CN115687973A - Mechanical equipment fault diagnosis method based on multi-signal fusion - Google Patents
Mechanical equipment fault diagnosis method based on multi-signal fusion Download PDFInfo
- Publication number
- CN115687973A CN115687973A CN202211306223.1A CN202211306223A CN115687973A CN 115687973 A CN115687973 A CN 115687973A CN 202211306223 A CN202211306223 A CN 202211306223A CN 115687973 A CN115687973 A CN 115687973A
- Authority
- CN
- China
- Prior art keywords
- signal
- fault diagnosis
- layer
- dimensional
- mechanical equipment
- 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
Images
Landscapes
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention discloses a mechanical equipment fault diagnosis method based on multi-signal fusion, which belongs to the field of equipment fault diagnosis method design and comprises the following steps: establishing a running signal setUAnd set of operating conditionsV(ii) a Determining sampling frequency according to the characteristics of the operation signals, wherein the operation signals sensitive to time change adopt larger sampling frequency, and the operation signals insensitive to time change adopt smaller sampling frequency; carrying out noise reduction processing on the running signal; normalizing the operation signal to make the operation signal be a decimal between 0 and 1; processing the operating signals to construct a sample set; constructing a fault diagnosis model, and setting parameters of the fault diagnosis model;training a fault diagnosis model; compared with the traditional method for judging the fault of the mechanical equipment by using a single signal, the method for diagnosing the fault of the mechanical equipment by using multiple signals has the advantages that the fault diagnosis is carried out by taking historical data into consideration and the current condition into consideration, and the result is more accurate.
Description
Technical Field
The invention relates to the technical field of equipment fault diagnosis, in particular to a mechanical equipment fault diagnosis method based on multi-signal fusion.
Background
The fault diagnosis technology is a technology for judging whether the mechanical equipment normally operates or not and finding faults in time according to the operating state of the mechanical equipment, and is a powerful guarantee for safe production and efficient operation of the mechanical equipment. In the automatic production, more accurate and more intelligent fault diagnosis is carried out on mechanical equipment, and the method has very important significance for improving the industrial production efficiency and the economic benefit. With the rapid development of the industry, the structures of various production devices are more and more complex, the functions are more and more diversified, and the failure rate of the devices is inevitably increased. Because these complex devices usually have the characteristics of multiple sources, complexity and concealment of faults, finding the cause of the fault through the conventional fault diagnosis method becomes increasingly difficult. When mechanical equipment fails, various symptoms such as changes of signals such as vibration, temperature and noise can be reflected, most of the existing fault diagnosis methods are based on a single vibration signal to judge, the influence of other signals of the equipment on a diagnosis result is not considered, and the diagnosis result is lack of accuracy.
Patent document CN109506921B discloses an online diagnosis and fault early warning method for a rotating machine, which is used for acquiring vibration data of a monitored rotating machine, analyzing and processing the vibration data to obtain characteristic parameters of each component in the operation process of the rotating machine, performing fault diagnosis on mechanical equipment by using the vibration data of the machine, and not using a multi-aspect diagnosis signal set, so that the diagnosis result is still not accurate enough.
Disclosure of Invention
The invention aims to provide a mechanical equipment fault diagnosis method based on multi-signal fusion, which enables a diagnosis result to be more scientific and reasonable and solves the defect of low accuracy of judgment based on a single vibration signal in the prior art.
In order to achieve the purpose, the invention adopts the following technical scheme: a mechanical equipment fault diagnosis method based on multi-signal fusion comprises the following steps:
s1: establishing an operation signal set U and an operation state set V;
s2: determining a sampling frequency according to the self characteristics of the operation signals, wherein the operation signals sensitive to time variation adopt a larger sampling frequency, and the operation signals insensitive to time variation adopt a smaller sampling frequency;
s3: performing noise reduction processing on the operation signal subjected to the S2 processing;
s4: normalizing the operation signal processed by the S3 to make the operation signal normalized to a decimal between 0 and 1;
s5: processing the operation signal processed by the S4 to construct a sample set;
s6: constructing a fault diagnosis model, and setting parameters of the fault diagnosis model;
s7: training a fault diagnosis model;
s8: and diagnosing the fault and outputting a fault diagnosis result.
The operation signal set U and the operation state set V in the S1 are established according to the structural characteristics of the diagnosed mechanical equipment;
operating signal set U = { U = 1 ，u 2 ，…，u m In which u 1 ，u 2 ，...，u m Representing different operation signals of the equipment, wherein m is the number of the operation signals; establishing an operating state set V = { V = { V = } 1 ，v 2 ，…，v n }，v 1 ，v 2 ，...，v n Indicating the operating state, n being the number of operating states.
The denoising processing in S3 is specifically:
s3-1: decomposing the collected operation signal by using Complete Adaptive Noise Complete set Empirical Mode Decomposition (CEEMDAN) to obtain n intrinsic Mode functions and a residual signal, wherein the n intrinsic Mode functions are respectively expressed as IMF 1 ，IMF 2 ，...，IMF n The residual signal is denoted r;
s3-2: and respectively calculating correlation coefficients of the n intrinsic mode functions and the original signal by adopting a Pearson correlation coefficient. The Pearson correlation coefficient calculation method is shown in formula (1)
In the formula: x represents the original signal, IMF i Representing the IMF component of the ith stage. According to the formula (1), when the original signal and the IMF component exhibit negative correlation, the value of the correlation coefficient is negative, and when the linear correlation is higher, the value of the correlation coefficient is larger;
s3-3: reconstructing the signal to obtain a denoised signal;
correlation coefficient between intrinsic mode function and original signal
When the intrinsic mode function is smaller than a set threshold value, the intrinsic mode function is determined to be generated by noise and is removed;
correlation coefficient between intrinsic mode function and original signal
And when the intrinsic mode function is larger than or equal to the set threshold, the intrinsic mode function is determined to be generated by the signal, the intrinsic mode function is reserved, and the reserved intrinsic mode function and the residual signal are reconstructed to form the denoised signal.
The specific construction mode of the sample set in the S5 is as follows:
s5-1: the sampling frequency of the operation signals sensitive to the time variation is larger, the number of the sampled data is more, a plurality of continuous data are used as a sample from the starting point of acquisition, the sampling frequency of the operation signals insensitive to the time variation is smaller, the number of the sampled data is less, and each data is used as a sample;
s5-2: processing the running signal sample sensitive to the time change to obtain a two-dimensional time-frequency graph, decomposing the running signal sensitive to the time change by adopting the CEEMDAN, and performing Hilbert transform on each IMF component decomposed by the CEEMDAN by utilizing Hilbert Transform (HT) to generate the two-dimensional time-frequency graph.
The fault diagnosis model in S6 includes: a 2D-CNN channel, a 1D-CNN channel and a fusion channel;
the 2D-CNN channel comprises an input layer input2D, a two-dimensional convolution layer conv2D, a two-dimensional maximum pooling layer max _ pooling2D, a two-dimensional convolution layer conv2D _1, a two-dimensional maximum pooling layer max _ pooling2D _1, an attention mechanism module layer se _ attribution _ layer, a two-dimensional convolution layer conv2D _2, a two-dimensional maximum pooling layer max _ pooling2D _2 and a tiling layer flatten;
the 1D-CNN comprises an input layer input1D, a one-dimensional convolution layer conv1D, a one-dimensional maximum pooling layer max _ pooling1D, a one-dimensional convolution layer conv1D _1, a one-dimensional maximum pooling layer max _ pooling1D _1, an attention mechanism module layer attention _ layer, a one-dimensional convolution layer conv1D _2, a one-dimensional maximum pooling layer max _ pooling1D _2 and a tiled layer flatten _1;
the fusion channel comprises a splicing layer concatenate, a discarding layer dropout, a dense layer feature and an output layer output;
processing the time-variation sensitive running signal to obtain a two-dimensional time-frequency graph, and splicing the two-dimensional time-frequency graph to be used as the input of a 2D-CNN channel input layer input 2D; splicing the operation signals and then using the spliced operation signals as the input1D of the input layer of the 1D-CNN channel; the 2D-CNN channel tiling layer flatten and the 1D-CNN channel tiling layer flatten _1 are used as the input of the splice layer concatenate.
The diagnostic model building of S7 comprises the following steps:
s7-1: dividing the sample set into a training set and a testing set according to a certain proportion, wherein the proportion of the number of samples in the training set to the number of samples in the testing set is 7:3;
s7-2: and setting training parameters, and updating network parameters by adopting an Adam optimizer, wherein the Epoch, the batch size and the learning rate of an Adam algorithm during model training are respectively 60, 64 and 0.001.
S7-3: and (3) model training, namely initializing the parameters of the fault diagnosis model in a random mode, inputting training set samples into the fault diagnosis model in batches for model training, inputting test samples into the trained fault diagnosis model for testing, and outputting fault diagnosis accuracy so as to judge the feasibility of the model. And after the training is finished, the trained fault diagnosis model is stored.
And in the S8, the fault diagnosis specifically inputs the operation signal set U to be diagnosed into the trained fault diagnosis model, diagnoses the fault diagnosis model and outputs a diagnosis result.
According to the technical scheme, the invention at least has the following beneficial effects:
1. in the process of diagnosing the fault of the mechanical equipment, the invention considers the influence of three signals of temperature, vibration and noise on the fault diagnosis accuracy rate, and avoids the defect of low fault diagnosis accuracy rate caused by the fact that the current fault diagnosis of the mechanical equipment is based on a single vibration signal.
2. The invention collects three signals of vibration, temperature and noise, considers the sensitivity of three variables to time change, adopts different sampling frequencies, not only ensures the integrity of collected data, but also reduces the calculated amount.
3. The double-channel convolutional neural network fault diagnosis model disclosed by the invention not only learns the correlation between adjacent intervals, but also learns the correlation between non-adjacent intervals, so that the fault diagnosis accuracy is improved, and the defect that the fault diagnosis accuracy is low because the current mechanical equipment fault diagnosis model adopts a single channel is overcome.
4. When fault diagnosis is carried out, historical data and the current condition are considered, and compared with the traditional fault diagnosis method based on the current condition, the method has the advantage that the result is more accurate.
5. According to the double-channel convolution neural network fault diagnosis model disclosed by the invention, a CEEMDAN decomposition and HT combination method is adopted by a time-frequency transformation module of a two-dimensional convolution channel, so that the problems that the mode aliasing phenomenon is caused by decomposing an original signal by adopting EMI) in the traditional HHT transformation, or the proper wavelet base is difficult to determine by the wavelet transformation and the like are solved.
Drawings
FIG. 1 is a flowchart of the fault diagnosis steps of the present example;
FIG. 2 is a diagram of a two-channel fault diagnosis model of the present example;
FIG. 3 is a two-dimensional time-frequency plot of a sample vibration signal during normal operation;
FIG. 4 is a two-dimensional time-frequency plot of a sample vibration signal during operation with an end face worn condition;
FIG. 5 is a two-dimensional time-frequency diagram of a vibration signal sample during operation with a worn slipper;
FIG. 6 is a two-dimensional time-frequency plot of a sample vibration signal during operation with a swashplate in a worn condition;
FIG. 7 is a two-dimensional time-frequency plot of a sample vibration signal during valve plate wear operation;
FIG. 8 is a two-dimensional time-frequency plot of a sample vibration signal during operation with a center spring in a failed condition;
FIG. 9 is a graph of loss function values versus accuracy change for a training set sample in accordance with the present invention;
fig. 10 is a comparison graph of the diagnostic result accuracy of different fault diagnosis methods.
Detailed Description
In order to make the technical means, the creation characteristics, the achievement purposes and the effects of the invention easy to understand, the invention is further explained by combining the specific embodiments.
The invention provides a mechanical equipment fault diagnosis method based on multi-signal fusion, a flow chart of which is shown in figure 1, and the following takes fault diagnosis of a certain hydraulic pump as an example to explain specific implementation steps:
s1, establishing an operation signal set U and an operation state set V.
According to the structural characteristics of the diagnosed mechanical equipment, an operation signal set is established, wherein U = { U = { (U) } 1 ，u 2 ，...，u m In which u 1 ，u 2 ，...，u m Representing different operation signals of the equipment, wherein m is the number of the operation signals; operating state set V = { V = 1 ，v 2 ，...，v n }，v 1 ，v 2 ，...，v n Representing the operating state and n is the number of operating states.
For the hydraulic pump, the operation signal includes a temperature signal, a vibration signal, and a noise signal. I.e. the operating signal set U = { U = 1 ，u 2 ，...，u m = { temperature signal, vibration signal, noise signal }.
The hydraulic pump is used for conveyingAbnormal states which may occur during the running process include end face abrasion, slipper abrasion, swash plate abrasion, port plate abrasion, central spring failure and normal states. So the set of operating states V = { V = { V = 1 ，v 2 ，...，v n = { normal state, end face wear, slipper wear, swash plate wear, port plate wear, center spring failure }.
And S2, determining a sampling frequency according to the characteristics of the operation signals, wherein the operation signals sensitive to time variation adopt a larger sampling frequency, and the operation signals insensitive to time variation adopt a smaller sampling frequency.
For the fault diagnosis of the hydraulic pump in the embodiment, the vibration signal and the noise signal are sensitive to time variation, the sampling frequency of the vibration signal and the sampling frequency of the noise signal are both 0.1K Hz, the temperature signal is insensitive to the time variation, and the sampling frequency of the temperature signal is 0.1Hz.
S3, noise reduction processing of running signals
S3-1, decomposing the collected operation signal by using Complete Adaptive Noise Complete set Empirical Mode Decomposition (CEEMDAN) to obtain n Intrinsic Mode Functions (IMF) and a residual signal, wherein the n Intrinsic Mode functions are respectively expressed as IMF 1 ，IMF 2 ，...，IMF n The residual signal is denoted r;
s3-2, correlation coefficients of n Intrinsic Mode Functions (IMFs) and the original signal are respectively calculated by adopting the Pearson correlation coefficient. The Pearson correlation coefficient calculation method is shown in formula (1);
in the formula: x represents the original signal, IMF i Representing the IMF component of the ith stage. According to the formula (1), when the original signal and the IMF component exhibit negative correlation, the value of the correlation coefficient is negative, and when the linear correlation is higher, the value of the correlation coefficient is larger;
and S3-3, signal reconstruction is carried out to obtain a denoised signal. When eigenmodeCorrelation coefficient between state function and original signal
When the intrinsic mode function is smaller than a set threshold value, the intrinsic mode function is determined to be generated by noise and is removed; correlation coefficient between intrinsic mode function and original signal
And when the intrinsic mode function is larger than or equal to the set threshold, the intrinsic mode function is determined to be generated by the signal, the intrinsic mode function is reserved, and the reserved intrinsic mode function and the residual signal are reconstructed to form the denoised signal.
For the fault diagnosis of the hydraulic pump in the embodiment, after the vibration signal and the noise signal are subjected to CEEMDAN decomposition, respective intrinsic mode functions of the vibration signal and the noise signal are obtained, correlation coefficients of the intrinsic mode functions are respectively calculated, one tenth of the maximum correlation coefficient in the correlation coefficient sequence is taken as a threshold value, the intrinsic mode function generated by the noise is removed, and the vibration signal and the noise signal after being denoised are obtained after reconstruction.
S4, normalizing the operation signals to change the operation signals into decimal numbers between (0, 1);
s5, processing the running signals to construct a sample set;
s5-1, the sampling frequency of the operation signal sensitive to the time variation is higher, more data are sampled, and a plurality of continuous data are used as a sample from the acquisition starting point; the sampling frequency of the running signal which is insensitive to the time change is small, the sampled data is less, and each data is used as a sample;
for the fault diagnosis of the hydraulic pump in this embodiment, 1024 data are taken as one sample from the starting point of the vibration signal and the noise signal respectively for each state, then 1024 data are taken as the next sample, and so on, 600 samples are taken for each state, the hydraulic pump has 6 states, 3600 samples are taken for the vibration signal, and 3600 samples are taken for the noise signal. For each state, the temperature signal starts from a starting point, 1 data is taken as one sample, 600 samples are taken in each state, the hydraulic pump has 6 states, and 3600 samples are taken from the temperature signal;
and S5-2, processing the time-variation-sensitive running signal sample to obtain a two-dimensional time-frequency graph. Decomposing the operation signals sensitive to time change by adopting CEEMDAN, and performing Hilbert transform on each IMF component decomposed by the CEEMDAN by utilizing Hilbert Transform (HT) to generate a two-dimensional time-frequency graph;
in this example, the CEEMDAN decomposition is performed on the vibration signal sample and the noise signal sample, and then Hilbert Transform (HT) is used to perform Hilbert transform on each IMF component after the CEEMDAN decomposition, so as to generate a two-dimensional time-frequency graph.
S6, constructing a fault diagnosis model, and setting parameters of the fault diagnosis model;
as shown in FIG. 2, the fault diagnosis model comprises a 2D-CNN channel, a 1D-CNN channel and a fusion channel.
The 2D-CNN channel comprises an input layer input2D, a two-dimensional convolution layer conv2D, a two-dimensional maximum pooling layer max _ pooling2D, a two-dimensional convolution layer conv2D _1, a two-dimensional maximum pooling layer max _ pooling2D _1, an attention mechanism module layer se _ attribution _ layer, a two-dimensional convolution layer conv2D _2, a two-dimensional maximum pooling layer max _ pooling2D _2 and a tiling layer flatten.
The 1D-CNN comprises an input layer input1D, a one-dimensional convolution layer convld, a one-dimensional maximum pooling layer max _ posing 1D, a one-dimensional convolution layer conv1D _1, a one-dimensional maximum pooling layer max _ posingld _1, an attention mechanism module layer attention _ layer, a one-dimensional convolution layer conv1D _2, a one-dimensional maximum pooling layer max _ posing 1D _2 and a tiling layer flatten _1.
The fusion channel comprises a splice layer concatenate, a drop layer dropout, a dense layer feature and an output layer output.
Processing the time-variation sensitive running signal to obtain a two-dimensional time-frequency graph, and splicing the two-dimensional time-frequency graph to be used as the input of a 2D-CNN channel input layer input 2D; splicing the operation signals and then using the spliced operation signals as the input1D of the input layer of the 1D-CNN channel; the 2D-CNN channel tiling layer flatten and the 1D-CNN channel tiling layer flatten _1 are used as the input of the splice layer conticatenate.
In this embodiment, the size of the input layer input2d is 32 × 32 × 6; the convolution kernel size of the two-dimensional convolution layer conv2d is 3 multiplied by 3, the number of the convolution kernels is 4, the activation function is ReLU, the step length is 1, and the filling mode is same; the size of the max _ pooling2d pooling kernel of the two-dimensional maximum pooling layer is 2 multiplied by 2, and the step length is 2; the convolution kernel size of the two-dimensional convolution layer conv2d _1 is 5 multiplied by 5, the number of the convolution kernels is 8, the activation function is ReLU, the step length is 1, and the filling mode is same; the size of the two-dimensional maximum pooling layer max _ pooling2d _1 pooling kernel is 2 × 2, and the step length is 2; the attention mechanism module layer se _ attention _ layer adopts SEnet; the convolution kernel size of the two-dimensional convolution layer conv2d _2 is 3 multiplied by 3, the number of the convolution kernels is 16, the activation function is ReLU, the step length is 1, and the filling mode is same; the two-dimensional maximum pooling layer max _ pooling2d _2 pooling kernel size is 2 × 2, step size is 2.
In this embodiment, the input1d has a size of 2049 × 1; the convolution kernel size of the one-dimensional convolution layer conv1d is 7 multiplied by 1, the number of convolution kernels is 2, the activation function is ReLU, the step length is 1, and the filling mode is same; the size of the one-dimensional maximum pooling layer max _ pooling1d pooling core is 3 multiplied by 1, and the step length is 3; the convolution kernel size of the one-dimensional convolution layer conv1d _1 is 5 multiplied by 1, the number of convolution kernels is 4, the activation function is ReLU, the step length is 1, and the filling mode is same; the size of the one-dimensional maximum pooling layer max _ pooling1d _1 pooling kernel is 2 multiplied by 1, and the step length is 2; attention is paid to the fact that a module layer of an attention mechanism adopts Bahdanauthention, the size of a convolution kernel of a one-dimensional convolution layer conv1d _2 is 3 multiplied by 1, the number of the convolution kernels is 8, an activation function is ReLU, a step length is 1, and a filling mode is same; the one-dimensional maximum pooling layer max _ pooling1d _2 pooling kernel size is 4 × 1, step size is 4.
In this embodiment, the number of neurons in the splice layer is 936, the discard rate of the discard layer dropout is 0.5, the number of dense layer dense neurons is 128, the activation function is ReLU, the number of output neurons in the output layer is 6, and the activation function is Softmax.
In this example, a two-dimensional time-frequency diagram of a vibration signal and a noise signal is spliced to form a feature diagram which is used as the input of the 2D-CNN channel input layer input2D, and a data vector of a temperature signal, a vibration signal sample and a noise signal which are spliced is used as the input of the 1D-CNN channel input layer input 1D.
S7, training a fault diagnosis model;
s7-1, dividing a data sample into a training set and a testing set according to a certain proportion;
in this example, each state has 600 samples, there are 6 states in total, for the two-dimensional time-frequency graphs of the vibration signal and the noise signal, 3600 feature graphs are formed after splicing, and the 3600 data vectors after splicing the temperature signal, the vibration signal sample and the noise signal have a value of 7:3, dividing by proportion, wherein the number of samples in the training set is 2520, and the number of samples in the test set is 180.
S7-2, setting training parameters;
in the example, an Adam optimizer is adopted to update network parameters, and the Epoch, the batch size and the learning rate of the Adam algorithm during model training are 60, 64 and 0.001 respectively.
S7-3 model training;
initializing the parameters of the fault diagnosis model in a random mode, inputting training set samples into the fault diagnosis model in batches for model training, inputting test samples into the trained fault diagnosis model for testing, and outputting the fault diagnosis accuracy so as to judge the feasibility of the model. And after the training is finished, storing the trained fault diagnosis model.
S8, fault diagnosis, and outputting a fault diagnosis result;
and inputting the temperature signal, the vibration signal and the noise signal which need to be diagnosed into the trained fault diagnosis model, diagnosing and outputting a diagnosis result.
The effects of the present invention can be further illustrated by the following comparative experiments:
1) The hydraulic pump respectively runs for a period of time under the normal state, the end surface abrasion, the sliding shoe abrasion, the swash plate abrasion, the valve plate abrasion and the central spring failure state, the temperature sensor samples the temperature signal according to the sampling frequency of 0.1Hz, the acceleration sensor samples the vibration signal according to the sampling frequency of 0.1KHz, and the noise sensor samples the noise signal according to the sampling frequency of 0.1 KHz.
2) The method comprises the steps of carrying out noise reduction processing on collected vibration signals and noise signals, firstly decomposing the collected vibration signals and noise signals through CEEEMDAN, then calculating the Pearson correlation coefficient of each IMF component after decomposition and the corresponding original signal by using a formula (1), taking one tenth of the maximum correlation coefficient in a correlation coefficient sequence as a threshold value, removing the IMF component of which the Pearson correlation coefficient is smaller than the threshold value, and reconstructing the reserved component.
3) And carrying out normalization processing on the reconstructed decomposition to form a normalized temperature signal, a normalized vibration signal and a normalized noise signal.
4) 3600 temperature signal samples, 3600 vibration signal samples and 3600 noise signal samples are collected.
For the normalized temperature signals continuously acquired by the hydraulic pump in normal operation, from a starting point, every 1 data is taken as a sample, and 600 temperature signal samples in normal operation are taken. For the normalized temperature signals continuously collected by the hydraulic pump in the end surface abrasion state, from the starting point, every 1 data is taken as a sample, and 600 temperature signal samples in the end surface abrasion state are taken in total. For the normalized temperature signals continuously acquired by the hydraulic pump in the sliding shoe abrasion state, from the starting point, every 1 data is taken as a sample, and 600 temperature signal samples in the sliding shoe abrasion state are taken in total. For the normalized temperature signals continuously collected by the hydraulic pump in the operation of the swash plate in the wear state, from a starting point, every 1 data is taken as a sample, and 600 temperature signal samples in the operation of the swash plate in the wear state are taken in total. For the normalized temperature signals continuously collected by the hydraulic pump in the running state of the valve plate in the wearing state, from the starting point, every 1 data is taken as a sample, and 600 temperature signal samples in the running state of the central spring in the failure state are taken in total. For the normalized temperature signals continuously collected by the hydraulic pump in the running state of the valve plate in the wearing state, from the starting point, every 1 data is taken as a sample, and 600 temperature signal samples in the running state of the central spring in the failure state are taken in total.
For the vibration signals after normalization continuously collected by the hydraulic pump in normal state operation, 1024 data are taken as a sample from the starting point, then 1024 data are taken as the next sample, and so on, and 600 vibration signal samples in normal state operation are taken in total. For the normalized vibration signals continuously acquired by the hydraulic pump in the end surface wear state, 1024 data are taken as one sample from the starting point, then 1024 data are taken as the next sample, and so on, and 600 vibration signal samples in the end surface wear state are taken. And for the normalized vibration signals continuously acquired by the hydraulic pump in the sliding shoe abrasion state, 1024 data are taken as one sample from the starting point, then 1024 data are taken as the next sample, and the like, and 600 vibration signal samples in the sliding shoe abrasion state are taken in total. For the normalized vibration signal continuously collected by the hydraulic pump in the operation of the swash plate in the wear state, 1024 data are taken as a sample from the starting point, the 1024 data are taken as the next sample, and the like, and 600 vibration signal samples in the operation of the swash plate in the wear state are taken in total. 1024 pieces of data are taken as a sample from a starting point, then 1024 pieces of data are taken as a next sample, and so on, for the vibration signal after normalization continuously acquired by the hydraulic pump in the running state of the valve plate in the wearing state, and 600 vibration signal samples in the running state of the central spring in the failure state are taken in total. 1024 pieces of data are taken as a sample from a starting point, then 1024 pieces of data are taken as a next sample, and so on, for the vibration signal after normalization continuously acquired by the hydraulic pump in the running state of the valve plate in the wearing state, and 600 vibration signal samples in the running state of the central spring in the failure state are taken in total.
For the normalized noise signals continuously collected by the hydraulic pump in the normal state, 1024 data are taken as one sample from the starting point, 1024 data are taken as the next sample, and the like, and 600 noise signal samples in the normal state are taken in total. And for the normalized noise signals continuously acquired by the hydraulic pump in the end surface abrasion state, 1024 data are taken as one sample from the starting point, then 1024 data are taken as the next sample, and so on, and 600 noise signal samples in the end surface abrasion state are taken in total. For a normalized noise signal continuously acquired by a hydraulic pump in the sliding shoe abrasion state, 1024 data are taken as one sample from a starting point, then 1024 data are taken as the next sample, and the like, and 600 noise signal samples in the sliding shoe abrasion state are taken totally. For the normalized noise signal continuously collected by the hydraulic pump in the operation of the swash plate in the wear state, 1024 data are taken as one sample from the starting point, then 1024 data are taken as the next sample, and so on, and 600 noise signal samples in the operation of the swash plate in the wear state are taken in total. For the normalized noise signals continuously collected by the hydraulic pump in the running state of the valve plate in the wear state, 1024 data are taken as one sample from the starting point, then 1024 data are taken as the next sample, and so on, and 600 noise signal samples in the running state of the central spring in the failure state are taken in total. For a normalized noise signal continuously acquired by a hydraulic pump in the running state of a valve plate in a wear state, 1024 data are taken as one sample from a starting point, then 1024 data are taken as the next sample, and the like, and 600 noise signal samples in the running state of a central spring in a failure state are taken.
5) Respectively carrying out CEEMDAN decomposition on 3600 temperature signal samples and 3600 vibration signal samples, and carrying out Hilbert transform on each IMF component subjected to the CEEMDAN decomposition by using Hilbert Transform (HT) to generate a two-dimensional time-frequency diagram. Two-dimensional time-frequency diagrams of some of the vibration signal samples are shown in fig. 3-8.
6) And the 3600 temperature signal sample two-dimensional time-frequency graphs are respectively spliced with the corresponding 3600 vibration signal sample two-dimensional time-frequency graphs, a bicubic interpolation algorithm is adopted to compress the images to form 3600 32 multiplied by 6 input feature graphs, and the 3600 temperature signal samples, the 3600 vibration signal samples and the 3600 noise signal samples are spliced to form 3600 2049 multiplied by 1 data vectors.
7) 3600 32 x 6 input feature maps and 3600 2049 x 1 data vectors were randomly divided into a training set and a test set in a 7:3 ratio.
8) Inputting a training set sample into a fault diagnosis model, as shown in fig. 9, training the model, after the training is finished, testing the training effect of the model by using a test set, wherein the training effect is shown in table 1, wherein a state 1, a state 2, a state 3, a state 4, a state 5 and a state 6 respectively represent a normal state, an end face abrasion, a slipper abrasion, a swash plate abrasion, a valve plate abrasion and a central spring failure state, and the correctness of the diagnosis result is 99.44%.
Table 1 test set sample test results table
As shown in fig. 10, M1, M2, M3, M4, M5, M6, M7, M8, and M9 respectively represent a fault diagnosis method, a multi-information fusion 2D-CNN single-channel fault diagnosis method, a multi-information fusion 1D-CNN single-channel fault diagnosis method, a vibration signal dual-channel fault diagnosis method, a noise signal dual-channel fault diagnosis method, a vibration signal 2D-CNN single-channel fault diagnosis method, a vibration signal 1D-CNN single-channel fault diagnosis method, a noise signal 2D-CNN single-channel fault diagnosis method, and a noise signal 1D-CNN single-channel fault diagnosis method disclosed in the present invention. In order to embody the advance of the double-channel fault diagnosis method disclosed by the invention relative to a single-channel fault diagnosis model, the same data set respectively carries out experimental tests on the 2D-CNN channel model and the 1D-CNN channel model, and the diagnosis accuracy rates are respectively 94.44% and 95.56%. In order to embody the advancement of the fault diagnosis method disclosed by the invention relative to a fault diagnosis model based on a single vibration signal or a single noise signal, the single vibration signal and the single noise signal are respectively used as sample sets to carry out experimental tests on the fault diagnosis model disclosed by the invention, and the diagnosis accuracy is 93.33% and 92.22% respectively. In order to embody the overall advancement of the fault diagnosis model disclosed by the invention, the single vibration signal and the single noise signal respectively carry out experimental tests on the 2D-CNN channel model and the 1D-CNN channel model, and the diagnosis accuracy rates are respectively 88.33%, 87.22%, 84.44% and 83.89%.
Claims (7)
1. A mechanical equipment fault diagnosis method based on multi-signal fusion is characterized by comprising the following steps:
s1: establishing an operation signal set U and an operation state set V;
s2: determining a sampling frequency based on whether the operating signal is sensitive to time variations;
s3: performing noise reduction processing on the operation signal processed by the S2;
s4: normalizing the operation signal processed by the S3 to make the operation signal normalized to a decimal between 0 and 1;
s5: processing the operation signal processed by the S4 to construct a sample set;
s6: constructing a fault diagnosis model, and setting parameters of the fault diagnosis model;
s7: training a fault diagnosis model;
s8: and diagnosing the fault and outputting a fault diagnosis result.
2. The mechanical equipment fault diagnosis method based on multi-signal fusion is characterized in that the operation signal set U and the operation state set V in the s1 are established according to the structural characteristics of the diagnosed mechanical equipment;
operating signal set U = { U = 1 ，u 2 ，...，u m In which u 1 ，u 2 ，...，u m Representing different operation signals of the equipment, wherein m is the number of the operation signals; establishing an operating state set V = { V = { V = } 1 ，v 2 ，...，v n }，v 1 ，v 2 ，...，v n Representing the operating state and n is the number of operating states.
3. The method for diagnosing the fault of the mechanical equipment based on the multi-signal fusion as claimed in claim 1, wherein the noise reduction processing in S3 is specifically:
s3-1: decomposing the collected operation signal by using complete self-adaptive noise complete set empirical mode decomposition to obtain n intrinsic mode functions and a residual signal, wherein the n intrinsic mode functions are respectively expressed as IMF 1 ，IMF 2 ，...，IMF n The residual signal is denoted r;
s3-2: and respectively calculating correlation coefficients of the n intrinsic mode functions and the original signal by adopting a Pearson correlation coefficient. The Pearson correlation coefficient calculation method is shown in formula (1)
In the formula: x represents the original signal, IMF i Representing the IMF component of the ith stage. According to the formula (1), when the original signal and the IMF component exhibit negative correlation, the value of the correlation coefficient is negative, and when the linear correlation is higher, the value of the correlation coefficient is larger;
s3-3: reconstructing the signal to obtain a denoised signal;
correlation coefficient between intrinsic mode function and original signal
When the intrinsic mode function is smaller than a set threshold value, the intrinsic mode function is determined to be generated by noise and is removed;
correlation coefficient between intrinsic mode function and original signal
And when the intrinsic mode function is larger than or equal to the set threshold, the intrinsic mode function is determined to be generated by the signal, the signal is reserved, and the reserved intrinsic mode function and the residual signal are reconstructed to form the denoised signal.
4. The mechanical equipment fault diagnosis method based on multi-signal fusion as claimed in claim 1, wherein the specific configuration manner of the sample set in S5 is as follows:
s5-1: determining sample data according to whether the operating signal is sensitive to time variation;
s5-2: processing the running signal sample sensitive to the time change to obtain a two-dimensional time-frequency graph, decomposing the running signal sensitive to the time change by adopting CEEMDAN, and performing Hilbert transform on each IMF component decomposed by the CEEMDAN by using Hilbert Transform (HT) to generate the two-dimensional time-frequency graph.
5. The method for diagnosing the fault of the mechanical equipment based on the multi-signal fusion as claimed in claim 1, wherein the fault diagnosis model in the S6 comprises: a 2D-CNN channel, a 1D-CNN channel and a fusion channel;
the 2D-CNN channel comprises an input layer input2D, a two-dimensional convolution layer conv2D, a two-dimensional maximum pooling layer max _ pooling2D, a two-dimensional convolution layer conv2D _1, a two-dimensional maximum pooling layer max _ pooling2D _1, an attention module layer se _ attention _ layer, a two-dimensional convolution layer conv2D _2, a two-dimensional maximum pooling layer max _ pooling2D _2 and a tiling layer flatten;
the 1D-CNN comprises an input layer input1D, a one-dimensional convolution layer conv1D, a one-dimensional maximum pooling layer max _ pooling1D, a one-dimensional convolution layer conv1D _1, a one-dimensional maximum pooling layer max _ pooling1D _1, an attention mechanism module layer attention _ layer, a one-dimensional convolution layer conv1D _2, a one-dimensional maximum pooling layer max _ pooling1D _2 and a tiled layer flatten _1;
the fusion channel comprises a splicing layer concatenate, a discarding layer dropout, a dense layer feature and an output layer output;
processing the time-variation sensitive running signal to obtain a two-dimensional time-frequency graph, and splicing the two-dimensional time-frequency graph to be used as the input of a 2D-CNN channel input layer input 2D; splicing the operation signals to be used as the input of the input layer inputld of the 1D-CNN channel; the 2D-CNN channel tiling layer flatten and the 1D-CNN channel tiling layer flatten _1 are used as the input of the splice layer concatenate.
6. The mechanical equipment fault diagnosis method based on multi-signal fusion according to any one of claims 1 to 5, characterized in that the diagnostic model building of S7 comprises the following steps:
s7-1: dividing the sample set into a training set and a testing set according to a certain proportion, wherein the proportion of the number of samples in the training set to the number of samples in the testing set is 7:3;
s7-2: and setting training parameters, and updating network parameters by adopting an Adam optimizer, wherein the Epoch, the batch size and the learning rate of an Adam algorithm during model training are respectively 60, 64 and 0.001.
S7-3: and (3) model training, namely initializing the parameters of the fault diagnosis model in a random mode, inputting a training set sample into the fault diagnosis model in batches for model training, inputting a test sample into the trained fault diagnosis model for testing, outputting the fault diagnosis accuracy rate to judge the feasibility of the model, and storing the trained fault diagnosis model after training.
7. The mechanical equipment fault diagnosis method based on multi-signal fusion as claimed in claim 1, wherein in the S8 fault diagnosis, an operation signal set U to be diagnosed is specifically input into a trained fault diagnosis model for diagnosis, and a diagnosis result is output.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211306223.1A CN115687973A (en) | 2022-10-24 | 2022-10-24 | Mechanical equipment fault diagnosis method based on multi-signal fusion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211306223.1A CN115687973A (en) | 2022-10-24 | 2022-10-24 | Mechanical equipment fault diagnosis method based on multi-signal fusion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115687973A true CN115687973A (en) | 2023-02-03 |
Family
ID=85099171
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211306223.1A Pending CN115687973A (en) | 2022-10-24 | 2022-10-24 | Mechanical equipment fault diagnosis method based on multi-signal fusion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115687973A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116049725A (en) * | 2023-03-29 | 2023-05-02 | 淄博热力有限公司 | Rotary machine fault diagnosis method based on improved deep learning classification model |
| CN116484268A (en) * | 2023-06-21 | 2023-07-25 | 西安黑石智能科技有限公司 | Intelligent industrial equipment fault diagnosis system based on machine learning |
| CN116846688A (en) * | 2023-08-30 | 2023-10-03 | 南京理工大学 | Interpretable flow intrusion detection method based on CNN |
-
2022
- 2022-10-24 CN CN202211306223.1A patent/CN115687973A/en active Pending
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116049725A (en) * | 2023-03-29 | 2023-05-02 | 淄博热力有限公司 | Rotary machine fault diagnosis method based on improved deep learning classification model |
| CN116049725B (en) * | 2023-03-29 | 2023-12-29 | 淄博热力有限公司 | Rotary machine fault diagnosis method based on improved deep learning classification model |
| CN116484268A (en) * | 2023-06-21 | 2023-07-25 | 西安黑石智能科技有限公司 | Intelligent industrial equipment fault diagnosis system based on machine learning |
| CN116484268B (en) * | 2023-06-21 | 2023-09-05 | 西安黑石智能科技有限公司 | Intelligent industrial equipment fault diagnosis system based on machine learning |
| CN116846688A (en) * | 2023-08-30 | 2023-10-03 | 南京理工大学 | Interpretable flow intrusion detection method based on CNN |
| CN116846688B (en) * | 2023-08-30 | 2023-11-21 | 南京理工大学 | Interpretable flow intrusion detection method based on CNN |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | Deep residual learning-based fault diagnosis method for rotating machinery | |
| CN111340238B (en) | Fault diagnosis method, device, equipment and storage medium of industrial system | |
| CN115687973A (en) | Mechanical equipment fault diagnosis method based on multi-signal fusion | |
| CN109858352B (en) | Fault diagnosis method based on compressed sensing and improved multi-scale network | |
| CN112254964A (en) | Rolling bearing fault diagnosis method based on rapid multi-scale convolution neural network | |
| CN110705525A (en) | Method and device for diagnosing rolling bearing fault | |
| CN113177537B (en) | Fault diagnosis method and system for rotary mechanical equipment | |
| CN116226646B (en) | Method, system, equipment and medium for predicting health state and residual life of bearing | |
| CN113642754A (en) | Complex industrial process fault prediction method based on RF noise reduction self-coding information reconstruction and time convolution network | |
| CN113110398B (en) | Industrial process fault diagnosis method based on dynamic time consolidation and graph convolution network | |
| CN113822139A (en) | Equipment fault diagnosis method based on improved 1DCNN-BilSTM | |
| Li et al. | Gear pitting fault diagnosis using raw acoustic emission signal based on deep learning | |
| CN113627375A (en) | Planetary gear fault diagnosis method and system, storage medium and computing device | |
| CN113642414A (en) | Method for predicting residual service life of rolling bearing based on Transformer model | |
| CN115859077A (en) | Multi-feature fusion motor small sample fault diagnosis method under variable working conditions | |
| CN115290326A (en) | Rolling bearing fault intelligent diagnosis method | |
| Neupane et al. | Deep learning-based bearing fault detection using 2-D illustration of time sequence | |
| van den Hoogen et al. | An improved wide-kernel cnn for classifying multivariate signals in fault diagnosis | |
| Lee et al. | An explainable deep learning-based prognostic model for rotating machinery | |
| CN114964782A (en) | Rolling bearing fault diagnosis acoustic emission signal processing method and system based on compressed sensing | |
| CN113159088A (en) | Fault monitoring and diagnosis method based on multi-feature fusion and width learning | |
| CN116595465A (en) | High-dimensional sparse data outlier detection method and system based on self-encoder and data enhancement | |
| CN115130500A (en) | Motor fault diagnosis system and method based on SDAE algorithm | |
| CN112485001A (en) | Fault diagnosis method for intelligent elevator | |
| Chakraborty et al. | Bearing Fault Detection based on Internet of Things using Convolutional 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 |