device is shown as Figure 11 shown, the electronic device may include: a processor 1110, a communication interface 1120, a memory 1130, and a communication bus 1140. Among them, the processor 1110, the communication interface 1120, and the memory 1130 communicate with each other through the communication bus 1140. The processor 1110 can call the logical instructions in the memory 1130 to execute the above method.

[0105] In addition, when the logical instructions in the above-mentioned memory 1130 are implemented in the form of a software functional unit and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical disks, and other various media that can store program codes.

[0106] On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the above method.

[0107] On the other hand, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above method is implemented.

[0108] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative work.

[0109] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rotor fault diagnosis method for rotating equipment, characterized in that: The method comprises: Acquire original vibration data of the device to be diagnosed, and perform data processing and feature fusion on the original vibration data to obtain fused data; Inputting the fused data into a pre-trained fault diagnosis model to obtain a diagnosis result output by the fault diagnosis model; Among them, the fault diagnosis model is obtained by training a pre-built VGG-SVM model using sample data and corresponding diagnosis result labels, the sample data is fused data obtained after fusing multi-source original data, and the sample data is multi-source original data obtained at multiple target sampling points of rotating equipment samples.

2. The rotor fault diagnosis method for rotating equipment according to claim 1, characterized in that: The raw vibration data is processed and feature-fused to obtain fused data, specifically including: The original vibration data is sequentially subjected to data division, data cleaning, normalization and data enhancement to obtain fused data of different fault types with consistent data sizes.

3. The rotor fault diagnosis method for rotating equipment according to claim 1, characterized in that: The pre-built VGG-SVM model is trained using the fused sample data and the corresponding diagnosis result labels to obtain the fault diagnosis model, specifically including: Multi-source raw data obtained at multiple target sampling points of the rotating device sample, and the sample data obtained by fusing the multi-source raw data; Constructing a sample data set using paired sample data and corresponding diagnostic result labels, and dividing the sample data set into a training set and a validation set; Using the sample data in the training set to train the pre-built VGG network to obtain an initial model; The pre-built SVM network is trained twice using feature data obtained after extracting features from several blocks in the initial model to obtain an integrated fault diagnosis model.

4. The rotor fault diagnosis method for rotating equipment according to claim 3, characterized in that: The network architecture of the VGG network specifically includes 5 convolution blocks, wherein the first two convolution blocks include 2 convolution layers and 1 pooling layer, and the last three convolution blocks include 3 convolution layers and 1 pooling layer, and are finally connected through 3 fully connected layers.

5. The rotor fault diagnosis method for rotating equipment according to claim 4, characterized in that: During the training process, the VGG network adopts the stochastic gradient descent algorithm, with an initial learning rate of 0.01, a momentum of 0.9, a weight decay coefficient of 1e-4, and 50 training batches.

6. The rotor fault diagnosis method for rotating equipment according to claim 4, characterized in that: During the secondary training of the SVM network, feature data is obtained by extracting features through the convolution block of the VGG network, and input into the SVM network for a new round of feature extraction and training. By adjusting the penalty coefficient C of the SVM, the optimal solution of the support vector is found to obtain an integrated fault diagnosis model.

7. The rotor fault diagnosis method for rotating equipment according to claim 3, characterized in that: The integrated fault diagnosis model is obtained, which also includes: Randomly extracting multiple groups of data from the multi-source original data as test sets; The test set is input into the trained fault diagnosis model, and the fault diagnosis model is tested through the test results.

8. A rotor fault diagnosis device for rotating equipment, characterized in that: The device comprises: A data acquisition unit, used to acquire original vibration data of the device to be diagnosed, and perform data processing and feature fusion on the original vibration data to obtain fused data; A result generating unit, used for inputting the fusion data into a pre-trained fault diagnosis model to obtain a diagnosis result output by the fault diagnosis model; Among them, the fault diagnosis model is obtained by training a pre-built VGG-SVM model using sample data and corresponding diagnosis result labels, the sample data is fused data obtained after fusing multi-source original data, and the sample data is multi-source original data obtained at multiple target sampling points of rotating equipment samples.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the rotor fault diagnosis method for rotating equipment as described in any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the rotor fault diagnosis method for rotating equipment as claimed in any one of claims 1 to 7 is implemented.

Images