Intelligent state monitoring and fault diagnosis system and method for die cutting gilding equipment

Abstract

The application provides a die cutting and hot stamping equipment intelligent state monitoring and fault diagnosis system and method, and relates to the field of intelligent monitoring.The method comprises the following steps: collecting working parameters of multiple key parts of the die cutting and hot stamping equipment, constructing a time sequence collection window, slidingly collecting the working parameters, and obtaining characteristic information reflecting the equipment state; based on the characteristic information, constructing an anomaly detection model, performing anomaly detection on the equipment state, obtaining an anomaly score, and judging whether the equipment state is abnormal according to the anomaly score; for the characteristic information judged as abnormal, constructing a fault diagnosis model based on the fault type to which the characteristic information belongs, performing fault diagnosis on the equipment state, and generating a diagnosis result; and generating a comprehensive diagnosis and operation and maintenance decision report according to the diagnosis result.The application realizes comprehensive perception of the internal state of a closed host through multi-sensor collaborative monitoring and dynamic time sequence collection, breaks through the limitations of traditional monitoring, and provides accurate data basis for early fault warning and predictive maintenance.

Description

Technical Field

[0001] This invention relates to the field of post-press equipment, and in particular to an intelligent status monitoring and fault diagnosis system and method for die-cutting and hot stamping equipment. Background Technology

[0002] Post-press equipment includes various types such as die-cutting, hot stamping, and creasing. Among them, die-cutting and hot stamping equipment are the core equipment in post-press processing. Their main unit bears the functions of die-cutting and hot stamping and must withstand the periodic impact of hundreds of tons during high-speed operation while maintaining micron-level precision. However, the main unit adopts a closed structure design to ensure stability, but this also makes it difficult to directly observe and monitor the condition of its internal core transmission components. Routine maintenance relies on manual experience, and its effectiveness is difficult to assess. Currently, the industry generally adopts a passive approach of repairing after a failure occurs, lacking effective early warning methods, resulting in frequent unplanned downtime and high maintenance costs.

[0003] Existing monitoring technologies suffer from three main limitations: ① Monitoring methods are limited and not deeply adapted to the equipment. They often rely on individual sensors for threshold alarms, failing to comprehensively and continuously collect signals from multiple key components within the host unit, and even less able to capture the subtle characteristics of progressive faults such as early pitting corrosion in bearings and wear on support shafts. ② Diagnostic methods are outdated, lacking the ability to deeply integrate and analyze multi-source heterogeneous data such as vibration and temperature. They still rely on manual feature extraction or fixed threshold judgments, unable to adapt to the dynamic changes of equipment under complex operating conditions such as different speeds and pressures. This leads to inaccurate fault location, high false alarm rates, and complete inability to achieve intelligent fault classification and remaining life prediction. ③ Even with the introduction of artificial intelligence methods, limitations such as the scarcity of fault samples in industrial settings, high data noise, and insufficient computing power of edge computing devices make it difficult to train high-precision, lightweight diagnostic models and achieve real-time, reliable online deployment. Summary of the Invention

[0004] This invention provides an intelligent status monitoring and fault diagnosis system and method for die-cutting and hot stamping equipment to solve at least one of the above-mentioned technical problems.

[0005] On the one hand, the present invention provides an intelligent status monitoring and fault diagnosis method for die-cutting and hot stamping equipment, comprising the following steps:

[0006] The system collects working parameters from multiple key components of the die-cutting and hot stamping equipment. These working parameters include multiple physical quantities distributed along the equipment's power transmission path and load-bearing structure. The physical quantities include at least one or more of vibration, temperature, displacement, angle, pressure, torque, current, flow rate, and abrasive particles. A time-series acquisition window is constructed to perform sliding acquisition of the working parameters, thereby obtaining characteristic information reflecting the equipment's status.

[0007] Based on the aforementioned feature information, an anomaly detection model is constructed to detect anomalies in the equipment status. The anomaly detection model is trained using an unsupervised contrastive learning approach, utilizing only the feature information of the equipment under normal operating conditions as training samples. Positive sample pairs are constructed and input into the contrastive learning network for training to learn the feature representation of the normal state. The contrastive learning network adopts a dual-channel symmetric structure, and a loss function is constructed by calculating the cosine similarity between the feature vectors output by the two channels and the predicted vector to characterize the directional consistency of positive sample pairs in the feature space.

[0008] Based on the trained anomaly detection model, anomaly detection is performed on the real-time collected feature information to be tested, an anomaly score is calculated and compared with a preset anomaly judgment threshold. If the anomaly score exceeds the anomaly judgment threshold, the equipment status is determined to be abnormal. The anomaly score is constructed based on the loss function and is obtained by performing a nonlinear transformation on the loss value, which is used to quantify the degree of deviation between the current state and the normal state.

[0009] For feature information that is judged to be abnormal, a fault diagnosis model is constructed based on the fault type to which it belongs, and fault diagnosis is performed on the equipment status to generate diagnosis results.

[0010] In one embodiment, the step of acquiring the operating parameters includes at least one of the following steps:

[0011] Vibration acceleration and / or surface temperature information are collected on the cam housing;

[0012] Radial vibration and / or temperature information is collected in the bearing area on the flywheel side of the main drive shaft and / or at the paper output end cap.

[0013] The axial movement of the power shaft is collected at the end cap of the paper output end of the main drive shaft.

[0014] Vibration and / or temperature information is collected at the half-shaft flange of the main worm gear drive shaft;

[0015] Vibration, temperature, angle and / or axial displacement information of the rotation angle transmission shaft is collected at the end of the worm gear transmission shaft of the host machine;

[0016] The axial movement of the drive shaft is collected on the shaft end face distributed along the axial direction of the main worm gear drive shaft;

[0017] Operating temperature information is collected on one side of the support shaft and the corresponding base of the elbow mechanism;

[0018] Calculate the working pressure by measuring micro-deformation at the top of the printing platform;

[0019] The output torque and / or operating current are collected from the bearings in the main motor housing.

[0020] Collect information on the temperature, flow rate, and / or abrasive particles of the lubricating oil in the lubricating oil sump of the lubrication system.

[0021] In one embodiment, the step of obtaining the feature information includes:

[0022] The vibration signal in the operating parameters is converted into a time-frequency characteristic map;

[0023] The temperature, displacement, angle, torque, current, flow rate and / or abrasive information in the working parameters are combined with the working condition information and converted into time series features;

[0024] The operating condition information includes at least one of the following: time, initial temperature, ambient temperature, equipment power-on time, equipment operating time, imprinting platform pressure, shaft speed, or production efficiency.

[0025] In one embodiment, converting the vibration signal into a time-frequency feature map includes the following specific steps:

[0026] The acquired vibration signal is resampled based on the rotational speed and sampling frequency;

[0027] The resampled samples are filtered to eliminate the influence of high-frequency noise on the fault characteristics. The filtering of the resampled vibration signal includes: determining a preset fault characteristic frequency band based on the fault characteristic frequency of the bearing or transmission component, and using a bandpass filter to retain the signal components within the fault characteristic frequency band to suppress high-frequency noise and low-frequency interference.

[0028] The filtered samples are standardized so that their mean is 0 and their variance is 1.

[0029] The standardized samples are subjected to continuous wavelet transform to obtain the time-frequency characteristic map of the vibration signal; the continuous wavelet transform of the standardized samples includes the following steps:

[0030] Based on the mother wavelet function, scale and translation parameters are set, and the mother wavelet function is scaled and translated to form a family of wavelet functions;

[0031] Wavelet coefficients are calculated for the standardized samples and the wavelet function family, and the wavelet coefficients are normalized using a normalization factor.

[0032] A time-frequency characteristic map of the vibration signal is constructed based on the amplitude or energy of the wavelet coefficients, wherein the time-frequency characteristic map is plotted with time as the horizontal axis and scale or frequency as the vertical axis.

[0033] In one embodiment, the specific process of constructing an anomaly detection model and determining whether the equipment status is abnormal includes:

[0034] Use the feature information of the die-cutting and hot stamping equipment under normal operating conditions as training samples;

[0035] The training samples are constructed into multiple positive sample pairs;

[0036] Each positive sample pair is input into an improved SimSiam model for training; the improved SimSiam model uses a deep separable convolutional neural network as an encoder, a support vector machine as a predictor, and alternately updates the gradients of the two channels using a gradient cutoff method.

[0037] Based on the output of the model to healthy sample pairs during the training phase, anomaly scores are constructed using loss values, and the anomaly determination threshold is determined using the P-quantile method. The anomaly determination threshold using the P-quantile method includes: inputting an independent healthy validation set into the trained anomaly detection model to obtain an anomaly score set, statistically analyzing the distribution of the anomaly score set, and selecting the anomaly score corresponding to a preset quantile as the anomaly determination threshold.

[0038] During online detection, based on the real-time collected feature information to be tested, a healthy baseline sample is selected from the healthy sample library composed of feature information under normal operation according to preset rules, and the feature information to be tested and the healthy baseline sample are combined to form an anomaly detection model trained by the input of the sample to be tested.

[0039] The anomaly score of the test sample pair is calculated and compared with the anomaly judgment threshold determined by the P quantile method. If the anomaly score exceeds the anomaly judgment threshold, the equipment status is determined to be abnormal, and the test feature information is entered into the fault diagnosis model for processing as anomaly feature information.

[0040] In one embodiment, the loss value of the anomaly detection model is obtained by symmetrically combining the cosine similarity term between the first feature vector A1 output by the first channel encoder and the second prediction vector B2 output by the second channel predictor, and the cosine similarity term between the second feature vector A2 output by the second channel encoder and the first prediction vector B1 output by the first channel predictor, so that the loss value tends to be smaller when the feature representation of the healthy sample pair in the two channels tends to be consistent.

[0041] The cosine similarity is calculated as follows: the inner product is obtained by multiplying the corresponding elements of the two vectors involved in the calculation, and the inner product is divided by the product of the magnitudes of the two vectors to characterize the degree of consistency in the directions of the two vectors.

[0042] The anomaly score is constructed based on the loss value, and the anomaly score is the logarithmic result of summing the loss value with a preset small amount.

[0043] In one embodiment, the specific process of constructing a fault diagnosis model includes:

[0044] Using feature information that includes normal operating status data and fault operating status data, the equipment operating status dataset is constructed by classifying it according to fault type.

[0045] The equipment operation status dataset is divided into a training set, a validation set, and a test set;

[0046] Based on the MobileNetV3 model, a multi-classification model is constructed with a deep separable convolutional neural network as the encoder and fully connected layers and Softmax layers as classifiers.

[0047] The model is trained using the training set, and its parameters are optimized using the cross-entropy loss function so that it can identify and classify the fault categories of the input signal.

[0048] In one embodiment, the intelligent status monitoring and fault diagnosis method for die-cutting and hot stamping equipment further includes the following step: generating a comprehensive diagnosis and operation and maintenance decision report based on the diagnosis results;

[0049] The comprehensive diagnosis and maintenance decision report is generated based on the large language model of the Transformer architecture, according to the diagnosis results and historical maintenance data. The comprehensive diagnosis and maintenance decision report includes at least one of the following: fault location, fault type, fault level, early warning prompt or maintenance suggestion.

[0050] On the other hand, the present invention provides an intelligent status monitoring and fault diagnosis system for die-cutting and hot stamping equipment, comprising:

[0051] The signal acquisition module is used to acquire working parameters of multiple key parts of the die-cutting and hot stamping equipment. The working parameters include multiple physical quantities distributed along the power transmission path and load-bearing structure of the equipment. The physical quantities include at least one or more of vibration, temperature, displacement, angle, pressure, torque, current, flow rate, and abrasive particles.

[0052] The information processing module is used to process the working parameters, construct a time-series acquisition window and perform sliding acquisition to obtain feature information reflecting the equipment status;

[0053] The intelligent diagnostic module is used to perform anomaly detection and fault diagnosis on the equipment status based on the aforementioned feature information, and generate diagnostic results. The anomaly detection is implemented using an unsupervised contrastive learning approach, utilizing only feature information from the equipment's normal operating state as training samples. Positive sample pairs are constructed and input into the contrastive learning network for training, learning the feature representation of the normal state. The contrastive learning network employs a dual-channel symmetrical structure, constructing a loss function by calculating the cosine similarity between the feature vectors output from the two channels and the predicted vector, to characterize the directional consistency of the positive sample pairs in the feature space. Based on the trained anomaly detection model, anomaly detection is performed on the real-time collected feature information, calculating an anomaly score and comparing it with a preset anomaly judgment threshold. If the anomaly score exceeds the anomaly judgment threshold, the equipment status is determined to be abnormal. The anomaly score is constructed based on the loss function, obtained by performing a nonlinear transformation on the loss value, and is used to quantify the degree of deviation between the current state and the normal state.

[0054] In one embodiment, the signal acquisition module includes at least one of the following acquisition units:

[0055] A cam acquisition unit, including a first vibration sensor and / or a first temperature sensor, is arranged at the cam housing to acquire vibration and / or temperature information at the cam.

[0056] The power shaft acquisition unit includes a set of second vibration sensors, a set of second temperature sensors and / or a first micro-displacement sensor, which are respectively arranged at the end caps of both ends of the power shaft near the bearings, for acquiring vibration, temperature and / or axial displacement information of the power shaft;

[0057] The transmission shaft acquisition unit includes a third vibration sensor, a third temperature sensor, an angle sensor, and / or a second micro-displacement sensor. The third vibration sensor and / or the third temperature sensor are adapted to be arranged at the half-shaft flange of the main machine worm gear transmission shaft to acquire vibration and / or temperature information of the transmission shaft. The angle sensor is adapted to be set at the shaft end of the transmission shaft to acquire the rotation angle of the transmission shaft. The second micro-displacement sensor is adapted to be arranged axially on the shaft end face of the transmission shaft to acquire the axial movement of the transmission shaft.

[0058] The data acquisition unit for the support shaft or elbow mechanism includes multiple fourth temperature sensors arranged on one side of the elbow mechanism support shaft and the base, for collecting temperature information of the support shaft or elbow mechanism support shaft and the base.

[0059] The platform acquisition unit, including a micro-deformation sensor, is located at the top center of the imprinting platform to acquire pressure information of the imprinting platform.

[0060] The main motor acquisition unit includes a fourth vibration sensor and / or torque sensor and a current sensor arranged in the housing near the bearing of the main motor output shaft, for acquiring torque, current and / or vibration information of the main motor;

[0061] The lubrication system acquisition unit, including an oil temperature sensor, a flow sensor, and / or an abrasive sensor, is arranged in the lubricating oil sump to collect information on the temperature, flow rate, and / or abrasive particles of the lubricating oil.

[0062] In one embodiment, the information processing module includes:

[0063] The signal reading and transmission unit is used to read the signals collected by each unit of the signal acquisition module at a set frequency.

[0064] The feature extraction unit is used to convert the read signal into a time-frequency feature map for diagnosis and / or a time-series feature combined with operating condition information.

[0065] In one embodiment, the system further includes a cloud collaboration module, which comprises:

[0066] A cloud computing unit is used to train and optimize the artificial intelligence model used by the intelligent diagnostic module;

[0067] The cloud storage unit is used to store equipment operation information, feature information, historical fault maintenance information, artificial intelligence models and their weight data;

[0068] The system also includes:

[0069] The operation and maintenance decision module is used to generate a comprehensive diagnosis and operation and maintenance decision report based on historical fault information and the diagnostic results.

[0070] The cloud push unit is used to push the comprehensive diagnostic and operation and maintenance decision report generated by the operation and maintenance decision module to the user terminal;

[0071] The operation and maintenance decision module adopts a large language model based on the Transformer architecture to generate a comprehensive diagnosis and operation and maintenance decision report based on the diagnostic results and historical maintenance data. The comprehensive diagnosis and operation and maintenance decision report includes at least one of the following: fault location, fault type, fault level, early warning prompt or maintenance suggestion.

[0072] In one embodiment, the intelligent diagnostic module includes:

[0073] The anomaly detection unit is used to deploy an anomaly detection algorithm based on the improved SimSiam model, calculate anomaly scores based on feature information, and determine whether the device's operating status is abnormal.

[0074] The fault diagnosis unit is used to deploy a fault diagnosis model based on the MobileNetV3 framework, and to diagnose the fault type and location when an anomaly is detected.

[0075] Beneficial effects:

[0076] (1) The intelligent status monitoring and fault diagnosis system and method for die-cutting and hot stamping equipment provided by this invention effectively overcomes the core problems in the prior art, such as the difficulty in observing the internal status due to the closed structure of the host, the single traditional monitoring method, the backward diagnostic method, and the difficulty in intelligent deployment. Through in-depth analysis of the power transmission path and core force-bearing components of the die-cutting and hot stamping equipment, the system innovatively constructs an indirect monitoring network based on the synergy of multiple physical quantities such as vibration, temperature, and displacement. Combined with the time-series acquisition window and sliding acquisition mechanism, it realizes comprehensive, continuous, and dynamic perception of the key operating status inside the closed host. This multi-dimensional and high-frequency collaborative monitoring method fundamentally breaks through the limitations of traditional single sensor or fixed threshold alarm in terms of coverage and sensitivity, laying a solid data foundation for the capture of early weak fault characteristics and predictive maintenance.

[0077] (2) The intelligent status monitoring and fault diagnosis system and method for die-cutting and hot stamping equipment provided by this invention, in the signal processing and feature extraction stage, introduces advanced preprocessing algorithms such as continuous wavelet transform to efficiently convert multi-source heterogeneous original signals into standardized feature information characterizing the health status of the equipment—in particular, generating high-resolution vibration time-frequency feature maps, which can significantly highlight the transient and frequency domain evolution characteristics of faults in non-stationary signals. At the same time, the system integrates multiple signals such as temperature, displacement, and current with operating condition information to construct time series features combined with operating conditions, realizing in-depth mining of the dynamic characteristics of the equipment under complex operating conditions. This process completely eliminates the dependence on manual experience to extract features, significantly improves the identification and diagnostic accuracy of fault features, and effectively solves the problems of unclear fault location and high false alarm rate caused by insufficient feature extraction and weak data correlation in traditional methods.

[0078] (3) The intelligent status monitoring and fault diagnosis system and method for die-cutting and hot stamping equipment provided by this invention creatively constructs a hierarchical, lightweight, and collaborative AI model architecture at the intelligent diagnosis level. In the anomaly detection stage, an improved SimSiam unsupervised contrastive learning model is adopted, which only requires normal equipment operation data to complete training, perfectly solving the bottleneck of scarce fault samples in industrial sites. This model combines depthwise separable convolution and gradient cutoff training strategies, which significantly reduces computational complexity while maintaining high detection sensitivity, realizing lightweight deployment on edge devices and real-time, low-latency anomaly warning. In the fault diagnosis stage, a lightweight multi-classification convolutional neural network model based on the MobileNetV3 architecture is adopted to quickly and accurately classify and locate abnormal signals. This efficient collaborative mechanism of anomaly detection and fault diagnosis completely changes the traditional passive mode that relies on post-event maintenance, realizes true intelligent predictive maintenance, effectively prevents sudden downtime, and significantly reduces maintenance costs and production losses.

[0079] (4) The intelligent status monitoring and fault diagnosis system and method for die-cutting and hot stamping equipment provided by the present invention realizes centralized management, continuous optimization and iterative learning of data, models and knowledge through a cloud-based collaborative mechanism. The cloud is not only used to store historical data and training models, but also supports large language models based on the Transformer architecture. It can automatically generate a comprehensive diagnosis and operation and maintenance decision report containing fault location, type, level and specific maintenance suggestions based on real-time diagnosis results and historical maintenance records, which greatly improves the scientificity, operability and response efficiency of operation and maintenance decisions. Attached Figure Description

[0080] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the content of the embodiments of the present invention and these drawings without creative effort.

[0081] Figure 1 This is a schematic diagram of the intelligent status monitoring and fault diagnosis system for die-cutting and hot stamping equipment of the present invention;

[0082] Figure 2 This is a flowchart illustrating the intelligent status monitoring and fault diagnosis method for die-cutting and hot stamping equipment according to the present invention. Detailed Implementation

[0083] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0084] In the description of this invention, unless otherwise explicitly specified and limited, the terms "connected," "linked," and "fixed" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0085] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0086] In the description of this embodiment, the terms "upper," "lower," "left," and "right," etc., refer to the orientation or positional relationship shown in the accompanying drawings. They are used only for ease of description and simplification of operation, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. In addition, the terms "first" and "second" are used only for distinction in description and have no special meaning.

[0087] Die-cutting and hot stamping equipment, as core production equipment in the post-press processing field, functions primarily to precisely die-cut paper or cardboard and transfer and hot stamp metal foil through high-precision, high-pressure periodic stamping. The main unit of this equipment is crucial for achieving these functions, typically containing a complex transmission system and high-speed, heavy-duty bearing components. During operation, it must withstand impact loads of hundreds of tons while maintaining extremely high precision and stability. To ensure accuracy and stability during long-term high-speed operation, its core transmission components and platform are often designed within a closed wall panel structure.

[0088] Currently, the industry mainly relies on periodic manual inspections, single-point threshold alarms, or simple vibration monitoring for condition monitoring of this type of equipment. These traditional technologies have limited monitoring dimensions and incomplete coverage, making it difficult to effectively capture and warn of early, progressive failures of internal core components within enclosed structures. Data collection points are sparse and isolated, lacking collaborative analysis and in-depth mining of information from multiple physical quantities. This results in fault diagnosis heavily relying on human experience, making it impossible to achieve intelligent, predictive, and precise maintenance.

[0089] Current equipment maintenance primarily relies on a reactive approach of repairing after a failure occurs. The root cause lies in the lack of effective real-time condition monitoring methods. Traditional equipment lacks continuous monitoring sensors for vibration, temperature, and current, making it unable to detect subtle characteristics of early-stage faults such as bearing pitting and wear. Once a fault occurs, on-site troubleshooting by technicians is required, and the lack of automated diagnostic tools leads to long downtimes and high maintenance costs.

[0090] Furthermore, existing technologies suffer from severely inadequate predictive maintenance capabilities. Maintenance plans are often based on fixed cycles rather than actual wear and tear, which can easily lead to over-maintenance or sudden failures. Because existing methods rely on human experience for fault feature extraction, they struggle to perform in-depth time-frequency analysis of vibration signals and other data. Simultaneously, equipment operating parameters, environmental data, and sensor data are fragmented, lacking a comprehensive analytical model that integrates multi-source information, making it impossible to accurately assess the true health status of components.

[0091] Intelligent technologies face multiple obstacles in practical industrial deployments. On the one hand, equipment operates in a healthy state for extended periods, resulting in a scarcity of fault sample data. Furthermore, the data collected on-site is noisy and of poor quality, posing significant challenges to training data-driven AI models. On the other hand, deep learning models are computationally complex, while the computing power of on-site edge devices is limited and cannot handle the demands. Relying on cloud computing introduces unacceptable network latency, leading to poor real-time fault diagnosis and a high false alarm rate.

[0092] The intelligent status monitoring and fault diagnosis system and method for die-cutting and hot stamping equipment provided in this application have the following significant technological advancements and beneficial effects compared to existing technologies:

[0093] First, this application constructs a multi-dimensional sensor network covering key components of the die-cutting and hot stamping equipment's main unit, achieving indirect yet comprehensive and continuous signal capture of the internal operating status of the enclosed main unit. By strategically deploying various types of sensors, including vibration, temperature, angle, and pressure sensors, at key locations such as the flywheel side, paper output end, crankshaft end, elbow mechanism support shaft and base, and the top of the imprinting platform, the system effectively collects multi-source heterogeneous data on vibration, temperature, axial displacement, and pressure from core components such as the power shaft, transmission shaft, elbow mechanism, and imprinting platform. To further enhance the temporal correlation and state continuity of signal capture, the system also introduces a temporal acquisition window and a sliding acquisition mechanism to synchronously acquire and time-align multi-sensor data, thereby constructing temporal characteristic information that reflects the dynamic operation of the equipment. This design not only fundamentally solves the problem of unobservable internal states of enclosed main units but also provides a high-density, highly correlated data foundation for capturing weak and progressive fault characteristics such as early bearing pitting and support shaft wear, realizing a fundamental shift from post-fault repair to pre-fault perception and state tracking.

[0094] Secondly, this application creatively adopts a hierarchical, collaborative, and lightweight intelligent diagnostic architecture and algorithm. At the information processing level, the system utilizes techniques such as continuous wavelet transform to convert the original vibration signal into a high-resolution time-frequency feature map, and combines operating condition information to construct time-series features from multiple signals such as temperature, displacement, and angle, achieving deep fusion and feature enhancement of non-stationary signals and multi-source data. In the anomaly detection stage, the system employs an improved SimSiam unsupervised contrastive learning model. By constructing sample pairs under normal conditions and combining depthwise separable convolution and gradient cutoff training strategies, it achieves an efficient anomaly detection mechanism that requires only normal data for training. The model outputs an anomaly score as a state evaluation criterion and uses the P-quantile method to dynamically determine the anomaly judgment threshold, significantly improving the sensitivity and robustness of identifying unknown anomaly states. Through lightweight network design and localized deployment, the system achieves low-latency, highly reliable real-time anomaly warning on edge devices.

[0095] Furthermore, this application constructs a complete technology chain from anomaly detection to fault diagnosis and intelligent operation and maintenance. In the fault diagnosis stage, the system employs a lightweight multi-classification convolutional neural network model based on the MobileNetV3 architecture to quickly and accurately classify and locate labeled anomaly signals, achieving intelligent identification of fault type, location, and level. Combined with a cloud-based collaborative mechanism, the system supports continuous model optimization and data backtracking analysis, and introduces a large language model based on the Transformer architecture, capable of automatically generating structured and actionable maintenance suggestions and early warning reports based on real-time diagnostic results and historical operation and maintenance records. This end-to-end intelligent coverage from "state perception → anomaly detection → fault characterization → operation and maintenance decision-making" not only significantly improves the accuracy, response speed, and scientific nature of fault diagnosis, but also upgrades equipment maintenance from passive response and planned maintenance to predictive intelligent operation and maintenance based on the actual health status of the equipment, thereby minimizing unplanned downtime, reducing maintenance costs, and ensuring production continuity and product quality stability.

[0096] like Figure 1 As shown, this embodiment provides a method for intelligent status monitoring and fault diagnosis of die-cutting and hot stamping equipment, which includes the following steps:

[0097] S1. Collect working parameters of multiple key parts of the die-cutting and hot stamping equipment, construct a time-series acquisition window, and perform sliding acquisition of the working parameters to obtain characteristic information reflecting the equipment status.

[0098] This step aims to comprehensively and continuously perceive the status of critical components of the equipment through a high-density sensor network, and specifically includes the following steps:

[0099] S11. Sensor Deployment and Signal Acquisition: Deploy multiple types of sensors in key parts of the die-cutting and hot stamping equipment to form a collaborative monitoring network.

[0100] Specifically, a first vibration sensor and a first temperature sensor are arranged on the cam housing to collect its vibration acceleration and surface temperature.

[0101] A second set of vibration sensors and a second set of temperature sensors are respectively arranged on the flywheel side of the main drive shaft and the end cap at the paper output end to collect the radial vibration and temperature of the bearing areas at both ends; a first micro-displacement sensor is arranged along the axial direction on the end cap at the paper output end to collect the axial movement of the drive shaft.

[0102] A third vibration sensor and a third temperature sensor are installed at the half-shaft flange of the worm gear drive shaft of the main unit; an angle sensor is installed at the shaft end to collect the rotation angle (speed); a second micro-displacement sensor is arranged along the axial direction on the shaft end face to collect the axial movement of the drive shaft.

[0103] A fourth temperature sensor is placed on one side of the support shaft and the corresponding base of the elbow mechanism to monitor its operating temperature.

[0104] Ideally, a micro-deformation sensor should be placed at the top center of the imprinting platform to indirectly calculate the working pressure by measuring the micro-deformation of the platform.

[0105] A fourth vibration sensor is placed near the bearing of the main motor housing; the output torque or operating current is collected by a torque sensor or a current sensor.

[0106] Oil temperature sensor, flow sensor and abrasive sensor are installed in the lubricating oil sump of the lubrication system.

[0107] S12. Timing Acquisition and Window Construction: All sensors acquire data synchronously using a unified clock source and a set sampling frequency. The system constructs a timing acquisition window of fixed length and performs sliding acquisition on the signals of each channel to ensure that the acquired data are continuous time signal segments with temporal context.

[0108] To balance the continuity of data acquisition with the storage and computing resources of edge devices, this embodiment adopts a sliding acquisition strategy that combines time-triggered and event-triggered methods:

[0109] (1) Time-triggered continuous acquisition mode:

[0110] The system defaults to continuous sliding data acquisition at fixed time intervals. Preferably, the acquisition window length is set to an integer multiple of the rotation cycle of the key components of the equipment. For example, for the main drive shaft, the window length is set to 10 complete rotation cycles (approximately 0.5-2 seconds, depending on the rotation speed). Adjacent windows overlap by 50%-75% to ensure continuous tracking of fault characteristics. The window slides forward in fixed steps (e.g., 0.1 seconds) to achieve near real-time monitoring of the equipment status.

[0111] (2) Event-triggered encrypted data collection mode:

[0112] When the following trigger conditions are detected, the system automatically initiates high-frequency encrypted data acquisition, shortening the acquisition window length to 2-3 rotation cycles and increasing the sampling frequency to 2-5 times that of normal mode, in order to capture the high-frequency components of transient fault characteristics:

[0113] a. Significant changes occur in the equipment's operating conditions, such as sudden pressure changes on the imprinting platform exceeding a preset threshold (e.g., exceeding 20% ​​of the normal fluctuation range), a sudden increase in main motor current exceeding 15% of the rated current, or speed fluctuations exceeding ±5%;

[0114] b. A certain monitored physical quantity exceeds its healthy baseline by a certain multiple, such as the effective value of vibration exceeding 1.5 times the baseline value, or the rate of temperature change exceeding 2℃ / minute;

[0115] c. When the equipment reaches a critical action node, such as the moment each printing action is completed or when the cam mechanism passes the dead point, the system triggers short-term high-frequency acquisition to capture the periodic impact response;

[0116] d. Operators manually trigger diagnostic commands through the human-machine interface.

[0117] By constructing a time-series acquisition window and a sliding acquisition mechanism, the above steps not only acquire the instantaneous state but also capture the dynamic process of the state evolving over time. This provides crucial time-series information for subsequent analysis of transient faults, gradual wear, and operational condition correlations, overcoming the limitations of traditional single-point snapshot monitoring.

[0118] S13. Feature information extraction and standardization: Convert the original multi-source heterogeneous time series signals into standardized features that can be used for AI model analysis.

[0119] Among them, vibration signals are processed into time-frequency feature maps, and other signals are processed into time series features.

[0120] Specifically, the vibration signal processing into a time-frequency feature map includes the following steps:

[0121] a. Resampling: Based on the real-time rotational speed of the equipment (which can be obtained by an angle sensor) and the fault characteristic frequency, the vibration signals of the collected power shaft, transmission shaft or main motor are resampled in the angular domain so that each sample contains at least two complete rotational cycle information, eliminating the influence of rotational speed fluctuations.

[0122] b. Filtering: A bandpass filter with a frequency 0-3 times the maximum fault characteristic frequency of the bearing is used for filtering to retain the fault characteristic frequency band and suppress high-frequency noise and low-frequency interference.

[0123] c. Standardization: Perform Z-score standardization on the filtered signal to make its mean 0 and variance 1.

[0124] d. Continuous Wavelet Transform (CWT): Perform CWT on the standardized signal to generate a time-frequency plot. Plot time on the horizontal axis and scale (corresponding frequency) on the vertical axis; the color intensity represents the energy of the wavelet coefficients. CWT can provide good time and frequency domain resolution at the same time, making it very suitable for capturing transient characteristics of faults such as shock and modulation in non-stationary vibration signals.

[0125] The continuous wavelet transform method is as follows:

[0126]

[0127] In the formula, Indicates the original signal Continuous wavelet transform, Represents the original signal. Indicates the translation parameter. Indicates the scale parameter. This represents the wavelet function after translation and scaling. This is the normalization factor.

[0128] Continuous wavelet transform possesses adaptive resolution adjustment capabilities in the time and frequency domains, making it particularly suitable for analyzing non-stationary signals with transient, abrupt, or time-varying spectral characteristics. Its core advantage lies in its adaptability—it can analyze low-frequency persistent phenomena at a large scale, capture high-frequency transient events at a small scale, and is highly sensitive to signal singularities.

[0129] Specifically, other signal processing for time series features includes the following steps:

[0130] Signals such as temperature, displacement, angle, torque, current, flow rate, and abrasive particle concentration are aligned and combined with their corresponding operating conditions (such as acquisition timestamp, ambient temperature, continuous operating time of the equipment, current pressure of the imprinting platform, and worm gear shaft speed / production efficiency) to form a multi-dimensional time series feature vector. For example, the feature vector at a certain moment can be [time, support shaft temperature, base temperature, temperature difference, axial displacement, speed, pressure, current, ...].

[0131] The above steps effectively highlight the fault frequency components and their occurrence times in the vibration signal through the time-frequency feature map generated by CWT. Furthermore, the time-series features, which integrate operating condition information, deeply correlate the mechanical state of the equipment with factors such as operating load and environment, achieving a fusion representation of multiple physical quantities. This greatly enriches the descriptive dimensions of the health status and lays the foundation for accurate diagnosis.

[0132] S2. Based on the feature information, construct an anomaly detection model to detect anomalies in the equipment status, obtain an anomaly score, and determine whether the equipment status is abnormal based on the anomaly score.

[0133] This step is based on the improved SimSiam model for anomaly detection and score evaluation. It adopts an unsupervised learning approach that only requires normal samples to achieve sensitive perception of unknown abnormal states. Specifically, it includes the following steps:

[0134] S21. Model building and training, specifically including the following steps:

[0135] a. Data preparation: Collect a large amount of time-series feature information (time-frequency graphs and / or time series vectors) of the equipment under various healthy operating conditions as a training set.

[0136] b. Constructing an improved SimSiam model: This model improves upon the classic SimSiam structure by: ① Removing the data augmentation layer for images and directly inputting the original features; ② Employing dual-channel input, with each channel receiving a normal sample from the training set; ③ Using a dual-channel deep separable convolutional neural network where the two channels share network weights to extract high-dimensional feature vectors A1 and A2; ④ Using a support vector machine to map A1 and A2 to obtain B1 and B2 respectively.

[0137] c. Training Strategy: Employ gradient cutoff. In each training iteration, freeze (stop gradient) the encoder parameters of one channel (e.g., channel 2), and only update the encoder parameters of the other channel (channel 1) and the predictor parameters of both channels; the next iteration alternates between the two. This strategy effectively prevents model collapse.

[0138] d. Constructing the loss function: Using symmetric cosine similarity loss.

[0139]

[0140] in, This represents the loss value of the anomaly detection model. The feature vector output by the first layer encoder The feature vector output by the second-layer predictor cosine similarity, The feature vector output by the first layer encoder The feature vector output by the second-layer predictor The cosine similarity. The training objective is to maximize... That is, to make the feature representations of the same pair of normal samples as similar as possible after encoding and prediction in two channels.

[0141] The cosine similarity is calculated as follows:

[0142]

[0143] In the formula, For feature vectors and eigenvectors cosine similarity, For feature vectors The elements in For feature vectors The elements in.

[0144] S22. Calculation of outlier scores and determination of thresholds, specifically including the following steps:

[0145] a. Definition of outlier score: After training is completed, the outlier score is defined. In the formula These are abnormal scores. Add to the model's loss This is to avoid the case where the logarithmic operation variable is zero. When the input is a pair of highly similar healthy samples, Approaching 1, The closer to 0, the lower the similarity. The smaller, The larger.

[0146] b. Threshold Determination (P-quantile Method): A set of independent health validation sets is input into the trained model to obtain a threshold. The anomaly detection threshold is determined using the P-quantile method. During the training phase, the overall distribution of the model's output anomaly scores on the validation set is examined, allowing for a certain number of anomalous sample points. A P-quantile is set based on the overall distribution of the output anomaly scores on the validation set, and a threshold is determined using this quantile to classify sample points of P% as normal samples.

[0147] S23. Online detection:

[0148] The test sample, collected in real time, is paired with a baseline sample randomly selected from a healthy database, and then input into the trained anomaly detection model. The calculated... If the preset threshold is exceeded, the current state is determined to be abnormal and a fault diagnosis process is triggered; otherwise, it is determined to be normal.

[0149] This method requires only healthy data for training, perfectly solving the problem of scarce fault samples in industrial settings. The improved SimSiam combined with DS-CNN achieves a lightweight model, suitable for edge deployment. Anomaly scores provide a continuous quantitative indicator of state degradation, which is more refined than binary judgment; the threshold determined by the P-quantile method has statistical robustness and can adapt to the normal fluctuation range of different devices.

[0150] S3. For feature information that is judged to be abnormal, a fault diagnosis model is constructed based on the fault type to which it belongs, and fault diagnosis is performed on the equipment status to generate diagnosis results.

[0151] This step is based on MobileNetV3's fault diagnosis and classification. After step S2 detects an anomaly, this step accurately locates and characterizes the anomaly, specifically including the following steps:

[0152] S31. Model building and training, specifically including the following steps:

[0153] a. Dataset Construction: Collect historical data, including characteristic information of various known fault types (such as bearing outer ring pitting, inner ring wear, gear tooth breakage, poor lubrication, etc.) at the time of occurrence, and assign detailed fault category labels (e.g., "power shaft - flywheel side - deep groove ball bearing - outer ring pitting - early stage"). Together with a large amount of health data, this constitutes a labeled equipment operating status dataset.

[0154] b. Model architecture construction: The lightweight MobileNetV3 network is used as the basic framework and modified. Its depthwise separable convolutional blocks are used as the backbone feature extractor (encoder), followed by global average pooling layers, fully connected layers, and softmax classification layers.

[0155] c. Training: Divide the dataset into training, validation, and test sets in a 7:2:1 ratio. Use the training set with the cross-entropy loss function. To optimize the objective, the model is trained using backpropagation and gradient descent algorithms, enabling it to accurately distinguish between different fault categories. This represents the loss value of the fault diagnosis model. Number of fault categories For real labels, To predict probabilities.

[0156] S32. Online Diagnosis: Input the feature information (time-frequency graph or time series vector) marked as abnormal into the trained fault diagnosis model. The model outputs a probability distribution vector, where each position corresponds to the confidence level of a fault category. The category with the highest confidence level is taken as the final diagnosis result, for example, outputting "Diagnosis result: Drive shaft - worm gear shaft end - tapered roller bearing - insufficient lubrication - abnormal temperature rise".

[0157] MobileNetV3 is a lightweight network designed for mobile devices. Combined with depthwise separable convolutions, it significantly reduces the number of parameters and computational cost while maintaining high classification accuracy, meeting the real-time requirements of edge devices. The above steps bridge the gap between anomaly detection and fault identification, providing direct evidence for precise maintenance.

[0158] S4. Based on the diagnostic results, generate a comprehensive diagnostic and maintenance decision report.

[0159] This step transforms the diagnostic results into executable maintenance instructions. Specifically, the operations and maintenance decision module incorporates a Large Language Model (LLM) based on the Transformer architecture. This LLM is trained to learn from historical fault reports, maintenance work orders, expert knowledge, and other textual data. Upon receiving the diagnostic results (faulty component, type, level) and current operating information (runtime, load, etc.) from the intelligent diagnostic module, the LLM automatically generates a structured "Comprehensive Diagnosis and Operations and Maintenance Decision Report."

[0160] The report should include, but is not limited to: ① Fault summary: time, location, type, severity level; ② Cause analysis: brief explanation of the mechanism; ③ Impact assessment: potential impact on production and safety; ④ Maintenance recommendations: specific operating procedures, required tools, spare parts, estimated working hours and personnel requirements; ⑤ Early warning reminder: suggested next inspection time or remaining safe operating time.

[0161] The report is sent in real time to managers' PCs, mobile apps, or workshop dashboards via cloud push unit. Simultaneously, all data from this incident (raw signals, features, diagnostic results, and reports) is uploaded to the cloud storage unit to enrich the training dataset. The cloud computing unit can periodically use new data to incrementally learn or retrain the anomaly detection and fault diagnosis models, achieving continuous model optimization.

[0162] By automatically generating professional and easy-to-understand maintenance reports through large language models, technical diagnostic results are directly transformed into actionable work instructions, greatly improving the efficiency of operation and maintenance response and the scientific nature of decision-making. Cloud collaboration enables closed-loop data flow and continuous knowledge accumulation, giving the system self-evolution capabilities.

[0163] On the other hand, this embodiment provides an intelligent status monitoring and fault diagnosis system for die-cutting and hot stamping equipment, including:

[0164] The signal acquisition module consists of sensors and data acquisition units deployed in various key parts of the equipment, and is used to collect working parameters of multiple key parts of the die-cutting and hot stamping equipment.

[0165] The information processing module receives the raw data stream from the signal acquisition module and processes the operating parameters to obtain characteristic information reflecting the equipment status.

[0166] The intelligent diagnostic module, which is the core AI analysis engine, is deployed on an edge computing gateway or industrial control computer close to the device. It is used to detect anomalies and diagnose faults in the equipment status based on the feature information and generate diagnostic results.

[0167] The operation and maintenance decision module, deployed in the cloud or on a factory-level server, is used to generate a comprehensive diagnosis and operation and maintenance decision report by calling a large language model based on the Transformer architecture and combining it with historical maintenance records in the knowledge base, based on historical fault information and the diagnostic results.

[0168] This embodiment first deploys multiple types of sensors at key functional parts of the equipment through a signal acquisition module, forming a collaborative data acquisition network. This network can continuously and comprehensively collect multi-dimensional operating parameters reflecting the operational status of the equipment's core components, thus overcoming the limitations of traditional single-point monitoring or manual inspection and providing a solid data foundation for subsequent intelligent analysis. Next, the information processing module fuses and extracts features from the collected multi-source heterogeneous raw data. It converts the raw signals into standardized feature information that can characterize the equipment's health status and evolution trends, such as time-frequency feature maps and time series features. This process effectively uncovers deep correlations and subtle early signs of faults in the data. Subsequently, the intelligent diagnosis module, based on the extracted feature information, uses a hierarchical artificial intelligence model to intelligently analyze the equipment's status. This module first achieves efficient anomaly information discovery through an anomaly detection model, and then uses a fault diagnosis model to accurately diagnose and classify the anomaly information. Finally, the operation and maintenance decision module automatically generates actionable early warning prompts and maintenance suggestions based on the results output by the intelligent diagnosis module. This changes the traditional experience-based passive maintenance model and drives the shift in operation and maintenance strategies towards data-driven predictive maintenance, thereby significantly improving the reliability and safety of equipment operation.

[0169] Furthermore, the signal acquisition module includes at least one of the following acquisition units:

[0170] The cam acquisition unit, including a first vibration sensor and / or a first temperature sensor, is arranged at the cam housing and is used to acquire vibration and temperature information at the cam.

[0171] The power shaft acquisition unit includes a set of second vibration sensors, a set of second temperature sensors, and / or a first micro-displacement sensor, which are respectively arranged at the end caps near the bearings at both ends of the power shaft to acquire vibration, temperature, and / or axial displacement information of the power shaft.

[0172] The transmission shaft acquisition unit includes a third vibration sensor, a third temperature sensor, an angle sensor, and / or a second micro-displacement sensor. The third vibration sensor and / or the third temperature sensor are adapted to be arranged at the half-shaft flange of the main machine worm gear transmission shaft to acquire vibration and / or temperature information of the transmission shaft. The angle sensor is adapted to be set at the shaft end of the transmission shaft to acquire the rotation angle of the transmission shaft. The second micro-displacement sensor is adapted to be arranged axially on the shaft end face of the transmission shaft to acquire the axial movement of the transmission shaft.

[0173] The data acquisition unit for the support shaft or elbow mechanism includes multiple fourth temperature sensors arranged on one side of the elbow mechanism support shaft and base to collect temperature information of the support shaft or elbow mechanism support shaft and base.

[0174] The platform acquisition unit, including a micro-deformation sensor, is located at the top center of the die-cutting stationary platform and is used to acquire pressure information from the imprinting platform.

[0175] The main motor acquisition unit includes a fourth vibration sensor and / or torque sensor and current sensor arranged in the housing near the main motor output shaft bearing, for acquiring torque, current and / or vibration information of the main motor.

[0176] The lubrication system acquisition unit, including an oil temperature sensor, a flow sensor, and / or an abrasive sensor, is arranged in the lubricating oil sump to collect information on the temperature, flow rate, and / or abrasive particles of the lubricating oil.

[0177] Each of the above units is responsible for converting physical signals into digital signals, synchronously acquiring them according to a unified timing clock, and outputting the raw working parameter data stream.

[0178] Specifically, the cam acquisition unit is used to collect vibration and temperature information at the cam. Sensors are installed at locations where the cam vibration status can be accurately collected to identify the cam's vibration state. When the vibration value exceeds a threshold range, it is considered that the cam is about to malfunction or has already malfunctioned. The unit is also set to collect the temperature of the cam during operation. The temperature sensor identifies the temperature value to determine whether there is an oil temperature rise problem, which may be due to a blockage in the oil delivery pipeline or a malfunction in the cam structure. The information can be combined with that from other acquisition units to make a definitive diagnostic conclusion.

[0179] The power shaft acquisition unit is used to collect vibration, temperature, and / or axial displacement information of the power shaft. The die-cutting power shaft has numerous bearings, making it a frequent site of failure. When a failure occurs at any point, its vibration signal is transmitted along the shaft. Monitoring the vibration signal can reflect the fault characteristics of the shaft and bearings during operation. Since the die-cutting power shaft and the die-cutting drive shaft are connected by a worm gear transmission mechanism, the die-cutting power shaft is subjected to axial force during power transmission, which may cause axial movement. Simultaneously, abnormal friction caused by the failure leads to localized temperature increases; monitoring temperature changes can indicate the location of the fault. Because the portion of the die-cutting power shaft inside the main unit is completely immersed in lubricating fluid, vibration signal sensors cannot operate in lubricating fluid. Therefore, vibration signal sensors are installed radially at the flywheel side end cap and the outer end cap of the paper output end of the main unit, and a micro-displacement sensor is installed axially at the outer end cap of the paper output end of the main unit. Temperature sensors are installed at the flywheel side end cap and the inner end cap of the paper output end of the main unit. This is used to monitor the operating information of multiple bearings and shafts on the die-cutting power shaft.

[0180] The drive shaft acquisition unit is used to collect vibration, temperature, angle, and / or axial displacement information of the drive shaft. The die-cutting drive shaft transmits power from the die-cutting power shaft to the imprinting platform, and the rotational frequency of the die-cutting drive shaft is the same as the operating frequency of the imprinting platform. Therefore, in addition to installing vibration signal sensors and temperature sensors at the half-shaft flange of the die-cutting drive shaft, an angle sensor is installed at the shaft end of the die-cutting drive shaft to monitor the operating information of the bearings, shafts, and imprinting platform on the die-cutting drive shaft. Furthermore, since the die-cutting power shaft and the die-cutting drive shaft are connected by a worm gear transmission mechanism, an axial force is applied to the die-cutting drive shaft during power transmission, causing axial movement. Prolonged operation can lead to severe wear between the shaft shoulder and the base, resulting in failure. Therefore, a micro-displacement sensor is installed axially on the end face of the die-cutting drive shaft to monitor its axial displacement.

[0181] The toggle mechanism data acquisition unit is used to collect temperature information from the toggle mechanism's support shaft and base. As crucial support components for the imprinting platform, the toggle mechanism's support shaft and base are prone to failure due to insufficient lubrication or excessive pressure leading to wear. When the support shaft fails, the primary symptom is an abnormal temperature increase. Therefore, temperature sensors are installed at both the toggle mechanism's support shaft and base to monitor the mechanism's operational information.

[0182] The platform acquisition unit is used to collect pressure information from the embossing platform. The embossing platform is the main platform for the die-cutting machine to complete the die-cutting work. Because the upper and lower platforms need to install cutting tools and rapidly feed paper during operation, it is impossible to directly monitor the pressure value on the platform's working surface. Therefore, a micro-deformation sensor is installed at the center of the upper part of the platform to calculate the pressure information during the embossing platform's operation. Simultaneously, the embossing platform operates cyclically, and the working frequency and pressure information can be comprehensively used to determine the current working status of the embossing platform.

[0183] The main motor acquisition unit is used to collect torque, current, and / or vibration information of the main motor. As the power source of the die-cutting machine, the torque and current information of the main motor reflects the operating status of the equipment under load. Therefore, the motor's current and torque are important indicators of the main motor's operating condition. At the same time, the operating status of the main motor's output shaft and bearings is also an important indicator of the main motor's operating condition. Vibration sensors are installed near the bearings to monitor the health of the main motor's output shaft and bearings.

[0184] The lubrication system data acquisition unit is used to collect information on the temperature, flow rate, and / or abrasive particles of the lubricating oil. Abnormalities in the lubrication system can accelerate wear in the transmission system. Monitoring the lubricating oil flow rate can reflect whether the lubrication system is functioning properly. The operation of the lubrication system directly affects the smooth operation of the transmission system. When a transmission system malfunctions, the lubricating oil temperature typically rises due to abnormal wear at the fault location. The abrasive particles detached due to abnormal wear are also important information for identifying early-stage faults. Therefore, collecting information on lubricating oil temperature and abrasive particles within the lubricating oil is crucial for monitoring whether abnormalities have occurred in the transmission system.

[0185] It should be noted that the signal types monitored by the above acquisition units (such as vibration, temperature, displacement, etc.) can be combined in an "and / or" manner in actual implementation. That is, each acquisition unit can select to acquire one or more signals according to actual monitoring needs. For example, the power shaft acquisition unit can acquire only vibration signals, or it can acquire any two or all of vibration, temperature, and axial displacement simultaneously.

[0186] Meanwhile, the signal acquisition module includes acquisition data from the power shaft, drive shaft, support shaft, platform, main motor, and lubrication system. These signals from different locations can be combined in various ways. For example, signals from the power shaft can be combined with signals from the platform to determine whether the overall machine operation quality can be guaranteed; signals from the lubrication system can be combined with signals from the main motor to determine whether the main motor can operate continuously and normally; signals from the support shaft, combined with signals from the platform and main motor, can be used to determine if insufficient platform pressure is caused by wear, or if the main motor's operating pressure is increasing. Therefore, combining signals from different locations can comprehensively assess the problem and provide early warnings; this is not limited to combining all signals from all locations.

[0187] Furthermore, the information processing module includes:

[0188] The signal reading and transmission unit is used to read the signals collected by each unit of the signal acquisition module at a set frequency. This unit reads the buffered data of each acquisition unit at the set frequency and performs packetization and timestamp alignment.

[0189] The feature extraction unit is used to convert the read signals into time-frequency feature maps and / or time-series features combined with operating condition information for diagnostic purposes. This unit performs the operation of step S13 in the method embodiment, converting the raw data stream into a standardized time-frequency feature map and time-series feature vector. This unit is also responsible for managing the sliding and updating of the time-series acquisition window.

[0190] The signal acquisition module's sensors collect signals at a set sampling frequency and temporarily store them in the corresponding acquisition unit's temporary storage with a specific data length. The feature extraction unit extracts the operating signals collected by the signal acquisition module from the temporary storage and processes them. When reading the operating signals, the feature extraction unit locks the temporary storage, thereby reading the operating signals stored in the temporary storage before the locking time. The signal reading and transmission unit transmits the processed information to the intelligent diagnostic module and the cloud collaboration module via Ethernet.

[0191] Furthermore, this system also includes a cloud-based collaboration module for providing computing power and storage support. An intelligent diagnostic module is connected to the cloud-based collaboration module to deploy an artificial intelligence model, receive operational information from the information processing module, and analyze the operational information through the artificial intelligence model to intelligently diagnose the equipment's operational status. The cloud-based collaboration module includes:

[0192] The cloud computing unit provides powerful GPU computing power for training and optimizing the artificial intelligence model used by the intelligent diagnostic module.

[0193] The cloud storage unit stores equipment operation information, characteristic information, historical fault maintenance information, artificial intelligence models and their weight data, forming an enterprise equipment health management database. Specifically, the cloud storage unit stores operation information including vibration, temperature, pressure, displacement, torque, current, flow rate, and abrasive particle data collected by various acquisition units of the signal acquisition module over time; historical fault maintenance information includes fault time, fault location, fault type, number of maintenance personnel, maintenance method, maintenance hours, and whether any new types of faults were discovered; and the artificial intelligence model and its weight information includes the training source code, execution source code, and weight data of each node within the trained artificial intelligence model.

[0194] The cloud push unit, acting as the information distribution hub, generates comprehensive diagnostic and operational decision reports based on the operation and maintenance decision module. This reports form operational status information, early warning alerts, and / or maintenance suggestions, which are then pushed to the user terminal to display the equipment's operational status to enterprises and users. The operational status information includes operating condition information, working time, and time-varying graphs of data from various monitoring points. Early warning alerts include the warning time, operating condition information, warning reason, fault time, fault location, fault type, and estimated remaining usable time. Maintenance suggestions include recommended maintenance methods, recommended maintenance personnel and number of personnel, and recommended maintenance time.

[0195] Maintenance personnel can perform timely and accurate equipment maintenance based on the operational status information, early warning information, and maintenance suggestions released by the cloud push unit. Simultaneously, maintenance personnel need to periodically add continuously collected new data to the training database of the artificial intelligence model and train the model, enabling it to continuously adapt to changes in equipment operating status, improve diagnostic accuracy, and reduce false alarm rates.

[0196] In this embodiment, the information output by the operation and maintenance decision module can be published to users and enterprises through the cloud push unit. Operation and maintenance personnel can promptly check and maintain the equipment based on the published information, avoiding sudden downtime caused by equipment failure, reducing safety hazards, and saving equipment maintenance costs.

[0197] Specifically, the operation and maintenance decision module adopts a large language model based on the Transformer architecture to generate a comprehensive diagnostic and operation and maintenance decision report based on diagnostic results and historical maintenance data. The comprehensive diagnostic and operation and maintenance decision report includes at least one of the following: fault location, fault type, fault level, early warning prompts or maintenance recommendations.

[0198] Furthermore, the intelligent diagnostic module includes:

[0199] The anomaly detection unit is used to deploy an anomaly detection algorithm based on the improved SimSiam model. It calculates anomaly scores based on feature information and determines whether the equipment's operating status is abnormal. If an anomaly is found, its corresponding fault type is marked.

[0200] The fault diagnosis unit is used to deploy a fault diagnosis model based on the MobileNetV3 framework. When an anomaly is detected, it diagnoses the fault type and location and generates diagnostic results.

[0201] During equipment operation, the signal acquisition module continuously acquires data, which is then converted into feature information by the information processing module. The intelligent diagnostic module performs real-time anomaly detection and fault diagnosis. Once a fault is diagnosed, the diagnostic results are uploaded to the operation and maintenance decision-making module, which automatically generates a report and notifies relevant personnel via cloud push. Simultaneously, this data loop is fed back to the cloud for system iteration and optimization.

[0202] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will be able to make various obvious changes, readjustments, and substitutions without departing from the scope of protection of the present invention. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent status monitoring and fault diagnosis of die-cutting and hot stamping equipment, characterized in that, Includes the following steps: The system collects working parameters from multiple key components of the die-cutting and hot stamping equipment. These working parameters include multiple physical quantities distributed along the equipment's power transmission path and load-bearing structure. The physical quantities include at least one or more of vibration, temperature, displacement, angle, pressure, torque, current, flow rate, and abrasive particles. A time-series acquisition window is constructed to perform sliding acquisition of the working parameters, thereby obtaining characteristic information reflecting the equipment's status. Based on the aforementioned feature information, an anomaly detection model is constructed to detect anomalies in the equipment status. The anomaly detection model is trained using an unsupervised contrastive learning approach, utilizing only the feature information of the equipment under normal operating conditions as training samples. Positive sample pairs are constructed and input into the contrastive learning network for training to learn the feature representation of the normal state. The contrastive learning network adopts a dual-channel symmetric structure, and a loss function is constructed by calculating the cosine similarity between the feature vectors output by the two channels and the predicted vector to characterize the directional consistency of positive sample pairs in the feature space. Based on the trained anomaly detection model, anomaly detection is performed on the real-time collected feature information to be tested, an anomaly score is calculated and compared with a preset anomaly judgment threshold. If the anomaly score exceeds the anomaly judgment threshold, the equipment status is determined to be abnormal. The anomaly score is constructed based on the loss function and is obtained by performing a nonlinear transformation on the loss value, which is used to quantify the degree of deviation between the current state and the normal state. For feature information that is judged to be abnormal, a fault diagnosis model is constructed based on the fault type to which it belongs, fault diagnosis is performed on the equipment status, and diagnostic results are generated. The specific process of constructing an anomaly detection model and determining whether the equipment status is abnormal includes: Use the feature information of the die-cutting and hot stamping equipment under normal operating conditions as training samples; The training samples are constructed into multiple positive sample pairs; Each positive sample pair is input into an improved SimSiam model for training; the improved SimSiam model uses a deep separable convolutional neural network as an encoder, a support vector machine as a predictor, and alternately updates the gradients of the two channels using a gradient cutoff method. Based on the output of the model to healthy sample pairs during the training phase, anomaly scores are constructed using loss values, and the anomaly determination threshold is determined using the P-quantile method. The anomaly determination threshold using the P-quantile method includes: inputting an independent healthy validation set into the trained anomaly detection model to obtain an anomaly score set, statistically analyzing the distribution of the anomaly score set, and selecting the anomaly score corresponding to a preset quantile as the anomaly determination threshold. During online detection, based on the real-time collected feature information to be tested, a healthy baseline sample is selected from the healthy sample library composed of feature information under normal operation according to preset rules, and the feature information to be tested and the healthy baseline sample are combined to form an anomaly detection model trained by the input of the sample to be tested. The anomaly score of the test sample pair is calculated and compared with the anomaly judgment threshold determined by the P quantile method. If the anomaly score exceeds the anomaly judgment threshold, the equipment status is determined to be abnormal, and the test feature information is entered into the fault diagnosis model for processing as anomaly feature information. The loss value of the anomaly detection model is obtained by symmetrically combining the cosine similarity term between the first feature vector A1 output by the first channel encoder and the second prediction vector B2 output by the second channel predictor, and the cosine similarity term between the second feature vector A2 output by the second channel encoder and the first prediction vector B1 output by the first channel predictor. This results in a smaller loss value when the feature representations of the healthy sample pair in the two channels tend to be consistent. The cosine similarity is calculated as follows: the inner product is obtained by multiplying the corresponding elements of the two vectors involved in the calculation, and the inner product is divided by the product of the magnitudes of the two vectors to characterize the degree of consistency in the directions of the two vectors. The anomaly score is constructed based on the loss value, and the anomaly score is the logarithmic result of summing the loss value with a preset small amount.

2. The intelligent status monitoring and fault diagnosis method for die-cutting and hot stamping equipment according to claim 1, characterized in that, The steps for collecting the operating parameters include at least one of the following steps: Vibration acceleration and / or surface temperature information are collected on the cam housing; Radial vibration and / or temperature information is collected in the bearing area on the flywheel side of the main drive shaft and / or at the paper output end cap. The axial movement of the power shaft is collected at the paper output end cap of the main drive shaft. Vibration and / or temperature information is collected at the half-shaft flange of the main worm gear drive shaft; The rotation angle is collected at the end of the worm gear drive shaft of the main unit; The axial movement of the drive shaft is collected on the shaft end face distributed along the axial direction of the main worm gear drive shaft; Operating temperature information is collected on one side of the support shaft and the corresponding base of the elbow mechanism; Calculate the working pressure by measuring micro-deformation at the top of the printing platform; The output torque and / or operating current are collected from the bearings in the main motor housing. Collect information on the temperature, flow rate, and / or abrasive particles of the lubricating oil in the lubricating oil sump of the lubrication system.

3. The intelligent status monitoring and fault diagnosis method for die-cutting and hot stamping equipment according to claim 1, characterized in that, The steps for obtaining the feature information include: The vibration signal in the operating parameters is converted into a time-frequency characteristic map; The temperature, displacement, angle, torque, current, flow rate and / or abrasive information in the working parameters are combined with the working condition information and converted into time series features; The operating condition information includes at least one of the following: time, initial temperature, ambient temperature, equipment power-on time, equipment operating time, imprinting platform pressure, shaft speed, or production efficiency.

4. The intelligent status monitoring and fault diagnosis method for die-cutting and hot stamping equipment according to claim 3, characterized in that, Converting the vibration signal into a time-frequency feature map includes the following specific steps: The acquired vibration signal is resampled based on the rotational speed and sampling frequency; The resampled samples are filtered to eliminate the influence of high-frequency noise on the fault characteristics. The filtering of the resampled vibration signal includes: determining a preset fault characteristic frequency band based on the fault characteristic frequency of the bearing or transmission component, and using a bandpass filter to retain the signal components within the fault characteristic frequency band to suppress high-frequency noise and low-frequency interference. The filtered samples are standardized so that their mean is 0 and their variance is 1. The standardized samples are subjected to continuous wavelet transform to obtain the time-frequency characteristic map of the vibration signal; the continuous wavelet transform of the standardized samples includes the following steps: Based on the mother wavelet function, scale and translation parameters are set, and the mother wavelet function is scaled and translated to form a family of wavelet functions; Wavelet coefficients are calculated for the standardized samples and the wavelet function family, and the wavelet coefficients are normalized using a normalization factor. A time-frequency characteristic map of the vibration signal is constructed based on the amplitude or energy of the wavelet coefficients, wherein the time-frequency characteristic map is plotted with time as the horizontal axis and scale or frequency as the vertical axis.

5. The intelligent status monitoring and fault diagnosis method for die-cutting and hot stamping equipment according to claim 1, characterized in that, The specific process of constructing a fault diagnosis model includes: Using feature information that includes normal operating status data and fault operating status data, the equipment operating status dataset is constructed by classifying the faults according to their categories. The equipment operation status dataset is divided into a training set, a validation set, and a test set; Based on the MobileNetV3 model, a multi-classification model is constructed with a deep separable convolutional neural network as the encoder and fully connected layers and Softmax layers as classifiers. The model is trained using the training set, and its parameters are optimized using the cross-entropy loss function so that it can identify and classify the fault categories of the input signal.

6. The intelligent status monitoring and fault diagnosis method for die-cutting and hot stamping equipment according to claim 1, characterized in that, It also includes the following steps: Based on the diagnostic results, a comprehensive diagnostic and maintenance decision report is generated; The comprehensive diagnosis and maintenance decision report is generated based on the large language model of the Transformer architecture, according to the diagnosis results and historical maintenance data. The comprehensive diagnosis and maintenance decision report includes at least one of the following: fault location, fault type, fault level, early warning prompt or maintenance suggestion.

7. An intelligent status monitoring and fault diagnosis system for die-cutting and hot stamping equipment, characterized in that, The intelligent status monitoring and fault diagnosis method for die-cutting and hot stamping equipment as described in any one of claims 1-6 includes: The signal acquisition module is used to acquire working parameters of multiple key parts of the die-cutting and hot stamping equipment. The working parameters include multiple physical quantities distributed along the power transmission path and load-bearing structure of the equipment. The physical quantities include at least one or more of vibration, temperature, displacement, angle, pressure, torque, current, flow rate, and abrasive particles. The information processing module is used to process the working parameters, construct a time-series acquisition window and perform sliding acquisition to obtain feature information reflecting the equipment status; The intelligent diagnostic module is used to perform anomaly detection and fault diagnosis on the equipment status based on the aforementioned feature information, and generate diagnostic results. The anomaly detection is implemented using an unsupervised contrastive learning approach, utilizing only feature information from the equipment's normal operating state as training samples. Positive sample pairs are constructed and input into the contrastive learning network for training, learning the feature representation of the normal state. The contrastive learning network employs a dual-channel symmetrical structure, constructing a loss function by calculating the cosine similarity between the feature vectors output from the two channels and the predicted vector, to characterize the directional consistency of the positive sample pairs in the feature space. Based on the trained anomaly detection model, anomaly detection is performed on the real-time collected feature information, calculating an anomaly score and comparing it with a preset anomaly judgment threshold. If the anomaly score exceeds the anomaly judgment threshold, the equipment status is determined to be abnormal. The anomaly score is constructed based on the loss function, obtained by performing a nonlinear transformation on the loss value, and is used to quantify the degree of deviation between the current state and the normal state.

8. The intelligent status monitoring and fault diagnosis system for die-cutting and hot stamping equipment according to claim 7, characterized in that, The signal acquisition module includes at least one of the following acquisition units: A cam acquisition unit, including a first vibration sensor and / or a first temperature sensor, is arranged at the cam housing to acquire vibration and / or temperature information at the cam. The power shaft acquisition unit includes a set of second vibration sensors, a set of second temperature sensors and / or a first micro-displacement sensor, which are respectively arranged at the end caps of both ends of the power shaft near the bearings, for acquiring vibration, temperature and / or axial displacement information of the power shaft; The transmission shaft acquisition unit includes a third vibration sensor, a third temperature sensor, an angle sensor, and / or a second micro-displacement sensor. The third vibration sensor and / or the third temperature sensor are adapted to be arranged at the half-shaft flange of the main machine worm gear transmission shaft to acquire vibration and / or temperature information of the transmission shaft. The angle sensor is adapted to be set at the shaft end of the transmission shaft to acquire the rotation angle of the transmission shaft. The second micro-displacement sensor is adapted to be arranged axially on the shaft end face of the transmission shaft to acquire the axial movement of the transmission shaft. The elbow mechanism acquisition unit includes multiple fourth temperature sensors arranged on one side of the elbow mechanism support shaft and base to collect temperature information of the elbow mechanism support shaft and base. The platform acquisition unit, including a micro-deformation sensor, is located at the top center of the imprinting platform to acquire pressure information of the imprinting platform. The main motor acquisition unit includes a fourth vibration sensor and / or torque sensor and a current sensor arranged in the housing near the bearing of the main motor output shaft, for acquiring torque, current and / or vibration information of the main motor; The lubrication system acquisition unit, including an oil temperature sensor, a flow sensor, and / or an abrasive sensor, is arranged in the lubricating oil sump to collect information on the temperature, flow rate, and / or abrasive particles of the lubricating oil.

9. The intelligent status monitoring and fault diagnosis system for die-cutting and hot stamping equipment according to claim 8, characterized in that, The information processing module includes: The signal reading and transmission unit is used to read the signals collected by each unit of the signal acquisition module at a set frequency. The feature extraction unit is used to convert the read signal into a time-frequency feature map for diagnosis and / or a time-series feature combined with operating condition information.

10. The intelligent status monitoring and fault diagnosis system for die-cutting and hot stamping equipment according to claim 7, characterized in that, The system also includes a cloud-based collaboration module, which comprises: A cloud computing unit is used to train and optimize the artificial intelligence model used by the intelligent diagnostic module; The cloud storage unit is used to store equipment operation information, feature information, historical fault maintenance information, artificial intelligence models and their weight data; The system also includes: The operation and maintenance decision module is used to generate a comprehensive diagnosis and operation and maintenance decision report based on historical fault information and the diagnostic results. The cloud push unit is used to push the comprehensive diagnostic and operation and maintenance decision report generated by the operation and maintenance decision module to the user terminal; The operation and maintenance decision module adopts a large language model based on the Transformer architecture to generate a comprehensive diagnosis and operation and maintenance decision report based on the diagnostic results and historical maintenance data. The comprehensive diagnosis and operation and maintenance decision report includes at least one of the following: fault location, fault type, fault level, early warning prompt or maintenance suggestion.

11. The intelligent status monitoring and fault diagnosis system for die-cutting and hot stamping equipment according to claim 7, characterized in that, The intelligent diagnostic module includes: The anomaly detection unit is used to deploy an anomaly detection algorithm based on the improved SimSiam model, calculate anomaly scores based on feature information, and determine whether the device's operating status is abnormal. The fault diagnosis unit is used to deploy a fault diagnosis model based on the MobileNetV3 framework, and to diagnose the fault type and location when an anomaly is detected.

Images