Wire feeding processing intelligent operation and maintenance system based on cloud computing

Abstract

The invention discloses a wire feeding processing intelligent operation and maintenance system based on cloud computing, and relates to the technical field of industrial automation and cloud computing. The problems that in an existing system, due to data transmission delay, abnormal response of equipment is not timely, system stability is affected by unreasonable computing resource scheduling, and the operation and maintenance effect is poor due to poor adaptability of a dynamic maintenance strategy are solved. Comprising an edge layer, a transmission layer, a scheduling layer, a cloud layer and an application layer. The edge layer collects data and makes an emergency response; the transmission layer optimizes data interaction based on network parameters; the scheduling layer balances task allocation through a deep reinforcement learning algorithm; the cloud layer constructs a prediction model and generates a dynamic maintenance strategy according to a result; the application layer visually displays and distributes operation and maintenance tasks; according to the method, the fault response speed, the resource scheduling balance and the maintenance strategy adaptivity are greatly improved, and the production continuity and the product quality are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial automation and cloud computing, and more particularly to a cloud computing-based feeding line processing intelligent operation and maintenance system. BACKGROUND

[0002] With the accelerated digital transformation, cloud computing has become a key technology to improve business efficiency and flexibility. The cloud computing-based feeding line processing intelligent operation and maintenance system emerges as the times require, aiming to improve the efficiency and quality of the feeding line processing process through intelligent operation and maintenance technology, to meet the urgent needs of modern industry for intelligent and automated operation and maintenance.

[0003] Feeding line processing refers to a system for automatically supplying, scheduling and distributing materials in the production process, and is widely used in automobile manufacturing, electronic assembly and high-end equipment manufacturing, etc. In the traditional feeding line processing operation and maintenance, it mainly relies on manual inspection and simple equipment monitoring system. With the explosive growth of data and the increasing demand for operation and maintenance efficiency, this approach is gradually unable to meet the needs. The emergence of cloud computing technology makes it possible to solve these problems. It virtualizes physical resources into multiple virtual resources through virtualization technology, achieving efficient use of resources; and uses distributed storage technology to store massive device operation data, fault information, etc.

[0004] For example, in the literature "Design and Research of Big Data Intelligent Operation and Maintenance System Based on Cloud Computing", device operation data is collected by sensors and uploaded to the cloud for analysis and processing, realizing real-time monitoring of device status and fault warning; secondly, the patent CN112732402B discloses a resource scheduling method, device and equipment of a cloud computing center, which allocates cloud tasks represented by a cloud task information set to virtual machines represented by a virtual machine information set, reducing the execution time and resource consumption of cloud computing center resource scheduling; however, the above literature still has some shortcomings in the practical application of the feeding line processing intelligent operation and maintenance system.

[0005] Firstly, since the data needs to be transmitted to the cloud for processing and then returns the control instruction, there is a delay in this process. Especially when the feeding line equipment suddenly fails, such as motor overload, the feeding line speed needs to be adjusted immediately, but network delay and cloud processing time may cause the control instruction to be issued in time, which cannot quickly respond to equipment abnormalities, seriously affecting the continuity of production and product quality. Secondly, although edge computing can reduce data processing delay, due to the limited computing resources of edge devices, if the scheduling strategy is unreasonable, it may cause idle cloud computing resources and overload edge computing tasks, affecting the stability of the whole system. Thirdly, the failure mode and operation and maintenance demand in the feeding line processing process are dynamic, but the existing dynamic maintenance strategy often cannot adapt to such changes, resulting in poor operation and maintenance effect. SUMMARY

[0006] In view of the deficiencies of the prior art, the application discloses a cloud computing-based intelligent operation and maintenance system for feeding line processing, and aims at solving the problems in the background art.

[0007] In order to achieve the above technical effects, the application adopts the following technical solutions: The cloud computing-based intelligent operation and maintenance system for feeding line processing comprises an edge layer, a transmission layer, a scheduling layer, a cloud layer and an application layer. The edge layer is used for collecting and uploading feeding line equipment operation data, and performing emergency response processing and feeding line equipment control, and comprises a collection module, a processing module and a control module. The transmission layer is used for determining and optimizing data interaction between the edge and the cloud based on link network delay, bandwidth utilization and data packet loss rate parameters by using an adaptive transmission optimization algorithm. The scheduling layer is used for dynamically adjusting the distribution of computing tasks according to the complexity and urgency of the tasks by using a load balancing scheduling algorithm based on deep reinforcement learning based on edge node and cloud server resource monitoring data and task urgency indexes. The cloud layer is used for the edge layer to construct a fault prediction model, a cored wire addition amount prediction model and a nodular cast iron defect prediction model by using a long short-term memory network and a convolutional neural network based on historical operation and maintenance data, real-time monitoring data and environmental variable data, to determine whether the equipment operation state, the cored wire addition amount and the nodular cast iron defect prediction results exceed a preset risk threshold, if yes, to generate an adaptive dynamic maintenance strategy by using a reinforcement learning algorithm based on a Markov decision process in combination with deep data analysis results and real-time states of the equipment, and to output a control instruction to the edge layer, and if no, to output a warning report and maintenance suggestions after fusing the current prediction results with historical data by statistical analysis and archiving methods. The application layer is used for displaying the equipment operation data, the deep data analysis results and the dynamic maintenance strategy by using a data visualization tool, and generating an operation and maintenance task list according to the dynamic maintenance strategy, and assigning the tasks to corresponding maintenance personnel by using a workflow management system.

[0008] As a further technical scheme of the present application, the acquisition module is used for real-time acquisition and preprocessing of the feeding device operation data, including motor current, voltage, speed, feeding speed, and device temperature parameters, and the preprocessed feeding device operation data is sent to the processing module; the processing module is used for data cleaning, time sequence alignment and feature extraction of the feeding device operation data through a real-time stream processing method, and judges whether the data is abnormal through a fault detection model, if yes, transmits an emergency handling instruction to the control module through a local emergency handling mechanism, and a local diagnosis result is sent to the cloud layer through the transmission layer; if not, the preprocessed data is sent to the cloud layer through the transmission layer; the control module is used for receiving the emergency handling instruction sent by the processing module and the control instruction sent by the cloud layer, and controlling the feeding device actuator, including the motor driver and the speed regulator, through a programmable logic controller.

[0009] As a further technical scheme of the present application, the working method of the processing module is: first, signal preprocessing and noise suppression are performed through Kalman filtering and wavelet denoising, dynamic time warping algorithm is used to correct the timestamp error of each data source, and a self-encoder is used to reduce dimension and output a low-dimensional feature vector to a fault detection model, the fault detection model calculates the deviation between the current feature vector and the historical normal state feature in real time through the Euclidean distance and Mahalanobis distance measurement mechanism, and performs online weight update using the fast gradient descent algorithm; if the deviation result exceeds the preset threshold, an emergency handling instruction is generated through an embedded rule engine, and is sent to the control module in JSON format through the UDP / TCP protocol, and the local diagnosis result is uploaded to the cloud layer in the MQTT protocol; if the deviation result is lower than the preset threshold, the normalized feature data is transmitted to the cloud data warehouse through the HTTP / 2 protocol.

[0010] As a further technical scheme of the present application, the working method steps of the adaptive transmission optimization algorithm include: S1, based on the real-time collected link delay , bandwidth utilization , packet loss rate and network jitter data, the collected data is smoothed by the exponential weighted moving average method to obtain the smoothed value 、 、 and ; S2, construct a transmission cost function to measure the link quality, and the formula expression of the transmission cost function is: (1) In formula (1), This is the normalized value for bandwidth utilization. , , , The adjustment coefficients represent the effects of delay, jitter, packet loss, and bandwidth utilization on transmission performance, respectively. S3, Based on transmission cost function By fitting the network state change curve using the least squares method and discretizing the network state using a fuzzy logic control mechanism, the optimal data transmission rate is generated. With group size If the transmission cost function If the threshold is exceeded, a reinforcement learning scheduling algorithm is used, based on the state vector ( , , ) and action sets Calculate the state action function and update the transmission strategy in real time; if the transmission cost function If the value is below the threshold, the current parameter configuration is maintained, and the parameter configuration strategy is issued through the feedback loop.

[0011] As a further technical solution of the present invention, the working method of the load balancing scheduling algorithm based on deep reinforcement learning is as follows: First, the resource status data of edge nodes and cloud servers are collected in real time through the resource monitoring module, including CPU utilization. Memory usage Network bandwidth and task queue length The algorithm extracts features from resource state data using a temporal convolutional network to generate a resource state vector X. Simultaneously, the deep reinforcement learning-based load balancing scheduling algorithm evaluates task complexity using a task classifier. and urgency Generate task feature vectors Based on resource state vectors and task feature vectors, the execution value of each task on edge nodes or cloud servers is calculated using the Q-value function. Where A represents the task assignment action, and the update formula for the Q-value function is: (2) In formula (2), For instant rewards, For learning rate, As a discount factor; if the urgency of the task... If the task exceeds a preset threshold, a high-priority scheduling mechanism is triggered, prioritizing the allocation of the task to edge nodes with available resources; if the task complexity... If the threshold is exceeded, the cloud collaborative computing mode will be activated, and the task will be split and distributed to the cloud server in parallel.

[0012] As a further technical solution of the application, the fault prediction model performs multi-scale decomposition on historical operation data and wire feeding equipment operation data through a time series feature extraction mechanism to construct a separated representation of short-period trend items and long-term evolution items; in the data preprocessing process, the fault prediction model constructs a probabilistic graph model using equipment sensor data based on an anomaly detection self-supervised learning framework, calculates the dynamic correlation between each operation parameter, and assigns feature weights through a graph attention mechanism; based on the processed feature data, the fault prediction model adopts an incremental learning strategy to construct a fault prediction network, and dynamically adjusts the parameters of long short-term memory networks and convolutional neural networks through an adaptive gradient optimization algorithm; in the prediction calculation process, the fault prediction model calculates the short-period trend offset through an autoregressive moving average model, and extracts long-term evolution patterns in combination with an attention-guided time difference method.

[0013] As a further technical solution of the application, the working method of the core-spun yarn addition amount prediction model is as follows: first, a time series representation learning method of self-supervised learning is used to extract latent key variables in the smelting process through a variational autoencoder, and principal component analysis is used to reduce the dimension of the data; based on the reduced data, the core-spun yarn addition amount prediction model adopts a multi-scale convolutional neural network to extract time series features, capture local short-period fluctuation characteristics of core-spun yarn consumption, and calculate the influence weight of key smelting parameters on core-spun yarn addition amount in combination with an attention mechanism; in the prediction calculation stage, the core-spun yarn addition amount prediction model adopts a long short-term memory network based on Bayesian optimization to construct a nonlinear time series prediction model; if the prediction error exceeds the set threshold, an error compensation mechanism of incremental learning is adopted, a prediction residual is calculated through an autoregressive integral sliding average model, and a recursive neural network is used to regress and optimize the historical deviation; if the prediction result exceeds the preset risk threshold, an adjustment strategy is generated and an optimization control instruction is output to the edge layer.

[0014] As a further technical scheme of the application, the spheroidal graphite cast iron defect prediction model is based on historical smelting process parameters, real-time production data and environmental variable data, and first constructs a high-dimensional feature space through a heterogeneous data fusion mechanism; in the feature construction process, a smelting process parameter correlation modeling method based on a graph attention network is used to extract the nonlinear coupling relationship between variables, and a nested variational autoencoder is used for feature dimension reduction of the data; in the defect prediction stage, the spheroidal graphite cast iron defect prediction model extracts high-dimensional features from the temperature field, stress field and composition fluctuation data in the casting process through a spatio-temporal fusion convolutional neural network, obtains the contribution of different smelting stages to defect formation, and uses a multi-scale residual network to capture the short-time dynamic characteristics of the defects, thereby improving the timeliness of the prediction model; in the prediction and reasoning process, the spheroidal graphite cast iron defect prediction model uses a double-layer long short-term memory network balance model based on Bayesian optimization to balance long-term defect trend prediction and short-term defect fluctuation; if the prediction result exceeds the set risk threshold, a defect avoidance strategy based on a Markov decision process is triggered, the amount of spheroidizing agent added and the cooling rate are optimized through reinforcement learning method, and dynamic adjustment instructions are output to the edge layer; if the prediction error exceeds the threshold, the model parameters are adjusted in real time based on an adaptive incremental learning mechanism.

[0015] As a further technical scheme of the application, the working method of the reinforcement learning algorithm based on the Markov decision process is as follows: first, a probabilistic graphical model is used to represent high-dimensional features of the wire feeding equipment operation data to construct a state transition matrix, and the state transition probability is dynamically adjusted in combination with a Bayesian update mechanism; then, an attention mechanism-based policy gradient method is used to calculate the optimal execution path of different maintenance strategies, and an experience replay mechanism is used to store historical strategy execution results; if the current prediction result exceeds the risk threshold, a double Q learning mechanism is used to evaluate the impact of different actions on long-term returns; if the prediction error exceeds the set threshold during strategy updating, a trust region optimization strategy is used to constrain the strategy distribution; finally, an A3C architecture based on adaptive entropy regulation is used to parallelly optimize the dynamic maintenance strategy, and control instructions are issued to the edge layer.

[0016] As a further technical scheme of the application, the application layer is also used to trigger an alarm mechanism based on a rule engine when an abnormal state or potential failure risk occurs in the device, and notify the operation and maintenance personnel through short message, email or system pop-up window.

[0017] Based on the above technical scheme, the application has the following positive and beneficial effects: The processing module of the edge layer can analyze the feeding device operation data in real time. When a sudden equipment failure is detected, such as a sudden motor overload, the processing module can immediately transmit emergency processing instructions to the control module through a local emergency processing mechanism, directly adjusting the feeding speed. This local decision and control method bypasses the process of transmitting data to the cloud and returning instructions, avoiding network delays and cloud processing time, allowing for quick response to equipment abnormalities, ensuring production continuity and reducing damage to product quality due to delayed fault handling. Second, the transmission layer uses an adaptive transmission optimization algorithm based on link network latency, bandwidth utilization, and packet loss rate parameters to determine and optimize data interaction between the edge and the cloud. This algorithm can dynamically adjust data transmission strategies based on real-time network status, reducing data transmission delays and improving data transmission efficiency and stability, further reducing response delays due to network problems.

[0018] The scheduling layer uses a load balancing scheduling algorithm based on deep reinforcement learning, combining edge node and cloud server resource monitoring data and task urgency indicators to dynamically adjust the allocation of computing tasks based on task complexity and urgency. This algorithm can real-time perceive resource usage of edge devices and cloud servers, reasonably allocate complex tasks to cloud servers and simple tasks to edge nodes, avoiding idle cloud computing resources and edge computing task overload, improving overall system computing resource utilization and enhancing system stability.

[0019] The cloud layer uses long short-term memory networks and convolutional neural networks to build a fault prediction model, a cored wire addition amount prediction model, and a ductile iron defect prediction model, which can accurately predict equipment operating status, cored wire addition amount, and ductile iron defect conditions. When the prediction result exceeds the preset risk threshold, a reinforcement learning algorithm based on Markov decision process is used to generate an adaptive dynamic maintenance strategy based on deep data analysis results and real-time equipment status. This strategy can be adjusted in real-time based on the actual operation of the equipment and the dynamic changes in the fault mode, effectively adapting to the dynamic nature of the operation and maintenance requirements during the feeding process, improving operation and maintenance effectiveness, and reducing equipment failure rates and maintenance costs. In addition, even if the prediction result does not exceed the preset risk threshold, the cloud layer will fuse the current prediction result with historical data through statistical analysis and archiving methods to output an early warning report and maintenance recommendations. This helps operation and maintenance personnel to identify potential problems in advance and take preventive measures, further improving system reliability and operation and maintenance efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description only represent some of the embodiments of the present application, and all other embodiments obtained by a person of ordinary skill in the art without creative labor based on these accompanying drawings also belong to the protection scope of the present application. Figure 1 The figure is a framework of the intelligent operation and maintenance system for feeding line processing based on cloud computing. Figure 2 The figure is a processing framework of the edge layer. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor also belong to the protection scope of the present application.

[0022] As shown in the figure, an intelligent operation and maintenance system for feeding line processing based on cloud computing includes an edge layer, a transmission layer, a scheduling layer, a cloud layer and an application layer. Figure 1 In the edge layer, the collection module collects the operation data of the feeding line equipment by means of various sensors, covering motor current, equipment temperature, feeding speed, etc. The processing module preliminarily processes the collected raw data, and uses real-time stream analysis technology to identify abnormal data. Once an abnormality is found, the local emergency response mechanism is triggered immediately, and the control module quickly adjusts the equipment parameters. The processed normal data and abnormal situation report are prepared to be uploaded to the transmission layer. The transmission layer receives the data uploaded by the edge layer, and uses an adaptive transmission optimization algorithm to dynamically adjust the data transmission mode and priority according to the real-time status of the network, such as link network delay, bandwidth utilization and data packet loss rate. For example, the high-speed channel is used to transmit the emergency abnormal situation report in priority, so as to ensure that important information can be sent to the cloud layer in time.

[0023] The scheduling layer monitors the resource usage of the edge node and the cloud server in real time, combines the emergency degree and complexity of the task, and uses a load balancing scheduling algorithm based on deep reinforcement learning. For complex data analysis tasks, they are allocated to the cloud server with sufficient computing resources; for simple tasks, they are arranged to be processed in the edge node, so as to realize the reasonable allocation of computing tasks.

[0024]

[0025] ​The cloud layer receives data from the transport layer and performs fault prediction, core wire addition prediction, and ductile iron defect prediction. The application layer receives early warning reports, maintenance suggestions, and dynamic maintenance strategies from the cloud layer and presents this information to maintenance personnel in intuitive charts and reports using data visualization tools. Simultaneously, it generates a detailed maintenance task list based on the dynamic maintenance strategy, assigns tasks to appropriate maintenance personnel using a workflow management system, and tracks task progress.

[0026] Among them, such as Figure 2 As shown, the edge layer is used to collect and upload operating data of the wire feeding equipment, and to perform emergency response processing and control of the wire feeding equipment. It includes a collection module, a processing module, and a control module. The collection module collects and preprocesses the operating data of the wire feeding equipment in real time, including motor current, voltage, speed, wire feeding speed, and equipment temperature parameters, and sends the preprocessed operating data to the processing module. The processing module performs data cleaning, time-series alignment, and feature extraction on the operating data of the wire feeding equipment using a real-time streaming processing method. It also uses a fault detection model to determine if the data is abnormal. If so, it transmits emergency processing instructions to the control module through a local emergency processing mechanism, and sends the local diagnostic results to the cloud layer through the transport layer. If not, it sends the preprocessed data to the cloud layer through the transport layer. The control module receives emergency processing instructions sent by the processing module and control instructions sent by the cloud layer, and... The feeder device actuator, including a motor driver and a speed regulator, is controlled by a programmable logic controller. The processing module operates as follows: first, signal preprocessing and noise suppression are performed using Kalman filtering and wavelet denoising; dynamic time warping is used to correct timestamp errors from various data sources; and a low-dimensional feature vector is output to the fault detection model via an autoencoder. The fault detection model uses Euclidean and Mahalanobis distance metrics to calculate the deviation between the current feature vector and historical normal state features in real time, and performs online weight updates using a fast gradient descent algorithm. If the deviation exceeds a preset threshold, an emergency processing command is generated by the embedded rule engine, sent to the control module in JSON format via UDP / TCP protocol, and the local diagnostic results are uploaded to the cloud layer via MQTT protocol. If the deviation is below the preset threshold, the normalized feature data is transmitted to the cloud data warehouse via HTTP / 2 protocol.

[0027] In specific implementation, the processing module adopts Kalman filtering and wavelet denoising for signal preprocessing and noise suppression. Kalman filtering is a recursive optimal estimation algorithm that predicts and updates the system state, uses the current measurement value and the estimated value at the last time to obtain the optimal estimation at the current time. In the data processing of the wire feeding equipment, Kalman filtering can accurately estimate the real signal value according to the historical trend of the data and the current measurement value, and effectively suppress the interference of random noise. Wavelet denoising is based on the multi-resolution analysis characteristics of wavelet transform, which decomposes the signal into different frequency subbands, then performs threshold processing on each subband to remove the high-frequency subband information where the noise is located, and then reconstructs the denoised signal through wavelet inverse transform. Dynamic time warping (DTW) algorithm is used to correct the timestamp error of each data source. In actual data collection of the wire feeding equipment, the sampling frequency and time of different sensors may be different, resulting in inconsistent timestamps of the data. The DTW algorithm finds the optimal matching path between two time series, so that they can be aligned in time, thereby providing a unified time reference for subsequent analysis. Autoencoder is an unsupervised learning neural network model, which maps input data to a low-dimensional feature space through an encoder, and then reconstructs the low-dimensional feature vector to the original input data through a decoder. In this process, the autoencoder learns the main features of the data, reducing the high-dimensional data to a low-dimensional feature vector. Through autoencoder dimension reduction, not only the storage space and computational complexity of the data can be reduced, but also the essential features of the data can be extracted, improving the accuracy of fault detection.

[0028] The fault detection model calculates the deviation between the current feature vector and the historical normal state feature in real time through the Euclidean distance and Mahalanobis distance measurement mechanism. Euclidean distance is the most commonly used distance measurement method, which calculates the straight-line distance between two vectors, reflecting the absolute difference of the vectors in space. Mahalanobis distance considers the covariance information of the data, which can better handle the correlation and distribution characteristics of the data. For data with different variances and correlations, Mahalanobis distance can more accurately measure the difference between them. Fast gradient descent algorithm is used for online weight update. The basic idea of gradient descent algorithm is to update the weights of the model along the negative gradient direction of the objective function, so that the value of the objective function is constantly reduced. Fast gradient descent algorithm optimizes the traditional gradient descent algorithm by adjusting the learning rate and other methods, which speeds up the convergence speed and can find the optimal model weight in a short time, improving the real-time performance and accuracy of the fault detection model.

[0029] When the deviation result exceeds the preset threshold, the embedded rule engine generates emergency treatment instructions according to the preset rules. The embedded rule engine is a lightweight rule execution system that can match and reason according to the input conditions and rules in the rule base, quickly generating corresponding decision results. The generated emergency treatment instructions are issued to the control module in JSON format through the UDP / TCP protocol. JSON is a lightweight data exchange format with the characteristics of simplicity, readability, and easy parsing, suitable for data transmission between different systems. UDP protocol has the characteristics of fast transmission speed and small overhead, suitable for high real-time requirement of emergency treatment instruction transmission; TCP protocol has the characteristics of reliable transmission, which can guarantee the integrity of data. The local diagnosis result is uploaded to the cloud layer through the MQTT protocol. MQTT is a lightweight message transmission protocol with low bandwidth occupation and easy implementation, suitable for data transmission between resource-limited edge devices and the cloud. When the deviation result is below the preset threshold, the normalized feature data is transmitted to the cloud data warehouse through the HTTP / 2 protocol. HTTP / 2 protocol has greatly improved performance, using binary framing, multiplexing and other technologies to improve data transmission efficiency and speed.

[0030] In the implementation of the cloud computing-based thread feeding intelligent operation and maintenance system, the working method of the edge layer is as follows: first, the acquisition module continuously and real-time acquires the running data of the thread feeding equipment and performs preliminary preprocessing on it. The preprocessed data is sent to the processing module, which further processes the data according to the established process. First, remove noise by Kalman filtering and wavelet denoising, then correct timestamp error using dynamic time warping algorithm, then use autoencoder for dimensionality reduction to get low-dimensional feature vector. The low-dimensional feature vector is input into the fault detection model, which calculates the deviation through Euclidean distance and Mahalanobis distance measurement mechanism, and updates the weight online using fast gradient descent algorithm. According to the comparison of the deviation result and the preset threshold, different decisions are made. If the deviation exceeds the threshold, the embedded rule engine generates emergency treatment instructions, which are issued to the control module through the UDP / TCP protocol, and the local diagnosis result is uploaded to the cloud layer through the MQTT protocol; if the deviation is below the threshold, the normalized feature data is transmitted to the cloud data warehouse through the HTTP / 2 protocol. The control module receives the control instructions from the processing module and the cloud layer, and accurately controls the actuators of the thread feeding equipment such as motor driver and speed regulator through the programmable logic controller, to ensure the normal operation of the equipment.

[0031] Compared with the prior art, the edge layer in the cloud computing-based feeding line processing intelligent operation and maintenance system has significant advantages and characteristics. First, its efficient real-time data processing capability enables the system to timely discover and respond to abnormal states of the equipment, greatly improving the safety and stability of production. Second, through the combination of local emergency processing mechanism and remote cloud monitoring, intelligent operation and maintenance of the equipment is realized, reducing the cost and risk of manual intervention. Finally, the collaborative work of the edge layer and the cloud layer builds a complete data acquisition, processing, analysis and control closed loop, providing strong technical support for the whole life cycle management of the equipment.

[0032] The transmission layer is configured to determine and optimize data interaction between the edge and the cloud based on link network latency, bandwidth utilization and data packet loss rate parameters, and an adaptive transmission optimization algorithm is used; the working method steps of the adaptive transmission optimization algorithm include: S1, based on real-time collected link latency , bandwidth utilization , data packet loss rate and network jitter data, the collected data is smoothed by an exponential weighted moving average method to obtain smoothed values , , and ; S2, constructing a transmission cost function to measure link quality, and the formula expression of the transmission cost function is: (1) In formula (1), is the normalized value of bandwidth utilization; , , , is an adjustment coefficient, respectively representing the influence of latency, jitter, packet loss and bandwidth utilization on transmission performance; S3, based on the transmission cost function , a network state change curve is fitted by a least squares method, a network state is discretized by a fuzzy logic control mechanism, and an optimal data transmission rate and packet size are generated; if the transmission cost function exceeds a preset threshold value, a reinforcement learning scheduling algorithm is used to calculate a state-action function based on a state vector , , and an action set to update the transmission strategy in real time; if the transmission cost function If the threshold is not met, the current parameter configuration is maintained, and the parameter configuration strategy is issued through the feedback loop.

[0033] In specific implementation, the core function of the transmission layer is to optimize the data interaction between the edge layer and the cloud layer, to reduce network delay, reduce data loss and improve transmission efficiency. The transmission layer first collects the key network parameters of the link in real time through the network monitoring module, including link delay, bandwidth utilization, packet loss rate and network jitter. Since the network state has high time-varying nature, direct use of raw data may lead to unstable transmission strategy. Therefore, the system uses the Exponentially Weighted Moving Average (EWMA) method to smooth the network state data.

[0034] The Exponentially Weighted Moving Average method is a recursive filtering method, whose core idea is to assign a higher weight to the latest data point and an exponentially decaying weight to the earlier data point, so that the smoothed value can quickly respond to the latest network changes while reducing the impact of short-term jitter. It can effectively reduce short-term noise while maintaining the trend characteristics of the data, so that the system can make subsequent decisions based on more stable network state information.

[0035] After obtaining the smoothed network parameters, the system needs to evaluate the current link quality to decide whether to adjust the transmission strategy. For this purpose, the transmission layer constructs a transmission cost function to measure the overall transmission performance of the link. This function considers factors such as delay, packet loss, jitter and bandwidth utilization, and uses a normalization method to uniformly process indicators of different dimensions. Through this function, the system can quantify the link quality and select appropriate transmission parameters through the optimal strategy.

[0036] After calculating the transmission cost function, the system needs to adjust the data transmission strategy based on the network state. Specifically, the transmission layer uses the Least Squares Method (LSM) to fit the network state change curve to predict the network trend in the short term. At the same time, the system introduces the Fuzzy Logic Control (FLC) mechanism, which divides the network state into four levels: excellent, good, medium and poor, and dynamically adjusts the data transmission rate and packet size according to different levels.

[0037] When the network state is poor (i.e., the transmission cost function exceeds the preset threshold), the system further optimizes the transmission strategy through a scheduling algorithm based on Reinforcement Learning (RL). Reinforcement Learning builds a state-action-reward model to constantly explore the optimal decision scheme. Specifically, the system defines: State vector: represents the current network environment, Action set: including strategies such as reducing data rate, adjusting data packet size, switching backup link, etc. State-action function: used to evaluate the benefits of performing a certain action under a specific network state.

[0038] The system uses Q-learning algorithm or Deep Q-Network (DQN) method to optimize the state-action function according to historical transmission data and update the transmission strategy in real time, thereby improving the network adaptability.

[0039] If the transmission cost function is lower than the set threshold, the system considers that the current network state is stable, maintains the existing transmission strategy, and periodically updates the parameters through the feedback loop to ensure that the system can adapt to network changes in the long term.

[0040] Compared with the prior art, the transmission layer in the cloud computing-based feeding line processing intelligent operation and maintenance system has shown significant advantages and characteristics. First, by monitoring and analyzing link parameters in real time, the algorithm can dynamically adjust the transmission strategy to adapt to the changing network environment, improving the stability and reliability of data transmission. Second, by constructing the transmission cost function and introducing fuzzy logic control and reinforcement learning algorithm, the algorithm can more accurately understand the network state and generate the optimal transmission strategy, further improving the efficiency and quality of data transmission. Finally, the algorithm also has certain adaptability and scalability, which can be flexibly adjusted and optimized according to the actual needs of different application scenarios, providing a strong guarantee for the efficient operation of the feeding line processing intelligent operation and maintenance system.

[0041] The scheduling layer is used to monitor the resource data of edge nodes and cloud servers and the task urgency index, and a load balancing scheduling algorithm based on deep reinforcement learning is used to dynamically adjust the allocation of computing tasks according to the complexity and urgency of the tasks. The working method of the load balancing scheduling algorithm based on deep reinforcement learning is as follows: first, the resource state data of edge nodes and cloud servers is collected in real time by the resource monitoring module, including CPU utilization , memory occupancy , network bandwidth and task queue length , and the resource state data is feature extracted based on time sequence convolution network to generate resource state vector X; at the same time, the load balancing scheduling algorithm of deep reinforcement learning evaluates the complexity and urgency of the task through the task classifier to generate task feature vector ; based on the resource state vector and the task feature vector, the execution value of each task on the edge node or the cloud server is calculated through the Q value function where A represents the task allocation action, and the update formula of the Q-value function is: (2) In formula (2), is the immediate reward, is the learning rate, is the discount factor; if the urgency of the task exceeds the preset threshold, a high-priority scheduling mechanism is triggered to preferentially allocate the task to the edge node with idle resources; if the complexity of the task exceeds the preset threshold, a cloud collaborative computing mode is enabled to split and parallelly allocate the task to the cloud server.

[0042] At the scheduling layer, the system first collects the key resource state data of the edge nodes and the cloud server in real time through the resource monitoring module, including: CPU utilization (measuring computing power load); memory occupancy rate (measuring available memory resources); network bandwidth (measuring data transmission capacity); task queue length (measuring the load of the current computing node).

[0043] Since these data have time series correlation, directly using the original data may lead to inaccurate resource state evaluation. Therefore, the system uses a temporal convolutional network (TCN) to extract features from the resource state data. The temporal convolutional network uses 1D causal convolution to ensure that the calculation at the current time step does not depend on the data at future time steps, thereby satisfying the time series causality. And use the extended causal convolution (Dilated Causal Convolution), through the exponentially growing receptive field to efficiently capture long-term dependencies, so that the system can learn the long-term trend of resource state. After processing by multiple convolution layers, a resource state vector is finally obtained, which is used for subsequent task scheduling decisions.

[0044] Then, the system needs to evaluate the complexity and urgency of each task to determine its optimal computing location. The specific method is as follows: Task complexity evaluation: use a task classifier to classify tasks based on their computing requirements (such as compute-intensive vs. I / O-intensive). Extract task features through a deep neural network (DNN) and generate a task feature vector TT, which includes computing time, data input size, etc.

[0045] Task urgency evaluation: by setting a task timeout risk, calculate the probability that the task needs to be completed within a certain time, and quantify it as an urgency indicator. When the task urgency exceeds the threshold, the task is marked as a high-priority task and the fast scheduling mechanism is triggered.

[0046] Finally, the system forms a resource state vector and a task feature vector, providing input for subsequent deep reinforcement learning decisions.

[0047] During task scheduling decision-making, the system uses a Q-learning-based reinforcement learning method to continuously optimize task allocation strategies. The decision logic of reinforcement learning is as follows: Calculate Q value: For each task, calculate its Q value on different computing nodes to measure its execution value.

[0048] If the task urgency exceeds the set threshold, the system prioritizes resource-free edge nodes to ensure the task can be executed quickly. If the task complexity is high (such as requiring long-time calculation), the system uses a cloud collaborative computing mode to split the task into multiple sub-tasks and schedule them to multiple cloud servers in parallel, improving processing efficiency. Through reinforcement learning, the system continuously updates Q values to optimize task scheduling strategies, making future decisions more efficient.

[0049] In implementation, traditional load balancing methods usually use round-robin scheduling or static allocation strategies, which cannot adapt to dynamic computing environments. This scheme is based on reinforcement learning and can intelligently schedule tasks according to task requirements and computing resource states, avoiding waste of computing resources and improving computing efficiency. In addition, existing scheduling strategies usually cannot distinguish the priority of tasks, leading to high-priority tasks being blocked by low-priority tasks. This system assesses task urgency to ensure that urgent tasks can be prioritized and processed, improving the system's response capability. Secondly, traditional methods usually only consider the execution of tasks on a single computing node, while this scheme uses a cloud collaborative computing mode to split high-complexity tasks for parallel execution, improving computing throughput.

[0050] The cloud layer is used for the edge layer to adopt long short-term memory network and convolutional neural network to construct a fault prediction model, a cored wire addition amount prediction model and a nodular cast iron defect prediction model based on historical operation and maintenance data, real-time monitoring data and environmental variable data, to judge whether the equipment operation state, cored wire addition amount and nodular cast iron defect prediction results exceed a preset risk threshold, if yes, through a reinforcement learning algorithm based on Markov decision process combined with deep data analysis results and real-time state of the equipment, an adaptive dynamic maintenance strategy is generated, and a control instruction is output to the edge layer, if no, a warning report and maintenance suggestion are output after the current prediction result is fused with historical data through statistical analysis and archiving method; the fault prediction model performs multi-scale decomposition on historical operation data and wire feeding equipment operation data through a time series feature extraction mechanism, to construct a separated representation of short-period trend items and long-term evolution items; in a data preprocessing process, the fault prediction model uses device sensor data to construct a probabilistic graph model based on a self-supervised learning framework of anomaly detection, to calculate dynamic correlation between each operation parameter, and give feature weights through a graph attention mechanism; based on the processed feature data, the fault prediction model adopts an incremental learning strategy to construct a fault prediction network, and dynamically adjusts parameters of the long short-term memory network and the convolutional neural network through an adaptive gradient optimization algorithm; in a prediction calculation process, the fault prediction model calculates short-period trend deviation through an autoregressive moving average model, and extracts long-term evolution patterns combined with an attention-guided time series difference method.

[0051] In specific implementation, due to the existence of noise, missing values and non-synchronization in original data, an anomaly detection self-supervised learning framework is adopted to construct a probabilistic graph model to calculate the dynamic correlation between various operating parameters, and a graph attention mechanism (GAT) is used to give different feature weights and improve data quality. In addition, the system uses wavelet transform to decompose time series data at multiple scales, distinguishing between short-term trend items and long-term evolution items, and enhancing the understanding of time series patterns. In order to correct the timestamp error of the data, the dynamic time warping (DTW) method is used to align historical data and real-time data, avoiding the prediction bias caused by uneven sampling intervals. In the feature extraction stage, the fault prediction model uses LSTM and CNN to capture the time-dependent features and spatial pattern features of the device operating state. LSTM has strong memory ability and can effectively process data with long-term dependencies, using the collaborative action of forget gate, input gate and output gate to extract long-term evolution patterns of device operating data, while CNN uses one-dimensional convolution (1D-CNN) to extract local patterns from sensor data and combines with the pooling layer to reduce noise interference and improve the generalization ability of the model. At the same time, the model uses an incremental learning strategy to dynamically adjust the parameters of LSTM and CNN through an adaptive gradient optimization algorithm (AdaGrad), so that the model can adapt to changes in the operating environment and improve long-term prediction capability. In the prediction calculation process, in order to enhance the detection ability of short-term trends, the system uses an autoregressive moving average model (ARIMA), in which the autoregressive part is used to capture the correlation of historical data, the moving average part is used to smooth fluctuations, and the difference operation is used to eliminate long-term trends to ensure data stationarity. In addition, combined with the attention-guided time difference method, the system automatically selects key time step data through the attention mechanism to improve the extraction ability of long-term patterns, and calculates the rate of change of device state through time difference to enhance the detection ability of sudden failures. When the prediction result exceeds the preset risk threshold, the system automatically generates a maintenance strategy and optimizes the maintenance strategy through Markov decision process (MDP). MDP takes the real-time state of the device as input through the state-action-reward mechanism, and continuously optimizes the maintenance decision through reinforcement learning, so that the maintenance strategy can adapt to different operating environments. If the prediction result shows that the device is about to fail, the system will send control instructions to the edge layer in advance to adjust the device operating parameters, or suggest performing preventive maintenance to reduce the risk of unplanned downtime. If the prediction result shows that the device is in normal operating state, the system will archive the current prediction result and fuse it with historical data to optimize future operation and maintenance analysis.

[0052] Compared with the traditional fixed cycle maintenance method, the system can dynamically adjust the maintenance plan, avoid wasting resources due to premature maintenance, and reduce production losses caused by failure to predict in time. In addition, the traditional fault prediction method usually only relies on statistical models or single feature extraction methods, while the system combines deep learning, probability statistics and reinforcement learning technology, which can more comprehensively capture the running mode of the equipment and improve the accuracy of prediction. In particular, the combination of self-supervised learning framework and probabilistic graph model significantly improves the anomaly detection capability, enhancing the robustness and reliability of the system. At the same time, the application of incremental learning strategy enables the model to continuously adapt to new data distribution in long-term operation, avoiding the problem of decreased prediction ability of traditional static models when facing new data. In addition, the intelligent maintenance strategy of the system can dynamically adjust according to real-time prediction results, improving maintenance efficiency and reducing operating costs. Overall, the cloud computing-based wire feeding intelligent operation and maintenance system realizes the intelligentization of equipment state prediction, fault detection and maintenance strategy optimization through the integration of deep learning, reinforcement learning and probability statistics and other advanced technologies, making operation and maintenance management more accurate and highly automated, and providing effective technical support for enterprises to improve production efficiency and reduce maintenance costs.

[0053] In addition, the working method of the core-spun yarn addition amount prediction model is as follows: first, the time series representation learning method of self-supervised learning is used to extract the potential key variables in the smelting process by using the variational autoencoder, and the principal component analysis method is used to reduce the dimension of the data; based on the reduced dimension data, the core-spun yarn addition amount prediction model uses a multi-scale convolutional neural network to extract time series features, captures the local short-period fluctuation characteristics of core-spun yarn consumption, and combines the attention mechanism to calculate the influence weight of key smelting parameters on core-spun yarn addition amount; in the prediction calculation stage, the core-spun yarn addition amount prediction model uses a long short-term memory network based on Bayesian optimization to build a nonlinear time series prediction model; if the prediction error exceeds the set threshold, an incremental learning error compensation mechanism is used to calculate the prediction residual by using an autoregressive integral moving average model, and a recurrent neural network is used to optimize the historical deviation; if the prediction result exceeds the preset risk threshold, an adjustment strategy is generated and an optimized control instruction is output to the edge layer.

[0054] The ductile iron defect prediction model is based on historical smelting process parameters, real-time production data and environmental variable data. First, a high-dimensional feature space is constructed through a heterogeneous data fusion mechanism. In the feature construction process, a smelting process parameter correlation modeling method based on a graph attention network is used to extract the nonlinear coupling relationship between variables, and a nested variational autoencoder is used for feature dimension reduction. In the defect prediction stage, the ductile iron defect prediction model extracts high-dimensional features from the temperature field, stress field and composition fluctuation data during the casting process through a spatio-temporal fusion convolutional neural network, obtains the contribution of different smelting stages to defect formation, and uses a multi-scale residual network to capture the short-term dynamic characteristics of defect generation, thereby improving the timeliness of the prediction model. In the prediction reasoning process, the ductile iron defect prediction model uses a double-layer long short-term memory network balance model based on Bayesian optimization to balance long-term defect trend prediction and short-term defect fluctuation. If the prediction result exceeds the set risk threshold, a defect avoidance strategy based on Markov decision process is triggered, the addition amount of spherulite and the cooling rate are optimized through reinforcement learning method, and dynamic adjustment instructions are output to the edge layer. If the prediction error exceeds the threshold, the model parameters are adjusted in real time based on an adaptive incremental learning mechanism.

[0055] In specific implementation, the cored wire addition amount prediction model and the ductile cast iron defect prediction model jointly constitute the core link of intelligent optimization control and quality prediction. The core objective of the cored wire addition amount prediction model is to accurately estimate the actual addition amount of alloying elements based on historical operation data, real-time monitoring data and environmental variable data, so as to ensure that the final alloy composition meets the process requirements. The model relies on a deep learning architecture combining long short-term memory network (LSTM) and convolutional neural network (CNN), and constructs high-dimensional time series feature representation in view of the complexity of multi-factor coupling in the smelting process. First, in the data preprocessing stage, the model decomposes the input data through wavelet transform (Wavelet Transform) to separate the original time series into sub-signals of different scales, extracts short-term fluctuations and long-term trend features, and improves the prediction stability. At the same time, the dynamic time warping (DTW) method is used to correct the timestamp deviation between different data sources, ensuring the time sequence consistency of the input data. In the feature extraction stage, CNN is responsible for extracting local features from multi-dimensional input data, such as wire feeding speed, current, voltage, temperature, cored wire type, etc., and through convolution kernel weight adaptive learning, the relationship between different input variables is learned, while LSTM is mainly used to model the time dependence of the cored wire addition process, using gate unit (Gate Unit) to control the forgetting and remembering of historical information, and improving the learning ability of long-term trends. In addition, in order to enhance the generalization ability of the prediction model, the system adopts a self-supervised learning framework to build an anomaly detection mechanism, calculates the dynamic correlation between wire feeding parameters and alloy composition through a probabilistic graphical model (Probabilistic Graphical Model, PGM), and uses an attention mechanism (Attention Mechanism) to give higher weights to key features to improve prediction accuracy. Based on the above processed feature data, the model uses an incremental learning strategy, combined with an adaptive gradient optimization algorithm (Adaptive Gradient Optimizer, AdaGrad) to continuously adjust the model parameters of LSTM and CNN, so that they can adapt to the working condition changes of different heats, and improve the long-term prediction ability. In the prediction calculation process, the model uses an autoregressive moving average model (ARIMA) to calculate the short-period trend deviation of cored wire addition amount, and combines an attention-guided temporal differencing method (Attention-guided Temporal Differencing) to extract long-term evolution patterns, to reduce prediction error. When the prediction result exceeds the preset deviation threshold, the system will automatically adjust the wire feeding speed or alarm the operator to correct the parameters, ensuring the composition qualification rate.

[0056] On the other hand, the core objective of the ductile iron defect prediction model is to predict the likelihood of internal defects in castings based on real-time production data, raw material composition data, and process parameters, and to provide defect avoidance optimization recommendations. Since the quality of ductile iron is influenced by multiple factors such as carbon silicon ratio, inoculation treatment, magnesium treatment, cooling rate, etc., this model uses a multi-task learning (MTL) framework to predict the probability distribution of multiple defects such as shrinkage, graphite floating, and inclusions in a parallel manner. In the data processing stage, the model first pre-processes the input signal through wavelet denoising and Kalman filtering to eliminate random noise in the sensor data and improve data quality. Subsequently, using time series feature extraction methods, LSTM combined with CNN is used to extract key time series features in the smelting process, and graph attention network (GAT) is used to analyze the complex nonlinear relationship between raw material elements and defect occurrence probability. Since the formation of ductile iron defects is influenced by historical process data, the model uses a Transformer Encoder in the prediction process to establish a global association between historical data and current data through a self-attention mechanism, improving long-term pattern recognition capabilities. To optimize the prediction model, the system uses Bayesian optimization (Bayesian Optimization) to search for the best combination of hyperparameters, while incorporating an incremental learning strategy to enable the model to continuously optimize with new data. In practical applications, when the prediction results show that the occurrence probability of a certain type of defect exceeds the pre-set risk threshold, the system will generate the best process adjustment recommendations through reinforcement learning strategies, such as adjusting the inoculant dosage, optimizing the cooling rate, or modifying the magnesium treatment parameters, to reduce the defect occurrence rate and improve the quality of castings.

[0057] In implementation, traditional core-spun yarn addition estimation methods are mostly based on linear regression or static empirical models, which are difficult to adapt to changes in different furnace conditions, resulting in low prediction accuracy. The system uses a deep learning model combining LSTM and CNN to dynamically learn the complex nonlinear relationships in the feeding process, improving prediction accuracy. In addition, the introduction of incremental learning strategy enables the model to continuously adapt to changes in different production batches, avoiding the problem of performance degradation over time. Traditional cast iron defect prediction methods mainly rely on expert experience or simple statistical analysis, while the system predicts the probability of multiple defects through a multi-task learning framework, and uses a Transformer encoder to model long-term process data, making the prediction model more robust and adaptable. In addition, the system combines reinforcement learning to optimize defect avoidance strategies, making maintenance and adjustment more accurate. Compared to traditional methods that rely on manual intervention, it can effectively reduce human error and improve production consistency. In summary, the intelligent operation and maintenance system under the cloud computing architecture greatly improves the prediction accuracy of core-spun yarn addition and the accuracy of nodular cast iron defect prediction, making the feeding process more intelligent and stable, thereby significantly improving production efficiency and product quality.

[0058] The working method of the reinforcement learning algorithm based on Markov decision process is as follows: first, a probabilistic graphical model is used to represent high-dimensional features of the feeding equipment operation data to construct a state transition matrix, and a Bayesian update mechanism is used to dynamically adjust the state transition probability; then, the optimal execution path of different maintenance strategies is calculated through a policy gradient method based on attention mechanism, and the historical strategy execution results are stored through an experience replay mechanism; if the current prediction result exceeds the risk threshold, the influence of different actions on long-term returns is evaluated through a double Q learning mechanism; if the prediction error exceeds the set threshold during strategy updating, the strategy distribution is constrained through a trust region optimization strategy; finally, the dynamic maintenance strategy is optimized in parallel through an A3C architecture based on adaptive entropy regulation, and control instructions are issued to the edge layer.

[0059] In specific implementation, the probabilistic graphical model can mine the potential dependency relationships between various operating variables and store them in a graph structure to support subsequent state transition modeling. Due to the complexity of the feeding equipment operating environment and the presence of various random disturbance factors, a Markov decision process is used to model the state evolution process of the equipment, where the state transition probability is driven by real-time monitoring data and dynamically adjusted through a Bayesian update mechanism, enabling the system to perform autonomous learning and strategy optimization in an uncertain environment.

[0060] In the reinforcement learning strategy optimization phase, the system uses an attention mechanism-based policy gradient method to calculate the optimal execution path of different maintenance strategies. Traditional reinforcement learning methods usually use value iteration or policy iteration methods for decision optimization, while policy gradient-based methods can directly optimize the policy function, making the optimization of maintenance strategies more efficient and stable. In addition, the attention mechanism can help the model identify the most influential key variables in high-dimensional feature space, thereby improving the accuracy of strategy generation. To further improve decision quality, the system introduces an experience replay mechanism, which can store historical maintenance strategy execution results and use batch sampling for training to break the temporal correlation between samples and improve the model's generalization ability. When the prediction result exceeds the set risk threshold, the system triggers a double Q-learning mechanism to evaluate the impact of different maintenance actions on long-term returns. The core idea of double Q-learning is to use two independent Q functions to update alternately to reduce the bias of Q value estimation and improve the stability of decision-making. In addition, during the strategy update process, the system uses a trust region optimization strategy to constrain the policy distribution and prevent the policy update from being too large, which can cause unstable decision-making. This method constrains the KL divergence between the new and old strategies to keep the distribution change within a reasonable range, thereby improving the convergence speed and stability of strategy optimization.

[0061] In the final dynamic maintenance strategy optimization phase, the system employs an A3C (Asynchronous Advantage Actor-Critic) architecture based on adaptive entropy regulation for parallel optimization. A3C is an asynchronous parallel reinforcement learning method that simultaneously learns policy through multiple threads to improve training efficiency and model convergence speed. Compared to traditional single-thread DDPG (Deep Deterministic Policy Gradient) or DQN (Deep QNetwork) methods, A3C can adapt more quickly to complex dynamic environments. During the A3C training process, the system uses an adaptive entropy regulation mechanism (Adaptive Entropy Regularization) to dynamically adjust the policy entropy weight, balancing the relationship between exploration and exploitation. When the system's prediction uncertainty is high, the exploration weight is increased to encourage the system to try new maintenance strategies; when the system's understanding of the current state is sufficient, the exploration weight is reduced to enhance the use of existing optimal strategies. Finally, the optimized dynamic maintenance strategy is issued to the edge layer, and the edge computing nodes perform specific maintenance operations, such as adjusting the feeding speed, optimizing the core yarn addition parameters, or performing device health self-checks, etc.

[0062] Compared with the prior art, the reinforcement learning algorithm based on Markov decision process has significant advantages in intelligent maintenance strategy generation and optimization. Traditional equipment maintenance strategies are mostly based on fixed rule bases or expert experience, which are difficult to adapt to the dynamic changes of complex working conditions, and the update cycle of the maintenance strategy is long, resulting in high lag of equipment maintenance decision. The system uses MDP for state modeling, and combines with the Bayesian update mechanism, so that the system can dynamically adjust the state transition probability, thereby improving the adaptive ability of the maintenance strategy. In addition, the policy gradient method combined with the attention mechanism enables the system to more accurately identify key influencing factors and improve the optimization efficiency of the maintenance strategy. Traditional reinforcement learning methods often have the problem of overestimation of Q value, and the introduction of double Q learning mechanism effectively reduces the strategy bias in the maintenance decision process, making the system's decision more robust. The application of trust region optimization strategy makes the strategy update more stable, avoids the problem of strategy oscillation, and improves the reliability of maintenance optimization. Compared with traditional single-threaded reinforcement learning algorithms, A3C uses asynchronous parallel optimization, which can converge to the optimal strategy faster, improve the online decision-making ability of the system, and the introduction of adaptive entropy regulation mechanism enables the system to flexibly adjust the balance between exploration and utilization, thereby providing efficient maintenance decisions under different working conditions. In addition, the system combines with the experience replay mechanism, so that the maintenance decision can fully utilize the historical data, improve the generalization ability and long-term optimization effect of the model. In summary, the reinforcement learning algorithm based on Markov decision process not only improves the intelligent maintenance capability of the feeding line equipment, but also greatly improves the dynamic optimization efficiency of the maintenance strategy, making the maintenance process more accurate, stable and efficient, thereby effectively reducing the equipment failure rate and improving the stability and production efficiency of the production line.

[0063] The application layer is used to display the equipment operation data, deep data analysis results and dynamic maintenance strategy through a data visualization tool; and generate an operation and maintenance task list according to the dynamic maintenance strategy, and assign the tasks to corresponding maintenance personnel through a workflow management system; the application layer is also used to trigger an alarm mechanism based on a rule engine when the equipment is in an abnormal state or has a potential failure risk, and notify the operation and maintenance personnel through short message, email or system pop-up window.

[0064] In specific implementations, the data visualization module displays the device operation data, deep data analysis results, and dynamic maintenance strategies stored in the cloud to users in the form of charts, dashboards, and interactive interfaces through visualization tools such as Grafana, ECharts, or D3.js. This process involves key technologies such as data cleaning, feature extraction, and visualization mapping. Data cleaning handles outliers, missing values, and duplicate data through the ETL (Extract-Transform-Load) process to ensure data integrity and consistency. Feature extraction relies on aggregation calculations, sliding window statistics, and dimensionality reduction methods to ensure data interpretability and trend analysis capabilities in time series. Visualization mapping uses the Layered Abstraction Model (LAM) to select appropriate chart types based on data types, such as line charts for time series data, heat maps for anomaly detection results, and dashboards for device operation status.

[0065] Secondly, the Workflow Management System (WMS) of the application layer is responsible for the automatic allocation of maintenance tasks. After generating the maintenance strategy, the system will allocate tasks based on the urgency, complexity of the task, and the skill tags of the maintenance personnel, using task scheduling algorithms such as priority scheduling, intelligent scheduling based on reinforcement learning, etc. Specifically, the priority of the task is determined by the severity of the device anomaly, the potential impact range, and the predicted failure time, and a weighted scoring mechanism is used to calculate the priority score. Task matching is based on genetic algorithm (GA) optimization to allocate the scheme, so that the load of the task in the maintenance team is balanced, and the processing efficiency is maximized. For example, for urgent tasks, the system can use preemptive scheduling strategies to reallocate resources from low-priority tasks to high-priority tasks to ensure that the maintenance of critical devices is given priority. At the same time, the WMS also integrates mobile applications, allowing maintenance personnel to accept tasks at any time and communicate in real-time through remote collaboration tools such as WebRTC or MQTT protocol, improving the efficiency of collaborative work.

[0066] In addition, the system triggers an alarm according to a predefined rule set when it detects an abnormal device state or potential failure risk. The rule engine uses forward chaining, which triggers rules based on real-time data streams and performs corresponding notification actions. The maintenance of the rule base uses Knowledge Graph technology, which continuously optimizes and updates the rule set based on historical failure data, expert experience, and data mining results, improving the accuracy of alarms. The triggering conditions of the alarm can be based on set thresholds, pattern recognition, or anomaly detection algorithms, such as using Isolation Forest or Local Outlier Factor (LOF) models to detect abnormal data points, and combining Markov Chain to predict failure probability. When the alarm is triggered, the notification mechanism uses a multi-channel delivery strategy to push the alarm information to relevant operation and maintenance personnel, supporting SMS, email, system pop-up, and other methods, and combining natural language processing (NLP) technology to intelligently summarize the alarm information to improve the efficiency of operation and maintenance personnel in understanding the problem.

[0067] In the implementation process, the application layer first interacts with the cloud database through the data interface, regularly pulls the device running state, prediction analysis results, and maintenance suggestions, and uses a stream data processing framework (such as Apache Flink or KafkaStreams) to perform real-time processing and analysis of data. Then, based on RESTful API or GraphQL interface, the data is sent to the front-end visualization system, and the page is dynamically rendered to provide interactive data analysis functions. At the same time, the WMS listens to the event stream of the maintenance strategy generation module, and once a new maintenance task is generated, it triggers the task scheduling process, assigns tasks according to the set rules or reinforcement learning model, and notifies relevant personnel through the mobile application or Web. For the abnormal alarm process, the system continuously monitors device sensor data and uses CEP (Complex Event Processing) technology to perform pattern matching on data streams. Once an abnormal pattern is detected, the rule engine triggers the alarm process, and finally pushes the alarm information to the designated recipient.

[0068] While specific embodiments of the application have been described above, it will be appreciated that those skilled in the art within the scope of the application can make modifications, substitutions and changes in the form and details of the method and system described above without departing from the spirit and scope of the application. For example, it is well within the scope and spirit of the application to combine various of the illustrative steps of the above methods to carry out essentially the same function to achieve essentially the same result in substantially the same way. Accordingly, the scope of the application should be determined sole by the appended claims.

Claims

1. A cloud-based intelligent operation and maintenance system for wire feeding; characterized in that: The system comprises: an edge layer, a transport layer, a scheduling layer, a cloud layer, and an application layer; The edge layer is used to collect and upload the operating data of the wire feeding equipment, and to perform emergency response processing and control of the wire feeding equipment, including a collection module, a processing module and a control module; The transport layer is used to determine and optimize data interaction between the edge and the cloud based on parameters such as link network latency, bandwidth utilization, and packet loss rate, using an adaptive transmission optimization algorithm. The scheduling layer is used to dynamically adjust the allocation of computing tasks based on the complexity and urgency of tasks, using a load balancing scheduling algorithm based on deep reinforcement learning, based on edge node and cloud server resource monitoring data and task urgency indicators. The cloud layer is used by the edge layer to construct fault prediction models, cored wire addition prediction models, and ductile iron defect prediction models based on historical operation and maintenance data, real-time monitoring data, and environmental variable data, using long short-term memory networks and convolutional neural networks. It determines whether the equipment operating status, cored wire addition amount, and ductile iron defect prediction results exceed preset risk thresholds. If they exceed the thresholds, an adaptive dynamic maintenance strategy is generated by combining deep data analysis results and the real-time status of the equipment with a reinforcement learning algorithm based on Markov decision process, and control commands are output to the edge layer. If they do not exceed the thresholds, a warning report and maintenance suggestions are output after fusing the current prediction results with historical data through statistical analysis and archiving methods. The application layer is used to display equipment operation data, in-depth data analysis results, and dynamic maintenance strategies through data visualization tools; and to generate a list of operation and maintenance tasks based on the dynamic maintenance strategies, and to assign the tasks to the corresponding maintenance personnel through the workflow management system.

2. The intelligent operation and maintenance system for cable feeding based on cloud computing according to claim 1, characterized in that: The acquisition module is used to acquire and preprocess the operating data of the wire feeding equipment in real time, including motor current, voltage, speed, wire feeding speed, and equipment temperature parameters, and send the preprocessed operating data of the wire feeding equipment to the processing module. The processing module is used to perform data cleaning, time-series alignment, and feature extraction on the operating data of the wire feeding equipment using a real-time stream processing method. It also uses a fault detection model to determine whether the data is abnormal. If so, it transmits emergency processing instructions to the control module through a local emergency processing mechanism, and sends the local diagnostic results to the cloud layer through the transport layer. If not, it sends the pre-processed data to the cloud layer through the transport layer. The control module is used to receive emergency processing instructions sent by the processing module and control instructions sent by the cloud layer, and controls the actuators of the wire feeding equipment, including a motor driver and a speed regulator, through a programmable logic controller.

3. The intelligent operation and maintenance system for cable feeding based on cloud computing according to claim 1, characterized in that: The processing module operates as follows: First, signal preprocessing and noise suppression are performed using Kalman filtering and wavelet denoising. Dynamic time warping is used to correct timestamp errors from various data sources. Then, a low-dimensional feature vector is output to the fault detection model via an autoencoder for dimensionality reduction. The fault detection model uses Euclidean and Mahalanobis distance metrics to calculate the deviation between the current feature vector and historical normal state features in real time, and performs online weight updates using a fast gradient descent algorithm. If the deviation exceeds a preset threshold, an emergency processing command is generated through an embedded rule engine, sent to the control module in JSON format via UDP / TCP protocol, and the local diagnostic results are uploaded to the cloud layer via MQTT protocol. If the deviation result is lower than the preset threshold, the normalized feature data will be transmitted to the cloud data warehouse via HTTP / 2 protocol.

4. The intelligent operation and maintenance system for cable feeding based on cloud computing according to claim 1, characterized in that: The working steps of the adaptive transmission optimization algorithm include: S1, Link delay based on real-time data acquisition Bandwidth utilization Packet loss rate And network jitter The data was smoothed using an exponentially weighted moving average method to obtain smoothed values. , , and ; S2. Construct the transmission cost function The transmission cost function is used to measure link quality. The formula expression is: （1） In formula (1), This is the normalized value for bandwidth utilization. , , , The adjustment coefficients represent the effects of delay, jitter, packet loss, and bandwidth utilization on transmission performance, respectively. S3, Based on transmission cost function By fitting the network state change curve using the least squares method and discretizing the network state using a fuzzy logic control mechanism, the optimal data transmission rate is generated. With group size If the transmission cost function If the threshold is exceeded, a reinforcement learning scheduling algorithm is used, based on the state vector ( , , ) and action sets Calculate the state action function and update the transmission strategy in real time; if the transmission cost function If the value is below the threshold, the current parameter configuration is maintained, and the parameter configuration strategy is issued through the feedback loop.

5. The intelligent operation and maintenance system for cable feeding based on cloud computing according to claim 1, characterized in that: The working method of the load balancing scheduling algorithm based on deep reinforcement learning is as follows: First, the resource monitoring module collects resource status data of edge nodes and cloud servers in real time, including CPU utilization. Memory usage Network bandwidth and task queue length The algorithm extracts features from resource state data using a temporal convolutional network to generate a resource state vector X. Simultaneously, the deep reinforcement learning-based load balancing scheduling algorithm evaluates task complexity using a task classifier. and urgency Generate task feature vectors Based on resource state vectors and task feature vectors, the execution value of each task on edge nodes or cloud servers is calculated using the Q-value function. Where A represents the task assignment action, and the update formula for the Q-value function is: （2） In formula (2), For instant rewards, For learning rate, As a discount factor; if the urgency of the task If the task exceeds a preset threshold, a high-priority scheduling mechanism is triggered, prioritizing the allocation of the task to edge nodes with available resources; if the task complexity... If the threshold is exceeded, the cloud collaborative computing mode will be activated, and the task will be split and distributed to the cloud server in parallel.

6. The intelligent operation and maintenance system for cable feeding based on cloud computing according to claim 1, characterized in that: The fault prediction model decomposes historical operating data and feeder equipment operating data into multiple scales through a time-series feature extraction mechanism, constructing a separate representation of short-term trend terms and long-term evolution terms. During data preprocessing, the model utilizes a self-supervised learning framework based on anomaly detection to construct a probabilistic graphical model using equipment sensor data, calculates the dynamic correlation between various operating parameters, and assigns feature weights through a graph attention mechanism. Based on the processed feature data, the model employs an incremental learning strategy to construct a fault prediction network, dynamically adjusting the parameters of the long short-term memory network and convolutional neural network through an adaptive gradient optimization algorithm. During prediction calculation, the model calculates short-term trend shifts using an autoregressive moving average model and extracts long-term evolution patterns using an attention-guided time-series difference method.

7. The intelligent operation and maintenance system for cable feeding based on cloud computing according to claim 1, characterized in that: The working method of the cored wire addition prediction model is as follows: First, a variational autoencoder is used to extract potential key variables in the smelting process through a self-supervised learning temporal representation learning method, and the data is reduced in dimensionality using principal component analysis. Based on the dimensionality-reduced data, the cored wire addition prediction model uses a multi-scale convolutional neural network to extract temporal features, capture the local short-cycle fluctuation characteristics of cored wire consumption, and calculates the influence weights of key smelting parameters on the amount of cored wire added by combining an attention mechanism. In the prediction calculation stage, the cored wire addition prediction model uses a long short-term memory network based on Bayesian optimization to construct a nonlinear time series prediction model. If the prediction error exceeds a set threshold, an incremental learning error compensation mechanism is used to calculate the prediction residual through an autoregressive integral moving average model, and a recurrent neural network is used to regress and optimize the historical deviation. If the prediction result exceeds a preset risk threshold, an adjustment strategy is generated and an optimization control command is output to the edge layer.

8. The intelligent operation and maintenance system for cable feeding based on cloud computing according to claim 1, characterized in that: The defect prediction model for ductile iron is based on historical smelting process parameters, real-time production data and environmental variable data. First, a high-dimensional feature space is constructed through a heterogeneous data fusion mechanism. exist In the feature construction process, a graph attention network-based smelting process parameter association modeling method is used to extract the nonlinear coupling relationship between variables, and a nested variational autoencoder is used to reduce the feature dimensionality of the data. In the defect prediction stage, the ductile iron defect prediction model uses a spatiotemporal fusion convolutional neural network to extract high-dimensional features from the temperature field, stress field, and composition fluctuation data during the casting process. This allows for the acquisition of the contribution of different smelting stages to defect formation. A multi-scale residual network is also used to capture the short-term dynamic features of defect generation, improving the timeliness of the prediction model. During the prediction inference process, the ductile iron defect prediction model employs a Bayesian-optimized two-layer long short-term memory network to balance long-term defect trend prediction with short-term defect fluctuations. If the prediction result exceeds a set risk threshold, a defect avoidance strategy based on a Markov decision process is triggered. This strategy optimizes the amount of spheroidizing agent added and the cooling rate using reinforcement learning, and outputs dynamic adjustment instructions to the edge layer. If the prediction error exceeds the threshold, the model parameters are adjusted in real time based on an adaptive incremental learning mechanism.

9. The intelligent operation and maintenance system for cable feeding based on cloud computing according to claim 1, characterized in that: The working method of the reinforcement learning algorithm based on Markov decision process is as follows: First, a probabilistic graphical model is used to represent the operation data of the feeding equipment in high dimension to construct a state transition matrix, and the state transition probability is dynamically adjusted by combining a Bayesian update mechanism; then, the optimal execution path of different maintenance strategies is calculated by a policy gradient method based on an attention mechanism, and the historical strategy execution results are stored by combining an experience replay mechanism; if the current prediction result exceeds the risk threshold, the impact of different actions on long-term returns is evaluated by a double Q learning mechanism. If the prediction error exceeds the set threshold during the policy update process, the policy distribution is constrained by optimizing the policy through the trust domain. Finally, the dynamic maintenance policy is optimized in parallel through the A3C architecture based on adaptive entropy control, and control commands are sent to the edge layer.

10. The intelligent operation and maintenance system for cable feeding based on cloud computing according to claim 1, characterized in that: The application layer is also used to trigger an alarm mechanism based on the rule engine when the device is in an abnormal state or has potential failure risks, and to notify the operation and maintenance personnel via SMS, email or system pop-up.

Images