CN117009916A - Actuator fault diagnosis method based on multi-sensor information fusion and transfer learning - Google Patents

Actuator fault diagnosis method based on multi-sensor information fusion and transfer learning Download PDF

Info

Publication number
CN117009916A
CN117009916A CN202310753177.8A CN202310753177A CN117009916A CN 117009916 A CN117009916 A CN 117009916A CN 202310753177 A CN202310753177 A CN 202310753177A CN 117009916 A CN117009916 A CN 117009916A
Authority
CN
China
Prior art keywords
sensor information
transfer learning
model
information fusion
trained
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
Application number
CN202310753177.8A
Other languages
Chinese (zh)
Inventor
薛松
潘成辉
王从思
陈李辉
连培园
许谦
孔德庆
赵武林
彭海波
王晓洁
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xidian University
Original Assignee
Xidian University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xidian University filed Critical Xidian University
Priority to CN202310753177.8A priority Critical patent/CN117009916A/en
Publication of CN117009916A publication Critical patent/CN117009916A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M13/00—Testing of machine parts
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • 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
    • G06N3/096—Transfer learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

The application provides an actuator fault diagnosis method based on multi-sensor information fusion and transfer learning, which comprises the steps of obtaining a vibration acceleration signal of an actuator to be diagnosed in a working state; carrying out signal feature extraction, signal feature fusion and fault type prediction on the vibration acceleration signal through a pre-trained fault diagnosis model to obtain probability distribution of fault types output by the pre-trained fault diagnosis model; the fault type with the highest probability is selected and determined as the fault type of the actuator to be diagnosed. The fault diagnosis model adopts the multi-sensor information fusion and transfer learning technology, so that the multi-dimensional information of the actuator transmission system can be fully utilized, and the accuracy and reliability of fault diagnosis are improved. In addition, the knowledge of the existing data set can be effectively utilized based on the transfer learning method, and the generalization capability of the actuator transmission system fault diagnosis model is improved, so that the method is suitable for more complex working conditions and environments.

Description

Actuator fault diagnosis method based on multi-sensor information fusion and transfer learning
Technical Field
The application belongs to the technical field of intelligent fault diagnosis of traditional systems, and particularly relates to an actuator fault diagnosis method based on multi-sensor information fusion and transfer learning.
Background
Along with the continuous improvement of the industrial automation degree in China, the transmission system is taken as an important component part in a mechanical system, and the reliability and the stability of the transmission system have important influences on the aspects of production efficiency, equipment running cost and the like. However, since the transmission system has various fault types, such as gear damage, bearing failure, tooth surface fatigue, etc., the fault diagnosis thereof is a key technology in mechanical maintenance.
With the continuous development of sensor technology, sensor data acquisition and diagnosis methods in the fault diagnosis of a transmission system are more and more abundant, and in the traditional fault diagnosis method, a signal processing method is generally used for fault diagnosis, and fault characteristics of signals such as wavelet transformation, fourier transformation, EMD decomposition and the like are extracted. Bin et al propose a method for extracting fault feature frequency based on Wavelet Packet Decomposition (WPD) and Empirical Mode Decomposition (EMD), denoising by using the WPD method, obtaining an inherent mode function by using the EMD method, and proposing energy moment of the IMF as a feature vector to effectively express fault features. However, the method mainly depends on expert experience, has more man-made subjectivity, is not good in fault recognition effect, and cannot guarantee accuracy and robustness; and then, combining a machine learning shallow layer network to perform diagnosis, and performing recognition diagnosis on the processed signal characteristics through the shallow layer network.
Rohit et al research is based on advanced signal processing methods and artificial intelligence techniques such as Artificial Neural Networks (ANN) and K-nearest neighbors (KNN) for bearing failure classification. The self-adaptive algorithm based on wavelet transformation is adopted to extract bearing fault classification characteristics from the time domain signals, and then the bearing fault classification characteristics are used as input of an ANN model, and the same characteristics are also used for KNN. Although this approach provides some improvement, a single signal processing approach is not effective in extracting features. With the development of deep learning, the field of fault diagnosis is further improved. Chen et al have used convolutional neural networks (convolutional neural networks, CNN) and extreme learning machines to accomplish the gear and bearing fault classification problem for deep learning bearing fault diagnosis problems. The deep learning network has rapid and strong self-adaptive feature extraction capability, and performs feature extraction and fault classification on the acquired signals in an end-to-end mode. The defects of the traditional diagnosis method are overcome effectively, and the diagnosis efficiency is improved effectively.
Deep transfer learning is to integrate knowledge of deep learning network and transfer learning. The model and the parameter migration can train a deep neural network model on the source field, and then migrate the model to the target field by adjusting part of parameters of the model so as to adapt to new tasks. zhao et al propose a fault diagnosis method based on VGG16 convolutional neural network and transfer learning by applying the fine-tuned VGG16 model to fault diagnosis; the domain self-adaptive transfer learning is to adjust partial parameters of the model to adapt to the task of the target domain under the condition that the deep neural network model is trained on the source domain. Lei et al propose a Deep Convolutional Transfer Learning Network (DCTLN) fault diagnosis method. The one-dimensional CNN learning domain invariant feature is aided by conditionally identifying auto-learning features and identifying the health of the machine and domain-adaptive minimized probability distribution distance. Domain adaptive migration learning more emphasizes domain differences between source and target domains. The training time of the model can be reduced, the data with the labels can be greatly reduced, and the high-efficiency diagnosis result can be achieved. Some students introduce transfer learning in fault diagnosis.
The above method results in poor model generalization performance only for fault information of a single sensor, so that the accuracy of fault diagnosis of an actuator in a transmission system is not high.
Disclosure of Invention
In order to solve the problems in the prior art, the application provides an actuator fault diagnosis method based on multi-sensor information fusion and transfer learning. The technical problems to be solved by the application are realized by the following technical scheme:
the application provides an actuator fault diagnosis method based on multi-sensor information fusion and transfer learning, which comprises the following steps:
s100, obtaining a vibration acceleration signal of an actuator to be diagnosed in a working state;
s200, carrying out signal feature extraction, signal feature fusion and fault type prediction on the vibration acceleration signal of the S100 through a pre-trained fault diagnosis model to obtain probability distribution of fault types output by the pre-trained fault diagnosis model;
s300, selecting the fault type with the highest probability from the probability distribution of the fault types obtained in S200 to determine the fault type of the actuator to be diagnosed.
The application provides an actuator fault diagnosis device based on multi-sensor information fusion and transfer learning, which comprises:
the acquisition device is used for acquiring a vibration acceleration signal of the actuator to be diagnosed in the working state;
the predicting device is used for extracting signal characteristics, fusing the signal characteristics and predicting the fault types of the vibration acceleration signals of the acquiring device through a pre-trained fault diagnosis model to obtain probability distribution of the fault types output by the pre-trained fault diagnosis model;
and a diagnostic device for determining the fault type with the highest probability of selection from the probability distribution of fault types as the fault type of the actuator to be diagnosed.
1. The application has complementarity, accuracy and robustness: the application adopts multi-sensor data fusion to fully utilize the information of different sensors so as to complement the information on fault diagnosis. The data from the plurality of sensors is then processed in multiple levels, aspects, and layers to produce meaningful new information that is not available from any single sensor and which is effective in eliminating errors due to single type sensor failure. The multi-sensor information fusion can improve the robustness of a fault diagnosis system, and the system can still keep high accuracy even if certain sensors fail or data are abnormal; the transfer learning can learn knowledge from data in other related fields, so that the diagnosis model is more accurate;
2. the application can reduce the labeling data requirement: the transfer learning method can utilize the data and the labeling information in the source field to reduce the required labeling data quantity in the target field, thereby reducing the data acquisition and labeling cost, providing useful information and experience and improving the robustness of the model.
3. The application has good real-time performance and diagnosis efficiency: the application can autonomously extract useful information without prior dependence, saves labor cost and saves diagnosis time; the adopted migration network parameters are few, and the diagnosis speed is high. The application can output the diagnosis result only by inputting the signal network, and does not need to manually extract useful information from huge data like the traditional diagnosis method, thereby meeting the real-time requirement of the related field and having high diagnosis speed.
The present application will be described in further detail with reference to the accompanying drawings and examples.
Drawings
FIG. 1 is a flow chart of an actuator fault migration learning method based on multi-sensor information fusion provided by the application;
FIG. 2 is a schematic diagram of an implementation process of the actuator fault migration learning method based on multi-sensor information fusion;
FIG. 3 is a schematic diagram of a parallel convolutional network (PCNN);
FIG. 4 is a schematic diagram of a transfer learning method;
FIG. 5 is a flow chart of a VGG16 model fine tuning process;
FIG. 6 is a flow chart of a network training process;
FIG. 7 is a schematic diagram of the result of wavelet transform CWT;
FIG. 8 is a schematic diagram of a multisensor information fusion and transfer learning (Fmd-VGG 16) based approach;
FIG. 9 is a block diagram of an actuator experiment platform;
FIG. 10 is a plurality of sets of histograms of the diagnostic results of a plurality of single sensor methods and methods presented herein;
FIG. 11 is a diagram of a confusion matrix for diagnostic results of each sensor;
FIG. 12 is a plurality of sets of histograms of diagnostic results without performing a shift learning method and the methods presented herein;
fig. 13 is a diagnostic result confusion matrix diagram of the transfer learning and non-transfer learning multisensor data fusion method.
Detailed Description
The present application will be described in further detail with reference to specific examples, but embodiments of the present application are not limited thereto.
Referring to fig. 1 and 2, the application provides an actuator fault diagnosis method based on multi-sensor information fusion and migration learning, which comprises the following steps:
s100, obtaining a vibration acceleration signal of an actuator to be diagnosed in a working state;
s200, carrying out signal feature extraction, signal feature fusion and fault type prediction on the vibration acceleration signal of the S100 through a pre-trained fault diagnosis model to obtain probability distribution of fault types output by the pre-trained fault diagnosis model;
the fault diagnosis model comprises a pre-trained multi-sensor information fusion model and a pre-trained transfer learning convolution network model, signal characteristics are extracted through the pre-trained multi-sensor information fusion model, and the signal characteristics are fused in a cascading mode to obtain fused multi-sensor information; and outputting fault type probability distribution of the actuator through a pre-trained transfer learning convolutional network model.
Referring to fig. 2, the multi-sensor information fusion model includes two layers of CNN, each of which includes a convolution layer and a max-pooling layer; the convolution layer is used for sequentially carrying out convolution operation, LReLU operation and batch normalization operation on each input time-frequency diagram to obtain signal characteristic diagrams of different convolution kernels; the maximum pooling layer is used for pooling all the signal feature images and selecting the maximum local feature to obtain the depth feature of each time-frequency image, and splicing all the depth features to obtain the fusion result of the information of the multiple sensors; the transfer learning convolutional network model comprises: the input layer, 5 convolution blocks, 5 biggest pooling layers and a full-connection layer, the convolution blocks are alternately connected with the biggest pooling layers, the output of the last biggest pooling layer is connected with the full-connection layer, and the full-connection layer is used for outputting probability distribution of fault types.
Referring to fig. 3 and 5, a CNN layer is constructed using python's tensorflow2.3 library, parallel Convolutional Neural Network (PCNN) is designed to extract features in parallel from information converted from multiple sensors, and the extracted features are cascaded to realize multi-sensor information fusion.
First, two layers of CNNs are established, and condition discrimination features are extracted from each sensor image data, respectively, to receive and extract features from the converted image. The proposed CNN framework follows the general CNN structure and each layer of CNNs includes convolution operations (Conv), lrlu operations, batch normalization operations, and maximum pool operations (pooling). And then splicing the depth features of each data to obtain the fused depth features so as to realize multi-sensor information fusion.
Convolution layer: in the convolution layer, a plurality of convolution kernels are convolved with input data to generate features for the next layer, and different feature maps can be obtained by different convolution kernels. Mathematically, the convolution operation can be written as:
wherein,a j-th feature map representing an i-th layer; />K output features representing the previous layer; />Is a convolution operation; m is M j Representing the size of the input feature; />Representing the corresponding bias.
Pooling layer: the pooling layer may reduce the spatial size of the feature map and further minimize the number of model parameters while maintaining translational invariance. A maximum pool operation is selected to operate each feature map and a maximum local feature is selected. Mathematically, it can be expressed as:
where S is the pool size of the sliding window and X is the input feature map.
The method adopts a parallel convolution neural network to receive and extract the characteristics from the converted images through 2 layers of convolution and pooling operations, and then splices the depth characteristic images of each data to obtain the fused depth characteristics so as to realize multi-sensor information fusion.
Referring to fig. 4 and 5, the migration learning uses a depth model to learn knowledge from a certain domain (source domain) to migrate or apply to other different but similar domains (target domains), and is a machine learning method capable of migrating knowledge learned from other domains to another similar domain. The method can solve the difficult problem of data cross-domain diagnosis under different working conditions, and effectively reduces the requirement of sample number.
The VGG16 convolutional neural network model was proposed by oxford university in 2014, and among numerous VGG variants, VGG16 performs well in both image classification and target detection tasks. The VGG16 model consists of feature extractors of 13 convolutional layers and 5 pooled layers, followed by 3 fully connected and softmax layers to accomplish the classification task.
S300, selecting the fault type with the highest probability from the probability distribution of the fault types obtained in S200 to determine the fault type of the actuator to be diagnosed.
The method is suitable for fault diagnosis of the transmission system of the antenna actuator, and can be applied to fault diagnosis of all fields in which the sensor can collect the characterization signals, such as the fields of antennas, wind driven generators, aerospace, electric and the like. The application has wide application in the field of traditional systems which can replace actuators in actual operation.
The application provides an actuator fault diagnosis method based on multi-sensor information fusion and transfer learning, which comprises the steps of obtaining a vibration acceleration signal of an actuator to be diagnosed in a working state; carrying out signal feature extraction, signal feature fusion and fault type prediction on the vibration acceleration signal through a pre-trained fault diagnosis model to obtain probability distribution of fault types output by the pre-trained fault diagnosis model; the fault type with the highest probability is selected and determined as the fault type of the actuator to be diagnosed. The fault diagnosis model adopts the multi-sensor information fusion and transfer learning technology, so that the multi-dimensional information of the actuator transmission system can be fully utilized, and the accuracy and reliability of fault diagnosis are improved. In addition, the knowledge of the existing data set can be effectively utilized based on the transfer learning method, and the generalization capability of the actuator transmission system fault diagnosis model is improved, so that the method is suitable for more complex working conditions and environments. As an alternative embodiment of the present application, before S100, the actuator fault diagnosis method based on multi-sensor information fusion and transfer learning further includes:
s110, collecting vibration acceleration signals of different positions or different directions of the actuator transmission part bearing under different working states through a plurality of sensors;
according to the application, vibration acceleration signals of the upper bearings of the antenna actuators and the planetary gear boxes under different loads can be acquired through the acceleration sensor, the data are preprocessed to ensure the same distribution of the data, and then the data are divided according to the working state.
S120, preprocessing the vibration acceleration signals acquired in the S110, and dividing the vibration acceleration signals into training samples belonging to a source domain and test samples belonging to a target domain according to different working states; wherein, the working states of the training sample and the test sample are different; the pretreatment steps corresponding to the application are as follows:
the data average value calculation formula:
data variance calculation formula:
the pretreatment formula:
sample sequence after pretreatment: x'. n ＝[x 1 ',x' 2 ,…,x' n ] T
The proportion of the training set and the test set is 80% and 20% respectively.
And S130, training a preset multi-sensor information fusion model and a preset transfer learning convolutional network model by using a training sample and a test sample to obtain a pre-trained multi-sensor information fusion model and a pre-trained transfer learning convolutional network model.
As an alternative embodiment of the present application, S130 includes:
s131, training a preset multi-sensor information fusion model and a preset transfer learning convolutional network model by using a training sample to obtain a pre-trained multi-sensor information fusion model and a pre-trained transfer learning convolutional network model;
and S132, updating weight parameters in the pre-trained multi-sensor information fusion model and the pre-trained transfer learning convolutional network model by the test sample to obtain a trained multi-sensor information fusion model and a trained transfer learning convolutional network model.
Referring to fig. 6, as an alternative embodiment of the present application, S131 includes:
s1311, converting training samples into a first time-frequency diagram by using wavelet transform CWT;
noteworthy are: the fault diagnosis directly using the one-dimensional time domain signal cannot fully express the fault characteristics of the signal, which leads to insufficient prominence of the characteristics extracted from the data by using the deep learning model, increases the model training difficulty, and finally leads to lower model recognition rate. The wavelet transformation can decompose signals in time and frequency, process non-stationary and nonlinear signals, provide more detailed time-frequency information and describe the change of the signals more accurately.
The application uses wavelet transformation to convert signals collected by a plurality of sensors into a RGB three-channel time-frequency diagram, converts one-dimensional vibration acceleration signals into multi-channel image data, and effectively converts the multi-channel image data into image recognition good for a deep learning model.
Expression of CWT function:
wherein, ψ is a,b Representing a family of wavelet functions. a, b represent the scale parameter and the time shift parameter, respectively.
Because the vibration signal fault diagnosis waves of the Morlet wavelet time domain waveform domain actuator transmission system are similar, the Morlet wavelet basis function is selected, and the expression is as follows:
the method converts the divided data set into two-dimensional data through wavelet transform (CWT), directly uses one-dimensional time domain signals to carry out fault diagnosis, has noise interference, further extracts characteristics through wavelet transform and removes noise. The fault characteristics obtained by the time-frequency domain analysis are more obvious, and the analysis based on the time-frequency domain is more comprehensive and accurate, so that the model training can be more accurate by converting one-dimensional data into a time-frequency domain image and inputting the time-frequency domain image into a neural network model through the CWT, and the converted signals of the CWT are shown in figure 7.
S1312, inputting the first time-frequency diagram into a preset multi-sensor information fusion model to extract first signal features through the multi-sensor information fusion model, and cascading the extracted first signal features to obtain a first fusion result of multi-sensor information;
s1313, inputting the first fusion result into a preset transfer learning convolutional network model so as to output a predicted fault type of the training sample through the transfer learning convolutional network model;
s1314, comparing the predicted fault type of the training sample with the real fault type, and adjusting weight parameters of the multi-sensor information fusion model and the transfer learning convolutional network model;
s1315, repeating S1311 to S1314 until the iteration times are reached, and obtaining a pre-trained multi-sensor information fusion model and a pre-trained transfer learning convolutional network model.
S132 includes:
s1321, converting the test sample into a second time-frequency diagram by using wavelet transform CWT;
the implementation process of S1321 in the present application is the same as that of S1311, and will not be described here again.
S1322, inputting a second time-frequency diagram into a pre-trained multi-sensor information fusion model to extract second signal features through the multi-sensor information fusion model, and cascading the extracted second signal features to obtain a second fusion result of multi-sensor information;
s1323, calculating the maximum average difference loss of the first fusion result and the second fusion result, and constructing a total loss function according to MMD loss and classification loss;
MMD loss is a measure of the distance between two probability distributions. As shown in fig. 5, in the VGG16 network model, an MMD loss layer is added to the full connection layer, the feature representations of the source domain and the target domain are mapped into a common hidden space, and the MMD distance between them is calculated. MMD distance can be added as an additional penalty term to the total penalty of the network, and optimized along with conventional classification penalty to improve domain adaptation performance of the network, the MMD penalty is expressed as follows:
wherein n is s Is the number of training samples from the source domain, n t Is the number of test samples from the target domain, I.I H Is the reproduction kernel Hilbert space, y j Representing a second fusion result of the jth test sample; x is x i Representing a first fusion result of the ith training sample;
according to the application, the maximum average difference (Maximum Mean Discrepancy, MMD) loss is added into the training of the fine-tuned VGG16 model, and the model is subjected to domain self-adaptive training by using a back propagation algorithm, so that the MMD distance between a source domain data set and a target domain data set can be minimized, and the adaptability of the model to different data distribution is enhanced.
The total loss function is expressed as:
L c ＝loss+λD (2)；
where λ is a trade-off parameter, loss is a classification loss calculated from the error between the predicted fault type and the actual fault type, expressed as:
wherein y is m Represents the mth sampleThe true probability distribution of the fault type of the present,is the probability distribution of the fault type predicted by the m-th sample, and the total category number of the fault type is n.
S1324, inputting a second fusion result into a pre-trained transfer learning convolutional network model, and fine-tuning weight parameters;
the fine adjustment of the weight parameters comprises fine adjustment of the weight parameters of each layer in the multi-sensor information fusion model by using a second fusion result, and fine adjustment of the weight parameters of an input layer, a fourth convolution block, a fifth convolution block, a maximum pooling layer and a full connection layer in the transfer learning convolution network model.
With continued reference to fig. 2, the fine-tuning transfer learning convolutional network model of the present application first keeps the input layer unchanged, i.e., the top layer unchanged. And freeze the first three convolutions of the transfer learning convolutional network model during training to avoid destroying their feature representation capabilities. The latter two convolved blocks are then trimmed so that their weights can be updated in each cycle of the training model. The model is trimmed in the new dataset, i.e. the test sample of the target domain, and the weights of the network are updated by training, the trimming process is shown in fig. 8. This enables the model to adapt to specific fault diagnosis tasks and improves its performance.
According to the application, the fine adjustment transmission of the model and the parameters is carried out through transfer learning, so that the time for applying the deep learning model diagnosis to the transmission part is effectively reduced, and meanwhile, the accuracy of fault diagnosis is improved.
S1325, with the aim of minimizing the total loss function, repeating the processes of S1321 to S1324 by using a back propagation algorithm, so as to perform domain self-adaptive iterative training on the pre-trained transfer learning convolutional network model until reaching an iteration cut-off condition, and obtaining a trained sensor information fusion model and a transfer learning convolutional network model.
In the training process, the network parameters need to be adjusted through the optimization algorithm so as to enable the predicted result of the network to be as close as possible to the target output value. The back propagation algorithm improves the predictive and generalization capabilities of the network by calculating the error between the network output and the target output and then back propagating the error signal to update the weights and biases of each neuron.
Table 1 network training parameters
The application adopts the characteristic of multi-source data fusion to input the fine-tuned domain self-adaptation (Fmd-VGG 16) transfer learning model, and then inputs the model into a classifier for classification. The method not only can reduce the number of network parameters and prevent the problems of over fitting and the like of the network, but also can improve the generalization capability of the network, thereby improving the fault diagnosis precision of the transmission system.
The application provides an actuator fault diagnosis device based on multi-sensor information fusion and transfer learning, which comprises:
the acquisition device is used for acquiring a vibration acceleration signal of the actuator to be diagnosed in the working state;
the predicting device is used for extracting signal characteristics, fusing the signal characteristics and predicting the fault types of the vibration acceleration signals of the acquiring device through a pre-trained fault diagnosis model to obtain probability distribution of the fault types output by the pre-trained fault diagnosis model;
and a diagnostic device for determining the fault type with the highest probability of selection from the probability distribution of fault types as the fault type of the actuator to be diagnosed.
The present application has hereinafter devised experiments from two angles. The single sensor diagnostic method is first compared to the proposed method to demonstrate the effectiveness of the multi-sensor collaborative diagnosis. And then compared with a non-migratory learning multisensor fusion method to demonstrate the advantages of the proposed method.
In order to verify the validity and accuracy of the transmission system fault diagnosis method for multi-sensor information fusion and transfer learning, data sets of different loads are acquired on an actuator through a plurality of sensors. The actuator is mainly composed of a servo motor, a planetary gear box, a bearing, a lead screw and a load end, as shown in fig. 9. The planetary gear box is provided with two vibration acceleration sensors, and the bearing is provided with one vibration acceleration sensor for collecting vibration signals. The sampling time was 10s and the sampling frequency was 10240Hz. A total of 7 different health conditions were simulated as shown in table 2. During data acquisition, the different loads in 3, 0,1,2KN, were set to simulate different domains.
Table 2 actuator health and label
Specific relevant task settings are given in table 3 for the tasks of the different domain transfer learning fault diagnosis.
TABLE 3 Cross-domain diagnostic tasks
1) Comparison with Single sensor diagnostics
The planetary gear box has two vibration acceleration sensor signals, and one vibration acceleration sensor signal is arranged on the bearing. The diagnostic experiments are performed on the individual sensors separately, and the network structure for the single sensor approach is consistent with the proposed approach to ensure uniformity of verification. The parameters of the network training are consistent with those of table 1, and the fault recognition rate is obtained as shown in table 4 and fig. 10.
TABLE 4 diagnosis results of various sensor diagnosis (%)
Among these six tasks, the highest accuracy of single sensor signal diagnostics is 0.83, 0.85, 0.82, 0.83, 0.84, 0.82, respectively. The diagnosis rates of the actuator fault migration learning method based on multi-sensor information fusion provided by the application are 0.97, 0.99, 0.98, 0.97, 0.98 and 0.97 respectively. By data comparison, the method provided obviously improves the fault diagnosis precision of the actuator transmission system, and proves that the method is superior to other single-sensor signal diagnosis methods.
To further reflect the performance advantages of multi-sensor information fusion and transfer-learned driveline fault diagnostics, confusion matrices are used to display the diagnostic accuracy of the actuator health. Taking the first experiment of task C1 as an example, the all sensor method confusion matrix is shown in FIG. 11. As can be seen from the figure, there is a difference in diagnostic accuracy for individual sensors under certain health conditions. The proposed multi-sensor collaborative diagnosis method can integrate information of each sensor, thereby improving diagnosis results.
2) Comparison with non-migratory learning multisensor fusion methods
To verify the diagnostic superiority of MMD loss to add to the fine-tuned VGG16 transition learning model (Fmmd-VGG 16). A simple VGG16 model (without MMD loss and fine tuning) is set for comparison and verification with the method provided by the application. The parameters of model training are consistent with table 1 and the final diagnostic results are shown in table 5 and fig. 12.
TABLE 5 transfer learning diagnostic results (%)
Through the table comparison, the accuracy of the method provided by the application is obviously higher than that of a model which is not subjected to migration learning, and the validity of the method for cross-domain diagnosis of different working condition data is proved, so that the problem that the cross-domain diagnosis data is difficult to acquire can be solved, and the requirement of a smaller sample number is effectively met. All the results show that the diagnostic performance of the proposed method is superior to the other most advanced methods.
To further illustrate the advantages of the proposed method, a confusion matrix is used to illustrate the diagnostic accuracy of the actuator health condition. Taking the first experiment of task C1 as an example, by not performing transfer learning and performing verification by the present method, the confusion matrix is shown in fig. 13. The method provided by the application is obviously superior to a model which does not carry out transfer learning, and is beneficial to improving the diagnosis result.
In summary, the method for fault migration and learning of the actuator based on multi-sensor information fusion provides a more accurate and rapid intelligent diagnosis method, inputs the image data converted by the CWT into a proposed network model, rapidly outputs a diagnosis result, and provides an end-to-end fault diagnosis solution, which can be widely applied to the fields of actuator transmission systems, antennas, aerospace, electric power and other mechanical equipment.
Furthermore, the terms "first," "second," and the like, are used for descriptive purposes only and are not to be construed as indicating or implying a relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include one or more such feature. In the description of the present application, the meaning of "a plurality" is two or more, unless explicitly defined otherwise.
Although the application is described herein in connection with various embodiments, other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed application, from a study of the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other elements or steps, and the "a" or "an" does not exclude a plurality.
The foregoing is a further detailed description of the application in connection with the preferred embodiments, and it is not intended that the application be limited to the specific embodiments described. It will be apparent to those skilled in the art that several simple deductions or substitutions may be made without departing from the spirit of the application, and these should be considered to be within the scope of the application.

Claims (10)

1. An actuator fault diagnosis method based on multi-sensor information fusion and transfer learning is characterized by comprising the following steps:
s100, obtaining a vibration acceleration signal of an actuator to be diagnosed in a working state;
s200, carrying out signal feature extraction, signal feature fusion and fault type prediction on the vibration acceleration signal of the S100 through a pre-trained fault diagnosis model to obtain probability distribution of fault types output by the pre-trained fault diagnosis model;
s300, selecting the fault type with the highest probability from the probability distribution of the fault types obtained in S200 to determine the fault type of the actuator to be diagnosed.
2. The actuator fault diagnosis method based on multi-sensor information fusion and transfer learning according to claim 1, wherein the fault diagnosis model comprises a pre-trained multi-sensor information fusion model and a pre-trained transfer learning convolution network model, signal features are extracted through the pre-trained multi-sensor information fusion model, and the signal features are fused in a cascading manner to obtain fused multi-sensor information; and outputting fault type probability distribution of the actuator through a pre-trained transfer learning convolutional network model.
3. The multi-sensor information fusion and transfer learning-based actuator failure diagnosis method according to claim 2, characterized in that, before S100, the multi-sensor information fusion and transfer learning-based actuator failure diagnosis method further comprises:
s110, collecting vibration acceleration signals of different positions or different directions of the actuator transmission part bearing under different working states through a plurality of sensors;
s120, preprocessing the vibration acceleration signals acquired in the S110, and dividing the vibration acceleration signals into training samples belonging to a source domain and test samples belonging to a target domain according to different working states;
wherein, the working states of the training sample and the test sample are different;
and S130, training a preset multi-sensor information fusion model and a preset transfer learning convolutional network model by using the training sample and the test sample to obtain a pre-trained multi-sensor information fusion model and a pre-trained transfer learning convolutional network model.
4. The method for actuator fault diagnosis based on multi-sensor information fusion and transfer learning according to claim 3, wherein S130 comprises:
s131, training a preset multi-sensor information fusion model and a preset transfer learning convolutional network model by using the training sample to obtain a pre-trained multi-sensor information fusion model and a pre-trained transfer learning convolutional network model;
and S132, updating weight parameters in the pre-trained multi-sensor information fusion model and the pre-trained transfer learning convolutional network model by the test sample to obtain a trained multi-sensor information fusion model and a trained transfer learning convolutional network model.
5. The method for diagnosing an actuator failure based on multi-sensor information fusion and transfer learning of claim 4, wherein S131 comprises:
s1311, converting training samples into a first time-frequency diagram by using wavelet transform CWT;
s1312, inputting the first time-frequency diagram into a preset multi-sensor information fusion model, extracting first signal features through the multi-sensor information fusion model, and cascading the extracted first signal features to obtain a first fusion result of multi-sensor information;
s1313, inputting the first fusion result into a preset transfer learning convolutional network model so as to output a predicted fault type of the training sample through the transfer learning convolutional network model;
s1314, comparing the predicted fault type of the training sample with the real fault type, and adjusting weight parameters of the multi-sensor information fusion model and the transfer learning convolutional network model;
s1315, repeating S1311 to S1314 until the iteration times are reached, and obtaining a pre-trained multi-sensor information fusion model and a pre-trained transfer learning convolutional network model.
6. The method for actuator fault diagnosis based on multi-sensor information fusion and transfer learning according to claim 4, wherein S132 comprises:
s1321, converting the test sample into a second time-frequency diagram by using wavelet transform CWT;
s1322, inputting the second time-frequency diagram into a pre-trained multi-sensor information fusion model to extract second signal features through the multi-sensor information fusion model, and cascading the extracted second signal features to obtain a second fusion result of multi-sensor information;
s1323, calculating the maximum average difference loss of the first fusion result and the second fusion result, and constructing a total loss function according to MMD loss and classification loss;
s1324, inputting a second fusion result into a pre-trained transfer learning convolutional network model, and fine-tuning weight parameters;
s1325, with the aim of minimizing the total loss function, repeating the processes of S1321 to S1324 by using a back propagation algorithm, so as to perform domain self-adaptive iterative training on the pre-trained transfer learning convolutional network model until reaching an iteration cut-off condition, and obtaining a trained sensor information fusion model and a transfer learning convolutional network model.
7. The method for actuator fault diagnosis based on multi-sensor information fusion and transfer learning according to claim 5, wherein,
the multi-sensor information fusion model comprises two layers of CNN, wherein each layer of CNN comprises a convolution layer and a maximum pooling layer;
the convolution layer is used for sequentially carrying out convolution operation, LReLU operation and batch normalization operation on each input time-frequency diagram to obtain signal characteristic diagrams of different convolution kernels; the maximum pooling layer is used for pooling all the signal feature images and selecting the maximum local feature to obtain the depth feature of each time-frequency image, and splicing all the depth features to obtain the fusion result of the information of the multiple sensors;
the transfer learning convolutional network model comprises: the input layer, 5 convolution blocks, 5 biggest pooling layers and a full-connection layer, the convolution blocks are alternately connected with the biggest pooling layers, the output of the last biggest pooling layer is connected with the full-connection layer, and the full-connection layer is used for outputting probability distribution of fault types.
8. The method for actuator fault diagnosis based on multi-sensor information fusion and transfer learning according to claim 7, wherein fine tuning the weight parameters in S1324 comprises fine tuning the weight parameters of each layer in the multi-sensor information fusion model and fine tuning the weight parameters of the input layer, the fourth convolution block, the fifth convolution block, the max pooling layer and the full connection layer in the transfer learning convolution network model using the second fusion result.
9. The method for diagnosing an actuator failure based on multi-sensor information fusion and transfer learning of claim 6, wherein the maximum average difference loss in S1323 is expressed as:
wherein n is s Is the number of training samples from the source domain, n t Is the number of test samples from the target domain, I.I H Is the reproduction kernel Hilbert space, y j Representing a second fusion result of the jth test sample; x is x i Representing a first fusion result of the ith training sample;
the total loss function is expressed as:
L c ＝loss+λD (2)；
where λ is a trade-off parameter, loss is a classification loss calculated from the error between the predicted fault type and the actual fault type, expressed as:
wherein y is m A true probability distribution representing the fault type of the mth sample,is the probability distribution of the fault type predicted by the m-th sample, and the total category number of the fault type is n.
10. An actuator fault diagnosis device based on multi-sensor information fusion and transfer learning is characterized by comprising:
the acquisition device is used for acquiring a vibration acceleration signal of the actuator to be diagnosed in the working state;
the predicting device is used for extracting signal characteristics, fusing the signal characteristics and predicting the fault types of the vibration acceleration signals of the acquiring device through a pre-trained fault diagnosis model to obtain probability distribution of the fault types output by the pre-trained fault diagnosis model;
and a diagnostic device for determining the fault type with the highest probability of selection from the probability distribution of fault types as the fault type of the actuator to be diagnosed.
CN202310753177.8A 2023-06-25 2023-06-25 Actuator fault diagnosis method based on multi-sensor information fusion and transfer learning Pending CN117009916A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310753177.8A CN117009916A (en) 2023-06-25 2023-06-25 Actuator fault diagnosis method based on multi-sensor information fusion and transfer learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310753177.8A CN117009916A (en) 2023-06-25 2023-06-25 Actuator fault diagnosis method based on multi-sensor information fusion and transfer learning

Publications (1)

Publication Number Publication Date
CN117009916A true CN117009916A (en) 2023-11-07

Family

ID=88570049

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310753177.8A Pending CN117009916A (en) 2023-06-25 2023-06-25 Actuator fault diagnosis method based on multi-sensor information fusion and transfer learning

Country Status (1)

Country Link
CN (1) CN117009916A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117892258A (en) * 2024-03-12 2024-04-16 沃德传动（天津）股份有限公司 Bearing migration diagnosis method based on data fusion, electronic equipment and storage medium
CN118013289A (en) * 2024-04-09 2024-05-10 北京理工大学 Variable working condition small sample fault diagnosis method, device, medium and product based on information fusion element transfer learning
CN117892258B (en) * 2024-03-12 2024-06-07 沃德传动（天津）股份有限公司 Bearing migration diagnosis method based on data fusion, electronic equipment and storage medium

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117892258A (en) * 2024-03-12 2024-04-16 沃德传动（天津）股份有限公司 Bearing migration diagnosis method based on data fusion, electronic equipment and storage medium
CN117892258B (en) * 2024-03-12 2024-06-07 沃德传动（天津）股份有限公司 Bearing migration diagnosis method based on data fusion, electronic equipment and storage medium
CN118013289A (en) * 2024-04-09 2024-05-10 北京理工大学 Variable working condition small sample fault diagnosis method, device, medium and product based on information fusion element transfer learning

Similar Documents

Publication Publication Date Title
CN111898634B (en) Intelligent fault diagnosis method based on depth-to-reactance-domain self-adaption
CN109765053B (en) Rolling bearing fault diagnosis method using convolutional neural network and kurtosis index
CN109635677B (en) Compound fault diagnosis method and device based on multi-label classification convolutional neural network
CN112084974B (en) Multi-label rolling bearing fault diagnosis method based on meta-learning
CN111914883B (en) Spindle bearing state evaluation method and device based on deep fusion network
CN110657984B (en) Planetary gearbox fault diagnosis method based on reinforced capsule network
CN112434643A (en) Classification and identification method for low-slow small targets
CN114048769A (en) Multi-source multi-domain information entropy fusion and model self-optimization method for bearing fault diagnosis
CN112784920A (en) Cloud-side-end-coordinated dual-anti-domain self-adaptive fault diagnosis method for rotating part
CN113780242A (en) Cross-scene underwater sound target classification method based on model transfer learning
CN117009916A (en) Actuator fault diagnosis method based on multi-sensor information fusion and transfer learning
CN112749675A (en) Potato disease identification method based on convolutional neural network
CN114491828A (en) Compound fault diagnosis method based on multi-gating hybrid expert multi-task learning
CN113076920B (en) Intelligent fault diagnosis method based on asymmetric domain confrontation self-adaptive model
Shahid et al. Performance comparison of 1D and 2D convolutional neural networks for real-time classification of time series sensor data
CN116401603A (en) Multi-mode bearing fault intelligent diagnosis method based on transfer learning
CN114357372A (en) Aircraft fault diagnosis model generation method based on multi-sensor data driving
Rethik et al. Attention Based Mapping for Plants Leaf to Classify Diseases using Vision Transformer
CN117312948A (en) Rotary machine fault diagnosis method based on multi-sensor fusion and ECA-CNN
CN116399592A (en) Bearing fault diagnosis method based on channel attention dual-path feature extraction
CN116644273A (en) Fault diagnosis method and system based on interpretability multiplication convolution network
CN116070126A (en) Aviation plunger pump oil distribution disc abrasion detection method and system based on countermeasure self-supervision
CN115640531A (en) Fault diagnosis method based on residual error learning and attention mechanism fusion
Ojeda et al. Multivariate time series imaging for short-term precipitation forecasting using convolutional neural networks
CN113959719A (en) SOT-SVM-based variable working condition rolling bearing fault diagnosis method

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