CN114970598B - Mechanical health state monitoring method and device - Google Patents
Mechanical health state monitoring method and device Download PDFInfo
- Publication number
- CN114970598B CN114970598B CN202210174862.0A CN202210174862A CN114970598B CN 114970598 B CN114970598 B CN 114970598B CN 202210174862 A CN202210174862 A CN 202210174862A CN 114970598 B CN114970598 B CN 114970598B
- Authority
- CN
- China
- Prior art keywords
- network
- machine
- signal
- target machine
- mechanical
- 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.)
- Active
Links
- 230000036541 health Effects 0.000 title claims abstract description 108
- 238000012544 monitoring process Methods 0.000 title claims abstract description 103
- 238000000034 method Methods 0.000 title claims abstract description 72
- 238000012549 training Methods 0.000 claims abstract description 53
- 238000013507 mapping Methods 0.000 claims abstract description 17
- 230000003862 health status Effects 0.000 claims description 22
- 230000008569 process Effects 0.000 claims description 16
- 238000001228 spectrum Methods 0.000 claims description 16
- 238000010606 normalization Methods 0.000 claims description 9
- 238000001514 detection method Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000012806 monitoring device Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 abstract description 11
- 238000007781 pre-processing Methods 0.000 abstract description 7
- 238000012360 testing method Methods 0.000 description 8
- 238000002474 experimental method Methods 0.000 description 5
- 230000006870 function Effects 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 238000005070 sampling Methods 0.000 description 4
- 238000000605 extraction Methods 0.000 description 3
- 238000009776 industrial production Methods 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 230000007246 mechanism Effects 0.000 description 3
- 238000009827 uniform distribution Methods 0.000 description 3
- 230000008485 antagonism Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000003745 diagnosis Methods 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 230000001133 acceleration Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000007797 corrosion Effects 0.000 description 1
- 238000005260 corrosion Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000011089 mechanical engineering Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 230000005236 sound signal Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
Classifications
-
- 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
- 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
-
- 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
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E10/00—Energy generation through renewable energy sources
- Y02E10/70—Wind energy
- Y02E10/72—Wind turbines with rotation axis in wind direction
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Signal Processing (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
The invention provides a method and a device for monitoring the health state of a machine, comprising the following steps: establishing a challenge representation learning network comprising a first encoder subnetwork for mapping the input signal to a first implicit spatial code, a decoder subnetwork for reconstructing the first implicit spatial code to a signal space, and a second encoder subnetwork for mapping the reconstructed signal in the signal space to a second implicit spatial code; taking the characteristic data of the healthy machine running signals as network training data to conduct countermeasure learning training on a countermeasure representation learning network to obtain a machine health state monitoring network; and (3) inputting the operation signals of the target machine acquired in real time into a machine health state monitoring network after preprocessing to obtain the mechanical damage index of the target machine. The invention can solve the problems that the mechanical health state is difficult to monitor under the condition of missing fault operation data, the effect of the existing method is poor, and the like.
Description
Technical Field
The invention relates to the technical field of mechanical engineering, in particular to a method and a device for monitoring mechanical health state.
Background
In modern industrial systems, mechanical equipment plays a very critical role, and force and motion can be converted through combination of a mechanism and a machine, so that an industrial production process is completed, the mechanical health condition has a direct influence on the operation efficiency of the industrial system, the industrial system is stopped due to mechanical failure, huge economic loss is caused, and serious mechanical failure also causes casualties, so that the life safety of production personnel is endangered. With the improvement of the mechanical design and manufacturing level, the current mechanical equipment presents a development trend of complexity, large size and integration, which further aggravates the severity of the consequences caused by mechanical faults, so how to accurately monitor the health state of the machine and give reliable early warning before the mechanical faults worsen becomes a problem to be solved.
In the mechanical operation process, a plurality of operation signals, such as vibration signals, sound signals, current signals of electric components and the like, can be measured, and the signals are closely related to the mechanical operation process, so that the mechanical health state can be reflected, and the mechanical health state can be identified by collecting and processing the signals. In order to avoid catastrophic consequences due to serious mechanical failure, a number of mechanical health monitoring methods based on mechanical operating signals have been proposed. For simple machines, indexes such as root mean square value (J.Igba,K.Alemzadeh,C.Durugbo,et al.Analysing RMS and peak values of vibration signals for condition monitoring of wind turbine gearboxes[J].Renewable Energy,2016,91(3):90-106.)、 kurtosis (CN 201610562478), fault characteristic frequency (CN 201711086012) and the like can be adopted to monitor the health state of the machine, the health state of the machine is judged according to the numerical values of the indexes, and most of the indexes are established based on a machine fault mechanism and do not have self-adaptive adjustment capability for different working conditions, so that the machine can only play a role in specific scenes, and the health state of the complex machine is difficult to accurately reveal. With the development of information science and artificial intelligence, machine learning-based mechanical fault diagnosis techniques (CN 202010911939, CN 201811197982) are proposed, which can mine mechanical health status information from historical data, overcome the difficulty of modeling the mechanism of complex mechanical faults, but the above method can only perform supervised fault diagnosis, require that all possible mechanical fault types be defined in advance, and require that the model be trained by operation data under different health conditions. However, mechanical faults are rare compared with normal operation conditions in actual conditions, and mechanical long-time fault operation is not allowed for ensuring production safety, so that mechanical operation data under the fault conditions are very rare, and the practicability of the machine learning method is limited by the scarcity of the fault data. In order to perform monitoring of the mechanical health status under the condition of missing fault data, an anomaly detection method, such as support vector data description (Mao W,J Chen,Liang X,et al.A New Online Detection Approach for Rolling Bearing Incipient Fault via Self-Adaptive Deep Feature Matching[J].IEEE Transactions on Instrumentation and Measurement,2020,69(2):443-456.)、, may be used to generate the self-encoder (Jiang G,Xie P,He H,et al.Wind Turbine Fault Detection Using Denoising Autoencoder with Temporal Information[J].IEEE/ASME Transactions on Mechatronics,2017:1-1.) of the countermeasure network (Peng C A,Yu L A,Kw A,et al.A threshold self-setting condition monitoring scheme for wind turbine generator bearings based on deep convolutional generative adversarial networks[J].Measurement,2020,167.)、, and the methods only need the operation data under the health status in the model training stage, extract the characteristics in the mechanical health operation signal by learning the health operation data, calculate the difference between the test data and the extracted health characteristics in the test stage, and further identify the mechanical health status. The method is established under the assumption that the healthy operation signals are in a single mode and distributed, and only data in a healthy state are needed in the training process, only the characteristics in the healthy operation signals are extracted, and the mechanical operation signals in a fault state are not highlighted in the test stage. However, due to the complexity of the mechanical equipment, the mechanical operation signal in the healthy state has a non-stationary, multi-modal characteristic, which cannot meet the applicable conditions and assumptions of the above-mentioned method, and these problems affect the mechanical health monitoring effect of the existing method, so that the mechanical health cannot be effectively monitored only by the existing method.
Disclosure of Invention
In view of the above problems, the present invention aims to provide a method and an apparatus for monitoring a mechanical health state, so as to solve the problems that it is difficult to monitor the mechanical health state under the condition of missing fault operation data, and it is difficult to overcome the defect that the existing method has poor effect under the condition of unstable operation data and multiple modes, so that the reliability of the mechanical health state monitoring result is low.
The invention provides a method for monitoring the health state of a machine, which comprises the following steps:
Establishing a challenge representation learning network comprising a first encoder subnetwork for mapping an input signal into a first implicit spatial code, a decoder subnetwork for reconstructing the first implicit spatial code into a signal space, and a second encoder subnetwork for mapping a reconstructed signal in the signal space into a second implicit spatial code;
Taking the characteristic data of the healthy machine running signals as network training data to perform countermeasure learning training on the countermeasure representation learning network, and taking the relative error of the first hidden space code and the second hidden space code as a mechanical damage index after the countermeasure learning training reaches a preset convergence condition to obtain a mechanical health state monitoring network;
The operation signals of the target machinery, which are collected in real time, are preprocessed and then input into the machinery health state monitoring network, so that the mechanical damage index of the target machinery is obtained;
And determining a monitoring result of the mechanical health state of the target machine according to the mechanical damage index of the target machine based on a preset mechanical damage index threshold.
Furthermore, it is preferred that the challenge representation learning network further comprises a discriminator subnetwork for generating uniformly distributed random variables from the second implicit spatial code.
Furthermore, preferably, the preset convergence condition is:
the differences between the operating signal of the healthy machine in the characteristic data and the reconstructed signal in the signal space, the differences between the first and second hidden space codes, the differences between the second hidden space code and the uniformly distributed random variable are minimized to converge the parameters of the first encoder sub-network, the decoder sub-network, the second encoder sub-network and the arbiter sub-network to a steady state.
Furthermore, it is preferable that the first encoder sub-network, the decoder sub-network, the second encoder sub-network, and the discriminator sub-network are each formed by connecting a convolution layer, a full connection layer, and a normalization layer.
In addition, preferably, in the process of establishing the countermeasure representation learning network,
And splicing the feature graphs of the same level in the first encoder sub-network and the decoder sub-network along the feature direction, and taking the spliced feature graph as the feature graph of the corresponding level in the decoder sub-network so as to enable the feature graph in the first encoder sub-network to be shared in the feature graph of the decoder sub-network.
In addition, preferably, the method for acquiring the characteristic data of the operation signal of the healthy machine comprises the following steps:
Acquiring an original running signal x r of the machine in a healthy state through a sensor;
Converting the original operating signal x r to a frequency domain by using discrete fourier transform, so as to obtain a frequency spectrum s=dft (x r) of the original operating signal of the machine in a healthy state; wherein DFT () is a discrete fourier transform;
Normalizing the frequency spectrum to ensure that the amplitude values of the obtained operation signals are in the range of 0 and 1, thereby obtaining the characteristic data of the operation signals of the healthy machinery; wherein, the normalization process is expressed as:
Wherein,
S max is the element with the largest amplitude in s, s min is the element with the smallest amplitude in s, and x is the feature data.
In addition, it is preferable that the original operating signal x r of the machine in the healthy state is any one of a vibration signal, a current signal and an acoustic signal.
In addition, preferably, the preprocessing the operation signal of the target machine acquired in real time and inputting the operation signal to the machine health status monitoring network, and obtaining the machine damage index of the target machine includes:
acquiring an operation signal x t of the target machine in real time through a sensor;
Converting the operating signal x t of the target machine to a frequency domain using a discrete fourier transform, resulting in a frequency spectrum s=dft (x t) of the operating signal of the target machine, wherein DFT () is a discrete fourier transform;
Normalizing the frequency spectrum of the operation signal of the target machine to ensure that the amplitude values of the obtained operation signal are all in the range of [0,1] to obtain the operation signal characteristic data of the preprocessed target machine;
And inputting the operation signal characteristic data of the target machine to the machine health state monitoring network to obtain the mechanical damage index of the target machine.
Furthermore, preferably, the inputting the operation signal characteristic data of the target machine to the machine health status monitoring network, and obtaining the machine damage index of the target machine includes:
Mapping the operation signal characteristic data of the target machine into a first hidden space code of the target machine through a first encoder sub-network of the machine health status monitoring network;
decoding the first hidden space code of the target machine into a reconstructed signal of the target machine through a decoder subnetwork of the machine health status monitoring network;
Transmitting the reconstructed signal of the target machine to a second hidden space code of the target machine through a second encoder sub-network of the machine health status monitoring network;
Calculating a mechanical damage index of the target machine according to the first hidden space code of the target machine and the second hidden space code of the target machine; wherein, the calculation formula of the mechanical damage index is as follows:
where z (x) is the first hidden space encoding of the target machine,/> For the second hidden space encoding of the target machine, DI is an indicator of mechanical damage to the target machine.
The invention provides a mechanical health state monitoring device, comprising:
A learning network establishment module for establishing a challenge representation learning network comprising a first encoder sub-network for mapping an input signal into a first hidden space code, a decoder sub-network for reconstructing the first hidden space code into a signal space, and a second encoder sub-network for mapping a reconstructed signal in the signal space into a second hidden space code;
The network training module is used for performing countermeasure learning training on the countermeasure representation learning network by taking the characteristic data of the healthy machine running signal as network training data, and taking the relative error of the first hidden space code and the second hidden space code as a mechanical damage index after the countermeasure learning training reaches a preset convergence condition to obtain a machine health state monitoring network;
the mechanical damage index detection module is used for inputting the operation signals of the target machine acquired in real time into the mechanical health state monitoring network after pretreatment to obtain mechanical damage indexes of the target machine;
And the monitoring result generation module is used for determining the monitoring result of the mechanical health state of the target machine according to the mechanical damage index of the target machine based on a preset mechanical damage index threshold.
According to the technical scheme, the method and the device for monitoring the mechanical health state provided by the invention have the advantages that the built countermeasure representation learning network is trained by the characteristic data of the mechanical operation signals in the health state, the problem of training difficulty caused by mechanical failure is solved, the mechanical health state is indicated by building damage indexes in the test stage, and the method and the device have good practicability in the actual industrial production scene; compared with the prior art, the invention reconstructs the mechanical operation signals in the signal space and the hidden space, can adapt to the non-stable and multi-mode operation conditions of the machinery, and can more effectively extract the characteristics related to the mechanical health state from the operation signals, thereby obtaining better health state monitoring effect; the invention provides constraint for the separability of the healthy operation signal and the fault operation signal in the monitoring stage, and inhibits the model parameters from falling into the local optimum in a way of resisting evolutionary training, thereby improving the accuracy of monitoring the mechanical health state under the complex working condition.
To the accomplishment of the foregoing and related ends, one or more aspects of the invention comprise the features hereinafter fully described. The following description and the annexed drawings set forth in detail certain illustrative aspects of the invention. These aspects are indicative, however, of but a few of the various ways in which the principles of the invention may be employed. Furthermore, the invention is intended to include all such aspects and their equivalents.
Drawings
Other objects and attainments together with a more complete understanding of the invention will become apparent and appreciated by referring to the following description taken in conjunction with the accompanying drawings. In the drawings:
FIG. 1 is a flow chart of a method for monitoring a state of mechanical health according to an embodiment of the present invention;
FIG. 2 is a flow chart illustrating an implementation of a method for monitoring machine health status according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a mechanical health monitoring network according to a mechanical health monitoring method of an embodiment of the present invention;
Fig. 4 is a graph comparing the results of the mechanical health monitoring performed by the mechanical health monitoring method and the comparison method according to the embodiment of the invention.
In the drawings, like reference numerals designate similar or corresponding features or functions.
Detailed Description
In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of one or more embodiments. It may be evident, however, that such embodiment(s) may be practiced without these specific details.
Aiming at the problems that the mechanical health state monitoring is difficult to be carried out under the condition of missing fault operation data at present and the effect is poor under the condition of non-stable operation data and multiple modes in the existing method, the reliability of the mechanical health state monitoring result is low and the like are solved, the mechanical health state monitoring method and device are provided.
Specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
In order to illustrate the method and the device for monitoring the mechanical health state provided by the invention, fig. 1 shows a flow of the method for monitoring the mechanical health state according to an embodiment of the invention; FIG. 2 illustrates an implementation flow of a method for machine health monitoring in accordance with an embodiment of the present invention; FIG. 3 illustrates a structure of a machine health monitoring network of a machine health monitoring method according to an embodiment of the present invention; fig. 4 shows a comparison of the results of the mechanical health monitoring by the mechanical health monitoring method according to the embodiment of the present invention and the comparison method.
As shown in fig. 1 to fig. 4 together, the method for monitoring the mechanical health status provided by the invention comprises the following steps:
s1, establishing a countermeasure representation learning network, wherein the countermeasure representation learning network comprises a first encoder sub-network for mapping an input signal into a first hidden space code, a decoder sub-network for reconstructing the first hidden space code into a signal space, and a second encoder sub-network for mapping a reconstructed signal in the signal space into a second hidden space code.
Specifically, a countermeasure representation learning network is established, training data (i.e., characteristic data of a healthy machine operation signal) is input into a first encoder sub-network E R and a decoder sub-network D R which are connected in series, the first encoder sub-network E R maps the input signal (i.e., the characteristic data of the healthy machine operation signal) into a first hidden space code, thereby extracting hidden space characteristics of the machine operation signal under healthy conditions, and the decoder sub-network D R reconstructs the first hidden space code into a signal space, thereby ensuring effectiveness of the extraction of the hidden space characteristics, wherein an output of the decoder sub-network is represented as:
Where z (x) =e R (x) is the first hidden space coding output by the first encoder subnetwork E R.
Due to the complexity of the machine in the actual industrial scene, the machine operation signal in the healthy state has the characteristics of non-stability and multi-mode, so that the difficulty is brought to the feature extraction of the machine operation signal in the healthy condition, and the machine health condition cannot be judged directly according to the error before and after the signal passes through the first encoder sub-network E R and the decoder sub-network D R to reconstruct. The hidden space coding has the characteristics of low dimensionality and less redundancy, can adapt to mechanical non-stationary operation characteristics, and highlights components related to health states in operation signals, so that in the built countermeasure representation learning network, a second encoder sub-network E D maps the reconstructed signals into second hidden space codes:
As a preferred aspect of the invention, the challenge representation learning network further comprises a discriminator sub-network for generating uniformly distributed random variables from the second implicit spatial encoding.
Specifically, to overcome the complexity of mechanical failure and the rarity of failure signal samples, network training is performed using only the operation signals in the healthy state, and it is assumed that through the training process, the built countermeasure indicates that the learning network can reconstruct the mechanical operation signals in the healthy state in the signal space and the hidden space, and the mechanical operation signals in the failed state cannot be effectively reconstructed because they do not appear in the training process. The reliability of the mechanical health monitoring is therefore related to the distribution characteristics of the hidden spatial coding of the operating signal in the healthy state. In order to inhibit the reconstruction effect of the mechanical operation signal in the mechanical health monitoring process under the fault state and improve the recognition sensitivity to the mechanical fault, in the built countermeasure representation learning network, the arbiter sub-network E S enhances the similarity between the first hidden space code and the uniformly distributed random variables U-U (a, b) in a manner of generating countermeasure training, and the objective function of the generated countermeasure training process is as follows:
Wherein/> Is a mathematical expectation.
As a preferred embodiment of the present invention, the first encoder sub-network, the decoder sub-network, the second encoder sub-network and the arbiter sub-network are each formed by connecting a convolution layer, a full connection layer and a normalization layer.
Specifically, each sub-network is formed by connecting a series of convolution layers, full-connection layers and normalization layers, feature extraction is carried out through the convolution layers, extracted features are spliced through the full-connection layers, and finally the spliced features are normalized through the normalization layers.
As a preferred aspect of the present invention, in establishing the antagonism means learning network,
And splicing the feature graphs of the same level in the first encoder sub-network and the decoder sub-network along the feature direction, and taking the spliced feature graphs as the feature graphs of the corresponding level in the decoder sub-network, so that the feature graphs in the first encoder sub-network are shared in the feature graphs of the decoder sub-network.
Specifically, for mechanical health monitoring, the hidden space coding of the mechanical operating signal contains less redundant information than the original signal, but during the coding and decoding processes, the detail information related to the health state in the signal is lost due to the downsampling of the coding, and once the information is lost, the information cannot be recovered only through the upsampling of the decoding, and the loss of the detail information can reduce the accuracy of mechanical health monitoring. To solve this problem, the feature map of the first encoder sub-network E R is shared in a spliced manner among the feature maps of the decoder sub-network D R, the feature maps of the same hierarchy in the first encoder sub-network E R and the decoder sub-network D R are spliced along the feature direction, and the spliced feature map is used as the feature map of the corresponding hierarchy in the decoder sub-network D R, so that the detail information in the first encoder sub-network E R is transferred to the decoder sub-network D R.
S2, feature data of the healthy machine running signals are used as network training data to conduct countermeasure learning training on a countermeasure representation learning network, and when the countermeasure learning training reaches a preset convergence condition, the relative error of the first hidden space code and the second hidden space code is used as a mechanical damage index to obtain a machine health state monitoring network.
Specifically, characteristic data (i.e., input signals) of the operation signals of the healthy machine are input into the countermeasure representation learning network, the countermeasure representation learning network is subjected to countermeasure learning training through processing such as coding, signal reconstruction, recoding and uniformly distributed random variables of each sub-network, and after a preset convergence condition is reached, a machine health state monitoring network is obtained, and in the machine health state monitoring network, the relative error of the first hidden space coding and the second hidden space coding is used as a machine damage index.
The method comprises the following steps:
Step ①: initializing the total training epoch number I, the training epoch number J of the second encoder sub-network E D, the batch training data size K, the parameter updating step length l, the gradient penalty parameter lambda, the penalty function weights omega sig、ωcode、ωadv and omega lat and the hidden space boundaries a and b;
Step ②: randomly initialized parameters θ ER of the first encoder sub-network E R, parameters θ DR of the decoder sub-network D R, parameters θ ED of the second encoder sub-network E D, parameters θ ES of the arbiter sub-network E S;
step ③: let the overall training count variable i=1, e D training count variable j=1.
Step ④: a batch of data x k (k=1, 2,) of number K is selected among the characteristic data of the healthy machine operation signal, K), data x k is input to the first encoder sub-network E R to obtain z k＝ER(xk), z k is input to D R to obtainWill/>Input E D gets/>
Step ⑤: sampling a sample a in a uniform distribution within the range of (0, 1), calculating weighted dataInputting the weighted data x s into the second encoder subnetwork E D results in/>Updating the value of θ ED with RMSProp optimizer:
Wherein,
M is the dimension of the weighted data x s;
Step ⑥: if J < J, let j=j+1 and return to step ④.
Step ⑦: updating the value of θ ER with RMSProp optimizer:
Wherein,
Updating values of θ DR using RMSProp optimizerSampling a sample u k of the same dimension as x k in a uniform distribution over the range of (a, b), sampling a sample β in a uniform distribution over the range of (0, 1), updating the value of θ ES with a RMSProp optimizer:
Step ⑧: if I < I, let i=i+1 and return to step ⑤.
As a preferred aspect of the present invention, the preset convergence condition is:
The difference between the operating signal of the healthy machine in the characteristic data and the reconstructed signal in the signal space, the difference between the first hidden space coding and the second hidden space coding, the difference between the second hidden space coding and the uniformly distributed random variable are minimized to converge the parameters of the first encoder sub-network, the decoder sub-network, the second encoder sub-network and the arbiter sub-network to a steady state.
As a preferred embodiment of the present invention, a method for acquiring characteristic data of an operation signal of a healthy machine includes:
Acquiring an original running signal x r of the machine in a healthy state through a sensor;
Converting the original operation signal x r to a frequency domain by using discrete fourier transform to obtain a frequency spectrum s=dft (x r) of the original operation signal of the machine in a healthy state; wherein DFT () is a discrete fourier transform;
normalizing the frequency spectrum to ensure that the amplitude values of the obtained operation signals are all in the range of [0,1] to obtain the characteristic data of the operation signals of the healthy machinery; the normalization process is expressed as follows:
Wherein,
S max is the element with the largest amplitude in s, s min is the element with the smallest amplitude in s, and x is the feature data.
Specifically, the original operation signal of the machine in the health state is collected, and the operation signal characteristics of the health machine can be highlighted after the signal preprocessing is performed. The method comprises the steps of collecting original running signals of a machine in a healthy state by using a sensor, preprocessing the original running signals in the healthy state, and for most rotating machines, reflecting fault characteristics better than those of a time domain, so that the original signals are converted into the frequency domain by using discrete Fourier transform to obtain frequency spectrums of the original running signals in the healthy state, and carrying out normalization processing on the frequency spectrums of the original running signals in the healthy state, so that signal amplitude values are in the range of 0 and 1, and the influence of the signal amplitude values on health state monitoring results is eliminated.
As a preferred embodiment of the present invention, the original operation signal of the machine in the healthy state is any one of a vibration signal, a current signal, and an acoustic signal.
S3, inputting the operation signals of the target machine acquired in real time into a machine health state monitoring network after preprocessing, and obtaining the mechanical damage index of the target machine.
Specifically, a sensor is adopted to collect an operation signal of the target machine in real time, namely, an operation signal of the machine to be monitored for health status, signal data which can highlight the operation signal characteristics of the target machine is obtained after preprocessing, and the signal data is input into a machine health status monitoring network, so that a machine damage index of the target machine is obtained.
As a preferred scheme of the invention, the operation signals of the target machine, which are collected in real time, are input into a machine health state monitoring network after being preprocessed, and the mechanical damage index of the target machine is obtained, which comprises the following steps:
acquiring an operation signal x t of the target machine in real time through a sensor;
Converting the operating signal x t of the target machine to the frequency domain using a discrete fourier transform, resulting in a frequency spectrum s=dft (x t) of the operating signal of the target machine, wherein DFT () is the discrete fourier transform;
normalizing the frequency spectrum of the operation signal of the target machine to ensure that the amplitude values of the obtained operation signal are all in the range of [0,1] to obtain the operation signal characteristic data of the target machine after pretreatment;
and inputting the operation signal characteristic data of the target machine into a machine health state monitoring network to obtain the mechanical damage index of the target machine.
Specifically, a sensor is used for collecting an operation signal x t of a target machine in real time, the operation signal x t is preprocessed, the operation signal x t is converted into a frequency domain by using discrete Fourier transform to obtain a frequency spectrum s=DFT (x t) of the operation signal of the target machine, normalization processing is carried out on the frequency spectrum s=DFT (x t) to ensure that signal amplitudes are all in the range of [0,1], so that the influence of the signal amplitudes on a health state monitoring result is eliminated, the preprocessed operation signal characteristic data of the target machine is input into a machine health state monitoring network to obtain a first hidden space code z (x) =E R (x) and a second hidden space codeAnd constructing a mechanical damage index DI according to the relative error of the first hidden space code and the second hidden space code.
As a preferred embodiment of the present invention, inputting the operation signal characteristic data of the target machine into the machine health status monitoring network, obtaining the machine damage index of the target machine includes:
mapping the operation signal characteristic data of the target machine into a first hidden space code of the target machine through a first encoder sub-network of the machine health status monitoring network;
decoding the first hidden space code of the target machine into a reconstructed signal of the target machine through a decoder subnetwork of the machine health status monitoring network;
Transmitting the reconstructed signal of the target machine to a second hidden space code of the target machine through a second encoder sub-network of the machine health status monitoring network;
Calculating a mechanical damage index of the target machine according to the first hidden space code of the target machine and the second hidden space code of the target machine; wherein, the calculation formula of the mechanical damage index is:
where z (x) is the first hidden space encoding of the target machine,/> For the second hidden space encoding of the target machine, DI is an indicator of mechanical damage to the target machine.
The machine damage indicator DI may indicate a health condition of the machine, with a greater DI value indicating a greater probability of occurrence of damage in the machine and a lesser DI value indicating a healthier machine.
And S4, determining a monitoring result of the mechanical health state of the target machine according to the mechanical damage index of the target machine based on a preset mechanical damage index threshold.
Specifically, the mechanical damage index threshold value can be preset according to the actual situation, the mechanical damage index of the target machine is compared with the preset mechanical damage index threshold value, and when the mechanical damage index threshold value is larger than the preset mechanical damage index threshold value, a monitoring result that the mechanical health state of the target machine is the damage state is generated, otherwise, the monitoring result is the health state monitoring result. In order to give a better warning, damage warning information can be generated according to the monitoring result of the damage state.
The invention provides a mechanical health state monitoring device, comprising:
A learning network establishment module for establishing a countermeasure representation learning network comprising a first encoder sub-network for mapping the input signal to a first implicit spatial code, a decoder sub-network for reconstructing the first implicit spatial code to a signal space, and a second encoder sub-network for mapping the reconstructed signal in the signal space to a second implicit spatial code;
The network training module is used for performing countermeasure learning training on the countermeasure representation learning network by taking the characteristic data of the healthy mechanical operation signal as network training data, and taking the relative error of the first hidden space code and the second hidden space code as a mechanical damage index after the countermeasure learning training reaches a preset convergence condition to obtain a mechanical health state monitoring network;
The mechanical damage index detection module is used for inputting the operation signals of the target machine acquired in real time into the mechanical health state monitoring network after pretreatment to obtain mechanical damage indexes of the target machine;
The monitoring result generation module is used for generating a monitoring result of the mechanical health state of the target machine according to the mechanical damage index of the target machine based on a preset mechanical damage index threshold value.
In order to better embody the practicability of the embodiment of the invention, the technical scheme in the embodiment of the invention is described and verified by utilizing a bearing disclosure data set published by Paderbern university. The data set is collected on a bearing experiment table, the experiment table comprises a motor, a coupler, a bearing experiment device, a flywheel and a load, the type of the bearing is 6203, bearing vibration signals in the experiment process are collected by using an acceleration sensor, repeated experiments are carried out under the conditions of various loads, rotating speeds and axial forces so as to reflect mechanical system responses under different working conditions, and the sampling frequency of the vibration signals is 64kHz.
In the model training stage, vibration signals of the healthy bearing under four different working conditions (1500 RPM, 0.7N.m load and 1000N axial force, 900RPM, 0.7N.m load and 1000N axial force, 1500RPM, 0.1 N.m load and 1000N axial force, 1500RPM, 7 N.m load and 400N axial force) are taken as training samples, a mechanical health state monitoring network is obtained by establishing an countermeasure representation learning network, preprocessing training sample data, the countermeasure representation learning network and carrying out network training, and the network construction and training processes are implemented by using Python language and PyTorch frames. Because the training samples contain data under different working conditions, the embodiment can embody the multi-mode characteristic of the complex mechanical running state. In the model training process, the total training epoch number i=100, the E D training epoch number j=10, the batch training data size k=128, the parameter update step l=0.0001, the gradient penalty parameter λ=1, the penalty function weight ω sig＝10、ωcode＝1、ωadv＝1、ωlat =0.1, and the hidden space boundary a=0, b=1.
In the model test stage, the operation data under the conditions of the bearing inner ring fault and the bearing outer ring fault are adopted, the fault forms of the bearing outer ring fault and the bearing inner ring fault are pitting corrosion, the fault size is smaller than 2mm, and the bearing outer ring fault and the bearing inner ring fault belong to faults naturally formed in the accelerated life test process. In order to embody the complexity of the mechanical operation state, the operation data under four different working conditions are mixed and used as test samples together, and the working conditions considered are consistent with the working conditions of the model training stage. Comparing the technical solution of the embodiment of the invention with the existing mechanical health state monitoring method, the comparison method comprises generating the description of the antagonism network, the self-encoder, the kurtosis and the support vector data, and the area (area under the curve of THE RECEIVER operating characteristic, AUCROC) under the operation characteristic curve of the receiver obtained by different methods is shown in fig. 4, it can be seen that the technical solution of the embodiment of the invention has the best mechanical health state monitoring effect, can obtain the monitoring accuracy of 99.98%, can remarkably improve the monitoring capability of the mechanical health state compared with the existing method, and is suitable for the non-steady operation characteristics of complex machinery. For generating the countermeasure network, the method can realize the monitoring of the mechanical health state by learning the distribution characteristics of the mechanical operation signals under the health state, and establish the connection between the random variable and the mechanical operation signals, but has great difficulty in generating the effective training of the countermeasure network and has insufficient adaptability to complex working conditions. The self-encoder can extract hidden space features of the healthy running signal, but the change of the mechanical health state can not be accurately reflected only through pixel-by-pixel errors, so that the health monitoring effect is poor. The support vector data description is a statistical learning method, and compared with a deep learning method, the method has strong dependence on characteristic engineering, so that a better effect can be obtained in the aspects of simple working conditions and simple mechanical health monitoring, but the method has insufficient adaptability to complex mechanical health monitoring, and the effect is inferior to the technical scheme of the embodiment of the invention. Kurtosis can be used for identifying impact components in signals, but inherent impact exists between kinematic pairs in the mechanical operation process, and the mechanical health condition cannot be accurately judged only by the criterion. Based on the above results and analysis, it can be seen that the embodiments of the present invention have better practicality, can evaluate the mechanical health status in actual situations, and can obtain better results compared with the existing methods.
According to the method and the device for monitoring the mechanical health state, disclosed by the invention, the built countermeasure representation learning network is trained by the characteristic data of the mechanical operation signal in the health state, so that the problem of training difficulty caused by mechanical failure is solved, the mechanical health state is indicated by building the damage index in the test stage, and the method and the device have good practicability in an actual industrial production scene; compared with the prior art, the invention reconstructs the mechanical operation signals in the signal space and the hidden space, can adapt to the non-stable and multi-mode operation conditions of the machinery, and can more effectively extract the characteristics related to the mechanical health state from the operation signals, thereby obtaining better health state monitoring effect; the invention provides constraint for the separability of the healthy operation signal and the fault operation signal in the monitoring stage, and inhibits the model parameters from falling into the local optimum in a way of resisting evolutionary training, thereby improving the accuracy of monitoring the mechanical health state under the complex working condition.
The method and apparatus for monitoring the state of health of a machine according to the present invention are described above by way of example with reference to the accompanying drawings. It will be appreciated by those skilled in the art that various modifications may be made to the method and apparatus for monitoring a mechanical health condition as set forth above without departing from the spirit of the invention. Accordingly, the scope of the invention should be determined from the following claims.
Claims (4)
1. A method for monitoring the health status of a machine, comprising the steps of:
establishing a challenge representation learning network comprising a first encoder sub-network for mapping an input signal into a first hidden space code, a decoder sub-network for reconstructing the first hidden space code into a signal space, a second encoder sub-network for mapping a reconstructed signal in the signal space into a second hidden space code, and a arbiter sub-network for generating uniformly distributed random variables from the second hidden space code;
Taking the characteristic data of the healthy machine running signals as network training data to perform countermeasure learning training on the countermeasure representation learning network, and taking the relative error of the first hidden space code and the second hidden space code as a mechanical damage index after the countermeasure learning training reaches a preset convergence condition to obtain a mechanical health state monitoring network; wherein, the preset convergence condition is: minimizing differences between the operating signal of the healthy machine in the characteristic data and the reconstructed signal in the signal space, between the first and second hidden space codes, between the second hidden space code and the uniformly distributed random variable to converge parameters of the first encoder sub-network, the decoder sub-network, the second encoder sub-network and the arbiter sub-network to a steady state;
the method for acquiring the characteristic data of the operation signal of the healthy machine comprises the following steps: acquiring an original running signal x r of the machine in a healthy state through a sensor; converting the original operating signal x r to a frequency domain by using discrete fourier transform, so as to obtain a frequency spectrum s=dft (x r) of the original operating signal of the machine in a healthy state; wherein DFT () is a discrete fourier transform; normalizing the frequency spectrum to ensure that the amplitude values of the obtained operation signals are in the range of 0 and 1, thereby obtaining the characteristic data of the operation signals of the healthy machinery; wherein, the normalization process is expressed as: Wherein,
S max is the element with the largest amplitude in s, s min is the element with the smallest amplitude in s, and x is the feature data; the original running signal x r of the machine in a healthy state is any one of a vibration signal, a current signal and an acoustic signal;
the operation signals of the target machinery, which are collected in real time, are preprocessed and then input into the machinery health state monitoring network, so that the mechanical damage index of the target machinery is obtained; the method for obtaining the mechanical damage index of the target machine comprises the following steps of:
acquiring an operation signal x t of the target machine in real time through a sensor;
Converting the operating signal x t of the target machine to a frequency domain using a discrete fourier transform, resulting in a frequency spectrum s=dft (x t) of the operating signal of the target machine, wherein DFT () is a discrete fourier transform;
Normalizing the frequency spectrum of the operation signal of the target machine to ensure that the amplitude values of the obtained operation signal are all in the range of [0,1] to obtain the operation signal characteristic data of the preprocessed target machine;
inputting the operation signal characteristic data of the target machine to the machine health state monitoring network to obtain a mechanical damage index of the target machine, wherein the operation signal characteristic data comprises the following steps:
Mapping the operation signal characteristic data of the target machine into a first hidden space code of the target machine through a first encoder sub-network of the machine health status monitoring network;
decoding the first hidden space code of the target machine into a reconstructed signal of the target machine through a decoder subnetwork of the machine health status monitoring network;
Transmitting the reconstructed signal of the target machine to a second hidden space code of the target machine through a second encoder sub-network of the machine health status monitoring network;
Calculating a mechanical damage index of the target machine according to the first hidden space code of the target machine and the second hidden space code of the target machine; wherein, the calculation formula of the mechanical damage index is as follows:
where z (x) is the first hidden space encoding of the target machine,/> The second hidden space of the target machine is coded, and DI is a mechanical damage index of the target machine;
And determining a monitoring result of the mechanical health state of the target machine according to the mechanical damage index of the target machine based on a preset mechanical damage index threshold.
2. The method of claim 1, wherein the first encoder sub-network, the decoder sub-network, the second encoder sub-network, and the arbiter sub-network are each connected by a convolutional layer, a fully-connected layer, and a normalized layer.
3. The method of claim 2, wherein, during the establishing of the challenge representation learning network,
And splicing the feature graphs of the same level in the first encoder sub-network and the decoder sub-network along the feature direction, and taking the spliced feature graph as the feature graph of the corresponding level in the decoder sub-network so as to enable the feature graph in the first encoder sub-network to be shared in the feature graph of the decoder sub-network.
4. A machine health monitoring device, characterized in that it is capable of implementing the steps of the machine health monitoring method according to claim 1, comprising:
A learning network establishment module for establishing a challenge representation learning network comprising a first encoder sub-network for mapping an input signal into a first hidden space code, a decoder sub-network for reconstructing the first hidden space code into a signal space, and a second encoder sub-network for mapping a reconstructed signal in the signal space into a second hidden space code;
The network training module is used for performing countermeasure learning training on the countermeasure representation learning network by taking the characteristic data of the healthy machine running signal as network training data, and taking the relative error of the first hidden space code and the second hidden space code as a mechanical damage index after the countermeasure learning training reaches a preset convergence condition to obtain a machine health state monitoring network;
the mechanical damage index detection module is used for inputting the operation signals of the target machine acquired in real time into the mechanical health state monitoring network after pretreatment to obtain mechanical damage indexes of the target machine;
And the monitoring result generation module is used for determining the monitoring result of the mechanical health state of the target machine according to the mechanical damage index of the target machine based on a preset mechanical damage index threshold.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210174862.0A CN114970598B (en) | 2022-02-24 | 2022-02-24 | Mechanical health state monitoring method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210174862.0A CN114970598B (en) | 2022-02-24 | 2022-02-24 | Mechanical health state monitoring method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114970598A CN114970598A (en) | 2022-08-30 |
| CN114970598B true CN114970598B (en) | 2024-04-30 |
Family
ID=82976028
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210174862.0A Active CN114970598B (en) | 2022-02-24 | 2022-02-24 | Mechanical health state monitoring method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114970598B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108647786A (en) * | 2018-07-10 | 2018-10-12 | 电子科技大学 | The rotating machinery on-line fault monitoring method of neural network is fought based on depth convolution |
| CN113158814A (en) * | 2021-03-26 | 2021-07-23 | 清华大学 | Bearing health state monitoring method based on convolution self-encoder |
| CN113435321A (en) * | 2021-06-25 | 2021-09-24 | 西安交通大学 | Method, system and equipment for evaluating state of main shaft bearing and readable storage medium |
| US11153566B1 (en) * | 2020-05-23 | 2021-10-19 | Tsinghua University | Variable bit rate generative compression method based on adversarial learning |
-
2022
- 2022-02-24 CN CN202210174862.0A patent/CN114970598B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108647786A (en) * | 2018-07-10 | 2018-10-12 | 电子科技大学 | The rotating machinery on-line fault monitoring method of neural network is fought based on depth convolution |
| US11153566B1 (en) * | 2020-05-23 | 2021-10-19 | Tsinghua University | Variable bit rate generative compression method based on adversarial learning |
| CN113158814A (en) * | 2021-03-26 | 2021-07-23 | 清华大学 | Bearing health state monitoring method based on convolution self-encoder |
| CN113435321A (en) * | 2021-06-25 | 2021-09-24 | 西安交通大学 | Method, system and equipment for evaluating state of main shaft bearing and readable storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114970598A (en) | 2022-08-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Pandarakone et al. | Deep neural network based bearing fault diagnosis of induction motor using fast Fourier transform analysis | |
| CN105758644A (en) | Rolling bearing fault diagnosis method based on variation mode decomposition and permutation entropy | |
| CN109583386A (en) | A kind of intelligent rotating mechanical breakdown depth network characterization discrimination method | |
| CN110823576A (en) | Mechanical anomaly detection method based on generation of countermeasure network | |
| CN112487910A (en) | Fault early warning method and system for nuclear turbine system | |
| Yang et al. | Intelligent diagnosis of rotating machinery faults-a review | |
| Sun et al. | A robust approach of multi-sensor fusion for fault diagnosis using convolution neural network | |
| CN112949402A (en) | Fault diagnosis method for planetary gear box under minimum fault sample size | |
| CN213149750U (en) | State evaluation and fault prediction system for steam turbine generator unit | |
| CN113468728B (en) | Variable pitch system fault prediction method based on neural network | |
| Souad et al. | Fault diagnosis of rolling element bearings using artificial neural network | |
| Hajnayeb et al. | Vibration measurement for crack and rub detection in rotors | |
| CN114970598B (en) | Mechanical health state monitoring method and device | |
| CN113776818B (en) | Offshore wind turbine generator system gear system fault diagnosis method based on Park transformation | |
| CN116611184A (en) | Fault detection method, device and medium for gear box | |
| Yang et al. | Intelligent diagnosis technology of wind turbine drive system based on neural network | |
| Jung et al. | Fault Diagnosis of Inter-turn Short Circuit in Permanent Magnet Synchronous Motors with Current Signal Imaging and Semi-Supervised Learning | |
| CN115753103A (en) | Fault diagnosis method and system based on standard self-learning data enhancement | |
| CN113447267B (en) | Vibration signal analysis-based complete gearbox state evaluation method and system | |
| Wang | Fault early warning of wind turbine generator based on residual autoencoder network | |
| Zheng et al. | Wavelet packet decomposition and neural network based fault diagnosis for elevator excessive vibration | |
| KR20210132486A (en) | Fault diagnosis method fof rotor system using multi task learning | |
| Zhao et al. | Health indicator selection and health assessment of rolling element bearing | |
| Yang et al. | Self-attention Parallel Fusion Network for Wind Turbine Gearboxes Fault Diagnosis | |
| Dou et al. | An Improved Fault Detection Method based on Canonical Variate Analysis for Tricky Faults of Wind Turbine |
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 |