Printing machine fault diagnosis method based on acoustic emission detection

Abstract

The invention relates to the technical field of printing machine faults, and discloses a printing machine fault diagnosis method based on acoustic emission detection, which comprises the following steps of: constructing a printing machine acoustic emission signal acquisition system which comprises a plurality of acoustic emission sensors, a signal preprocessing module and a data transmission unit, the plurality of acoustic emission sensors are distributed at a key transmission part, a printing execution part and a supporting structure of the printing machine so as to synchronously collect acoustic emission original signals at different positions; the method comprises the following steps: performing multi-dimensional preprocessing on an acquired acoustic emission original signal, wherein the preprocessing comprises noise reduction processing based on a self-adaptive wavelet threshold value; acoustic emission signals of key parts of the printing machine are synchronously collected through distributed acoustic emission sensors, time domain, frequency domain and time-frequency domain features are extracted in combination with multi-dimensional preprocessing, features and position weights are dynamically given by using a fault diagnosis model fused with an attention mechanism, and accurate recognition of the fault type and position of the printing machine is achieved.

Description

Technical Field

[0001] This invention relates to the field of printing press fault technology, specifically to a printing press fault diagnosis method based on acoustic emission detection. Background Technology

[0002] As core equipment in industries such as packaging printing and book printing, the operating status of printing presses directly affects product quality and production efficiency. With the printing industry's development towards higher speeds and greater precision, printing presses are becoming increasingly complex, incorporating multiple key subsystems including transmission systems, printing cylinder assemblies, paper feeding systems, and control systems. During long-term high-load operation, these components are prone to malfunctions due to wear, fatigue, and assembly errors. Failure to diagnose and address these issues promptly can lead to equipment downtime, product scrapping, or even safety accidents, resulting in significant economic losses. Therefore, achieving accurate and efficient fault diagnosis for printing presses is crucial for ensuring production continuity and reducing maintenance costs.

[0003] Currently, printing press fault diagnosis technology mainly relies on a combination of traditional detection methods and intelligent diagnostic tools. Traditional methods involve manual inspections, which rely on experience to judge based on observations of abnormal noises, vibrations, and ink adhesion. However, this is limited by the professional level and reaction speed of the inspectors, making it difficult to accurately identify early faults and carries the risk of missed detections and misdiagnosis. Vibration detection technology analyzes fault characteristics by collecting vibration signals from the equipment. While widely used in diagnosing rotating parts, the printing press operating environment is subject to multi-source vibration interference from paper feeding, roller compression, etc., which can easily drown out useful signals, especially making it less sensitive to early, minor faults. Visual inspection technology identifies printing quality defects through image recognition, indirectly reflecting equipment abnormalities, but it cannot directly locate potential faults in mechanical components and is significantly affected by environmental factors such as light and oil contamination.

[0004] With the development of sensor technology and artificial intelligence, acoustic emission detection technology is gradually being applied to the field of equipment fault diagnosis. Acoustic emission is the phenomenon of materials or structures releasing stress waves during processes such as deformation, crack propagation, and frictional collisions. Its signals contain rich mechanical dynamic information, are highly sensitive to early-stage minor faults, and are not strongly affected by steady-state vibrations, making it suitable for condition monitoring of complex mechanical systems. However, when applying acoustic emission detection technology to printing press fault diagnosis, the following technical bottlenecks still exist:

[0005] Challenges in signal acquisition and preprocessing: During printing press operation, paper-to-roll friction, ink transfer, and motor operation generate a large amount of background noise, resulting in a low signal-to-noise ratio of acoustic emission signals. Furthermore, the acoustic emission characteristic frequencies of different fault types overlap. Traditional fixed threshold noise reduction methods easily filter out key impact features, and single-domain feature extraction is difficult to fully characterize fault information, affecting the accuracy of subsequent diagnosis.

[0006] Insufficient fault location and multi-feature fusion: The key components of the printing press are densely distributed, and the acoustic emission signal is attenuated and reflected when it propagates in the structure, making it difficult for a single sensor to determine the fault location; existing diagnostic models mostly use single feature input or simple feature splicing, without considering the differences in importance of features of different dimensions (time domain, frequency domain, time-frequency domain) and the spatial correlation of sensor locations, resulting in poor adaptability of the model to complex multi-fault scenarios, and easy to misjudge the same symptoms but different faults or the same fault but different symptoms.

[0007] Limitations in sample library construction and model generalization ability: Printing press failures are diverse and sporadic, and historical failure data are often scarce and uneven in type. Diagnostic models trained solely on actual failure data are prone to overfitting. At the same time, existing research on failure simulations mostly targets single components or fixed failure degrees, lacking sample coverage of the entire failure development process. This results in the model's insufficient ability to identify unseen failure types or minor failures in actual production, making it difficult to meet the needs of engineering applications. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a printing press fault diagnosis method based on acoustic emission detection, which solves the problems mentioned in the background section.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a printing press fault diagnosis method based on acoustic emission detection, comprising the following steps:

[0010] A printing press acoustic emission signal acquisition system is constructed. The acquisition system includes multiple acoustic emission sensors, a signal preprocessing module, and a data transmission unit. The multiple acoustic emission sensors are distributed in key transmission components, printing execution components, and support structures of the printing press to synchronously acquire raw acoustic emission signals from different locations.

[0011] The acquired acoustic emission raw signal is preprocessed in multiple dimensions. The preprocessing includes noise reduction based on adaptive wavelet threshold, signal decomposition based on empirical mode decomposition, and extraction of time-domain features, frequency-domain features, and time-frequency-domain features of the signal to form a multi-feature set.

[0012] A fault diagnosis model with an integrated attention mechanism is constructed. The model takes the obtained multi-feature set as input, assigns dynamic weights to features of different dimensions through a feature attention module, assigns position weights to signal features collected by different sensors through a spatial attention module, and then performs feature fusion and fault classification through a deep learning network.

[0013] A fault sample library is constructed based on historical fault data and simulated fault test data of printing presses. The sample library contains labeled samples under normal operating conditions and various typical fault conditions. The constructed fault diagnosis model is trained and optimized using the sample library.

[0014] The acoustic emission signal, which is collected and preprocessed in real time, is input into the trained fault diagnosis model, which outputs the fault type and fault location information of the printing press and generates a fault warning signal.

[0015] Preferably, the key transmission components include a gearbox, a transmission shaft, and a bearing assembly; the printing execution components include a printing plate cylinder, a rubber cylinder, and an impression cylinder; and the acoustic emission sensors are arranged such that at least one sensor is provided in the radial and axial directions of each key transmission component, at least one sensor is provided at each of the bearing seats at both ends of the printing execution components, and the signal coverage ranges of adjacent sensors have a preset degree of overlap.

[0016] Preferably, the noise reduction processing of the adaptive wavelet threshold is specifically as follows: the wavelet decomposition layer and threshold function are dynamically adjusted by analyzing the noise characteristics of the original acoustic emission signal. The threshold function adopts an adaptive threshold based on the signal kurtosis value to preserve the impulse feature components in the signal.

[0017] Preferably, the time-domain features include peak value, kurtosis, impulse exponent, and root mean square value; the frequency-domain features include center frequency, frequency variance, and power spectrum peak value; and the time-frequency-domain features include wavelet packet energy entropy, Hilbert marginal spectrum features, and high-frequency energy proportion of short-time Fourier transform.

[0018] Preferably, the fault diagnosis model with the fusion attention mechanism includes an input layer, a feature attention layer, a spatial attention layer, a feature fusion layer, and a classification output layer connected in sequence; the feature attention layer learns the importance weights of different feature dimensions through a fully connected network and performs weighted summation; the spatial attention layer extracts sensor location-related features through convolution operations and generates a location weight matrix.

[0019] Preferably, the construction of the fault sample library further includes: introducing a controllable fault simulation device under normal printing press operation to generate fault signals of different degrees and types. The controllable fault simulation device can simulate gear wear, bearing outer ring peeling, roller imbalance and transmission belt slippage faults, and each fault type contains sample data of multiple fault development stages.

[0020] Preferably, the method also includes verifying and optimizing the output of the fault diagnosis model. Specifically, auxiliary verification data is obtained through vibration signal detection, visual detection and manual inspection. The auxiliary verification data is compared with the model output. If there is a deviation, the fault sample library is expanded based on the deviation data, and the fault diagnosis model is retrained and optimized.

[0021] Preferably, the fault types include transmission system faults, printing roller faults, paper feeding system faults, and control system faults. The fault location information is determined by fusing sensor spatial distribution information with the location probability distribution output by the model. When the fault probability exceeds a preset threshold, a fault warning signal is triggered. The warning signal includes the fault type, location, and estimated severity.

[0022] Preferably, in the residual network structure used in the feature fusion layer, each residual block contains two convolutional layers and one skip connection. The kernel size of the convolutional layers is dynamically adjusted based on the frequency characteristics of the acoustic emission signal. An attention gating mechanism is introduced in the skip connection to enhance the transmission efficiency of key features.

[0023] This invention provides a method for diagnosing printing press faults based on acoustic emission detection. It has the following beneficial effects:

[0024] 1. This invention synchronously collects acoustic emission signals from key components of a printing press using distributed acoustic emission sensors. It combines multi-dimensional preprocessing to extract time-domain, frequency-domain, and time-frequency-domain features, and dynamically assigns feature and location weights using a fault diagnosis model with an attention fusion mechanism. This achieves accurate identification of printing press fault types and locations. Simultaneously, by constructing a sample library containing historical and simulated fault data and introducing a verification feedback mechanism, the generalization ability and diagnostic reliability of the model are effectively improved. This solves the problems of high signal interference, incomplete feature extraction, and low positioning accuracy in traditional printing press fault diagnosis, providing accurate fault warning and diagnostic support for efficient operation and maintenance of printing presses.

[0025] 2. This invention employs a preprocessing method that combines adaptive wavelet threshold denoising with empirical mode decomposition, which can selectively retain fault impact features in acoustic emission signals and avoid the loss of useful information. The deep learning model that integrates attention mechanisms enhances the extraction of key fault features and the utilization of sensor location correlation information through dual weighting of feature attention and spatial attention. Compared with traditional diagnostic methods, it is more adaptable to the multi-fault type identification needs in the complex operating environment of printing presses, significantly improving the timeliness and accuracy of fault diagnosis and reducing production losses caused by printing press downtime due to faults. Attached Figure Description

[0026] Figure 1 This is the overall flowchart of the present invention;

[0027] Figure 2 This is a flowchart of the signal processing and feature extraction process of the present invention;

[0028] Figure 3 This is a flowchart of the model training and optimization process of the present invention. Detailed Implementation

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Please see the appendix Figure 1 - Appendix Figure 3 This invention provides a printing press fault diagnosis method based on acoustic emission detection, comprising the following steps:

[0031] An acoustic emission signal acquisition system for a printing press is constructed. The acquisition system includes multiple acoustic emission sensors, a signal preprocessing module, and a data transmission unit. The multiple acoustic emission sensors are distributed at key transmission components, printing execution components, and support structures of the printing press to synchronously acquire raw acoustic emission signals from different locations. The key transmission components include gearboxes, drive shafts, and bearing assemblies. The printing execution components include printing plate cylinders, rubber cylinders, and impression cylinders. The arrangement of the acoustic emission sensors satisfies the following conditions: at least one sensor is installed in the radial and axial directions of each key transmission component, and at least one sensor is installed at each of the two bearing seats at both ends of the printing execution component. Furthermore, the signal coverage areas of adjacent sensors have a predetermined degree of overlap.

[0032] Specifically, the core function of the printing press acoustic emission signal acquisition system is to accurately and synchronously capture the raw acoustic emission signals of key components of the equipment, providing high-quality data support for subsequent fault diagnosis. Its design principles and structural layout are closely based on the mechanical characteristics of the printing press and the propagation patterns of fault signals. From a system composition perspective, multiple acoustic emission sensors, signal preprocessing modules, and data transmission units work collaboratively. The sensors, as the signal acquisition front end, need to be distributed according to the fault sensitivity characteristics of different functional components of the printing press. Key transmission components (gearbox, drive shaft, bearing assembly) are high-risk areas for mechanical wear and fatigue failures, and their radial and axial vibrations and stress wave propagation differ. Therefore, at least one sensor is installed in both the radial and axial directions of each key transmission component to comprehensively capture acoustic emission signals generated by component meshing and rotation in different directions. Faults in printing execution components (plate cylinder, rubber cylinder, impression cylinder) are often related to bearing failure and cylinder imbalance. Since the bearing housing is the main propagation node for acoustic emission signals, sensors are installed at both ends of the bearing housing to directly acquire abnormal signals during cylinder operation. Meanwhile, the signal coverage areas of adjacent sensors are pre-defined with a certain overlap to avoid blind spots caused by signal attenuation or structural obstruction, ensuring that no critical areas of the equipment are missed. The signal preprocessing module can perform preliminary conditioning (such as amplification and filtering) on ​​the raw signals collected by the sensors to reduce transmission interference. The data transmission unit realizes the synchronous transmission and aggregation of signals from multiple sensors, laying the foundation for subsequent multi-dimensional preprocessing and fault analysis. Through targeted component coverage and layout optimization, the overall system achieves comprehensive, synchronous, and high-quality acquisition of acoustic emission signals from the printing press.

[0033] The acquired acoustic emission raw signal undergoes multi-dimensional preprocessing, including noise reduction based on adaptive wavelet thresholding, signal decomposition based on empirical mode decomposition, and extraction of time-domain, frequency-domain, and time-frequency-domain features to form a multi-feature set. The time-domain features include peak value, kurtosis, impulse exponent, and root mean square value; the frequency-domain features include center frequency, frequency variance, and power spectrum peak value; and the time-frequency-domain features include wavelet packet energy entropy, Hilbert marginal spectrum features, and high-frequency energy proportion of short-time Fourier transform.

[0034] Specifically, the noise reduction processing based on adaptive wavelet thresholding differs from traditional fixed thresholding methods. It dynamically adjusts the wavelet decomposition level and threshold function by analyzing the noise characteristics of the original signal in real time. The threshold function uses an adaptive threshold based on the signal kurtosis value. The kurtosis value is sensitive to the impact components in the signal; determining the threshold accordingly can filter out background noise such as paper friction and motor operation while preserving the impact characteristics generated by faults such as gear wear and bearing peeling to the greatest extent possible, avoiding the loss of useful signals. The signal decomposition processing based on empirical mode decomposition can decompose the non-stationary and nonlinear acoustic emission signal into multiple intrinsic mode functions, each corresponding to signal components at different frequency scales, thereby separating the fault-related characteristic frequency band signals and reducing signal complexity. After noise reduction and decomposition, three types of features—time domain, frequency domain, and time-frequency domain—are further extracted to form a multi-feature set: Time domain features directly reflect the amplitude changes and statistical characteristics of the signal in the time dimension, and can quickly capture signal mutations caused by faults; Frequency domain features transform the signal to the frequency domain through Fourier transform, revealing the frequency distribution law of fault features, such as gear faults often accompanied by sidebands of specific frequencies; Time-frequency domain features combine time and frequency dimension information, solving the problem that single-domain features cannot fully characterize non-stationary fault signals. For example, wavelet packet energy entropy can reflect the uniformity of signal energy distribution in different time-frequency units, and Hilbert marginal spectrum can accurately extract the instantaneous frequency features of fault signals. The three types of features complement and fuse to completely cover the multi-dimensional information of fault signals, providing sufficient basis for fault type and location identification.

[0035] The noise reduction process using adaptive wavelet thresholding is as follows: by analyzing the noise characteristics of the original acoustic emission signal, the wavelet decomposition level and threshold function are dynamically adjusted. The threshold function adopts an adaptive threshold based on the signal kurtosis value to preserve the impulse feature components in the signal.

[0036] A fault diagnosis model with a fusion attention mechanism is constructed. The model takes the obtained multi-feature set as input, assigns dynamic weights to features of different dimensions through a feature attention module, and assigns position weights to signal features collected by different sensors through a spatial attention module. Then, a deep learning network is used for feature fusion and fault classification. The fault diagnosis model with a fusion attention mechanism includes an input layer, a feature attention layer, a spatial attention layer, a feature fusion layer, and a classification output layer connected in sequence. The feature attention layer learns the importance weights of different feature dimensions through a fully connected network and performs weighted summation. The spatial attention layer extracts sensor position-related features through convolution operations and generates a position weight matrix.

[0037] Specifically, the original acoustic emission signal from a printing press is mixed with background noise such as paper friction and motor operation. Furthermore, the noise intensity and frequency distribution vary under different operating conditions. Traditional fixed-threshold noise reduction methods often filter out critical fault signals or leave excessive noise. Therefore, this method first analyzes the noise characteristics of the original signal and dynamically adjusts the number of wavelet decomposition layers accordingly. Too few decomposition layers lead to incomplete noise separation, while too many may damage signal integrity. Simultaneously, an adaptive threshold function based on signal kurtosis is employed. Since kurtosis is highly sensitive to impulsive components in the signal, determining the threshold based on kurtosis effectively filters out stable background noise while maximizing the retention of impulsive feature components crucial for fault diagnosis, laying a high-quality data foundation for subsequent feature extraction. The fault diagnosis model incorporating an attention mechanism is designed around accurately focusing on key information. Using multiple feature sets as input, it optimizes the utilization of features and spatial information through a hierarchical structure: after the input layer receives multiple feature sets, the feature attention layer learns the importance weights of different feature dimensions through a fully connected network.

[0038] A fault sample library was constructed based on historical fault data and simulated fault test data of the printing press. The sample library contains labeled samples under normal operating conditions and various typical fault conditions. The fault diagnosis model was trained and optimized using the sample library. The construction of the fault sample library also includes: introducing a controllable fault simulation device under normal operating conditions of the printing press to generate fault signals of different degrees and types. The controllable fault simulation device can simulate gear wear, bearing outer ring peeling, roller imbalance and transmission belt slippage faults, and each fault type contains sample data of multiple fault development stages.

[0039] Specifically, the sample library needs to cover labeled samples of both normal equipment operation and various typical fault states. Normal state samples provide the model with fault-free baseline features, while fault state samples contain feature information of different fault types. The combination of the two allows the model to learn the feature difference boundary between normal and fault states. Among these, the introduction of a controllable fault simulation device is a key design. Since actual printing press faults are difficult to trigger actively and their severity is uncontrollable, this device can accurately simulate common faults such as gear wear, bearing outer ring peeling, roller imbalance, and drive belt slippage during normal equipment operation through human intervention. It can also control the development process of faults from minor to severe, thereby generating sample data of different degrees and types covering multiple fault development stages. This compensates for the lack of minor fault and new fault samples in historical data, ensuring the completeness and gradient of the sample library.

[0040] The acoustic emission signal, which is collected and preprocessed in real time, is input into the trained fault diagnosis model, which outputs the fault type and fault location information of the printing press and generates a fault warning signal.

[0041] It also includes verifying and optimizing the output of the fault diagnosis model. Specifically, auxiliary verification data is obtained through vibration signal detection, visual inspection and manual inspection. The auxiliary verification data is compared with the model output. If there is a deviation, the fault sample library is expanded based on the deviation data and the fault diagnosis model is retrained and optimized.

[0042] Specifically, the signal, after multi-dimensional preprocessing, has removed redundant noise and retained key fault features. The model can quickly match the correspondence between signal features and fault types based on the feature attention weights and spatial attention weights learned during the training phase. At the same time, by combining the spatial distribution information of the sensor and the position probability distribution of the model output, the fault location can be accurately determined. When the fault probability exceeds a preset threshold, an early warning signal containing the fault type, location, and estimated severity is generated, realizing a real-time response from signal input to fault warning, and meeting the needs of timely fault handling in continuous production of printing presses.

[0043] The fault types include transmission system faults, printing roller faults, paper feeding system faults, and control system faults. The fault location information is determined by fusing sensor spatial distribution information with the location probability distribution output by the model. When the fault probability exceeds a preset threshold, a fault warning signal is triggered. The warning signal includes the fault type, location, and estimated severity.

[0044] Specifically, the determination of fault location information relies on the fusion logic of sensor spatial distribution and model location probability distribution. On the one hand, based on the distribution characteristics of key components of the printing press, acoustic emission sensors have been distributed in key transmission components, printing execution components, and other locations. The monitoring range and spatial coordinates of each sensor are clear, allowing for preliminary identification of the signal source area. On the other hand, the fault diagnosis model, which integrates an attention mechanism, outputs fault probability distributions for different locations based on the characteristic differences of signals collected by each sensor. By fusing these two methods, the positioning deviation caused by signal attenuation and interference from a single sensor can be effectively compensated, improving the accuracy of location determination. The triggering of fault warning signals is based on a preset fault probability threshold. This threshold is set based on historical fault data of the printing press and safe operation requirements. When the probability of a certain type of fault output by the model exceeds this threshold, the system automatically triggers a warning. The warning signal includes the fault type, location, and estimated severity. The estimated severity is calculated comprehensively using parameters such as the intensity and duration of the fault characteristic signal. This helps maintenance personnel quickly determine the urgency of the fault, prioritize high-risk faults, and minimize equipment downtime losses and safety hazards.

[0045] In the residual network structure used in the feature fusion layer, each residual block contains two convolutional layers and one skip connection. The kernel size of the convolutional layers is dynamically adjusted based on the frequency characteristics of the acoustic emission signal. An attention gating mechanism is introduced in the skip connection to enhance the transmission efficiency of key features.

[0046] Specifically, each residual block contains two convolutional layers responsible for feature extraction and transformation. The kernel size is not fixed but dynamically adjusted based on the frequency characteristics of the acoustic emission signal. Since the frequency distribution of acoustic emission signals generated by different printing press faults varies significantly, dynamically adjusting the kernel size allows the convolutional layers to more accurately match features in different frequency ranges, improving the ability to extract features from various faults. Simultaneously, the skip connections in the residual block are not simple signal transmissions but incorporate an attention-gated mechanism. This mechanism learns to judge the importance of features, assigning higher transmission weights to key fault features and suppressing redundant or interfering features, thereby significantly enhancing the transmission efficiency of key features in the deep network and avoiding the attenuation of useful features due to increased network layers. The overall structure, through the synergy of dynamic kernel adaptation and attention-gated skip connections, achieves efficient fusion of multi-dimensional acoustic emission fault features from the printing press, laying the foundation for accurate subsequent fault classification.

[0047] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A printing press fault diagnosis method based on acoustic emission detection, characterized in that, Includes the following steps: A printing press acoustic emission signal acquisition system is constructed. The acquisition system includes multiple acoustic emission sensors, a signal preprocessing module, and a data transmission unit. The multiple acoustic emission sensors are distributed in key transmission components, printing execution components, and support structures of the printing press to synchronously acquire raw acoustic emission signals from different locations. The acquired acoustic emission raw signal is preprocessed in multiple dimensions. The preprocessing includes noise reduction based on adaptive wavelet threshold, signal decomposition based on empirical mode decomposition, and extraction of time-domain features, frequency-domain features, and time-frequency-domain features of the signal to form a multi-feature set. A fault diagnosis model with an integrated attention mechanism is constructed. The model takes the obtained multi-feature set as input, assigns dynamic weights to features of different dimensions through a feature attention module, assigns position weights to signal features collected by different sensors through a spatial attention module, and then performs feature fusion and fault classification through a deep learning network. A fault sample library is constructed based on historical fault data and simulated fault test data of printing presses. The sample library contains labeled samples under normal operating conditions and various typical fault conditions. The constructed fault diagnosis model is trained and optimized using the sample library. The acoustic emission signal, which is collected and preprocessed in real time, is input into the trained fault diagnosis model, which outputs the fault type and fault location information of the printing press and generates a fault warning signal.

2. The printing press fault diagnosis method based on acoustic emission detection according to claim 1, characterized in that, The key transmission components include a gearbox, a drive shaft, and a bearing assembly. The printing execution components include a printing plate cylinder, a rubber cylinder, and an impression cylinder. The acoustic emission sensors are arranged to satisfy the following conditions: at least one sensor is provided in the radial and axial directions of each key transmission component, at least one sensor is provided at each of the bearing seats at both ends of the printing execution component, and the signal coverage range of adjacent sensors has a preset overlap.

3. The printing press fault diagnosis method based on acoustic emission detection according to claim 1, characterized in that, The noise reduction process of the adaptive wavelet threshold is specifically as follows: the wavelet decomposition level and threshold function are dynamically adjusted by analyzing the noise characteristics of the original acoustic emission signal. The threshold function adopts an adaptive threshold based on the signal kurtosis value to preserve the impulse feature components in the signal.

4. The printing press fault diagnosis method based on acoustic emission detection according to claim 1, characterized in that, The time-domain features include peak value, kurtosis, impulse exponent, and root mean square value; the frequency-domain features include center frequency, frequency variance, and power spectrum peak value; and the time-frequency-domain features include wavelet packet energy entropy, Hilbert marginal spectrum features, and high-frequency energy proportion of short-time Fourier transform.

5. The printing press fault diagnosis method based on acoustic emission detection according to claim 1, characterized in that, The fault diagnosis model based on the fusion attention mechanism includes an input layer, a feature attention layer, a spatial attention layer, a feature fusion layer, and a classification output layer connected in sequence. The feature attention layer learns the importance weights of different feature dimensions through a fully connected network and performs a weighted summation. The spatial attention layer extracts sensor location-related features through convolution operations and generates a location weight matrix.

6. The printing press fault diagnosis method based on acoustic emission detection according to claim 1, characterized in that: The construction of the fault sample library also includes: introducing a controllable fault simulation device under normal printing press operation to generate fault signals of different degrees and types. The controllable fault simulation device can simulate gear wear, bearing outer ring peeling, roller imbalance and transmission belt slippage faults, and each fault type contains sample data of multiple fault development stages.

7. The printing press fault diagnosis method based on acoustic emission detection according to claim 1, characterized in that, It also includes verifying and optimizing the output of the fault diagnosis model. Specifically, auxiliary verification data is obtained through vibration signal detection, visual inspection and manual inspection. The auxiliary verification data is compared with the model output. If there is a deviation, the fault sample library is expanded based on the deviation data and the fault diagnosis model is retrained and optimized.

8. The printing press fault diagnosis method based on acoustic emission detection according to claim 1, characterized in that, The fault types include transmission system faults, printing roller faults, paper feeding system faults, and control system faults. The fault location information is determined by fusing sensor spatial distribution information with the location probability distribution output by the model. When the fault probability exceeds a preset threshold, a fault warning signal is triggered. The warning signal includes the fault type, location, and estimated severity.

9. A printing press fault diagnosis method based on acoustic emission detection according to claim 1, characterized in that, In the residual network structure used in the feature fusion layer, each residual block contains two convolutional layers and one skip connection. The kernel size of the convolutional layers is dynamically adjusted based on the frequency characteristics of the acoustic emission signal. An attention gating mechanism is introduced in the skip connection to improve the transmission efficiency of key features.

Images