CN112001314A - Early fault detection method for variable speed hoist - Google Patents
Early fault detection method for variable speed hoist Download PDFInfo
- Publication number
- CN112001314A CN112001314A CN202010861405.XA CN202010861405A CN112001314A CN 112001314 A CN112001314 A CN 112001314A CN 202010861405 A CN202010861405 A CN 202010861405A CN 112001314 A CN112001314 A CN 112001314A
- Authority
- CN
- China
- Prior art keywords
- signal
- time domain
- vibration signal
- fault
- dcnn
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 29
- 238000000034 method Methods 0.000 claims abstract description 82
- 238000001228 spectrum Methods 0.000 claims abstract description 22
- 238000011176 pooling Methods 0.000 claims description 27
- 238000001914 filtration Methods 0.000 claims description 26
- 230000036541 health Effects 0.000 claims description 23
- 238000012952 Resampling Methods 0.000 claims description 21
- 238000004364 calculation method Methods 0.000 claims description 17
- 238000004422 calculation algorithm Methods 0.000 claims description 15
- 238000012545 processing Methods 0.000 claims description 14
- 238000007781 pre-processing Methods 0.000 claims description 12
- 238000010606 normalization Methods 0.000 claims description 10
- 230000004913 activation Effects 0.000 claims description 7
- 238000005259 measurement Methods 0.000 claims description 6
- 230000002123 temporal effect Effects 0.000 claims 1
- 238000005070 sampling Methods 0.000 abstract description 23
- 238000003745 diagnosis Methods 0.000 abstract description 20
- 239000000284 extract Substances 0.000 abstract description 7
- 230000006870 function Effects 0.000 description 19
- 230000008569 process Effects 0.000 description 15
- 238000012544 monitoring process Methods 0.000 description 11
- 238000013527 convolutional neural network Methods 0.000 description 10
- 230000008859 change Effects 0.000 description 9
- 230000000737 periodic effect Effects 0.000 description 9
- 238000012549 training Methods 0.000 description 9
- 238000005311 autocorrelation function Methods 0.000 description 7
- 238000000605 extraction Methods 0.000 description 7
- 238000004458 analytical method Methods 0.000 description 6
- 238000013135 deep learning Methods 0.000 description 6
- 230000009467 reduction Effects 0.000 description 6
- 238000012360 testing method Methods 0.000 description 6
- 238000004088 simulation Methods 0.000 description 5
- 230000003595 spectral effect Effects 0.000 description 5
- 230000001052 transient effect Effects 0.000 description 5
- 238000004519 manufacturing process Methods 0.000 description 4
- 238000005457 optimization Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 230000004044 response Effects 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000015556 catabolic process Effects 0.000 description 2
- 238000009749 continuous casting Methods 0.000 description 2
- 238000000354 decomposition reaction Methods 0.000 description 2
- 238000006731 degradation reaction Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000001208 nuclear magnetic resonance pulse sequence Methods 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 238000005096 rolling process Methods 0.000 description 2
- 239000004576 sand Substances 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000009529 body temperature measurement Methods 0.000 description 1
- 239000003638 chemical reducing agent Substances 0.000 description 1
- 239000003245 coal Substances 0.000 description 1
- 230000001149 cognitive effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000003862 health status Effects 0.000 description 1
- 238000012880 independent component analysis Methods 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
- 239000010687 lubricating oil Substances 0.000 description 1
- 239000003550 marker Substances 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012821 model calculation Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000877 morphologic effect Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 239000003921 oil Substances 0.000 description 1
- 230000010287 polarization Effects 0.000 description 1
- 238000007430 reference method Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 230000001629 suppression Effects 0.000 description 1
- 238000011426 transformation method Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Computational Mathematics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Algebra (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Computational Linguistics (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Signal Processing (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Operations Research (AREA)
- Probability & Statistics with Applications (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses an early fault detection method of a variable speed hoist, which estimates the instantaneous rotation speed of the variable speed hoist by using a CPP (constant-current position propagation) method, performs equal-angle sampling on a vibration signal, eliminates the influence of the speed on a fault-related vibration signal, obtains an optimal model parameter, adaptively extracts weak fault impact characteristics from the vibration signal, avoids the problem that the fault impact characteristics are not obvious due to improper parameter setting, considers the successful application of DCNN (distributed computing network) and deep residual error learning in fault diagnosis, fuses residual error connection into DCNN (distributed computing network) and improves the performance of the DCNN model on fault classification. Time domain and frequency domain characteristics are extracted from the filtered signals respectively, time domain statistical characteristics are extracted, Hilbert spectrums of instantaneous frequency and amplitude characteristics and signal envelopes are extracted based on a Teager operator and input to different convolution layers of the DCNN respectively, useful information in the frequency domain characteristics and the time domain characteristics is extracted fully, and accuracy of early fault detection of a detection model is improved.
Description
Technical Field
The invention relates to the technical field of fault detection, in particular to an early fault detection method for a variable speed hoist.
Background
The mine hoist is key core equipment for coal mine production and mainly comprises a motor, a speed reducer, a roller, a head sheave, a hydraulic brake and the like. The method extracts relevant fault information from mechanical operation parameters such as vibration, pressure, temperature and the like, realizes monitoring of the operation state of the mine hoist, and is the main content of current hoist fault monitoring research. A large number of production practices and theoretical studies have shown that more than 70% of the faults are hidden in the vibration signal. However, the device temperature measurement signal has many important device state information, can reflect effective factors of the fault position and degree to a certain extent, and can be used for analyzing different fault types. For example, when a bearing fails, the return oil temperature and the vibration signal are often changed. The insufficient amount of lubricating oil leads to friction, some internal and external environments or working conditions and misalignment all can influence the temperature value, therefore synthesize and use temperature signal can improve equipment failure monitoring's accuracy and reliability.
Many parts of the rotating machine, such as gears and bearings, generate vibration signals with periodic impact during operation, and the periodic impact signals generate abnormal changes when the mechanical parts are defective. Therefore, noise suppression, signal-to-noise ratio improvement, and fault feature extraction research are attracting attention. Wavelet analysis and improvement techniques, Hilbert-Huang transform, fractal morphology, chirp transform, Local feature-scale decomposition (LCD), and the like. Wiggins proposed that Minimum Entropy Deconvolution (MED) can simultaneously extract the fault pulse and minimize noise, and can still effectively detect even in severe noise situations. Aiming at the problem that MED can only deconvolve single pulse characteristics, Maximum Correlation Kurtosis Deconvolution (MCKD) selects a finite impulse response filter to maximize the correlation kurtosis of a signal, and meanwhile enhances the periodicity of a specified periodic signal without the prior deconvolution in an autoregressive model stage. The MCKD can display periodic fault transients and locate the spectral kurtosis of the transients in the frequency domain to extract weak fault features masked by background noise, thereby achieving good effects in practical applications. Whereas the MCKD performance is heavily dependent on parameter selection. The Multipoint Optimal Minimum Entropy Deconvolution Adjusted (MOMEDA) method can realize that a non-iterative method directly solves an Optimal filter, can extract a pulse sequence and is used for determining the period of the fault occurrence by the Multipoint kurtosis value. MOMEDA can distinguish the health state and the fault state, has performance superior to MCKD and MED, and is suitable for fault detection of rotary machines. Although MOMEDA overcomes some of the MCKD problems, this approach is still affected by the pulse period, filter length selection.
In practice, the mechanical operation usually has the characteristics of speed change, load change, complex operation environment and the like, the vibration signal has the characteristics of complexity, nonlinearity, non-stationarity, low signal-to-noise ratio and the like, and the early fault characteristic is too weak and is often submerged in the environmental noise. Therefore, using multiple signal processing methods is a viable way to process complex vibration signals. For example, Shao Haidong (2020) extracts multi-fault features using an enhanced depth gate recursion unit and complex wavelets and diagnoses using a support vector machine. The method for extracting the vibration signal characteristics of the variable speed rotating machinery is characterized in that a method combining morphological component analysis and linear frequency modulation wavelet path tracking (CPP) is adopted by a citizen (2013) to solve the problem of extracting the vibration signal characteristics of the variable speed rotating machinery: aiming at the influence of acceleration and deceleration of mechanical equipment on vibration signal faults, Jiesi Luo (2012) proposes a gear fault monitoring method based on multi-scale CPP and fractal Fourier transform. The rolling bearing has the problems of complex working condition and large noise interference, and an article firstly adopts a wavelet transformation method to construct time-frequency graphs of vibration signals under different states, then uses a Deep Convolution Auto Encoder (DCAE) to eliminate time-frequency image noise, and utilizes CNN to classify faults of the time-frequency graphs. The Zhixiong Li (2013) firstly uses kernel independent component analysis (kira) to select a characteristic vibration signal component related to a fault source, and then uses a nonlinear analysis method WPT and EMD to process nonlinear and non-stationary vibration noise to extract an original signal characteristic vector.
The deep learning has strong capability in the aspects of representation and learning, can process massive high-dimensional data, has great success in the aspects of physiological signal processing, cognitive science, natural language processing, mechanical equipment fault diagnosis, industrial process monitoring, soft measurement and the like, and is receiving more and more attention. Compared with the conventional shallow learning method, the deep learning method solves the problems of large data band processing, lack of frequency characteristic priori knowledge and the like by using hierarchical nonlinear change, and has wider application range. The common Deep learning algorithms mainly include Deep Belief Networks (DBN), Stacked Denoising auto-coders (SDAE), Long Short Term Memory networks (LSTM), Deep Convolutional Neural Networks (DCNN), and Deep residual error networks (DRN). Among them, DCNN has been used in the field of fault diagnosis with its powerful data processing capability, and has achieved great success in image recognition. Aiming at the problems that the working condition change of the actual industrial process and the noise of the working environment influence the performance of the intelligent fault diagnosis model, the original time signal is used as the deep learning model to be input in the text [4], so that not only is the noise elimination processing not required in advance, but also a domain self-adaptive method or target domain information is not required. The existing signal processing method is difficult to process nonlinear and non-stationary characteristics of a rolling bearing caused by a complex working environment, Shuhui Wang proposes a hidden Markov classification method based on CNN (hidden Markov), wherein the CNN extracts data characteristics and uses t-distribution random neighbor embedding (SNE) to perform dimension reduction, and HMM is used as a tool with stronger classification capability for fault diagnosis. Although deep learning has abundant results in the aspect of variable-working-condition mechanical fault diagnosis research, the signal frequency is changed unstably due to speed change, so that the data essence cannot be well represented by the characteristics extracted by the deep learning method, and the model performance is reduced. Aiming at operating equipment under variable working conditions, some scholars propose various improved deep learning fault diagnosis algorithms, for example, aiming at the condition that spectral features and statistical features cannot reflect mechanical characteristics of the variable working conditions, Sai Ma (2019) proposes fault diagnosis of a planetary gearbox through deep residual error learning of demodulation time-frequency features under the variable working conditions, wherein the deep residual error network training speed-frequency represents the spectral features, and a generalized demodulation operator eliminates rotation speed change. Aiming at the multi-scale characteristics of the vibration signals related to the faults, the Xiaoan Yan (2020) provides a multi-scale cascade DBN model for extracting the characteristics of the vibration signals in a sliding window on each scale, so that the wider representation of the signals is achieved, and the variable working condition signal characteristic representation capability is improved. Despite advances in variable condition fault diagnosis research, variable condition mechanical fault diagnosis remains a challenging problem.
Disclosure of Invention
Aiming at the problems, the invention provides an early fault detection method for a variable speed hoist.
In order to achieve the purpose of the invention, the invention provides an early fault detection method of a variable speed hoisting machine, which comprises the following steps:
and S10, acquiring a health state vibration signal of the variable speed hoisting machine from the sensor, and performing normalization pretreatment on the health state vibration signal to obtain a first pretreatment signal.
S20, acquiring the instantaneous speed of the first preprocessed signal by a CPP method, resampling the vibration signal in the health state by using a step ratio tracking method based on the instantaneous speed to obtain a first resampled signal, calling an IMOMEDA algorithm to perform filtering processing on the first resampled signal, determining a first MOMEDA model parameter, and obtaining an IMOMEDA filtered signal.
And S30, extracting the time domain statistical characteristics of the IMOMEDA filtering signal by adopting the first MOMEDA model parameters.
S50, using Hilbert spectrum values of IMOMEDA filtering signals as input of a preset DCNN, performing convolution operation and weight sharing with a local view field, performing maximum pooling layer calculation on the output of a convolution layer, using the output of a 1 st pooling layer and time domain statistical characteristics as input of a 2 nd-layer convolution, performing maximum pooling layer calculation on the output of the convolution layer, calculating errors of a target label and the output of a last layer, and adjusting model parameters according to an error back propagation method until convergence to obtain a detection model.
And S60, acquiring the vibration signal to be detected of the variable speed hoist, extracting the time domain statistical characteristics of the vibration signal to be detected to obtain the statistical characteristics to be detected, inputting the statistical characteristics to be detected into the detection model, and determining the health state of the variable speed hoist according to the output of the detection model.
In one embodiment, step S60 is preceded by:
constructing DCNN containing 3 convolution layers and pooling layers, selecting ReLU function as activation function, and initializing model parameters.
In one embodiment, the extracting the time-domain statistical characteristic of the IMOMEDA filtered signal using the first MOMEDA model parameters includes:
calculating the instantaneous amplitude of the IMOMEDA filtering signal according to the Teager operator, and calculating the amplitude entropy and the frequency entropy according to the instantaneous amplitude of the signal;
and resampling the time domain statistics of the vibration signal in the health state according to the amplitude entropy and the frequency entropy to obtain the time domain statistical characteristics of the IMOMEDA filtering signal.
In one embodiment, the normalization preprocessing of the state of health vibration signal comprises:
xn＝(xn-min(x))/(max(x)-min(x))，
wherein x isnRepresenting the first pre-processed signal, x representing the healthy state vibration signal, min () representing the minimum value, max () representing the maximum value.
In one embodiment, extracting the time domain statistical characteristics of the vibration signal to be detected comprises:
carrying out normalization preprocessing on the vibration signal to be detected to obtain a second preprocessing signal;
acquiring the instantaneous speed of a second preprocessing signal by adopting a CPP (constant-current position measurement) method, resampling the vibration to be detected by using a step ratio tracking method based on the instantaneous speed to obtain a second resampling signal, calling an IMOMEDA (inertial measurement architecture) algorithm to filter the second resampling signal, determining a second MOMEDA (model of average) model parameter, and obtaining a filtering signal to be detected;
and extracting the time domain statistical characteristics of the filtering signal to be detected by adopting the second MOMEDA model parameters to obtain the time domain statistical characteristics of the vibration signal to be detected.
The method for detecting the early failure of the variable speed elevator is an elevator health state diagnosis method based on an improved MOMEDA (improved MOMEDA) and a residual connection Deep Convolutional Neural Network (DCNN); the method comprises the steps of estimating the instantaneous rotation speed of the variable speed hoist by using a CPP (continuous casting process) method, carrying out equal-angle sampling on a vibration signal according to the rotation speed, eliminating the influence of the speed on a fault-related vibration signal, obtaining an optimal model parameter through iteration, extracting weak fault impact characteristics from the vibration signal in a self-adaptive manner, avoiding the problem that the fault impact characteristics are not obvious due to improper parameter setting, integrating residual connection into a DCNN (distributed computing network) by considering the successful application of the DCNN and deep residual learning in fault diagnosis, and improving the performance of the DCNN model on fault classification. Time domain and frequency domain characteristics are extracted from the filtered signals respectively, time domain statistical characteristics are extracted, transient frequency and amplitude characteristics and Hilbert spectrums enveloped by the signals are extracted based on a Teager operator and input to different convolution layers of the DCNN respectively, useful information in the frequency domain characteristics and the time domain characteristics is extracted fully, and accuracy of early fault detection of a detection model is improved.
Drawings
FIG. 1 is a flow diagram of an embodiment of a method for early fault detection of a variable speed hoist;
FIG. 2 is a flow diagram of an embodiment elevator condition monitoring implementation;
fig. 3 is a schematic structural diagram of a DCNN network according to an embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.
Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. It is explicitly and implicitly understood by one skilled in the art that the embodiments described herein can be combined with other embodiments.
Referring to fig. 1, fig. 1 is a flowchart of an early failure detection method for a variable speed hoist according to an embodiment, including the following steps:
and S10, acquiring a health state vibration signal of the variable speed hoisting machine from the sensor, and performing normalization pretreatment on the health state vibration signal to obtain a first pretreatment signal.
In one embodiment, the normalization preprocessing of the state of health vibration signal comprises:
xn＝(xn-min(x))/(max(x)-min(x))，
wherein x isnRepresenting the first pre-processed signal, x representing the healthy state vibration signal, min () representing the minimum value, max () representing the maximum value.
S20, acquiring the instantaneous speed of the first preprocessed signal by a CPP method, resampling the vibration signal in the health state by using a step ratio tracking method based on the instantaneous speed to obtain a first resampled signal, calling an IMOMEDA algorithm to perform filtering processing on the first resampled signal, determining a first MOMEDA model parameter, and obtaining an IMOMEDA filtered signal.
And S30, extracting the time domain statistical characteristics of the IMOMEDA filtering signal by adopting the first MOMEDA model parameters.
In one embodiment, the extracting the time-domain statistical characteristic of the IMOMEDA filtered signal using the first MOMEDA model parameters includes:
calculating the instantaneous amplitude of the IMOMEDA filtering signal according to the Teager operator, and calculating the amplitude entropy and the frequency entropy according to the instantaneous amplitude of the signal;
and resampling the time domain statistics of the vibration signal in the health state according to the amplitude entropy and the frequency entropy to obtain the time domain statistical characteristics of the IMOMEDA filtering signal.
Specifically, the calculation process of the magnitude entropy and the frequency entropy may include:
in the formula, esiRepresenting the magnitude entropy, ai(t) denotes signal amplitude, N denotes signal length, efiRepresenting frequency entropy, fi(t) represents the frequency of the signal.
The time domain statistic sampling process of the resampled vibration signal comprises the following steps:
Crest factor:maxyi/RMS，
P-P:ymax-ymin，
where RMS represents the standard deviation of the resampled signal, yiRepresenting the ith sample point of the resampled vibration signal.
S50, using Hilbert spectrum values of IMOMEDA filtering signals as input of a preset DCNN, performing convolution operation and weight sharing with a local view field, performing maximum pooling layer calculation on the output of a convolution layer, using the output of a 1 st pooling layer and time domain statistical characteristics as input of a 2 nd-layer convolution, performing maximum pooling layer calculation on the output of the convolution layer, calculating errors of a target label and the output of a last layer, and adjusting model parameters according to an error back propagation method until convergence to obtain a detection model.
The above steps can calculate the Hilbert spectrum characteristics of IMOMEDA filtering signals, and normalize the characteristics for subsequent processing.
And S60, acquiring the vibration signal to be detected of the variable speed hoist, extracting the time domain statistical characteristics of the vibration signal to be detected to obtain the statistical characteristics to be detected, inputting the statistical characteristics to be detected into the detection model, and determining the health state of the variable speed hoist according to the output of the detection model.
The step can adopt a time domain statistical characteristic obtaining process similar to the vibration signal in the health state to obtain the vibration signal to be detected of the variable speed lifter.
In one embodiment, extracting the time domain statistical characteristics of the vibration signal to be detected comprises:
carrying out normalization preprocessing on the vibration signal to be detected to obtain a second preprocessing signal;
acquiring the instantaneous speed of a second preprocessing signal by adopting a CPP (constant-current position measurement) method, resampling the vibration to be detected by using a step ratio tracking method based on the instantaneous speed to obtain a second resampling signal, calling an IMOMEDA (inertial measurement architecture) algorithm to filter the second resampling signal, determining a second MOMEDA (model of average) model parameter, and obtaining a filtering signal to be detected;
and extracting the time domain statistical characteristics of the filtering signal to be detected by adopting the second MOMEDA model parameters to obtain the time domain statistical characteristics of the vibration signal to be detected.
Specifically, the calculation process of the magnitude entropy and the frequency entropy may include:
in the formula, esiRepresenting the magnitude entropy, ai(t) denotes signal amplitude, N denotes signal length, efiRepresenting frequency entropy, fi(t) represents the frequency of the signal.
The time domain statistic sampling process of the resampled vibration signal comprises the following steps:
Crest factor:max|yi|/RMS，
P-P:ymax-ymin，
where RMS represents the standard deviation of the resampled signal yiRepresenting the ith sample point of the resampled signal.
In one embodiment, step S60 is preceded by:
constructing DCNN containing 3 convolution layers and pooling layers, selecting ReLU function as activation function, and initializing model parameters.
In this embodiment, a DCNN including 3 convolutional layers and pooling layers may be constructed according to the residual connection DCNN theory, a ReLU function is selected as the activation function, and then model parameters are initialized.
The method for detecting the early failure of the variable speed elevator is an elevator health state diagnosis method based on an improved MOMEDA (improved MOMEDA) and a residual connection Deep Convolutional Neural Network (DCNN); the method comprises the steps of estimating the instantaneous rotation speed of the variable speed hoist by using a CPP (continuous casting process) method, carrying out equal-angle sampling on a vibration signal according to the rotation speed, eliminating the influence of the speed on a fault-related vibration signal, obtaining an optimal model parameter through iteration, extracting weak fault impact characteristics from the vibration signal in a self-adaptive manner, avoiding the problem that the fault impact characteristics are not obvious due to improper parameter setting, integrating residual connection into a DCNN (distributed computing network) by considering the successful application of the DCNN and deep residual learning in fault diagnosis, and improving the performance of the DCNN model on fault classification. Time domain and frequency domain characteristics are extracted from the filtered signals respectively, time domain statistical characteristics are extracted, instantaneous frequency and amplitude characteristics and Kilbert spectrums of the signals are extracted based on a Teager operator and are input to different convolution layers of the DCNN respectively, useful information in the frequency domain characteristics and the time domain characteristics is extracted fully, and accuracy of early fault detection of a detection model is improved.
In one embodiment, the above method for detecting early failure of a variable speed hoist involves related theoretical knowledge including:
1.1 CPP-to-order ratio tracking based Signal resampling
The linear frequency modulation wavelet path tracking algorithm estimates the instantaneous rotation speed change, performs equal-angle resampling on non-stationary signals according to rotation speed signals, eliminates the influence of the rotation speed change on a vibration signal spectrogram, avoids the loss of some characteristic information in the vibration signals, and is an important technology for improving the vibration signal analysis and fault diagnosis performance of the rotary machine.
The linear frequency modulation wavelet path tracking algorithm is a method for detecting non-stationary signals from weak signals by self-adaptive signals. The method comprises the following steps of firstly establishing a linear frequency modulation wavelet atom library:
here, I ═ kN2-j,(k+1)N2-j]For dynamic support regions, N is the sample length, j is 0,1,2, …, log2N-5， k＝0,1,2,…,2j-1；1I(t) is a rectangular window function, 1 when t ∈ II(t) 1, otherwise 1I(t) is 0. Obviously, given the parameter aμAnd bμInstantaneous frequency if (t) ═ aμ+2bμt. According to the sampling theorem, aμ+2bμt≤fs/2，fsFor the sampling theorem, the sampling time t is n/fs＝nΔt，Δt＝1/fsT is 0,1, …, N-1. The chirp wavelet decomposition is the inner product of signal f (t) and chirp wavelet atoms, i.e.
It is noted that the support region I is a plurality of intervals evenly dividing the sample in the time domain.
Assuming that the vibration signal is x (t), assuming that the cycle number of each rotation, namely the order ratio is recorded as O, and the maximum order ratio is recorded as OmaxAt a rotational speed of nr(rpm/min), the relationship between the signal frequency f and the order ratio O is
Assume that the number of samples collected per revolution is NsThen the sampling angle interval is Δ θ ═ 2 π/NsSatisfy Δ θ ≧ π/DmaxWherein D ismaxIs the maximum analysis order ratio. As defined by the sampling frequency of the time domain signal, the sampling order ratio is the inverse of the sampling angle interval, i.e., Os1/Δ θ. It is clear that the following relationship exists between them:
given sampling order ratio OsSignal sampling frequency fsAnd a rotational speed nr(T) the total time length T of time domain samples of the signal, the signal length after resampling of the signal is calculated by
Given signal sampling frequency fsThe number of sampling time domain samples in the time T is NT＝fsAnd T. Assuming that the time domain sampling time is 0 moment, the phase-locked time of the time domain sampling sample is marked as Tn＝nTs(n＝0,1,…,NT-1). The time period corresponding to the angle equal interval sampling is TΔθT/N, corresponding to a phase-locked time scale of the resampled samples of
Given healthy phase time scale
The phase-healthy time marker close to the nearest time domain sampling point is TiAnd Ti+1Then the samples are resampled
Calculated using Lagrange's linear interpolation formula
The rotational speed of the rotor can be measured using a tachometer, and the CPP is used to estimate the rotational speed from the vibration signal. The method is based on the assumption that the transient frequency of a non-stationary signal exhibits an approximately linear transformation over a period of time, so that the transient frequency of the signal can be estimated in segments using a frequency-modulated wavelet yard with a linear transformation in frequency.
1.2Teager energy operator
The Teager energy operator demodulation method can effectively separate and not detect amplitude modulation and frequency modulation information of single components and modulation signals. The energy operator demodulation method can calculate the signal energy at any moment, has good time resolution and self-adaptive capacity to the transient change of the signal, and can enhance the weak transient impact component.
The modulated AM-FM signal x (t) for the time-varying amplitude a (t) and the time-varying phase phi (t) is denoted as
Here omegacFor the carrier frequency, a (t) is a slowly varying amplitude signal, q (u) is a normalized frequency modulation signal, i.e., | q (u) | ≦ 1, and the highest frequencies of a (t) and q (u) are both much less than the carrier frequency. Then the energy operator demodulation is defined as
Wherein
And
first and second derivatives of x (t), respectively. The discrete form of the above formula is psid[x(n)]＝(x(n))2-x (n-1) x (n + 1). Its instantaneous amplitude a (t) and instantaneous frequency omegai(t) is
Further, since the resampled signal still contains noise, we perform de-noising processing on the resampled signal using the following theory:
1 improved multipoint optimal minimum entropy deconvolution
For the fault cycle impact vibration signal y of the rotating machine, the sensor measures the output signal x, and the signals have the following relationship
x＝h*y+e (7)
Where h is the system response and e is the ambient noise interference. The goal of MOMEDA is to find an optimal FIR filter to perform deconvolution of multiple pulses at known locations
y＝f*x (8)
Wherein f ═ f1,f2,L,fL]TIs a vector of FIR filter impulse response coefficients. The above formula can be rewritten as a matrix representation
To obtain multiple pulse signals, the MOMEDA optimization problem has the following form
Where t is an invariant vector defining the position and weight of the deconvolved target pulse, whose length is equal to the length of the output signal y. For example, t is [0,0,1,0,0,1,0,0,0]TIndicating that 2 pulses need to be deconvolved in the output signal, one position is n-3 and the other position is n-6. By introducing the vector t to specify the location to deconvolve the impact signal, the window segment of the deconvolved separation, localization and processing signal can also be controlled. When used as a window for processing signals, the running machine may be allowed to run at a fluctuating speed. For example, for a periodic pulse sequence p (t), when t ═ 111111]P (t) indicates that the periodic sequence signal takes a rectangular window, so that only the sequences within the window are deconvolved.
Notice tTy/||t||||y||＝(t/||t||)T(y/| y |), the value of the function MDN (y, t) shown in formula () is [0,1 | ]]. It is clear that the optimal solution of the filter can be obtained when its value reaches 1. The extreme value of the optimization problem is obtained by the method of deriving f, namely
Here, the
Mk＝[xk+L-1,xk+L-2,L,xk]T. Notice t1M1+t2M2+L+tN-LMN-L＝X0t，y＝X0f equals 0, thus having
It is obvious that f is the optimal solution when the objective function value shown in equation (14) reaches the maximum value of 1. Note that t is constant, then the optimal solution for the MOMEDA optimization problem is
When filtering the noise signal using MOMEDA, the pulse signal period is first determined. In addition, the energy of the periodic impact signal is determined by the length of the filter, and improper setting of the length of the filter can increase or weaken the energy of the original impact signal, thereby causing misdiagnosis.
In practical application of MOMEDA, the following appropriate parameters need to be determined in advance:
(1) window function w: the selection of the window function value influences the spectrum clearness and the extraction precision of the fault impact characteristics, and generally does not exceed 5.
(2) The fault correlation period is determined using an autocorrelation function. The autocorrelation function is
rx(τ)＝∫x(t)x(t+τ)dt (15)
Wherein, the lag τ, rx(0) Representing a global maximum. When r isx(τ) from a plurality of local extrema points, the sequence can be considered periodic, taking the point τ corresponding to the local maximummaxCan be regarded as the reasonBarrier characteristic period Tc. Note that in practical applications, the value of τ is larger than the actual failure period. In addition, because the signals acquired in the field contain serious noise, the autocorrelation function has a plurality of local values, and the estimation is often approximate to the fault period value.
(3) Length of the filter: this parameter directly affects the accuracy of the extraction of the fault-impact sequence, the length of which is determined by the following formula
L＞nfs/fc＝nTc/Ts (16)
Wherein f iss＝1/TsRepresenting the sampling frequency of the vibration signal, fc＝1/TcIndicating a fault-related characteristic frequency, TcThe fault characteristic period is shown, and the factor n is more than or equal to 5 and is generally an integer. In experimental simulations, we found that the filter length satisfies the above condition, and generally, n is 5: 10.
Period of optimum failure TcThe determination method mainly comprises two methods: 1 is a theoretical calculation method; second using the following correlation entropy measure
And calculating an envelope autocorrelation function of the filtered signal y to determine the fault-related characteristic period.
Mechanical gears are important devices of mechanical systems and are also components that often fail. Therefore, to better determine the optimal failure period, we use the failure signature gear failure signature index for the determination, which is defined as
KR＝Kur(y)+λRf (18)
Wherein R isf(S (f) + S (2f) + S (3f))/S, f fault signature frequency, S is the total amplitude of the envelope spectrum; rfRepresenting the fault characteristic proportion of the Hilbert envelope spectrum; kur (·) represents a kurtosis value of the time-domain signal; λ is a weighting coefficient that prevents time domains Kur (-) and RfThe difference between them is too large, resulting in the problem that the bias is to occupy the dominant index and ignore the other indexes.
Taking into account TcFor MOMEDA propertyThe influence is great, an iterative method is used for solving a filtering signal y, then an autocorrelation function of y and KR are calculated, and when the KR is maximum, the corresponding tau is obtainedmaxIs the optimal fault characteristic cycle.
Since the original data has serious noise, it is obviously difficult to determine the optimal T simply by calculating the fault characteristic period according to the original datac. Method for determining T by searching methodcThe method has a large calculation amount. To reduce the amount of calculation, T is automatically identifiedcIn this embodiment, an iterative method is used to obtain the optimal value. The IMOMEDA algorithm is realized by the following detailed steps:
1) inputting a vibration signal x, determining the size of a window, a weighting coefficient lambda, the iteration number i equal to 1, and the maximum iteration number iter;
2) calculating the autocorrelation function of equation (15) to obtain TiComputing KR of signal xiIndexes;
3) calculating X0,
Determining t and a filter length L;
4) calculating filter impulse response coefficients according to equation (14);
5) filtered signal y according to equation (10);
6) calculating the autocorrelation function of y to determine Ti+1That is, its index KRi+1；
7) If i is less than iter, i is i +1, and the step 3) is switched to;
otherwise, stopping the algorithm;
8) selecting the T corresponding to the largest KRiAs an optimal fault characteristic period, the filtered signal is used as a final signal;
the fault is measured using a multi-point kurtosis index of the form:
2 connecting DCNN elevator state diagnosis based on IMOEDA and residual error
2.1 residual connected deep convolutional neural network
DCNN is a deep feedforward neural network model, the basic principle of which is similar to that of a convolutional neural network, and the essence of the model is that a self-learning mechanism realizes layer-by-layer extraction from input data through hidden layer processing. Compared with the traditional neural network, the neural network is composed of a plurality of convolution layers and a pooling layer (polarization layer), a Fully Connected Layer (FCL) and an output layer, wherein the convolution layers realize feature extraction by performing convolution calculation on input data and a plurality of convolution kernels, the low-layer convolution features represent a macroscopic data structure of the data, and the high-layer convolution features represent complex and abstract features of the data. The convolution calculation method is
Wherein,
representing the output image characteristics j, b of the current layer llThe offset is represented by the number of bits in the bit,
convolution kernel, RiRepresenting the input feature image set, f (-) represents the activation function, and there are 4 functions in common use:
ReLU function is used in simulation experiment to accelerate convergence speed of DCNN.
The pooling model is essentially a vertical sampling process for reducing the amount of computation and eliminating redundant features. The pooling model mainly has 3 forms of average, maximum and norm, and the maximum mode is adopted in the simulation experiment:
s denotes a window. The role of the pooling layer not only reduces the number of input nodes and training parameters of the next layer,but also keeps the maximum local characteristics and improves the popularization performance of the network. After the full connection layer is positioned at the last pooling layer, 1-2 FCLs are generally used for fusing local information of different types. FCL is typically implemented using a neural network whose mathematical representation is in the form of
Where σ (-) is the activation function,
for the connection weight, the remaining parameters are like equation (). The output layer is the last layer of the DCNN, which passes the FCL output to the output layer. The output layer is typically sorted using a softmax function. For the multi-class classification problem, assume that there are n classes and n corresponding class labels, and the training data set is
ylE.g. {0,1, L, n }. Given a training data sample x, then the probability that x belongs to class i is p (y ═ i | x), and softmax is returned as output
Wherein,
the effect is normalization. Model parameters
The following optimization problem is solved
The DCNN uses an error back propagation algorithm to optimize the main parameters including convolution kernel, bias, weights, etc. Loss function using mean squareAnd (4) error. Because the DCNN can generate a degradation phenomenon during training, residual connection is introduced into jump connection in the text, low-level output characteristics are introduced into high levels, the problem of vanishing gradient and degradation caused by too deep network is solved to a great extent, and network convergence is accelerated. In addition, the Hilbert spectrum h of the vibration signal is used as the 1D input of the DCNN, and the dimension of the Hilbert spectrum h is far greater than the time domain statistical characteristic ftDimension (d) of (a). Therefore, the input dimension reduction output h' is realized by carrying out convolution and pooling operations on the DCNN input, and then the input dimension reduction feature and the time domain feature f are combinedtFusion of [ h', ft]As the input of the next layer, the problem of unbalanced quantity of frequency domain features and time domain features is solved, and time domain and frequency domain feature information is better utilized. The 2 nd convolutional layer adopts 5-channel convolutional kernels, and the output result of the convolutional layer is 2D data. The 3 rd layer convolution layer adopts a degree multi-channel 2D convolution kernel, and the advantage of processing 2D data by DCNN is better utilized.
And carrying out envelope spectrum demodulation on the energy operator to obtain envelope signals of each frequency band, and then calculating corresponding envelope spectrums by using Fourier transform, thereby extracting mechanical fault envelope spectrum characteristics.
Further, referring to fig. 2, fig. 2 is a flow chart of a state monitoring implementation of the hoist (variable speed hoist) in the present embodiment. As can be seen from fig. 2, the condition monitoring process is roughly divided into 4 steps: (1) and (5) preprocessing a vibration signal. Considering that the working process of the hoister has the characteristic of variable speed, the vibration signal of the hoister acquired on site has strong nonlinearity and non-stationarity, and the CPP and order ratio analysis technology introduced by 1.1 is used for resampling the vibration signal to eliminate the non-stationarity problem caused by the vibration signal due to the rotating speed. (2) Data denoising and feature extraction. The original vibration signal is filtered by using the IMOMEDA method proposed in part 1 of this embodiment, so as to enhance the periodic impact component. Resampling the vibration signal to extract multi-order statistics, extracting instantaneous amplitude, instantaneous frequency entropy and Hilbert spectrum of the IMOMEDA filtered vibration signal based on a Teager operator (see step 5 of the discrete modeling stage below), and generating training samples and test sample data. In the sample feature set, the Hilbert spectrum is a frequency domain feature, and the rest are time domain statistical features, so that the essential attributes of faults can be revealed from multiple visual angles, and the extraction capability of early fault feature signals is improved. (3) Training of residual error connection DCNN model: according to the theory of residual connection deep convolutional neural network introduced in 2.1, a DCNN network structure is determined according to fig. 3, a vibration signal envelope spectrum processed by IMOMEDA is used as 1D input of DCNN, a convolutional layer 1 adopts 1 convolutional kernel and input data for calculation, dimension reduction of the input data is realized through pooling operation, and the convolutional layer and the resampling vibration signal time domain statistical characteristics are used as input of a convolutional layer of the next layer and are transmitted to a convolutional layer of the highest layer in a residual connection mode. And training the DCNN model according to an error back propagation algorithm. (4) And classifying the test samples according to the established residual connection DCNN model, and determining the state of the elevator. And (3) inputting the frequency domain characteristics (Hilbert spectrum of IMOMEDA filtering signals) of the test sample as a DCNN model, inputting the time domain statistical characteristics and the layer 1 pooling layer output in the test sample as a layer 2 convolution layer together, and obtaining the classification output shown in the formula (22) through model calculation. The class to which the test sample belongs is determined from the output of equation (22) by
Here, i ∈ {0,1, L, K }, where K is the number of categories. c is the state category of the elevator. The residual connection DCNN shown in fig. 3 is composed of 3 convolutional layers and pooling layers, 1 fully-connected layer, and a softmax regression layer. The input of fig. 3 is 1D, taking the Hilbert spectrum h of the vibration signal as the DCNN input, the convolution sum is also a 1D convolution kernel. The dimension reduction (denoted by h') of the input data is thus achieved by performing convolution and pooling operations on the DCNN input. Then, inputting the dimension reduction feature and the time domain statistical feature ftFusion of [ h', ft]As the input of the next layer, the problem of unbalanced quantity of frequency domain features and time domain features is solved, and time domain and frequency domain feature information is better utilized. The 2 nd convolution layer adopts a multi-channel convolution kernel, and the output result of the convolution layer is 2D data. The 3 rd convolution layer adopts a multi-channel 2D convolution kernel, and the advantage of processing 2D data by DCNN is better utilized. Model layer 2 convolution output and layer 3 convolution outputAre all 2D data. The layer 3 convolution input is the result of adding the layer 2 convolution input to each column of the layer 2 convolution output. The node numbers of the Softermax layer and the full connection layer need to be repeatedly tested in practice to obtain an optimal value.
Further, the implementation steps of the state diagnosis (early fault detection of the variable speed hoist) of the variable speed hoist based on the IMOEDA and the DCNN are divided into two stages, namely an off-line model training stage and an on-line model, and the detailed implementation steps are as follows:
an off-line modeling stage:
1) acquiring a health state vibration signal of the elevator from a sensor, and carrying out normalization pretreatment:
xn＝(xn-min(x))/(max(x)-min(x))
2) acquiring the instantaneous velocity of the signal using the CPP method described in section 1.1;
3) the original signal is resampled based on the instantaneous velocity using the order ratio tracking method described in section 1.1.
4) And calling an IMOMEDA algorithm provided in section 1 of the part to filter the resampled signal and determine MOMEDA model parameters.
5) Extracting time domain statistical characteristics, which are defined as follows:
calculating IMOMEDA filtering signal instantaneous amplitude value calculation amplitude entropy and frequency entropy according to Teager operator
Time domain statistics of the resampled vibration signal:
Crest factor:max|yi|/RMS
P-P:ymax-ymin
hilbert spectral features of the IMOMEDA filtered signal are calculated and normalized.
6) According to 2.1, a DCNN comprising 3 convolutional layers and pooling layers is constructed by introducing a residual connecting DCNN theory, a ReLU function is selected as an activation function, and model parameters are initialized by a reference method.
7) The Hilbert spectral values of the filtered signal are used as input to the DCNN, convolved with the local field of view and weight-shared. The convolutional layer output is subjected to maximum pooling layer calculation.
8) And taking the output of the layer 1 pooling layer and the extracted features of the step 5 as the input of the layer 2 convolution, performing convolution operation with a local view field and sharing weight. The convolutional layer output is subjected to maximum pooling layer calculation.
9) And calculating the error of the target label and the output of the last layer.
10) And adjusting the model parameters according to an error back propagation method until convergence.
And (3) an online monitoring stage:
as can be seen from fig. 2, the online monitoring phase is the same as the offline modeling 1) -5).
6) And inputting the extracted features into the DCNN model established off line, and calculating the model output.
7) And determining the health state of the elevator according to the model output.
The present embodiment provides an elevator health status diagnosis method based on an improved MOMEDA (improved MOMEDA) and a residual connected Deep Convolutional Neural Network (DCNN). The method estimates the instantaneous rotating speed of the hoister by using a CPP method, and carries out equal-angle sampling on the vibration signals according to the rotating speed, so as to eliminate the influence of the speed on the fault-related vibration signals. The IMOMEDA method is provided, and the method extracts weak fault impact characteristics from a vibration signal in a self-adaptive manner by iteratively solving optimal model parameters, so that the problem that the fault impact characteristics are not obvious due to improper parameter setting is avoided. Considering the successful application of the DCNN and the deep residual learning in fault diagnosis, residual connection is merged into the DCNN, and the performance of the DCNN model on fault classification is improved. Respectively extracting time domain and frequency domain characteristics from the filtered signals, namely, instant domain statistical characteristics, extracting instantaneous frequency and amplitude characteristics and Kilbert spectrums of the signals based on a Teager operator, respectively inputting the instantaneous frequency and amplitude characteristics and the Kilbert spectrums into different convolution layers of the DCNN, fully extracting useful information in the frequency domain characteristics and the time domain characteristics, and improving the accuracy of model diagnosis. And finally, introducing the construction of a test platform and carrying out experimental simulation, wherein the simulation experiment result verifies the effectiveness of the monitoring method provided by the embodiment.
The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, however, as long as there is no such combination, the scope of the present description should be considered as being described in the present specification.
It should be noted that the terms "first \ second \ third" referred to in the embodiments of the present application merely distinguish similar objects, and do not represent a specific ordering for the objects, and it should be understood that "first \ second \ third" may be interchanged under a specific order or sequence where permitted. It should be understood that "first \ second \ third" distinct objects may be interchanged under appropriate circumstances such that the embodiments of the application described herein may be implemented in an order other than those illustrated or described herein.
The terms "comprising" and "having" and any variations thereof in the embodiments of the present application are intended to cover non-exclusive inclusions. For example, a process, method, apparatus, product, or device that comprises a list of steps or modules is not limited to the listed steps or modules, but may alternatively include other steps or modules not listed or inherent to such process, method, product, or device.
The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent shall be subject to the appended claims.
Claims (5)
1. An early failure detection method for a variable speed hoist is characterized by comprising the following steps:
s10, acquiring a health state vibration signal of the variable speed hoist from a sensor, and performing normalization pretreatment on the health state vibration signal to obtain a first pretreatment signal;
s20, acquiring the instantaneous speed of the first preprocessed signal by a CPP method, resampling the vibration signal in the health state by using a step ratio tracking method based on the instantaneous speed to obtain a first resampled signal, calling an IMOMEDA algorithm to perform filtering processing on the first resampled signal, determining a first MOMEDA model parameter, and obtaining an IMOMEDA filtered signal;
s30, extracting time domain statistical characteristics of the IMOMEDA filtering signal by adopting the first MOMEDA model parameters;
s50, using Hilbert spectrum values of IMOMEDA filtering signals as input of a preset DCNN, performing convolution operation and weight sharing with a local view field, performing maximum pooling layer calculation on the output of a convolution layer, using the output of a 1 st pooling layer and time domain statistical characteristics as input of a 2 nd-layer convolution, performing maximum pooling layer calculation on the output of the convolution layer, calculating errors of a target label and the output of a last layer, and adjusting model parameters according to an error back propagation method until convergence to obtain a detection model;
and S60, acquiring the vibration signal to be detected of the variable speed hoist, extracting the time domain statistical characteristics of the vibration signal to be detected to obtain the statistical characteristics to be detected, inputting the statistical characteristics to be detected into the detection model, and determining the health state of the variable speed hoist according to the output of the detection model.
2. The method for detecting the early failure of the variable speed hoist according to claim 1, wherein step S60 is preceded by the step of:
constructing DCNN containing 3 convolution layers and pooling layers, selecting ReLU function as activation function, and initializing model parameters.
3. The method of early fault detection for variable speed elevators of claim 1, wherein extracting the temporal statistical characteristics of the imomda filtered signal using the first MOMEDA model parameters comprises:
calculating the instantaneous amplitude of the IMOMEDA filtering signal according to the Teager operator, and calculating the amplitude entropy and the frequency entropy according to the instantaneous amplitude of the signal;
and resampling the time domain statistics of the vibration signal in the healthy state according to the amplitude entropy and the frequency entropy to obtain the time domain statistical characteristics of the IMOMEDA filtering signal.
4. The method of early fault detection for variable speed hoists of claim 1, wherein normalizing the state of health vibration signal comprises:
xn＝(xn-min(x))/(max(x)-min(x))，
wherein x isnRepresenting the first pre-processed signal, x representing the healthy state vibration signal, min () representing the minimum value, max () representing the maximum value.
5. The method of claim 1, wherein extracting the time domain statistical characteristic of the vibration signal to be detected comprises:
carrying out normalization preprocessing on the vibration signal to be detected to obtain a second preprocessing signal;
acquiring the instantaneous speed of a second preprocessing signal by adopting a CPP (constant-current position vector) method, resampling the vibration to be detected by using a step ratio tracking method based on the instantaneous speed to obtain a second resampling signal, calling an IMOMEDA (inertial measurement architecture) algorithm to filter the second resampling signal, determining a second MOMEDA (model of average) model parameter, and obtaining a filtering signal to be detected;
and extracting the time domain statistical characteristics of the filtering signal to be detected by adopting the second MOMEDA model parameters to obtain the time domain statistical characteristics of the vibration signal to be detected.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010861405.XA CN112001314A (en) | 2020-08-25 | 2020-08-25 | Early fault detection method for variable speed hoist |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010861405.XA CN112001314A (en) | 2020-08-25 | 2020-08-25 | Early fault detection method for variable speed hoist |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112001314A true CN112001314A (en) | 2020-11-27 |
Family
ID=73470670
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010861405.XA Pending CN112001314A (en) | 2020-08-25 | 2020-08-25 | Early fault detection method for variable speed hoist |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112001314A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113033601A (en) * | 2021-02-02 | 2021-06-25 | 湖南科技大学 | DBN parameter selection method for identifying double-rotor misalignment fault |
| CN113345399A (en) * | 2021-04-30 | 2021-09-03 | 桂林理工大学 | Method for monitoring sound of machine equipment in strong noise environment |
| CN113418685A (en) * | 2021-06-01 | 2021-09-21 | 中广核工程有限公司 | Vibration monitoring device and fault analysis system and method based on vibration monitoring |
| CN113485261A (en) * | 2021-06-29 | 2021-10-08 | 西北师范大学 | CAEs-ACNN-based soft measurement modeling method |
| CN114459760A (en) * | 2021-12-31 | 2022-05-10 | 南京理工大学 | Rolling bearing fault diagnosis method and system under strong noise environment |
| CN114647909A (en) * | 2022-04-02 | 2022-06-21 | 江苏科技大学 | Method for determining rolling bearing degradation point based on maximum value of spectral kurtosis characteristic |
| CN114935357A (en) * | 2022-03-14 | 2022-08-23 | 浙江倍时信息科技有限公司 | Equipment health monitoring system based on entropy value calculation |
| CN116257747A (en) * | 2022-12-22 | 2023-06-13 | 中国人民解放军海军工程大学 | Early abnormality detection method for rotor system of fusion kurtosis and convolution self-encoder |
| CN116861218A (en) * | 2023-07-25 | 2023-10-10 | 上海华菱电站成套设备股份有限公司 | Mine winder key equipment state monitoring early warning system |
| CN117040983A (en) * | 2023-09-28 | 2023-11-10 | 联通(江苏)产业互联网有限公司 | Data sharing method and system based on big data analysis |
| CN117804811A (en) * | 2024-03-01 | 2024-04-02 | 青岛理工大学 | Health diagnosis system and method for deep neural network port machinery equipment of attention mechanism |
| CN118051860A (en) * | 2024-04-15 | 2024-05-17 | 西安瓦力机电科技有限公司 | Intelligent high-precision gear reducer monitoring and diagnosing system |
Citations (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| ES270165U (en) * | 1983-02-03 | 1983-07-16 | Ruiz Hurtado Eduardo | Machine for counting perfected. (Machine-translation by Google Translate, not legally binding) |
| US20100139403A1 (en) * | 2008-12-04 | 2010-06-10 | University Of Ottawa | Parameter independent detection of rotating machinery faults |
| WO2012038567A1 (en) * | 2010-09-24 | 2012-03-29 | Universidad Politécnica De Valencia | Device and method for diagnosing breakdown of induction motors |
| CN102422127A (en) * | 2009-03-31 | 2012-04-18 | 法雷奥电机设备公司 | Method and device for diagnosing malfunctions of sensors for determining the angular position of a rotor of a polyphase rotating electric machine |
| CN104076730A (en) * | 2014-06-12 | 2014-10-01 | 武汉海王机电工程技术公司 | Main push control method for propeller and engine matched controllable pitch propeller |
| WO2015086959A1 (en) * | 2013-12-09 | 2015-06-18 | Centre Technique Des Industries Mecaniques | Method and device for detecting mechanical faults in a variable speed rotating machine |
| CN105068644A (en) * | 2015-07-24 | 2015-11-18 | 山东大学 | Method for detecting P300 electroencephalogram based on convolutional neural network |
| CN107506710A (en) * | 2017-08-15 | 2017-12-22 | 河北建设集团股份有限公司 | A kind of rolling bearing combined failure extracting method |
| CN107526853A (en) * | 2016-06-22 | 2017-12-29 | 北京航空航天大学 | Rolling bearing fault mode identification method and device based on stacking convolutional network |
| CN109633270A (en) * | 2019-01-02 | 2019-04-16 | 江苏科技大学 | The identification of fault energy zone boundary and feature extracting method based on instantaneous spectrum entropy and noise energy difference |
| CN109916626A (en) * | 2019-03-14 | 2019-06-21 | 华北电力大学（保定） | The method and terminal device that rolling bearing combined failure determines |
| CN110516742A (en) * | 2019-08-28 | 2019-11-29 | 广东电网有限责任公司 | A kind of distribution terminal fault distinguishing method and system based on Combination neural network model |
| CN110686890A (en) * | 2019-10-23 | 2020-01-14 | 中船动力有限公司 | On-line diesel engine air valve state detection method |
| CN110779746A (en) * | 2019-10-24 | 2020-02-11 | 西安理工大学 | Diagnosis method for improving composite fault of deep sparse self-encoder network rotating machinery |
| CN110849627A (en) * | 2019-11-27 | 2020-02-28 | 哈尔滨理工大学 | Width migration learning network and rolling bearing fault diagnosis method based on same |
| CN110926594A (en) * | 2019-11-22 | 2020-03-27 | 北京科技大学 | Method for extracting time-varying frequency characteristics of rotary machine signal |
| US20200200648A1 (en) * | 2018-02-12 | 2020-06-25 | Dalian University Of Technology | Method for Fault Diagnosis of an Aero-engine Rolling Bearing Based on Random Forest of Power Spectrum Entropy |
| CN111353385A (en) * | 2020-01-23 | 2020-06-30 | 广东工业大学 | Pedestrian re-identification method and device based on mask alignment and attention mechanism |
| US20200225655A1 (en) * | 2016-05-09 | 2020-07-16 | Strong Force Iot Portfolio 2016, Llc | Methods, systems, kits and apparatuses for monitoring and managing industrial settings in an industrial internet of things data collection environment |
-
2020
- 2020-08-25 CN CN202010861405.XA patent/CN112001314A/en active Pending
Patent Citations (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| ES270165U (en) * | 1983-02-03 | 1983-07-16 | Ruiz Hurtado Eduardo | Machine for counting perfected. (Machine-translation by Google Translate, not legally binding) |
| US20100139403A1 (en) * | 2008-12-04 | 2010-06-10 | University Of Ottawa | Parameter independent detection of rotating machinery faults |
| CN102422127A (en) * | 2009-03-31 | 2012-04-18 | 法雷奥电机设备公司 | Method and device for diagnosing malfunctions of sensors for determining the angular position of a rotor of a polyphase rotating electric machine |
| WO2012038567A1 (en) * | 2010-09-24 | 2012-03-29 | Universidad Politécnica De Valencia | Device and method for diagnosing breakdown of induction motors |
| WO2015086959A1 (en) * | 2013-12-09 | 2015-06-18 | Centre Technique Des Industries Mecaniques | Method and device for detecting mechanical faults in a variable speed rotating machine |
| CN104076730A (en) * | 2014-06-12 | 2014-10-01 | 武汉海王机电工程技术公司 | Main push control method for propeller and engine matched controllable pitch propeller |
| CN105068644A (en) * | 2015-07-24 | 2015-11-18 | 山东大学 | Method for detecting P300 electroencephalogram based on convolutional neural network |
| US20200225655A1 (en) * | 2016-05-09 | 2020-07-16 | Strong Force Iot Portfolio 2016, Llc | Methods, systems, kits and apparatuses for monitoring and managing industrial settings in an industrial internet of things data collection environment |
| CN107526853A (en) * | 2016-06-22 | 2017-12-29 | 北京航空航天大学 | Rolling bearing fault mode identification method and device based on stacking convolutional network |
| CN107506710A (en) * | 2017-08-15 | 2017-12-22 | 河北建设集团股份有限公司 | A kind of rolling bearing combined failure extracting method |
| US20200200648A1 (en) * | 2018-02-12 | 2020-06-25 | Dalian University Of Technology | Method for Fault Diagnosis of an Aero-engine Rolling Bearing Based on Random Forest of Power Spectrum Entropy |
| CN109633270A (en) * | 2019-01-02 | 2019-04-16 | 江苏科技大学 | The identification of fault energy zone boundary and feature extracting method based on instantaneous spectrum entropy and noise energy difference |
| CN109916626A (en) * | 2019-03-14 | 2019-06-21 | 华北电力大学（保定） | The method and terminal device that rolling bearing combined failure determines |
| CN110516742A (en) * | 2019-08-28 | 2019-11-29 | 广东电网有限责任公司 | A kind of distribution terminal fault distinguishing method and system based on Combination neural network model |
| CN110686890A (en) * | 2019-10-23 | 2020-01-14 | 中船动力有限公司 | On-line diesel engine air valve state detection method |
| CN110779746A (en) * | 2019-10-24 | 2020-02-11 | 西安理工大学 | Diagnosis method for improving composite fault of deep sparse self-encoder network rotating machinery |
| CN110926594A (en) * | 2019-11-22 | 2020-03-27 | 北京科技大学 | Method for extracting time-varying frequency characteristics of rotary machine signal |
| CN110849627A (en) * | 2019-11-27 | 2020-02-28 | 哈尔滨理工大学 | Width migration learning network and rolling bearing fault diagnosis method based on same |
| CN111353385A (en) * | 2020-01-23 | 2020-06-30 | 广东工业大学 | Pedestrian re-identification method and device based on mask alignment and attention mechanism |
Non-Patent Citations (6)
| Title |
|---|
| CHAOGE WANG等: ""Identification of planetary gearbox weak compound fault based on parallel dual-parameter optimized resonance sparse decomposition and improved MOMEDA"", 《MEASUREMENT》, pages 1 - 18 * |
| MIAO Y H等: ""Application of an improved maximum correlation kurtosis deconvolution method for fault diagosis of rolling element bearings"", 《MECHANICAL SYSTEMS AND SIGNAL PROCESSING》, no. 92, pages 173 - 195 * |
| 乔志城等: ""基于Teager能量算子的改进MOMEDA方法 在铁路轮对轴承故障诊断中的应用"", 《测量与仪器》, no. 4, pages 43 - 50 * |
| 任世锦等: ""基于ELMD与改进SMSVM的机械故障诊断方法"", 《南京航空航天大学学报》, vol. 51, no. 5, pages 693 - 703 * |
| 赵小强: ""使用改进残差神经网络的滚动轴承变工况故障诊断方法"", 《西安交通大学学报》, pages 1 - 9 * |
| 陈向民等: ""基于CPP与S变换的自适应时频滤波及其在滚动轴承故障诊断中的应用"", 《航空动力学报》, vol. 33, no. 1, pages 147 - 155 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113033601B (en) * | 2021-02-02 | 2023-11-10 | 湖南科技大学 | DBN parameter selection method for double-rotor misalignment fault identification |
| CN113033601A (en) * | 2021-02-02 | 2021-06-25 | 湖南科技大学 | DBN parameter selection method for identifying double-rotor misalignment fault |
| CN113345399A (en) * | 2021-04-30 | 2021-09-03 | 桂林理工大学 | Method for monitoring sound of machine equipment in strong noise environment |
| CN113418685A (en) * | 2021-06-01 | 2021-09-21 | 中广核工程有限公司 | Vibration monitoring device and fault analysis system and method based on vibration monitoring |
| CN113485261A (en) * | 2021-06-29 | 2021-10-08 | 西北师范大学 | CAEs-ACNN-based soft measurement modeling method |
| CN113485261B (en) * | 2021-06-29 | 2022-06-28 | 西北师范大学 | CAEs-ACNN-based soft measurement modeling method |
| CN114459760A (en) * | 2021-12-31 | 2022-05-10 | 南京理工大学 | Rolling bearing fault diagnosis method and system under strong noise environment |
| CN114459760B (en) * | 2021-12-31 | 2024-05-24 | 南京理工大学 | Rolling bearing fault diagnosis method and system in strong noise environment |
| CN114935357A (en) * | 2022-03-14 | 2022-08-23 | 浙江倍时信息科技有限公司 | Equipment health monitoring system based on entropy value calculation |
| CN114647909A (en) * | 2022-04-02 | 2022-06-21 | 江苏科技大学 | Method for determining rolling bearing degradation point based on maximum value of spectral kurtosis characteristic |
| CN116257747A (en) * | 2022-12-22 | 2023-06-13 | 中国人民解放军海军工程大学 | Early abnormality detection method for rotor system of fusion kurtosis and convolution self-encoder |
| CN116257747B (en) * | 2022-12-22 | 2024-08-02 | 中国人民解放军海军工程大学 | Early abnormality detection method for rotor system of fusion kurtosis and convolution self-encoder |
| CN116861218A (en) * | 2023-07-25 | 2023-10-10 | 上海华菱电站成套设备股份有限公司 | Mine winder key equipment state monitoring early warning system |
| CN117040983A (en) * | 2023-09-28 | 2023-11-10 | 联通(江苏)产业互联网有限公司 | Data sharing method and system based on big data analysis |
| CN117040983B (en) * | 2023-09-28 | 2023-12-22 | 联通(江苏)产业互联网有限公司 | Data sharing method and system based on big data analysis |
| CN117804811A (en) * | 2024-03-01 | 2024-04-02 | 青岛理工大学 | Health diagnosis system and method for deep neural network port machinery equipment of attention mechanism |
| CN118051860A (en) * | 2024-04-15 | 2024-05-17 | 西安瓦力机电科技有限公司 | Intelligent high-precision gear reducer monitoring and diagnosing system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112001314A (en) | Early fault detection method for variable speed hoist | |
| Ma et al. | Deep residual learning with demodulated time-frequency features for fault diagnosis of planetary gearbox under nonstationary running conditions | |
| Kankar et al. | Rolling element bearing fault diagnosis using autocorrelation and continuous wavelet transform | |
| Wuxing et al. | Classification of gear faults using cumulants and the radial basis function network | |
| Hao et al. | Milling chatter detection with WPD and power entropy for Ti-6Al-4V thin-walled parts based on multi-source signals fusion | |
| Yu et al. | A novel improved full vector spectrum algorithm and its application in multi-sensor data fusion for hydraulic pumps | |
| Li et al. | Gear fault detection and diagnosis under speed-up condition based on order cepstrum and radial basis function neural network | |
| CN114838936B (en) | Rolling bearing fault diagnosis method and system for sample unbalance under variable rotating speed working condition | |
| CN113139278A (en) | Deep learning-based bearing residual life prediction method and system | |
| Shang et al. | An intelligent fault diagnosis system for newly assembled transmission | |
| Jiang et al. | A novel method for self-adaptive feature extraction using scaling crossover characteristics of signals and combining with LS-SVM for multi-fault diagnosis of gearbox | |
| Alsalaet et al. | Bearing fault diagnosis using normalized diagnostic feature-gram and convolutional neural network | |
| Al Tobi et al. | Machinery faults diagnosis using support vector machine (SVM) and Naïve Bayes classifiers | |
| CN115371988A (en) | Engineering machinery fault diagnosis method and system based on multi-feature fusion | |
| Chemseddine et al. | Gear fault feature extraction and classification of singular value decomposition based on Hilbert empirical wavelet transform | |
| CN108844735A (en) | Epicyclic gearbox fault detection method based on convolution coder and Min formula distance | |
| Peng et al. | A visual vibration characterization method for intelligent fault diagnosis of rotating machinery | |
| Liu et al. | An interpretable multiplication-convolution network for equipment intelligent edge diagnosis | |
| CN113627268A (en) | Model training method, and method and device for detecting fault of speed reducer for mine hoist | |
| Yao et al. | A low-frequency fault detection method for low-speed planetary gearbox based on acoustic signals | |
| CN114970624A (en) | Asynchronous fault diagnosis system and method for double-motor driven ratchet-pawl speed reducer | |
| CN115830368A (en) | Vehicle shock absorber defect diagnosis method, system, equipment and medium | |
| Beering et al. | Estimation of rotational speeds based on gearbox vibrations via artificial neural networks | |
| Wang et al. | Energy weighting method and its application to fault diagnosis of rolling bearing | |
| CN114638265A (en) | Milling flutter discrimination method based on signal convolution neural network |
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 |