Acoustic emission on-line monitoring method for early fault diagnosis of large press gearbox

Abstract

The invention discloses an acoustic emission online monitoring method for early fault diagnosis of a large press gearbox, and relates to the technical field of industrial equipment fault diagnosis, and the method comprises the steps: obtaining a monitoring reference parameter according to a read mold number; the collected data is separated into periodic and impact components by using an extreme value driven sliding time window algorithm; decomposing the signal through dual-tree complex wavelet packet transformation and extracting a standardized multi-dimensional feature vector; inputting the feature vector into a multi-model fusion discrimination system to obtain a result; and calculating the position of a fault source based on the signal strength sequence and the monitoring time difference when the fault is judged. According to the method, the corresponding monitoring reference parameters are retrieved and acquired according to the mold numbers, and the monitoring data are separated into the periodic components and the impact components by using the extreme value driven sliding time window algorithm, so that the problem that a fixed threshold cannot adapt to variable loads is solved, and the sensitivity of capturing early fault signals of the gearbox is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment fault diagnosis, and particularly relates to an acoustic emission online monitoring method for early fault diagnosis of a large press gear box. BACKGROUND

[0002] A large press is a core equipment for stamping forming of a cover part and a vehicle body structure part in the automobile manufacturing industry, and the running state of a gear box of the large press is directly related to transmission stability and working reliability of the whole machine. Under the combined action of continuous operation and impact load, components such as rolling bearings and meshing gears inside the gear box are prone to local fatigue, and then cracks or abnormal wear are caused. Since stamping is the first process of an automatic production line, unplanned downtime caused by gear box failure will affect subsequent processes such as welding, painting and assembly, increase the maintenance cost of parts and cause production interruption. The traditional offline detection method relies on manual inspection using an endoscope and other equipment under the condition of shutdown, which has the problems of limited detection field of view, complex operation and subjective evaluation, and cannot meet the real-time sensing demand of equipment state in modern production. Therefore, online monitoring by using sensors to collect physical characteristics and combining signal processing technology has become a main means to ensure the safe operation of the large press.

[0003] At present, vibration monitoring is a common method for gear box fault diagnosis, which captures the characteristic frequency caused by component damage for judgment. However, the large press has the characteristics of low speed and heavy load, and the rotating speed of the shaft system inside is usually low, which makes the fault characteristic frequency of the rolling bearing and the gear in the low frequency band, and it is easy to be disturbed by the low frequency sensitivity of the sensor and the environmental noise. At the same time, in order to accurately capture low frequency vibration information, a long sampling time window is usually needed, but the actual working condition of the large press contains alternating periodic components of idling and impact components of doing work, and the long sampling window will cause signal aliasing under different working conditions, which is difficult to realize effective separation and feature extraction through a single vibration analysis method. In contrast, the acoustic emission technology focuses on the high-frequency elastic waves released by local damage of materials, which can effectively avoid low-frequency noise interference and has physical advantages in early fault perception.

[0004] Although acoustic emission technology has been applied in the diagnosis of conventional rotating machinery, there are still technical limitations in its direct application to large press gearboxes. First, existing monitoring methods usually set a fixed alarm threshold, while large presses frequently change molds during production, and different molds correspond to different load references and acoustic emission responses. The fixed threshold is easy to cause false alarms. Second, traditional data-driven diagnostic models often rely on sufficient fault samples for supervised training, but industrial sites lack complete fault label data, and it is difficult to build accurate discrimination boundaries relying only on normal samples. In addition, conventional signal decomposition methods lack shift invariance when dealing with non-stationary impact signals, resulting in insufficient stability of the extracted fault features. Therefore, how to establish an acoustic emission online monitoring method that can adapt to mold changing conditions, use normal data for adaptive discrimination due to the lack of fault samples, and accurately locate the fault source is a problem that needs to be solved in this field. SUMMARY

[0005] The purpose of the present application is to provide an acoustic emission online monitoring method for early fault diagnosis of large press gearboxes, which at least solves one of the technical problems in the prior art that the fixed alarm threshold cannot adapt to the frequent mold changes of large presses, resulting in changes in working conditions and loads, and that it is difficult to build an accurate discrimination model due to the lack of complete fault sample data in industrial sites.

[0006] The present application provides an acoustic emission online monitoring method for early fault diagnosis of large press gearboxes, which adopts the following technical solution:

[0007] An acoustic emission online monitoring method for early fault diagnosis of large press gearboxes, comprising the following steps:

[0008] S10, using the industrial control computer of the acoustic emission online monitoring system to read the clutch state and mold number of the PLC controller of the large press, retrieving the historical monitoring database in the local storage unit according to the mold number to obtain the monitoring reference parameters corresponding to the current mold;

[0009] S20, when the clutch state indicates engagement, collecting the monitoring data of each acoustic emission sensor arranged on the gearbox of the large press, processing the monitoring data using an extreme value driven sliding time window algorithm to separate the periodic component and the impact component;

[0010] S30, using dual-tree complex wavelet packet transform to decompose the periodic component and the impact component respectively to obtain signal components in different frequency bands;

[0011] S40, for each signal component, a multi-dimensional feature vector containing waveform features, time domain features, frequency domain features, envelope features and inter-band features is extracted, and Z-score standardization processing is performed on the monitoring reference parameters to generate a standardized multi-dimensional feature vector;

[0012] S50, the standardized multi-dimensional feature vector is input into a multi-model fusion discrimination system containing Lyapunov rule, one-class support vector machine and isolated forest, single discrimination results output by each model are obtained, voting statistics are performed on the single discrimination results, and a final fault discrimination result is determined according to the result of the voting statistics;

[0013] S60, when the fault discrimination result indicates that the gearbox exists a fault, the position of the fault source is calculated based on the intensity ranking and monitoring time difference of the multi-point fault characteristic acoustic emission signal, and a diagnostic result containing fault position information is output.

[0014] By adopting the above technical scheme, the application establishes a self-adaptive monitoring system oriented to the mold, which can adapt to the characteristics of complex working conditions and variable load of large presses. The monitoring system reads the mold number and calls the corresponding monitoring reference parameters, eliminating the influence of the differences in gearbox vibration and acoustic emission response reference under different mold working conditions, thereby reducing the false alarm rate. At the same time, the signal separation technology is used to decouple the periodic components representing the mechanical action rhythm and the impact components representing the component damage, and the multi-domain features are extracted by combining the dual-tree complex wavelet packet transform, which overcomes the defect that a single feature is not sensitive to early weak faults. In addition, the combination of the unsupervised discrimination strategy of multi-model fusion and the positioning method based on intensity correction can realize abnormal identification and accurate positioning of the fault source in the absence of fault sample training, thereby improving the accuracy and robustness of early fault diagnosis of the gearbox of the large press.

[0015] Further, in the S10 step, acquiring the monitoring reference parameters corresponding to the current mold specifically includes:

[0016] The acoustic emission online monitoring system searches in the local storage unit according to the acquired mold number;

[0017] If the corresponding historical monitoring database exists, the feature distribution parameters and discrimination model parameters corresponding to the mold number are called as the monitoring reference parameters;

[0018] If the corresponding historical monitoring database does not exist, a new database entry is created, and a self-learning mode is entered;

[0019] In the self-learning mode, acoustic emission raw signals of a set number of stamping strokes are continuously collected, multi-domain feature vectors of reference periodic components and reference impact components are extracted as an initial sample set, stored in the new database entry, and the monitoring reference parameters are calculated based on the initial sample set.

[0020] By adopting the technical solution, the system has the working condition self-adaptive capability. For a new die or a device after major repair, the system can automatically construct a health reference through the self-learning mode without manual recalibration of thresholds, thereby reducing the complexity of system deployment and maintenance and ensuring monitoring continuity in a multi-variety and small-batch production mode.

[0021] Further, the S10 step further comprises performing a database update based on a sliding window:

[0022] A stroke count window of a fixed length is set, and feature data judged as fault-free is continuously recorded by the acoustic emission online monitoring system;

[0023] When the acoustic emission online monitoring system does not have any alarm in a set update period and all discrimination results are normal, the latest fault-free feature data collected in the current update period is added to the sample set while the oldest data in the sample set is removed according to the first-in first-out principle;

[0024] The acoustic emission online monitoring system recalculates statistical parameters of features based on the updated sample set and incrementally trains or re-trains the fault discrimination model.

[0025] By adopting the technical solution, the system can dynamically adapt to signal reference drift caused by running-in state change, lubricating oil temperature change and mechanical aging of the gearbox due to long-term operation. By continuously introducing the latest health data and eliminating old data, the discrimination model maintains the best fitting degree to the current device state, suppresses false alarms caused by time-varying characteristics, and ensures monitoring stability throughout the life cycle.

[0026] Further, the sliding time window algorithm driven by extreme values to separate the periodic components and the impact components specifically comprises: selecting monitoring data of the acoustic emission sensor arranged near the idler shaft as a reference signal, and setting the length of the initial sampling time window as twice the standard production beat time.

[0027] The reference signal is subjected to Hilbert transform to construct an analytical signal and calculate an envelope, and a low-pass filter is applied to the envelope for smoothing processing.

[0028] In the initial sampling time window of the smoothed envelope, four extreme points with the largest amplitude are searched and labeled in chronological order.

[0029] Selecting a second extreme point and a third extreme point located in the middle position, finding a minimum point of the envelope in a time interval of the second extreme point and the third extreme point, and determining a time corresponding to the minimum point as a truncation point time.

[0030] Taking the truncation point time as a boundary, defining a signal segment before the truncation point time as the impact component, and defining a low-energy smooth segment near the truncation point time as the periodic component.

[0031] By adopting the technical scheme, the signal envelope is extracted by using the Hilbert transform, and the physical segmentation of the non-stationary acoustic emission signal is realized in combination with the physical extreme value characteristics of the press stamping cycle. The impact segment containing the main working information and the periodic segment containing the background transmission noise can be accurately separated by using the method, signal truncation or aliasing caused by fixed time window segmentation is avoided, and a pure data source is provided for subsequent targeted extraction of fault features.

[0032] Further, the obtaining of the signal components in different frequency bands specifically includes: constructing a dual-tree complex filter bank structure including a real part tree and an imaginary part tree; in first layer decomposition, a delay of one sampling point is maintained between filters of the real part tree and the imaginary part tree; in second layer and subsequent layer decomposition, a delay of half a sampling point is maintained between filter banks of the real part tree and the imaginary part tree.

[0033] Setting a decomposition layer number, iteratively decomposing the periodic component and the impact component to obtain real part wavelet packet coefficients and imaginary part wavelet packet coefficients.

[0034] Performing a single-branch reconstruction operation, outputting a real part tree reconstructed signal and an imaginary part tree reconstructed signal by using a comprehensive filter bank, and calculating a comprehensive response of the real part tree reconstructed signal and the imaginary part tree reconstructed signal as a smooth envelope signal.

[0035] By adopting the technical scheme, the approximate shift invariance of the dual-tree complex wavelet packet transform is used to solve the problem of unstable energy distribution of decomposition coefficients when a signal is slightly shifted in the traditional discrete wavelet transform. By calculating the comprehensive response of the real part tree reconstructed signal and the imaginary part tree reconstructed signal, the transient impact energy caused by early failure of a gear or a bearing can be more stably captured, the feature extraction error caused by different sampling phases is reduced, and the robustness of the feature is improved.

[0036] Further, in the S40, the generating of the standardized multi-dimensional feature vector specifically includes:

[0037] The inter-band feature specifically includes energy entropy, singularity entropy, dispersion entropy, conditional entropy, and cross entropy.

[0038] The Z-score standardization method is used to perform normalization calculation on each feature element constituting the multi-dimensional feature vector; the mean and standard deviation used in the normalization calculation are included in the monitoring reference parameters.

[0039] By using the above technical solution, the introduction of multi-class entropy feature quantifies the complexity and disorder degree of signal energy distribution, enhancing the sensitivity to weak damage. At the same time, the mean and standard deviation based on historical mold reference are used for Z-score standardization, which maps features of different physical dimensions to a relative deviation space, eliminating the influence of different mold loads on the absolute amplitude of the features, so that the subsequent discrimination model can focus on identifying abnormal deviation from the reference state, and the discrimination scale under different working conditions is unified.

[0040] Further, the multi-model fusion discrimination system performs the following processing logic:

[0041] Pareto principle discrimination logic: set a discrimination threshold interval, if more than a certain proportion of feature components in the normalized multi-dimensional feature vector fall outside the discrimination threshold interval, it is determined to be abnormal.

[0042] Single-class support vector machine discrimination logic: radial basis function is used as the kernel function, and the optimal hyperplane that separates the origin from the normal samples is determined by solving the dual optimization problem; if the normalized multi-dimensional feature vector falls outside the optimal hyperplane, it is determined to be abnormal.

[0043] Isolation forest discrimination logic: a forest composed of multiple random binary trees is constructed, the average path length of the normalized multi-dimensional feature vector in the tree is calculated and converted into an anomaly score; if the anomaly score is greater than a preset anomaly threshold, it is determined to be abnormal.

[0044] By using the above technical solution, an unsupervised discrimination system combining statistical models, geometric models and density models is constructed. This system does not need to train fault samples, but only uses normal data to construct a health boundary. The three algorithms describe the data distribution from different dimensions: the Pareto principle is sensitive to amplitude mutation, the single-class support vector machine is good at handling high-dimensional nonlinear boundaries, and the isolation forest has high identification ability for sparse outliers. The three complement each other, effectively solving the problem of lack of fault label data in industrial field, and improving the coverage ability for multiple fault modes.

[0045] Further, the fault discrimination result is obtained according to the following logic: the sum of the discrimination results of the Pareto principle discrimination logic, the single-class support vector machine discrimination logic and the isolation forest discrimination logic is calculated; if the sum is greater than or equal to 2, it is determined that a fault exists.

[0046] The standardized multi-dimensional feature vector contains all features extracted for the periodic component and the impact component; a discrimination result of non-fault is generated only when the relevant features of the periodic component and the impact component are all judged as normal; and a discrimination result of fault is generated if the features of any of the periodic component and the impact component show abnormality, resulting in the sum value showing fault.

[0047] By using the above technical solution, the false alarm rate caused by algorithm bias of a single model is reduced by using the majority voting mechanism, and the credibility of the diagnosis result is improved. At the same time, combined with the component separation result for joint discrimination, it is ensured that the abnormality occurring in the working impact stage or the idling stable stage can be effectively captured, avoiding missed judgment, and realizing the health monitoring of the gearbox in the whole cycle.

[0048] Further, in the S60 step, before calculating the fault source position, an equivalent sound speed matrix is further constructed:

[0049] The sound propagation characteristics of the gearbox are calibrated on site by using the Hsu-Nielsen lead breaking sound source experiment method;

[0050] A standard lead breaking excitation is implemented at a preset coordinate position, and the propagation time of sound waves reaching each acoustic emission sensor position is recorded;

[0051] Based on the ratio of the preset known distance to the propagation time, the single-point equivalent sound speed is calculated, and through multiple measurements at multiple points, an equivalent sound speed matrix covering different monitoring areas is constructed.

[0052] Further, in the S60 step, calculating the fault source position specifically includes:

[0053] The signal energy features of the monitoring data of all acoustic emission sensors in the current time window are extracted, and the three acoustic emission sensors with the largest energy values are identified;

[0054] The search range of the fault source is locked in a triangular region with the three acoustic emission sensors as the vertices or the neighborhood of the triangular region;

[0055] For the three selected acoustic emission sensors, the accurate time of arrival of the fault transient impact signal at each acoustic emission sensor is determined, and the arrival time difference between the acoustic emission sensors is calculated;

[0056] According to the corresponding equivalent sound speed of the triangular region matched from the equivalent sound speed matrix, a nonlinear hyperbolic equation set taking the arrival time difference as the observation is established, the Newton-Raphson iteration method is used for numerical solution, and the estimated coordinates of the fault source are obtained.

[0057] By adopting the technical scheme, the region is preliminarily screened based on the energy intensity sorting before positioning calculation, the signal interference caused by the reflection wave and the multipath effect of the sensor far away from the fault source is excluded, and the solution space is reduced. The anisotropy influence of the sound speed caused by the casting structure of the gearbox and the oil film medium is corrected by using the calibrated equivalent sound speed. The position of the fault source is accurately calculated by solving the hyperbolic equation set based on the time difference, and a clear spatial orientation is provided for equipment maintenance.

[0058] Through the above scheme, the following beneficial technical effects are obtained:

[0059] The application retrieves and obtains the corresponding monitoring reference parameters according to the mold number, and separates the monitoring data into periodic components and impact components by using a sliding time window algorithm driven by extreme values, establishes a differentiated monitoring reference for the load characteristics of a large press under different mold working conditions, overcomes the problem that a single fixed threshold value cannot adapt to variable loads, decouples the periodic components representing the mechanical action rhythm from the impact components representing component damage, avoids the masking of weak fault features by strong background noise, and improves the sensitivity of early fault signal capture of the gearbox.

[0060] The application constructs a multi-model fusion discrimination system including the PauTa rule, a single-class support vector machine and an isolated forest, and uses the standardized multi-dimensional feature vectors for voting discrimination, realizes unsupervised adaptive fault diagnosis of only using data in a normal state to construct a health boundary, solves the problem that the model is difficult to train due to the scarcity of fault samples in an industrial field, uses the complementarity of multiple algorithms to jointly determine from the statistical, geometric and density dimensions, reduces the false positive rate and the false negative rate caused by the deviation of a single model, and improves the accuracy of the diagnosis result.

[0061] The application uses a dual-tree complex wavelet packet transform to decompose the periodic components and the impact components to obtain signal components, uses the shift invariance to maintain the stability of the energy distribution of the decomposition coefficients, provides a high-quality data basis for feature extraction, and combines the intensity sorting based on the multi-point fault representation of the acoustic emission signal and the positioning calculation of the monitoring time difference to output the specific fault source position when it is determined that there is a fault, realizes a complete diagnosis process from fault discovery to fault positioning, and provides a clear spatial orientation for targeted maintenance of the equipment. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 It is a structure schematic diagram of a large press gearbox acoustic emission monitoring system according to an embodiment of the application.

[0063] Figure 2 It is a whole flow chart of a large press gearbox early fault diagnosis acoustic emission online monitoring method according to an embodiment of the application. It is a whole flow chart of a large press gearbox early fault diagnosis acoustic emission online monitoring method according to an embodiment of the application.

[0064] Figure 3 This is a flowchart illustrating a signal component separation method based on an extreme value-driven sliding time window according to an embodiment of the present invention.

[0065] Figure 4 This is a schematic diagram illustrating the anti-interference signal decomposition principle based on dual-tree complex wavelet packet transform according to an embodiment of the present invention.

[0066] Figure 5 This is a schematic diagram illustrating the principle of a multi-domain feature extraction and adaptive fault discrimination system according to an embodiment of the present invention.

[0067] Figure 6 This is a flowchart illustrating a fault source localization method based on intensity ranking and sound velocity correction according to an embodiment of the present invention.

[0068] Among them, 1. Rear bearing of intermediate shaft A; 2. Rear gear of intermediate shaft A; 3. Intermediate shaft A; 4. High-speed shaft; 5. Front gear of intermediate shaft A; 6. Front bearing of intermediate shaft A; 7. Large gear of intermediate shaft A; 8. Intermediate bearing of high-speed shaft; 9. High-speed shaft gear; 10. Front bearing of high-speed shaft; 11. Flywheel system; 12. Idler gear; 13. Idler gear shaft bearing; 14. Idler gear shaft; 15. Large gear of intermediate shaft B; 16. Front gear of intermediate shaft B; 17. Front bearing of intermediate shaft B; 18. Gearbox base and bearing housing; 19. Rear bearing of intermediate shaft B; 20. Rear gear of intermediate shaft B; 21. Intermediate shaft B; 22. Acoustic emission sensor; 23. Rear bearing of high-speed shaft. Detailed Implementation

[0069] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0070] See attached document Figure 1 This invention provides an online acoustic emission monitoring system for early fault diagnosis of gearboxes in large presses.

[0071] The transmission structure of the large press gear box contains a power input part and four main shaft systems. The power input part is a flywheel system 11, which releases stored energy to drive the high-speed shaft system to rotate when the clutch is in the engaged state. The high-speed shaft system contains a high-speed shaft 4, a high-speed shaft front bearing 10, a high-speed shaft rear bearing 23, a high-speed shaft intermediate bearing 8, and a high-speed shaft gear 9. One side of the high-speed shaft system directly drives the intermediate shaft A system through meshing, and the other side drives the intermediate shaft B system through the idler shaft system. The idler shaft system contains an idler shaft 14, an idler shaft bearing 13, and an idler 12. The intermediate shaft A system contains an intermediate shaft A 3, an intermediate shaft A large gear 7, an intermediate shaft A front gear 5, an intermediate shaft A rear gear 2, an intermediate shaft A front bearing 6, and an intermediate shaft A rear bearing 1. The intermediate shaft B system contains an intermediate shaft B 21, an intermediate shaft B large gear 15, an intermediate shaft B front gear 16, an intermediate shaft B rear gear 20, an intermediate shaft B front bearing 17, and an intermediate shaft B rear bearing 19.

[0072] The monitoring system is configured with 8 acoustic emission sensors 22. These acoustic emission sensors 22 are arranged on the key nodes of the gear box base and bearing seat 18 respectively. The arrangement position follows the principle of close to the rotating mechanism, that is, the sensor needs to be close to the bearing seat position where the meshing gear or rolling bearing is located. This arrangement shortens the signal propagation path and reduces the energy attenuation of high-frequency vibration caused by material distortion in the transmission process.

[0073] The acoustic emission sensor 22 is rigidly coupled to the gear box base and bearing seat 18. For uneven installation surfaces, the sensor is fixed by welding or threaded connection of the waveguide rod. For flat installation surfaces, it is fixed by strong magnetic seat combined with acoustic coupling agent, and mechanical pressing device is added. The acoustic coupling agent fills the micro voids between the sensor end face and the surface of the measured body, ensuring the effective transmission of high-frequency elastic wave signals.

[0074] Each acoustic emission sensor 22 is connected with a signal conditioning unit. The signal conditioning unit integrates a preamplifier and a bandpass filter. The preamplifier amplifies the weak acoustic emission signal for gain and drives long-distance signal cable transmission. The passband of the bandpass filter is set to 1 kHz to 1 MHz, which filters out the low-frequency mechanical vibration of the press body and the environmental background noise before analog-to-digital conversion.

[0075] The conditioned analog signal is connected to a high-speed data acquisition card. The sampling rate of the data acquisition card is set to 2 MHz or higher to meet the Nyquist sampling theorem and ensure complete acquisition of the high-frequency transient characteristics of the acoustic emission signal. The data acquisition card is connected to an industrial control computer, and the computer executes subsequent signal processing and fault diagnosis algorithms.

[0076] The industrial control computer of the monitoring system is communicatively connected with the PLC controller of the large press through a field bus. The monitoring system reads the clutch engagement state signal from the PLC controller in real time as a collection trigger source, and reads the current mold number information. The mold number information is used to index the historical monitoring database required for subsequent data processing.

[0077] Referring to the drawings Figure 2 The present application provides an acoustic emission online monitoring method for early fault diagnosis of a large press gear box, comprising the following steps:

[0078] S10, obtaining the current press state and initializing the monitoring parameters: reading the clutch state and mold number of the PLC controller of the large press through the industrial control computer of the monitoring system, and calling the corresponding historical monitoring database according to the mold number;

[0079] S20, collecting acoustic emission signals and performing component separation: after the clutch is engaged, the monitoring data of each acoustic emission sensor is collected, and the monitoring data is separated into periodic components and impact components by using an extreme value driven sliding time window algorithm;

[0080] S30, frequency band decomposition of the separated signals: using a dual-tree complex wavelet packet transform to decompose the periodic components and the impact components respectively, to obtain signal components in different frequency bands;

[0081] S40, extracting multi-domain fault features and performing standardization: for each signal component, a multi-dimensional feature vector containing waveform features, time domain features, frequency domain features, envelope features and inter-band features is extracted, and Z-score standardization processing is performed;

[0082] S50, performing adaptive fault discrimination: inputting the standardized feature vectors into a multi-model fusion discrimination system containing the Laplace rule, one-class support vector machine and isolated forest, and determining whether the gear box has a fault according to the voting result;

[0083] S60, fault source positioning and result output: if it is determined that there is a fault, the fault source position is calculated based on the intensity ranking of the multi-point fault representation acoustic emission signal and the monitoring time difference, and the diagnostic result is output.

[0084] The specific embodiments of each of the above steps are described in detail below.

[0085] The specific embodiments of the historical monitoring database construction and dynamic updating guided by the mold structure in step S10 are as follows.

[0086] Since large presses need to replace different structures of molds when producing different products, the production pace (SPM) and the stamping load curve corresponding to different molds are different, which leads to different vibration and acoustic emission response references of the gear box under normal working conditions. To avoid false positives or false negatives caused by a single discrimination standard, the present embodiment adopts a mold management monitoring strategy.

[0087] Step S11, mold identity recognition and database indexing. The monitoring system reads the formula register address of the press PLC controller in real time through the field bus to obtain the unique identification code of the currently loaded mold. The system searches whether there is a corresponding historical monitoring database in the local storage unit according to the identification code. If there is, the characteristic distribution parameters and discrimination model parameters corresponding to the mold are called as the current monitoring reference; if there is not, a new database entry is created and marked as an initialization state.

[0088] Step S12, acquisition of reference data and construction of feature space. For a new mold marked as an initialization state, or during the running-in period after equipment overhaul, the system enters a self-learning mode. In this mode, the system assumes that the current equipment is in a healthy state, and continuously acquires a set number (for example, 50 to 100 stamping strokes) of acoustic emission raw signals. For these raw signals, the multi-domain feature vectors of periodic components and impact components are obtained according to the signal separation and feature extraction method described below. These feature vectors are stored in the historical monitoring database of the mold as an initial sample set, which is used to train an initial fault discrimination model.

[0089] Step S13, dynamic updating mechanism of historical monitoring database. Considering that the lubricating oil temperature change, the tooth surface running-in state change and the environmental temperature fluctuation during the long-term operation of the gear box will cause the baseline of the acoustic emission signal to slowly drift, in order to ensure the timeliness of the discrimination standard, the system adopts a database updating strategy based on a sliding window.

[0090] The specific updating strategy is as follows: a fixed length time window or stroke count window (for example, the last 10,000 stamping strokes) is set. The monitoring system continuously records the feature data judged as fault-free while performing fault diagnosis tasks. When the system does not occur any alarm within the set updating period (for example, every 168 hours of continuous operation), and all discrimination results are normal, the system triggers the database updating operation. The updating operation follows the first-in-first-out (FIFO) principle, adds the latest fault-free feature data collected in the current updating period to the sample set, and removes the oldest data in the sample set at the same time.

[0091] Step S14, reconstruction of model parameters. After the completion of the sliding update of the sample set, the system recalculates the statistical parameters of the features (such as the mean vector and the covariance matrix) based on the updated sample set, and incrementally trains or re-trains the fault discrimination model (such as the one-class support vector machine and the isolation forest described below). In this way, the monitoring system can automatically adapt to the gradual changes in the physical characteristics of the gearbox over time, suppressing false alarms caused by equipment aging without reducing fault sensitivity.

[0092] For the database storage structure, bus communication protocol analysis and other conventional engineering implementation means involved in the above steps, those skilled in the art can select SQL database, non-relational database and Profinet, Modbus and other standard protocol implementation according to the actual industrial control environment, and the specific code implementation belongs to the public technical knowledge in the art, which will not be described here.

[0093] Referring to the accompanying drawings Figure 3 , the specific implementation of the signal component separation method based on the extreme value driven sliding time window in step S20 is as follows.

[0094] Since the single stamping stroke of a large press contains different stages such as the slider descending after the clutch engagement, contacting the workpiece to do work, the slider returning and the flywheel idling to store energy during the operation process, the acoustic emission signal presents non-stationarity in the time domain. In order to realize fine diagnosis, it is necessary to accurately separate the impact component representing the loaded working condition from the periodic component representing the idling working condition.

[0095] Step S21, reference signal selection and initial sampling window setting. After receiving the clutch engagement trigger signal sent by the press PLC, the monitoring system starts synchronous acquisition. The system selects the monitoring data of the acoustic emission sensor arranged near the idler shaft as the reference signal. The reason for selecting the idler shaft signal as the reference is that the idler shaft is located in the middle link of the transmission chain, and its stress state can clearly reflect the energy fluctuation of the press stamping cycle, and the periodic characteristics of its signal envelope are more obvious than those of the end effector. The system sets the length of the initial sampling time window to be twice the length of the single standard production beat time (T) , that is, 2T, to ensure that the window contains at least two complete stamping cycles, so as to capture the boundary features at the periodic alternation.

[0096] Step S22, envelope extraction and smoothing of the reference signal. In order to overcome the interference of high-frequency burrs in the original acoustic emission signal caused by random crack propagation or transient impact on extreme value positioning, the system does not directly search for extreme values on the original waveform, but first performs envelope demodulation. The Hilbert transform is performed on the reference signal to construct an analytical signal​ , whose expression is:

[0097] ;

[0098] wherein, is the Hilbert transform of

[0099] ;

[0100] The envelope of the signal is further calculated as :

[0101] ;

[0102] Subsequently, a low-pass filter is applied to the envelope signal for smoothing. The cut-off frequency of the low-pass filter is set to 10 to 20 times of the press stroke frequency (SPM) (e.g. 2.5 Hz to 5 Hz for a 15 SPM press) to filter out the gear meshing frequency components and only keep the low frequency profile reflecting the press mechanical action stroke.

[0103] Step S23, extreme point positioning and time sequence sorting. Within the initial sampling time window of the smoothed envelope signal , local maximum points are searched. Since the time window covers two periods, the system identifies the four extreme points with the largest amplitude, which usually correspond to the slider bottom dead center or maximum contact force moment and its accompanying secondary oscillation peak in two consecutive press strokes. The system records the positions of the four extreme points on the time axis and labels them in chronological order as , , , .

[0104] Step S24, cut-off point determination and signal physical segmentation. The second extreme point and the third extreme point located in the middle position are selected. These two points respectively represent the tail feature of the previous press cycle and the head feature of the next press cycle (or the main work peak). A search neighborhood is set in the interval to find the minimum point of the envelope signal , and the time corresponding to the minimum point is determined as the cut-off point time . Physically, the cut-off point corresponds to the lowest energy point after the completion of a press and before the start of the next press, i.e. the dividing point between the periodic component and the impact component.

[0105] ​Step S25, full-channel synchronous truncation and component definition. Since all channels of the monitoring system are driven by the same clock source, the sensor signals have strict time synchronization. The truncation point time determined in step S24 is applied to all 8 raw monitoring data of acoustic emission sensors. With the truncation point as the boundary, the signal segment before the truncation point is defined as the impact component containing the complete stamping working process, and the low-energy smooth segment near the truncation point is defined as the periodic component reflecting the background noise of the mechanical transmission chain.

[0106] Step S26, sliding window iterative separation. For continuous monitoring process, the system adopts sliding window strategy. After completing the separation of the current time window, the sampling window slides backward by one single-beat time . The latter half of the data determined in the last window is taken as the former half of the new window, and steps S23 to S25 are repeatedly executed, so as to realize real-time and cycle-by-cycle separation of acoustic emission data in continuous production process.

[0107] Referring to the accompanying Figure 4 , the specific implementation of the anti-interference signal decomposition based on double-tree complex wavelet packet transform in step S30 is as follows.

[0108] Since the early failure of the gearbox is usually manifested as weak transient impact, and the complex working environment of the press causes the collected acoustic emission signals to have nonlinear and non-stationary characteristics, the traditional signal decomposition method (such as discrete wavelet transform) lacks shift invariance, and when the input signal has a slight time shift, the energy distribution of the sub-band obtained by decomposition will change, resulting in unstable feature extraction. Therefore, the embodiment adopts double-tree complex wavelet packet transform (DTCWPT) to perform multi-layer frequency band decomposition on the periodic component and the impact component separated in step S20.

[0109] Step S31, construct double-tree complex filter bank structure. The structure contains two parallel wavelet packet decomposition trees, which are defined as real part tree (Tree-R) and imaginary part tree (Tree-I) respectively. The real part tree adopts low-pass filter and high-pass filter , and the imaginary part tree adopts low-pass filter and high-pass filter . In order to realize approximate shift invariance and suppress spectral aliasing, the two groups of filters need to meet the specific delay condition: in the first layer of decomposition, the filters of the real part tree and the imaginary part tree maintain a delay of one sampling point; in the second layer and subsequent layers of decomposition, the filter groups of the real part tree and the imaginary part tree maintain a delay of half a sampling point. This delay design makes the wavelet basis function of the imaginary part tree approximately the Hilbert transform of the wavelet basis function of the real part tree.

[0110] Step S32, perform multi-layer wavelet packet decomposition. Set the decomposition layer number as​ (For example or The input acoustic emission signal components are iteratively decomposed using a binary tree. In each decomposition level, the signal is passed through low-pass and high-pass filters of the real and imaginary parts trees, respectively, and downsampled at intervals. For the ... layer( ) Nodes ( The real part wavelet packet coefficients were obtained respectively. imaginary part wavelet packet coefficients .

[0111] Step S33, single-branch reconstruction of frequency band components. To recover physically meaningful time-domain signal components from the decomposition coefficients for subsequent feature extraction, the system performs a single-branch reconstruction operation. For the... The first layer For each node, only the real coefficient of that node is retained. and imaginary part coefficient The coefficients of all other nodes are set to zero, and then a dual-tree complex wavelet packet inverse transform is performed using the corresponding synthetic filter bank. The real-part tree inverse transform outputs the reconstructed signal. The inverse transform of the imaginary part tree outputs the reconstructed signal. .

[0112] Step S34: Combine modulus extraction and signal component acquisition. Utilize the complementarity of the reconstruction results from the real part tree and the imaginary part tree to calculate the... Layer Complex signal components of each frequency band node The envelope magnitude or real part projection is used as the input for subsequent feature extraction. In this embodiment, the complementary properties of the two trees are utilized: the real part tree is equivalent to sampling at the peak of the signal, and the imaginary part tree is equivalent to sampling at the valley of the signal. The combined response is calculated using the following formula, thereby eliminating the oscillations caused by the initial phase difference of the signal:

[0113] ;

[0114] in, That is, the first one obtained after decomposition The smooth envelope signal of each frequency band component. This signal is invariant to the slight translation of the original acoustic emission signal, ensuring that the energy and waveform characteristics of the decomposed frequency band components remain stable in different stamping cycles, even if there are slight jitters at the moment of fault impact.

[0115] Through the above decomposition process, the original acoustic emission signal is finely divided into multiple sub-signal components covering different frequency ranges, and the high-frequency fault impact information is concentrated in the specific high-frequency node component, while the low-frequency mechanical interference is separated into the low-frequency node component, providing a high signal-to-noise ratio data basis for subsequent feature extraction. For the specific filter coefficient design of the dual-tree complex wavelet transform (such as the Q-shift filter designed by Kingsbury), it belongs to the existing mathematical tools in the field of signal processing, and the relevant algorithm library can be directly called by the person skilled in the art.

[0116] Referring to the accompanying drawings Figure 5 , the specific implementation of the multi-domain feature extraction and data standardization processing in step S40 is as follows.

[0117] Although each frequency band component obtained through step S30 separates noise and fault information at the physical level, high-dimensional time series data is not suitable for direct input into a machine learning model. In order to fully characterize the subtle changes of the gearbox in different health states, the present embodiment constructs a high-dimensional feature space from four dimensions of time domain statistics, frequency domain distribution, envelope characteristics and information entropy for each reconstructed frequency band component, and standardizes the feature matrix to eliminate the influence of dimension.

[0118] Step S41, extract time domain statistical features and waveform features. For the first layer th frequency band component ( is the sample point index), the effective value (RMS) reflecting the overall energy level of the signal and the kurtosis (Kurtosis) reflecting the early fault impact characteristics are calculated.

[0119] The calculation formula of the effective value is:

[0120] ;

[0121] The kurtosis is extremely sensitive to outliers with large amplitudes and is a key indicator for early crack detection, and its calculation formula is:

[0122] ;

[0123] Among them is the mean value of the signal. In addition, an adaptive threshold based on the background noise level (for example, 6dB higher than the background noise baseline) is set, the number of waveform oscillations exceeding the threshold is counted as the ring count, and the time from crossing the threshold to reaching the peak value is extracted as the rise time to represent the transient characteristics of the impact waveform.

[0124] Step S42, extract frequency domain statistical features. For the frequency band component Fast Fourier Transform (FFT) is performed to obtain power spectral density The center frequency (f0) and root mean square frequency (fr) of the power spectrum are calculated for monitoring the center shift of the spectrum energy caused by tooth surface wear or bearing spalling. The center frequency f0is calculated by The formula is as follows:

[0125] ;

[0126] wherein f is the frequency variable, and fNis the Nyquist frequency. Step S43, extracting inter-band information entropy features. The signal is decomposed into multiple orthogonal subspaces by using dual-tree complex wavelet packet transform, the distribution probability of the energy of each sub-band in the total energy is calculated, and then the energy entropy is solved to quantify the complexity and disorder degree of the signal energy distribution.

[0127] First, the energy Ei of the ith frequency band node is calculated

[0128] Then, the proportion of the node energy in the total energy of all nodes in the same layer is calculated :

[0129] ;

[0130] Based on this, the Shannon energy entropy H is calculated :

[0131] ;

[0132] When the gearbox components are damaged, the energy originally concentrated in certain specific meshing frequencies will spread to a wide frequency band, causing the energy entropy value to change.

[0133] In addition to the features detailed above, the feature space constructed in this embodiment can also include other multi-domain features mentioned in the disclosure to enhance the sensitivity to different fault modes. Specifically, they include:

[0134] Time domain and waveform features: mean, variance, maximum, peak-to-peak value, kurtosis (similar to kurtosis), duration, arrival time, signal energy and amplitude, etc.

[0135] Frequency domain features: average frequency, frequency variance, etc.

[0136] Envelope features: envelope spectrum energy, envelope spectrum center frequency, envelope spectrum variance, envelope spectrum extreme value, etc.

[0137] ​​​Band characteristics: In addition to energy entropy, it also includes singular entropy, dispersion entropy, conditional entropy and cross entropy.

[0138] The specific calculation methods of these features are existing mathematical tools in the field of signal processing, and those skilled in the art can select part or all of the features to construct a feature vector according to actual needs.

[0139] Step S44, construction of multi-dimensional feature vector and Z-score standardization. Arrange all the feature values extracted in the above steps in order to form an original feature vector F , which can represent the state of the equipment in the current time window. is the total dimension of the features. Since features with different physical meanings (such as energy entropy and kurtosis) differ significantly in numerical magnitude, directly inputting the discriminant model based on distance measurement will cause the model to fail to converge or the weight to deviate. Therefore, the Z-score standardization method is used to map the features to a standard normal distribution.

[0140] For the current collected feature value , the standardized value is calculated as follows:

[0141] ;

[0142] , wherein and are not only calculated from the current sample, but also directly call the mean and standard deviation of the corresponding mold reference data in the historical monitoring database established in step S10. This processing method ensures that the monitoring data is the degree of deviation relative to the historical health reference of the equipment, thereby realizing the unification of the discrimination scale under different mold working conditions. The feature vector after standardization is transmitted to the subsequent fault discrimination module.

[0143] The specific implementation of the adaptive fault discrimination system based on multi-model fusion in step S50 is as follows.

[0144] Due to the weak early fault characteristics of large press gearboxes and the lack of clear fault label samples, it is difficult to balance high detection rate and low false alarm rate by relying on a single discriminant algorithm. This embodiment constructs a hybrid discrimination system containing a statistical model, a distance-based geometric model and a density-based ensemble learning model. This system only uses historical data for single-class (One-Class) training, and discriminates abnormal states through the logic of multi-model voting.

[0145] Step S51, construction and execution of the Laplace rule (3σ criterion) discrimination module. This module serves as a statistical benchmark to identify abnormalities using the fluctuation trend of historical features. For each standardized feature value Since it has been Z-score transformed by the historical failure-free data, it should theoretically follow a standard normal distribution under healthy condition . The discrimination threshold interval is set as . If more than a certain proportion (e.g. 30%) of the feature components in the current feature vector fall outside the interval, i.e. , the module determines that the sample belongs to an abnormal distribution statistically, and outputs the discrimination result (abnormal); otherwise, it outputs (normal). The module has fast response capability for amplitude mutation type failures.

[0146] Step S52, construction and execution of the one-class support vector machine (OC-SVM) discrimination module. The module is an unsupervised geometric model, aiming to find a minimum hypersphere or hyperplane that can tightly enclose the normal samples in high-dimensional space. Radial basis function (RBF) is used as the kernel function to map the low-dimensional nonlinear features to high-dimensional feature space. By solving the dual optimization problem, the optimal hyperplane that separates the origin from the normal samples is determined:

[0147] ;

[0148] where is the Lagrange multiplier. In the monitoring stage, the current feature vector is substituted into the trained decision function . If , it indicates that the sample falls in the abnormal area outside the hyperplane, and the module outputs the discrimination result ; otherwise, it outputs . The kernel parameter and the penalty factor νν in the model are determined by 5-fold cross-validation of historical samples to ensure that the fitting of the model to the normal boundary is neither too tight nor too loose.

[0149] Step S53, construction and execution of the isolation forest discrimination module. The module is based on the idea of ensemble learning, and uses the sparse distribution of abnormal data points in feature space. A forest composed of multiple random binary trees (iTrees) is constructed. In the training stage, the feature and the split point are randomly selected to recursively cut the sample space until each sample is isolated. Since abnormal samples are usually sparsely distributed and far away from high-density areas, they are often located in the shallow layer of the tree and have a shorter average path length . The abnormal score of the current sample is calculated as:

[0150] ;

[0151] where is the sample size the average path length normalization factor of the time. If the anomaly score is close to 1, it indicates that the sample is extremely easy to be isolated, and is determined as an anomaly, the module outputs ; if the score is much less than 0.5, the output . This module has extremely high sensitivity to complex nonlinear fault patterns in high-dimensional feature space.

[0152] Step S54, multi-model voting fusion decision. In order to reduce false alarms caused by algorithm bias of a single model, the system introduces a voting fusion mechanism. The discrimination results of the three basic modules are collected , , , and the weighted or unweighted sum value is calculated. A voting threshold is set. When and only when , that is, at least two basic modules simultaneously determine that the current state is abnormal, the monitoring system formally sends out a fault warning signal and triggers the subsequent fault source positioning process. If , it is considered that the current abnormal feature belongs to incidental noise interference, and the system remains silent and continues to monitor the next period. This mechanism guarantees the fault detection rate while improving the anti-interference ability in the industrial field environment.

[0153] It should be noted that the above feature vector contains all the features extracted for periodic components and impact components. Therefore, the logic of the above multi-model fusion discrimination essentially follows the following principles:

[0154] When and only when the relevant features of the periodic component and the impact component both meet the health criteria (i.e. are determined to be normal), the system indicates that the gearbox does not have a fault; if the features of either component show abnormality (resulting in a voting result showing a fault), the system determines that there is a fault and triggers subsequent positioning analysis.

[0155] Referring to the accompanying Figure 6 , the specific implementation of the fault source positioning method based on intensity ranking and sound speed correction in step S60 is as follows.

[0156] When the adaptive fault discrimination system determines that the gearbox has an anomaly, the system needs to further determine the specific physical location of the fault occurrence in order to guide the maintenance personnel to carry out precise maintenance on the specific shaft or bearing. Due to the complex structure of the large press gearbox, which contains a large number of reinforcing ribs, oil medium and components of different materials, the propagation speed of sound waves in it is anisotropic, and there is a serious reflection and attenuation phenomenon. Therefore, this embodiment adopts a time difference positioning method combining energy intensity space constraint and sound speed correction.

[0157] Step S61, Calibration and Construction of the Equivalent Sound Velocity Matrix. During the commissioning phase before the monitoring system's formal operation, or during equipment downtime maintenance, the Hsu-Nielsen lead-broken sound source test method is used to calibrate the sound propagation characteristics of the gearbox on-site. This is done at preset coordinate positions of the gearbox's main bearing seats and key transmission nodes. A standard lead-breaking excitation method (e.g., using a 0.5mm diameter 2H lead core broken at a 30-degree angle) is applied to generate a simulated acoustic emission source. The system records the arrival times of the sound waves at each sensor location. transmission time Based on the known distance and propagation time, calculate the equivalent speed of sound after considering the effects of the casting structure and lubricating oil film. :

[0158] ;

[0159] in The moment the lead broke is recorded. Multiple measurements at multiple points are taken, and the average value is used to construct an equivalent sound velocity matrix or a global average equivalent sound velocity covering different monitoring areas, which is then used to correct the deviation between the theoretical and actual sound velocities.

[0160] Step S62, initial screening of fault areas based on signal strength. When a fault event is detected, the system first extracts the signal energy characteristics (such as root mean square value RMS) of all 8 sensor channels within the current time window. Based on the physical characteristic that acoustic emission signals attenuate with increasing distance in a medium (the attenuation model is approximately...),... The sensor closest to the fault source theoretically receives the strongest signal energy. The system sorts the energy values ​​of the eight channels from largest to smallest, identifying the top three sensor numbers based on their energy levels, and denoting them as follows: , , The system narrows the search range for the fault source to a triangular region or its neighborhood with these three sensors as vertices, thereby eliminating interference caused by reflected waves from sensors far from the fault source and reducing the solution space.

[0161] Step S63, precise extraction of wavefront arrival time (TOA). This is performed on the first three selected sensors. , , The waveform data were used to determine the precise arrival time of the fault transient impact signal at each sensor using either the long short time window energy ratio (STA / LTA) method or the Akaike information criterion (AIC) method, denoted as follows: , , Calculate the time difference of arrival (TDOA) between sensors:

[0162] ;

[0163] Step S64, solving of hyperbolic positioning equation set. Based on the plane positioning assumption, a nonlinear hyperbolic equation set is established with time difference as the observation. Let the to-be-solved coordinates of the fault source be , the known coordinates of the th sensor be , and the equivalent sound speed calibrated in step S61 be combined, the following equation is established:

[0164] ;

[0165] For an array composed of three sensors, the system simultaneously solves the above equation set, and uses the Newton-Raphson iteration method or the least square method for numerical solution to obtain the estimated coordinates of the fault source .

[0166] Step S65, physical mapping and result output. The calculated estimated coordinates are mapped to the three-dimensional CAD digital model or two-dimensional development drawing of the gearbox. The system determines which mechanical component (such as the high-speed shaft rear bearing, the intermediate shaft A gear meshing area, etc.) the coordinates fall within, thereby determining the specific fault component. Finally, the monitoring system highlights the fault position on the human-computer interaction interface and generates a diagnostic report containing the fault type (determined by feature discrimination) and the fault position.

[0167] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An online acoustic emission monitoring method for early fault diagnosis of large press gearboxes, characterized in that, Includes the following steps: S10. The industrial control computer of the acoustic emission online monitoring system reads the clutch status and mold number of the large press PLC controller, and retrieves the historical monitoring database in the local storage unit according to the mold number to obtain the monitoring benchmark parameters corresponding to the current mold. S20. When the clutch status indicator is engaged, the monitoring data of each acoustic emission sensor arranged on the gearbox of the large press is collected, and the monitoring data is processed using the extreme value driven sliding time window algorithm to separate the periodic component and the impact component. S30. The periodic component and the impulsive component are decomposed by dual-tree complex wavelet packet transform to obtain signal components in different frequency bands. S40. For each of the signal components, extract a multi-dimensional feature vector containing waveform features, time domain features, frequency domain features, envelope features and inter-band features, and perform Z-score normalization processing using the monitoring reference parameters to generate a normalized multi-dimensional feature vector. S50. Input the standardized multidimensional feature vector into a multi-model fusion discrimination system that includes Laida's rule, single-class support vector machine and isolated forest to obtain the individual discrimination results output by each model. Perform voting statistics on the individual discrimination results and determine the final fault discrimination result based on the voting statistics results. S60. When the fault diagnosis result indicates that there is a fault in the gearbox, the fault source location is calculated based on the intensity ranking of the acoustic emission signals representing the multi-point fault and the monitoring time difference, and a diagnostic result containing fault location information is output.

2. The online acoustic emission monitoring method for early fault diagnosis of large press gearboxes according to claim 1, characterized in that, In step S10, obtaining the monitoring benchmark parameters corresponding to the current mold specifically includes: The acoustic emission online monitoring system retrieves the obtained mold number from the local storage unit; If a corresponding historical monitoring database exists, the feature distribution parameters and discrimination model parameters corresponding to the mold number are retrieved as the monitoring benchmark parameters. If the corresponding historical monitoring database does not exist, a new database entry is created and the system enters self-learning mode. In the self-learning mode, a set number of acoustic emission raw signals from stamping strokes are continuously acquired, and the multi-domain feature vectors of the reference periodic component and the reference impact component are extracted as an initial sample set, stored in the new database entry, and the monitoring reference parameters are calculated based on the initial sample set.

3. The online acoustic emission monitoring method for early fault diagnosis of large press gearboxes according to claim 2, characterized in that, Step S10 further includes performing a sliding window-based database update: A fixed-length travel counting window is set, and the acoustic emission online monitoring system continuously records the characteristic data that are determined to be fault-free; When the acoustic emission online monitoring system does not trigger any alarms within the set update cycle and all judgment results are normal, it follows the first-in-first-out principle to add the latest fault-free feature data collected in the current update cycle to the sample set, while removing the oldest equivalent amount of data from the sample set. The acoustic emission online monitoring system recalculates the statistical parameters of the features based on the updated sample set and performs incremental training or retraining on the fault discrimination model.

4. The online acoustic emission monitoring method for early fault diagnosis of large press gearboxes according to claim 1, characterized in that, In step S20, the separation of periodic and impulsive components using an extremum-driven sliding time window algorithm specifically includes: The signal from the monitoring data corresponding to the acoustic emission sensor located near the idler shaft is selected as the reference signal, and the length of the initial sampling time window is set to twice the standard production cycle time. The reference signal is subjected to Hilbert transform to construct an analytic signal, and its envelope is calculated; the envelope is then smoothed using a low-pass filter. Within the initial sampling time window of the smoothed envelope, search for the four extreme points with the largest amplitudes and mark them in chronological order; Select the second and third extreme points located in the middle, find the minimum point of the envelope within the time interval of the second and third extreme points, and determine the time corresponding to the minimum point as the cutoff point time; Using the cutoff time as a boundary, the signal segment before the cutoff time is defined as the impulsive component, and the low-energy stable segment near the cutoff time is defined as the periodic component.

5. The online acoustic emission monitoring method for early fault diagnosis of large press gearboxes according to claim 1, characterized in that, In step S30, obtaining signal components in different frequency bands specifically includes: Construct a dual-tree complex filter bank structure containing real part trees and imaginary part trees; In the first-level decomposition, a delay of one sampling point is maintained between the filters of the real part tree and the imaginary part tree; In the second and subsequent decomposition layers, a delay of half a sampling point is maintained between the filter banks of the real part tree and the imaginary part tree; Set the number of decomposition levels, and iteratively decompose the input periodic component and the impulsive component to obtain the real part wavelet packet coefficients and the imaginary part wavelet packet coefficients; Perform a single-branch reconstruction operation, use a synthesis filter bank to output the real part tree reconstruction signal and the imaginary part tree reconstruction signal respectively, and calculate the synthesis response of the real part tree reconstruction signal and the imaginary part tree reconstruction signal as a smooth envelope signal.

6. The online acoustic emission monitoring method for early fault diagnosis of large press gearboxes according to claim 1, characterized in that, In step S40, generating the standardized multidimensional feature vector specifically includes: The inter-band features specifically include energy entropy, singularity entropy, dispersion entropy, conditional entropy, and cross entropy; The Z-score standardization method is used to perform standardization calculations on each feature element that constitutes the multidimensional feature vector; the mean and standard deviation used in the standardization calculation are included in the monitoring benchmark parameters.

7. The online acoustic emission monitoring method for early fault diagnosis of large press gearboxes according to claim 1, characterized in that, In step S50, the multi-model fusion discrimination system performs the following processing logic: Laida's rule discrimination logic: Set a discrimination threshold range. If more than a set proportion of feature elements in the standardized multidimensional feature vector fall outside the discrimination threshold range, it is judged as abnormal. The single-class support vector machine discrimination logic is as follows: the radial basis function is used as the kernel function, and the optimal hyperplane that separates the origin from the normal samples is determined by solving the dual optimization problem; if the standardized multidimensional feature vector falls in the abnormal region outside the optimal hyperplane, it is judged as an anomaly. Isolated Forest Detection Logic: Construct a forest composed of multiple random binary trees, calculate the average path length of the standardized multidimensional feature vector in the trees and convert it into anomaly score; if the anomaly score is greater than a preset anomaly threshold, it is determined to be an anomaly.

8. The online acoustic emission monitoring method for early fault diagnosis of large press gearboxes according to claim 7, characterized in that, In step S50, the fault determination result is obtained according to the following logic: Calculate the sum of the discrimination results of the three basic discrimination logics: the Laida rule discrimination logic, the single-class support vector machine discrimination logic, and the isolated forest discrimination logic; if the sum is greater than or equal to 2, then a fault is determined to exist. The standardized multidimensional feature vector contains all features extracted for the periodic component and the impact component; a judgment result indicating no fault is generated if and only if the relevant features of the periodic component and the impact component are all judged to be normal; if the features of any component of the periodic component and the impact component show abnormalities that cause the sum value display to malfunction, a judgment result indicating a fault is generated.

9. The online acoustic emission monitoring method for early fault diagnosis of large press gearboxes according to claim 1, characterized in that, In step S60, before calculating the location of the fault source, the equivalent sound velocity matrix is ​​also constructed: The sound propagation characteristics of the gearbox were calibrated on-site using the Hsu-Nielsen lead-broken sound source test method. A standard lead-breaking excitation was performed at a preset coordinate position, and the propagation time of the sound wave to each of the acoustic emission sensor positions was recorded. The equivalent sound velocity at a single point is calculated based on the ratio of the preset known distance to the propagation time, and an equivalent sound velocity matrix covering different monitoring areas is constructed through multiple measurements at multiple points.

10. The online acoustic emission monitoring method for early fault diagnosis of large press gearboxes according to claim 9, characterized in that, In step S60, calculating the location of the fault source specifically includes: Extract the signal energy characteristics of the monitoring data of all the acoustic emission sensors within the current time window, and identify the three acoustic emission sensors with the highest energy values; The search range for the fault source is narrowed down to the triangular region formed by the three acoustic emission sensors or the neighborhood of the triangular region. For the three selected acoustic emission sensors, the precise time when the fault transient impact signal arrives at each of the acoustic emission sensors is determined, and the arrival time difference between the acoustic emission sensors is calculated. Based on the equivalent sound velocity matched from the equivalent sound velocity matrix in the triangular region, a nonlinear hyperbolic equation system with the arrival time difference as the observation quantity is established. The Newton-Raphson iterative method is used for numerical solution to obtain the estimated coordinates of the fault source.

Images