Intelligent management method and system for die-cutting workshop production line

Abstract

The application provides a die-cutting workshop production line intelligent management method and system, the method comprising: collecting multi-dimensional state data to form an original matrix, preprocessing and fusing the original matrix to generate a fused matrix; inputting the fused matrix into a pre-constructed deep perception model to output a state coding vector and a channel attention vector; inputting the state coding vector as an environment state to construct a deep reinforcement learning model to output an action sequence and a dynamic strategy distribution; combining the action sequence and the dynamic strategy distribution to construct a dynamic process offset index to generate a process offset value; when a parameter correction value exceeds a preset threshold, adjusting a current process parameter through a parameter correction network; when the real-time monitoring index continuously exceeds a dynamic threshold for a certain period, triggering a fault response program to output a real-time early warning instruction sequence. The application realizes automation and optimization of production scheduling, equipment fault prediction and quality control.

Description

Technical Field

[0001] This invention belongs to the field of intelligent production line management, and in particular relates to an intelligent management method and system for a die-cutting workshop production line. Background Technology

[0002] With the continuous improvement of automation in manufacturing, die-cutting workshops, as a crucial link in industrial production, face challenges in various aspects, including production efficiency, product quality, and equipment reliability. The core task of a die-cutting workshop is to efficiently and accurately cut raw materials using die-cutting machines to produce products that meet quality requirements. However, existing die-cutting workshop production lines still have some problems that urgently need to be addressed in actual operation, particularly in areas such as production scheduling, equipment failure prediction, quality control, and intelligent management.

[0003] Currently, production scheduling in die-cutting workshops primarily relies on manual operation or rule-based scheduling algorithms. These traditional scheduling methods are often unable to cope with the dynamic changes and complexities within the production line, leading to low efficiency in production task execution. Especially in environments with multiple devices working collaboratively, multiple tasks being scheduled, and complex production environments, traditional production scheduling methods struggle to effectively respond to changes in equipment status, material fluctuations, and urgent adjustments to production tasks, resulting in low production efficiency, excessive equipment downtime, and frequent production bottlenecks. Furthermore, equipment failures and downtime remain significant factors in the production process. Although some workshops have adopted sensor-based fault detection systems, these systems largely rely on preset rules and thresholds, lacking sufficient flexibility and predictability to identify potential equipment failures in advance, leading to frequent production line downtime and maintenance. More seriously, equipment failures often impact subsequent production processes, delaying delivery dates and affecting customer satisfaction.

[0004] Regarding quality control, although some workshops have introduced automated quality inspection equipment, such as vision inspection systems, most still rely on manual inspection or experience-based judgment to ensure product quality. This method has a certain degree of uncertainty, especially when product quality fluctuates during production. Traditional quality control methods often fail to identify and adjust production processes in a timely manner, leading to the accumulation of quality problems and an increase in scrap rates. The current shortcomings of quality management are mainly manifested in the following aspects: low inspection efficiency, delayed feedback, and inability to adjust process parameters in real time to adapt to changes in production.

[0005] Therefore, existing technologies still face many bottlenecks in intelligent production line management, necessitating innovative solutions to improve production efficiency, equipment reliability, and quality control in die-cutting workshops. Traditional scheduling algorithms and quality management methods have limited effectiveness in dynamic production environments, failing to achieve precise scheduling of complex production tasks and comprehensive prediction of equipment status. Furthermore, they lack sufficient adaptive capabilities and real-time feedback mechanisms in quality control. Therefore, how to improve production efficiency, reduce downtime, and optimize quality control through intelligent means, particularly data-driven automated control methods, has become a pressing issue for current die-cutting workshop management systems. Summary of the Invention

[0006] The purpose of this invention is to design an intelligent management method and system for die-cutting workshop production lines. Based on advanced intelligent sensing and adaptive control technologies, the system achieves automation and optimization of production scheduling, equipment fault prediction, and quality control by monitoring the production line status in real time and combining technologies such as deep learning and reinforcement learning.

[0007] To achieve the above objectives, a first aspect of the present invention provides an intelligent management method for a die-cutting workshop production line, the method comprising:

[0008] Multiple sensors are deployed at key workstations in the die-cutting workshop production line to collect multi-dimensional status data to form an original matrix. The original matrix is ​​then preprocessed and fused to generate a fused matrix.

[0009] The fused matrix is ​​input into a pre-constructed deep perception model, local features are extracted using dual-domain perception convolution kernels, and processed by a channel attention module that combines physical prior knowledge, outputting a state encoding vector and a channel attention vector.

[0010] Using the state encoding vector as the environmental state input, a deep reinforcement learning model is constructed, and a reward function containing a task-resource cross-fit index is designed to output the action sequence and dynamic policy distribution.

[0011] Combining the action sequence and dynamic strategy distribution, a dynamic process offset index is constructed, and a process offset value is generated. When the parameter correction value exceeds a preset threshold, the current process parameters are adjusted through a parameter correction network, and the adjusted new parameter values ​​are output, forming a parameter correction sequence.

[0012] A real-time monitoring index is constructed based on the process offset value and parameter correction sequence, and a dynamic threshold is set. When the real-time monitoring index is continuously higher than the dynamic threshold for a certain period of time, a fault response program is triggered, and a real-time early warning instruction sequence is output.

[0013] Further, the preprocessing and fusion of the original matrix to generate the fused matrix specifically includes:

[0014] Obtain the original matrix;

[0015] The original matrix is ​​time-aligned using linear interpolation for sensor data at different frequencies. Dimensional differences are eliminated through normalization. Then, based on the original matrix, the Frobenius norm and... Regular expressions are used to perform sparse recovery calculations to generate the fused matrix.

[0016] Furthermore, the dual-domain perceptual convolutional kernel includes a short perceptual domain and a long perceptual domain, which respectively capture abnormal perturbations and trend changes.

[0017] Furthermore, the process of inputting the fused matrix into a pre-constructed deep perception model, extracting local features using a dual-domain perceptual convolutional kernel, and processing it using a channel attention module incorporating prior physical knowledge, outputting a state encoding vector and a channel attention vector, specifically includes:

[0018] Obtain the fused matrix;

[0019] The fused matrix is ​​convolved using the short and long receptive domains respectively, and then channel concatenation is performed to generate local features.

[0020] The local features are input into the channel attention model, which outputs a state encoding vector and a channel attention vector; wherein, the loss function of the channel attention model is... Represented as:

[0021] ;

[0022] in, For the current training process Channel attention coefficients of each sensor, This represents the prior probability of the sensor's physical importance as given in the process drawings or experience database. Kullback-Leibler divergence measures the deviation between the attention distribution learned by the model and the physical preset. This is the weight adjustment factor for each channel, used to control the intensity of its penalty.

[0023] Furthermore, the deep reinforcement learning model consists of two fully connected layers, each equipped with a ReLU activation function and a normalization layer. Its output is a probability distribution of a set of discrete actions, with each action corresponding to a set of scheduling decisions.

[0024] The scheduling strategy of the deep reinforcement learning model is trained through deep reinforcement learning and updated using the PPO algorithm.

[0025] Furthermore, the reward function is expressed as:

[0026] ;

[0027] in, It is the first Instant rewards for each step; It is the scheduling action generated by the policy network at the current time step; This represents the resource utilization score, which measures the equipment utilization and average waiting time under the current action. The adaptability score measures how well the selected task matches the device's capabilities and sensitive channels in the current state. and These are adjustable weight parameters.

[0028] Furthermore, the step of constructing a dynamic process offset index and generating a process offset value by combining the action sequence and the dynamic strategy distribution specifically includes:

[0029] Multidimensional data acquisition is performed using the various sensors to generate fusion features for the current task;

[0030] Obtain the optimal feature of the current task within a historical stable output cycle;

[0031] Based on the optimal features and fused features, the deviation of the current process state from its historical optimal operating condition is calculated and used as the process offset value.

[0032] Furthermore, the real-time monitoring metrics are used to reflect the system's recent performance based on parameter correction amounts. The dynamic threshold is obtained based on the mean and standard deviation of the real-time monitoring indicators during the stable operation period of the system.

[0033] Furthermore, the real-time early warning instructions include: sending load reduction / shutdown instructions to the target equipment, pushing early warning notifications to on-site personnel, automatically dispatching alternative equipment to take over subsequent tasks, and packaging and archiving fault-related segments.

[0034] A second aspect of the present invention provides an intelligent management system for a die-cutting workshop production line, the system comprising:

[0035] The data acquisition and fusion module is used to deploy multiple sensors at key workstations in the die-cutting workshop production line to collect multi-dimensional status data to form an original matrix. The original matrix is ​​then preprocessed and fused to generate a fused matrix.

[0036] The deep perception modeling module is used to input the fused matrix into a pre-constructed deep perception model, extract local features using a dual-domain perception convolution kernel, process them using a channel attention module with prior physical knowledge, and output a state encoding vector and a channel attention vector.

[0037] The scheduling optimization module is used to take the state encoding vector as the environmental state input, construct a deep reinforcement learning model, design a reward function that includes a task-resource cross-fitness index, and output action sequences and dynamic policy distributions.

[0038] The quality control module is used to combine the action sequence and dynamic strategy distribution to construct a dynamic process offset index, generate a process offset value, and when the parameter correction value exceeds a preset threshold, adjust the current process parameters through a parameter correction network, output the adjusted new parameter values, and form a parameter correction sequence.

[0039] The fault early warning module is used to construct real-time monitoring indicators based on the process offset value and parameter correction sequence, and set dynamic thresholds. When the real-time monitoring indicators are continuously higher than the dynamic threshold for a certain period of time, the fault response program is triggered and a real-time early warning instruction sequence is output.

[0040] The beneficial technical effects of the present invention are at least as follows:

[0041] First, by utilizing multi-dimensional sensor data fusion and deep sensing networks, the system comprehensively perceives and monitors the production status of the die-cutting workshop in real time. By integrating data from multiple sensors such as temperature, humidity, pressure, vibration, and vision, the system can accurately capture the health status of equipment, changes in the production environment, and product quality during the production process, solving the problem of lacking comprehensive real-time monitoring and dynamic data analysis in traditional technologies.

[0042] Secondly, by introducing reinforcement learning algorithms to dynamically optimize production scheduling, the system can autonomously adjust the priority of production tasks and the allocation of production line resources based on real-time feedback data during the production process, ensuring the efficient execution of production tasks. This method overcomes the problems of traditional scheduling algorithms being unable to respond to changes in the production environment in real time and struggling to cope with complex scheduling needs, thereby improving the overall operating efficiency of the production line.

[0043] Finally, this invention incorporates an adaptive control mechanism to achieve real-time optimization of die-cutting process parameters. The system can automatically adjust process parameters (such as tool pressure and cutting speed) based on real-time data feedback during production, thereby achieving precise quality control. This innovative method solves the problems of slow response and difficulty in timely adjustment associated with traditional quality control methods, significantly improving product quality consistency and production efficiency.

[0044] In summary, this invention provides an intelligent solution by combining deep perceptual networks and reinforcement learning. This solution dynamically optimizes production scheduling, accurately predicts equipment failures, and adaptively controls quality during the production process, ensuring efficient and stable operation of the die-cutting workshop in complex and ever-changing production environments. This solution not only improves production efficiency and reduces downtime but also optimizes quality control, demonstrating a high level of intelligence and practical value. Attached Figure Description

[0045] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0046] Figure 1 This is a flowchart of an intelligent management method for a die-cutting workshop production line according to the present invention. Detailed Implementation

[0047] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0048] In one or more embodiments, such as Figure 1 As shown, an intelligent management method for a die-cutting workshop production line is disclosed, the method comprising the following S1~S5:

[0049] S1. Deploy multiple sensors at key workstations in the die-cutting workshop production line to collect multi-dimensional status data to form an original matrix. Preprocess and fuse the original matrix to generate a fused matrix.

[0050] Specifically, this step involves deploying sensors of various types, including temperature, pressure, vibration, and vision sensors, at multiple key workstations on the die-cutting production line to collect multi-dimensional status data in real time during the production process. Considering that die-cutting equipment is simultaneously affected by physical pressure, thermal deformation, and mechanical vibration during operation, single-dimensional data is insufficient to fully describe its state. Therefore, we need to construct a data structure that integrates multi-dimensional information to lay the data foundation for subsequent state-aware modeling and scheduling optimization. Raw sensor data is transmitted in real time to the central data system via serial port, Modbus, or image acquisition card. At a specific point in time... For example, from The data collected by each sensor forms a feature vector. The observation data at all time points form the following matrix:

[0051] ;

[0052] in, Indicates the first The data collected at the [number]th time point Data values ​​from each sensor; It is the number of sampling times. This refers to the total number of sensor types (such as pressure, temperature, image features, vibration RMS value, etc.).

[0053] The original matrix contains data with significant differences in dimensions, frequency, and noise characteristics across different dimensions, necessitating unified preprocessing and fusion. First, time alignment is performed: for example, a visual sensor acquires images every 5 seconds, while a vibration sensor acquires data every second. Linear interpolation is used to upscale the low-frequency data to a second-level resolution, ensuring all sensor data correspond to the same time axis. Then, normalization is applied to the data across different dimensions to unify their numerical scale and eliminate model bias caused by differences in dimensions.

[0054] Next, to enhance the discriminative power of the data in subsequent models, we introduce a structurally sparse regularization term during the fusion stage to emphasize key signal channels and suppress redundant interference signals. The preprocessed data matrix is ​​as follows:

[0055] ;

[0056] in, It is the new matrix after fusion; It is the first A vector corresponding to each sensor column; It is the Frobenius norm, used to measure the error in data reconstruction; the second term is... Regular expressions are used to sparsify channels. This is a coefficient that balances data retention with channel selection. The introduction of this regularization term is designed for the unique scenario of a die-cutting workshop: when the state of die-cutting equipment fluctuates, typically only some key sensors (such as vibration and pressure) show a significant response to the state change, while other sensors may simply be sources of noise. Through structural sparsity fusion, we enable the model to automatically compress redundant inputs while preserving the main signal, improving subsequent modeling efficiency.

[0057] Among them, the fused matrix The output of this step has the following characteristics: 1) time sequence uniformity, 2) data standardization, 3) key features are highlighted, and 4) it is easy to input into subsequent deep models.

[0058] For example, in a typical production cycle, the vibration signal collected per second may exhibit a spike the instant the die-cutting blade contacts the material; while the output of the pressure sensor may reflect whether the die-cutting pressure head is in the correct position. By unifying and aligning these physically correlated high-frequency information and introducing a sparse fusion structure, the discrimination capability in the state modeling stage can be significantly improved.

[0059] S2. Input the fused matrix into the pre-constructed deep perception model, extract local features using dual-domain perception convolution kernels, process them using a channel attention module with prior physical knowledge, and output state encoding vector and channel attention vector.

[0060] Specifically, this step aims to utilize the fusion matrix output in step 1. Our goal is to construct a deep perception model that can accurately and finely reflect the status of a die-cutting production line. This model must not only be able to identify minute fluctuations in equipment operation in real time, but also possess a certain level of physical constraint awareness to distinguish the impact of operational variations and environmental disturbances. Our objective is not to build a general neural network, but rather to design a perception framework that combines the state response characteristics of the die-cutting process, possesses strong structural awareness, clear physical interpretation, and self-constraint capabilities.

[0061] The input is the standardized fusion data matrix generated in step 1. ,in Number of sampling times The effective number of sensor channels is represented by each column of the matrix, which is a normalized time series of a certain type of sensor (such as pressure, vibration RMS, image texture entropy, etc.), and each row represents a multi-channel observation at a certain moment.

[0062] Considering the two typical characteristics of the operating state of die-cutting equipment: high-frequency and detailed local fluctuations, and low-frequency and stable global trends, we use a set of dual-domain perceptual convolutional kernels for local feature extraction. These convolutional kernels are divided into short perceptual domains ( ) and long perceptual field ( There are two types, which capture abnormal disturbances and trend changes respectively. The network front-end... Perform parallel multi-sensory domain convolution:

[0063] ;

[0064] in, Indicates local features, This indicates a channel splicing operation. and These are convolution operations performed on short and long sensing kernels, respectively. This design can simultaneously extract high-frequency oscillations on the order of seconds caused by die-cutting head imbalance, and pressure attenuation trends on the order of hours caused by die-cutting mold aging.

[0065] After the above features are processed by batch normalization and activation functions, they are fed into a channel attention module that incorporates prior physical knowledge. This module, based on the importance of different signals in the physical model of the die-cutting device, incorporates a penalty term during the learning process to constrain the physical rationality of the attention weights. This regularization term is constructed as follows:

[0066] ;

[0067] in, For the current training process Channel attention coefficients of each sensor, This represents the prior probability of the sensor's physical importance as given in the process drawings or experience database. Kullback-Leibler divergence measures the deviation between the attention distribution learned by the model and the physical preset. This is the weight adjustment factor for each channel, used to control the intensity of its penalty.

[0068] Understandably, the introduction of this regularization term addresses the problem of "misfocusing" in traditional attention mechanisms during multi-sensor tasks. Taking the die-cutting pressure channel as an example, it may have a weak correlation with output quality in most situations, but it is highly decisive in the microseconds of the die's entry and exit. If relying entirely on training data, the model might ignore this channel due to its rarity and instead include it. Subsequently, the network will be forced to learn the importance of the channel during critical periods, thereby more realistically simulating the experience and judgment of human craftsmen.

[0069] The final output of the model is a state encoding vector. Its dimensions much smaller . It is the state embedding of the current moment of die-cutting production in the low-dimensional feature space, and has the following characteristics: (1) it reflects the overall picture of the equipment operation; (2) it can calculate the difference with the historical state; (3) it is convenient to input into the subsequent scheduling and optimization module.

[0070] Also output is the channel attention vector. This vector can be used by system operators to interpret the model's decision-making logic, such as whether the fluctuation of high-weight channels is related to the current decrease in die-cutting accuracy.

[0071] For example, in practical applications, we observed significant electrostatic accumulation in a certain batch of materials, leading to an abnormally high level of adhered residue after die-cutting. Traditional methods struggle to distinguish between abnormal mechanical intervention and material changes, while this model... The changing trends related to visual signals in the dimensions, combined with the phenomenon of increased image channel weights in channel attention, successfully prompted the system to switch the static eliminator and adjust the surface treatment, demonstrating the model's high sensitivity and interpretability to complex production conditions.

[0072] S3. Using the state encoding vector as the environmental state input, construct a deep reinforcement learning model, design a reward function that includes a task-resource cross-fit index, and output the action sequence and dynamic policy distribution.

[0073] Specifically, this step is based on the state encoding vector output in step 2. and channel attention vector This paper proposes a reinforcement learning strategy model for multi-process scheduling and resource allocation in a die-cutting workshop. This model aims to address the typical characteristics of a die-cutting workshop, such as frequent changes in equipment health status, simultaneous scheduling of multiple tasks, and strong resource competition coupling. Traditional fixed scheduling strategies or heuristic job ordering algorithms often only maintain local optima under given conditions and struggle to cope with sudden disturbances in the production line, such as excessive equipment vibration or temperature overload. Therefore, this paper proposes a state-driven, self-evolving reinforcement learning system as a bridge between the perception and control modules, enabling real-time adaptive optimization of key aspects such as die-cutting task sequence, equipment configuration, and personnel scheduling.

[0074] The model obtained in step 2 (Indicating the current overall operating status of the equipment and environment) is used as the environmental status input, and utilized This indicates the importance of each sensor channel, thus inferring which process variables are most sensitive in the current task. For example, if the channel weight of the pressure sensor is significantly higher than other channels at a certain moment, the system should prioritize completing the precision die-cutting task that is highly relevant to pressure control. Scheduling actions are generated by a policy network consisting of two fully connected layers, each equipped with a ReLU activation function and a normalization layer. Its output is a probability distribution of a set of discrete actions, each corresponding to a set of scheduling decisions, such as "assign task A to device 3" or "pause device 2 for maintenance."

[0075] Furthermore, to enhance the model's understanding of the associated costs between different resources, we designed a task-resource cross-fitness index in the reward function to measure whether the current action fully considers the coupling between equipment status and task process characteristics. The reward function takes the following form:

[0076] ;

[0077] in It is the first Instant rewards for each step; It is the scheduling action generated by the policy network at the current time step; This represents the resource utilization score, which measures the equipment utilization and average waiting time under the current action. The adaptability score measures how well the selected task matches the device's capabilities and sensitive channels in the current state. and These are adjustable weight parameters.

[0078] Furthermore, taking a real-world scenario as an example, assume that die-cutting tasks include three types: Type A (requiring high-pressure precision), Type B (long-cycle, large-mold), and Type C (requiring high image recognition capabilities). The state vector at a certain moment... The display shows a slight increasing trend in vibration of die-cutting machine 1; channel attention. This indicates a significant increase in the weight of the vibration channel, leading the system to conclude that the equipment exhibits a tendency towards structural abnormalities. At this point, the reinforcement learning model will automatically reduce the probability of equipment 1 receiving type A tasks, prioritizing its allocation to type B tasks, which have lower vibration sensitivity. It will also schedule type A tasks to the more stable equipment 2 or 3, while simultaneously prompting the system to schedule a short shutdown for maintenance on equipment 1.

[0079] Furthermore, the scheduling strategy is trained using deep reinforcement learning, and the Proximal Policy Optimization (PPO) algorithm is used for policy updates. To improve sample utilization and training stability, we introduce the following mechanisms: First, an experience replay pool stores recent decision trajectories and reward feedback to avoid overfitting to recent states during training; second, a dynamic sampling mechanism increases the sampling frequency of samples with high process sensitivity (such as equipment malfunctions or task critical points) to make training more focused on key scenarios; and third, a long-term average reward baseline is used to dynamically adjust model preferences to avoid misleading the overall strategy due to short-term efficient behavior.

[0080] The model's two key outputs are:

[0081] Action sequence : Represents scheduling behavior under continuous time steps, used to directly drive workshop execution, including specific operations such as task sequence, equipment selection, and mold change scheduling;

[0082] Dynamic policy distribution It can be used to evaluate the current model's preference for scheduling behavior under different states, and assist manual scheduling in strategy correction.

[0083] The core innovation of this step is reflected in three aspects: First, it systematically integrates the die-cutting process perception and task scheduling system, avoiding the problem of the "state recognition-scheduling control" disconnect; second, by introducing sensor channel attention to participate in the construction of scheduling preferences, the model has the ability to interpret the process dimension; finally, through the dual reward structure and policy stabilization mechanism, it realizes the long-term optimal exploration of the dynamic task-equipment mapping relationship, adapting to the highly changing environment and resource coupling relationship in die-cutting production.

[0084] S4. Combining the action sequence and dynamic strategy distribution, construct a dynamic process offset index, generate a process offset value, and when the parameter correction value exceeds a preset threshold, adjust the current process parameters through the parameter correction network, output the adjusted new parameter value, and form a parameter correction sequence.

[0085] Specifically, this step builds upon the previous reinforcement learning scheduling module, incorporating its output sequence of scheduling actions. With state-policy distribution This technology enables online quality feedback identification and adaptive adjustment of process parameters for die-cutting production scenarios. Its goal is to establish a control mechanism that can dynamically respond to fluctuations in operating conditions and proactively correct key process parameters during production, thereby solving typical pain points in traditional die-cutting production such as "difficulty in real-time detection of quality fluctuations," "rigid and static parameter settings," and "delayed correction of process anomalies."

[0086] First, the two outputs from step 3 are fully utilized in this step: It represents the specific scheduling decisions made by the system at each moment, involving execution behaviors such as task allocation, equipment selection, and mold activation; The decision confidence level, representing the strategy under the current state, is used to assess whether the current scheduling decision is in a sensitive state of "high uncertainty," posing a potential risk. Based on these two outputs and combined with real-time quality monitoring data, we identify quality fluctuations and drive the generation and deployment of parameter adjustment actions.

[0087] In die-cutting production, defective products are often caused by certain key parameters not being adapted to the current operating conditions in a timely manner. For example, the cutting pressure may not be adjusted according to the material thickness, the feeding cycle may not be adjusted according to the equipment vibration status, and the scraping angle and pressure may not match, resulting in missed cuts or adhesion. Therefore, we construct a "dynamic process deviation index" in real time. This quantifies the deviation of the current process state from its historical best operating condition. The construction method is as follows:

[0088] ;

[0089] in, It is the first The fusion features (such as image sharpness, edge contours, pressure consistency indicators, etc.) are extracted from the quality monitoring module in real time. It is the optimal feature reference corresponding to this task in the historical stable output cycle. This represents the length of the sliding window. This metric is updated once per second and is used to determine whether to trigger parameter adjustments.

[0090] Furthermore, when Exceeding the set threshold And the corresponding strategy distribution If a significant skewness occurs (e.g., the distribution entropy exceeds a set confidence limit), the system determines that it may be in a process risk state characterized by "uncertain scheduling strategy + significant quality fluctuations," triggering a fine-tuning mechanism. We construct a parameter correction network, which is a two-layer perceptron, with inputs including... Scheduling actions Based on the policy distribution characteristics (such as entropy, maximum probability value), output a process parameter correction vector. The correction method is as follows:

[0091] ;

[0092] in This refers to the current process parameters (such as die-cutting pressure, feed rate, cutter frequency, etc.). The direction and magnitude of the model prediction tuning. These are the adjusted new parameter values. Parameter correction values ​​can be tailored to physical constraints (such as maximum pressure limits) to ensure equipment safety.

[0093] Understandably, in specific implementation processes, for example, if task A is assigned to device 2, after running continuously for 10 seconds, the quality monitoring module detects repeated burrs on the cut edge contour through image recognition, indicating a quality deviation index. It reached 0.85, far exceeding the normal value of 0.35. Meanwhile, the scheduling strategy distribution... The high entropy distribution during this time period indicates insufficient confidence in the scheduling choice. After combining these two signals, the model determines there is a potential mismatch risk, activates the correction network, and outputs... This indicates that the die-cutting pressure should be reduced by 0.2 units, the feed rate increased by 0.05 units, and the scraping frequency kept constant. The adjusted parameters are as follows. The information was immediately written to the PLC control unit and took effect. The quality monitoring module reported a reduction in edge burrs after 5 seconds. A drop to 0.42 indicates a successful adjustment.

[0094] The two outputs of this step are:

[0095] Parameter correction sequence It is used to replace the original parameter-driven process execution and is a responsive control for quality fluctuations after the strategy behavior.

[0096] Dynamic offset curve It is used to monitor in real time whether the current production task deviates from the historical normal working conditions, and can also be used as an evaluation index of scheduling stability.

[0097] It is important to note that the innovation of this step is reflected in the following three aspects:

[0098] First, it introduces policy distribution characteristics as a judgment trigger condition, enabling the system to "intervene in uncertain states" rather than just reacting to results. Second, by comparing the sliding window dynamic offset index with the historical best feature reference point, the quality evaluation becomes more targeted and real-time. Finally, the parameter correction mechanism is not based on manually set rules, but on regression prediction by building a neural network by fusing multiple input states, which has adaptive and sustainable evolution capabilities.

[0099] S5. Construct real-time monitoring indicators based on the process offset value and parameter correction sequence, and set dynamic thresholds. When the real-time monitoring indicators are continuously higher than the dynamic threshold for a certain period of time, trigger the fault response program and output a real-time early warning command sequence.

[0100] Specifically, this step involves adjusting the parameter correction sequence output in the previous stage. and Based on this, an online fault early warning mechanism and closed-loop optimization feedback control process are constructed for the die-cutting production system. Its core objective is to practically monitor the status of the system in operation, identify potential fault trends in a timely manner, and drive actual production intervention operations, thereby forming a final control closed loop with "status judgment + risk triggering + control execution" as the main line.

[0101] Regarding the input, firstly, the parameter correction sequence... This reflects the process adjustments the system made to maintain quality during past production cycles. Frequent and significant adjustments are often early signs of abnormal operating conditions. Secondly, the quality deviation trend... The fluctuation range of product quality within a certain time window was quantified. The trend of continuous deviation from the historical stable state indicates that there are anomalies in the current process that are difficult to repair through conventional adjustments.

[0102] To integrate these two inputs into a decision-oriented fault warning signal, we construct a real-time monitoring metric. The comprehensive reflection system in recent State stability over a period of time:

[0103] ;

[0104] in, Indicates the first The parameter adjustment amount at any given time. This is a quality offset indicator. Adjust the weighting balance coefficients for parameter adjustment and mass offset terms (e.g.) ), This is the length of the time window (e.g., 30 seconds). Because... and All have been normalized in previous steps (including maximum and minimum scaling and zero-mean standardization), therefore, in the construction... The units are already consistent, so no additional unit alignment is needed, avoiding deviations caused by comparisons across physical quantities.

[0105] Furthermore, to ensure that the early warning mechanism can both cover significant anomalies and avoid frequent false alarms, we have... Set dynamic threshold Its initial value can be adjusted during the stable operation period of the system. mean with standard deviation Statistical data, set as follows:

[0106] ;

[0107] in The value is usually taken as to It can be adjusted according to the stability of different process lines. During subsequent operation, It can be updated on a rolling basis based on weekly / monthly operating data, adapting to the long-term degradation modes of different task groups and different devices, ensuring that the early warning mechanism remains sensitive and robust in different scenarios.

[0108] Among them, when Continuously above the threshold Reaching the number of warning cycles (e.g., 3 consecutive cycle times) The system then triggers the actual fault response procedure. The response mechanism no longer remains at the modeling or inference level, but directly enters the motion control execution stage, initiating the following operations:

[0109] Send a load reduction / shutdown command to the target device;

[0110] Send early warning notifications to on-site personnel, including task number, equipment ID, and current status. The main source of the offset (e.g., frequent parameter adjustments on a certain channel);

[0111] Automatically allocate and assign backup equipment to take over subsequent tasks, and the scheduling layer prioritizes the transfer of low-risk batches based on historical task allocation strategies;

[0112] Package and archive the fault-related segments for use as subsequent expert review or maintenance records.

[0113] For example, in a multi-batch cutting task, a die-cutting machine began to frequently exhibit pressure correction behavior after completing the 5th batch of tasks. The mean gradually increases, and the quality deviation index It exhibits a linear growth trend. The system has cumulatively calculated... It is far higher than the set threshold. The system had been operating for two production cycles, and an early warning was triggered upon entering the third cycle. The control system immediately issued a "Stop Equipment Task 4" command, and the on-site dashboard displayed "Pressure Abnormality + Quality Drift." After inspection, maintenance personnel confirmed that the mold pressure head had become loose. In this case, the early warning mechanism not only played a role in early detection but also improved accuracy through parameter-quality joint judgment, avoiding unnecessary line stoppages or false alarms.

[0114] Finally, the output is a sequence of real-time warning instructions. Used at the direct control level, including control commands, notification content, emergency dispatch requests, etc.; fault data fragment record set. The system automatically archives the key parameter change trajectories and quality offset trends before and after triggering the warning, for use in subsequent analysis or training enhancement.

[0115] In one or more embodiments, an intelligent management system for a die-cutting workshop production line is disclosed, the system comprising:

[0116] The data acquisition and fusion module is used to deploy multiple sensors at key workstations in the die-cutting workshop production line to collect multi-dimensional status data to form an original matrix. The original matrix is ​​then preprocessed and fused to generate a fused matrix.

[0117] The deep perception modeling module is used to input the fused matrix into a pre-constructed deep perception model, extract local features using a dual-domain perception convolution kernel, process them using a channel attention module with prior physical knowledge, and output a state encoding vector and a channel attention vector.

[0118] The scheduling optimization module is used to take the state encoding vector as the environmental state input, construct a deep reinforcement learning model, design a reward function that includes a task-resource cross-fitness index, and output action sequences and dynamic policy distributions.

[0119] The quality control module is used to combine the action sequence and dynamic strategy distribution to construct a dynamic process offset index, generate a process offset value, and when the parameter correction value exceeds a preset threshold, adjust the current process parameters through a parameter correction network, output the adjusted new parameter values, and form a parameter correction sequence.

[0120] The fault early warning module is used to construct real-time monitoring indicators based on the process offset value and parameter correction sequence, and set dynamic thresholds. When the real-time monitoring indicators are continuously higher than the dynamic threshold for a certain period of time, the fault response program is triggered and a real-time early warning instruction sequence is output.

[0121] It is worth noting that the specific workflow of the intelligent management system for a die-cutting workshop production line provided in this embodiment of the invention is the same as that of the intelligent management method for a die-cutting workshop production line described in the above embodiment, and will not be repeated here.

[0122] This invention also provides an intelligent management device for a die-cutting workshop production line, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps described in the above embodiment of an intelligent management method for a die-cutting workshop production line, for example... Figure 1 The steps S1 to S5 described above; or, when the processor executes the computer program, it implements the functions of each module in the above system embodiments.

[0123] For example, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the intelligent management equipment for a die-cutting workshop production line.

[0124] The intelligent management equipment for a die-cutting workshop production line can be a desktop computer, laptop, handheld computer, or cloud server, etc. This equipment may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the intelligent management equipment may also include input / output devices, network access devices, buses, etc.

[0125] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the intelligent management equipment for the die-cutting workshop production line, connecting various parts of the equipment via various interfaces and lines.

[0126] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the intelligent management equipment for the die-cutting workshop production line by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created based on the operation of the air conditioning controller, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart memory card (SmartMedaCard, SMC), secure digital card (SecureDagatal, SD), flash memory card, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0127] If the integrated module of the intelligent management equipment for a die-cutting workshop production line is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0128] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0129] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for intelligent management of a die-cutting workshop production line, characterized in that, The method includes: Multiple sensors are deployed at key workstations in the die-cutting workshop production line to collect multi-dimensional status data to form an original matrix. The original matrix is ​​then preprocessed and fused to generate a fused matrix. The fused matrix is ​​input into a pre-constructed deep perception model, and local features are extracted using a dual-domain perceptual convolutional kernel. This dual-domain perceptual convolutional kernel includes a short perceptual domain and a long perceptual domain, which respectively capture abnormal perturbations and trend changes. The model is then processed using a channel attention module incorporating prior physical knowledge to output a state encoding vector and a channel attention vector. Specifically, the process of inputting the fused matrix into the pre-constructed deep perception model, extracting local features using a dual-domain perceptual convolutional kernel, and processing it using a channel attention module incorporating prior physical knowledge to output a state encoding vector and a channel attention vector includes: Obtain the fused matrix; The fused matrix is ​​convolved using the short and long receptive domains respectively, and then channel concatenation is performed to generate local features. The local features are input into the channel attention model, which outputs a state encoding vector and a channel attention vector; wherein, the loss function of the channel attention model is... Represented as: in, For the current training process Channel attention coefficients of each sensor, This represents the prior probability of the sensor's physical importance as given in the process drawings or experience database. Kullback-Leibler divergence measures the deviation between the attention distribution learned by the model and the physical preset. This is the weight adjustment factor for each channel, used to control the intensity of its penalty. This represents the total number of sensor types. The state encoding vector is used as the environmental state input to construct a deep reinforcement learning model. A reward function containing a task-resource cross-fit index is designed to output the action sequence and dynamic policy distribution. Combining the action sequence and dynamic strategy distribution, a dynamic process offset index is constructed, and a process offset value is generated. When the parameter correction value exceeds a preset threshold, the current process parameters are adjusted through a parameter correction network, and the adjusted new parameter values ​​are output, forming a parameter correction sequence. A real-time monitoring index is constructed based on the process offset value and parameter correction sequence, and a dynamic threshold is set. When the real-time monitoring index is continuously higher than the dynamic threshold for a certain period of time, a fault response program is triggered, and a real-time early warning command sequence is output.

2. The intelligent management method for a die-cutting workshop production line according to claim 1, characterized in that, The preprocessing and fusion of the original matrix to generate the fused matrix specifically includes: Obtain the original matrix; The original matrix is ​​time-aligned using linear interpolation of sensor data at different frequencies. Dimensional differences are eliminated through normalization. Then, based on the original matrix, the Frobenius norm and... Regular expressions are used to perform sparse recovery calculations to generate the fused matrix.

3. The intelligent management method for a die-cutting workshop production line according to claim 1, characterized in that, The deep reinforcement learning model consists of two fully connected layers, each equipped with a ReLU activation function and a normalization layer. Its output is a probability distribution of a set of discrete actions, and each action corresponds to a set of scheduling decisions. The scheduling strategy of the deep reinforcement learning model is trained through deep reinforcement learning and updated using the PPO algorithm.

4. The intelligent management method for a die-cutting workshop production line according to claim 1, characterized in that, The reward function is expressed as follows: in, It is the first Instant rewards for each step; It is the scheduling action generated by the policy network at the current time step; This represents the resource utilization score, which measures the equipment utilization and average waiting time under the current action. The adaptability score measures how well the selected task matches the device's capabilities and sensitive channels in the current state. and These are adjustable weight parameters; A is the state encoding vector; A is the attention vector.

5. The intelligent management method for a die-cutting workshop production line according to claim 1, characterized in that, The step of combining the action sequence and dynamic strategy distribution to construct a dynamic process offset index and generate a process offset value specifically includes: Multidimensional data acquisition is performed using the various sensors to generate fusion features for the current task; Obtain the optimal feature corresponding to the current task in the historical stable output cycle; Based on the optimal features and fused features, the deviation of the current process state from its historical optimal operating condition is calculated and used as the process offset value.

6. The intelligent management method for a die-cutting workshop production line according to claim 1, characterized in that, The real-time monitoring metrics are used to reflect the system's recent performance based on parameter correction amounts. The dynamic threshold is obtained based on the mean and standard deviation of the real-time monitoring indicators during the stable operation period of the system.

7. The intelligent management method for a die-cutting workshop production line according to claim 1, characterized in that, The real-time early warning commands include: sending load reduction / shutdown commands to the target equipment, pushing early warning notifications to on-site personnel, automatically dispatching alternative equipment to take over subsequent tasks, and packaging and archiving fault-related segments.

8. A system for implementing the intelligent management method for a die-cutting workshop production line as described in claim 1, characterized in that, The system includes: The data acquisition and fusion module is used to deploy multiple sensors at key workstations in the die-cutting workshop production line to collect multi-dimensional status data to form an original matrix. The original matrix is ​​then preprocessed and fused to generate a fused matrix. The deep perception modeling module is used to input the fused matrix into a pre-constructed deep perception model, extract local features using a dual-domain perception convolution kernel, process them using a channel attention module with prior physical knowledge, and output a state encoding vector and a channel attention vector. The scheduling optimization module is used to take the state encoding vector as the environmental state input, construct a deep reinforcement learning model, design a reward function that includes a task-resource cross-fit index, and output an action sequence and dynamic policy distribution. The quality control module is used to combine the action sequence and dynamic strategy distribution to construct a dynamic process offset index, generate a process offset value, and when the parameter correction value exceeds a preset threshold, adjust the current process parameters through a parameter correction network, output the adjusted new parameter values, and form a parameter correction sequence. The fault early warning module is used to construct real-time monitoring indicators based on the process offset value and parameter correction sequence, and set dynamic thresholds. When the real-time monitoring indicators are continuously higher than the dynamic threshold for a certain period of time, the fault response program is triggered and a real-time early warning instruction sequence is output.

Images