Printing equipment vibration signal acquisition and equipment fault diagnosis method and system

Abstract

The invention provides a printing equipment vibration signal acquisition and equipment fault diagnosis method and system, and relates to the technical field of printing machines, and the method comprises the steps: arranging at least two vibration sensors at each key part of printing equipment, and collecting vibration signals of the printing equipment in the operation process; multiple groups of vibration signals reflecting the vibration states of different parts of the printing equipment are obtained; wavelet energy features, wavelet entropy features and wavelet coefficients corresponding to different parts are extracted, and a multi-dimensional vibration feature vector is constructed; performing trend prediction on the vibration feature vector of each key component by using a curve fitting algorithm to obtain a prediction curve reflecting the vibration trend of the key component, and judging whether the prediction curve exceeds a preset fault red line or not; when it is judged that the health state of a certain key component is abnormal, the current vibration feature vector of the component is input into the fault diagnosis model for fault type recognition, and the fault type is determined.

Description

Technical Field

[0001] The invention relates to printing machine technology, and in particular to a printing equipment vibration signal collection and equipment fault diagnosis method and system. Background Art

[0002] Printing equipment is the core equipment of the modern printing industry, and its operating status directly affects the quality of printed products and production efficiency. However, due to its complex mechanical structure and high-speed operation, printing equipment is susceptible to various faults, such as mechanical wear, loose parts, poor lubrication, etc., which lead to reduced equipment performance and product quality. In order to ensure the reliable operation and efficient production of printing equipment, it is necessary to monitor and diagnose the equipment in real time to detect and deal with potential faults in a timely manner.

[0003] Traditional printing equipment fault diagnosis mainly relies on manual inspection and regular maintenance, which observes and records the operating status of the equipment and judges whether there is a fault based on experience. This method has problems such as low diagnostic efficiency, strong subjectivity, and high misdiagnosis rate, and it is difficult to adapt to the rapid development needs of the modern printing industry. With the continuous advancement of sensor technology and data analysis technology, methods based on machine status monitoring and fault diagnosis have received more and more attention.

[0004] However, there are still some shortcomings in the existing printing equipment fault diagnosis methods. First, most methods only consider the signal of a single vibration sensor, ignoring the correlation between vibration signals of different parts, and the comprehensiveness and accuracy of diagnosis need to be improved. Second, the existing methods mainly target certain specific types of faults and lack the evaluation and prediction of the overall health status of printing equipment. In addition, the vibration signal of printing equipment is non-stationary and nonlinear. Traditional signal processing methods are difficult to fully mine the fault information contained therein, and more advanced signal analysis technology and intelligent diagnosis algorithms need to be introduced. Summary of the invention

[0005] The embodiments of the present invention provide a printing equipment vibration signal collection and equipment fault diagnosis method and system, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provides printing equipment vibration signal collection and equipment fault diagnosis methods, including:

[0008] A plurality of vibration sensors are arranged on key components of the printing equipment, wherein the key components include at least one of a printing motor, a transmission gear, and a printing cylinder, and each key component is provided with at least two vibration sensors for collecting vibration signals of the printing equipment during operation; according to the position and collection direction of each vibration sensor, the collected vibration signals are grouped and marked and synchronously processed to obtain a plurality of groups of vibration signals reflecting the vibration states of different components of the printing equipment;

[0009] Each group of vibration signals is subjected to denoising processing based on wavelet threshold, the wavelet threshold is adaptively determined, and the high-frequency noise in the vibration signal is filtered out to obtain the denoised vibration signal; multi-scale wavelet decomposition is used to perform time-frequency analysis on the denoised vibration signal, and wavelet energy features, wavelet entropy features and wavelet coefficients at different scales are extracted to construct a multi-dimensional vibration feature vector;

[0010] A curve fitting algorithm is used to predict the trend of the historical vibration feature vectors of each key component to obtain a prediction curve reflecting the vibration trend of the key components, and the health status of the key components is judged according to the changes in the prediction curve. When it is judged that the health status of a key component is abnormal, the current vibration feature vector of the component is input into the fault diagnosis model to identify the fault type and determine the fault type. The fault diagnosis model is composed of multiple stacked autoencoders, and the vibration characteristics of different fault types are deeply encoded and classified through unsupervised learning.

[0011] In an optional embodiment,

[0012] According to the position and collection direction of each vibration sensor, the collected vibration signals are grouped and identified and synchronously processed to obtain multiple groups of vibration signals reflecting the vibration status of different parts of the printing equipment, including:

[0013] According to the installation position of the vibration sensor on the printing equipment, the sensors are divided into different groups, each group corresponds to a key component, and at least two vibration sensors are installed on each key component to collect the vibration signal of the component in different directions;

[0014] The control system sends synchronous sampling instructions to each sensor at a fixed time interval. After receiving the instructions, the sensor synchronously collects vibration signals and caches data. The sampling frequency is adaptively adjusted according to the operating speed and vibration frequency characteristics of the equipment.

[0015] After each synchronous sampling period, the control system collects the vibration data cached by each sensor and divides the vibration data into different data packets according to the sensor group and acquisition timestamp. Each data packet corresponds to the vibration signal of a key component in a sampling period.

[0016] Each group of vibration signals is normalized and the vibration amplitude is mapped to the interval [0,1] to eliminate the influence of sensor sensitivity difference and range mismatch and obtain the normalized grouped vibration signals.

[0017] In an optional embodiment,

[0018] Each group of vibration signals is subjected to denoising based on wavelet threshold, the wavelet threshold is adaptively determined, and high-frequency noise in the vibration signal is filtered out. The denoised vibration signal includes:

[0019] Perform wavelet decomposition on the vibration signal to obtain wavelet coefficients at different scales, and calculate the standard deviation of the wavelet coefficients at each scale;

[0020] The frequency domain is divided into several sub-bands according to the working frequency range of the device, and the proportion of the energy of each sub-band to the total energy is calculated;

[0021] Adaptively determine the wavelet threshold at each scale. The wavelet threshold is determined by multiplying the standard deviation of the wavelet coefficient by the Gaussian noise threshold, and then multiplying it by the proportion of each sub-band energy to the total energy.

[0022] The wavelet coefficients smaller than the wavelet threshold are set to zero to obtain the processed wavelet coefficients; the processed wavelet coefficients are subjected to inverse wavelet transform to reconstruct the denoised vibration signal.

[0023] In an optional embodiment,

[0024] Multi-scale wavelet decomposition is used to perform time-frequency analysis on the denoised vibration signal, extract wavelet energy features, wavelet entropy features and wavelet coefficients at different scales, and construct a multi-dimensional vibration feature vector including:

[0025] The denoised vibration signal is subjected to multi-scale wavelet decomposition, and Daubechies wavelet is selected as the wavelet basis function. According to the frequency range and sampling frequency of the vibration signal, the optimal decomposition layer number is adaptively determined to obtain approximate components and detail components at different scales. The approximate components reflect the low-frequency trend of the signal, and the detail components reflect the high-frequency details of the signal at different scales.

[0026] The wavelet energy of the detail component at each scale is calculated, and the wavelet energy eigenvector is obtained by normalization. The high-order energy statistics of energy spectrum entropy and energy spectrum curvature are introduced.

[0027] The wavelet entropy of the detail component at each scale is calculated to obtain a multi-dimensional wavelet entropy feature vector to capture the irregular changes in vibration caused by the fault;

[0028] Extracting wavelet coefficient features of detail components at each scale, including at least one of mean, variance, skewness and kurtosis, and at least one nonlinear feature of maximum value, minimum value, median and dynamic range of wavelet coefficients, to form a multidimensional wavelet coefficient feature vector;

[0029] The extracted wavelet energy, wavelet entropy and wavelet coefficient features are reduced in dimension to obtain a multi-dimensional vibration feature vector after dimension reduction.

[0030] In an optional embodiment,

[0031] The curve fitting algorithm is used to predict the trend of the historical vibration characteristic vector of each key component, and a prediction curve reflecting the vibration trend of the key component is obtained. The health status of the key component is judged according to the change of the prediction curve, including:

[0032] Construct a time series data set of historical multi-dimensional vibration feature vectors of key components, arrange them in chronological order, and reflect the historical change trend of vibration characteristics of key components;

[0033] The trend of the historical multi-dimensional vibration feature vector of each feature dimension is predicted by using a curve fitting algorithm to obtain a prediction function representing the future change trend of each feature dimension, which constitutes the healthy prediction value of the vibration characteristics of key components;

[0034] By calculating the rate of change of each health prediction value, monitoring whether there are abnormal changes in the rate of change, such as sudden rise or fall, an early warning will be triggered when it exceeds the preset health threshold.

[0035] In an optional embodiment,

[0036] When it is determined that the health status of a key component is abnormal, the current vibration feature vector of the component is input into the fault diagnosis model to identify the fault type and determine the fault type. The fault diagnosis model is composed of a plurality of stacked autoencoders, and the vibration features of different fault types are deeply encoded and classified by unsupervised learning, including:

[0037] Construct a fault diagnosis model composed of multiple stacked autoencoders. The number of autoencoders is no less than three. Each autoencoder includes an encoder and a decoder. The encoder maps the input data to the latent space to obtain its compressed representation, and the decoder maps the compressed representation back to the original data space to reconstruct the input data.

[0038] The stacked autoencoder is unsupervisedly pre-trained using a layer-by-layer greedy training strategy. Unlabeled vibration feature data is used to minimize the reconstruction error between the autoencoder output and input, so that each layer of the autoencoder learns a hierarchical compressed representation of the vibration features.

[0039] Add a softmax classifier to the top layer of the pre-trained stacked autoencoder, use the vibration feature data with fault type labels, perform end-to-end supervised fine-tuning on the entire network, optimize the classifier parameters and autoencoder parameters, so that the network can effectively identify different fault types;

[0040] The current multi-dimensional vibration features of the critical components detected to be abnormal are input into the trained fault diagnosis model, and after layer-by-layer encoding and softmax classifier, the probability distribution vector of the fault type to which the current multi-dimensional vibration features belong is obtained;

[0041] Select the fault type with the highest probability as the diagnosis result, and output the fault type corresponding to the key component.

[0042] In the second aspect of the embodiments of the present application,

[0043] Provide printing equipment vibration signal collection and equipment fault diagnosis system, including:

[0044] The first unit is used to set a plurality of vibration sensors on key components of the printing device, wherein the key components include at least one of a printing motor, a transmission gear, and a printing cylinder, and each key component is provided with at least two vibration sensors for collecting vibration signals of the printing device during operation; according to the position and collection direction of each vibration sensor, the collected vibration signals are grouped and marked and synchronously processed to obtain a plurality of groups of vibration signals reflecting the vibration states of different components of the printing device;

[0045] The second unit is used to perform denoising processing based on wavelet threshold for each group of vibration signals, adaptively determine the wavelet threshold, filter out high-frequency noise in the vibration signal, and obtain the denoised vibration signal; use multi-scale wavelet decomposition to perform time-frequency analysis on the denoised vibration signal, extract wavelet energy features, wavelet entropy features and wavelet coefficients at different scales, and construct a multi-dimensional vibration feature vector;

[0046] The third unit is used to use a curve fitting algorithm to perform trend prediction on the historical vibration feature vectors of each key component, obtain a prediction curve reflecting the vibration trend of the key component, and judge the health status of the key component according to the change of the prediction curve; when it is judged that the health status of a key component is abnormal, the current vibration feature vector of the component is input into the fault diagnosis model to identify the fault type and determine the fault type. The fault diagnosis model is composed of a plurality of stacked autoencoders, and the vibration characteristics of different fault types are deeply encoded and classified by unsupervised learning.

[0047] According to a third aspect of the embodiments of the present invention,

[0048] An electronic device is provided, comprising:

[0049] processor;

[0050] a memory for storing processor-executable instructions;

[0051] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0052] A fourth aspect of the embodiments of the present invention is:

[0053] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0054] This application adopts an adaptive denoising method based on wavelet threshold, which can effectively filter out high-frequency noise in vibration signals, improve signal quality, and lay the foundation for subsequent feature extraction and fault diagnosis. Multi-scale wavelet decomposition is used to perform time-frequency analysis on vibration signals, extract wavelet energy, wavelet entropy, and wavelet coefficients at different scales, construct a multi-dimensional vibration feature vector, fully explore the fault feature information contained in the vibration signal, and enhance the sensitivity and accuracy of fault diagnosis.

[0055] The curve fitting algorithm is introduced to predict the trend of the historical vibration characteristics of key components. By analyzing the changes in the prediction curve, the health status of key components can be evaluated and predicted, which helps to discover potential faults in advance and avoid unplanned equipment downtime and major accidents. A fault diagnosis model based on stacked autoencoders is designed. The vibration characteristics of different fault types are deeply encoded and classified through unsupervised learning, and the specific fault type of the equipment is automatically identified, which reduces the dependence on expert experience and improves the intelligence level of fault diagnosis.

[0056] Combining health status prediction with fault diagnosis, a complete set of printing equipment status monitoring and fault diagnosis strategies has been formed, which can monitor the operating status of the equipment in real time, detect abnormalities in time, accurately diagnose the cause of the fault, provide reliable decision support for equipment maintenance and overhaul, and improve equipment reliability and production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of a method for collecting vibration signals of printing equipment and diagnosing equipment faults according to an embodiment of the present invention;

[0058] Figure 2 It is a schematic diagram of the structure of the printing equipment vibration signal collection and equipment fault diagnosis system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0061] Figure 1 FIG. 1 is a flow chart of a method for collecting vibration signals of a printing device and diagnosing device faults according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0062] S101. A plurality of vibration sensors are arranged on key components of the printing device, wherein the key components include at least one of a printing motor, a transmission gear, and a printing cylinder, and each key component is provided with at least two vibration sensors for collecting vibration signals of the printing device during operation; according to the position and collection direction of each vibration sensor, the collected vibration signals are grouped and identified and synchronously processed to obtain multiple groups of vibration signals reflecting the vibration states of different components of the printing device;

[0063] In an optional embodiment,

[0064] According to the position and collection direction of each vibration sensor, the collected vibration signals are grouped and identified and synchronously processed to obtain multiple groups of vibration signals reflecting the vibration status of different parts of the printing equipment, including:

[0065] According to the installation position of the vibration sensor on the printing equipment, the sensors are divided into different groups, each group corresponds to a key component, and at least two vibration sensors are installed on each key component to collect the vibration signal of the component in different directions;

[0066] The control system sends synchronous sampling instructions to each sensor at a fixed time interval. After receiving the instructions, the sensor synchronously collects vibration signals and caches data. The sampling frequency is adaptively adjusted according to the operating speed and vibration frequency characteristics of the equipment.

[0067] After each synchronous sampling period, the control system collects the vibration data cached by each sensor and divides the vibration data into different data packets according to the sensor group and acquisition timestamp. Each data packet corresponds to the vibration signal of a key component in a sampling period.

[0068] Each group of vibration signals is normalized and the vibration amplitude is mapped to the interval [0,1] to eliminate the influence of sensor sensitivity difference and range mismatch and obtain the normalized grouped vibration signals.

[0069] For example, when acquiring the vibration signal of a printing device, it is necessary to consider the vibration characteristics of different components and the differences in sensor layout positions. In order to more accurately reflect the vibration state of the device, this paper proposes a vibration signal grouping identification and synchronous processing method based on sensor position and acquisition direction.

[0070] First, the sensors are divided into different groups according to their installation positions on the printing equipment. Each group corresponds to a key component, such as a printing motor, transmission gear, printing cylinder, etc. At least two vibration sensors are installed on each key component to collect the vibration signals of the component in different directions. The arrangement of the sensors should fully consider the vibration characteristics and transmission direction of the components to obtain the most representative vibration information.

[0071] Secondly, in order to ensure the time synchronization of vibration signals collected by different sensors, this paper designs a synchronous sampling mechanism based on time triggering. During the operation of the printing equipment, the control system sends synchronous sampling instructions to each sensor at a fixed time interval. After receiving the instructions, the sensor synchronously collects vibration signals and caches data. The sampling frequency can be adaptively adjusted according to the operating speed and vibration frequency characteristics of the equipment to balance the data volume and diagnostic accuracy.

[0072] Then, after each synchronous sampling cycle, the control system will collect the vibration data cached by each sensor and organize them into groups. According to the sensor group and acquisition timestamp, the vibration data is divided into different data packets. Each data packet corresponds to the vibration signal of a key component in a sampling cycle. Through this grouping method, the vibration status of different components can be clearly presented, laying the foundation for subsequent feature extraction and fault diagnosis.

[0073] Finally, in order to further improve the quality and consistency of vibration signals, this paper normalizes each group of vibration signals. By mapping the vibration amplitude to the [0,1] interval, the effects of sensor sensitivity differences and range mismatch can be eliminated. The normalized vibration signal is more stable, and the vibration characteristics between different groups are more comparable.

[0074] Through the above steps, this paper realizes the grouping identification and synchronous processing of the vibration signal of the printing equipment. This method fully considers the rationality of the sensor layout and the time synchronization of data acquisition, ensuring that the acquired vibration signal can truly reflect the operating status of the equipment. At the same time, through the grouping and normalization of the vibration signal, it provides high-quality data support for subsequent feature extraction and fault diagnosis.

[0075] S102. Performing denoising processing based on wavelet threshold for each group of vibration signals, adaptively determining the wavelet threshold, filtering out high-frequency noise in the vibration signal, and obtaining a denoised vibration signal; performing time-frequency analysis on the denoised vibration signal using multi-scale wavelet decomposition, extracting wavelet energy features, wavelet entropy features and wavelet coefficients at different scales, and constructing a multi-dimensional vibration feature vector;

[0076] In an optional embodiment,

[0077] Each group of vibration signals is subjected to denoising based on wavelet threshold, the wavelet threshold is adaptively determined, and high-frequency noise in the vibration signal is filtered out. The denoised vibration signal includes:

[0078] Perform wavelet decomposition on the vibration signal to obtain wavelet coefficients at different scales, and calculate the standard deviation of the wavelet coefficients at each scale;

[0079] The frequency domain is divided into several sub-bands according to the working frequency range of the device, and the proportion of the energy of each sub-band to the total energy is calculated;

[0080] Adaptively determine the wavelet threshold at each scale. The wavelet threshold is determined by multiplying the standard deviation of the wavelet coefficient by the Gaussian noise threshold, and then multiplying it by the proportion of each sub-band energy to the total energy.

[0081] The wavelet coefficients smaller than the wavelet threshold are set to zero to obtain the processed wavelet coefficients; the processed wavelet coefficients are subjected to inverse wavelet transform to reconstruct the denoised vibration signal.

[0082] For example, the collected vibration signal often contains a large amount of high-frequency noise components due to the influence of environmental interference and the equipment's own noise. In order to improve the accuracy of fault diagnosis, the vibration signal needs to be denoised. This paper adopts an adaptive denoising method based on wavelet threshold. By adaptively determining the wavelet threshold, the high-frequency noise in the vibration signal is effectively filtered out to obtain pure fault feature information.

[0083] Wavelet threshold denoising is a commonly used signal denoising method. Its basic principle is to decompose the signal into wavelet coefficients of different scales through wavelet transform, then perform threshold processing on the wavelet coefficients to filter out high-frequency noise, and finally reconstruct the denoised signal through inverse wavelet transform.

[0084] The selection of threshold is the key to wavelet denoising. The appropriate threshold can suppress noise to the greatest extent while retaining the signal characteristics. Traditional threshold selection methods usually use fixed thresholds. However, fixed thresholds ignore the time-varying and frequency characteristics of the signal and are difficult to adapt to complex vibration signals.

[0085] In order to better remove noise from vibration signals, this paper proposes an adaptive threshold determination method. This method combines the statistical characteristics of vibration signals and the frequency domain energy distribution, adaptively adjusts the threshold value, and achieves the best denoising effect.

[0086] First, each group of vibration signals is decomposed by wavelet to obtain wavelet coefficients at different scales. At each scale, the standard deviation of the wavelet coefficients is calculated to reflect the intensity of the noise at that scale.

[0087] Secondly, analyze the energy distribution of vibration signals in different frequency bands. According to the working frequency range of the equipment, the frequency domain is divided into several sub-bands. For each sub-band, calculate the proportion of its energy to the total energy to reflect the importance of the frequency band.

[0088] Then, the threshold at each scale is adaptively determined by comprehensively considering the noise intensity and frequency band importance:

[0089]

[0090] Among them, λ j represents the j-th scale wavelet threshold, σ j represents the standard deviation of the j-th scale wavelet coefficient, N represents the signal length, Σ i p i Represents the sum of the energy proportions of all sub-bands, p i represents the energy proportion of the i-th sub-band, Indicates the energy proportion of the most important sub-band.

[0091] Based on the above-mentioned adaptive threshold determination method, the specific steps of wavelet threshold denoising for each group of vibration signals are as follows:

[0092] The vibration signal is decomposed by wavelet to obtain wavelet coefficients at different scales; at each scale, the standard deviation of the wavelet coefficients is calculated; according to the working frequency range of the equipment, the frequency domain is divided into several sub-bands, and the energy proportion of each sub-band is calculated; the adaptive threshold at each scale is calculated; the wavelet coefficients at each scale are threshold processed to obtain the processed wavelet coefficients; the processed wavelet coefficients are inversely transformed to reconstruct the denoised vibration signal.

[0093] Through the above steps, adaptive wavelet threshold denoising processing is achieved for each group of vibration signals. This method can adaptively adjust the threshold size according to the actual characteristics of the vibration signal, effectively filter out high-frequency noise, and retain useful fault feature information to the greatest extent. The quality of the denoised vibration signal is significantly improved, laying a good foundation for subsequent feature extraction and fault diagnosis.

[0094] The adaptive denoising method based on wavelet threshold proposed in this application fully considers the non-stationarity and frequency domain characteristics of the vibration signal, and realizes effective noise suppression and retention of signal characteristics by determining the adaptive threshold. Compared with the traditional fixed threshold denoising, this method has stronger adaptability and robustness, and can cope with complex and changeable vibration signals. At the same time, this method is simple to calculate and easy to implement in engineering, and is suitable for real-time online fault diagnosis systems.

[0095] In an optional embodiment,

[0096] Multi-scale wavelet decomposition is used to perform time-frequency analysis on the denoised vibration signal, extract wavelet energy features, wavelet entropy features and wavelet coefficients at different scales, and construct a multi-dimensional vibration feature vector including:

[0097] The denoised vibration signal is subjected to multi-scale wavelet decomposition, and Daubechies wavelet is selected as the wavelet basis function. According to the frequency range and sampling frequency of the vibration signal, the optimal decomposition layer number is adaptively determined to obtain approximate components and detail components at different scales. The approximate components reflect the low-frequency trend of the signal, and the detail components reflect the high-frequency details of the signal at different scales.

[0098] The wavelet energy of the detail component at each scale is calculated, and the wavelet energy eigenvector is obtained by normalization. The high-order energy statistics of energy spectrum entropy and energy spectrum curvature are introduced.

[0099] The wavelet entropy of the detail component at each scale is calculated to obtain a multi-dimensional wavelet entropy feature vector to capture the irregular changes in vibration caused by the fault;

[0100] Extracting wavelet coefficient features of detail components at each scale, including at least one of mean, variance, skewness and kurtosis, and at least one nonlinear feature of maximum value, minimum value, median and dynamic range of wavelet coefficients, to form a multidimensional wavelet coefficient feature vector;

[0101] The extracted wavelet energy, wavelet entropy and wavelet coefficient features are reduced in dimension to obtain a multi-dimensional vibration feature vector after dimension reduction.

[0102] For example, wavelet decomposition is an effective time-frequency analysis tool that can decompose signals into approximate components and detail components at different scales. By selecting appropriate wavelet basis functions and decomposition levels, multi-resolution analysis of signals can be performed in both time and frequency dimensions to extract feature information of different frequency bands.

[0103] Assume that the denoised vibration signal is Select Daubechies wavelet as the wavelet basis function and perform L-layer wavelet decomposition on it. The wavelet decomposition process can be expressed as:

[0104]

[0105] Among them, A L (t) represents the approximate component of the Lth layer, reflecting the low-frequency trend of the signal; D j (t) represents the detail component of the jth layer, reflecting the high-frequency details of the signal at different scales.

[0106] After obtaining the denoised vibration signal, it is necessary to further extract the fault feature information contained therein to provide input for the training and prediction of the intelligent diagnosis model. This paper adopts a feature extraction method based on multi-scale wavelet decomposition. By performing time-frequency analysis on the vibration signal, wavelet energy features, wavelet entropy features and wavelet coefficient features at different scales are extracted, and a multi-dimensional vibration feature vector is constructed to comprehensively characterize the time-frequency domain characteristics of the vibration signal.

[0107] At each scale, the energy distribution of the vibration signal reflects the vibration intensity and frequency components at that scale. Faults usually cause abnormal changes in energy at certain scales, so the fault characteristics can be characterized by extracting wavelet energy features.

[0108] For the detail component D of the jth layer j (t), its energy can be expressed as:

[0109] E j =Σ k |d j,k | 2 ;

[0110] Among them, d j,k is the kth wavelet coefficient of the jth layer detail component. Normalize the energy of each layer detail component to obtain the wavelet energy eigenvector:

[0111] E=[E1 / E total ,E2 / E total ,…,E L / E total ];

[0112] Among them, E totalis the total energy. The wavelet energy eigenvector reflects the energy distribution of the vibration signal in different frequency bands and can effectively characterize the energy changes caused by faults.

[0113] In addition to energy distribution, the complexity and irregularity of vibration signals at different scales can also reflect fault characteristics. Wavelet entropy is an indicator of signal complexity and can be used to measure the uncertainty and disorder of vibration signals.

[0114] For the detail component of the jth layer, its wavelet entropy can be defined as:

[0115] H j =-Σ k p j,k logp j,k ;

[0116] Among them, p j,k =|d j,k | 2 / E j is the energy probability of the kth wavelet coefficient of the jth layer detail component. The larger the wavelet entropy, the higher the complexity and uncertainty of the vibration signal at this scale. The wavelet entropy of each layer detail component is combined to obtain the wavelet entropy eigenvector; the wavelet entropy eigenvector reflects the complexity distribution of the vibration signal at different scales and can capture the vibration mode changes caused by faults.

[0117] In addition to energy and entropy features, the wavelet coefficients themselves also contain rich fault information. Different fault types usually show obvious wavelet coefficient anomalies at certain specific scales. Therefore, the wavelet coefficients can be directly extracted as fault features.

[0118] For the detail component of the jth layer, the statistical characteristics of its wavelet coefficients, such as mean, variance, skewness and kurtosis, are selected to form the wavelet coefficient feature vector:

[0119] C j =[μ j ,σ j ,γ i ,κ j ];

[0120] Among them, μ j ,σ j ,γ j ,κ j are the mean, variance, skewness and kurtosis of the wavelet coefficients of the jth layer detail component respectively. The wavelet coefficient eigenvectors of each layer are cascaded to obtain a complete wavelet coefficient eigenvector. The wavelet coefficient eigenvector directly describes the time-frequency domain characteristics of the vibration signal at different scales and can sensitively reflect the local waveform changes caused by the fault.

[0121] Through the above steps, the time-frequency domain features of the vibration signal are extracted from multiple perspectives, including wavelet energy features, wavelet entropy features, and wavelet coefficient features. In order to fully characterize the characteristics of the vibration signal, these features are combined to construct a multidimensional vibration feature vector.

[0122] The multidimensional vibration feature vector integrates the energy distribution, complexity distribution and time-frequency local characteristics of the vibration signal. It can characterize the vibration anomaly caused by the fault from multiple dimensions and improve the expression ability of the fault characteristics and the diagnostic accuracy.

[0123] The vibration feature extraction method based on multi-scale wavelet decomposition proposed in this application makes full use of the time-frequency localization characteristics of wavelet transform, and extracts the characteristic information of vibration signals at different time and frequency scales through multi-scale decomposition. Compared with traditional time domain or frequency domain feature extraction methods, this method can more comprehensively characterize the non-stationary and instantaneous characteristics of vibration signals and capture weak anomalies caused by faults. At the same time, by constructing a multi-dimensional vibration feature vector, the energy, complexity and local characteristics of the vibration signal are comprehensively considered, which improves the discrimination of fault features and the reliability of diagnosis.

[0124] S103. Use a curve fitting algorithm to predict the trend of the historical vibration feature vectors of each key component, obtain a prediction curve reflecting the vibration trend of the key component, and judge the health status of the key component according to the change of the prediction curve; when it is judged that the health status of a key component is abnormal, input the current vibration feature vector of the component into the fault diagnosis model to identify the fault type and determine the fault type. The fault diagnosis model is composed of a plurality of stacked autoencoders, and the vibration characteristics of different fault types are deeply encoded and classified by unsupervised learning.

[0125] In an optional embodiment,

[0126] The curve fitting algorithm is used to predict the trend of the historical vibration characteristic vector of each key component, and a prediction curve reflecting the vibration trend of the key component is obtained. The health status of the key component is judged according to the change of the prediction curve, including:

[0127] Construct a time series data set of historical multi-dimensional vibration feature vectors of key components, arrange them in chronological order, and reflect the historical change trend of vibration characteristics of key components;

[0128] The trend of the historical multi-dimensional vibration feature vector of each feature dimension is predicted by using a curve fitting algorithm to obtain a prediction function representing the future change trend of each feature dimension, which constitutes the healthy prediction value of the vibration characteristics of key components;

[0129] By calculating the rate of change of each health prediction value, monitoring whether there are abnormal changes in the rate of change, such as sudden rise or fall, an early warning will be triggered when it exceeds the preset health threshold.

[0130] For example, after obtaining the multi-dimensional vibration characteristic vector of key components, the historical characteristic data can be further used to predict and evaluate the health status of key components. This paper proposes a health status prediction method based on curve fitting. By performing trend analysis and curve fitting on the historical vibration characteristic vector of each key component, a prediction curve reflecting the vibration trend of the key component is obtained. The health status of the key component is judged according to the change of the prediction curve, realizing predictive maintenance and fault warning of the equipment.

[0131] Assume that a key component has M vibration feature vectors at historical time points, and the corresponding timestamps are t1, t2, …, t M . Arrange these feature vectors in chronological order and construct a time series dataset:

[0132] X={F1,F2,…,F i};

[0133] Among them, F i is the multidimensional vibration feature vector extracted at the i-th time point. The time series data set reflects the historical change trend of the vibration characteristics of key components and can be used for subsequent trend prediction and health status assessment.

[0134] For time series data sets, curve fitting algorithms can be used to predict the trend of data in each feature dimension. Commonly used curve fitting algorithms include polynomial fitting, exponential fitting, logarithmic fitting, and power function fitting. Here, we take polynomial fitting as an example and perform n-order polynomial fitting on the data of the j-th feature dimension:

[0135] f j (t) = a 0,j +a 1,j t+a 2,j t 2 +…+a n,j t n ;

[0136] Among them, f j (t) is the prediction function of the jth feature dimension, a 0,j ,a 1,j ,…,a n,j is the polynomial coefficient. Through optimization algorithms such as the least squares method, the optimal polynomial coefficient can be estimated so that the prediction function f j (t) The fitting error with the actual data point is the smallest.

[0137] Curve fitting is performed on all feature dimensions to obtain a set of prediction functions, which constitute the prediction curve of the vibration characteristics of key components. The prediction curve reflects the change trend of each vibration characteristic of key components in the future and can be used for health status assessment and early warning.

[0138] According to the changes in the prediction curve, it can be determined whether the health status of key components has changed significantly and whether there is a potential risk of failure. Specifically, the following strategies can be used for health status assessment and early warning:

[0139] Abnormal deviation of characteristic values: Monitor whether the predicted value of each characteristic dimension has abnormal deviation. If the deviation exceeds the preset threshold, an early warning is triggered. The preset threshold can be determined based on the statistical distribution of historical data.

[0140] Abnormal feature change rate: Calculate the change rate of the predicted value of each feature dimension, such as the first-order derivative or second-order derivative. Monitor whether the change rate is abnormal, such as sudden change, sharp rise or sharp drop, and trigger an early warning if it exceeds the preset threshold.

[0141] Multi-feature comprehensive evaluation: Comprehensively consider the changes in the prediction curves of multiple feature dimensions, build health indicators or health metrics, and quantify the overall health status of key components. When the health indicator or health metric exceeds the preset threshold, an early warning is triggered.

[0142] Remaining life prediction: Use the prediction curve to extrapolate to a longer time scale in the future, combined with the failure threshold, to predict the remaining life of key components. When the predicted remaining life is lower than the preset threshold, an early warning is triggered and maintenance is arranged.

[0143] When an early warning is triggered, corresponding maintenance decisions can be taken according to the warning level, such as strengthening monitoring, adjusting operating parameters, arranging maintenance or replacing parts, etc., thereby achieving predictive maintenance and fault prevention of equipment.

[0144] The key component health status prediction method based on curve fitting proposed in this application uses historical vibration feature data, through trend analysis and curve fitting, to construct a prediction curve to reflect the future change trend of the vibration characteristics of key components. Compared with simple threshold detection, this method can capture the gradual trend and dynamic change of features, and provide prediction capabilities on a longer time scale. At the same time, by comprehensively analyzing the prediction curves of multiple feature dimensions, considering the correlation and interaction between features, the health status of key components is comprehensively evaluated to improve the reliability and accuracy of early warning.

[0145] This prediction method requires less computing resources and is easy to implement in engineering. The prediction curve is intuitive and clear, which is easy for engineers to understand and make decisions. At the same time, the prediction model can be updated and adaptively adjusted online according to new monitoring data to continuously improve the prediction performance. Combining this prediction method with the previous feature extraction method to form a complete set of equipment health monitoring and predictive maintenance solutions based on vibration signals can significantly improve the reliability and availability of equipment, extend equipment life, and reduce unexpected downtime and maintenance costs.

[0146] In an optional embodiment,

[0147] When it is determined that the health status of a key component is abnormal, the current vibration feature vector of the component is input into the fault diagnosis model to identify the fault type and determine the fault type. The fault diagnosis model is composed of a plurality of stacked autoencoders, and the vibration features of different fault types are deeply encoded and classified by unsupervised learning, including:

[0148] Construct a fault diagnosis model composed of multiple stacked autoencoders. The number of autoencoders is no less than three. Each autoencoder includes an encoder and a decoder. The encoder maps the input data to the latent space to obtain its compressed representation, and the decoder maps the compressed representation back to the original data space to reconstruct the input data.

[0149] The stacked autoencoder is unsupervisedly pre-trained using a layer-by-layer greedy training strategy. Unlabeled vibration feature data is used to minimize the reconstruction error between the autoencoder output and input, so that each layer of the autoencoder learns a hierarchical compressed representation of the vibration features.

[0150] Add a softmax classifier to the top layer of the pre-trained stacked autoencoder, use the vibration feature data with fault type labels, perform end-to-end supervised fine-tuning on the entire network, optimize the classifier parameters and autoencoder parameters, so that the network can effectively identify different fault types;

[0151] The current multi-dimensional vibration features of the critical components detected to be abnormal are input into the trained fault diagnosis model, and after layer-by-layer encoding and softmax classifier, the probability distribution vector of the fault type to which the current multi-dimensional vibration features belong is obtained;

[0152] Select the fault type with the highest probability as the diagnosis result, and output the fault type corresponding to the key component.

[0153] For example, when the health status prediction module determines that the health status of a key component is abnormal, it is necessary to further identify and determine the specific fault type in order to take targeted maintenance measures. This paper proposes a fault diagnosis model based on stacked autoencoders, which deeply encodes and classifies the vibration characteristics of different fault types through unsupervised learning to achieve automatic identification of fault types.

[0154] Stacked Autoencoder (SAE) is a deep neural network model composed of multiple stacked Autoencoders (AE). Autoencoder is an unsupervised learning model that aims to learn the compressed representation of input data and reconstruct the original input from the compressed representation. The Autoencoder consists of two parts: the encoder and the decoder. The encoder maps the input data to the hidden space to obtain its compressed representation, and the decoder maps the compressed representation back to the original data space to reconstruct the input data.

[0155] The stacked autoencoder constructs a deep network structure by stacking multiple autoencoders. The hidden layer output of the first autoencoder is used as the input of the second autoencoder, and so on, forming a multi-level feature learning framework. The stacked autoencoder can learn the hierarchical feature representation of the input data layer by layer, from low-level local features to high-level global abstract features, enhancing the richness and discriminability of the feature representation.

[0156] The fault diagnosis model designed in this paper adopts a three-layer stacked autoencoder structure. The input of the first layer of autoencoder is the multi-dimensional vibration feature vector of the key component, which is mapped to the first latent space by the encoder to obtain the first layer of compressed representation of the vibration feature. The second layer of autoencoder takes the hidden layer output of the first layer as input to learn the second layer of compressed representation of the vibration feature. The third layer of autoencoder takes the hidden layer output of the second layer as input to learn the third layer of compressed representation of the vibration feature as the deep feature of fault diagnosis.

[0157] The training process of the stacked autoencoder is divided into two stages: unsupervised pre-training and supervised fine-tuning. The unsupervised pre-training stage adopts a layer-by-layer greedy training strategy, training one autoencoder at a time, while fixing the parameters of other layers. Specifically, the first layer of autoencoder is first trained with unlabeled vibration feature data to learn the first layer of compressed representation of vibration features. Then, the hidden layer output of the first layer of autoencoder is used as the input of the second layer of autoencoder to train the second layer of autoencoder. And so on, each autoencoder is trained layer by layer to obtain a pre-trained stacked autoencoder model.

[0158] The reconstruction error is used as the optimization target in the pre-training phase, that is, to minimize the difference between the output and input of the autoencoder. Commonly used reconstruction error metrics include Mean Squared Error (MSE) and Cross-Entropy Error. By minimizing the reconstruction error, the autoencoder can learn an effective compressed representation of the input data and capture the intrinsic structure and key features of the data.

[0159] After pre-training, the stacked autoencoder needs to be supervised and fine-tuned to adapt to specific fault diagnosis tasks. The fine-tuning stage adds a softmax classifier to the highest layer of the pre-trained model to map deep features to fault type labels. Then, the entire network is end-to-end supervised training using vibration feature data with fault type labels to optimize classifier parameters and autoencoder parameters so that the network can effectively identify different fault types. The fine-tuning process usually uses the back-propagation algorithm and gradient descent optimization method, with cross entropy loss or classification error as the optimization target, and continuously adjusts network parameters to improve fault classification performance.

[0160] After pre-training and fine-tuning, a complete fault diagnosis model is obtained, which can map the vibration characteristics of key components to the corresponding fault types. When the health status prediction module detects that a key component is abnormal, its current multi-dimensional vibration feature vector is input into the trained fault diagnosis model, and after layer-by-layer encoding and softmax classifier, the fault type to which the vibration feature belongs is obtained.

[0161] The output of the fault diagnosis model is a probability distribution vector, which indicates the probability that the input vibration feature belongs to each fault type. The fault type with the highest probability is selected as the diagnosis result.

[0162] The fault diagnosis method based on stacked autoencoders has the following advantages:

[0163] It is possible to learn the deep representation of vibration features through unsupervised pre-training, explore the intrinsic correlation and high-level semantics between features, and enhance the discriminative ability of features. The layer-by-layer greedy training strategy is adopted to alleviate the optimization difficulty of deep networks and help learn more effective feature representations. Supervised fine-tuning is performed on the basis of pre-training, and the model is tuned using fault label information to improve the accuracy of fault classification. The model has strong generalization ability, can adapt to fault diagnosis tasks under different working conditions and environments, and also has a certain recognition ability for unknown fault types.

[0164] The fault diagnosis model proposed in this application is combined with the previous health status prediction method to form a complete equipment health monitoring and fault diagnosis process. When the prediction module detects that the health status of key components is abnormal, the fault diagnosis module is triggered, and the stacked autoencoder model is used to identify the specific fault type, thereby realizing automatic fault diagnosis of the equipment and providing support for subsequent maintenance decisions. This method makes full use of the advantages of deep learning technology in feature learning and pattern recognition, and improves the intelligence level and practicality of fault diagnosis.

[0165] Figure 2 FIG. 1 is a schematic diagram of the structure of a system for collecting vibration signals of printing equipment and diagnosing equipment faults according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0166] The first unit is used to set a plurality of vibration sensors on key components of the printing device, wherein the key components include at least one of a printing motor, a transmission gear, and a printing cylinder, and each key component is provided with at least two vibration sensors for collecting vibration signals of the printing device during operation; according to the position and collection direction of each vibration sensor, the collected vibration signals are grouped and marked and synchronously processed to obtain a plurality of groups of vibration signals reflecting the vibration states of different components of the printing device;

[0167] The second unit is used to perform denoising processing based on wavelet threshold for each group of vibration signals, adaptively determine the wavelet threshold, filter out high-frequency noise in the vibration signal, and obtain the denoised vibration signal; use multi-scale wavelet decomposition to perform time-frequency analysis on the denoised vibration signal, extract wavelet energy features, wavelet entropy features and wavelet coefficients at different scales, and construct a multi-dimensional vibration feature vector;

[0168] The third unit is used to use a curve fitting algorithm to perform trend prediction on the historical vibration feature vectors of each key component, obtain a prediction curve reflecting the vibration trend of the key component, and judge the health status of the key component according to the change of the prediction curve; when it is judged that the health status of a key component is abnormal, the current vibration feature vector of the component is input into the fault diagnosis model to identify the fault type and determine the fault type. The fault diagnosis model is composed of a plurality of stacked autoencoders, and the vibration characteristics of different fault types are deeply encoded and classified by unsupervised learning.

[0169] According to a third aspect of the embodiments of the present invention,

[0170] An electronic device is provided, comprising:

[0171] Processor;

[0172] a memory for storing processor-executable instructions;

[0173] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0174] A fourth aspect of the embodiments of the present invention is:

[0175] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0176] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0177] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A printing equipment vibration signal collection and equipment fault diagnosis method, characterized in that: include: A plurality of vibration sensors are arranged on key components of the printing equipment, wherein the key components include at least one of a printing motor, a transmission gear, and a printing cylinder, and each key component is provided with at least two vibration sensors for collecting vibration signals of the printing equipment during operation; according to the position and collection direction of each vibration sensor, the collected vibration signals are grouped and marked and synchronously processed to obtain a plurality of groups of vibration signals reflecting the vibration states of different components of the printing equipment; Each group of vibration signals is subjected to denoising processing based on wavelet threshold, the wavelet threshold is adaptively determined, and the high-frequency noise in the vibration signal is filtered out to obtain the denoised vibration signal; multi-scale wavelet decomposition is used to perform time-frequency analysis on the denoised vibration signal, and wavelet energy features, wavelet entropy features and wavelet coefficients at different scales are extracted to construct a multi-dimensional vibration feature vector; A curve fitting algorithm is used to predict the trend of the historical vibration feature vectors of each key component to obtain a prediction curve reflecting the vibration trend of the key components, and the health status of the key components is judged according to the changes in the prediction curve. When it is judged that the health status of a key component is abnormal, the current vibration feature vector of the component is input into the fault diagnosis model to identify the fault type and determine the fault type. The fault diagnosis model is composed of multiple stacked autoencoders, and the vibration characteristics of different fault types are deeply encoded and classified through unsupervised learning.

2. The method according to claim 1, characterized in that According to the position and collection direction of each vibration sensor, the collected vibration signals are grouped and identified and synchronously processed to obtain multiple groups of vibration signals reflecting the vibration status of different parts of the printing equipment, including: According to the installation position of the vibration sensor on the printing equipment, the sensors are divided into different groups, each group corresponds to a key component, and at least two vibration sensors are installed on each key component to collect the vibration signal of the component in different directions; The control system sends synchronous sampling instructions to each sensor at a fixed time interval. After receiving the instructions, the sensor synchronously collects vibration signals and caches data. The sampling frequency is adaptively adjusted according to the operating speed and vibration frequency characteristics of the equipment. After each synchronous sampling period, the control system collects the vibration data cached by each sensor and divides the vibration data into different data packets according to the sensor group and acquisition timestamp. Each data packet corresponds to the vibration signal of a key component in a sampling period. Each group of vibration signals is normalized and the vibration amplitude is mapped to the interval [0,1] to eliminate the influence of sensor sensitivity difference and range mismatch and obtain the normalized grouped vibration signals.

3. The method according to claim 1, characterized in that Each group of vibration signals is subjected to denoising based on wavelet threshold, the wavelet threshold is adaptively determined, and high-frequency noise in the vibration signal is filtered out. The denoised vibration signal includes: Perform wavelet decomposition on the vibration signal to obtain wavelet coefficients at different scales, and calculate the standard deviation of the wavelet coefficients at each scale; The frequency domain is divided into several sub-bands according to the working frequency range of the device, and the proportion of the energy of each sub-band to the total energy is calculated; Adaptively determine the wavelet threshold at each scale. The wavelet threshold is determined by multiplying the standard deviation of the wavelet coefficient by the Gaussian noise threshold, and then multiplying it by the proportion of each sub-band energy to the total energy. The wavelet coefficients smaller than the wavelet threshold are set to zero to obtain the processed wavelet coefficients; the processed wavelet coefficients are subjected to inverse wavelet transform to reconstruct the denoised vibration signal.

4. The method according to claim 3, characterized in that Multi-scale wavelet decomposition is used to perform time-frequency analysis on the denoised vibration signal, extract wavelet energy features, wavelet entropy features and wavelet coefficients at different scales, and construct a multi-dimensional vibration feature vector including: The denoised vibration signal is subjected to multi-scale wavelet decomposition, and Daubechies wavelet is selected as the wavelet basis function. According to the frequency range and sampling frequency of the vibration signal, the optimal decomposition layer number is adaptively determined to obtain approximate components and detail components at different scales. The approximate components reflect the low-frequency trend of the signal, and the detail components reflect the high-frequency details of the signal at different scales. The wavelet energy of the detail component at each scale is calculated, and the wavelet energy eigenvector is obtained by normalization. The high-order energy statistics of energy spectrum entropy and energy spectrum curvature are introduced. The wavelet entropy of the detail component at each scale is calculated to obtain a multi-dimensional wavelet entropy feature vector to capture the irregular vibration changes caused by the fault; Extracting wavelet coefficient features of detail components at each scale, including at least one of mean, variance, skewness and kurtosis, and at least one nonlinear feature of maximum value, minimum value, median and dynamic range of wavelet coefficients, to form a multidimensional wavelet coefficient feature vector; The extracted wavelet energy, wavelet entropy and wavelet coefficient features are reduced in dimension to obtain a multi-dimensional vibration feature vector after dimension reduction.

5. The method according to claim 1, characterized in that The curve fitting algorithm is used to predict the trend of the historical vibration characteristic vector of each key component, and a prediction curve reflecting the vibration trend of the key component is obtained. The health status of the key component is judged according to the change of the prediction curve, including: Construct a time series data set of historical multi-dimensional vibration feature vectors of key components, arrange them in chronological order, and reflect the historical change trend of vibration characteristics of key components; The curve fitting algorithm is used to predict the trend of the historical multi-dimensional vibration feature vector of each feature dimension, and the prediction function representing the future change trend of each feature dimension is obtained to form the health prediction value of the vibration characteristics of key components; By calculating the rate of change of each health prediction value, monitoring whether there are abnormal changes in the rate of change, such as sudden rise or fall, an early warning will be triggered when it exceeds the preset health threshold.

6. The method according to claim 1, characterized in that When it is determined that the health status of a key component is abnormal, the current vibration feature vector of the component is input into the fault diagnosis model to identify the fault type and determine the fault type. The fault diagnosis model is composed of a plurality of stacked autoencoders, and the vibration features of different fault types are deeply encoded and classified by unsupervised learning, including: Construct a fault diagnosis model composed of multiple stacked autoencoders. The number of autoencoders is no less than three. Each autoencoder includes an encoder and a decoder. The encoder maps the input data to the latent space to obtain its compressed representation, and the decoder maps the compressed representation back to the original data space to reconstruct the input data. The stacked autoencoder is unsupervisedly pre-trained using a layer-by-layer greedy training strategy. Unlabeled vibration feature data is used to minimize the reconstruction error between the autoencoder output and input, so that each layer of the autoencoder learns a hierarchical compressed representation of the vibration features. Add a softmax classifier to the top layer of the pre-trained stacked autoencoder, use the vibration feature data with fault type labels, perform end-to-end supervised fine-tuning on the entire network, optimize the classifier parameters and autoencoder parameters, so that the network can effectively identify different fault types; The current multi-dimensional vibration features of the critical components detected to be abnormal are input into the trained fault diagnosis model, and after layer-by-layer encoding and softmax classifier, the probability distribution vector of the fault type to which the current multi-dimensional vibration features belong is obtained; Select the fault type with the highest probability as the diagnosis result, and output the fault type corresponding to the key component.

7. A printing equipment vibration signal collection and equipment fault diagnosis system, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to set a plurality of vibration sensors on key components of the printing device, wherein the key components include at least one of a printing motor, a transmission gear, and a printing cylinder, and each key component is provided with at least two vibration sensors for collecting vibration signals of the printing device during operation; according to the position and collection direction of each vibration sensor, the collected vibration signals are grouped and marked and synchronously processed to obtain a plurality of groups of vibration signals reflecting the vibration states of different components of the printing device; The second unit is used to perform denoising processing based on wavelet threshold for each group of vibration signals, adaptively determine the wavelet threshold, filter out high-frequency noise in the vibration signal, and obtain the denoised vibration signal; use multi-scale wavelet decomposition to perform time-frequency analysis on the denoised vibration signal, extract wavelet energy features, wavelet entropy features and wavelet coefficients at different scales, and construct a multi-dimensional vibration feature vector; The third unit is used to use a curve fitting algorithm to perform trend prediction on the historical vibration feature vectors of each key component, obtain a prediction curve reflecting the vibration trend of the key component, and judge the health status of the key component according to the change of the prediction curve; when it is judged that the health status of a key component is abnormal, the current vibration feature vector of the component is input into the fault diagnosis model to identify the fault type and determine the fault type. The fault diagnosis model is composed of a plurality of stacked autoencoders, and the vibration characteristics of different fault types are deeply encoded and classified by unsupervised learning.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images