Data monitoring method for circular knife film cutting machine for flexible cover plate machining

Abstract

The invention relates to the technical field of data monitoring, in particular to a circular knife film cutting machine data monitoring method for flexible cover plate machining, which comprises the following steps: acquiring physical quantities detected by various sensors and product surface images to form multi-source time sequence data; constructing an equipment state map based on a physical causal relationship in the multi-source time sequence data, and extracting an equipment state spatio-temporal feature vector; inputting the spatial-temporal feature vector into a parallel prediction structure to obtain a processing quality prediction result, and generating a hidden space manifold representation of the current processing state; by calculating the distance between the fault prototype manifold and each preset fault prototype manifold, the fault prototype manifold is classified to the fault type corresponding to the fault prototype closest to the fault prototype manifold. A more reliable and intelligent solution is provided for guaranteeing the machining quality and stability of the flexible cover plate, and the pursuit of machining of the flexible cover plate for the ultimate quality and efficiency is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data monitoring. More particularly, the present application relates to a data monitoring method for a circular knife film cutting machine for flexible cover plate processing. BACKGROUND

[0002] As the core component of flexible electronic devices such as foldable screen mobile phones, the processing precision and quality of flexible cover plates directly determine the performance and reliability of the terminal product. Circular knife film cutting is a key process in the forming process of flexible cover plates. This process involves the use of a high-speed rotating knife roller to accurately punch a multi-layer composite film material. The process window is narrow, and the stability of the equipment state and process parameters is extremely high. In actual production, multiple factors such as knife roller drive motor, knife die pressure, material belt tension, and tool wear are dynamically coupled and interact with each other. Small parameter shifts or equipment abnormalities can cause processing defects such as burrs, glue overflow, delamination, and size out-of-tolerance, resulting in serious yield loss. Traditional quality control methods mainly rely on periodic sampling and offline measurement by human beings. This approach has a lagging response and cannot detect problems in real time. It is highly dependent on the experience of operators and is difficult to cope with complex and changing production conditions. It is impossible to achieve real-time, accurate monitoring and active intervention of the processing process.

[0003] With the development of intelligent manufacturing technology, researchers have begun to use sensors and data-driven models to monitor the film cutting process. However, existing technologies still have many limitations. First, in terms of data fusion, the film cutting process produces multi-source heterogeneous data including motor current, pressure, tension, and visual images. Existing methods usually simply concatenate or independently process these data, ignoring the internal coupling relationship between physical quantities based on device structure and process mechanism, making it difficult to fully and accurately represent the overall running state of the equipment. Secondly, in terms of quality prediction and fault diagnosis, simple physical mechanism models cannot accurately describe the complex nonlinear cutting process, while purely data-driven "black box" models lack physical interpretability, have weak generalization ability, and require a large amount of labeled data for training. When a fault is detected, existing methods often stop at the fault classification level and cannot further provide quantitative and optimal process parameter adjustment strategies. They fail to form a closed-loop management from state monitoring, quality prediction to fault diagnosis and optimization control, making it difficult to meet the pursuit of extreme quality and efficiency in flexible cover plate processing. SUMMARY

[0004] To solve the technical problem of how to meet the pursuit of extreme quality and efficiency in flexible cover plate processing, the present application provides the following solutions.

[0005] A kind of data monitoring method for round knife film cutting machine of flexible cover plate processing, comprising: S1, the physical quantity detected by various sensors and the product surface image photographed by industrial camera are acquired;Image characteristics of the product surface image are extracted, and all the physical quantity and image characteristics constitute multi-source time series data;S2, device state atlas is constructed based on the physical causality in the multi-source time series data, the device state atlas is processed by the cascade model of graph attention network and gated recurrent unit network, and device state space-time feature vector is extracted;S3, the space-time feature vector is input to a parallel prediction structure, the parallel prediction structure includes a physical mechanism model branch for calculating theoretical cutting state, and a data-driven residual model branch for calculating nonlinear deviation between theoretical and actual cutting states, the output of the two is fused to obtain processing quality prediction result, and the hidden space flow form representation of current processing state is generated;S4, when the distance between the hidden space flow form representation and the preset normal state flow form exceeds threshold value, it is classified to the fault type corresponding to the fault prototype with the closest distance by calculating the distance between it and each preset fault prototype flow form.

[0006] Further, when the determined fault type is parameter deviation, a group of candidate state representations are generated in the hidden space for the parameter, and the processing quality corresponding to each candidate state is deduced using the parallel prediction structure, and the parameter adjustment value corresponding to the candidate state with the optimal processing quality is selected.

[0007] Further, the physical quantity detected by the various sensors includes: the current and torque of the knife roller driving motor, the knife mold down pressure and the material belt tension during the operation of the round knife film cutting machine.

[0008] Further, the device state atlas is constructed based on the physical causality in the multi-source time series data, comprising: each physical quantity and image feature is taken as a node, and directed edges are established between the nodes according to the device dynamics transmission and process influence relationship, to construct the device state atlas.

[0009] Further, the device state space-time feature vector is extracted by the cascade model of graph attention network and gated recurrent unit network, comprising: at each time step, the node state vector fused with space coupling information is generated using the features of each node and its adjacent nodes in the device state atlas;The node state vector sequence formed by the node state vector is input to the gated recurrent unit network to capture its time dependence, and the device state space-time feature vector is output.

[0010] Further, the physical mechanism model branch adopts Ernst-merchant cutting model.

[0011] Further, the data-driven residual model branch uses a deep neural network.

[0012] Further, the physical mechanism model branch is used to calculate a theoretical cutting state; the data-driven residual model branch is used to predict a nonlinear deviation between the theoretical cutting state and an actual cutting state; and the theoretical cutting state and the nonlinear deviation are fused to obtain a final machining quality prediction result.

[0013] Further, the determination of the fault type comprises the following steps: collecting data in a normal state and in a plurality of known single fault modes in advance, respectively generating a normal state prototype manifold and a plurality of fault prototype manifolds through the cascade model and the parallel prediction structure; performing abnormality detection by calculating a distance between a hidden space manifold representation of a current state and the normal state prototype manifold; and when the distance exceeds a threshold value, further calculating distances between the current state and each fault prototype manifold, and attributing the fault to a fault type corresponding to a fault prototype with the closest distance.

[0014] Further, the selection of the parameter adjustment value corresponding to the candidate state with the optimal machining quality comprises the following steps: identifying a hidden space dimension associated with a parameter to be adjusted; generating a group of candidate state representations around a current state representation in the hidden space according to a preset strategy; performing inference prediction of machining quality on each group of candidate state representations by using the parallel prediction structure; selecting a candidate state representation with the optimal predicted machining quality, and decoding the candidate state representation into a specific parameter adjustment value.

[0015] Compared with the prior art, the beneficial effects of the present application comprise the following: by constructing a device state atlas reflecting physical causality and combining a graph attention network and a gated recurrent unit network, the inherent coupling relationship between multiple source data can be fundamentally revealed, and a more comprehensive and accurate device state representation can be obtained than traditional methods. The present application fuses a physical mechanism model and a data-driven residual model in parallel, which not only guarantees the physical interpretability and generalization ability of the quality prediction result, but also compensates for the nonlinear deviation of the theoretical model through the data model, thereby improving the prediction accuracy. The present application not only enables accurate fault diagnosis through the hidden space manifold distance, but also, when a parameter deviation occurs, utilizes the high-precision model to inversely deduce a quantitative and optimal process adjustment strategy, thereby forming a complete technical closed loop from monitoring, prediction, diagnosis to optimization control, and providing a more reliable and intelligent solution for guaranteeing the machining quality and stability of the flexible cover plate. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a schematic diagram of a parallel prediction structure according to an embodiment of the present application; Figure 2 is a schematic diagram of calculating distances between a hidden space manifold representation and a normal state prototype manifold and a fault prototype manifold according to an embodiment of the present application. DETAILED DESCRIPTION

[0017] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0018] A data monitoring method for a circular knife film cutting machine for flexible cover plate processing, comprising: S1, acquiring physical quantities detected by various sensors and product surface images captured by an industrial camera, the physical quantities detected by the various sensors including: current and torque of a knife roller driving motor, knife mold down pressure, and material belt tension during operation of the circular knife film cutting machine; extracting image characteristics of the product surface images, and constructing multi-source time series data from all the physical quantities and image characteristics; Specifically, the motor current and torque can be collected by a Hall sensor, the knife mold down pressure can be acquired by a pressure sensor, the material belt tension can be monitored by a tension sensor, the product surface images can be captured by an industrial camera, and the image characteristics (i.e., image feature vectors) of the product surface images can be extracted by a convolutional neural network such as ResNet, and all the physical quantities and the image feature vectors can be aligned in time stamps to form unified and heterogeneous multi-dimensional time series data.

[0019] For example, to comprehensively monitor the operating state of the circular knife film cutting machine, a current sensor and a vibration accelerometer are installed on the knife roller driving motor to collect current and vibration signals at a frequency of 1000 Hz, for example, recording an average operating current of 5.1 amperes and a vibration main frequency of 120 Hz. A pressure sensor is installed at the knife mold down mechanism to monitor the real-time knife mold down pressure, which can be stabilized at 3.2 megapascals. A tension sensor is installed at the material belt tension control system to measure a value of 25 newtons. At the same time, a line array camera is deployed at the product outlet to capture product surface images at a speed of 50 frames per second. All physical quantities and product surface images (image data) are synchronized by the network time protocol to ensure that the data is aligned in time.

[0020] S2, constructing a device state graph based on the physical causal relationship in the multi-source time series data, processing the device state graph by a cascaded model of a graph attention network and a gated recurrent unit network, and extracting a device state spatiotemporal feature vector; In an optional embodiment, the acquisition of the multi-source time series data and the construction of the device state graph include: taking each physical quantity and image feature as a node, and establishing a directed edge between the nodes according to the device dynamics transmission and process influence relationship, to construct the device state graph.

[0021] When constructing the device state graph, each physical quantity is abstracted as a node. For example, motor current becomes a node, motor vibration spectrum becomes another node, and die down pressure and material belt tension also become independent nodes, respectively. Image features extracted by algorithms, such as scratch number and burr height, also become nodes, respectively. Subsequently, directed edges between nodes are established according to the physical logic of device operation. For example, fluctuations in motor current directly affect the speed stability of the knife roll, which in turn affects the cutting force, so a directed edge is established between the motor current node and the die down pressure node. Similarly, abnormal material belt tension affects the material positioning accuracy, resulting in scratches on the finished surface, so a directed edge is established between the material belt tension node and the scratch number node, forming a dynamic directed graph representing the state of each component of the device and their mutual influence relationship.

[0022] In an optional embodiment, the device state spatiotemporal feature vector is extracted by a cascaded model of a graph attention network and a gated recurrent unit network, comprising: At each time step, the graph attention network is used to aggregate the features of each node and its adjacent nodes in the device state graph, to generate a node state vector that integrates spatial coupling information; the sequence of node state vectors formed in time sequence is input into the gated recurrent unit network to capture its time dependence, and the device state spatiotemporal feature vector is output.

[0023] Specifically, at any sampling time point, for example, at 10.05 seconds, the graph attention network receives the device state graph at this moment as input. For the die down pressure node in the graph, its current measurement value is 3.2 MPa. The graph attention network calculates the influence weight of the adjacent nodes, such as the motor vibration node and the material belt tension node, on the die down pressure node. Assuming that the attention weight of the vibration node is calculated to be 0.6 and the weight of the material belt tension node is calculated to be 0.2, then the updated die down pressure node state vector not only contains its own 3.2 MPa information, but also weightedly integrates the spatial coupling information that the vibration is larger and the tension is stable at this moment, forming a higher-dimensional feature representation. This process is performed simultaneously for all nodes in the device state graph, thereby generating a snapshot of the entire device state at this moment.

[0024] The above process is repeated at each sampling time point, for example, from the 1st second to the 20th second, generating a node state vector of the full device state map every 0.01 seconds. This forms a vector sequence containing 2000 time steps. The vector sequence is input into the gated recurrent unit network. The gated recurrent unit network learns and captures the patterns of the evolution of these state vectors over time through its internal update gate and reset gate structures. For example, it can learn that the motor vibration frequency continuously rises from 120 Hz to 180 Hz within 3 seconds, which usually indicates that the cutting quality is about to decline due to the cutter. Finally, at the 20th second, the gated recurrent unit network outputs a fixed-length device state spatiotemporal feature vector that condenses the spatial correlation and temporal evolution of all component states of the device within the past 20 seconds.

[0025] S3, inputting the spatiotemporal feature vector into a parallel prediction structure, which includes a physical mechanism model branch for calculating a theoretical cutting state and a data-driven residual model branch for calculating a nonlinear deviation between the theoretical cutting state and the actual cutting state, fusing the outputs of the two branches to obtain a machining quality prediction result and generating a latent space manifold representation of the current machining state; As shown in Figure 1 Specifically, the physical mechanism model branch uses the Ernst-Schmidt cutting model, for example, an empirical formula based on cutting mechanics can be used to calculate the theoretical cutting force according to the input related components such as force and speed. For example, the component value corresponding to the tool speed is 300 revolutions per minute, the component value corresponding to the pressing force is 3.2 megapascals, and the component value corresponding to the material thickness is 0.1 millimeters. Substituting these values into the empirical formula, a theoretical burr height of 0.02 millimeters is calculated. This result is a theoretical cutting state based on ideal working conditions, which has a clear physical meaning but ignores many real-world complexities.

[0026] The data-driven residual model branch is used to predict the nonlinear deviation between the theoretical cutting state and the actual cutting state. The theoretical cutting state and the nonlinear deviation are fused to obtain the final machining quality prediction result.

[0027] Specifically, the data-driven residual model branch uses a deep neural network (multi-layer perceptron) whose input is the complete spatio-temporal feature vector. The deep neural network is trained to learn the difference between the predicted theoretical burr height and the actual burr height, i.e. the residual. For example, according to the current complete spatio-temporal feature vector, it predicts that the additional burr height caused by the combined effects of tool wear and ambient temperature change is 0.015 mm. Finally, the calculation result of the physical mechanism model branch 0.02 mm is added to the residual 0.015 mm predicted by the data-driven residual model branch, and the final machining quality prediction result is 0.035 mm, which takes into account the stability of the physical law and the accuracy of the data-driven.

[0028] The deep neural network, the output of the last hidden layer of which is the current state's latent space manifold representation.

[0029] S4, when the distance between the latent space manifold representation and the preset normal state manifold exceeds the threshold, the distance between the current state representation and each preset fault prototype manifold is calculated, and the current state is classified into the fault type corresponding to the fault prototype with the smallest distance.

[0030] Specifically, time series data under normal production and various known faults such as tool wear, pressure misalignment, and tension abnormalities are collected in advance, and their latent space manifold representation sets are extracted to form the normal state manifold and each fault prototype manifold. In real-time monitoring, the difference between the current state representation and the normal state manifold distribution is calculated using the bulldozer distance (Wasserstein distance). Once the distance exceeds the limit, the distances between the current representation and each fault prototype manifold are calculated, and the fault label corresponding to the fault prototype with the smallest distance is selected as the current fault type.

[0031] In an optional embodiment, the determination of the fault type comprises: Data under normal state and various known single fault modes are collected in advance, and the normal state prototype manifold and various fault prototype manifolds are generated respectively through the cascade model and the parallel prediction structure; abnormality detection is performed by calculating the distance between the current state's latent space manifold representation and the normal state prototype manifold, and when the distance exceeds the threshold, the distance between the current state and each fault prototype manifold is further calculated, and the fault is attributed to the fault type corresponding to the fault prototype manifold with the smallest distance.

[0032] Specifically, the circular knife film cutting machine was first operated under normal conditions for 10 hours. Data was collected and encoded into a series of feature vectors in the latent space using the model. These feature vectors together constitute a distribution region, the normal state prototype manifold. Next, typical faults were artificially introduced. For example, replacing a worn tool with a worn tool for two hours generated a tool wear fault prototype manifold; another example was intentionally loosening the web tension controller for two hours to generate a tension abnormality fault prototype manifold. The normal state prototype manifold and each fault prototype manifold represent the data distribution pattern of a specific device state in the latent space.

[0033] During the real-time monitoring of the equipment, the system continuously encodes the current operating data into a latent space manifold representation. Figure 2 As shown, the Wasserstein distance between the manifold representation of the latent space and the prototype manifold of the normal state is first calculated. Assume that the preset abnormality threshold is 0.8. If the calculated distance is 0.4, it is judged that the equipment is operating normally. If the distance calculated at a certain moment is 1.5, which is greater than the threshold 0.8, the system determines that an abnormality has occurred. At this time, the system will then calculate the Wasserstein distance between the manifold representation of the latent space and each fault prototype manifold. For example, the distance between it and the tool wear prototype manifold (such as the fault prototype tool wear in the figure) is 0.3, and the distance between it and the pressure misalignment prototype manifold (such as Figure 2 The distance between the prototype (pressure misalignment) and the tool wear prototype is 1.2. Since 0.3 is the minimum distance, the tool wear prototype corresponds to the tool wear fault. The system diagnoses the current fault type as tool wear and issues an early warning to the operator.

[0034] Furthermore, step S5 is also included. When the fault type determined is parameter offset, a set of candidate state representations are generated in the latent space for the parameter, the parallel prediction structure is used to deduce the processing quality corresponding to each candidate state, and the parameter adjustment value corresponding to the candidate state that optimizes the processing quality is selected.

[0035] Parameters refer to controllable machining parameters. Specifically, if the fault is diagnosed as excessive die pressure, the system explores the latent space along directions related to pressure changes, generating a series of candidate state representations representing pressure decreasing from the current value. These candidate state representations are input one by one into the trained parallel prediction structure to determine the machining quality corresponding to each candidate pressure value. Finally, the candidate state representation with the best machining quality is selected, and the corresponding pressure value is the optimal adjustment value recommended to the operator or control system.

[0036] In an optional embodiment, selecting the parameter adjustment value corresponding to the candidate state that optimizes the processing quality includes: identifying the latent space dimensions associated with the parameters to be adjusted; within the latent space, generating a set of candidate state representations around the current state representation according to a preset strategy; performing inference prediction of the processing quality for each set of candidate state representations using the parallel prediction structure; selecting the candidate state representation with the optimal predicted processing quality and decoding it into a specific parameter adjustment value.

[0037] Specifically, when the current processing quality, e.g. burr height, is detected to be 0.05mm, which does not meet the requirement, the system initiates the optimization procedure. The system knows from the analysis model that the 10th to 12th dimensions of the latent space manifold representation (a feature vector) are mainly associated with the roll speed, while the 25th to 27th dimensions are mainly associated with the depth of pressing. The values of the latent space manifold representation of the current state on these dimensions can be 0.58 and 0.73 respectively. The system then explores around the current latent space manifold representation along these controllable dimensions to generate 100 candidate state representations. For example, one candidate state representation can become 0.62 on the roll speed related dimensions and 0.71 on the depth of pressing dimensions.

[0038] The 100 candidate state representations are fed into the previously established parallel prediction structure one by one. The processing quality is inferred for the virtual equipment state represented by each candidate state representation. For example, the burr height is predicted to be 0.04mm for candidate state representation one, 0.03mm for candidate state representation two, and the optimal burr height of 0.015mm is predicted for candidate state representation eighty-three. The system then selects candidate state representation eighty-three as the optimization target. Through the decoder network trained in advance, candidate state representation eighty-three, an abstract mathematical vector, is decoded back into specific, executable equipment parameters. The decoding result can be to increase the roll speed from 300rpm to 310rpm and to fine-tune the depth of pressing from 3.2MPa to 3.15MPa.

[0039] Many modifications, changes and substitutes will occur to those skilled in the art without departing from the spirit and principles of the present application. It should be understood that various alternatives to the embodiments of the present application described herein can be employed in practicing the present application.

Claims

1. A data monitoring method for a circular knife film cutting machine for processing flexible cover sheets, characterized in that: include: S1, obtaining physical quantities detected by various sensors and product surface images captured by industrial cameras; extracting image characteristics of the product surface images, and constructing all the physical quantities and image features into multi-source time series data; S2, constructing a device state graph based on the physical causal relationship in the multi-source time series data, processing the device state graph through a cascade model of a graph attention network and a gated recurrent unit network, and extracting a device state spatiotemporal feature vector; S3, inputting the spatiotemporal feature vector into a parallel prediction structure, which includes a physical mechanism model branch for calculating the theoretical cutting state and a data-driven residual model branch for calculating the nonlinear deviation between the theoretical and actual cutting states. The outputs of the two are combined to obtain a machining quality prediction result and generate a latent space manifold representation of the current machining state; S4, when the distance between the latent space manifold representation and the preset normal state manifold exceeds a threshold, the distance between the latent space manifold and each preset fault prototype manifold is calculated to classify the manifold into the fault type corresponding to the nearest fault prototype.

2. The method according to claim 1, characterized in that When the fault type determined is parameter offset, a set of candidate state representations are generated in the latent space for the parameter, the parallel prediction structure is used to deduce the processing quality corresponding to each candidate state, and the parameter adjustment value corresponding to the candidate state that optimizes the processing quality is selected.

3. The method according to claim 2, characterized in that The physical quantities detected by the various sensors include: the current and torque of the knife roller driving motor, the knife die pressing pressure and the material strip tension during the operation of the circular knife film cutting machine.

4. The method according to claim 3, characterized in that Constructing a device state graph based on the physical causal relationship in the multi-source time series data includes: Each physical quantity and image feature is taken as a node, and directed edges are established between the nodes based on the relationship between equipment dynamics transmission and process influence, so as to construct the equipment state map.

5. The method according to claim 4, characterized in that The device state spatiotemporal feature vector is extracted by using a cascade model of a graph attention network and a gated recurrent unit network, including: At each time step, the characteristics of each node and its adjacent nodes in the device state graph are used to generate a node state vector that incorporates spatial coupling information; the node state vector sequence composed of the node state vectors is input into the gated recurrent unit network to capture its time dependency, and the device state spatiotemporal feature vector is output.

6. The method according to claim 1, characterized in that The physical mechanism model branch adopts the Ernst-Schönter cutting model.

7. The method according to claim 6, characterized in that The data-driven residual model branch uses a deep neural network.

8. The method according to claim 7, characterized in that The physical mechanism model branch is used to calculate the theoretical cutting state; the data-driven residual model branch is used to predict the nonlinear deviation between the theoretical cutting state and the actual cutting state; the theoretical cutting state and the nonlinear deviation are integrated to obtain the final processing quality prediction result.

9. The method according to claim 1, characterized in that Determination of the fault type, including: Collect data in normal state and multiple known single fault modes in advance, and generate normal state prototype manifolds and multiple fault prototype manifolds respectively through the cascade model and parallel prediction structure; Anomaly detection is performed by calculating the distance between the latent space manifold representation of the current state and the normal state prototype manifold. When the distance exceeds a threshold, the distance between it and each fault prototype manifold is further calculated, and the fault is attributed to the fault type corresponding to the nearest fault prototype.

10. The method according to claim 2, characterized in that The parameter adjustment value corresponding to the candidate state that optimizes the processing quality is selected, including: Identify the latent space dimensions associated with the parameters to be adjusted; In the latent space, a set of candidate state representations is generated based on a preset strategy around the current state representation; Using the parallel prediction structure to perform deductive prediction of processing quality for each group of candidate state representations; The candidate state representation that best predicts the machining quality is selected and decoded into specific parameter adjustment values.

Images