CN110333074A - Multi-measuring point drive failure diagnostic method and system based on convolutional neural networks - Google Patents
Multi-measuring point drive failure diagnostic method and system based on convolutional neural networks Download PDFInfo
- Publication number
- CN110333074A CN110333074A CN201910692350.1A CN201910692350A CN110333074A CN 110333074 A CN110333074 A CN 110333074A CN 201910692350 A CN201910692350 A CN 201910692350A CN 110333074 A CN110333074 A CN 110333074A
- Authority
- CN
- China
- Prior art keywords
- neural networks
- convolutional neural
- measuring point
- layer
- data
- 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.)
- Granted
Links
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
- 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
- G06N3/045—Combinations of 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
- G06N3/08—Learning methods
Abstract
The present invention provides a kind of multi-measuring point drive failure diagnostic method and system based on convolutional neural networks, solves existing efficiency of fault diagnosis and accuracy rate is not suitable with the technical issues of transmission mechanism develops.Method includes: optimization vibrating sensor installation position, forms parallel vibration data corresponding with vibrating sensor;The convolutional layer, pond layer and full articulamentum for establishing initial convolutional neural networks adjust initial convolutional neural networks and verify the fault diagnosis precision of initial convolutional neural networks, and adjust the degree of fitting of initial convolutional neural networks, form fault diagnosis model.The automated analysis process to the numerous measuring point datas of complicated transmission mechanism is formd, noise information can automatically, be accurately filtered out and isolates fault signature;It, can be by adjusting the purpose of the scale adaptation processing large-scale data of network, to characterize complex mapping relation for a large amount of data to be processed.Meet the efficiency and reliability of complicated transmission mechanism Life cycle internal fault early warning.
Description
Technical field
The present invention relates to fault diagnosis signal processing technology fields, and in particular to a kind of more surveys based on convolutional neural networks
Point drive failure diagnostic method and system.
Background technique
In multi-measuring point drive failure diagnostic field, the existing fault diagnosis technology fortune current by monitoring transmission mechanism
The position and analysis failure Producing reason, prevention that row state detection failure occurs generate catastrophic effect.Currently used event
Hinder diagnostic techniques process are as follows: firstly, carrying out fault signature extraction by signal processing technology;Then, using machine learning techniques
Carry out fault diagnosis.The technology has four: one, due to the complexity of transmission mechanism itself, needing to set in different location
It sets multiple measuring points and is monitored simultaneously, and the effective fault diagnosis of comprehensive multiple measuring point information progress is very difficult;Two, it needs
Operator has the signal processing background knowledge of profession, and the internal structure in conjunction with transmission mechanism is artificially analyzed and extracts spy of being out of order
Levy the processing for causing inefficiency to be unsuitable for a large amount of characteristics;Three, as transmission mechanism is towards high speed, high-precision, efficient direction
Development, internal structure become increasingly complex, and collected data cover the bulk information that different physical resources give off, and are mingled with
Noise information, traditional signal processing technology be difficult accurately and effectively to separate fault signature that extract accurate failure special
Sign;Four, increasing with equipment measuring point, the increase of sample frequency begin to use the data collection to end-of-life to go through from mechanism
Duration data volume is huge, and traditional signal processing technology model is difficult collected information and equipment event in the case of characterization big data
Complicated mapping relations between barrier situation cause finally to influence fault diagnosis precision.Therefore need one kind can be from multi-measuring point, big
The fault diagnosis technology of accurate adaptive extraction fault signature in scale complex data.In the prior art, convolutional neural networks
The network structure having has well adapting to property to input data, can carry out feature extraction and classification, these features simultaneously
There is a possibility that solving above-mentioned data analyzing defect.
Summary of the invention
In view of the above problems, the embodiment of the present invention provides a kind of multi-measuring point drive failure based on convolutional neural networks
Diagnostic method and system, solve existing multi-measuring point drive failure diagnosis efficiency and accuracy rate is not suitable with transmission mechanism development
Technical problem.
The multi-measuring point drive failure diagnostic method based on convolutional neural networks of the embodiment of the present invention, comprising:
Optimize vibrating sensor installation position, forms parallel vibration data corresponding with the vibrating sensor;
The convolutional layer, pond layer and full articulamentum for establishing initial convolutional neural networks utilize the initial convolution nerve net
The network processing parallel vibration data;
The initial convolutional neural networks are adjusted in parallel vibration data treatment process and verify the initial convolution mind
Fault diagnosis precision through network, and the degree of fitting of the initial convolutional neural networks is adjusted, form fault diagnosis model.
In one embodiment of the invention, the optimization vibrating sensor installation position is formed corresponding with the vibrating sensor
Parallel vibration data include:
Determine the crucial measurement position of transmission mechanism;
The vibrating sensor of measuring point is laid in the crucial measurement position;
It is grouped according to vibrating sensor of the freedom degree to the measuring point, phase is carried out to the measuring point vibration data in same group
Coefficients comparison is closed, and related coefficient reliability is determined according to significant difference method of inspection;
Whether the related coefficient for judging each measuring point in the same group is more than correction threshold;
The related coefficient is less than the correction threshold and then determines measuring point and corresponding vibration in the same group
Data;
The related coefficient is more than that then the measuring point high to related coefficient carries out the correction threshold in the same group
Single screening;
Measuring point and vibration data described in determining each group corresponding with crucial measurement position, repeatedly related coefficient compare-it is mono-
One screening process completes the measuring point screening.
In one embodiment of the invention, the correction threshold of the related coefficient is 80%.
In one embodiment of the invention, further includes:
It establishes data time sequence-and the vibration data-of each measuring point is encoded according to the frequency acquisition of vibration data according to data time sequence
The time series data-for forming each measuring point carries out zero-mean normalized to time series data, forms the time series data of each measuring point
Parallel input of one data channel as convolutional neural networks.
In one embodiment of the invention, the convolutional layer for establishing initial convolutional neural networks, pond layer and full articulamentum, benefit
The parallel vibration data described in the initial convolution Processing with Neural Network includes:
Construct the first convolutional layer and the first pond layer: the convolution kernel size of the first convolutional layer is 1 × 5 × n, and n indicates measuring point
Number, convolution kernel number are 64;The size of the pond window of first pond layer is 1 × 1 × 2 × 1, step-length 2;
Construct the second convolutional layer and the second pond layer: the convolution kernel size of the second convolutional layer is 1 × 5 × 64, convolution nucleus number
Mesh is 64;The pond window size of second pond layer is 1 × 1 × 2 × 1, step-length 2;
Construct third convolutional layer and third pond layer: the convolution kernel size of third convolutional layer is 1 × 5 × 64, convolution nucleus number
Mesh is 64;The pond window size of third pond layer is 1 × 1 × 2 × 1, step-length 2;
Construct Volume Four lamination and the 4th pond layer: the convolution kernel size of Volume Four lamination is 1 × 5 × 64, convolution nucleus number
Mesh is 64;The pond window size of 4th pond layer is 1 × 1 × 2 × 1, step-length 2;
Construct the 5th convolutional layer and the 5th pond layer: the convolution kernel size of the 5th convolutional layer is 1 × 5 × 64, convolution nucleus number
Mesh is the data category number of training data, and the pond window size of the 5th pond layer is 1 × 1 × 750 × 1, step-length 750;
It forms full articulamentum and handles the matrix data after the convolutional layer and pond layer operation, and exported by softmax
Layer forms fault diagnosis classification results.
In one embodiment of the invention, further includes:
Error back propagation is formed using cross entropy, while classification check is carried out to fault diagnosis classification results.
It is described that the initial convolutional neural networks are adjusted in parallel vibration data treatment process in one embodiment of the invention
And the fault diagnosis precision of the initial convolutional neural networks is verified, and adjust the degree of fitting of the initial convolutional neural networks,
Forming fault diagnosis model includes:
Gradient vector combination gradient descent method is formed to each convolutional layer and pond using cross entropy in data processing
The weight parameter of layer carries out the iteratively faster adjustment of initial convolutional neural networks;
Judged whether to meet fault diagnosis precision according to previous iteration;
Judge whether over-fitting occur when meeting the fault diagnosis precision;
Judge whether poor fitting occur when being unsatisfactory for the fault diagnosis precision.
In one embodiment of the invention, further includes:
The convolutional layer and pond layer are adjusted by increasing regular terms when meeting the fault diagnosis precision and over-fitting
Weight parameter quantity;
The power of the convolutional layer and pond layer is formed by error back propagation when being unsatisfactory for the fault diagnosis precision
Weight parameter adaptive adjustment;
Pass through the size of the adjusting convolutional layer and pond layer when being unsatisfactory for the fault diagnosis precision and poor fitting
The complexity of the initial convolutional neural networks is improved with number.
The multi-measuring point drive failure diagnostic system based on convolutional neural networks of the embodiment of the present invention, feature exist
In, comprising:
Memory, for storing the above-mentioned multi-measuring point drive failure diagnostic method processing based on convolutional neural networks
The program code of process;
Processor, for executing said program code.
The multi-measuring point drive failure diagnostic system based on convolutional neural networks of the embodiment of the present invention, comprising:
Sensor optimization device is formed corresponding with the vibrating sensor for optimizing vibrating sensor installation position
Parallel vibration data;
Convolutional neural networks establish device, for establishing the convolutional layer, pond layer and full connection of initial convolutional neural networks
Layer, utilizes parallel vibration data described in the initial convolution Processing with Neural Network;
Fault diagnosis model optimizes device, for adjusting the initial convolutional Neural in parallel vibration data treatment process
Network and the fault diagnosis precision for verifying the initial convolutional neural networks, and adjust the fitting of the initial convolutional neural networks
Degree forms fault diagnosis model.
A kind of multi-measuring point drive failure diagnostic method and system based on convolutional neural networks of the embodiment of the present invention
The automated analysis process to the numerous measuring point datas of complicated transmission mechanism is formd using convolutional neural networks, for vibration data
Regular recessive traits carry out automated analysis, pass through the suitable of the design of the functional layer of fault diagnosis model and layer weight parameter
Match, automated analysis is allowed to form efficient stealthy data characteristics extraction and classification for the vibration data of magnanimity, and
Form data mapping relations corresponding with fault state.It can automatically, accurately filter out noise information and isolate fault signature;
It, can be by adjusting the purpose of the scale adaptation processing large-scale data of network, to characterize for a large amount of data to be processed
Complex mapping relation between data information and fault state.Meet the effect of complicated transmission mechanism Life cycle internal fault early warning
Rate and reliability.
The technical program carries out fault signature extraction, then carries out fault diagnosis, be during which not required to very important person according to input data
To intervene, does not need to carry out signal processing to data yet, greatly reduce the requirement to operator's signal processing technology;The skill
Art, according to the value of parameters in the adaptive adjustment model of input data, is carried out using fault diagnosis accuracy rate as final goal
Fault signature extracts and fault diagnosis, by adjusting parameter value repeatedly, reaches defined fault diagnosis accuracy rate.For any
The similar troubleshooting issue for needing to establish complex mapping relation between characterization failure information and fault state can apply it
Principle.
Detailed description of the invention
Fig. 1 show multi-measuring point drive failure diagnostic method of the one embodiment of the invention based on convolutional neural networks
Configuration diagram.
Fig. 2 show one embodiment of the invention based in the multi-measuring point drive failure diagnostic method of convolutional neural networks
The method flow schematic diagram of tested spot optimization.
Fig. 3 show one embodiment of the invention based in the multi-measuring point drive failure diagnostic method of convolutional neural networks
Establish the method flow schematic diagram of initial convolutional neural networks.
Fig. 4 show one embodiment of the invention based in the multi-measuring point drive failure diagnostic method of convolutional neural networks
Adjust the method flow schematic diagram of initial convolutional neural networks degree of fitting.
Fig. 5 show multi-measuring point drive failure diagnostic system of the one embodiment of the invention based on convolutional neural networks
Configuration diagram.
Specific embodiment
To be clearer and more clear the objectives, technical solutions, and advantages of the present invention, below in conjunction with attached drawing and specific embodiment party
The invention will be further described for formula.Obviously, described embodiments are only a part of the embodiments of the present invention, rather than all
Embodiment.Based on the embodiments of the present invention, those of ordinary skill in the art institute without creative efforts
The every other embodiment obtained, shall fall within the protection scope of the present invention.
Multi-measuring point drive failure diagnostic method such as Fig. 1 institute based on convolutional neural networks of one embodiment of the invention
Show, in Fig. 1, the present embodiment includes:
Step 100: optimization vibrating sensor installation position forms parallel vibration data corresponding with vibrating sensor.
Vibrating sensor installation position is corresponding with tested point, and the correlation and difference of data are acquired according to vibrating sensor
Conspicuousness carries out the iteration optimization that vibrating sensor lays quantity and installation position, can make the position characteristic of vibrating sensor
Not only the vibration information of complicated transmission mechanism can have more fully been fed back, but also the concurrently acquisition data flow simplified can be formed.
Step 200: establishing the convolutional layer, pond layer and full articulamentum of initial convolutional neural networks, utilize initial convolution mind
Through the parallel vibration data of network processes.
Using convolutional layer adaptation data input channel and the recessive vibration performance in input data is acquired, it is right using pond layer
Recessive vibration performance data carry out the simplification of necessary dimensionality reduction and common trait, using full articulamentum by recessive vibration performance data
Form the Type mapping between vibration performance and failure identification.
Step 300: adjusting initial convolutional neural networks in parallel vibration data treatment process and verify initial convolution mind
Fault diagnosis precision through network, and the degree of fitting of initial convolutional neural networks is adjusted, form fault diagnosis model.
Adjust the weight parameter of each layer of initial convolutional neural networks and using verifying input data be iterated verifying until
Meet fault diagnosis required precision.It is adjusted in each layer number and layer according to the degree of fitting defect of initial convolutional neural networks simultaneously
Feature weight number of parameters is iterated verifying using wider verifying input data and wants until meeting fault diagnosis precision
It asks, ultimately forms the fault diagnosis model of convolutional neural networks.
The multi-measuring point drive failure diagnostic method based on convolutional neural networks of the embodiment of the present invention utilizes convolution mind
The automated analysis process to the numerous measuring point datas of complicated transmission mechanism is formd through network, it is hidden for the regularity of vibration data
Property characteristic carry out automated analysis, by the adaptation of the functional layer of fault diagnosis model design and layer weight parameter, so that automatically
Efficient stealthy data characteristics extraction and classification can be formed for the vibration data of magnanimity by changing analysis, and be formed and failure shape
The corresponding data mapping relations of condition.It can automatically, accurately filter out noise information and isolate fault signature;For largely wait locate
The data of reason, can be by adjusting the purpose of the scale adaptation processing large-scale data of network, thus characterize data information and event
Complex mapping relation between barrier situation.Meet the efficiency and reliability of complicated transmission mechanism Life cycle internal fault early warning.
According to input data, fault signature extraction is carried out, fault diagnosis is then carried out, does not during which need human intervention, also not needed pair
Data carry out signal processing, greatly reduce the requirement to operator's signal processing technology;The technology is accurate with fault diagnosis
Rate is final goal, according to the value of parameters in the adaptive adjustment model of input data, carry out fault signature extraction and
Fault diagnosis reaches defined fault diagnosis accuracy rate by adjusting parameter value repeatedly.Characterization is established for any similar needs
The troubleshooting issue of complex mapping relation can apply its principle between fault message and fault state.
Measured point in the multi-measuring point drive failure diagnostic method based on convolutional neural networks of one embodiment of the invention
Bit optimization is as shown in Figure 2.In Fig. 2, tested spot optimization includes:
Step 110: determining the crucial measurement position of transmission mechanism.
Crucial measurement position include but is not limited to transmission mechanism bearing, gearbox, transmission shaft, differential mechanism and semiaxis, belt,
The base position of the transmission parts such as chain, gear, cam and connecting rod or link position.
Step 120: laying the vibrating sensor of measuring point in crucial measurement position.
Vibrational degrees of freedom includes at least lateral, axial, vertical three directions, therefore the same crucial measurement position can wrap
Measuring point corresponding with freedom degree, and the vibrating sensor of corresponding measuring point are included, the corresponding tested point of each vibrating sensor is (i.e.
Measuring point).Vibrating sensor can use linear oscillator acceleration transducer, and each vibrating sensor forms one group of measuring point data.
Step 130: the vibrating sensor of measuring point is grouped according to freedom degree, to the measuring point vibration data in same group into
Correlation series compare, and determine related coefficient reliability according to significant difference method of inspection.
Compared by the correlation between measuring point vibration data related between examining determining sensing data to significant difference
Property and correlation the degree of reliability, establish in same Freedom Types group between measuring point correlation quantization basis.
Step 140: whether the related coefficient for judging each measuring point in same group is more than correction threshold.
Correction threshold is preferred are as follows: related coefficient is greater than 0.8 and has significant difference, for being more than the measuring point of correction threshold
It is marked.
Step 150: related coefficient is less than correction threshold and then determines measuring point and corresponding vibration data in same group.
Related coefficient be less than the measuring point of correction threshold vibration data it is corresponding with measuring point can form data flow, as volume
The input traffic arranged side by side of product neural network.
Step 160: related coefficient be more than correction threshold then the measuring point high to related coefficient in same group carry out it is single
Screening.
The measuring point excessively high for the degree of correlation in same group is screened out from it as one alternative group with physics general character
Unique measuring point retains corresponding vibration data, while excluding other measuring points and corresponding vibration data in alternative organize.
Step 170: determining each group measuring point corresponding with crucial measurement position and vibration data repeat step 130 to step
Rapid 160 related coefficient compares-and single screening process completes measuring point screening.
On the basis of carrying out the single screening of related measuring point in same group, measuring point grouping, vibration data phase are re-started
Coefficients comparison and the screening of further measuring point are closed, each group measuring point related coefficient is reached by iterative manner and is less than correction threshold,
The optimization of sensor installation position is formed, determines the final point position of transmission mechanism.Vibration data is corresponding with measuring point to be formed
Data flow, the input data channel as convolutional neural networks.
The multi-measuring point drive failure diagnostic method based on convolutional neural networks of the embodiment of the present invention passes through sensor
Optimization reduces the redundancy of acquisition information, while each measuring point can be regarded as to a channel, supports multi-channel data defeated
Out, the data transmission problems of comprehensive multiple measuring point information are solved, it can be directly using collected vibration data as convolution mind
Input through network.
It further include on that basi of the above embodiments following suitable to each group measuring point and vibration data in one embodiment of the invention
Sequence preprocessing process:
Step 180: establishing data time sequence-according to data time sequence and encode the vibration data-of each measuring point according to vibration data
The time series data-that frequency acquisition forms each measuring point, which carries out zero-mean normalized-to time series data, makes the timing of each measuring point
Data form parallel input of the data channel as convolutional neural networks.
The acquisition data of the multi-measuring point drive failure diagnostic method based on convolutional neural networks of the embodiment of the present invention
The substantially dynamic change that parallel sequential data are adapted to measuring point collection vibration data amount is formed by pretreatment, is conducive to adaptation volume
The scale adjustment of product neural network.
It is established just in the multi-measuring point drive failure diagnostic method based on convolutional neural networks of one embodiment of the invention
Beginning convolutional neural networks are as shown in Figure 3.In Fig. 3, the foundation of initial convolutional neural networks includes:
Step 210: the first convolutional layer of building and the first pond layer: the convolution kernel size of the first convolutional layer is 1 × 5 × n, volume
Product nucleus number mesh is 64;The size of the pond window of first pond layer is 1 × 1 × 2 × 1, step-length 2.N indicates measure-point amount,
In one embodiment of the invention, preferably 6.
The vibration data of input obtains 1 × m after the operation of the first convolutional layer1The matrix data of × 64 sizes, matrix data
The first pond layer operation is carried out after carrying out zero-mean normalization, obtains 1 × m2× 64 matrix data, matrix data is as next
The input data of layer convolutional layer.m1Indicate sensor sample frequency, m2Corresponding pond is indicated as a result, in one embodiment of the invention
In, m1Preferably 12kHz, corresponding m2For 6k.
Step 220: the second convolutional layer of building and the second pond layer: the convolution kernel size of the second convolutional layer is 1 × 5 × 64,
Convolution kernel number is 64;The pond window size of second pond layer is 1 × 1 × 2 × 1, step-length 2.
The vibration data of input obtains 1 × m after the second convolutional layer and the second pond layer operation3× 64 matrix data,
Input data of the matrix data as next layer of convolutional layer.In an embodiment of the present invention, m3For 3k.
Step 230: building third convolutional layer and third pond layer: the convolution kernel size of third convolutional layer is 1 × 5 × 64,
Convolution kernel number is 64;The pond window size of third pond layer is 1 × 1 × 2 × 1, step-length 2.
The vibration data of input obtains 1 × m after third convolutional layer and third pond layer operation4× 64 matrix data,
Input data of the matrix data as next layer of convolutional layer.In an embodiment of the present invention, m4For 1.5k.
Step 240: building Volume Four lamination and the 4th pond layer: the convolution kernel size of Volume Four lamination is 1 × 5 × 64,
Convolution kernel number is 64;The pond window size of 4th pond layer is 1 × 1 × 2 × 1, step-length 2.
The vibration data of input obtains 1 × m after Volume Four lamination and the 4th pond layer operation5× 64 matrix data,
Input data of the matrix data as next layer of convolutional layer.In an embodiment of the present invention, m5It is 750.
Step 250: the 5th convolutional layer of building and the 5th pond layer: the convolution kernel size of the 5th convolutional layer is 1 × 5 × 64,
Convolution kernel number is the data category number of training data, and the pond window size of the 5th pond layer is 1 × 1 × 750 × 1, step-length
It is 750.In one embodiment of the invention, convolution kernel number is 4.
Convolution kernel number is corresponding with the fault type of data, i.e. totally 3 class fault datas and normal data.The vibration number of input
According to 1 × 1 × 4 matrix data is obtained after the 5th convolutional layer and the 5th pond layer operation, matrix data is as full articulamentum
Input data.
Step 260: the matrix data after forming full articulamentum processing convolutional layer and pond layer operation, and pass through softmax
Output layer forms fault diagnosis classification results.
The probability for being directed to recessive character is formed by softmax output layer, can obtain the probability of each fault type
Data, cooperation realize that the quantization of classification results is explained.
The multi-measuring point drive failure diagnostic method based on convolutional neural networks of the embodiment of the present invention, which forms, to be used for
The convolutional neural networks of vibration fault type judgement, realize the automatically entering of vibration data, the stealth characteristics in vibration data
Automatically extract, stealth characteristics automatically analyze and fault type recognition, preferably realize characterization failure information and fault state
Between complex mapping relation accuracy rate.
In one embodiment of the invention, as shown in figure 3, on that basi of the above embodiments, further includes:
Step 270: forming error back propagation using cross entropy, while classification school is carried out to fault diagnosis classification results
It tests.
The multi-measuring point drive failure diagnostic method based on convolutional neural networks of the embodiment of the present invention is by softmax layers
Fault signature classification is carried out to input data, the true classification of fault diagnosis result and input data is compared, is missed using cross entropy
Poor back propagation adaptively adjusts the value of each convolution layer parameter.The process does not need manually to participate in, and fault diagnosis model is complete
Each layer weight parameter of model is adjusted according to the information of input data, until reaching fault diagnosis precision.It is necessary to avoid operator
It grasps signal processing technology and can accurately extract the artificial treatment defect of fault signature.
The multi-measuring point drive failure diagnostic method degree of fitting based on convolutional neural networks of one embodiment of the invention
Adjustment is as shown in Figure 4.The adjustment of degree of fitting includes: in Fig. 4
Step 310: forming gradient vector combination gradient descent method to each convolution using cross entropy in data processing
The weight parameter of layer and pond layer carries out the iteratively faster adjustment of initial convolutional neural networks.
Error back propagation, weight parameter derivation of the cross entropy to interlayer can be formed using cross entropy as loss function
The gradient decline compared with steep slope can be formed, the training effect for improving data processing is conducive to.
Step 320: being judged whether to meet fault diagnosis precision according to previous iteration.
Initial convolutional neural networks adaptively adjust each layer weight parameter through input data training in data processing
Fault diagnosis precision is gradually increased, according to whether meeting fault diagnosis precision determines whether fault diagnosis model is formed.
Step 330: judging whether over-fitting occur when meeting fault diagnosis precision.
Over-fitting shows that the fault diagnosis model to be formed is only applicable to part input data set, and there are model defects.
Step 340: judging whether poor fitting occur when being unsatisfactory for fault diagnosis precision.
Poor fitting shows that fault diagnosis model fails the stealth characteristics for being preferably fitted input data, and there are model defects.
Step 350: adjusting each convolutional layer and pond by increasing regular terms when meeting fault diagnosis precision and over-fitting
The weight parameter quantity of layer.
Simplify the weight parameter quantity of each convolutional layer and pond layer, using regular terms to adapt to bigger data set.
Step 360: each convolutional layer and pond layer being formed by error back propagation when being unsatisfactory for fault diagnosis precision
Weight parameter adaptively adjusts.
Each layer weight parameter is formed using the optimization of error back propagation adaptively to adjust.
Step 370: when being unsatisfactory for fault diagnosis precision and poor fitting by adjusting the big of each convolutional layer and pond layer
Small and number improves the complexity of initial convolutional neural networks.
The multi-measuring point drive failure diagnostic method based on convolutional neural networks of the embodiment of the present invention passes through technology hand
Section is exercised supervision and is adjusted to the self-adaptive processing process of initial convolutional neural networks, and convolutional neural networks model is led to
The mode for crossing change model complexity solves the problems, such as that big data is handled, and establishes complicated between characterize data information and fault state
Mapping relations, obtain final fault diagnosis model.
The multi-measuring point drive failure diagnostic system based on convolutional neural networks of one embodiment of the invention, comprising:
Memory, for storing the multi-measuring point drive failure diagnosis side based on convolutional neural networks of above-described embodiment
The program code of method treatment process;
Processor, for executing the multi-measuring point drive failure diagnosis side based on convolutional neural networks of above-described embodiment
The program code of method treatment process.
Processor can use DSP (Digital Signal Processing) digital signal processor, FPGA
(Field-Programmable Gate Array) field programmable gate array, MCU (Microcontroller Unit) system
Plate, SoC (system on a chip) system board or PLC (Programmable Logic Controller) including I/O are most
Mini system.
Multi-measuring point drive failure diagnostic system such as Fig. 5 institute based on convolutional neural networks of one embodiment of the invention
Show.In Fig. 5, the present embodiment includes:
Sensor optimization device 1100 is formed corresponding with vibrating sensor for optimizing vibrating sensor installation position
Parallel vibration data;
Convolutional neural networks establish device 1200, for establishing the convolutional layers of initial convolutional neural networks, pond layer and complete
Articulamentum utilizes the initial parallel vibration data of convolution Processing with Neural Network;
Fault diagnosis model optimizes device 1300, for adjusting initial convolutional Neural in parallel vibration data treatment process
Network and the fault diagnosis precision for verifying initial convolutional neural networks, and the degree of fitting of initial convolutional neural networks is adjusted, it is formed
Fault diagnosis model.
As shown in figure 5, in one embodiment of the invention, sensor optimization device 1100 includes:
Locating module 1110, for determining the crucial measurement position of transmission mechanism
Module 1120 is laid, for laying the vibrating sensor for determining freedom degree in crucial measurement position
Correlation comparison module 1130, for being grouped according to vibrating sensor of the freedom degree to measuring point, in same group
Measuring point vibration data carry out related coefficient comparison, and according to significant difference method of inspection determine related coefficient reliability
Threshold value comparison module 1140, for judging whether the related coefficient of each measuring point in same group is more than correction threshold
Preliminary measuring point confirmation module 1150, for related coefficient be less than correction threshold then determine in same group measuring point with
And corresponding vibration data
Measuring point screening module 1160, for related coefficient be more than correction threshold then in same group it is high to related coefficient
Measuring point carries out single screening
Measuring point reaffirms module 1170, for determining each group measuring point corresponding with crucial measurement position and vibration number
According to, repeat related coefficient compare-single screening process completes measuring point screening.
As shown in figure 5, in one embodiment of the invention, sensor optimization device 1100 further include:
Sequence preprocessing module 1180 encodes the vibration data-of each measuring point for establishing data time sequence-according to data time sequence
Carrying out zero-mean normalized-to time series data according to the time series data-that the frequency acquisition of vibration data forms each measuring point makes
The time series data of each measuring point forms parallel input of the data channel as convolutional neural networks.
As shown in figure 5, convolutional neural networks establish module 1200 and include: in one embodiment of the invention
First layer establishes module 1210, for constructing the first convolutional layer and the first pond layer: the convolution kernel of the first convolutional layer
Size is 1 × 5 × n, and convolution kernel number is 64;The size of the pond window of first pond layer is 1 × 1 × 2 × 1, step-length 2;n
Indicate measure-point amount, in an embodiment of the present invention, preferably 6.
The second layer establishes module 1220, for constructing the second convolutional layer and the second pond layer: the convolution kernel of the second convolutional layer
Size is 1 × 5 × 64, and convolution kernel number is 64;The pond window size of second pond layer is 1 × 1 × 2 × 1, step-length 2;
Third layer establishes module 1230, for constructing third convolutional layer and third pond layer: the convolution kernel of third convolutional layer
Size is 1 × 5 × 64, and convolution kernel number is 64;The pond window size of third pond layer is 1 × 1 × 2 × 1, step-length 2;
4th layer is established module 1240, for constructing Volume Four lamination and the 4th pond layer: the convolution kernel of Volume Four lamination
Size is 1 × 5 × 64, and convolution kernel number is 64;The pond window size of 4th pond layer is 1 × 1 × 2 × 1, step-length 2;
Layer 5 establishes module 1250, for constructing the 5th convolutional layer and the 5th pond layer: the convolution kernel of the 5th convolutional layer
Size is 1 × 5 × 64, and convolution kernel number is the data category number of training data, the pond window size of the 5th pond layer for 1 ×
1 × 750 × 1, step-length 750;In one embodiment of the invention, convolution kernel number is 4.
Full articulamentum establishes module 1260, the matrix function after being used to form full articulamentum processing convolutional layer and pond layer operation
According to, and fault diagnosis classification results are formed by softmax output layer.
As shown in figure 5, convolutional neural networks establish module 1200 in one embodiment of the invention further include:
Full articulamentum configuration module 1270, for forming error back propagation using cross entropy, while to fault diagnosis point
Class result carries out classification check.
As shown in figure 5, in one embodiment of the invention, fault diagnosis model optimization module 1300 includes:
Iteration adjustment module 1310, for being formed under gradient vector combination gradient using cross entropy in data processing
The iteratively faster that drop method carries out initial convolutional neural networks to the weight parameter of each convolutional layer and pond layer adjusts;
Precision judgment module 1320 meets fault diagnosis precision for judging whether according to previous iteration;
Over-fitting judgment module 1330, for judging whether over-fitting occur when meeting fault diagnosis precision;
Poor fitting judgment module 1340, for judging whether poor fitting occur when being unsatisfactory for fault diagnosis precision;
The first adjustment module 1350, for each by increasing regular terms adjustment when meeting fault diagnosis precision and over-fitting
The weight parameter quantity of convolutional layer and pond layer;
Second adjustment module 1360, for forming each convolution by error back propagation when being unsatisfactory for fault diagnosis precision
The weight parameter of layer and pond layer adaptively adjusts;
Third adjusts module 1370, for when being unsatisfactory for fault diagnosis precision and poor fitting by adjust each convolutional layer with
And the size and number of pond layer improve the complexity of initial convolutional neural networks.
The foregoing is only a preferred embodiment of the present invention, but scope of protection of the present invention is not limited thereto,
In the technical scope disclosed by the present invention, any changes or substitutions that can be easily thought of by anyone skilled in the art,
It should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with the protection model of claims
Subject to enclosing.
Claims (10)
1. a kind of multi-measuring point drive failure diagnostic method based on convolutional neural networks characterized by comprising
Optimize vibrating sensor installation position, forms parallel vibration data corresponding with the vibrating sensor;
The convolutional layer, pond layer and full articulamentum for establishing initial convolutional neural networks, at the initial convolutional neural networks
Manage the parallel vibration data;
The initial convolutional neural networks are adjusted in parallel vibration data treatment process and verify the initial convolution nerve net
The fault diagnosis precision of network, and the degree of fitting of the initial convolutional neural networks is adjusted, form fault diagnosis model.
2. the multi-measuring point drive failure diagnostic method based on convolutional neural networks, feature exist as described in claim 1
In the optimization vibrating sensor installation position, forming parallel vibration data corresponding with the vibrating sensor includes:
Determine the crucial measurement position of transmission mechanism;
The vibrating sensor of measuring point is laid in the crucial measurement position;
It is grouped according to vibrating sensor of the freedom degree to the measuring point, phase relation is carried out to the measuring point vibration data in same group
Number compares, and determines related coefficient reliability according to significant difference method of inspection;
Whether the related coefficient for judging each measuring point in the same group is more than correction threshold;
The related coefficient is less than the correction threshold and then determines measuring point and corresponding vibration data in the same group;
The related coefficient be more than the correction threshold then the measuring point high to related coefficient in the same group carry out it is single
Screening;
Measuring point and vibration data described in determining each group corresponding with crucial measurement position, repeatedly related coefficient compares-single sieve
Process is selected to complete the measuring point screening.
3. the multi-measuring point drive failure diagnostic method based on convolutional neural networks, feature exist as claimed in claim 2
In the correction threshold of the related coefficient is 80%.
4. the multi-measuring point drive failure diagnostic method based on convolutional neural networks, feature exist as claimed in claim 2
In, further includes:
Data time sequence-is established to be formed according to the vibration data-that data time sequence encodes each measuring point according to the frequency acquisition of vibration data
The time series data-of each measuring point carries out zero-mean normalized to time series data, and the time series data of each measuring point is made to form one
Parallel input of the data channel as convolutional neural networks.
5. the multi-measuring point drive failure diagnostic method based on convolutional neural networks, feature exist as described in claim 1
In the convolutional layer for establishing initial convolutional neural networks, pond layer and full articulamentum utilize the initial convolutional neural networks
Handling the parallel vibration data includes:
Construct the first convolutional layer and the first pond layer: the convolution kernel size of the first convolutional layer is 1 × 5 × n, and n indicates measure-point amount,
Convolution kernel number is 64;The size of the pond window of first pond layer is 1 × 1 × 2 × 1, step-length 2;
Construct the second convolutional layer and the second pond layer: the convolution kernel size of the second convolutional layer is 1 × 5 × 64, and convolution kernel number is
64;The pond window size of second pond layer is 1 × 1 × 2 × 1, step-length 2;
Construct third convolutional layer and third pond layer: the convolution kernel size of third convolutional layer is 1 × 5 × 64, and convolution kernel number is
64;The pond window size of third pond layer is 1 × 1 × 2 × 1, step-length 2;
Construct Volume Four lamination and the 4th pond layer: the convolution kernel size of Volume Four lamination is 1 × 5 × 64, and convolution kernel number is
64;The pond window size of 4th pond layer is 1 × 1 × 2 × 1, step-length 2;
Construct the 5th convolutional layer and the 5th pond layer: the convolution kernel size of the 5th convolutional layer is 1 × 5 × 64, and convolution kernel number is
The data category number of training data, the pond window size of the 5th pond layer are 1 × 1 × 750 × 1, step-length 750;
It forms full articulamentum and handles the matrix data after the convolutional layer and pond layer operation, and pass through softmax output layer shape
At fault diagnosis classification results.
6. the multi-measuring point drive failure diagnostic method based on convolutional neural networks, feature exist as claimed in claim 5
In, further includes:
Error back propagation is formed using cross entropy, while classification check is carried out to fault diagnosis classification results.
7. the multi-measuring point drive failure diagnostic method based on convolutional neural networks, feature exist as described in claim 1
In described to adjust the initial convolutional neural networks in parallel vibration data treatment process and verify the initial convolutional Neural
The fault diagnosis precision of network, and the degree of fitting of the initial convolutional neural networks is adjusted, forming fault diagnosis model includes:
Gradient vector combination gradient descent method is formed to each convolutional layer and pond layer using cross entropy in data processing
Weight parameter carries out the iteratively faster adjustment of initial convolutional neural networks;
Judged whether to meet fault diagnosis precision according to previous iteration;
Judge whether over-fitting occur when meeting the fault diagnosis precision;
Judge whether poor fitting occur when being unsatisfactory for the fault diagnosis precision.
8. the multi-measuring point drive failure diagnostic method based on convolutional neural networks, feature exist as claimed in claim 7
In, further includes:
The power of the convolutional layer and pond layer is adjusted by increasing regular terms when meeting the fault diagnosis precision and over-fitting
Weight number of parameters;
Joined when being unsatisfactory for the fault diagnosis precision by the weight that error back propagation forms the convolutional layer and pond layer
The adaptive adjustment of number;
Pass through the size sum number of the adjusting convolutional layer and pond layer when being unsatisfactory for the fault diagnosis precision and poor fitting
Mesh improves the complexity of the initial convolutional neural networks.
9. a kind of multi-measuring point drive failure diagnostic system based on convolutional neural networks characterized by comprising
Memory, for storing the multi-measuring point transmission mechanism event as described in any of the claims 1 to 8 based on convolutional neural networks
Hinder the program code of diagnostic method treatment process;
Processor, for executing said program code.
10. a kind of multi-measuring point drive failure diagnostic system based on convolutional neural networks characterized by comprising
Sensor optimization device is formed corresponding with the vibrating sensor parallel for optimizing vibrating sensor installation position
Vibration data;
Convolutional neural networks establish device, for establishing the convolutional layer, pond layer and full articulamentum of initial convolutional neural networks, benefit
The parallel vibration data described in the initial convolution Processing with Neural Network;
Fault diagnosis model optimizes device, for adjusting the initial convolutional neural networks in parallel vibration data treatment process
And the fault diagnosis precision of the initial convolutional neural networks is verified, and adjust the degree of fitting of the initial convolutional neural networks,
Form fault diagnosis model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910692350.1A CN110333074B (en) | 2019-07-30 | 2019-07-30 | Multi-measuring-point transmission mechanism fault diagnosis method and system based on convolutional neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910692350.1A CN110333074B (en) | 2019-07-30 | 2019-07-30 | Multi-measuring-point transmission mechanism fault diagnosis method and system based on convolutional neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110333074A true CN110333074A (en) | 2019-10-15 |
CN110333074B CN110333074B (en) | 2021-08-17 |
Family
ID=68147992
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910692350.1A Active CN110333074B (en) | 2019-07-30 | 2019-07-30 | Multi-measuring-point transmission mechanism fault diagnosis method and system based on convolutional neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110333074B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110703743A (en) * | 2019-11-12 | 2020-01-17 | 深圳市亲邻科技有限公司 | Equipment failure prediction and detection system and method |
CN111476299A (en) * | 2020-04-07 | 2020-07-31 | 国家电网有限公司华东分部 | Improved convolutional neural network and power grid intelligent alarm system based on same |
CN111899905A (en) * | 2020-08-05 | 2020-11-06 | 哈尔滨工程大学 | Fault diagnosis method and system based on nuclear power device |
CN112231134A (en) * | 2020-10-28 | 2021-01-15 | 地平线(上海)人工智能技术有限公司 | Fault processing method and device of neural network processor, equipment and storage medium |
CN112651491A (en) * | 2020-12-28 | 2021-04-13 | 浙江智昌机器人科技有限公司 | Industrial robot fault prediction method based on CNN convolution model |
CN116558824A (en) * | 2023-04-19 | 2023-08-08 | 华中科技大学 | Multi-channel-oriented bearing comprehensive index health monitoring method and system |
CN117367570A (en) * | 2023-11-02 | 2024-01-09 | 中国人民解放军海军工程大学 | Intelligent fault diagnosis method for air compressor based on single-point acoustic signals |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105898774A (en) * | 2016-03-30 | 2016-08-24 | 国家电网公司 | Distributed sensor cooperation selecting method |
CN106772695A (en) * | 2016-11-14 | 2017-05-31 | 中南大学 | A kind of Along Railway wind speed forecasting method for merging many air measuring station measured datas |
CN108053061A (en) * | 2017-12-08 | 2018-05-18 | 天津大学 | A kind of solar energy irradiation level Forecasting Methodology based on improvement convolutional neural networks |
CN108563906A (en) * | 2018-05-02 | 2018-09-21 | 北京航空航天大学 | A kind of short fiber reinforced composite macro property prediction technique based on deep learning |
CN108830127A (en) * | 2018-03-22 | 2018-11-16 | 南京航空航天大学 | A kind of rotating machinery fault feature intelligent diagnostic method based on depth convolutional neural networks structure |
CN109299705A (en) * | 2018-10-24 | 2019-02-01 | 电子科技大学 | Rotary machinery fault diagnosis method based on one-dimensional depth residual error convolutional neural networks |
-
2019
- 2019-07-30 CN CN201910692350.1A patent/CN110333074B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105898774A (en) * | 2016-03-30 | 2016-08-24 | 国家电网公司 | Distributed sensor cooperation selecting method |
CN106772695A (en) * | 2016-11-14 | 2017-05-31 | 中南大学 | A kind of Along Railway wind speed forecasting method for merging many air measuring station measured datas |
CN108053061A (en) * | 2017-12-08 | 2018-05-18 | 天津大学 | A kind of solar energy irradiation level Forecasting Methodology based on improvement convolutional neural networks |
CN108830127A (en) * | 2018-03-22 | 2018-11-16 | 南京航空航天大学 | A kind of rotating machinery fault feature intelligent diagnostic method based on depth convolutional neural networks structure |
CN108563906A (en) * | 2018-05-02 | 2018-09-21 | 北京航空航天大学 | A kind of short fiber reinforced composite macro property prediction technique based on deep learning |
CN109299705A (en) * | 2018-10-24 | 2019-02-01 | 电子科技大学 | Rotary machinery fault diagnosis method based on one-dimensional depth residual error convolutional neural networks |
Non-Patent Citations (1)
Title |
---|
张文风等: "基于Dropout-CNN的滚动轴承故障诊断研究", 《轻工机械》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110703743A (en) * | 2019-11-12 | 2020-01-17 | 深圳市亲邻科技有限公司 | Equipment failure prediction and detection system and method |
CN111476299A (en) * | 2020-04-07 | 2020-07-31 | 国家电网有限公司华东分部 | Improved convolutional neural network and power grid intelligent alarm system based on same |
CN111899905A (en) * | 2020-08-05 | 2020-11-06 | 哈尔滨工程大学 | Fault diagnosis method and system based on nuclear power device |
CN111899905B (en) * | 2020-08-05 | 2022-11-01 | 哈尔滨工程大学 | Fault diagnosis method and system based on nuclear power device |
CN112231134A (en) * | 2020-10-28 | 2021-01-15 | 地平线(上海)人工智能技术有限公司 | Fault processing method and device of neural network processor, equipment and storage medium |
CN112231134B (en) * | 2020-10-28 | 2023-08-08 | 地平线(上海)人工智能技术有限公司 | Fault processing method and device for neural network processor, equipment and storage medium |
CN112651491A (en) * | 2020-12-28 | 2021-04-13 | 浙江智昌机器人科技有限公司 | Industrial robot fault prediction method based on CNN convolution model |
CN116558824A (en) * | 2023-04-19 | 2023-08-08 | 华中科技大学 | Multi-channel-oriented bearing comprehensive index health monitoring method and system |
CN116558824B (en) * | 2023-04-19 | 2024-02-06 | 华中科技大学 | Multi-channel-oriented bearing comprehensive index health monitoring method and system |
CN117367570A (en) * | 2023-11-02 | 2024-01-09 | 中国人民解放军海军工程大学 | Intelligent fault diagnosis method for air compressor based on single-point acoustic signals |
CN117367570B (en) * | 2023-11-02 | 2024-04-12 | 中国人民解放军海军工程大学 | Intelligent fault diagnosis method for air compressor based on single-point acoustic signals |
Also Published As
Publication number | Publication date |
---|---|
CN110333074B (en) | 2021-08-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110333074A (en) | Multi-measuring point drive failure diagnostic method and system based on convolutional neural networks | |
CN106895975B (en) | Method for Bearing Fault Diagnosis based on Stacked SAE deep neural network | |
CN110032975B (en) | Seismic facies picking method | |
CN106769052B (en) | A kind of mechanical system rolling bearing intelligent failure diagnosis method based on clustering | |
CN111914883B (en) | Spindle bearing state evaluation method and device based on deep fusion network | |
CN110514924A (en) | The Winding in Power Transformer Fault Locating Method of depth convolutional neural networks fusion visual recognition | |
CN110070060B (en) | Fault diagnosis method for bearing equipment | |
CN110245745B (en) | Equipment residual service life prediction method based on integrated bidirectional cyclic neural network | |
WO2022156330A1 (en) | Fault diagnosis method for rotating device | |
CN108805269A (en) | A method of TRANSFORMATION RATIO is picked up based on LSTM Recognition with Recurrent Neural Network | |
CN110335617A (en) | A kind of noise analysis method in substation | |
CN110058294A (en) | A kind of tunnel micro seismic monitoring rock rupture event automatic identifying method | |
CN104156565B (en) | System mode analysis method based on offline CAN bus data and analytical equipment | |
CN107909118A (en) | A kind of power distribution network operating mode recording sorting technique based on deep neural network | |
CN110954326A (en) | Rolling bearing online fault diagnosis method capable of automatically learning feature expression | |
CN110515781B (en) | Complex system state monitoring and fault diagnosis method | |
CN110455512B (en) | Rotary mechanical multi-integration fault diagnosis method based on depth self-encoder DAE | |
CN110261746B (en) | Cable defect detection method based on periodic attenuation characteristics of oscillating wave voltage | |
CN113762405B (en) | Power network attack recognition system and recognition method thereof | |
CN110297469A (en) | The production line fault judgment method of Ensemble feature selection algorithm based on resampling | |
CN110318731A (en) | A kind of oil well fault diagnostic method based on GAN | |
CN114565038A (en) | Intelligent electric meter fault diagnosis method based on improved capsule network | |
CN115185937A (en) | SA-GAN architecture-based time sequence anomaly detection method | |
CN111898744A (en) | TDLAS trace gas concentration detection method based on pooled LSTM | |
CN114722886A (en) | Knowledge distillation-based crankshaft internal defect detection method and detection equipment |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |