CN116451133A - Heavy-duty robot speed reducer fault diagnosis method based on graph structure information - Google Patents
Heavy-duty robot speed reducer fault diagnosis method based on graph structure information Download PDFInfo
- Publication number
- CN116451133A CN116451133A CN202310323746.5A CN202310323746A CN116451133A CN 116451133 A CN116451133 A CN 116451133A CN 202310323746 A CN202310323746 A CN 202310323746A CN 116451133 A CN116451133 A CN 116451133A
- Authority
- CN
- China
- Prior art keywords
- layer
- graph
- data
- heavy
- node
- 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
- 238000000034 method Methods 0.000 title claims abstract description 76
- 239000003638 chemical reducing agent Substances 0.000 title claims abstract description 54
- 238000003745 diagnosis Methods 0.000 title claims abstract description 32
- 238000012549 training Methods 0.000 claims abstract description 17
- 238000012360 testing method Methods 0.000 claims abstract description 16
- 238000000547 structure data Methods 0.000 claims abstract description 11
- 230000007246 mechanism Effects 0.000 claims abstract description 8
- 238000002372 labelling Methods 0.000 claims abstract description 5
- 238000012423 maintenance Methods 0.000 claims abstract description 5
- 238000013528 artificial neural network Methods 0.000 claims abstract description 4
- 230000007704 transition Effects 0.000 claims description 27
- 238000010586 diagram Methods 0.000 claims description 15
- 230000006870 function Effects 0.000 claims description 15
- 238000012546 transfer Methods 0.000 claims description 15
- 239000011159 matrix material Substances 0.000 claims description 14
- 238000011176 pooling Methods 0.000 claims description 10
- 238000012545 processing Methods 0.000 claims description 10
- 238000005070 sampling Methods 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 6
- 230000002776 aggregation Effects 0.000 claims description 4
- 238000004220 aggregation Methods 0.000 claims description 4
- 230000004931 aggregating effect Effects 0.000 claims description 3
- 230000008033 biological extinction Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000012937 correction Methods 0.000 claims description 3
- 239000000203 mixture Substances 0.000 claims description 3
- 230000008859 change Effects 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000005299 abrasion Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000001174 ascending effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000007418 data mining Methods 0.000 description 1
- 238000012774 diagnostic algorithm Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 230000009347 mechanical transmission Effects 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000005236 sound signal Effects 0.000 description 1
Classifications
-
- 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
- G01M13/02—Gearings; Transmission mechanisms
- G01M13/028—Acoustic or vibration analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/10—Pre-processing; Data cleansing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
- G06N3/047—Probabilistic or stochastic networks
-
- 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
- G06N3/048—Activation functions
-
- 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
-
- 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
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Acoustics & Sound (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses a fault diagnosis method of a heavy-duty robot reducer based on graph structure information, which comprises the following steps: the method comprises the steps of collecting vibration data of the whole machine when a speed reducer of the heavy-duty robot normally operates and single faults occur, labeling all data, and carrying out downsampling pretreatment; converting the preprocessed data into graph structure data by a Markov field graph information method; dividing the graph structure data set into a training set and a testing set according to the proportion of 7:3; constructing a graph radix conservation attention network with residual connections, the network being a graph neural network based on a radix conservation attention mechanism; training a graph base number maintenance attention network with residual connection by using a training set until loss converges, and testing by using a testing set to obtain a final model; vibration data of the heavy-duty robot speed reducer is collected in real time, and is converted into graph structure information by using a Markov field graph information method, and the graph structure information is input into a final model to diagnose the state of the heavy-duty robot speed reducer.
Description
Technical Field
The invention relates to the technical field of fault diagnosis of industrial robots, in particular to a fault diagnosis method of a speed reducer of a heavy-duty robot based on graphic structure information.
Background
In recent years, with the rapid development of the mechanical industry, heavy-duty robot reducers have been widely used in various mechanical devices. The heavy-duty robot speed reducer is used as an important component of a mechanical transmission system and has the functions of torque transmission, speed reduction and speed change, steering change and the like. However, due to long-term operation of the heavy-duty robot reducer in severe environments such as high load, high speed, high temperature and the like, various faults such as gear abrasion, bearing abrasion, oil pollution and the like can occur due to the increase of the service life. These faults can lead to equipment damage, production accidents, increased production costs and other problems, and bring about serious losses to enterprises.
The fault diagnosis of the speed reducer of the heavy-duty robot is always one of hot spots in the research of the mechanical field. The traditional fault diagnosis method of the speed reducer of the heavy-duty robot is mainly used for diagnosing faults based on the change of physical quantities such as vibration signals, sound signals and the like. The method depends on professional detection equipment and detection personnel, and the method needs to stop for detection, so that the production efficiency is affected. Meanwhile, the traditional fault diagnosis method cannot accurately identify the fault type and position, and misdiagnosis and missed diagnosis are easy.
In recent years, with rapid development of data mining and machine learning technologies, a data-driven-based fault diagnosis method for a speed reducer of a heavy-duty robot has been attracting attention. According to the method, signals such as vibration, sound and current in the working state of the heavy-duty robot reducer are collected, and the characteristic parameters of the signals are analyzed, so that model training is performed by using a machine learning algorithm, and accurate, rapid and automatic fault diagnosis is realized.
However, most of the current data-driven-based fault diagnosis methods for the speed reducer of the heavy-duty robot still have some problems. First, these methods rely on a priori knowledge or model assumptions, which make it difficult to adapt to the diversified, complex failure modes of the heavy-duty robot decelerator. Secondly, these methods usually only focus on certain signal characteristics, and cannot describe the working state of the speed reducer of the heavy-duty robot comprehensively and accurately. Therefore, in order to improve the accuracy and reliability of the fault diagnosis of the speed reducer of the heavy-duty robot, a fault diagnosis method of the speed reducer of the heavy-duty robot based on the graph structure information needs to be studied.
Disclosure of Invention
In order to solve the technical problems, the invention aims to provide a fault diagnosis method of a heavy-duty robot reducer based on graph structure information, which converts an input vibration signal into a graph structure by a Markov field graph information method, can keep most of information in an original time sequence, can display state characteristics of a diagnosed object more clearly and comprehensively, and simultaneously performs fault diagnosis on the whole machine by collecting the vibration signal of the whole machine of the reducer.
The invention provides a fault diagnosis method of a heavy-duty robot reducer based on graph structure information, which comprises the following steps:
step 1: acquiring vibration data of the whole machine when the speed reducer of the heavy-duty robot normally operates and single faults occur, constructing a vibration data set, and labeling and downsampling all data;
step 2: converting the preprocessed data into graph structure data by a Markov field graph information method;
step 3: randomly scrambling the graph structure data set, and dividing the graph structure data set into a training set and a testing set according to the proportion of 7:3;
step 4: constructing a graph radix conservation attention network with residual connection, wherein the network is a graph neural network based on a radix conservation attention mechanism, and the attention mechanism based on radix conservation determines the sum of attention weights according to the number of neighbor nodes so as to reflect the connection relation between the nodes;
step 5: training a graph base number maintenance attention network with residual connection by using a training set until loss converges, and testing by using a testing set to obtain a final model;
step 6: vibration data of the heavy-duty robot speed reducer is collected in real time, and is converted into graph structure information by using a Markov field graph information method, and the graph structure information is input into a final model to diagnose the state of the heavy-duty robot speed reducer.
In the fault diagnosis method of the heavy-duty robot reducer based on the graph structure information, the single fault in the step 1 comprises the following steps: wave generator bearing failure, input shaft misalignment failure, soft gear tooth root pitting failure, and soft gear tooth tip pitting failure.
In the fault diagnosis method of the heavy-duty robot reducer based on the graph structure information, the step 1 specifically comprises the following steps:
step 1.1: respectively marking different labels on vibration data when normal operation and single fault occur by using onehot codes;
step 1.2: initializing an all-zero vector data1 with the length L1, and filling data1 according to the length L of the vibration data to enable the length of the vibration data after processing to be L1; namely, when L is more than or equal to L1, filling data1 by the first L1 elements of the vibration data; when L is smaller than L1, filling L elements of vibration data with data1, wherein the rest positions are zero;
step 1.3: defining a downsampled output matrix B N×K Sampling vibration data with the step length as m to sequentially fill matrix B with the sampled data N×K Each row contains K sampling data, and the downsampling processing of the vibration data is completed.
In the fault diagnosis method of the heavy-duty robot reducer based on the graph structure information, the step 2 specifically comprises the following steps:
step 2.1: quantifying vibration data, and setting the preprocessed vibration data as a time-series vibration signal X including n time points t ={x 1t ,x 2t ,…x it …,x nt Discretizing the vibration time sequence into B intervals according to the amplitude, and obtaining the data x of each time point it Mapped to the section number Q to which it belongs j Calculating the interval number Q according to j :
Wherein x is tmax ,x tmin Maximum amplitude and minimum amplitude of vibration data, j= [1,2,3 …, B];
Step 2.2: calculating interval probabilities, i.e. the number of occurrences of each interval and its probability in the whole time sequence:
where N is the number of time points in the time series, N j Represents the number of occurrences of the jth interval, P j Representing the probability of the jth interval in the whole time sequence, reflecting the frequency of occurrence of time sequence vibration data in the interval;
step 2.3: calculating transition probability, and counting each interval number Q j Number of occurrences N in the entire time series j The slave Q is calculated according to j Transfer to Q k Transition probability W of (2) jk :
Wherein N is jk Is Q j Transfer to Q k Number of transfers N of (2) jk I.e., the number of transitions from interval j to interval k, each time a pair of consecutive elements of the data are respectively located in two different intervals, the number of transitions from interval j to interval k increases by 1; w (w) jk Representing the slave Q j Transfer to Q k Reflecting the probability of a time seriesSimilarity of vibration signals between these two intervals, k= [1,2,3 …, B];
Step 2.4: the Markov transfer matrix W for the entire time series vibration data is obtained according to the following:
step 2.5: constructing a Markov field map using interval probabilities and transition probabilities, numbering each interval Q j Corresponding to a node, the attribute of the node is the probability P of the interval number j The method comprises the steps of carrying out a first treatment on the surface of the Constructing edges between nodes by using transition probabilities, wherein the weight of the edges is the transition probability w jk The method comprises the steps of carrying out a first treatment on the surface of the Obtain a Markov field diagram G t =[V t ,E t ]From a group of nodes V t And edge E t Composition; wherein each node represents an interval number Q j Each edge represents the transition probability w between two adjacent intervals jk 。
In the fault diagnosis method of the heavy-duty robot reducer based on the graph structure information, the graph base number keeping attention network with residual connection in the step 4 specifically comprises the following steps: an input layer, a CPAConv, a pooling layer and a full connection layer;
(1) Input layer: markov field map G from step 2 t =[V t ,E t ]As input;
(2) CPAConv layer: each CPAConv layer has h attention heads for calculating different attention weights, and for each pair of adjacent nodes u and v, an attention coefficient alpha is calculated uv The calculation formula of the leak ReLU nonlinear function as a weighted sum of node u and v features is as follows:
wherein,,is a learnable parameter, < >>Is the connection of hidden representations of the l-1 layer nodes u and v, and LeakyReLU is a leakage correction linear element activation function; aggregating the messages of the adjacent nodes by using an addition aggregation method to obtain a new feature representation ++of the first layer node u>
Wherein f l () Sum phi l () A nonlinear activation function; n (N) u Is a node set containing all neighbor nodes of node u, |N u I represents all cardinal information of node u;representing the hidden state of the node u in the l-1 layer; />Indicating the attention weight between the node u and its neighboring node v in the l-1 layer, as well as the element multiplication; for the l CPAConv layers, each layer processes the input feature matrix in sequence to generate new node feature representation;
(3) Pooling layer: extracting a graph level representation from the output of each CPAconv layer using a global pooling operation;
(4) Full tie layer: processing the graph level representations of all CPAConv layers through a full connection layer, and adding the graph level representations to obtain a final graph representation; the final graph representation is passed through a full connection layer and log-softmax function to obtain class probabilities.
In the heavy-duty robot decelerator fault diagnosis method based on the graph structure information of the present invention, a residual connection is added between each CPAConv layer for improving gradient extinction in the back propagation process, and the residual connection is defined as:
wherein,,and->Hidden representations of node u in layer l-1, layer l, and layer l+1, respectively; the residual connection adds the hidden representation of node u in layer l-1 to the hidden representation of node u in layer l as input, calculating the hidden expression of node u in layer l+1.
The fault diagnosis method of the heavy-duty robot speed reducer based on the graph structure information has the following beneficial effects:
(1) The problem of lack of interpretability of the current deep learning method is solved. The input vibration signal is converted into a graph structure by a Markov field graph information method, so that most of information in an original time sequence can be reserved, and the state characteristics of a diagnosed object can be displayed more clearly and comprehensively.
(2) The problem that fault diagnosis cannot be conducted on a complete heavy-load robot speed reducer at present is solved. The fault diagnosis method of the heavy-duty robot speed reducer based on the graph structure information constructs a graph structure by each component of the heavy-duty robot speed reducer and the interrelation thereof, and models and analyzes the working state of the heavy-duty robot speed reducer based on the graph structure. The method can effectively describe the association relation among all the parts of the speed reducer of the heavy-duty robot, thereby comprehensively and accurately describing the working state of the speed reducer of the heavy-duty robot. Meanwhile, the method can also adaptively learn the failure mode of the speed reducer of the heavy-duty robot, has better adaptability and expandability, has universality and higher accuracy, and can be used for the integral failure diagnosis of the speed reducer of the heavy-duty robot.
Drawings
FIG. 1 is a flow chart of a method for diagnosing faults of a speed reducer of a heavy-duty robot based on the structural information of the diagram;
FIG. 2 is a flow chart of the vibration data conversion diagram data of the Markov field diagram information method of the present invention;
fig. 3 is a diagram of a model structure of a graph base number keep-alive network (GCPAR network) with residual connection in the present invention.
Detailed Description
As shown in fig. 1, the fault diagnosis method of the heavy-duty robot reducer based on the graph structure information comprises the following steps:
step 1: acquiring vibration data of the whole machine when the speed reducer of the heavy-duty robot normally operates and single faults occur, constructing a vibration data set, and labeling and downsampling all data; wherein the single fault comprises: the method comprises the following steps of:
step 1.1: respectively marking different labels on vibration data when normal operation and single fault occur by using onehot codes;
the collected fault signals in 5 different states are marked with different labels by using onehot codes respectively, such as: normal state: [1, 0]; wave generator bearing failure: [0,1, 0]; input shaft misalignment failure: [0,1, 0]; soft tooth root pitting failure:
[0,1, 0]; soft tooth tip etch: [0,0,0,0,1].
After labeling, the signals are uniformly subjected to downsampling, and the step is used for carrying out noise reduction on the original data so as to better extract signal characteristics, and comprises the following specific steps:
step 1.2: initializing an all-zero vector data1 with the length L1, and filling data1 according to the length L of the vibration data to enable the length of the vibration data after processing to be L1; namely, when L is more than or equal to L1, filling data1 by the first L1 elements of the vibration data; when L is smaller than L1, filling L elements of vibration data with data1, wherein the rest positions are zero;
step 1.3: defining a downsampled output matrix B N×K Sampling vibration data with the step length as m to sequentially fill matrix B with the sampled data N×K Each row contains K sampling data, and the downsampling processing of the vibration data is completed.
In specific implementation, a new data sequence is obtained by taking a point every 47513 sampling points. Equivalent to reducing the sampling rate of the raw data by a factor of about 47513. Since this downsampling frequency is very low, the function is believed to act to noise-reduce the raw data to better extract the signal features.
Step 2: the preprocessed data is converted into graph structure data through a Markov field graph information method, as shown in fig. 2, and the method specifically comprises the following steps:
step 2.1: quantifying vibration data, and setting the preprocessed vibration data as a time-series vibration signal X including n time points t ={x 1t ,x 2t ,…x it …,x nt Discretizing the vibration time sequence into B intervals according to the amplitude, and obtaining the data x of each time point it Mapped to the section number Q to which it belongs j Calculating the interval number Q according to j :
Wherein x is tmax ,x tmin Maximum amplitude and minimum amplitude of vibration data, j= [1,2,3 …, B];=
For example, assume that a time series is discretized into 3 intervals (b=3), and x tmin =0,x tmax =10. Then the interval width is x tma -x tmin =10. If a certain point in time x it The value of (2) is 5.5, since 1/3<5.5/10<2/3 then it will be mapped to the 2 nd interval.
Step 2.2: calculating interval probabilities, i.e. the number of occurrences of each interval and its probability in the whole time sequence:
where N is the number of time points in the time series, N j Represents the number of occurrences of the jth interval, P j Representing the probability of the jth interval in the whole time sequence, reflecting the frequency of occurrence of time sequence vibration data in the interval;
continuing with the above example, if there is a time-series vibration signal x= [2,3,5,7,6,8,9,10,7,3 ] of length 10]And discretizes it into 3 intervals, then there are 1 time point in the 1 st interval, 5 time points in the 2 nd interval, and 4 time points in the 3 rd interval. Therefore, the probability of the 1 st interval is P 1 The probability of interval 2 is P =1/10=0.1 2 =5/10=0.5, probability of 3 rd interval is P 3 =4/10=0.4。
Step 2.3: calculating a transition probability matrix, and counting each interval number Q j Number of occurrences N in the entire time series j The slave Q is calculated according to j Transfer to Q k Transition probability W of (2) jk :
Wherein N is jk Is Q j Transfer to Q k Number of transfers N of (2) jk I.e., the number of transitions from interval j to interval k, each time a pair of consecutive elements of the data are respectively located in two different intervals, the number of transitions from interval j to interval k increases by 1; w (w) jk Representing the slave Q j Transfer to Q k Reflects the similarity of the time-series vibration signal between the two intervals, k= [1,2,3 …, B];
Step 2.4: the Markov transfer matrix W for the entire time series vibration data is obtained according to the following:
wherein w is 1B Representing a pair of consecutive elements x t And x t-1 Respectively located in two different intervals Q 1 And Q B At this time, the number of times of transition from section 1 to section B.
Step 2.5: constructing a Markov field map using interval probabilities and transition probabilities, numbering each interval Q j Corresponding to a node, the attribute of the node is the probability P of the interval number j The method comprises the steps of carrying out a first treatment on the surface of the Constructing edges between nodes by using transition probabilities, wherein the weight of the edges is the transition probability w jk The method comprises the steps of carrying out a first treatment on the surface of the Obtain a Markov field diagram G t =[V t ,E t ]From a group of nodes V t And edge E t Composition; wherein each node represents an interval number Q j Each edge represents the transition probability w between two adjacent intervals jk . By such a markov field map we can better understand the characteristics and structure of the vibration time series and can use the method of the map structure for further analysis and processing.
Step 3: randomly scrambling the graph structure data set, and dividing the graph structure data set into a training set and a testing set according to the proportion of 7:3;
step 4: constructing a graph radix conservation attention network (GCPAR network) with residual connection, wherein the network is a graph neural network based on a radix conservation attention mechanism, the attention mechanism based on radix conservation determines the sum of attention weights according to the number of neighbor nodes so as to reflect the connection relation between the nodes, and a network structure diagram is shown in figure 3;
in particular implementations, the graph base number keep-alive network with residual connection specifically includes: an input layer, a CPAConv, a pooling layer and a full connection layer;
(1) Input layer: markov field map G from step 2 t =[V t ,E t ]As input;
(2) CPAConv layer: each CPAConv layer has h attention heads for calculating different attention weights, and for each pair of adjacent nodes u and v, an attention coefficient alpha is calculated uv The calculation formula of the leak ReLU nonlinear function as a weighted sum of node u and v features is as follows:
wherein,,is a learnable parameter, < >>Is the connection of hidden representations of the l-1 layer nodes u and v, and LeakyReLU is a leakage correction linear element activation function; aggregating the messages of the adjacent nodes by using an addition aggregation method to obtain a new feature representation ++of the first layer node u>
Wherein f l () Sum phi l () A nonlinear activation function; n (N) u Is a node set containing all neighbor nodes of node u, |N u I represents all cardinal information of node u;representing the hidden state of the node u in the l-1 layer; />Indicating the attention weight between the node u and its neighboring node v in the l-1 layer, as well as the element multiplication; for the l CPAConv layers, each layer processes the input feature matrix in sequence to generate new node feature representation;
(3) Pooling layer: extracting a graph level representation from the output of each CPAconv layer using a global pooling operation;
(4) Full tie layer: processing the graph level representations of all CPAConv layers through a full connection layer, and adding the graph level representations to obtain a final graph representation; the final graph representation is passed through a full connection layer and log-softmax function to obtain class probabilities.
A residual connection was added between each CPAConv layer to improve the gradient extinction during back propagation, the residual connection being defined as:
wherein,,and->Hidden representations of node u in layer l-1, layer l, and layer l+1, respectively; the residual connection adds the hidden representation of node u in layer l-1 to the hidden representation of node u in layer l as input, calculating the hidden expression of node u in layer l+1.
Step 5: training a graph base number maintenance attention network with residual connection by using a training set until loss converges, and testing by using a testing set to obtain a final model;
the training process is as follows:
step 5.1: calculating an implicit representation of node u of the first layer according to equation (6)
Step 5.2: residual connection is carried out on the l-1 layer node u and the l layer node u according to formulas (7) and (8), so that a new hidden representation of the l layer node u is obtained
Step 5.3: steps 5.1 and 5.2 are circularly executed, and the node u is continuously updated through the neighbor nodes until the last layer is passed;
step 5.4: sequentially passing the output of each layer through a pooling layer, a batch normalization layer, a full connection layer and a relu function, adding all the outputs and obtaining a classification result through the full connection layer;
step 5.5: all training set data are used and recorded as one epoch at a time, and a trained final model is obtained after all epochs are completed.
Step 6: vibration data of the heavy-duty robot speed reducer are converted into graph structure information by using a Markov field graph information method, and the graph structure information is input into a final model to diagnose the state of the heavy-duty robot speed reducer.
The graph radix keeping attention network GCPAR with residual connection is developed based on CPA model of GNN, and an attention mechanism of keeping radix is used. It can assign different weights to different nodes during node information aggregation, enabling the network to self-focus on the most relevant nodes during training. To verify its superiority, the GCPAR model was compared with the other five GNN correlation algorithms, namely GCN model, attention-based SAGPool model, diffpool model, DGCNN model and GIN model.
We use a downsampling method to reduce the frequency of data acquisition so that the network can get more input. At the same time, the method can filter out mixed high-frequency noise signals in the data and reduce interference. The downsampling factor selected was 10, 10000 data samples were obtained from 1000 raw data samples after downsampling. After converting the 10000 data samples using the markov field map based information method, 10000 markov field maps were obtained and divided into A, B, C, D four data sets each containing 2500 markov field maps with corresponding rotational speeds of 500rpm, 1000rpm, 1500rpm and 2000 rpm. 10000 graphs of four speeds are combined together to form the data set E to verify the robustness of the proposed model. The five data sets are divided into a training set and a test set, respectively.
The average test accuracy for the five data sets is shown in table 1. The average test accuracy of the six models indicated that the GCN model was worst in either the single speed dataset A, B, C, D or the mixed speed dataset E. For single speed dataset a, the average test accuracy for the six models is lower than for the other single speed datasets B, C and D. This may be because the low frequency fault signal is disturbed by the fundamental frequency, the characteristics of the fault signal being insignificant. Experimental results show that the GCPAR model has superior accuracy when tested on all five data sets relative to the other five methods. Meanwhile, the GIN model and the DGCNN model also obtain relatively good classification results, and the effectiveness of using a Markov field diagram information method in fault diagnosis of the speed reducer of the heavy-duty robot is verified.
TABLE 1
The method of the invention adopts a Markov field diagram information method to convert the vibration signal into a diagram form, namely, one-dimensional vibration data is converted into two-dimensional diagram information in an ascending dimension, and partial noise can be separated in the process. Probability P of numbering sections of vibration data j And converting the node characteristics into the characteristics of the nodes in the Markov field diagram, constructing edges between the nodes by using transition probabilities, and representing the connectivity and edge characteristics of the nodes in the diagram structural data through a Markov transition matrix W. The characteristics of the nodes and edges determine the characteristics of the graph, and the connectivity (edges) of the nodes determine the graph topology. The spatial structure characteristics of the original vibration data are extracted in the conversion process, and the difference of the spatial structures of the original vibration data can be intuitively reflected through the difference of topological structures of the diagrams. And vibration data in different fault states have different spatial structures, so that the state of the speed reducer of the heavy-duty robot can be displayed explicitlyIs shown. This also allows the method of the present invention to have better interpretability than existing deep learning diagnostic algorithms. Compared with other methods for diagnosing all parts of the heavy-duty robot speed reducer, the method is wider in applicability. The method can be directly deployed on the heavy-duty robot reducer which works actually, and fault diagnosis is carried out on all components on the heavy-duty robot reducer.
The foregoing description of the preferred embodiments of the invention is not intended to limit the scope of the invention, but rather to enable any modification, equivalent replacement, improvement or the like to be made without departing from the spirit and principles of the invention.
Claims (6)
1. The fault diagnosis method of the heavy-duty robot speed reducer based on the graph structure information is characterized by comprising the following steps of:
step 1: acquiring vibration data of the whole machine when the speed reducer of the heavy-duty robot normally operates and single faults occur, constructing a vibration data set, and labeling and downsampling all data;
step 2: converting the preprocessed data into graph structure data by a Markov field graph information method;
step 3: randomly scrambling the graph structure data set, and dividing the graph structure data set into a training set and a testing set according to the proportion of 7:3;
step 4: constructing a graph radix conservation attention network with residual connection, wherein the network is a graph neural network based on a radix conservation attention mechanism, and the attention mechanism based on radix conservation determines the sum of attention weights according to the number of neighbor nodes so as to reflect the connection relation between the nodes;
step 5: training a graph base number maintenance attention network with residual connection by using a training set until loss converges, and testing by using a testing set to obtain a final model;
step 6: vibration data of the heavy-duty robot speed reducer is collected in real time, and is converted into graph structure information by using a Markov field graph information method, and the graph structure information is input into a final model to diagnose the state of the heavy-duty robot speed reducer.
2. The method for diagnosing a fault in a speed reducer of a heavy-duty robot based on the structural information of fig. 1, wherein the single fault in step 1 includes: wave generator bearing failure, input shaft misalignment failure, soft gear tooth root pitting failure, and soft gear tooth tip pitting failure.
3. The method for diagnosing faults of the speed reducer of the heavy-duty robot based on the graph structure information as set forth in claim 1, wherein the step 1 is specifically as follows:
step 1.1: respectively marking different labels on vibration data when normal operation and single fault occur by using onehot codes;
step 1.2: initializing an all-zero vector data1 with the length L1, and filling data1 according to the length L of the vibration data to enable the length of the vibration data after processing to be L1; namely, when L is more than or equal to L1, filling data1 by the first L1 elements of the vibration data; when L is smaller than L1, filling L elements of vibration data with data1, wherein the rest positions are zero;
step 1.3: defining a downsampled output matrix B N×K Sampling vibration data with the step length as m to sequentially fill matrix B with the sampled data N×K Each row contains K sampling data, and the downsampling processing of the vibration data is completed.
4. The fault diagnosis method for the heavy-duty robot decelerator based on the graphic structure information according to claim 1, wherein the step 2 is specifically:
step 2.1: quantifying vibration data, and setting the preprocessed vibration data as a time-series vibration signal X including n time points t ={x 1t ,x 2t ,…x it …,x nt Discretizing the vibration time sequence into B intervals according to the amplitude, and obtaining the data x of each time point it Mapped to the section number Q to which it belongs j Calculating the interval number Q according to j :
Wherein x is tmax ,x tmin The maximum amplitude and the minimum amplitude of the vibration data respectively,
j=[1,2,3…,B];
step 2.2: calculating interval probabilities, i.e. the number of occurrences of each interval and its probability in the whole time sequence:
where N is the number of time points in the time series, N j Represents the number of occurrences of the jth interval, P j Representing the probability of the jth interval in the whole time sequence, reflecting the frequency of occurrence of time sequence vibration data in the interval;
step 2.3: calculating transition probability, and counting each interval number Q j Number of occurrences N in the entire time series j The slave Q is calculated according to j Transfer to Q k Transition probability W of (2) jk :
Wherein N is jk Is Q j Transfer to Q k Number of transfers N of (2) jk I.e., the number of transitions from interval j to interval k, each time a pair of consecutive elements of the data are respectively located in two different intervals, the number of transitions from interval j to interval k increases by 1; w (w) jk Representing the slave Q j Transfer to Q k Reflects the similarity of the time-series vibration signal between the two intervals, k= [1,2,3 …, B];
Step 2.4: the Markov transfer matrix W for the entire time series vibration data is obtained according to the following:
step 2.5: constructing a Markov field map using interval probabilities and transition probabilities, numbering each interval Q j Corresponding to a node, the attribute of the node is the probability P of the interval number j The method comprises the steps of carrying out a first treatment on the surface of the Constructing edges between nodes by using transition probabilities, wherein the weight of the edges is the transition probability w jk The method comprises the steps of carrying out a first treatment on the surface of the Obtain a Markov field diagram G t =[V t ,E t ]From a group of nodes V t And edge E t Composition; wherein each node represents an interval number Q j Each edge represents the transition probability w between two adjacent intervals jk 。
5. The method for diagnosing faults of a speed reducer of a heavy-duty robot based on the structural information of the figure according to claim 1, wherein the figure base maintenance attention network with residual connection in the step 4 specifically comprises: an input layer, a CPAConv, a pooling layer and a full connection layer;
(1) Input layer: markov field map G from step 2 t =[V t ,E t ]As input;
(2) CPAConv layer: each CPAConv layer has h attention heads for calculating different attention weights, and for each pair of adjacent nodes u and v, an attention coefficient alpha is calculated uv The calculation formula of the leak ReLU nonlinear function as a weighted sum of node u and v features is as follows:
wherein,,is a learnable parameter, < >>Is the connection of hidden representations of the l-1 layer nodes u and v, and LeakyReLU is a leakage correction linear element activation function; aggregating the messages of the adjacent nodes by using an addition aggregation method to obtain a new feature representation ++of the first layer node u>
Wherein f l () Sum phi l () A nonlinear activation function; n (N) u Is a node set containing all neighbor nodes of node u, |N u I represents all cardinal information of node u;representing the hidden state of the node u in the l-1 layer; />Indicating the attention weight between the node u and its neighboring node v in the l-1 layer, as well as the element multiplication; for the l CPAConv layers, each layer processes the input feature matrix in sequence to generate new node feature representation;
(3) Pooling layer: extracting a graph level representation from the output of each CPAconv layer using a global pooling operation;
(4) Full tie layer: processing the graph level representations of all CPAConv layers through a full connection layer, and adding the graph level representations to obtain a final graph representation; the final graph representation is passed through a full connection layer and log-softmax function to obtain class probabilities.
6. The heavy duty robot decelerator fault diagnosis method based on the graph structure information of claim 5, wherein a residual connection is added between each CPAConv layer for improving gradient extinction during back propagation, the residual connection being defined as:
wherein,,and->Hidden representations of node u in layer l-1, layer l, and layer l+1, respectively; the residual connection adds the hidden representation of node u in layer l-1 to the hidden representation of node u in layer l as input, calculating the hidden expression of node u in layer l+1.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310323746.5A CN116451133A (en) | 2023-03-30 | 2023-03-30 | Heavy-duty robot speed reducer fault diagnosis method based on graph structure information |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310323746.5A CN116451133A (en) | 2023-03-30 | 2023-03-30 | Heavy-duty robot speed reducer fault diagnosis method based on graph structure information |
Publications (1)
Publication Number | Publication Date |
---|---|
CN116451133A true CN116451133A (en) | 2023-07-18 |
Family
ID=87119467
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310323746.5A Pending CN116451133A (en) | 2023-03-30 | 2023-03-30 | Heavy-duty robot speed reducer fault diagnosis method based on graph structure information |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116451133A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117892188A (en) * | 2023-12-19 | 2024-04-16 | 济南大学 | Fault classification method based on bearing vibration signal relation and graph neural network |
CN118051860A (en) * | 2024-04-15 | 2024-05-17 | 西安瓦力机电科技有限公司 | Intelligent high-precision gear reducer monitoring and diagnosing system |
-
2023
- 2023-03-30 CN CN202310323746.5A patent/CN116451133A/en active Pending
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117892188A (en) * | 2023-12-19 | 2024-04-16 | 济南大学 | Fault classification method based on bearing vibration signal relation and graph neural network |
CN118051860A (en) * | 2024-04-15 | 2024-05-17 | 西安瓦力机电科技有限公司 | Intelligent high-precision gear reducer monitoring and diagnosing system |
CN118051860B (en) * | 2024-04-15 | 2024-06-07 | 西安瓦力机电科技有限公司 | Intelligent high-precision gear reducer monitoring and diagnosing system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111709448B (en) | Mechanical fault diagnosis method based on migration relation network | |
CN116451133A (en) | Heavy-duty robot speed reducer fault diagnosis method based on graph structure information | |
CN112906644B (en) | Mechanical fault intelligent diagnosis method based on deep migration learning | |
Si et al. | Unsupervised deep transfer learning with moment matching: A new intelligent fault diagnosis approach for bearings | |
CN111562108A (en) | Rolling bearing intelligent fault diagnosis method based on CNN and FCMC | |
Li et al. | Joint attention feature transfer network for gearbox fault diagnosis with imbalanced data | |
CN113670616B (en) | Bearing performance degradation state detection method and system | |
CN116358871B (en) | Rolling bearing weak signal composite fault diagnosis method based on graph rolling network | |
CN116028876A (en) | Rolling bearing fault diagnosis method based on transfer learning | |
Lv et al. | Deep transfer network with multi-kernel dynamic distribution adaptation for cross-machine fault diagnosis | |
CN113537152A (en) | Flow field state fault detection method based on deep neural network | |
CN114462446A (en) | Rolling bearing fault diagnosis method based on vibration signal and electronic equipment | |
CN114819315A (en) | Bearing degradation trend prediction method based on multi-parameter fusion health factor and time convolution neural network | |
CN116702076A (en) | Small sample migration learning fault diagnosis method, system, computer and storage medium based on CNN feature fusion | |
CN115290326A (en) | Rolling bearing fault intelligent diagnosis method | |
Tang et al. | An efficient sequential embedding ConvNet for rotating machinery intelligent fault diagnosis | |
Li et al. | A fault diagnosis method for rotating machinery with semi-supervised graph convolutional network and images converted from vibration signals | |
Zhang et al. | Graph neural network-based bearing fault diagnosis using Granger causality test | |
CN113310689B (en) | Aeroengine transmission system fault diagnosis method based on domain self-adaptive graph convolution network | |
CN117056865B (en) | Method and device for diagnosing operation faults of machine pump equipment based on feature fusion | |
Han et al. | Deep subclass alignment transfer network based on time–frequency features for intelligent fault diagnosis of planetary gearboxes under time-varying speeds | |
Li et al. | Knowledge features enhanced intelligent fault detection with progressive adaptive sparse attention learning for high-power diesel engine | |
CN116738367A (en) | Rolling bearing fault diagnosis method based on multi-scale map transducer | |
CN114548153B (en) | Planetary gear box fault diagnosis method based on residual error-capsule network | |
CN117668623B (en) | Multi-sensor cross-domain fault diagnosis method for leakage of ship pipeline valve |
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 |