Fault diagnosis system and method for rolling mill gearbox
Technical Field
The invention relates to the technical field of fault diagnosis of a main transmission system of a strip rolling mill in the metallurgical industry, in particular to a fault diagnosis system and method of a rolling mill gear box.
Background
With the higher and higher requirements of the market on the quality of the plate strip, the precision degree and the automation level of the rolling mill equipment are more and more concerned. As a core component of a rolling mill, a main transmission system gear box of the rolling mill runs for a long time under complex and severe conditions of heavy load, fatigue, high temperature and the like, and various damages or faults are inevitably generated, such as the influence of damage forms of fatigue cracks, pitting corrosion and the like on gears and bearings. The defect formation process of critical parts of the gearbox often causes the aggravation of non-linear phenomena and the generation of non-stationary phenomena, so that the prior art may not provide enough time between warning and serious fault to implement safety procedures, and on the other hand, the low precision of fault diagnosis may cause false alarm and unnecessary maintenance shutdown.
Due to the complex running environment of the rolling mill gearbox, fault information of key parts of the rolling mill gearbox is often submerged in strong noise, so that the signal-to-noise ratio of the collected gearbox vibration signal is low; on the other hand, the types of parts in the gear box are more, the structure is more complex, vibration responses of the parts are mutually coupled, and transmission paths are complex and changeable, so that vibration signals acquired outside the gear box often have stronger nonlinearity and non-stationarity, the fault diagnosis of the bearing parts is researched more at present, and the common fault of the bearing can be judged more accurately depending on the standard type of the structure. The fault diagnosis for the whole gearbox is not provided with a mature scheme at present due to the complex structure and mapping relation of the gearbox, and with the increasing maturity of the field of artificial intelligence algorithms, a signal feature extractor represented by a neural network is widely concerned by people, so that the fault diagnosis for the gearbox system becomes possible by reasonably analyzing the vibration data acquired by a sensor through the algorithm.
In summary, practitioners in the metallurgical industry need to find a method capable of adaptively monitoring the operation state of a rolling mill gearbox in real time and rapidly diagnosing the faults of key parts of the rolling mill gearbox so as to improve the production efficiency and the production safety.
Disclosure of Invention
The invention aims to provide a fault diagnosis system and method for a rolling mill gearbox, which are used for accurately identifying the fault type of key parts of the rolling mill main transmission system gearbox. The method can solve the defects existing in the conventional fault diagnosis method of the rolling mill gearbox: the extraction of the mutual coupling characteristics of the vibration signal responses is difficult due to the complex structure of the rolling mill gearbox; the vibration signal has large noise energy, simple time-frequency analysis is difficult to analyze the fault information of the rolling mill gearbox, and the fault type identification rate of key parts of the gearbox in the operation process is low.
In order to achieve the purpose, the invention provides the following technical scheme:
a rolling mill gearbox fault diagnostic system comprising:
the data acquisition module is used for acquiring vibration signals generated in the running process of the rolling mill;
the signal analysis processor is connected with the signal acquisition card and is used for storing the acquired vibration signals and carrying out noise reduction processing and time-frequency diagram generation on the acquired vibration signals;
the signal analysis processor is connected with a human-computer interaction interface, and the human-computer interaction interface is used for displaying the original vibration signals acquired by the data acquisition module, the signals subjected to noise reduction processing by the signal analysis processor and a time-frequency image;
the fault classification processor is connected with the signal analysis processor and is used for classifying faults of key parts of the rolling mill gearbox.
Preferably, the data acquisition module comprises acceleration sensors and a signal acquisition card which are placed at a bearing seat of a gearbox of a main transmission system of the rolling mill in the vertical direction, the horizontal direction and the axial direction.
The signal analysis processor comprises a signal noise reduction module, a time-frequency image processing module and a signal storage module, preferably, the signal noise reduction module adopts a wavelet packet threshold noise reduction algorithm to perform noise reduction processing on the original signal so as to eliminate the influence of white noise on fault classification of key parts of a subsequent gearbox in the running process of the rolling mill.
Preferably, the time-frequency image processing module is responsible for converting the denoised signal into a spectrogram simultaneously containing signal time domain characteristics and frequency domain characteristic information, and the spectrogram is used as an input sample of a subsequent fault classification processor.
Preferably, the signal storage module is responsible for storing data of a fault classification model training sample library and gear box vibration signals collected by the data acquisition card in the running process of the rolling mill.
Preferably, the signal analysis processor is connected with a human-computer interaction interface, and displays the original vibration signal acquired by the acquisition card, the signal processed by the signal noise reduction module, the time-frequency image generated by the time-frequency image processing module and the like through the human-computer interaction interface.
The fault classification processor comprises a feature extraction module, a fault classification module and a training upgrading module, preferably, the feature extraction module adopts a deep convolution neural network to carry out convolution calculation on the time-frequency image sample to extract signal features, and the signal features are matched with corresponding fault labels.
Preferably, the fault classification module uses a softmax classifier to perform final classification on the gearbox fault.
Preferably, the training and upgrading module updates the parameters of the deep convolutional neural network through signals acquired by a data acquisition card in the signal storage module, so that the fault classification processor can perform self-adaptive adjustment according to different working environments.
And the human-computer interaction interface is responsible for visualizing the working state and fault diagnosis of the gear box of the main transmission system of the rolling mill.
In addition, the invention also discloses a fault diagnosis method for the rolling mill gearbox, which specifically comprises the following steps:
step 1, obtaining vibration signals of a bearing seat in the horizontal direction, the vertical direction and the axial direction under different states of a rolling mill gearbox, carrying out equal-length division on the signals and obtaining sampling frequency;
step 2, carrying out noise reduction processing on the original vibration signal by adopting a wavelet packet domain value algorithm;
step 3, converting the processed rolling mill gearbox vibration signal into a time-frequency expression form to obtain a spectrogram of a de-noising signal;
step 4, constructing a deep convolutional neural network classifier for identifying the fault type of the rolling mill gearbox;
step 5, processing the vibration signals of the rolling mill gearbox collected in real time into image samples through the steps 1 to 3, and inputting the image samples into the deep convolutional neural network classifier constructed in the step 4 for classification and identification;
and 6, transmitting the recognition result to a human-computer interaction interface for display, and if the system detects that a key part in the gearbox fails, lighting a related failure indicator lamp in a display, and updating the section of signal into a model library.
Specifically, step 1: the method comprises the steps of collecting vibration signals of a bearing seat in the horizontal direction, the vertical direction and the axial direction under nine different states of a rolling mill gear box by using an acceleration sensor, carrying out equal-length division on the signals and obtaining sampling frequency.
Step 2: the method comprises the following steps of carrying out noise reduction processing on an original vibration signal by utilizing the diametrically opposite transfer characteristics of a useful signal and a noise signal under different scale decomposition and adopting a wavelet packet threshold algorithm, wherein the noise reduction processing method specifically comprises the following steps:
s2.1, selecting a wavelet basis function to perform wavelet packet transformation on the input signal, wherein db4 wavelets are selected in the scheme; wavelet basis function Ψ (t) and scale function
The dual-scale equation of (a) is:
wherein k is a translation factor parameter, and t is a time parameter; h (n), g (n) represent the low-pass filter parameter and the high-pass filter parameter, respectively.
S2.2, the method is popularized according to the Mallat algorithm, and the wavelet packet coefficients of the vibration signals of the rolling mill gearbox at all positions of the wavelet packet tree are obtained
S2.3, selecting a threshold function to determine a wavelet packet denoising threshold value, and subtracting the wavelet packet coefficient higher than the threshold value from the threshold value on the basis of setting the wavelet packet coefficient lower than the threshold value to be zero.
S2.4, carrying out inverse transformation reconstruction on the processed wavelet packet coefficient, wherein the reconstruction formula of the wavelet packet coefficient is as follows:
wherein m is a translation factor, t is a time parameter, and j and p respectively represent the number of decomposition layers and the positions of wavelet packet coefficients; h (), g () denote low-pass filter parameters and high-pass filter parameters, respectively.
And step 3: converting the processed rolling mill gearbox vibration signal into a time-frequency expression form to obtain a spectrogram representation of a de-noised signal, wherein the spectrogram conversion process comprises the following steps:
s3.1 selecting a window function and determining a window function length NwinAnd the number N of segments, and determining the frame shift length N according to the data number of the length of each segment signalshiftThe window function selects a hamming window.
S3.2 Length N from Window functionwinPerforming frame division processing on the denoised signal, wherein the signal of the ith frame is Si。
S3.3 pairs of SiPerforming windowing to obtain S ═ Si×hamming(Nmin) Fourier transform of S to obtain transformed spectrum ZiThe phase of which is phii。
S3.4 calculates the signal energy density spectrum p for the ith frame by the equation p ═ Zi|2. This is converted to be expressed in decibel units dB and is taken as the ith column of matrix a.
S3.5, repeating the steps and completing all elements of the matrix A.
And S3.6, mapping the matrix A to a gray image with time as a horizontal axis and frequency as a vertical axis to finish the conversion of the spectrogram.
And 4, step 4: constructing a deep convolutional neural network classifier for identifying the fault type of the rolling mill gearbox; training the deep convolutional neural network classifier by using a training sample, and verifying the classification precision of the deep convolutional neural network classifier by using a test sample. The construction method of the deep convolutional neural network classifier training sample comprises the following specific steps:
s4.1, performing wavelet packet noise reduction on the acquired vibration signals to generate a time-frequency graph, and intercepting 500000 sampling points for each fault type to generate 50 two-dimensional time-frequency graphs.
S4.2 performs a preliminary process on the image to normalize it to a standard size of 227 × 227 × 3. Each fault type randomly generated 30 groups of samples as test sets and 20 groups of samples as validation sets.
And S4.3, inputting all samples into an AlexNet model with a changed structure for training and recognition, and continuously adjusting correction parameters.
Preferably, the deep convolutional neural network classifier learning rate is set to 0.0001, the maximum number of training rounds is set to 60, and the batch size is set to 30.
And 5: and processing the acquired vibration signals of the three channels of the rolling mill gearbox into image samples through steps 1 to 3, and inputting the image samples into the deep convolutional neural network classifier constructed in the step 4 for classification and identification.
Further, typical fault types of key parts of the rolling mill gearbox include: bearing rolling element failure, bearing cage failure, bearing inner race failure, bearing outer race failure, gear missing tooth failure, tooth flank wear failure, tooth root failure, and tooth flank pitting failure.
Furthermore, the deep convolutional neural network model refers to an AlexNet model architecture, and in order to fully utilize the computation capability of the GPU, the model is divided into two parts for computation, and the structure is as follows:
the first convolution layer maps the input layer into two 55 × 55 × 48 feature maps using two 11 × 11 × 3 × 48 convolution kernels.
The second convolutional layer processes the maximally pooled first layer map into two 27 x 128 feature maps using two 5 x 48 x 128 convolutional kernels.
The third convolutional layer processes the maximally pooled feature map of the previous layer into two 13 × 13 × 192 feature maps using two 3 × 3 × 256 × 384 convolutional kernels.
The fourth convolutional layer processes the previous layer feature map into two 13 × 13 × 192 feature maps using two 3 × 3 × 192 × 192 convolution kernels.
The fifth convolutional layer processes the previous layer feature map into two 13 × 13 × 128 feature maps using two 3 × 3 × 192 × 128 convolutional kernels.
The two groups of feature maps are connected with the full connection layer, and 9 groups of fault occurrence probability parameters (including the normal state of the gear box) are output after training.
The invention has the beneficial effects that:
1. the invention can solve a series of problems in the fault diagnosis process of the rolling mill gearbox in the prior art: for example, the phenomenon that useful signals are submerged due to overlarge noise energy in the operation process of a rolling mill gearbox is solved; the extraction of the non-stable and non-linear vibration signal characteristics of the rolling mill gear box is difficult, and the like.
2. The invention realizes the online monitoring of the running state of the rolling mill gearbox and key parts by utilizing an advanced signal analysis technology and an artificial intelligent network algorithm, greatly reduces the labor cost, has high detection sensitivity, more accurate detection result, convenient equipment arrangement and visual equipment fault diagnosis result.
3. The invention has acute identification to the weak fault of key parts of the rolling mill gearbox, can provide sufficient time to implement safety procedures before the weak fault of the gearbox is developed into serious fault, and prevents maintenance shutdown and personnel safety problems.
Drawings
FIG. 1 is a schematic diagram of the system of the present invention.
FIG. 2 is a schematic representation of the diagnostic principle of the present invention.
Fig. 3 is a spectrogram of a part of fault signals in the embodiment of the present invention.
FIG. 4 is a diagram of a deep convolutional neural network classifier in accordance with an embodiment of the present invention.
Fig. 5 is a schematic diagram of the layout position of the sensors in the embodiment of the invention.
FIG. 6 is a signal display interface after vibration data monitoring and denoising of the signal analysis processor according to the embodiment of the present invention.
FIG. 7 is a data classification clustering effect display interface of the deep convolutional neural network classifier of the fault classification processor in the embodiment of the present invention.
Fig. 8 is a diagnostic result display interface in an embodiment of the present invention.
Detailed Description
In order to make the technical solution and features of the present invention clearer, a detailed description will be given below of a specific implementation flow of the present invention with reference to the accompanying drawings.
As shown in FIG. 1, the fault diagnosis system for the rolling mill gearbox mainly comprises an acceleration sensor, a multi-channel data acquisition card, a signal analysis processor, a fault classification processor, a human-computer interaction interface and the like. The signal analysis processor comprises a signal noise reduction module, an image processing module and a signal storage module; the fault classification processor comprises a feature extraction module, a fault classification module and a training upgrading module; the human-computer interaction interface comprises a state monitoring interface and information input equipment, the acceleration sensor is adsorbed on measuring points in the vertical direction, the horizontal direction and the axial direction of a rolling mill gearbox bearing seat through magnetic attraction, a specific schematic diagram is shown in detail in figure 5, wherein 1, 2 and 3 are respectively marked with 3 sensors which are arranged.
And the signal noise reduction module adopts a wavelet packet threshold noise reduction algorithm to perform noise reduction processing on the original signal so as to eliminate the influence of white noise on fault classification of key parts of a subsequent gearbox in the running process of the rolling mill.
The signal storage module is responsible for storing data of a fault classification model training sample library and gear box vibration signals collected by the data acquisition card in the running process of the rolling mill.
The signal analysis processor is connected with the human-computer interaction interface and displays the original vibration signals collected by the collection card, the signals processed by the signal noise reduction module, the time-frequency images generated by the time-frequency image processing module and the like through the human-computer interaction interface.
And the fault classification processor comprises a feature extraction module, a fault classification module, a training upgrading module and the like, is connected with the signal analysis processor, and classifies the faults of key parts of the rolling mill gearbox.
And the characteristic extraction module performs convolution calculation on the time-frequency image sample by adopting a deep convolution neural network to extract signal characteristics, and matches the signal characteristics with corresponding fault labels.
And the fault classification module adopts a softmax classifier to perform final classification on the gearbox fault.
And the training upgrading module updates the parameters of the deep convolutional neural network through signals acquired by a data acquisition card in the signal storage module so that the fault classification processor can perform self-adaptive adjustment according to different working environments.
And the human-computer interaction interface is responsible for visualizing the working state and fault diagnosis of the gear box of the main transmission system of the rolling mill.
As shown in fig. 2, the basic principle of the system includes three main parts, i.e., signal noise reduction and preprocessing, spectrogram generation and fault classification.
The signal noise reduction and preprocessing realization mainly depends on a wavelet packet noise reduction technology, and the fundamental principle of the wavelet packet noise reduction is determined by utilizing the diametrically opposite transmission characteristics of a useful signal and a noise signal under different scale decomposition. The wavelet packet coefficient amplitude of the noise signal is reduced along with the increase of the wavelet scale, the wavelet packet coefficient amplitude of the useful signal is increased along with the increase of the wavelet scale, the wavelet packet coefficient lower than the threshold is processed, the wavelet packet coefficient higher than the threshold is reserved, then the newly obtained wavelet packet coefficient is inversely transformed to reconstruct a signal, and the obtained reconstructed signal is the de-noised signal.
Compared with wavelet transform, the wavelet packet has the advantages of being capable of dividing frequency bands in multiple levels, further decomposing high-frequency parts which are not subdivided by the wavelet analysis method, and adaptively selecting proper frequency bands according to the characteristics of signals to match with the frequencies of the signals, so that the resolution of the frequencies is improved.
According to the definition of the wavelet packet, it is assumed that { V } exists
J}
J∈zIs L
2(R) and satisfying wavelet basis function Ψ (t) and scale function of two-scale relation
The dual-scale equation of (a) is:
let Ψ (t) ═ u1(t),Ф(t)=u0(t), then the dual-scale equation can be written as:
wherein k is a translation factor parameter, and t is a time parameter; h (n), g (n) represent the low-pass filter parameter and the high-pass filter parameter, respectively.
Generalizing according to Mallat algorithm, the wavelet packet coefficient of the p-th position of the j-th layer is
The next layer is dividedThe wavelet packet coefficients of the solution are:
wherein m is a translation factor, t is a time parameter, and j and p respectively represent the number of decomposition layers and the positions of wavelet packet coefficients; h (), g () denote low-pass filter parameters and high-pass filter parameters, respectively.
And (3) adopting a soft threshold noise reduction method to carry out difference between the wavelet packet coefficient higher than the threshold and the threshold on the basis of setting the wavelet packet coefficient lower than the threshold to be zero, and constructing a new wavelet packet coefficient for reconstruction according to the result.
The spectrogram generation part realizes the window function analysis of discrete signals by using short-time Fourier transform and is realized by a method of mapping gray images by using an energy spectrum, and the method comprises the following basic steps:
s3.1, selecting a window function and determining the window function length NwinAnd the number N of segments, and determining the frame shift length N according to the data number of the length of each segment signalshiftSelection of Hamming Window by Window function, where NwinAnd N can be determined according to the following formula:
wherein length is the number of sampling points, Δ T is the total signal time length, and Δ T is the single-segment signal time length
S3.2 Length N from Window functionwinPerforming frame division processing on the denoised signal, wherein the signal of the ith frame is Si。
S3.3, to SiAs a windowAfter treatment, the product is obtained S ═ Si×hamming(Nmin) Fourier transform of S to obtain transformed spectrum ZiThe phase of which is phii。
S3.4, calculating the signal energy density spectrum p of the ith frame, wherein the calculation formula is p ═ Zi|2. This is converted to dB in dB and taken as the ith column of the matrix a, where the energy density spectrum in dB is:
p(n,k)(dB)=10log10(p(n,k))
and S3.5, repeating the steps and completing all elements of the matrix A.
And S3.6, mapping the matrix A to a gray image with time as a horizontal axis and frequency as a vertical axis to finish the conversion of the spectrogram.
The fault classification part is realized by depending on a deep convolutional neural network model, and the basic steps of the fault classification part comprise spectrogram standardization, feature extraction of a convolutional layer and a pooling layer and final diagnosis result output of a full-link layer and an output layer.
The method for realizing the spectrogram standardization comprises the steps of generating a time-frequency graph after wavelet packet denoising is carried out on collected vibration signals, intercepting 500000 sampling points for each fault type, generating 50 two-dimensional time-frequency graphs and carrying out primary processing on the images to enable the images to meet the input requirements of an AlexNet model of 227 multiplied by 3. Each fault type randomly generated 30 groups of samples as test sets and 20 groups of samples as validation sets. The normalized spectrogram is shown in FIG. 3.
The convolution layer is used for extracting local features of the input layer, different convolution kernels are different feature extractors, and the process of processing the image by using the convolution layer is that the convolution kernels filter the image to obtain a feature mapping cluster. The discrete convolution calculation can be expressed as:
where y (n) is the feature map, x (n) is the input data, and x (n) is the convolution kernel.
The pooling layer functions to reduce the number of characteristic parameters, reduce the amount of computation, and prevent the over-fitting phenomenon. The most common pooling methods are maximum pooling and average pooling, both of which process the mapping extracted by the convolution kernel through a fixed and common pooling window, except that the maximum pooling is a maximum value in the window, and the average pooling is a method of averaging values in the window to construct the pooling. Optionally, the invention adopts a 2 × 2 maximum pooling method to reduce the characteristic parameters.
The full-connection layer is responsible for the final classification work of the neural network, data after convolution pooling is input into the full-connection layer, and similar to a BP algorithm, the data can obtain a result through forward weighting calculation. The result is not necessarily a true value, and the error can be solved by subtracting the true result from the result, for example, by subtracting the true result from the result. And in the same way, the error is returned one way according to the matrix multiplication. Also in this process, each neuron and nerve line naturally gets some information. We can use this information to correct the parameters, using calculus to correct the weights, which is calculated as:
yk=f(wkxk-1+bk)
wherein, yk、xk-1Respectively the output and input of the full connection layer, f is the activation function, wkWeight of the full connection layer, bkIs a deviation.
The structure of the deep convolutional neural network is shown in fig. 4, and the structure is as follows:
the first convolution layer maps the input layer into two 55 × 55 × 48 feature maps using two 11 × 11 × 3 × 48 convolution kernels.
The second convolutional layer processes the maximally pooled first layer map into two 27 x 128 feature maps using two 5 x 48 x 128 convolutional kernels.
The third convolutional layer processes the maximally pooled feature map of the previous layer into two 13 × 13 × 192 feature maps using two 3 × 3 × 256 × 384 convolutional kernels.
The fourth convolutional layer processes the previous layer feature map into two 13 × 13 × 192 feature maps using two 3 × 3 × 192 × 192 convolution kernels.
The fifth convolutional layer processes the previous layer feature map into two 13 × 13 × 128 feature maps using two 3 × 3 × 192 × 128 convolutional kernels.
The two groups of feature maps are connected with the full connection layer, and 9 groups of fault occurrence probability parameters (including the normal state of the gear box) are output after training.
In order to better complete the content realization of the rolling mill gearbox fault diagnosis system and the rolling mill gearbox fault diagnosis method, a man-machine interaction interface is compiled by utilizing LabVIEW, and as shown in FIG. 6, a signal display interface after vibration data monitoring and denoising of a signal analysis processor comprises a time and date display, a monitoring equipment indicator light, a temperature display and vibration acceleration monitoring signals and denoising signals in three directions of a vertical direction 1, a horizontal direction 2 and an axial direction 3 of a rolling mill gearbox.
As shown in FIG. 7, the data classification clustering effect display interface of the deep convolutional neural network classifier comprises t-SNE clustering effect display of deep convolutional neural network activation layers in the axial direction, the horizontal direction and the vertical direction of the rolling mill gearbox, the larger the distance between different samples of the sample characteristics in the characteristic space is, the smaller the distance between the samples of the same type is, and the more obvious the distribution difference of fault samples in the characteristic space is, the better the model training effect is.
The rolling mill gearbox fault diagnosis system and method can further display sample training accuracy and loss on a data classification clustering effect display interface of the deep convolutional neural network classifier, and the higher the training accuracy is, the smaller the loss is, the better the model classification effect is.
As shown in fig. 8, the fault diagnosis result display interface of the rolling mill gearbox fault diagnosis system and method includes 8 types of common fault early warning lamps, namely, a bearing rolling element fault, a bearing retainer fault, a bearing inner ring fault, a bearing outer ring fault, a gear missing tooth, a tooth surface abrasion, a tooth root fracture and a tooth surface pitting corrosion, and when a corresponding fault of the rolling mill gearbox is detected, the corresponding indicator lamp lights up for early warning.
The fault diagnosis result display interface of the rolling mill gearbox fault diagnosis system and method is also provided with a fault signal storage button and an input signal button; the storage fault signal button is used for storing the section of vibration signal into the model training library to update the model when the rolling mill gearbox has a fault, so that the model classification is more accurate; the input signal button is used for calling an external database to perform model training so as to achieve the better effect of model classification.
The rolling mill gearbox fault diagnosis system and method provided by the invention have a wide application prospect in the field of fault diagnosis of key parts of rolling mill equipment, and on one hand, the equipment state monitoring and fault diagnosis method combining a big data intelligent algorithm and a neural network greatly saves labor force and reduces the operation cost of enterprises. On the other hand, the invention introduces wavelet packet noise reduction technology, deep convolutional neural network technology and other big data analysis technologies into the fault diagnosis of the rolling mill gearbox, and has extremely high fault diagnosis accuracy compared with manual and traditional online fault diagnosis technologies.