A hierarchical fault diagnosis method for mechanical equipment based on adaptive hierarchical clustering and Subset

Abstract

The present invention relates to the technical field of predicting mechanical failures using big data, and specifically to a mechanical equipment hierarchical fault diagnosis method based on adaptive hierarchical clustering and Subset, wherein training samples without fault degree distinction under different fault types are selected as inputs of adaptive hierarchical clustering, and adaptive hierarchical clustering is used to perform cluster analysis on the training samples without fault degree distinction under different fault types, and the clustering results of the fault types obtained according to the vibration signals of the mechanical equipment are used as output results; for different fault types identified by training, a SubDAE model is established according to the fault degree classification; pre-training and overall fine-tuning accurately locate the fault point, thereby improving the intelligence of the method; the use of optimal features simplifies the feature set, avoids the dimensionality disaster, reduces the calculation burden, and improves the efficiency of fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of predicting mechanical failures using big data, and specifically relates to a mechanical equipment hierarchical fault diagnosis method based on adaptive hierarchical clustering and Subset. Background Art

[0002] In recent years, the application of deep learning to solve the research of fault diagnosis of complex industrial systems has begun to emerge: Lei Yaguo et al. proposed a big data health monitoring method for mechanical equipment based on denoising autoencoder (DAE), which realized multiple fault diagnosis of planetary gears, reflecting the powerful ability of deep learning to extract mechanical vibration signal features. Yu W et al. effectively integrated DAE and EN to solve the noise interference problem in fault diagnosis, effectively detected abnormal samples in industrial processes, and isolated fault variables from normal variables. Nguyen V H et al. proposed a deep learning network composed of autoencoder and SoftMax classifier to identify bearing faults of different degrees. DBN is more combined with other technologies to solve fault diagnosis problems. Since CNN was used to identify bearing faults in 2016, the performance and scope of fault diagnosis have been continuously improved. Hoang D T et al. proposed a new method for rolling bearing fault diagnosis based on CNN, which used the effectiveness of CNN for image classification to achieve 100% diagnostic accuracy for the CWRU bearing data set. Based on ResNet-50, Wen L et al. proposed a transfer learning convolutional neural network TCNN for fault diagnosis, and its prediction accuracy is significantly better than other DL models and traditional diagnostic methods. The application of RNN in fault diagnosis has been revived since 2015. Abed W et al. used RNN for bearing fault diagnosis and achieved accurate detection and classification of bearing faults under non-stationary conditions. Pan H et al. proposed a method for bearing fault classification by combining one-dimensional CNN and LSTM, with an experimental test accuracy of 99.6%.

[0003] Patent No. CN103398843B discloses a planetary gearbox sun gear fault classification method based on a multi-classification relevance vector machine. The patent segments the original data of the acquired vibration signal, obtains the eigenvalues ​​of each segment respectively, selects the Gaussian radial basis as the kernel function of the multi-classification relevance vector machine, and then inputs the eigenvalues ​​into the multi-classification relevance vector machine for classification training; Patent No. CN104748962B discloses a planetary gearbox intelligent diagnosis method based on a stacked autoencoder. The patent inputs the original data into a stacked autoencoder classification model with a deep structure, and uses a weighted connection autoencoder to pre-train all hidden layers of the stacked autoencoder layer by layer, thereby realizing adaptive extraction of planetary gearbox fault features and intelligent diagnosis of fault status under large amounts of data.

[0004] In summary, although the algorithms in the existing technology have been applied in mechanical equipment fault diagnosis, the fault recognition rate still has a lot of room for improvement. However, the obstacle to improving the technical ideas is that due to the large number of measurement points, monitoring parameters and high sampling frequency of mechanical equipment, large, diverse and complex state monitoring big data is formed, causing mechanical equipment fault diagnosis to enter the "big data" era. In addition, the obstacle to technical improvement is that the faults of modern mechanical equipment are often latent and have weak dynamic responses, making them difficult to detect and process.

[0005] From the above analysis, it can be seen that the contradiction of intelligent monitoring of mechanical equipment failures in modern production enterprises lies in: repairing on demand according to the early signs of equipment failure to reduce downtime, while traditional intelligent fault diagnosis methods require a lot of signal processing technology to extract relatively accurate feature parameters. If only shallow models are used to characterize the relationship between signals and faults, the diagnostic and generalization capabilities are insufficient, and it is difficult to meet the actual needs of fault diagnosis under big data. Summary of the invention

[0006] The purpose of the present invention is to address the above-mentioned problems. The present invention provides a mechanical equipment hierarchical fault diagnosis method based on adaptive hierarchical clustering and Subset. Summary of the invention:

[0008] The present invention is based on proposing a hierarchical fault diagnosis algorithm model AHCAS-HFD based on adaptive hierarchical clustering and Subset: the primary diagnosis one-dimensional convolutional autoencoder is only used to distinguish large categories with large differences in fault severity, the secondary diagnosis convolutional neural network receives the recognition results of the primary diagnosis, and on this basis further identifies the fault type, that is, obtains small categories with less difference, and integrates the powerful feature self-learning ability of the convolutional neural network to realize an end-to-end fault diagnosis model with original vibration data as input and fault diagnosis results as output.

[0009] The technical solution for realizing the present invention is:

[0010] A mechanical equipment hierarchical fault diagnosis method based on adaptive hierarchical clustering and Subset, characterized by comprising the following steps:

[0011] Step 1: Select training samples without distinguishing the degree of fault under different fault types as input of adaptive hierarchical clustering, use adaptive hierarchical clustering to perform cluster analysis on the training samples without distinguishing the degree of fault under different fault types, and use the clustering result of the fault type obtained according to the vibration signal of the mechanical equipment as the output result; wherein the training samples include: the vibration signal of the known mechanical equipment, and the fault type corresponding to the vibration signal of the known mechanical equipment;

[0012] Step 2: For different fault types identified by training in step 1, a SubDAE model is established according to the fault degree classification; the training samples also include: the fault degree of the fault type corresponding to the vibration signal of the known mechanical equipment;

[0013] The vibration signal of the mechanical equipment refers to: a vibration signal that is affected by both the type of mechanical equipment fault and the degree of mechanical equipment fault, such as the vibration signal of the mechanical equipment gear, the vibration signal of the rolling bearing, etc. The horizontal axis of the vibration signal is time, and the vertical axis is measured by the acceleration sensor.

[0014] In the actual scene experiment of the present invention, the vibration signal of the actual mechanical equipment to be tested is input into the model, and the recognition rate of the fault type and fault degree can reach more than 99.4%, which can ensure the reliable, safe and stable operation of the mechanical equipment.

[0015] Preferably, according to the present invention, when the training sample is used as input, it needs to be preprocessed, and the method includes:

[0016] (1-1) Use the normalization formula to map the vibration signal data to the [0,1] interval:

[0017]

[0018] In formula (I), z i is the preprocessed data; x i is the frequency value of the vibration signal; and It is the minimum and maximum frequency in each group of vibration signals. F represents the number of vibration signals in a sample, f=1,2,L,F.

[0019] This technical feature uses Min-Max standardization to map all vibration signal data of each group to [0,1]: the ordinate changes, the abscissa remains unchanged, and it is still in the time domain. The introduction of this technical feature eliminates the dimensional influence between data, and conducts comprehensive comparative evaluation, making the operation efficiency and prediction accuracy of the neural network higher.

[0020] Preferably, according to the present invention, the method of adaptive hierarchical clustering in step 1 comprises:

[0021] (2-1) Given a system S =<X,Q,U> , where X = {x1, x2, Λ, x num} represents the training sample set, Q={q1,q2,L,q J} represents the feature set, U = {u1,u2,L,u C} represents the fault type set;

[0022]

[0023] In formula (II), is the characteristic value of the jth feature of the ith sample of the cth type fault; I is the number of samples of the cth type fault; C is the number of fault types; J also represents the number of features of each sample in each type of fault type, and the number of features of each sample in each type of fault type is the same as the data x i The number of is equal;

[0024] (2-2) Calculate the intra-class distance of the j-th feature of the c-th class:

[0025]

[0026] In formula (III), They represent the characteristic values ​​of the jth feature of the lth and mth samples of the cth type fault respectively;

[0027] (2-3) Calculate the average intra-class distance of the jth feature C fault class

[0028]

[0029] Calculate the average value u of the jth feature of the Ith sample of the cth class c,j :

[0030]

[0031] In formula (V), represents the characteristic value of the jth feature of the mth sample of the cth type fault;

[0032] Calculate the average of the inter-class distances of the jth feature C class faults

[0033]

[0034] In formula (VI), u e,j 、u c,j They represent the average value of the jth feature of the Ith sample in the eth class and the average value of the jth feature of the Ith sample in the cth class respectively;

[0035] (2-4) For the clustering of the training sample set described in steps (2-1) to (2-3), the silhouette coefficient s(k) is:

[0036]

[0037] In formula (VII), s iis the silhouette coefficient of the sample individual, num is the number of samples in the training sample set, and k is the number of clusters;

[0038]

[0039] In formula (VIII), a(i) represents the average distance between sample i and all other samples belonging to the same type of fault, and b(i) represents the minimum value of the average distance between sample i and all samples in each class of non-type faults;

[0040]

[0041]

[0042] n c Represents the number of samples in the cth category, C c represents the sample of the cth class; p Indicates a mark other than class c, n p Indicates the number of samples that are not of the cth class, C p Represents samples that are not of the cth category;

[0043] The principle of adaptive clustering described in this technical feature is to obtain a clustering result so that the distance between classes is as large as possible, the distance within classes is as small as possible, and the classes have good separability. This algorithm uses the clustering silhouette coefficient as an indicator for evaluating the effectiveness of clustering and establishes an adaptive hierarchical clustering algorithm.

[0044] (2-5) The closer the calculated silhouette coefficient s(k) is to -1, the worse the clustering result is; the closer it is to 1, the better the clustering result is. Find the maximum s(k), that is, S max That is, the k value closest to 1;

[0045] (2-6) Start clustering, let k = 1, S max =0;

[0046] (2-7) Let k = k + 1, with k as the number of clusters, and perform hierarchical clustering on the input training samples;

[0047] (2-8) Calculate s(k) according to steps (2-1)-(2-5):

[0048] When s(k) < 0, s_Index is used as the number of clusters (s_Index is a parameter used to record the number of clusters), and the clustering results are output;

[0049] When s(k)≥0, then judge s(k)>S max :

[0050] When s(k)>S max When s_Index=k,Smax =s(k), execute steps (2-9);

[0051] When s(k)<S max , then return to step (2-7);

[0052] (2-9) Determine whether k is less than n (n represents the number of data set types):

[0053] When k<n, s_Index is used as the number of clusters and the clustering results are output;

[0054] When k≥n, repeat steps (2-7).

[0055] Preferably, according to the present invention, the method for training the SubDAE model for each fault degree training sample in step 2 is as follows:

[0056] (3-1) Initialize DAE parameters;

[0057] (3-2) Determine the number of hidden layers N of the deep network, and train N DAEs layer by layer in an unsupervised manner, that is, use the hidden layer output of each DAE as the input of the next layer of DAE until the training of N DAEs is completed;

[0058] (3-3) Start training, set h = 1;

[0059] (3-4) Train the hth DAE and use the value of the hidden layer of the hth DAE as the input of the h+1th DAE;

[0060] (3-5) Determine whether h is less than N.

[0061] When h<N, h=h+1 and repeat steps (3-4);

[0062] (3-6) When h ≥ N, add an output layer according to the fault classification, and perform supervised training to fine-tune the DAE parameters as a whole.

[0063] (3-7) In the above supervised training, labeled data samples are used, and the labeled data samples are (x (1) ,y (1) )、(x (2) ,y (2) ),...,(x (m) ,y (m) ), x (i) represents the feature vector, y (i) ∈{1,2,Λ,t} represents the fault type,

[0064] The cost function of the DAE classification model is:

[0065]

[0066] Among them, J(θ) is the loss function; represents the weight matrix between the feature space and the output layer, r represents the labels of the nine types of faults; P(y (i) ) is the feature vector x (i) The corresponding fault label y (i) The probability of

[0067]

[0068] (3-8) Input: batch information epoch, batch_size, global learning rate α, initialization parameter θ0 T ;

[0069] Intermediate variable: probability P(y (i) ), loss function J(θ); where epoch represents the number of times the training sample set is trained, and batch_size represents the number of groups the training sample set is divided into;

[0070] (3-9) Randomly extract batch_size samples {x1, x2, ..., x batch_size}, and its corresponding output category {y1, y2, ... y batch_size};

[0071] Calculate the eigenvector x (i) The probability of the corresponding fault label P(y (i) );Calculate the loss function J(θ);

[0072] (3-10) Determine whether the difference between two adjacent loss functions J(θ) is less than 0.0001: for example, the difference between the loss function value obtained after running the data for i epochs and the loss function value obtained after running the data for i+1 epochs;

[0073] If it is not less than, the weight matrix θ is iteratively updated using the gradient descent method. The iterative formula is as follows:

[0074]

[0075] in

[0076] Update the weight matrix once and continue with step (3-9) to obtain P(y (i) ) and J(θ);

[0077] If it is less than, the model training ends and the trained SubDAE model is obtained.

[0078] In step 1, the fault type obtained is Ydi (i=1,2,……,s_Index), Yd i represents the i-th fault type; in step 2, the fault degree obtained is Yd j i (j=1, 2..., a, i=1, 2..., s_Index), Yd j i It represents the j-type fault degree under the i-th type fault.

[0079] Compared with the prior art, the present invention has the following beneficial effects:

[0080] 1. This algorithm directly converts the original vibration signal from the time domain signal to the frequency domain signal through Fourier transform through data preprocessing as the input of 1D-CNN, which can obtain the characteristics of the vibration signal to the greatest extent.

[0081] 2. Select training samples of all fault severity as input for adaptive hierarchical clustering, perform cluster analysis on fault types, and then classify each fault type according to its fault severity to establish a SubDAE model (one SubDAE model for each fault severity). Use the severity of a class of fault types to train a SubDAE model. Pre-training and overall fine-tuning accurately locate the fault point, which improves the intelligence of this method. The use of optimal features simplifies the feature set, avoids the dimensionality curse, reduces the computational burden, and improves the efficiency of fault diagnosis. Through experiments, it can be found that the proposed model can achieve a fault recognition rate of more than 99.4%, which can ensure the reliable, safe and stable operation of mechanical equipment.

[0082] 3. The AHCAS-HFD algorithm includes two parts: fault type identification and fault severity identification. For fault types with obvious feature differences, clustering is used for classification and identification to obtain coarse classification results; for fault severity with similar features, deep neural networks are used for classification and identification to achieve fine-grained identification of fault severity, improve the precision and accuracy of fault diagnosis, identify fault types can provide guidance for the maintenance and replacement of corresponding components, and identify fault severity can provide a basis for the remaining service life of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0084] Figure 1is a flow chart of the adaptive hierarchical clustering algorithm of the present invention,

[0085] Figure 2 is a fault diagnosis flow chart of the present invention,

[0086] Figure 3 is a flow chart of establishing a model for training samples of the AHCAS-HFD algorithm of the present invention,

[0087] Figure 4 Schematic diagram of a gearbox experimental system for collecting experimental data using the AHCAS-HFD algorithm of the present invention,

[0088] Figure 5 This is a schematic diagram of the fault of the gear used in the gearbox experimental system of the present invention.

[0089] Figure 6 The experimental sample information and corresponding category labels of the present invention using the vibration signals of gears in different states as sample data sets,

[0090] Figure 7 It is the iteration curve of the SubDAE model of the present invention using the vibration signals of gears in different states as the sample data set,

[0091] Figure 8 is the diagnostic accuracy of each model of the present invention at different epochs on the test set using the vibration signals of gears in different states as the sample data set,

[0092] Fig. 9 FIG. 4 is a flow chart of a diagnostic process of a test sample of the AHCAS-HFD algorithm of the present invention. DETAILED DESCRIPTION

[0093] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0094] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "provided with", "connected", etc. should be understood in a broad sense. For example, "connected" 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 indirectly connected 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 the present invention can be understood according to specific circumstances.

[0095] The present invention provides the following technical solution: a mechanical equipment hierarchical fault diagnosis method based on adaptive hierarchical clustering and Subset, comprising the following steps:

[0096] Step 1: Select training samples without distinguishing the degree of fault under different fault types as input of adaptive hierarchical clustering, use adaptive hierarchical clustering to perform cluster analysis on the training samples without distinguishing the degree of fault under different fault types, and use the clustering result of the fault type obtained according to the vibration signal of the mechanical equipment as the output result; wherein the training samples include: the vibration signal of the known mechanical equipment, and the fault type corresponding to the vibration signal of the known mechanical equipment;

[0097] Step 2: For different fault types identified by training in step 1, a SubDAE model is established according to the fault degree classification; the training samples also include: the fault degree of the fault type corresponding to the vibration signal of the known mechanical equipment;

[0098] The vibration signal of the mechanical equipment refers to: a vibration signal that is affected by both the type of mechanical equipment fault and the degree of mechanical equipment fault, such as the vibration signal of the mechanical equipment gear, the vibration signal of the rolling bearing, etc. The horizontal axis of the vibration signal is time, and the vertical axis is measured by the acceleration sensor.

[0099] In the actual scene experiment of the present invention, the vibration signal of the actual mechanical equipment to be tested is input into the model, and the recognition rate of the fault type and fault degree can reach more than 99.4%, which can ensure the reliable, safe and stable operation of the mechanical equipment.

[0100] When the training samples and test samples are used as input, they need to be preprocessed. Specifically, the experimental data of the data sets of the training samples and test samples in the specific embodiment of the present invention are collected from a benchmark two-stage gearbox, such as Figure 4 As shown in the figure, the gear speed change is controlled by a motor. The torque is provided by a magnetic brake and can be adjusted by changing its input voltage. A 32-tooth pinion (N1-32) and an 80-tooth pinion (N2-80) are installed on the first-stage input shaft (Input Shaft). The second stage (Idle Shaft) consists of a 48-tooth pinion (N3-48) and a 64-tooth pinion (N4-64). The input shaft (Input Shaft) speed is measured by a tachometer (Tachometer), and the gear vibration signal is measured by an accelerometer (Accelerometer), as shown in Figure 5 As shown, the pinion (N1-32) on the input shaft introduces nine different gear conditions, including normal, missing teeth, root cracks, spalling and tip cutting with five different severities.

[0101] Preprocessing methods include:

[0102] (1-1) Use the normalization formula to map the vibration signal data to the [0,1] interval:

[0103]

[0104] In formula (I), z i is the preprocessed data; x i is the frequency value of the vibration signal; and is the minimum and maximum frequency of each group of vibration signals, F represents the number of vibration signals in a sample, f = 1, 2, L, F. Figure 6 It means, F = 3600;

[0105] This technical feature uses Min-Max standardization to map all vibration signal data of each group to [0,1]: the ordinate changes, the abscissa remains unchanged, and it is still in the time domain. The introduction of this technical feature eliminates the dimensional influence between data, and conducts comprehensive comparative evaluation, making the operation efficiency and prediction accuracy of the neural network higher.

[0106] (1-2) Divide the training samples and test sample sets processed in step (1-1) to form training samples and test samples. This technical feature uses the pre-processed training samples and test samples as the input of the AHCAS-HFD model, making data processing more convenient and accelerating the network learning speed, so that the mean of the input signal of all samples is close to 0 or very small compared to its mean square error. Figure 6 As shown, all vibration signal data are randomly divided into training samples and verification samples. 85% of the samples in each state are randomly selected as training samples, and the rest are used as test samples. The total number of samples is marked by categories, as shown in Table 1:

[0107] Table 1. Category labels

[0108]

[0109]

[0110] The method of adaptive hierarchical clustering in step 1 includes:

[0111] (2-1) Given a system S =<X,Q,U> , where X = {x1, x2, L, x num} represents the training sample set, Q={q1,q2,L,q J} represents the feature set, U = {u1,u2,L,u C} represents the fault type set;

[0112]

[0113] In formula (II), is the characteristic value of the jth feature of the ith sample of the cth type fault; I is the number of samples of the cth type fault; C is the number of fault types; J is the number of features of each sample in each type of fault type, and the number of features of each sample in each type of fault type is the same as the data x i The number of is equal, according to Figure 6 It can be seen that J = 3600, num = 9*88, the total number of test samples is 16*9, I = 88;

[0114] (2-2) Calculate the intra-class distance of the j-th feature of the c-th class:

[0115]

[0116] In formula (III), They represent the characteristic values ​​of the jth feature of the lth and mth samples of the cth type fault respectively;

[0117] (2-3) Calculate the average intra-class distance of the jth feature C fault class

[0118]

[0119] Calculate the average value u of the jth feature of the Ith sample of the cth class c,j :

[0120]

[0121] In formula (V), represents the characteristic value of the jth feature of the mth sample of the cth type fault;

[0122] Calculate the average of the inter-class distances of the jth feature C class faults

[0123]

[0124] In formula (VI), u e,j 、u c,j They represent the average value of the jth feature of the Ith sample in the eth class and the average value of the jth feature of the Ith sample in the cth class respectively;

[0125] (2-4) For the clustering of the training sample set described in steps (2-1) to (2-3), the silhouette coefficient s(k) is:

[0126]

[0127] In formula (VII), s i is the silhouette coefficient of the sample individual, num is the number of samples in the training sample set, and k is the number of clusters;

[0128]

[0129] In formula (VIII), a(i) represents the average distance between sample i and all other samples belonging to the same type of fault, and b(i) represents the minimum value of the average distance between sample i and all samples in each class of non-type faults;

[0130]

[0131]

[0132] n c Represents the number of samples in the cth category, C c represents the sample of the cth class; p represents the mark of non-cth class, n p Indicates the number of samples that are not of the cth class, C p Represents samples that are not of the cth category;

[0133] The principle of adaptive clustering described in this technical feature is to obtain a clustering result so that the distance between classes is as large as possible, the distance within classes is as small as possible, and the classes have good separability. This algorithm uses the clustering silhouette coefficient as an indicator for evaluating the effectiveness of clustering and establishes an adaptive hierarchical clustering algorithm.

[0134] like Figure 1 As shown:

[0135] (2-5) The closer the calculated silhouette coefficient s(k) is to -1, the worse the clustering result is; the closer it is to 1, the better the clustering result is. Find the maximum s(k), that is, S max That is, the k value closest to 1;

[0136] (2-6) Start clustering, let k = 1, S max =0;

[0137] (2-7) Let k = k + 1, with k as the number of clusters, and perform hierarchical clustering on the input training samples or test samples;

[0138] (2-8) Calculate s(k) according to steps (2-1)-(2-5):

[0139] When s(k)<0, s_Index is used as the number of clusters and the clustering results are output;

[0140] When s(k)≥0, then judge s(k)>S max :

[0141] When s(k)>S max When s_Index=k,S max =s(k), execute steps (2-9);

[0142] When s(k)≤S max , then return to step (2-7);

[0143] (2-9) Determine whether k is less than n (n represents the number of data set types, n is the type of failure of the known mechanical equipment, in this specific embodiment, n=9):

[0144] When k<n, s_Index is used as the number of clusters, and the clustering result is output. In this embodiment, the output s_Index=5;

[0145] When k≥n, repeat steps (2-7).

[0146] The method for training the SubDAE model for each fault degree training sample in step 2 is as follows:

[0147] like Figure 2 As shown:

[0148] (3-1) Initialize DAE parameters;

[0149] (3-2) Determine the number of hidden layers N of the deep network, and train N DAEs layer by layer in an unsupervised manner, that is, use the hidden layer output of each DAE as the input of the next layer DAE until the training of N DAEs is completed; in this specific embodiment, the program is run first according to the training samples, and the number of hidden layers N determined by the program is 9;

[0150] (3-3) Start training, set h = 1;

[0151] (3-4) Train the hth DAE and use the value of the hidden layer of the hth DAE as the input of the h+1th DAE;

[0152] (3-5) Determine whether h is less than N.

[0153] When h<N, h=h+1 and repeat steps (3-4);

[0154] (3-6) When h ≥ N, add an output layer according to the fault classification, and perform supervised training to fine-tune the DAE parameters as a whole.

[0155] (3-7) In the above supervised training, labeled data samples are used (labeled data samples are samples with known fault types and fault degrees selected from training samples), and the labeled data samples are (x (1) ,y (1) )、(x(2) ,y (2) ),...,(x (m) ,y (m) ), x (i) represents the feature vector, y (i) ∈{1,2,Λ,t} represents the fault type, which is the known fault type and fault degree.

[0156] The cost function of the DAE classification model is:

[0157]

[0158] Among them, J(θ) is the loss function; represents the weight matrix between the feature space and the output layer, r represents the labels of the nine types of faults, P(y (i) ) is the feature vector x (i) The corresponding fault label y (i) The probability of

[0159]

[0160] (3-8) Input: batch information epoch, batch_size, global learning rate α, α = 0.0001 in this patent, initialization parameters Initially a random value;

[0161] Intermediate variable: probability P(y (i) ), loss function J(θ); where epoch represents the number of times the training sample set is trained, and batch_size represents the number of groups the training sample set is divided into;

[0162] (3-9) Randomly extract batch_size samples {x1, x2, ..., x batch_size}, and its corresponding output category {y1, y2, ... y batch_size};

[0163] Calculate the eigenvector x (i) The probability of the corresponding fault label P(y (i) );Calculate the loss function J(θ);

[0164] (3-10) Determine whether the difference between the upper and lower loss functions J(θ) is less than 0.0001: for example, the difference between the loss function value obtained after running the data for i epochs and the loss function value obtained after running the data for i+1 epochs;

[0165] If it is not less than, the weight matrix θ is iteratively updated using the gradient descent method. The iterative formula is as follows:

[0166]

[0167] in

[0168] Update the weight matrix once and continue with step (3-9) to obtain P(y (i) ) and J(θ);

[0169] If it is less than, the model training ends and the trained SubDAE model is obtained.

[0170] like Figure 3 As shown in the figure, based on the selected experimental samples, the AHCAS-HFD model used in this experiment is constructed. The primary diagnosis uses adaptive hierarchical clustering to obtain the rough classification results. The secondary diagnosis uses the Subset method for the fault degree with similar characteristics to establish the Subset-DAE model for fault degree identification. In step 1, the fault type obtained is Yd i (i=1,2,……,s_Index), Yd i represents the i-th fault type; in step 2, the fault degree is Yd j i (j=1, 2..., a, i=1, 2..., s_Index), Yd j i It represents the j-type fault degree under the i-th type fault.

[0171] like Fig. 9 As shown, step 3: use the test samples to test the model trained in step 1 and step 2: input the test samples without distinguishing the fault degree under different fault types into the adaptive hierarchical clustering, and output the fault type; input the test samples with different fault degrees into the trained SubDAE model, and output the fault degree identification result; the test samples include: the vibration signal of the known mechanical equipment, the fault type corresponding to the real-time signal of the known mechanical equipment, and the fault degree corresponding to the real-time signal of the known mechanical equipment.

[0172] It is worth pointing out that although the maximum number of iterations of model training is set to 50, this paper introduces the Earlystopping mechanism. By monitoring the changes of the training set loss function between adjacent iterations during training, Earlystopping can terminate the model training in a timely manner to prevent the model from overfitting. If there is no obvious change within a certain number of iterations, the model training is terminated. The joint operation with Dropout can effectively prevent the common overfitting phenomenon in training deep learning models and effectively improve the generalization ability of the model.

[0173] The model in this paper is built on the Python-based pytorch deep learning library. The computer hardware configuration information is i5 processor, 8GB running memory, Linux system. Each batch iteration takes an average of about 1 second, and the model reaches the termination condition after 60 iterations on average to end training. The test set is used to verify the effectiveness of each CNN model. The experimental results are shown in Figure 7 shown.

[0174] In order to compare the effects of different backend classifiers, the data of the same dataset is used as training data, and 15% of the data is used as the test set. SVM, ResCNN, CNN, CAECNN and AHCAS-HFD used in this paper are used as classifiers for experiments. The diagnostic accuracy of each model at different epochs on the test set is shown in the following figure. Figure 8 As shown in the figure, each diagnosis result in each model is the result of 3 trainings and predictions and the average value.

[0175] Depend on Figure 8 It can be seen that for the test set running 40 epochs, the diagnostic accuracy of each model is high. The diagnostic accuracy of the AHCAS-HFD model can reach 99.4%, and the diagnostic accuracy of the SVM, ResCNN, CNN, and CAECNN models can reach 86%, 94.4%, 92.6%, and 97.5%, respectively. It can be seen that in the experimental results of different backend classifiers, the AHCAS-HFD classifier can achieve a high recognition rate, and the AHCAS-HFD classifier has the highest recognition rate.

[0176] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered exemplary and non-restrictive in all respects, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A mechanical equipment hierarchical fault diagnosis method based on adaptive hierarchical clustering and Subset, characterized by: The following steps are involved: Step 1: Select training samples without distinguishing the degree of fault under different fault types as input of adaptive hierarchical clustering, use adaptive hierarchical clustering to perform cluster analysis on the training samples without distinguishing the degree of fault under different fault types, and use the clustering result of the fault type obtained according to the vibration signal of the mechanical equipment as the output result; wherein the training samples include: the vibration signal of the known mechanical equipment, and the fault type corresponding to the vibration signal of the known mechanical equipment; Step 2: For different fault types identified by training in step 1, a SubDAE model is established according to the fault severity classification. The method for training the SubDAE model for each fault severity training sample in step 2 is as follows: (3-1) Initialize DAE parameters; (3-2) Determine the number of hidden layers N of the deep network, and train N DAEs layer by layer in an unsupervised manner, that is, use the hidden layer output of each DAE as the input of the next layer of DAE until the training of N DAEs is completed; (3-3) Start training, set h = 1; (3-4) Train the hth DAE and use the value of the hidden layer of the hth DAE as the input of the h+1th DAE; (3-5) Determine whether h is less than N. When h<N, h=h+1 and repeat steps (3-4); (3-6) When h ≥ N, add an output layer according to the fault classification, and perform supervised training to fine-tune the DAE parameters as a whole. (3-7) In the above supervised training, labeled data samples are used, and the labeled data samples are (x (1) ,y (1) )、(x (2) ,y (2) ),...,(x (m) ,y (m) ), x (i) represents the feature vector, y (i) ∈{1,2,…,t} represents the fault type, The cost function of the DAE classification model is: Among them, J(θ) is the loss function; represents the weight matrix between the feature space and the output layer, r represents the fault label; P(y (i) ) is the feature vector x (i) The corresponding fault label y (i) The probability of (3-8) Input: batch information epoch, batch_size, global learning rate α, initialization parameters Intermediate variable: probability P(y (i) ), loss function J(θ); where epoch represents the number of times the training sample set is trained, and batch_size represents the number of groups the training sample set is divided into; (3-9) Randomly extract batch_size samples {x1, x2,……,x batch_size }, and its corresponding output category {y1, y2, ... y batch_size }; Calculate the eigenvector x (i) The probability of the corresponding fault label P(y (i) );Calculate the loss function J(θ); (3-10) Determine whether the difference between two adjacent loss functions J(θ) is less than 0.0001: If it is not less than, the weight matrix θ is iteratively updated using the gradient descent method. The iterative formula is as follows: in Update the weight matrix once and continue with step (3-9) to obtain P(y (i) ) and J(θ); If it is less than, the model training ends and the trained SubDAE model is obtained; The training samples also include: the degree of fault type corresponding to the vibration signal of the known mechanical equipment; The vibration signal of the mechanical equipment refers to a vibration signal that is affected by both the type of mechanical equipment failure and the degree of mechanical equipment failure.

2. According to claim 1, a mechanical equipment hierarchical fault diagnosis method based on adaptive hierarchical clustering and Subset is characterized in that: When the training sample is used as input, it needs to be preprocessed, and the method includes: (1-1) Use the normalization formula to map the vibration signal data to the [0,1] interval: In formula (I), z i is the preprocessed data; x i is the frequency value of the vibration signal; and It is the minimum and maximum frequency in each group of vibration signals, F represents the number of vibration signals in a sample, f = 1, 2, …, F.

3. The mechanical equipment hierarchical fault diagnosis method based on adaptive hierarchical clustering and Subset according to claim 2 is characterized by: The method of adaptive hierarchical clustering in step 1 includes: (2-1) Given a system S = <X,Q,U>, where X = {x1,x2,…,x num } represents the training sample set, Q = {q1,q2,…,q J } represents the feature set, U={u1,u2,…,u C } represents the fault type set; In formula (II), is the characteristic value of the jth feature of the i-th sample of the c-th type fault; I is the number of samples of the c-th type fault; C is the number of fault types; J represents the number of features of each sample in each type of fault; (2-2) Calculate the intra-class distance of the j-th feature of the c-th class: In formula (III), They represent the characteristic values ​​of the jth feature of the lth and mth samples of the cth type fault respectively; (2-3) Calculate the average intra-class distance of the jth feature C fault class Calculate the average value u of the jth feature of the Ith sample of the cth class c,j : In formula (V), represents the characteristic value of the jth feature of the mth sample of the cth type fault; Calculate the average of the inter-class distances of the jth feature C class faults In formula (VI), u e,j 、u c,j They represent the average value of the jth feature of the Ith sample in the eth class and the average value of the jth feature of the Ith sample in the cth class respectively; (2-4) For the clustering of the training sample set described in steps (2-1) to (2-3), the silhouette coefficient s(k) is: In formula (VII), s i is the silhouette coefficient of the sample individual, num is the number of samples in the training sample set, and k is the number of clusters; In formula (VIII), a(i) represents the average distance between sample i and all other samples belonging to the same type of fault, and b(i) represents the minimum value of the average distance between sample i and all samples in each class of non-type faults; n c Represents the number of samples in the cth category, C c represents the sample of the cth class; p represents the mark of non-cth class, n p Indicates the number of samples that are not of the cth class, C p Indicates samples that are not of the cth category; (2-5) Find the maximum s(k), that is, S max ; (2-6) Start clustering, let k = 1, S max =0; (2-7) Let k = k + 1, with k as the number of clusters, and perform hierarchical clustering on the input training samples; (2-8) Calculate s(k) according to steps (2-1)-(2-5): When s(k)<0, s_Index is used as the number of clusters and the clustering results are output; When s(k)≥0, then judge s(k)>S max : When s(k)>S max When s_Index=k,S max =s(k), execute steps (2-9); When s(k)≤S max , then return to step (2-7); (2-9) Determine whether k is less than n, where n represents the number of data set types: When k<n, s_Index is used as the number of clusters and the clustering results are output; When k≥n, repeat steps (2-7).

4. The method for mechanical equipment hierarchical fault diagnosis based on adaptive hierarchical clustering and Subset according to claim 3 is characterized in that: In step 1, the fault type obtained is Yd i (i=1,2,……,s_Index), Yd i represents the i-th fault type; in step 2, the fault degree is Yd j i (j=1, 2..., a, i=1, 2..., s_Index), Yd j i It represents the j-type fault degree under the i-th type fault.

Images