Steam turbine fault diagnosis method based on multi-channel time-frequency spectrum cloud model fusion

Abstract

The invention belongs to the field of fault diagnosis, and particularly relates to a steam turbine fault diagnosis method based on multi-channel time-frequency spectrum cloud model fusion, and the method comprises the steps: carrying out the time-frequency transformation through multi-channel sensor data, and obtaining a plurality of time-frequency spectrums; then, dividing the cloud model into a plurality of sub-regions along the frequency direction, overlapping the sub-regions, and calculating feature parameters of the cloud model of each sub-region, namely a mean value, entropy and hyper-entropy, so as to form three matrixes; three convolutional neural networks are established respectively, the matrixes serve as input of the convolutional neural networks respectively, decision layer weighted fusion is carried out on output, a known state data set is used for training a model, and the output after model fusion is the fault type of the steam turbine; three feature matrixes are calculated according to the above steps for the to-be-tested monitoring data, and the three feature matrixes are input into the trained neural network model to diagnose the fault type of the steam turbine. The fault diagnosis precision can be effectively improved, early warning is provided, and safe and stable operation of the steam turbine is ensured.

Description

Technical Field

[0001] The present invention belongs to the field of fault diagnosis, and in particular relates to a steam turbine fault diagnosis method based on multi-channel time-frequency spectrum cloud model fusion. Background Art

[0002] As a rotating mechanical device that converts the kinetic energy and thermal energy of fluid into mechanical energy, steam turbines are key rotating mechanical devices and are widely used in power generation, aviation, shipbuilding, industrial manufacturing and other fields. Their operating status directly affects the safety and production efficiency of the equipment. The core components of steam turbines include rotors, blades, bearings and casings. They run at extremely high speeds and are subjected to complex mechanical stresses. Long-term operation may lead to faults such as component wear and abnormal vibration. Therefore, the operating status monitoring and fault diagnosis of steam turbines are crucial to improving equipment safety and extending their life. Traditional fault diagnosis methods, such as spectrum analysis and time domain feature analysis, usually rely on single-channel sensor data, which is difficult to fully reflect the health status of the equipment under complex operating conditions. In addition, with the complexity of the operating environment and working conditions of steam turbines, fault signals show nonlinear and non-stationary characteristics, which further increases the difficulty of diagnosis. A single method cannot effectively capture the hidden fault characteristics in multi-channel signals, resulting in reduced accuracy and reliability of the diagnosis results.

[0003] In the prior art, for example, there is a Chinese invention patent application with the patent publication number CN118111709A and the name "A method for extracting main bearing fault features by weighted fusion of multi-channel vibration information", which discloses the following content: The present invention provides a method for extracting main bearing fault features by weighted fusion of multi-channel vibration information, which relates to the technical field of bearing fault diagnosis. First, the data measured by multiple vibration sensors are weighted and fused into a one-dimensional signal through the established multi-channel vibration information weighted fusion model, and then the fused signal is decomposed by CEEMDAN, and then the strong impact component is screened out according to the kurtosis-correlation coefficient screening criterion for reconstruction to obtain a vibration signal rich in bearing fault feature information, and finally the weak characteristics of the bearing fault are proposed through the envelope spectrum. The above methods all give the specific structure of the neural network related algorithm. Conventional CNN uses time-frequency diagrams as training data, and the network model is complex. Summary of the invention

[0004] In view of the above problems existing in the prior art, the present invention proposes a turbine fault diagnosis method based on multi-channel time-frequency spectrum cloud model fusion.

[0005] In order to achieve the above effects, the technical solution adopted by the present invention is as follows:

[0006] The steam turbine fault diagnosis method based on multi-channel time-frequency spectrum cloud model fusion includes the following steps:

[0007] Multi-channel sensor data is used for time-frequency transformation to obtain multiple time-frequency spectra; then, it is divided into multiple sub-areas along the frequency direction, and the sub-areas are overlapped. The characteristic parameters of the cloud model of each sub-area are calculated: mean, entropy and super entropy, and three matrices are formed; three convolutional neural networks are established respectively, and the three matrices are used as the input of the three convolutional neural networks respectively, and the outputs of the three networks are weighted fused at the decision layer, and the model is trained using a known state data set. The output of the fused model is the fault type of the turbine; finally, the three characteristic matrices are calculated according to the above steps for the monitoring data to be tested, and the three characteristic matrices are input into the trained neural network model to diagnose the fault type of the turbine.

[0008] Furthermore, the specific steps are:

[0009] Step 1). Calculation of time-frequency spectrum of training data and test data:

[0010] First, the training data and test data are transformed into time-frequency data. The time-frequency transformation includes short-time Fourier transform, wavelet transform or Hilbert-Huang transform to obtain the time-frequency spectrum H of the signal. i (f, t), where i represents different channel sensors, i = 1, 2, ... Nc, Nc is the number of sensor channels, f represents frequency information, and t represents time information; the time-frequency spectrum H i (f, t) is a two-dimensional matrix. When calculating the time-frequency spectrum of each channel, ensure that the time-frequency spectrum H obtained by each channel i The (f, t) matrix dimensions remain consistent;

[0011] Furthermore, the time-frequency matrix H i (f, t) has a dimension of M×N, and the elements in the matrix are h i (f m ,t n ), then the frequency f m is an element in the set F0 = {Δf, 2·Δf, 3·Δf, …, M·Δf}, m = 1, 2, 3…, M, M·Δf is the time-frequency spectrum H i The highest frequency component of (f,t), t n is an element in the set T0 = {Δt, 2·Δt, 3·Δt, …, N·Δt}, n = 1, 2, 3…, N, Δf is the time-frequency spectrum H i The frequency resolution of (f, t), Δt is the time spectrum H i The temporal resolution of (f,t), h i (f m ,t n ) is the frequency f m , time t n The corresponding time-frequency spectrum amplitude under these conditions.

[0012] Step 2). Time-frequency spectrum frequency sub-area division of training data and test data:

[0013] The training data and test data after time-frequency transformation are divided into multiple sub-regions along the frequency direction, and there is overlap between adjacent sub-regions:

[0014] Furthermore, for each channel’s time-frequency spectrum H i (f, t) is divided along the frequency direction, and a fixed overlap ratio α is set to form multiple sub-areas j = 1, 2, 3, ... K, K is the number of sub-regions; the difference between the upper and lower frequency boundaries of each frequency region, that is, the frequency width of the sub-region, is L f , then the upper and lower bounds of each sub-region are as follows:

[0015] The first sub-area The frequency f1 value range LF1 is: LF1∈[0,L f ];

[0016] The second sub-area The frequency f2 value range LF2 is: LF2∈[L f -αL f ,2L f -αL f ];

[0017] The jth sub-region The frequency f j Value range LF j For: LF j ∈[(j-1)L f -αL f ,jL f -αL f ];

[0018] Kth sub-region The frequency f K Value range LF K For: LF K ∈[(K-1)L f -αL f ,KL f -αL f ];

[0019] Among them, K is the KL f -αL f ≤MΔf is the maximum positive integer; the overlap ratio α satisfies 0≤α≤0.5. When α is equal to 0, the frequency regions do not overlap. When α is equal to 0.5, the overlapping part of adjacent frequency regions is half of the length of the frequency region; MΔf is the maximum value of the time-spectrum frequency f.

[0020] Time spectrum of sub-areas after dividing the area The dimension is P×N, P is the value range Lf j The number of elements in the intersection with set F0; The elements in the matrix are For frequency Time t n The corresponding time-frequency spectrum amplitude under the conditions; From the value range Lf j The value is taken from the intersection with the set F0, where p = 1, 2, 3, ... P, time spectrum The time dimension of the matrix does not change, and the value range of t does not change, and it remains an element in the T0 set;

[0021] Step 3). Calculation of sub-regional time-frequency spectrum cloud model characteristics:

[0022] Sub-region time spectrum All elements in the matrix Expand to a one-dimensional vector x i,j ,but Where, q = 1, 2, 3, ... Q, Q = P × N, The sub-region time spectrum The three cloud model features of the matrix are:

[0023] Sub-region time spectrum The cloud model expects E i,j

[0024]

[0025] Sub-region time spectrum The cloud model characteristic entropy H i,j

[0026]

[0027] Sub-region time spectrum The cloud model of super entropy He i,j

[0028]

[0029] Among them, S j For sample The second-order center distance is calculated as follows:

[0030]

[0031] Through the above calculations, a set of signals in each channel can obtain three cloud model parameter vectors E i ,H i ,He i ,in, Then a set of multi-sensor data samples can be calculated to obtain a set of cloud model matrices, as follows:

[0032]

[0033] Where Nc is the number of sensor channels;

[0034] At this point, each set of multi-sensor data can obtain a set of cloud model matrices, which are the expected E t , Entropy H t and super entropy He t .

[0035] Step 4) Convolutional neural network fusion model structure construction;

[0036] Convolutional neural network fusion model is constructed for fault diagnosis. Three convolutional neural networks with the same structure are constructed respectively. The input of convolutional neural network model is E t , Entropy H t and super entropy He t The matrix is ​​constructed, and the output layers of the three convolutional neural networks are weightedly fused as the output of the entire convolutional neural network fusion model.

[0037] Furthermore, the three convolutional neural networks have the same structure, which is a 2D convolutional neural network, including an input layer, an intermediate layer, and an output layer. The input layer of the model is the expected E t , Entropy H t and super entropy He t Matrix, the input layer dimension is Nc×K, Nc is the number of signal channels, K is the number of frequency sub-regions divided by the time-frequency spectrum; the intermediate layer contains several convolutional layers, activation functions, normalization layers and pooling layers, etc. The output layer contains a fully connected layer and a softmax layer. After several convolutional kernel pooling layers, low-dimensional feature information is obtained. The network nodes of the fully connected layer realize the mapping of feature information to classification labels. The output expression of the fully connected layer is as follows:

[0038]

[0039] where z i is the one-dimensional output of the fully connected layer, l corresponds to the number of target categories of the network, a j is the element in the vector after the one-dimensional expansion of the final feature layer matrix, n is a j The number of i,j is a fully connected j The corresponding connection weight, b i is the offset.

[0040] The output value of each neuron in the fully connected layer is passed to a classification decision layer that outputs a probability distribution. The softmax logistic regression function is used to classify the results, and its probability output p(z i ) is calculated as follows:

[0041]

[0042] The probability output of each convolutional neural network output layer is obtained through the above softmax; the input matrix is ​​E t The probability output of the convolutional neural network is recorded as y E (i), the input matrix is ​​H t The probability output of the convolutional neural network is denoted as y H (i), the input matrix is ​​He t The probability output of the convolutional neural network is recorded as y He (i), where i = 1, 2, ..., l, and l corresponds to the number of target categories in the network.

[0043] The last layer of the convolutional neural network fusion model is the fusion layer, which is responsible for weighted fusion of the softmax layer outputs of the three neural networks. The number of neurons in the output layer of the convolutional neural network fusion model is equal to the number of target categories. Figure 4 As shown, the fusion can obtain the output vector of the convolutional neural network fusion model The specific elements in the vector are calculated as follows:

[0044]

[0045] in, It is connected to the first CNN network (the input is the expected matrix E t )The coefficients of the i-th neuron in the output layer and the i-th neuron in the fusion layer; It is connected to the second CNN network (the input is the entropy matrix H t )The coefficients of the i-th neuron in the output layer and the i-th neuron in the fusion layer; It is connected to the third CNN network (the input is the entropy matrix H t )The coefficients of the i-th neuron in the output layer and the i-th neuron in the fusion layer;

[0046] The weight matrix W of the fusion layer can be obtained F as follows:

[0047]

[0048] in,

[0049] Step 5). Model training:

[0050] The model training process uses the feature set of known state data and the corresponding state output to train the model, which is used to determine the parameters of each node of the three convolutional neural network models in the convolutional neural network fusion model and the W of the fusion layer of the fusion model. F Matrix parameters.

[0051] According to step 1) to step 3), the expectation, entropy and super entropy matrices of each group of multi-channel data are calculated. The number of samples is Nt, the sample index is t, then t = 1, 2, 3, ... Nt; the matrices extracted from each group of multi-channel data are combined as shown below to form a training data set:

[0052] E train ={E1,E1,…,E t ,…,E Nt}, (12)

[0053] H train ={H1,H2,…,H t ,…,H Nt}, (13)

[0054] He train ={He1,He2,…,He t ,…,He Nt}, (14)

[0055] Among them, E train is the set of expected matrices extracted from the training data, H train is the set of entropy matrices extracted from the training data, He train is the set of super entropy matrices extracted from the training data, and the dataset E train As the first CNN training sample set, the dataset H train As the second CNN training sample set, He train As the training sample set of the third CNN, before training, the known data is labeled. Assuming that the t-th sample is in state i, the expected output value of the convolutional neural network fusion model corresponding to the t-th sample is defined as follows:

[0056]

[0057] The expected output The only element in YE t (i) is 1, and the rest are 0, that is, for the i-th type of data, only the i-th element in the corresponding expected state output vector is 1, and the rest are 0.

[0058] During the training process, the weighted fusion matrix W of the fusion layer F To limit, W F The vectors in satisfy Right now

[0059]

[0060] in is a unit vector. During the training process, for the t-th sample known to be in state i, it is assumed that the output of its convolutional neural network fusion model is YT t =[YT t (1),YT t (2),…YT t (i), ... YT t (l)], the training process aims to determine the network parameters through training so that the optimization objective function between the output of the overall model after training and the known state is minimized, and then determine the network structure parameters. During the training process, the optimization objective function of the training process is set to the mean square error function, and the mean square error objective function is as follows:

[0061]

[0062] Step 6). Use of the model.

[0063] When using the model, the cloud model feature matrix should be calculated according to steps 1) to 3) for a set of multi-channel signals used for testing, and a set of mean, entropy, and super entropy matrices as shown in equations (5), (6), and (7) are obtained, which are respectively recorded as expected E test , Entropy H test and super entropy He test Matrix, E test Connected to the first CNN input layer, H test Connected to the second CNN input layer, He test Connected to the third CNN input, the model will output For model output The maximum value is selected to be 1, and the other values ​​are set to 0 to generate a new output result. The judgment method of the state category corresponding to the output result is consistent with the definition of the expected output of the model of the known category training data, that is, the index i where the maximum value is located is recorded as the fault classification result.

[0064] The advantages of this application are:

[0065] 1. This application adopts multi-channel sensor fusion, multi-feature extraction and cloud model-based fault diagnosis methods, which can effectively improve the accuracy of fault diagnosis, provide early warning, and ensure the safe and stable operation of the turbine.

[0066] 2. The present application provides a multi-channel fusion fault diagnosis method, which uniformly converts the time-frequency spectrum of each channel into a cloud model expression, and the statistical parameters of the cloud model under the same state will be relatively close.

[0067] 3. The present application can improve fault sensitivity. The statistical characteristics of the same frequency sub-regions of the same state data will also be similar after the region is divided; among them, the sub-frequency region division of the time-frequency spectrum is intended to count the distribution of the amplitude cloud model of each region during the training process. If a device fails, the cloud model statistics in the corresponding frequency domain will change. The purpose of dividing the region is to subdivide the sensitive area. If no division is performed, the statistical characteristics of the cloud model are calculated for the entire time-frequency spectrum. When a fault occurs, the impact of the fault on the statistical characteristics of the overall time-frequency spectrum cloud model will be very weak. Dividing the sub-region can effectively improve the sensitivity to fault information. If the partial overlapping operation of the sub-frequency region is not adopted, when the area affected by the fault on the time-frequency spectrum falls between two sub-regions, the fault information will be split. The overlapping operation of the frequency sub-regions will further improve the coverage of the fault information and improve the accuracy of fault classification.

[0068] 4. This application can reduce the complexity of CNN network structure. The commonly used two-dimensional CNN training input directly uses the time-frequency spectrum matrix data as input, while the present invention performs sub-region division while counting the sub-region cloud model features, and uses three types of cloud model features as model input, which greatly reduces the complexity of the model structure and improves the model training efficiency and execution speed.

[0069] 5. This application can achieve dual fusion of feature layer and decision layer. The three parameter matrices of the cloud model of multi-channel sensor data are used as CNN inputs respectively. Each feature matrix contains the time-frequency spectrum cloud model features of the multi-channel sensor. Channel fusion is completed in the feature calculation process. The results of each CNN output layer are weighted and fused to complete the fusion of the output layer. This is a method of dual fusion of feature layer and decision layer. Therefore, this method is innovative. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 This is a flow chart of a steam turbine fault diagnosis method based on multi-channel time-frequency spectrum cloud model fusion.

[0071] Figure 2 Schematic diagram of overlapping frequency division of sub-areas.

[0072] Figure 3 A schematic diagram of the network structure.

[0073] Figure 4 Schematic diagram of convolutional neural network fusion. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0075] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0076] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0077] In the description of the present application, it should be noted that the terms "upper", "vertical", "inner", "outer", etc. indicate the orientation or position relationship based on the orientation or position relationship shown in the drawings, or the orientation or position relationship in which the product of the application is usually placed when in use, or the orientation or position relationship commonly understood by those skilled in the art. They are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0078] In the description of this application, it should also be noted that, unless otherwise clearly specified and limited, the terms "set", "install", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0079] Example 1

[0080] The steam turbine fault diagnosis method based on multi-channel time-frequency spectrum cloud model fusion includes the following steps: using multi-channel sensor data to perform time-frequency transformation to obtain multiple time-frequency spectra; then, dividing it into multiple sub-regions along the frequency direction, overlapping the sub-regions, calculating the characteristic parameters of the cloud model of each sub-region: mean, entropy and hyperentropy, and forming three matrices; establishing three convolutional neural networks respectively, using the three matrices as the input of the three convolutional neural networks respectively, performing decision-layer weighted fusion on the outputs of the three networks, using a known state data set to train the model, and the output of the fused model is the fault type of the steam turbine; finally, calculating three characteristic matrices according to the above steps for the monitoring data to be tested, and inputting the three characteristic matrices into the trained neural network model to diagnose the fault type of the steam turbine.

[0081] Example 2

[0082] The specific steps of the steam turbine fault diagnosis method based on multi-channel time-frequency spectrum cloud model fusion are as follows:

[0083] Step 1). Calculation of time-frequency spectrum of training data and test data:

[0084] First, the training data and test data are transformed into time-frequency data. The time-frequency transformation includes short-time Fourier transform, wavelet transform or Hilbert-Huang transform to obtain the time-frequency spectrum H of the signal. i (f, t), where i represents different channel sensors, i = 1, 2, ... Nc, Nc is the number of sensor channels, f represents frequency information, and t represents time information; the time-frequency spectrum H i (f, t) is a two-dimensional matrix. When calculating the time-frequency spectrum of each channel, ensure that the time-frequency spectrum H obtained by each channel i The (f, t) matrix dimensions remain consistent;

[0085] Step 2). Time-frequency spectrum frequency sub-area division of training data and test data:

[0086] The training data and test data after time-frequency transformation are divided into multiple sub-regions along the frequency direction so that there is overlap between adjacent sub-regions;

[0087] For each channel’s time-frequency spectrum H i (f, t) is divided along the frequency direction, and a fixed overlap ratio α is set to form multiple sub-areas , j = 1, 2, 3, ... K, K is the number of sub-regions; the difference between the upper and lower frequency limits of each frequency region, that is, the frequency width of the sub-region, is L f , then the upper and lower bounds of each sub-region are as follows:

[0088] The first sub-area The frequency f1 value range LF1 is: LF1∈[0,L f ];

[0089] The second sub-area The frequency f2 value range LF2 is: LF2∈[L f -αL f ,2L f -αL f ];

[0090] The jth sub-region The frequency f j Value range LF j For: LF j ∈[(j-1)L f -αL f ,jL f -αL f ];

[0091] Kth sub-region The frequency f K Value range LF K For: LF K ∈[(K-1)L f -αL f ,KL f -αL f ];

[0092] Among them, K is the KL f -αL f ≤MΔf is the maximum positive integer; the overlap ratio α satisfies 0≤α≤0.5. When α is equal to 0, the frequency regions do not overlap. When α is equal to 0.5, the overlapping part of adjacent frequency regions is half of the length of the frequency region. MΔf is the maximum value of the time-spectrum frequency f.

[0093] Step 3). Calculation of sub-regional time-frequency spectrum cloud model characteristics:

[0094] Sub-region time spectrum All elements in the matrix Expand to a one-dimensional vector x i,j ,but Where, q = 1, 2, 3, ... Q, Q = P × N, The sub-region time spectrum The three cloud model features of the matrix are:

[0095] Sub-region time spectrum The cloud model expects E i,j

[0096]

[0097] Sub-region time spectrum The cloud model characteristic entropy H i,j

[0098]

[0099] Sub-region time spectrum The cloud model of super entropy He i,j

[0100]

[0101] Among them, S j For sample The second-order center distance is calculated as follows:

[0102]

[0103] Through the above calculations, a set of signals for each channel can obtain three cloud model parameter vectors E i ,H i ,He i ,in, Then a set of multi-sensor data samples can be calculated to obtain a set of cloud model matrices, as follows:

[0104]

[0105] Where Nc is the number of sensor channels;

[0106] At this point, each set of multi-sensor data can obtain a set of cloud model matrices, which are the expected E t , Entropy H t and super entropy He t ;

[0107] Step 4) Convolutional neural network fusion model structure construction;

[0108] Convolutional neural network fusion model is constructed for fault diagnosis. Three convolutional neural networks with the same structure are constructed respectively. The input of convolutional neural network model is E t , Entropy H t and super entropy He t Matrix, and weighted fusion of the output layers of the three convolutional neural networks as the output of the entire convolutional neural network fusion model;

[0109] Step 5). Model training:

[0110] The model training process uses the feature set of known state data and the corresponding state output to train the model, which is used to determine the parameters of each node of the three convolutional neural network models in the convolutional neural network fusion model and the W of the fusion layer of the fusion model. F Matrix parameters;

[0111] Step 6). Use of the model.

[0112] Example 3

[0113] The steam turbine fault diagnosis method based on multi-channel time-frequency spectrum cloud model fusion includes the following steps:

[0114] Multi-channel sensor data is used for time-frequency transformation to obtain multiple time-frequency spectra; then, it is divided into multiple sub-areas along the frequency direction, and the sub-areas are overlapped. The characteristic parameters of the cloud model of each sub-area are calculated: mean, entropy and super entropy, and three matrices are formed; three convolutional neural networks are established respectively, and the three matrices are used as the input of the three convolutional neural networks respectively, and the outputs of the three networks are weighted fused at the decision layer, and the model is trained using a known state data set. The output of the fused model is the fault type of the turbine; finally, the three characteristic matrices are calculated according to the above steps for the monitoring data to be tested, and the three characteristic matrices are input into the trained neural network model to diagnose the fault type of the turbine.

[0115] The specific steps are:

[0116] Step 1). Calculation of time-frequency spectrum of training data and test data:

[0117] First, the training data and test data are transformed in time-frequency, and the time-frequency transformation includes but is not limited to short-time Fourier transform, wavelet transform or Hilbert-Huang transform, etc., to obtain the time-frequency spectrum H of the signal i (f, t), where i represents different channel sensors, i = 1, 2, ... Nc, Nc is the number of sensor channels, f represents frequency information, and t represents time information; the time-frequency spectrum H i (f, t) is a two-dimensional matrix. When calculating the time-frequency spectrum of each channel, ensure that the time-frequency spectrum H obtained by each channel i The (f, t) matrix dimensions remain consistent;

[0118] Assume that the time-frequency matrix H i (f, t) has a dimension of M×N, and the elements in the matrix are h i (f m ,t n ), then the frequency f m is an element in the set F0 = {Δf, 2·Δf, 3·Δf, …, M·Δf}, m = 1, 2, 3…, M, M·Δf is the time-frequency spectrum H i The highest frequency component of (f,t), t n is an element in the set T0 = {Δt, 2·Δt, 3·Δt, …, N·Δt}, n = 1, 2, 3…, N, Δf is the time-frequency spectrum H i The frequency resolution of (f, t), Δt is the time spectrum H i The temporal resolution of (f,t), h i (fm ,t n ) is the frequency f m , time t n The corresponding time-frequency spectrum amplitude under these conditions.

[0119] Step 2). Time-frequency spectrum frequency sub-area division of training data and test data:

[0120] The training data and test data after time-frequency transformation are divided into multiple sub-regions along the frequency direction, and there is a certain overlap between adjacent sub-regions:

[0121] like Figure 2 As shown, for each channel, the time spectrum H i (f, t) is divided along the frequency direction, and a fixed overlap ratio α is set to form multiple sub-areas j = 1, 2, 3, ... K, K is the number of sub-regions; Assume that the difference between the upper and lower frequency limits of each frequency region, that is, the frequency width of the sub-region, is L f , then the upper and lower bounds of each sub-region are as follows:

[0122] The first sub-area The frequency f1 value range LF1 is: LF1∈[0,L f ];

[0123] The second sub-area The frequency f2 value range LF2 is: LF2∈[L f -αL f ,2L f -αL f ];

[0124] The jth sub-region The frequency f j Value range LF j For: LF j ∈[(j-1)L f -αL f ,jL f -αL f ];

[0125] Kth sub-region The frequency f K Value range LF K For: LF K ∈[(K-1)L f -αL f ,KL f -αL f ];

[0126] Among them, K is the KL f -αL f≤MΔf is the maximum positive integer; the overlap ratio α satisfies 0≤α≤0.5. When α is equal to 0, the frequency regions do not overlap. When α is equal to 0.5, the overlapping part of adjacent frequency regions is half of the length of the frequency region; MΔf is the maximum value of the time-spectrum frequency f.

[0127] Time spectrum of sub-areas after dividing the area The dimension is P×N, P is the value range Lf j The number of elements in the intersection with the set F0; assuming The elements in the matrix are For frequency Time t n The corresponding time-frequency spectrum amplitude under the conditions; From the value range Lf j The value is taken from the intersection with the set F0, where p = 1, 2, 3, ... P, time spectrum The time dimension of the matrix does not change, and the value range of t does not change, and it remains an element in the T0 set;

[0128] Step 3). Calculation of sub-regional time-frequency spectrum cloud model characteristics:

[0129] Sub-region time spectrum All elements in the matrix Expand to a one-dimensional vector x i,j ,but Where, q = 1, 2, 3, ... Q, Q = P × N, The sub-region time spectrum The three cloud model features of the matrix are:

[0130] Sub-region time spectrum The cloud model expects E i,j

[0131]

[0132] Sub-region time spectrum The cloud model characteristic entropy H i,j

[0133]

[0134] Sub-region time spectrum The cloud model of super entropy He i,j

[0135]

[0136] Among them, S j For sample The second-order center distance is calculated as follows:

[0137]

[0138] Through the above calculations, a set of signals in each channel can obtain three cloud model parameter vectors E i ,H i ,He i ,in, Then a set of multi-sensor data samples can be calculated to obtain a set of cloud model matrices, as follows:

[0139]

[0140] Where Nc is the number of sensor channels;

[0141] At this point, each set of multi-sensor data can obtain a set of cloud model matrices, which are the expected E t , Entropy H t and super entropy He t .

[0142] Step 4) Convolutional neural network fusion model structure construction;

[0143] like Figure 3 As shown in the figure, a convolutional neural network fusion model is constructed for fault diagnosis. Three convolutional neural networks with the same structure are constructed respectively. The input of the convolutional neural network model is E t , Entropy H t and super entropy He t The matrix is ​​constructed, and the output layers of the three convolutional neural networks are weightedly fused as the output of the entire convolutional neural network fusion model.

[0144] The three convolutional neural networks have the same structure, which is a 2D convolutional neural network, including an input layer, an intermediate layer, and an output layer. The input layer of the model is the expected E t , Entropy H t and super entropy He t Matrix, the input layer dimension is Nc×K, Nc is the number of signal channels, K is the number of frequency sub-regions divided by the time-frequency spectrum; the intermediate layer can be designed according to the needs to ensure that the three convolutional neural networks have the same structure. Each convolutional neural network contains several convolutional layers, activation functions, normalization layers, and pooling layers. The output layer contains a fully connected layer and a softmax layer. After several convolution kernel pooling layers, low-dimensional feature information is obtained. The network nodes of the fully connected layer realize the mapping of feature information to classification labels. The output expression of the fully connected layer is as follows:

[0145]

[0146] where z i is the one-dimensional output of the fully connected layer, l corresponds to the number of target categories of the network, a jis the element in the vector after the one-dimensional expansion of the final feature layer matrix, n is a j The number of i,j is a fully connected j The corresponding connection weight, b i is the offset.

[0147] The output value of each neuron in the fully connected layer is passed to a classification decision layer that outputs a probability distribution. The softmax logistic regression function is used to classify the results, and its probability output p(z i ) is calculated as follows:

[0148]

[0149] The probability output of each convolutional neural network output layer is obtained through the above softmax; the input matrix is ​​E t The probability output of the convolutional neural network is denoted as y E (i), the input matrix is ​​H t The probability output of the convolutional neural network is recorded as y H (i), the input matrix is ​​He t The probability output of the convolutional neural network is recorded as y He (i), where i = 1, 2, ..., l, and l corresponds to the number of target categories in the network.

[0150] The last layer of the convolutional neural network fusion model is the fusion layer, which is responsible for weighted fusion of the softmax layer outputs of the three neural networks. The number of neurons in the output layer of the convolutional neural network fusion model is equal to the number of target categories. Figure 4 As shown, the fusion can obtain the output vector of the convolutional neural network fusion model The specific elements in the vector are calculated as follows:

[0151]

[0152] in, It is connected to the first CNN network (the input is the expected matrix E t )The coefficients of the i-th neuron in the output layer and the i-th neuron in the fusion layer; It is connected to the second CNN network (the input is the entropy matrix H t )The coefficients of the i-th neuron in the output layer and the i-th neuron in the fusion layer; It is connected to the third CNN network (the input is the entropy matrix H t )The coefficients of the i-th neuron in the output layer and the i-th neuron in the fusion layer;

[0153] The weight matrix W of the fusion layer can be obtained F as follows:

[0154]

[0155] in,

[0156] Step 5). Model training:

[0157] The model training process uses the feature set of known state data and the corresponding state output to train the model, which is used to determine the parameters of each node of the three convolutional neural network models in the convolutional neural network fusion model and the W of the fusion layer of the fusion model. F Matrix parameters.

[0158] According to step 1) to step 3), the expectation, entropy and super entropy matrices of each group of multi-channel data are calculated (as shown in equations (5), (6) and (7)). The number of samples is Nt, the sample index is t, then t = 1, 2, 3, ... Nt; the matrices extracted from each group of multi-channel data are combined as shown below to form a training data set:

[0159] E train ={E1,E1,…,E t ,…,E Nt}, (12)

[0160] H train ={H1,H2,…,H t ,…,H Nt}, (13)

[0161] He train ={He1,He2,…,He t ,…,He Nt}, (14)

[0162] Among them, E train is the set of expected matrices extracted from the training data, H train is the set of entropy matrices extracted from the training data, He train is the set of super entropy matrices extracted from the training data, and the dataset E train As the first CNN training sample set, the dataset H train As the second CNN training sample set, He train As the training sample set of the third CNN, before training, the known data is labeled. Assuming that the t-th sample is in state i, the expected output value of the convolutional neural network fusion model corresponding to the t-th sample is defined as follows:

[0163]

[0164] The expected output The only element in YE t(i) is 1, and the rest are 0, that is, for the i-th category data, only the i-th element in the corresponding expected state output vector is 1, and the rest are 0. For example, for the four-state fault classification diagnosis problem, it is known that the network expected output corresponding to the training data of category 1 is [1 0 0 0], the network expected output corresponding to the training data of category 2 is [0 1 0 0], the network expected output corresponding to the training data of category 3 is [00 1 0], and the network expected output corresponding to the training data of category 4 is [0 0 0 1].

[0165] During the training process, the weighted fusion matrix W of the fusion layer F To limit, W F The vectors in satisfy Right now

[0166]

[0167] in is a unit vector. During the training process, for the t-th sample known to be in state i, it is assumed that the output of its convolutional neural network fusion model is YT t =[YT t (1),YT t (2),…YT t (i), ... YT t (l)], the training process aims to determine the network parameters through training so that the optimization objective function between the output of the overall model after training and the known state is minimized, and then determine the network structure parameters. During the training process, the optimization objective function of the training process is set to the mean square error function, but it is not limited to the mean square error function. The mean square error objective function is as follows:

[0168] Step 6). Use of the model.

[0169] When using the model, the cloud model feature matrix should be calculated according to steps 1) to 3) for a set of multi-channel signals used for testing, and a set of mean, entropy, and super entropy matrices as shown in equations (5), (6), and (7) are obtained, which are respectively recorded as expected E test , Entropy H test and super entropy He test Matrix, E test Connected to the first CNN input layer, H test Connected to the second CNN input layer, He test Connected to the third CNN input, the model will output For model output The maximum value is selected to be 1, and the other values ​​are set to 0 to generate a new output result. The judgment method of the state category corresponding to the output result is consistent with the definition of the expected output of the model of the known category training data, that is, the index i where the maximum value is located is recorded as the fault classification result.

[0170] In summary, after reading the present invention document, ordinary technicians in this field can make various other corresponding transformation schemes based on the technical scheme and technical concept of the present invention without creative mental labor, which all fall within the scope of protection of the present invention.

Claims

1. A steam turbine fault diagnosis method based on multi-channel time-frequency spectrum cloud model fusion, characterized by: The steps include: Multi-channel sensor data is used for time-frequency transformation to obtain multiple time-frequency spectra; then, it is divided into multiple sub-areas along the frequency direction, and the sub-areas are overlapped. The characteristic parameters of the cloud model of each sub-area are calculated: mean, entropy and super entropy, and three matrices are formed; three convolutional neural networks are established respectively, and the three matrices are used as the input of the three convolutional neural networks respectively, and the outputs of the three networks are weighted fused at the decision layer, and the model is trained using a known state data set. The output of the fused model is the fault type of the turbine; finally, the three characteristic matrices are calculated according to the above steps for the monitoring data to be tested, and the three characteristic matrices are input into the trained neural network model to diagnose the fault type of the turbine.

2. The method for steam turbine fault diagnosis based on multi-channel time-frequency spectrum cloud model fusion according to claim 1 is characterized in that: The specific steps are: Step 1). Calculation of time-frequency spectrum of training data and test data: First, the training data and test data are transformed into time-frequency data. The time-frequency transformation includes short-time Fourier transform, wavelet transform or Hilbert-Huang transform to obtain the time-frequency spectrum H of the signal. i (f, t), where i represents different channel sensors, i = 1, 2, ... Nc, Nc is the number of sensor channels, f represents frequency information, and t represents time information; the time-frequency spectrum H i (f, t) is a two-dimensional matrix. When calculating the time-frequency spectrum of each channel, ensure that the time-frequency spectrum H obtained by each channel i The (f, t) matrix dimensions remain consistent; Step 2). Time-frequency spectrum frequency sub-area division of training data and test data: The training data and test data after time-frequency transformation are divided into multiple sub-regions along the frequency direction so that there is overlap between adjacent sub-regions; For each channel’s time-frequency spectrum H i (f, t) is divided along the frequency direction, and a fixed overlap ratio α is set to form multiple sub-areas K is the number of sub-regions; the difference between the upper and lower frequency limits of each frequency region, i.e. the frequency width of the sub-region, is L f , then the upper and lower bounds of each sub-region are as follows: The first sub-area The frequency f1 value range LF1 is: LF1∈[0,L f ]; The second sub-area The frequency f2 value range LF2 is: LF2∈[L f -αL f ,2L f -αL f ]; The jth sub-region The frequency f j Value range LF j For: LF j ∈[(j-1)L f -αL f ,jL f -αL f ]; Kth sub-region The frequency f K Value range LF K For: LF K ∈[(K-1)L f -αL f ,KL f -αL f ]; Among them, K is the KL f -αL f ≤MΔf is the maximum positive integer; the overlap ratio α satisfies 0≤α≤0.

5. When α is equal to 0, the frequency regions do not overlap. When α is equal to 0.5, the overlapping part of adjacent frequency regions is half of the length of the frequency region. MΔf is the maximum value of the time-spectrum frequency f. Step 3). Calculation of sub-regional time-frequency spectrum cloud model characteristics: Sub-region time spectrum All elements in the matrix Expand to a one-dimensional vector x i,j ,but Where, q = 1, 2, 3, ... Q, Q = P × N, The sub-region time spectrum The three cloud model features of the matrix are: Sub-region time spectrum The cloud model expects E i,j Sub-region time spectrum The cloud model characteristic entropy H i,j Sub-region time spectrum The cloud model of super entropy He i,j Among them, S j For sample The second-order center distance is calculated as follows: Through the above calculations, a set of signals for each channel can obtain three cloud model parameter vectors E i ,H i ,He i ,in, Then a set of multi-sensor data samples can be calculated to obtain a set of cloud model matrices, as follows: Where Nc is the number of sensor channels; At this point, each set of multi-sensor data can obtain a set of cloud model matrices, which are the expected E t , Entropy H t and super entropy He t ; Step 4) Convolutional neural network fusion model structure construction; Convolutional neural network fusion model is constructed for fault diagnosis. Three convolutional neural networks with the same structure are constructed respectively. The input of convolutional neural network model is E t , Entropy H t and super entropy He t Matrix, and weighted fusion of the output layers of the three convolutional neural networks as the output of the entire convolutional neural network fusion model; Step 5). Model training: The model training process uses the feature set of known state data and the corresponding state output to train the model, which is used to determine the parameters of each node of the three convolutional neural network models in the convolutional neural network fusion model and the W of the fusion layer of the fusion model. F Matrix parameters; Step 6). Use of the model.

3. The method for steam turbine fault diagnosis based on multi-channel time-frequency spectrum cloud model fusion according to claim 2 is characterized in that: In step 1), the time-frequency matrix H i (f, t) has a dimension of M×N, and the elements in the matrix are h i (f m ,t n ), then the frequency f m is an element in the set F0 = {Δf, 2·Δf, 3·Δf, …, M·Δf}, m = 1, 2, 3…, M, M·Δf is the time-frequency spectrum H i The highest frequency component of (f,t), t n is an element in the set T0 = {Δt, 2·Δt, 3·Δt, …, N·Δt}, n = 1, 2, 3…, N, Δf is the time-frequency spectrum H i The frequency resolution of (f, t), Δt is the time spectrum H i The temporal resolution of (f,t), h i (f m ,t n ) is the frequency f m , time t n The corresponding time-frequency spectrum amplitude under these conditions.

4. The method for steam turbine fault diagnosis based on multi-channel time-frequency spectrum cloud model fusion according to claim 2 is characterized in that: In step 2), the sub-region time spectrum after the region is divided The dimension is P×N, P is the value range Lf j The number of elements in the intersection with the set F0; assuming The elements in the matrix are For frequency Time t n The corresponding time-frequency spectrum amplitude under the conditions; From the value range Lf j The value is taken from the intersection with the set F0, where p = 1, 2, 3, ... P, time spectrum The time dimension of the matrix does not change, and the value range of t does not change, and it remains an element in the T0 set.

5. The method for steam turbine fault diagnosis based on multi-channel time-frequency spectrum cloud model fusion according to claim 2 is characterized in that: The three convolutional neural networks in step 4) have the same structure, which is a 2D convolutional neural network, including an input layer, an intermediate layer, and an output layer. The input layer of the model is the expected E t , Entropy H t and super entropy He t Matrix, the input layer dimension is Nc×K, Nc is the number of signal channels, K is the number of frequency sub-regions divided by the time-frequency spectrum; the intermediate layer contains several convolutional layers, activation functions, normalization layers and pooling layers, and the output layer contains a fully connected layer and a softmax layer. After several convolutional kernel pooling layers, low-dimensional feature information is obtained. The network nodes of the fully connected layer realize the mapping of feature information to classification labels. The output expression of the fully connected layer is as follows: where z i is the one-dimensional output of the fully connected layer, l corresponds to the number of target categories of the network, a j is the element in the vector after the one-dimensional expansion of the final feature layer matrix, n is a j The number of i,j is a fully connected j The corresponding connection weight, b i is the offset.

6. The method for steam turbine fault diagnosis based on multi-channel time-frequency spectrum cloud model fusion according to claim 5 is characterized in that: The output value of each neuron in the fully connected layer is passed to a classification decision layer that outputs a probability distribution. The softmax logistic regression function is used to classify the results, and its probability output p(z i ) is calculated as follows: The probability output of each convolutional neural network output layer is obtained through the above softmax; the input matrix is ​​E t The probability output of the convolutional neural network is recorded as y E (i), the input matrix is ​​H t The probability output of the convolutional neural network is recorded as y H (i), the input matrix is ​​He t The probability output of the convolutional neural network is recorded as y He (i), where i = 1, 2, ..., l, and l corresponds to the number of target categories in the network.

7. The method for steam turbine fault diagnosis based on multi-channel time-frequency spectrum cloud model fusion according to claim 6 is characterized in that: The last layer of the convolutional neural network fusion model is the fusion layer, which is responsible for weighted fusion of the softmax layer outputs of the three neural networks. The number of neurons in the output layer of the convolutional neural network fusion model is equal to the number of target categories. The fusion can obtain the output vector of the convolutional neural network fusion model. The specific elements in the vector are calculated as follows: in, It is connected to the first CNN network (the input is the expected matrix E t )The coefficients of the i-th neuron in the output layer and the i-th neuron in the fusion layer; It is connected to the second CNN network (the input is the entropy matrix H t )The coefficients of the i-th neuron in the output layer and the i-th neuron in the fusion layer; It is connected to the third CNN network (the input is the entropy matrix H t )The coefficients of the i-th neuron in the output layer and the i-th neuron in the fusion layer; The weight matrix W of the fusion layer can be obtained F as follows: in, 8. The method for steam turbine fault diagnosis based on multi-channel time-frequency spectrum cloud model fusion according to claim 2 is characterized in that: According to step 1) to step 3), the expectation, entropy and super entropy matrices of each group of multi-channel data are calculated. The number of samples is Nt, and the sample index is t, then t = 1, 2, 3, ... Nt; the matrices extracted from each group of multi-channel data are combined as shown below to form a training data set: HAVE BEEN train ={E1,E1,…,E t ,…,HAVE BEEN Nt }, (12) H train ={H1,H2,…,H t ,…,H Nt }, (13) He train ={He1,He2,…,He t ,…,He Nt }, (14) Among them, E train is the set of expected matrices extracted from the training data, H train is the set of entropy matrices extracted from the training data, He train is the set of super entropy matrices extracted from the training data, and the dataset E train As the first CNN training sample set, the dataset H train As the second CNN training sample set, He train As the training sample set of the third CNN, before training, the known data is labeled. Assuming that the t-th sample is in state i, the expected output value of the convolutional neural network fusion model corresponding to the t-th sample is defined as follows: The expected output The only element in YE t (i) is 1, and the rest are 0, that is, for the i-th type of data, only the i-th element in the corresponding expected state output vector is 1, and the rest are 0.

9. The method for steam turbine fault diagnosis based on multi-channel time-frequency spectrum cloud model fusion according to claim 8 is characterized in that: During the training process, the weighted fusion matrix W of the fusion layer F To limit, W F The vectors in satisfy Right now in is a unit vector. During the training process, for the t-th sample known to be in state i, it is assumed that the output of its convolutional neural network fusion model is YT t =[YT t (1),YT t (2),…YT t (i), ... YT t (l)], the training process aims to determine the network parameters through training so that the optimization objective function between the output of the overall model after training and the known state is minimized, and then determine the network structure parameters. During the training process, the optimization objective function of the training process is set to the mean square error function, and the mean square error objective function is as follows:

10. The method for steam turbine fault diagnosis based on multi-channel time-frequency spectrum cloud model fusion according to claim 2 is characterized in that: In step 6), when using the model, a set of multi-channel signals used for testing should be used to calculate the cloud model feature matrix according to steps 1) to 3) to obtain a set of mean, entropy, and super entropy matrices, which are recorded as expected E test , Entropy H test and super entropy He test Matrix, E test Connected to the first CNN input layer, H test Connected to the second CNN input layer, He test Connected to the third CNN input, the model will output For model output The maximum value is selected to be 1, and the other values ​​are set to 0 to generate a new output result. The judgment method of the state category corresponding to the output result is consistent with the definition of the expected output of the model of the known category training data, that is, the index i where the maximum value is located is recorded as the fault classification result.

Images