CN115392333A - Equipment fault diagnosis method based on improved end-to-end ResNet-BilSTM dual-channel model - Google Patents
Equipment fault diagnosis method based on improved end-to-end ResNet-BilSTM dual-channel model Download PDFInfo
- Publication number
- CN115392333A CN115392333A CN202210176360.1A CN202210176360A CN115392333A CN 115392333 A CN115392333 A CN 115392333A CN 202210176360 A CN202210176360 A CN 202210176360A CN 115392333 A CN115392333 A CN 115392333A
- Authority
- CN
- China
- Prior art keywords
- channel
- model
- resnet
- data
- bilstm
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses an equipment fault diagnosis method based on an improved end-to-end ResNet-BilSTM double-channel model, which specifically comprises the following steps: s1, acquiring original one-dimensional time sequence fault data by adopting an acceleration sensor, and taking the data after simple normalization as input of a model; s2, constructing an improved end-to-end ResNet-BiLSTM dual-channel fault diagnosis model, and inputting normalized signals into two channels of a ResNet model and a BiLSTM model; s3, constructing a 1DECANet module and connecting the DECANet module in series with a ResNet model channel; s4, fusing the dual-channel extraction features by adopting a conditioner mechanism, and outputting a diagnosis result of the equipment fault by using a Softmax classifier, which relates to the technical field of dual-channel model construction and industrial equipment fault diagnosis in deep learning. According to the equipment fault diagnosis method of the end-to-end double-channel model, the improved end-to-end ResNet-BilSTM double-channel fault diagnosis model is constructed by aiming at the problem that the nonequilibrium and the recessive characteristics of fault diagnosis data in the field of industrial internet are difficult to extract, and the fault diagnosis efficiency of mechanical equipment is effectively improved.
Description
Technical Field
The invention relates to the technical field of dual-channel model construction and industrial equipment fault diagnosis in deep learning, in particular to an equipment fault diagnosis method based on an improved end-to-end ResNet-BilSTM dual-channel model.
Background
With the rapid development of modern industry, mechanical equipment in an 'intelligent factory' is developed in an integrating and hybridization direction. Mechanical equipment is an indispensable important component in the industrial field, and along with continuous operation of the mechanical equipment, various faults of equipment components cannot be avoided. The reasons for component failure are complex and various, and the condition monitoring and failure diagnosis of mechanical equipment are important contents of failure diagnosis technology in the industrial field. If the treatment of the component faults cannot be timely and effectively realized, the mechanical equipment cannot operate to influence industrial production, so that the fault diagnosis of the mechanical equipment has very important significance for improving the industrial production efficiency and the economic benefit.
The mechanical equipment fault diagnosis research mainly focuses on two parts, namely characteristic value extraction and state identification, which are core parts for mechanical equipment state monitoring. In the long-term operation process, the monitoring and the operation parameter acquisition of the operation state of the mechanical equipment are realized by connecting various sensors with the mechanical equipment, wherein a vibration acceleration signal is a main object for acquisition and research. The deep neural network is an effective model for solving the diagnosis problem of a complex system, can directly model highly nonlinear, complex and multidimensional vibration acceleration signal data, and excavates the mapping relation between the data and a diagnosis target. An article [ Zhang Hongbin and the like, a bearing fault diagnosis method [ J ]. The university of Western-Ann traffic, 2020,54 (8): 58-66 ] of a multi-channel sample and a deep convolutional neural network is adopted, the time-frequency domain characteristics of bearing vibration signals at two ends of a rotor are extracted by using continuous wavelet transformation, single-channel two-dimensional graphic samples of 3 types of vibration signals are constructed and fused, and the single-channel two-dimensional graphic samples are input into a convolutional neural network CNN to realize characteristic extraction and accurate classification of bearing fault signals. With the significant breakthrough of the feedback neural network in the field of Natural Language Processing (NLP), an article [ Fan Yuxue, etc. ] researches [ J ] Noise and vibration control based on a small sample rolling bearing fault diagnosis method of BI-LSTM, 2020 (4): 108-113 ] decomposes a signal by a self-Adaptive white Noise mode (Complete EEMD with Adaptive Noise, CEEMDAN) and Fourier transform and inputs the decomposed signal into a bidirectional long short-term memory BilSTM neural network to realize fault diagnosis of a high-speed train wheel bearing. An article [ Tan et al, rolling bearing fault diagnosis based on Single gate recurrent network [ J ]. Journal of Physics Conference Series,2020, 1601. An article [ Chen Wei, and the like ] rolling bearing fault identification based on RS-LSTM [ J ] Chinese scientific and technological paper 2018,13 (10): 1134-1141 ] combines Random Search (RS) and LSTM neural network to realize intelligent classification of vibration acceleration signals of different fault types of the rolling bearing, and uses original fault data as input to verify that the algorithm has higher generalization capability and robustness.
The fault diagnosis model based on signal processing and deep learning realizes fault diagnosis and state monitoring of mechanical equipment such as mechanical bearings and gears in the industrial field. Aiming at the advantages and disadvantages of different scenes and different feature data and different diagnostic algorithms, how to fuse the advantages of the algorithms and reduce the human intervention process, an end-to-end diagnostic model is constructed to be applied to extraction and deep mining of features of mass feature data, and the speed and precision of equipment fault diagnosis in the industrial field are improved, which is still the main problem to be solved.
Disclosure of Invention
Technical problem to be solved
Aiming at the defects of the prior art, the invention provides an equipment fault diagnosis method based on an improved end-to-end ResNet-BilSTM dual-channel model. In order to realize cross-channel interaction of deep and shallow data features of the deep neural network, a 1DECANet module is constructed and is integrated into a ResNet feature extraction channel. And finally, inputting a full connection layer, fusing the characteristics extracted by the double channels, and realizing accurate identification of equipment faults by means of a Softmax classifier.
(II) technical scheme
In order to achieve the purpose, the invention is realized by the following technical scheme: an equipment fault diagnosis method based on an improved end-to-end ResNet-BilSTM dual-channel model specifically comprises the following steps:
s1, collecting one-dimensional time sequence data of original vibration acceleration by adopting an acceleration sensor, and simply normalizing the data to be used as input of an end-to-end diagnosis model;
s2, constructing a ResNet model channel to carry out deep excavation and extraction on the implicit unobvious features of the original signal, and avoiding the gradient diffusion phenomenon of a deep neural network;
s3, constructing a 1DECANet module, connecting the DECANet module with a ResNet model channel in series, and performing cross-channel interaction on the deep and shallow data characteristics to be used as an improved ResNet characteristic extraction channel;
s4, constructing a stack type BilSt model channel to extract time sequence correlation characteristics of original fault data, and adding a Batch Normalization (BN) layer and a Dropout layer to avoid overfitting of the model channel;
s5, building an end-to-end double-channel diagnosis model, and fusing data characteristics extracted by the improved ResNet model channel and the stack type BilTM model channel by adopting a Concatenate mechanism in a full connection layer;
s6, coordinating unbalance of training sample data by utilizing a Focal local Loss function, and mining implicit characteristics of difficultly-classified samples to realize parameter updating of dual-channel diagnostic model training;
and S7, classifying data of various fault signals by adopting a Softmax function, and realizing accurate diagnosis of unbalanced fault signals.
Preferably, in the steps S2 and S3, an improved ResNet feature extraction channel is set up to implement cross-channel interaction of deep and shallow feature data, and the specific steps are as follows:
t1, constructing a one-dimensional ResNet neural network channel, namely, superposing an identity mapping layer on the basis of a shallow network to extract local features of the normalized time sequence signal along the direction of a time axis, and performing convolution layer and identity mapping calculation according to a formula;
y lay ＝F(x lay ,W i )+W s x lay
F＝W 2 σ(W,x lay )
x lay+1 ＝f(y lay )
wherein x is lay Representing the input of the residual block, x lay+1 Represents the output of the residual block, W, W i 、W 2 All represent conventional one-dimensional convolution calculations, W s Represents 1*1 convolution calculation, sigma represents Sigmoid activation function, and f (-) represents ReLU activation function;
t2, constructing a 1DECANet module, and performing deep and cross-channel mining and extraction on a hidden fault in an original fault signal through a non-dimensionality-reduction local cross-channel interaction strategy;
wherein C represents the number of characteristic data channels of the network intermediate layer, k represents the size of a one-dimensional convolution kernel calculated by the self-adaptive adjusting function, and gamma =2, b =1;
and T3, performing series fusion on the constructed ResNet module and the 1DECANet module to jointly form an improved ResNet feature extraction channel, and performing fusion on the BilSTM feature extraction channel by adopting a conditioner mechanism to provide a classification feature parameter basis for the end-to-end diagnosis model.
Preferably, in step S6, the imbalance of the fault sample data is coordinated by using a Focal local Loss function, specifically:
e1, firstly, adding a modulation coefficient omega > 0 on the basis of a standard cross entropy loss function to reduce the loss weight of samples which are easy to classify and endow more weight to sample data which are difficult to classify and wrong to classify;
wherein L is 0 Represents the standard cross entropy loss function, L 1 Representing a cross entropy loss function after adding a modulation coefficient, X representing a fault sample, n representing the total number of samples, a representing an expected output, and y representing the actual output of a neuron;
e2, adding a balance factor lambda on the basis to balance the problem of uneven proportion of the positive and negative samples to form a Focal local Loss function;
preferably, in step S1, an acceleration sensor is used to collect one-dimensional time series data of an original vibration acceleration, and the collected data is used as an input of an end-to-end diagnosis model after simple normalization processing, and the method specifically includes the following steps:
p1, collecting fault time-sequence data under different working conditions by adopting a vibration acceleration sensor, and cutting the collected time-sequence data according to the size of a sliding window to generate fault diagnosis sample data
As shown in the following formula:
wherein the sliding windowLength of l sw Wherein n-l sw ×i≥l sw (ii) a i =1,2,3, each time point sample sequence truncated is represented as
Wherein
Are all d-dimensional vectors;
p2, simply normalizing the original time-sequence vibration acceleration sample data to be directly used as input data X of the end-to-end fault diagnosis model inputs ＝[X (1) ,X (2) ,...,X (n-1) ,X (n) ]As shown in the following formula:
preferably, in step S5, an end-to-end dual-channel diagnostic model is built, and the data features extracted from the improved ResNet model channel and the stack-type BiLSTM model channel are fused by using a Concatenate mechanism at the full connection layer, specifically:
a1, channel1 is set as an improved ResNet neural network model, and specifically comprises a ResNet module and A1 DECNet module which are connected in series, wherein the output of the ResNet module is used as the input of the 1DECANet module;
a2, channel2 is a Stacked bidirectional long-short term memory (Stacked Bi-LSTM) neural network model, the number of memory units of the BiLSTM network and the number of layers of the neural network are adjusted, and extraction of different dimensional signal characteristics is achieved;
a3, channel1 and Channel2 feature extraction channels are expanded by a Flatten mechanism and then fused by a Concatenate mechanism.
(III) advantageous effects
The invention provides an equipment fault diagnosis method based on an improved end-to-end ResNet-BilSTM double-channel model. Compared with the prior art, the method has the following beneficial effects:
(1) The equipment fault diagnosis method of the end-to-end double-channel model improves the three processes of end-to-end model construction, data characteristic extraction and mining and model parameter updating based on the ResNet neural network and the BiLSTM neural network by aiming at the problem that existing fault diagnosis sample data is unbalanced, time sequence characteristic and hidden fault is difficult to extract in the field of industrial internet. Compared with the traditional fault diagnosis method, the method not only increases the attention mechanism of a feature extraction channel, but also realizes the updating of parameters of the double-channel diagnosis model by means of a Focal Loss function so as to deal with non-balanced sample data, thereby solving the problem of fault diagnosis of mechanical equipment in the industrial field.
(2) Compared with the traditional diagnosis model based on signal processing and deep learning, the equipment fault diagnosis method of the end-to-end double-channel model has the advantages that: 1) The method has the advantages that the original vibration acceleration data are directly input into an end-to-end diagnosis model only through simple normalization processing, original characteristics of time sequence data are reserved to the maximum extent, and follow-up chain reaction caused by poor effect of a signal preprocessing technology is avoided; 2) The method for diagnosing the faults of the improved end-to-end ResNet-BilSTM double-channel model is provided, and context information of hidden difficultly-identified features and time sequence data in unbalanced data is fully mined; 3) Adding a Dropout layer and a batch normalization layer in the dual-channel model to standardize the characteristic data and avoid overfitting of the model; 4) And introducing a channel attention mechanism, constructing a 1DECNet module and embedding a residual error network channel, avoiding influence on feature extraction caused by intermediate layer data dimensionality reduction, and simultaneously realizing cross-channel interaction of deep and shallow layer data features in a deep neural network.
(3) The method for diagnosing the equipment fault of the end-to-end double-channel model is applied to a American West university storage bearing fault data set, and the effectiveness of the method for diagnosing the fault of the improved end-to-end ResNet-BilSTM double-channel model is verified through test analysis, wherein the diagnosis precision on 11 different unbalance ratio test sets is 97.67% at least, and the highest diagnosis precision is 99.52%. Compared with a ResNet-BilSTM dual-channel model without the 1DECANet module, the model convergence rate is obviously improved, the minimum loss value of the model is reduced to 0.0711, and accurate fault diagnosis of the bearing component of the mechanical equipment in the industrial field can be realized.
Drawings
FIG. 1 is a block diagram of a framework of an end-to-end neural network model of the present invention;
FIG. 2 is a schematic diagram of the residual error module ResNet of the present invention;
FIG. 3 is a schematic diagram of the present invention 1 DECANet;
FIG. 4 is a frame diagram of an improved ResNet feature extraction channel according to the present invention;
FIG. 5 is a frame diagram of a stacked BilSTM feature extraction channel according to the present invention;
FIG. 6 is a diagram of the overall framework of the improved end-to-end ResNet-BilSTM two-channel model of the present invention;
FIG. 7 is a flow chart of the dual channel model training and diagnosis provided by the present invention;
FIG. 8 is a visual diagram of vibration acceleration signals of normal and 9-class faults of the rolling bearing of the invention;
FIG. 9 is a visualization of the two-channel model input layer proposed by the present invention;
FIG. 10 is a visualization of a fusion layer T-SNE of the present invention;
FIG. 11 is a graph of dual channel model diagnostic accuracy as proposed by the present invention;
FIG. 12 is a graph of loss value degradation for the present invention;
FIG. 13 is a confusion matrix diagram of class 9 fault diagnosis classification results of the present invention;
FIG. 14 is a comparison of model training convergence before and after the present invention is incorporated into 1 DECANet.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Referring to fig. 1 to 14, an embodiment of the present invention provides a technical solution: an equipment fault diagnosis method based on an improved end-to-end ResNet-BilSTM dual-channel model takes industrial mechanical equipment fault diagnosis as a carrier, takes ResNet and BilSTM algorithms as a main algorithm framework, and the model is shown in figure 1, and specifically comprises the following steps:
s1, acquiring one-dimensional time sequence data of the original vibration acceleration by adopting an acceleration sensor, and simply normalizing the data to be used as the input of an end-to-end diagnosis model.
1-1) collecting fault time sequence data under different working conditions by adopting a vibration acceleration sensor, and cutting the collected time sequence data according to the size of a sliding window to generate fault diagnosis sample data
As shown in the following formula:
wherein the sliding window has a length of l sw Wherein n-l sw ×i≥l sw (ii) a i =1,2,3, each time point sample sequence truncated is represented as
Wherein
Are d-dimensional vectors.
1-2) simply normalizing the original time-sequence vibration acceleration sample data, directly using the sample data as input data of an end-to-end fault diagnosis model, wherein the frame structure of the end-to-end neural network model is shown in figure 1, and data X inputs ＝[X (1) ,X (2) ,...,X (n-1) ,X (n) ]As shown in the following formula:
s2, constructing a ResNet model channel to carry out deep excavation and extraction on the implicit unobvious characteristics of the original signal, and avoiding the gradient diffusion phenomenon of a deep neural network.
The main idea of the ResNet module is to overlay Identity maps (Identity maps) on the basis of a shallow network, so as to avoid the problem that the network degrades with depth increase, and the residual module principle of the ResNet neural network is shown in fig. 2. When the identity mapping function H (x) is directly fitted in the middle layer lay )＝x lay It is relatively difficult, therefore, to introduce a residual function of F (x) lay )＝H(x lay )-x lay When F (x) lay ) → 0, i.e. the identity map, where x is lay Representing the input of the residual module. Performing equal mapping through shortcut, wherein the addition of F (x) and x is a process of adding element by element, and the formula is as follows:
y lay ＝F(x lay ,W i )+x lay
F＝W 2 σ(W,x lay )
wherein x is lay Representing the input of the residual block, W, W i 、W 2 All represent one-dimensional convolution calculations and σ represents the Sigmoid activation function.
When the dimensions of F (x) and x are different, a linear mapping needs to be performed on x to match the dimensions, and there are generally two solutions for linear mapping: (1) The addition of dimension (2) by W matrix projection to new space directly by zero padding can be achieved by 1*1 convolution and changing the number of filters for 1*1 convolution. The invention selects a second scheme to calculate the output of the residual error module, and the formula is as follows:
y lay ＝F(x lay ,W i )+W s x lay
x lay+1 ＝f(y lay )
wherein W s Denotes 1*1 convolution calculation, x lay+1 F (-) is the ReLU activation function, which is the output of the residual module.
S3: and constructing a 1DECANet module, connecting the DECANet module with a ResNet model channel in series, and performing cross-channel interaction on the deep and shallow data characteristics to serve as an improved ResNet characteristic extraction channel.
2-1) construction of 1DECANet Module first one dimensional Global average Pooling Global Avg Pooling to get 1X 1C feature maps;
2-2) calculating through a self-adaptive adjusting function to obtain the size k of the convolution kernel, wherein the size k of the self-adaptive convolution kernel is calculated because the number of the middle-layer channels in the network is generally an index of 2, and the formula is as follows:
where C represents the number of intermediate layer data channels, γ =2, b =1.
2-3) the size k of the convolution kernel of the one-dimensional convolution realizes self-adaptive calculation adjustment by means of a function, so that layers with larger channel number can perform cross-channel interaction more. And applying k to one-dimensional convolution calculation, and obtaining the weight of each channel through an activation function Sigmoid to finally complete information interaction among the cross-channels. The principle of the constructed 1DECANet module is shown in FIG. 3;
2-4) connecting the constructed 1DECANet module and the ResNet feature extraction module in series, wherein the output of the ResNet module is used as the input of the 1DECANet module, and the 1DECANet + ResNet module formed in series jointly forms an improved ResNet feature extraction channel, as shown in FIG. 4.
S4, constructing a stack type BilSTM model channel to extract time sequence correlation characteristics of original fault data, and adding a Batch Normalization (BN) layer and a Dropout layer to avoid overfitting of the model channel.
LSTM unit cell calculation procedure was as follows:
f t ＝σ(W t [h t-1 ,x t ]+b t )
i t ＝σ(W i [h t-1 ,x t ]+b t )
o t ＝σ(W o [h t-1 ,x t ]+b o )
H t ＝o t tanhC t
wherein f, i and o respectively represent the calculation results of the forgetting gate, the input gate and the output gate,
value to be updated for the cellular status of LSTM, C t For updated cell state values, H t W and b represent weight matrix and offset vector participating in training in the memory cell unit respectively, which are final output values of the LSTM cell unit; σ and tanh represent Sigmoid functions and hyperbolic tangent functions, respectively, which serve as activation functions for different "gate" structures.
In order to realize the context judgment of the signal data, the invention realizes two processes of forward calculation and backward calculation by means of the hidden layer of the BilSTM neural network, and can provide data context information for the network, and the updating process of the BilSTM is shown as the following formula:
wherein LSTM + (·)、LSTM - (. H) is the LSTM cell unit operation, W, above hy And W h ' y Respectively weighing values of a BilSTM forward calculation layer and a reverse calculation layer; b y Is the bias vector of the output layer. In the two-way LSTM calculation process, the weight calculation is simultaneously carried out on the context information of the data, and more data features are learned compared with the traditional LSTM neural network, as shown in figure 5.
And S5, building an end-to-end double-channel diagnosis model, and fusing data characteristics extracted by the improved ResNet model channel and the stack type BilTM model channel by adopting a Concatenate mechanism in a full connection layer.
The constructed frame structure of the improved end-to-end ResNet-BilSt dual-Channel model is shown in FIG. 6, and features extracted by combining two channels are fused by a conditioner mechanism after passing through a Full Connection (FC) layer, so that more classification feature parameter bases are provided for the input of a classification layer.
S6, coordinating unbalance of training sample data and mining implicit characteristics of difficultly-divided samples by utilizing a Focal local Loss function to update parameters of the training of the dual-channel diagnostic model.
The loss function on the conventional cross is formulated as follows:
where X represents the faulty sample, n represents the total number of samples, a represents the desired output, and y represents the actual output of the neuron, which is calculated to be between 0 and 1 by the activation function.
The Focal local Loss function adds a modulation coefficient omega > 0 on the basis of a standard cross entropy Loss function to reduce the Loss weight of easily classified samples, and endows difficult and wrongly classified sample data with larger weight, wherein the formula is as follows:
on the basis, a balance factor lambda is added to balance the problem of uneven proportion of positive and negative samples, and the formula is as follows:
s7: and classifying various fault signal data by adopting a Softmax function, and realizing accurate diagnosis of unbalanced fault signals.
The classification layer takes the sample characteristic vector after the two channels are fused as input, and a Softmax classifier is adopted to obtain probability distribution of fault vibration acceleration time sequence signals belonging to different classes, wherein the probability distribution is shown as the following formula:
wherein m represents the number of output layer units of the classification layer, namely the number of fault signal classes.
Based on the steps, the method effectively solves the problem of fault diagnosis of mechanical equipment, and firstly, the method simply normalizes an original vibration acceleration signal to be used as input data of an end-to-end diagnosis model; secondly, aiming at the characteristic that the time sequence and the recessive characteristics of the unbalanced fault signal are difficult to extract, constructing a ResNet model Channel1 and a stack type BilSTM model Channel2 which are fused with a 1DECANet module respectively to carry out deep-level and cross-Channel excavation on the data characteristics; and finally, fusing the data characteristics extracted by the two channels and realizing accurate classification of the fault vibration signals on a classification layer. The method obviously improves the fault diagnosis precision of the mechanical equipment and effectively improves the convergence speed of model training.
The invention is based on the test verification of the fault diagnosis method of the improved end-to-end ResNet-BilSTM double-channel model:
1. description of data
The test data was derived from a rolling bearing failure vibration data set at the bearing data center of the university of Keiss West reservoir (CWRU) USA. The driving end and the fan end are respectively provided with 6205-2RS JEM SKF deep groove Ball bearings and 6203-2RS JEM SKF deep groove Ball bearings, and single point damages with different grades are respectively arranged at 3 different positions of an Inner ring (Inner radius), an Outer ring (Outer radius) and a rolling body (Ball) of the rolling bearing arranged at the driving end and the fan end by adopting an electric spark machining technology. The failure diameters were set at 0.007, 0.014, 0.021 inches, respectively, and the failure depths were set at 0.011, 0.050, 0.150 inches, respectively. Vibration data are collected at frequencies of 12kHz and 48kHz respectively by vibration acceleration sensors arranged at the driving end, the fan end and the base. The invention carries out fault diagnosis experiments on 3 types of faults with different fault diameters at 3 different fault positions of an inner ring, an outer ring and a rolling body of a driving end based on a sampling frequency of 12kHz, and each type of fault samples is divided into data samples with different imbalance proportions by taking 100 data points as time step length, wherein the information of the 9 types of bearing fault experiment samples is shown in Table 1:
TABLE 1 bearing failure experiment sample information of 9 types (failure diameter unit: foot)
The vibration acceleration data of normal and 9 different failure types of the rolling bearing are visualized in a specified range of 1000 time steps, as shown in fig. 8. Amplitude and period differences exist among the 9 types of vibration signals with different fault types, weak periodicity exists among the vibration acceleration signals, and the implicit characteristics of the signals are difficult to diagnose and identify directly through a data distribution rule.
2. Improved end-to-end ResNet-BilSTM double-channel model structure parameter
For the improved end-to-end ResNet-BilSTM double-channel fault diagnosis model, the original non-equilibrium time sequence data are simply normalized and then directly input into the model, the deep excavation and extraction of the non-equilibrium sample data implicit characteristic are realized by utilizing the overall fitting degree of the hidden layer of the end-to-end model, and the parameters of the end-to-end classification model are shown in a table 2. The stack-type BilSTM channel fully extracts the context information of the time-sequence signal, the ResNet channel fully extracts the recessive non-obvious characteristic of the local space of the unbalanced data, and the cross-channel communication of the shallow data and the intermediate layer characteristic data is realized. The improved ResNet channel is embedded into a 1DECANet module, so that the influence of the middle layer feature data dimension reduction on feature extraction is avoided, and the overall diagnosis precision of the model is improved on the basis of slightly increasing model calculation parameters.
TABLE 2 end-to-end Classification model parameters
3. Improved end-to-end ResNet-BilSTM double-channel model fault diagnosis effect
The method comprises the steps of randomly dividing collected failure signals after normalization of original one-dimensional time-sequence signals and corresponding class labels into a training set and a testing set according to 7:3, further randomly dividing the training set into training and verification data according to 9:1 for evaluation of a current training result of a model, and enabling proposed two-channel model training and diagnosis flow Cheng Rutu to be shown, wherein model training parameters are shown in table 3.
TABLE 3 model training Process description
In order to verify the effectiveness of the improved end-to-end ResNet-BilSTM diagnostic model in extracting the implicit characteristics of unbalanced data, the section constructs 11 unbalanced sample data sets and compares the diagnostic precision of a test data set with the loss value of the test set. And inputting the data of the test set into a model with the best training for fault diagnosis, wherein the diagnosis precision of the improved end-to-end ResNet-BilSTM double-channel diagnosis model on the unbalance test set is 97.67% at the lowest as shown in Table 4.
TABLE 4 diagnosis of unbalanced data set
In order to more intuitively show the effectiveness of the improved end-to-end ResNet-BilSTM double-channel fault diagnosis model on non-equilibrium data implicit feature extraction, an input layer and a fusion layer of the end-to-end model are respectively visualized by means of a T-SNE algorithm. A test sample set with the number of 6 is selected for testing, as shown in fig. 9, the visualization result of the input layer of the end-to-end model shows that 9 different fault feature data are randomly distributed and have a unbalance phenomenon, and the number of samples with the fault labels of 6 is much larger than that of other fault samples. Fig. 10 shows that the overlap rate between the data of different fault characteristics of the fusion layer 9 of the end-to-end model is low, and the model distinguishes different implicit characteristics of unbalanced sample data remarkably. Further proves that the improved end-to-end ResNet-BilSTM double-channel diagnosis model can effectively and fully extract the recessive characteristics of the unbalanced samples, and finally realizes the accurate classification of the unbalanced fault samples.
A ResNet module is arranged on a ResNet channel of the improved end-to-end ResNet-BilSTM dual-channel diagnosis model, namely when the ResNet channel is connected with a 1DECANet module in series after cross-channel communication of feature data is realized only once, the end-to-end model shows the optimal average diagnosis precision. Fig. 10 shows a model training accuracy and loss value reduction curve. The loss value descending curve shows that the model is basically converged when the training batch reaches about 20 times, the diagnosis results of the model on the test set are displayed by using a confusion matrix, as shown in fig. 11, the average value of the diagnosis results of 9 types of faults is taken as the overall diagnosis performance of the model, and the average diagnosis result of the 11 types of unbalanced fault sample test set reaches 98.74%.
To further verify the role of the 1DECANET module in the convergence process of the improved end-to-end ResNet-BilSTM dual-channel fault diagnosis model training, ablation experiments were performed on the model training loss values before and after the 1DECANET module was added, as shown in FIG. 12. The result shows that the model loss value is reduced more rapidly after the 1DECANet module is embedded, and the efficient channel attention mechanism obtains lower loss value and faster convergence speed under the premise of adding a small number of parameters, which means that the model can obtain higher diagnosis precision on the basis of not increasing the complexity of the model greatly by integrating the 1DECANet module.
4. Conclusion
In order to solve the problems that original fault signal data acquired under complex working conditions are lost, and hidden characteristics of different types of fault samples are difficult to extract due to unbalance, the invention provides an improved end-to-end ResNet-BilSTM double-channel fault diagnosis model. The cross-channel communication of the feature data and the mining of deep recessive features are realized by introducing a ResNet residual module, a 1DECANet module is constructed and embedded into a ResNet feature extraction channel, and the influence of the dimension reduction of the feature data on feature extraction in the training process is avoided. Meanwhile, the module also avoids the phenomena of rapid increase of model parameters and slow convergence speed of the model caused by the conventional complex attention mechanism. In order to verify the effectiveness of the extracted model on the implicit characteristic extraction of the unbalanced data, the invention constructs a plurality of unbalanced bearing sample sets with different proportions, inputs the unbalanced bearing sample sets into an end-to-end diagnosis model for fault diagnosis, and adopts a Focal local Loss function as a model convergence basis. Experimental results show that the improved end-to-end ResNet-BilSTM double-channel diagnosis model can effectively deal with various unbalanced sample data sets and can effectively extract the implicit characteristics of different types of unbalanced data sets.
And those not described in detail in this specification are well within the skill of those in the art.
It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.
Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.
Claims (5)
1. An equipment fault diagnosis method based on an improved end-to-end ResNet-BilSTM dual-channel model is characterized by comprising the following steps: the method specifically comprises the following steps:
s1, acquiring original vibration acceleration one-dimensional time sequence data by adopting an acceleration sensor, and taking the acquired data as input of an end-to-end diagnosis model after simple normalization processing;
s2, constructing a ResNet model channel to carry out deep excavation and extraction on the implicit unobvious features of the original signal;
s3, constructing a 1DECANet module, connecting the DECANet module in series with a ResNet model channel, and performing cross-channel interaction on the deep and shallow data characteristics to serve as an improved ResNet characteristic extraction channel;
s4, constructing a stack type BilSt model channel to extract time sequence correlation characteristics of original fault data, and adding a batch normalization layer and a Dropout layer to avoid overfitting of the model channel;
s5, building an end-to-end double-channel diagnosis model, and fusing data characteristics extracted by the improved ResNet model channel and the stack type BilTM model channel by adopting a Concatenate mechanism in a full connection layer;
s6, coordinating unbalance of training sample data and mining implicit characteristics of difficultly-divided samples by utilizing a Focal local Loss function, and updating parameters of the training of the dual-channel diagnosis model;
and S7, classifying data of various fault signals by adopting a Softmax function, and realizing accurate diagnosis of unbalanced fault signals.
2. The method for diagnosing the equipment fault based on the improved end-to-end ResNet-BilSTM dual-channel model according to claim 1, wherein: in the steps S2 and S3, an improved ResNet feature extraction channel is set up to realize cross-channel interaction of depth feature data, and the specific steps are as follows:
t1, constructing a one-dimensional ResNet neural network channel, namely superposing an identity mapping layer on the basis of a shallow network to extract local features of the normalized time sequence signal along the direction of a time axis, and performing convolution layer and identity mapping calculation according to a formula;
y lay ＝F(x lay ,W i )+W s x lay
F＝W 2 σ(W,x lay )
x lay+1 ＝f(y lay )
wherein x lay Representing the input of the residual block, x lay+1 Representing the output of the residual block, W, W i 、W 2 All represent conventional one-dimensional convolution calculations, W s Represents 1*1 convolution calculation, sigma represents Sigmoid activation function, and f (-) represents ReLU activation function;
t2, constructing a 1DECANet module, and performing deep and cross-channel mining and extraction on a latent fault in an original fault signal through a non-dimensionality-reduction local cross-channel interaction strategy;
wherein C represents the number of characteristic data channels of the network intermediate layer, k represents the size of a one-dimensional convolution kernel calculated by the self-adaptive adjusting function, and gamma =2, b =1;
and T3, performing series fusion on the constructed ResNet module and the 1DECANet module to jointly form an improved ResNet feature extraction channel, and performing fusion on the BilSTM feature extraction channel by adopting a conditioner mechanism to provide a classification feature parameter basis for the end-to-end diagnosis model.
3. The method for diagnosing the equipment fault based on the improved end-to-end ResNet-BilSTM dual-channel model according to claim 1, wherein: in the step S6, a Focal local Loss function is used to coordinate imbalance of fault sample data, specifically:
e1, firstly, adding a modulation coefficient omega > 0 on the basis of a standard cross entropy loss function to reduce the loss weight of samples which are easy to classify and endow more weight to difficult and misclassified sample data;
wherein L is 0 Represents the standard cross entropy loss function, L 1 Representing a cross entropy loss function after adding a modulation coefficient, X representing a fault sample, n representing the total number of samples, a representing an expected output, and y representing the actual output of a neuron;
e2, adding a balance factor lambda on the basis to balance the problem of uneven proportion of the positive and negative samples to form a Focal local Loss function;
4. the method for diagnosing the equipment fault based on the improved end-to-end ResNet-BilSTM dual-channel model according to claim 1, wherein: in the step S1, an acceleration sensor is used to collect one-dimensional time-sequence data of an original vibration acceleration, and the collected data is used as an input of an end-to-end diagnosis model after simple normalization processing, and the method specifically includes the following steps:
p1, collecting fault time-sequence data under different working conditions by adopting a vibration acceleration sensor, and cutting the collected time-sequence data according to the size of a sliding window to generate fault diagnosis sample data
As shown in the following formula:
wherein the sliding window has a length of l sw Wherein n-l sw ×i≥l sw (ii) a i =1,2,3, each time sample sequence taken is represented as
Wherein
Are all d-dimensional vectors;
p2, simply normalizing the original time-sequence vibration acceleration sample data to be directly used as input data X of the end-to-end fault diagnosis model inputs ＝[X (1) ,X (2) ,...,X (n-1) ,X (n) ]As shown in the following formula:
5. the method for diagnosing the equipment fault based on the improved end-to-end ResNet-BilSTM dual-channel model according to claim 1, wherein: in the step S5, an end-to-end dual-channel diagnostic model is built, and the data features extracted from the improved ResNet model channel and the stack-type BiLSTM model channel are fused by using a Concatenate mechanism in the full connection layer, specifically:
a1, channel1 is set as an improved ResNet neural network model, and specifically, the model is formed by connecting a ResNet module and A1 DECNet module in series, and the output of the ResNet module is used as the input of the 1DECANet module;
a2, channel2 is a stack type bidirectional long-short term memory neural network model, the number of memory units of a BilSTM network and the number of layers of a neural network are adjusted, and extraction of different dimensional signal characteristics is realized;
a3, channel1 and Channel2 feature extraction channels are expanded by a Flatten mechanism and then fused by a Concatenate mechanism.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210176360.1A CN115392333A (en) | 2022-02-24 | 2022-02-24 | Equipment fault diagnosis method based on improved end-to-end ResNet-BilSTM dual-channel model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210176360.1A CN115392333A (en) | 2022-02-24 | 2022-02-24 | Equipment fault diagnosis method based on improved end-to-end ResNet-BilSTM dual-channel model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115392333A true CN115392333A (en) | 2022-11-25 |
Family
ID=84115600
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210176360.1A Pending CN115392333A (en) | 2022-02-24 | 2022-02-24 | Equipment fault diagnosis method based on improved end-to-end ResNet-BilSTM dual-channel model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115392333A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116701918A (en) * | 2023-08-02 | 2023-09-05 | 成都星云智联科技有限公司 | Rolling bearing fault diagnosis method based on LightGBM feature extraction and BiLSTM |
| CN116880279A (en) * | 2023-07-12 | 2023-10-13 | 中冀电力集团股份有限公司 | Intelligent power consumption control management system |
| CN117056865A (en) * | 2023-10-12 | 2023-11-14 | 北京宝隆泓瑞科技有限公司 | Method and device for diagnosing operation faults of machine pump equipment based on feature fusion |
-
2022
- 2022-02-24 CN CN202210176360.1A patent/CN115392333A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116880279A (en) * | 2023-07-12 | 2023-10-13 | 中冀电力集团股份有限公司 | Intelligent power consumption control management system |
| CN116701918A (en) * | 2023-08-02 | 2023-09-05 | 成都星云智联科技有限公司 | Rolling bearing fault diagnosis method based on LightGBM feature extraction and BiLSTM |
| CN116701918B (en) * | 2023-08-02 | 2023-10-20 | 成都星云智联科技有限公司 | Rolling bearing fault diagnosis method based on LightGBM feature extraction and BiLSTM |
| CN117056865A (en) * | 2023-10-12 | 2023-11-14 | 北京宝隆泓瑞科技有限公司 | Method and device for diagnosing operation faults of machine pump equipment based on feature fusion |
| CN117056865B (en) * | 2023-10-12 | 2024-01-23 | 北京宝隆泓瑞科技有限公司 | Method and device for diagnosing operation faults of machine pump equipment based on feature fusion |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110162018B (en) | Incremental equipment fault diagnosis method based on knowledge distillation and hidden layer sharing | |
| Shao et al. | A novel approach of multisensory fusion to collaborative fault diagnosis in maintenance | |
| CN115392333A (en) | Equipment fault diagnosis method based on improved end-to-end ResNet-BilSTM dual-channel model | |
| CN110361176A (en) | A kind of intelligent failure diagnosis method for sharing neural network based on multitask feature | |
| CN110307982B (en) | Bearing fault classification method based on CNN and Adaboost | |
| CN108510153A (en) | A kind of multi-state rotary machinery fault diagnosis method | |
| CN111914883B (en) | Spindle bearing state evaluation method and device based on deep fusion network | |
| CN112577747A (en) | Rolling bearing fault diagnosis method based on space pooling network | |
| CN110657984B (en) | Planetary gearbox fault diagnosis method based on reinforced capsule network | |
| CN111721536A (en) | Rolling bearing fault diagnosis method for improving model migration strategy | |
| CN107784325A (en) | Spiral fault diagnosis model based on the fusion of data-driven increment | |
| CN106779069A (en) | A kind of abnormal electricity consumption detection method based on neutral net | |
| CN113822139B (en) | Equipment fault diagnosis method based on improved 1DCNN-BiLSTM | |
| CN112819059B (en) | Rolling bearing fault diagnosis method based on popular retention transfer learning | |
| Shen et al. | A deep multi-label learning framework for the intelligent fault diagnosis of machines | |
| CN110333076B (en) | Bearing fault diagnosis method based on CNN-Stacking | |
| CN114358123B (en) | Generalized open set fault diagnosis method based on deep countermeasure migration network | |
| Di et al. | Ensemble deep transfer learning driven by multisensor signals for the fault diagnosis of bevel-gear cross-operation conditions | |
| CN107967487A (en) | A kind of colliding data fusion method based on evidence distance and uncertainty | |
| CN111597991A (en) | Rehabilitation detection method based on channel state information and BilSTM-Attention | |
| CN112147432A (en) | BiLSTM module based on attention mechanism, transformer state diagnosis method and system | |
| CN113887136A (en) | Improved GAN and ResNet based electric vehicle motor bearing fault diagnosis method | |
| CN113887342A (en) | Equipment fault diagnosis method based on multi-source signals and deep learning | |
| You et al. | An efficient lightweight neural network using BiLSTM-SCN-CBAM with PCA-ICEEMDAN for diagnosing rolling bearing faults | |
| CN114429152A (en) | Rolling bearing fault diagnosis method based on dynamic index antagonism self-adaption |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |