A management method and system for a by-wire mes system

Abstract

The application discloses a kind of management method and system for drive-by-wire MES system, it is related to MES system technical field, the system is composed of several function modules, including: data acquisition module, real-time acquisition production line operation data, operation data include sensor data, production line data and process parameter data;Data processing module, by deploying lightweight algorithm in edge node, processing operation data, upload the operation data after processing to cloud for depth analysis, depth analysis includes trend analysis, bottleneck identification and prediction modeling;In combination with depth analysis result, real-time mapping physical production environment, simulate different scheduling strategy response;Task scheduling and optimization module, based on production line data, dynamically adjust task priority using reinforcement learning algorithm, while introducing buffer capacity adjustment mechanism to cope with load fluctuation;Receive the process limiting link information predicted by MES system, in combination with distributed path planning strategy, dynamically adjust transportation path.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of MES systems, in particular to a management method and system for a drive-by-wire MES system. BACKGROUND

[0002] With the development of technology and changes in market demand, MES systems have been continuously developed and improved, and have gradually been applied to various types of manufacturing industries, helping enterprises to improve production efficiency and flexibility while ensuring product quality. Through real-time data acquisition, production scheduling, quality control and equipment management, the transparency and intelligent management of the production process are realized; located between the enterprise resource planning system and the workshop control layer, it acts as a data bridge and focuses on the dynamic optimization of the execution layer; through sensors, PLCs, RFID and other technologies, data such as equipment status, process parameters and material flow are collected, and production plans are dynamically adjusted to respond to urgent orders or equipment failures; through a computerized management system oriented to the production process of the workshop, it is located between the upper-layer enterprise resource planning system and the bottom-layer production equipment control, and plays a role in connecting the upper and lower layers; the drive-by-wire MES system mainly focuses on the management and optimization of the production site to improve production efficiency, product quality and resource utilization; it can collect various data on the production line in real time, including equipment status, process parameters and material usage, and realize real-time monitoring of the production process; according to order requirements and actual production capacity, reasonable production scheduling and production scheduling are carried out, the production process is optimized, and waiting time is reduced.

[0003] In traditional drive-by-wire MES systems, the management method usually relies on preset fixed rules for task scheduling and decision making, which has obvious limitations in real-time data driving. At the same time, due to the delay in data acquisition and processing, the real-time performance of the system is insufficient, which leads to a lag in updating the production status, affecting the timeliness and accuracy of the decision, and the data accuracy problem is also prominent, especially in the aspects of equipment status monitoring and quality control, the traditional method cannot avoid noise interference and data loss, further weakening the reliability of the system; taking automobile manufacturing as an example, when a device failure or bottleneck occurs at a workstation on a production line, the existing MES system is difficult to redistribute production tasks to other available workstations through dynamic scheduling algorithms. SUMMARY

[0004] To achieve the above purpose, the application is implemented through the following technical solutions:

[0005] A management system for a drive-by-wire MES system, comprising:

[0006] a data acquisition module for real-time acquisition of production line operation data, the operation data including sensor data, production line data and process parameter data;

[0007] A data processing module processes the running data by deploying lightweight algorithms on the edge node, uploads the processed running data to the cloud for deep analysis, which includes trend analysis, bottleneck identification and prediction modeling; combined with the deep analysis results, the physical production environment is mapped in real time, and different scheduling strategies are simulated to respond;

[0008] A task scheduling and optimization module dynamically adjusts task priority based on production line data using reinforcement learning algorithms, and introduces a buffer capacity adjustment mechanism to cope with load fluctuations; receives process limiting link information predicted by the MES system, and dynamically adjusts the transportation path combined with the distributed path planning strategy;

[0009] A prediction and control module analyzes device historical data based on LSTM neural network, predicts potential failures and triggers preventive maintenance work orders, and monitors production quality in real time.

[0010] Further, the process of collecting production line running data in real time is:

[0011] By deploying near the production line, collecting sensor data; sensor data includes temperature sensor, pressure sensor and displacement sensor;

[0012] Production line data includes real-time material location, order urgency and material inventory;

[0013] Using standardized protocols, data collected by sensors is transmitted to the MES system for real-time monitoring and adjustment, and saved to process parameter data.

[0014] Further, the process of processing running data is:

[0015] Deploy low-complexity models on the edge node to clean and extract features from running data in real time; use distributed computing to process multiple data streams in parallel, combine LSTM neural network to predict device failure, and dynamically adjust task priority using reinforcement learning algorithm; use event-driven architecture to trigger immediate response; dynamically allocate resources through containerized deployment on the edge node, and upload key data to the cloud for deep analysis based on TSN time-sensitive network;

[0016] Clean the running data, and unify the running data from different sensors into a standard format, and structure the unstructured running data.

[0017] Further, the process of deep analysis is:

[0018] Trend analysis: Adopting moving average and wavelet transform to eliminate sensor noise, applying moving average with large window to the de-noised data to extract long-term trend, judging the trend stationarity through unit root test; identifying periodic components by combining autocorrelation function and partial autocorrelation function, using ARIMA model to predict periodic fluctuation law of the stationary sequence after difference, superimposing trend and periodic terms to generate complete trend curve, and updating model parameters based on real-time data of MES;

[0019] Bottleneck identification: Monitoring device OEE through Prometheus, combining Apriori algorithm to mine association rules between device failure and process parameters; calculating device comprehensive utilization rate based on sensor data and MES system indicators, and determining as a bottleneck if the device idle rate exceeds the threshold; otherwise, it does not reach the bottleneck; using RFID to monitor buffer inventory, and locating the accumulation source combined with the topology of the production line;

[0020] Prediction modeling: Extracting the characteristic frequency of device failure, using random forest to predict device failure; modeling device aging trend based on long-term dependence relationship of time series, quantifying trend direction and duration; embedding the failure prediction model and bottleneck rule into the MES system.

[0021] Further, the process of real-time mapping the physical production environment is:

[0022] Deploying a digital twin platform in the cloud, creating corresponding virtual entities, using industrial simulation software to build a runnable production line model, defining event-driven mechanisms, setting state transition logic, and configuring task scheduling strategies; collecting and processing running data through edge computing nodes, and uploading them to the cloud in real time. When new data is uploaded, the simulation engine updates the model state in real time.

[0023] Further, the process of dynamically adjusting task priority using reinforcement learning algorithm is:

[0024] S301: Obtain historical running data, use Plant Simulation to build a digital twin model, and simulate the dynamic behavior of the production line;

[0025] S302: Balance production efficiency, cost, and quality through discrete action space and continuous action space algorithms combined with entropy weight multi-objective decision method;

[0026] S303: Initialize the model, randomly generate a scheduling strategy, update the Q value according to the reward function, use experience replay to store historical state-action-reward data, and use target network to stabilize the training process; stop training when the model reaches the pre-set performance indicators in the simulation environment; verify the robustness of the model in extreme scenarios.

[0027] Further, the specific process of introducing a buffer capacity adjustment mechanism is:

[0028] A physical buffer is arranged before and after each key station, and the release rhythm of the upstream task is dynamically adjusted according to the current buffer queue length; if the buffer is close to full, the priority of the task before the process is reduced; otherwise, the release task is accelerated.

[0029] Further, the process of dynamically adjusting the transportation path is:

[0030] The global task is decomposed into multiple subtasks, and the subtasks are distributed to the AGV optimal route through a distributed algorithm; the AGV competes for the subtask according to its own state, and preferentially selects the optimal path.

[0031] Further, the process of predicting potential failures and triggering preventive maintenance work orders is:

[0032] The edge computing node receives sensor data in real time, inputs the LSTM model in the form of a sliding window, the sliding window length generates a prediction sequence, and outputs the fault category; the fault probability threshold is set according to historical data, and the threshold is dynamically adjusted; the prediction result is pushed to the MES system through an API interface, and a maintenance work order is automatically generated; through the Bi-LSTM+Attention mechanism, the long-term dependence in the production process is captured, the key abnormal points are identified by attention weight, and the model is regularly updated.

[0033] A management method for a line control MES system, comprising the following steps:

[0034] Step one, real-time acquisition of production line running data, the running data including sensor data, production line data and process parameter data;

[0035] Step two, processing the running data by deploying a lightweight algorithm on the edge node, uploading the processed running data to the cloud for deep analysis, the deep analysis including trend analysis, bottleneck identification and prediction modeling; combining the deep analysis results, real-time mapping of the physical production environment, simulation of different scheduling strategies;

[0036] Step three, based on the production line data, dynamically adjusting the task priority by using a reinforcement learning algorithm, and introducing a buffer capacity adjustment mechanism to cope with load fluctuations; receiving the process limiting link information predicted by the MES system, combining a distributed path planning strategy, and dynamically adjusting the transportation path;

[0037] Step four, predicting potential failures and triggering preventive maintenance work orders based on LSTM neural network analysis of device historical data, and real-time monitoring of production quality.

[0038] The management method and system for the line control MES system provided by the application have the following beneficial effects:

[0039] (1) The present application can perform real-time cleaning and feature extraction on production line operation data by deploying lightweight algorithms at the edge node, and utilize distributed computing to process multiple source data streams in parallel. This architecture not only reduces the time delay of data transmission to the cloud, but also ensures that critical data can be uploaded to the cloud for deep analysis with minimal delay, thereby improving the response speed and real-time performance of the overall system.

[0040] (2) The present application uses a reinforcement learning algorithm to dynamically adjust task priorities, and combines a buffer capacity adjustment mechanism to cope with load fluctuations. The MES system can flexibly adjust the task allocation strategy according to real-time production conditions; and based on historical data analysis of the LSTM neural network, the system can predict potential equipment failures and trigger preventive maintenance work orders.

[0041] (3) The present application uses the Bi-LSTM+Attention mechanism, the system can capture long-term dependencies in the production process, identify key abnormal points, and model and analyze quality detection data; through the cloud-deployed digital twin platform, a virtual entity of the physical production environment is created, the model state is updated in real time, and the response of different scheduling strategies is simulated; this allows managers to test the effects of various production plans in a virtual environment without actually interfering with the production process, providing support for making better production plans, and further improving the accuracy and efficiency of decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 is a system flowchart of the present application;

[0043] Figure 2 is a schematic diagram of the overall method of the present application. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application. Embodiment 1

[0045] Please refer to Figure 1 Embodiment 1 of the present application provides a management system for a drive-by-wire MES system, which comprises:

[0046] a data acquisition module that acquires real-time production line operation data, including sensor data, production line data and process parameter data;

[0047] Sensor data:

[0048] Sensor data includes temperature sensors, pressure sensors, and displacement sensors; deployed near the production line, responsible for collecting sensor data, making quick decisions, such as adjusting the gripping force to avoid damaging materials;

[0049] Temperature sensor: determines the surface temperature of an object by measuring the amount of infrared radiation emitted by the object, installed near the welding equipment, without direct contact with the measured object, monitors temperature changes during the welding process;

[0050] Pressure sensor: detects changes in the distance between two plates through a capacitive sensor, measures pressure by detecting changes in capacitance caused by pressure, has high sensitivity and good linearity; integrated into the robot's gripping device, used to measure gripping force to ensure safe handling;

[0051] Displacement sensor: calculates the distance between the target object and the sensor by emitting a laser beam and receiving the reflected light signal, provides high precision measurement, measures without contact, installed at the robot joints to monitor the position and movement distance of the mechanical arm, ensures the accuracy of the mechanical arm operation;

[0052] Production line data:

[0053] Production line data includes real-time material location, order urgency, and material inventory; obtained through visual cameras, RFID devices, and PLC technology;

[0054] Real-time material location: industrial vision cameras installed at key nodes on the production line or on robots capture material image information required by the production line, analyze image content using computer vision algorithms to determine the location coordinates of the material, for example, in an automobile parts assembly line, vision cameras can be used to identify the specific location of parts on the conveyor belt to guide the mechanical arm to accurately grasp; for moving materials, their positions can be updated in real time by continuous shooting and combining with motion tracking algorithms;

[0055] Order urgency: extracts order information from the ERP system, including delivery date, customer priority, etc., as a standard for judging order urgency; based on the extracted information, a series of rules are set, and the MES system automatically calculates the urgency of each order based on these rules, and adjusts the production plan and task allocation strategy accordingly;

[0056] Material Inventory: Through RFID devices and PLC, attach RFID tags on materials or pallets, store information about material type, quantity, batch, etc.; deploy RFID readers at warehouse entrances, exits, and key positions on the production line. When RFID-tagged items pass through, the readers can automatically identify and record their entry and exit, updating inventory status in real time; meanwhile, connect PLC to various sensors on the production line to monitor changes in material consumption, combined with RFID system data to obtain inventory levels;

[0057] Process Parameter Data:

[0058] Using standardized protocols, transmit sensor-collected data to the MES system, monitor and adjust these process parameters in real time, effectively improving the automation and intelligence level of the production line, ensuring efficient and stable production processes;

[0059] For example:

[0060] On a car body assembly production line in an automobile factory, multiple robots are responsible for transporting different types of parts to designated workstations for assembly. The MES system performs the following operations based on data obtained from various process parameter sensors:

[0061] When a car door needs to be transported to a welding workstation, a vision camera first confirms its position on the conveyor belt; the MES system accordingly guides the robot arm to the correct position for grasping; during grasping, pressure sensors continuously monitor the force applied to the car door, ensuring that it will not be damaged by excessive pressure; at the same time, the MES system decides whether to prioritize this transportation task based on the urgency of other tasks on the production line; if the temperature sensor in the welding area detects an abnormal temperature rise, the MES system will automatically reduce the welding speed or temporarily stop the welding operation to prevent equipment damage; throughout the process, all relevant data is recorded by the MES system for subsequent analysis and quality traceability.

[0062] Data Processing Module: Process running data by deploying lightweight algorithms on edge nodes, upload processed running data to the cloud for deep analysis, including trend analysis, bottleneck identification, and predictive modeling; combine deep analysis results to map the physical production environment in real time, simulate different scheduling strategies in response;

[0063] Processing running data:

[0064] Optimize low-complexity models on edge nodes through lightweight algorithms, perform real-time cleaning and feature extraction on raw sensor data; use distributed computing to process multiple data streams in parallel, combine LSTM neural networks to predict device failures, and dynamically adjust task priorities through reinforcement learning algorithms; use event-driven architecture to trigger immediate responses, such as triggering emergency shutdown when pressure sensors exceed limits, ensuring low-latency decision-making; dynamically allocate resources through containerized deployment on edge nodes, and upload critical data to the cloud based on TSN time-sensitive networks for deep analysis;

[0065] Low-complexity model: For resource-constrained edge devices, use low-complexity algorithms to adjust task priorities, dynamically allocate robot task priorities, for example, prioritize material handling for urgent orders; use Voronoi diagram algorithm to avoid obstacles in real time, and combine environmental data provided by MES system, such as obstacle location, to generate optimal path;

[0066] Cleaning and feature extraction: Use edge node rule engine to filter redundant data, such as uploading only alarm data with temperature exceeding threshold, rather than all raw data; aggregate high-frequency data in time window, merge vibration data every second into statistical value every minute;

[0067] Use sliding window averaging method, low-pass filter, etc. to smooth signals, use wavelet transform or Fourier transform for frequency domain filtering on high-frequency noise; detect if there are missing data packets or too large sampling intervals, if there are missing values, use interpolation method to fill in; identify outliers based on statistical methods, use machine learning models to identify sudden abnormal behavior, and remove or replace outliers with reasonable estimated values; unify data from different sensors into standard format, and structure unstructured data;

[0068] After completing the running data cleaning, further extract representative and operable features from the original data for subsequent modeling and analysis; perform time-domain feature extraction, use fast Fourier transform to convert signals to frequency domain, extract main frequency amplitude, frequency band energy distribution, harmonic components, etc.; use short-time Fourier transform, wavelet transform, etc.; analyze the frequency components of signals over time, suitable for non-stationary signals; and determine the device operating state based on sensor data, such as start, run, stop, failure, etc.; combine logical rules or classification models to identify specific events;

[0069] Upload critical data to the cloud: use lightweight encoding format to compress data volume, reduce transmission bandwidth requirements; upload processed data to the cloud through 5G / TSN network, support OPC UA / MQTT protocol, ensure seamless integration with MES system;

[0070] Deep analysis:

[0071] Trend analysis: Identify long-term trends using moving averages or wavelet transforms to remove sensor noise; use ARIMA models to analyze periodic fluctuations; apply large-window moving averages to the denoised data to filter out short-term fluctuations and extract long-term trend curves; use unit root tests to determine whether the trend term is stationary; identify periodic components through autocorrelation and partial autocorrelation functions, select the optimal parameter combination based on AIC / BIC criteria, and use ARIMA models to predict the stationary sequence after differencing, then back-propagate the periodic fluctuations of the original data; superimpose the long-term trend term and the periodic fluctuation term to generate a complete trend analysis result, periodically update the ARIMA model parameters based on real-time data feedback from the MES system;

[0072] Bottleneck identification: Locate performance bottlenecks in the production line, such as low equipment utilization and material accumulation; use the Apriori algorithm to find associations between device failures and specific process parameters based on operational data; monitor device OEE in real-time through Prometheus + Grafana to identify inefficient links; collect device operating status, material flow speed, and other data through sensors, and combine with MES system work order completion rate indicators to calculate device comprehensive utilization, for example, if the idle rate of a certain station device is consistently higher than 30%, it is determined as a potential bottleneck; use RFID or vision sensors to monitor buffer inventory levels, and trigger an alert when inventory exceeds a pre-set threshold, combined with the production line topology map to locate the source of accumulation;

[0073] Identify inefficient links through statistical analysis, for example, if the throughput of a certain process is significantly lower than upstream and downstream, and its CPU / memory utilization rate is consistently above 85%, it is determined as a performance bottleneck; classify fault types based on association rules, and verify rule effectiveness combined with device maintenance logs; embed association rules into the MES system to monitor parameter combination anomalies in real-time and trigger adaptive adjustments, such as automatically reducing device operating speed when "temperature > threshold" is detected to avoid overload failures;

[0074] Prediction modeling: Extract feature frequencies of device failures through fast Fourier transform; use random forests or XGBoost to predict device failures; extract trend components from noise to quantify the direction, speed, and duration of trends; capture long-term dependencies in time series through complex nonlinear trends, extract local features, predict device aging time by analyzing long-term trends in vibration and temperature data, identify bottleneck links, and improve overall efficiency;

[0075] Real-time mapping of physical production environment:

[0076] Obtain process path information from the MES system, use process number, name, sequence, equipment resources required for each process, standard working hours, material list, and quality control points, key parameter requirements;

[0077] In the cloud-deployed digital twin platform, a corresponding virtual entity is created for each device, each transportation chain, and each process node. Using industrial simulation software, a runnable production line model is constructed, an event-driven mechanism is defined, state transition logic is set, and task scheduling strategies are configured. Through edge computing nodes, collected and processed operation data is uploaded to the cloud in real time to update the state of the digital twin model and unify multi-source heterogeneous data into a structured format. Whenever new data is uploaded, the simulation engine updates the model state, updates device status, updates material location and flow status, refreshes task queues and scheduling priorities, modifies process parameters, and triggers abnormal events based on the following logic;

[0078] The task scheduling and optimization module dynamically adjusts task priorities based on production line data using reinforcement learning algorithms, while introducing a buffer capacity adjustment mechanism to cope with load fluctuations. It receives process restriction information predicted by the MES system and dynamically adjusts transportation paths using a distributed path planning strategy.

[0079] By inputting production line data, the output is the priority ranking of tasks, AGV transportation path planning, and resource allocation suggestions between processes.

[0080] The reinforcement learning algorithm is used to dynamically adjust task priorities.

[0081] The task scheduling problem is abstracted as a reinforcement learning environment.

[0082] Element Description Status Status of all current tasks (waiting, in process, completed), equipment occupancy, buffer queue length, AGV distribution, etc. Action Adjust the priority of a certain task (raise / lower), assign an AGV to perform a certain transportation task Reward Positive feedback such as time reduction for completing tasks, bottleneck relief, OEE improvement, and waiting time reduction; otherwise, negative reward Strategy Reinforcement learning model determines the next action based on the current state

[0083] The reinforcement learning model relies on a simulated environment, reward function design, and policy optimization to make optimal scheduling decisions in complex production environments.

[0084] Training process:

[0085] S301: Obtain historical operation data such as device failure records, task completion times, process paths provided by the MES system, bottleneck predictions, and real-time sensor data uploaded from the edge. Use Plant Simulation to construct a digital twin model to simulate production line dynamic behavior.

[0086] S302: Balance production efficiency, cost, quality, and other objectives using discrete action space and continuous action space algorithms combined with entropy weight multi-objective decision-making methods.

[0087] S303: Initialize the model, randomly generate the scheduling strategy, such as randomly assigning task priorities; execute actions in the Agen simulation environment, observe environmental feedback, update Q values or policy network parameters according to the reward function, use experience replay, store historical state-action-reward data, avoid overfitting, use target networks to stabilize the training process; stop training when the model reaches the preset performance indicators in the simulation environment; verify the robustness of the model in extreme scenarios;

[0088] Preset performance indicators: based on historical data analysis, industry standards, and enterprise's own goals; by collecting and analyzing existing data, setting specific improvement goals based on business needs, selecting appropriate KPIs as measurement standards, using Plant Simulation or other tools to simulate the performance of the model in different scenarios, and adjusting until the preset performance indicators are met;

[0089] Save as a deployable AI service, extract the optimal strategy of the model in a specific scenario, such as "when the welding station buffer is >80%, reduce the priority of upstream tasks";

[0090] Introduce buffer capacity adjustment mechanism:

[0091] Set virtual / physical buffers before and after each key workstation, dynamically adjust the release rhythm of upstream tasks according to the current buffer queue length; if the buffer is close to full, reduce the priority of tasks before this process, or guide AGV to other paths, the code formula is:

[0092] ```python

[0093] if buffer_level > threshold_high:

[0094] reduce_priority_of_upstream_tasks()

[0095] elif buffer_level < threshold_low:

[0096] increase_priority_of_upstream_tasks()

[0097] ```

[0098] Key workstation: the slowest link in the production process, its working speed determines the output capacity of the entire production line; for example, in automobile manufacturing, the robot at the welding station is slower than other workstations due to equipment aging, becoming a key workstation;

[0099] Dynamic adjustment of transportation path:

[0100] The global task is divided into multiple sub-tasks, such as "transporting materials from the warehouse to the welding station", "AGV1 starts from the warehouse → AGV1 passes through the buffer zone → AGV1 arrives at the welding station"; the sub-tasks are distributed to the most suitable AGV or robot through a distributed algorithm, AGV competes for sub-tasks according to its own state, such as power, load capacity, and selects the optimal path preferentially, AGV reaches consensus through negotiation to avoid resource contention;

[0101] When the production line has a sudden situation, such as equipment failure, task priority change, AGV needs to adjust the path in real time:

[0102] If the welding station is marked as a bottleneck by MES due to full buffer zone, AGV can temporarily change course to the spraying station to temporarily store materials;

[0103] If a path is blocked by an obstacle, AGV needs to find an alternative route through local re-planning;

[0104] AGV adjusts the trajectory by exchanging position and gradient information to maximize group efficiency, such as distance from target, path cost; AGV predicts the future behavior of other devices to avoid potential conflicts in advance, such as predicting potential collisions when an AGV is about to turn; balance efficiency and flexibility by combining centralized global planning and distributed local obstacle avoidance;

[0105] For example:

[0106] In the automobile manufacturing production line, the aging of equipment leads to a decrease in the processing capacity of the welding station, and the task backlog is serious; the buffer zone capacity is close to the upper limit, and AGV frequently waits, causing transportation path congestion; traditional fixed path planning cannot adapt to dynamic load changes, AGV forms "traffic congestion" around the welding station, and the welding station bottleneck causes the waiting time of downstream processes such as spraying and assembly to increase, and the overall production rhythm is delayed; RRT algorithm can quickly generate a feasible path that avoids obstacles by randomly sampling the environment, especially suitable for complex and dynamic production line environments; when the welding station buffer capacity exceeds the threshold, AGV dynamically plans a detour path through the RRT algorithm to avoid the congestion area of the welding station, for example:

[0107] Path 1: AGV starts from the warehouse → detours to the spraying station to temporarily store materials → returns after the welding station load decreases;

[0108] Path 2: AGV directly drives to other non-bottleneck stations such as stamping station to complete the task, reducing the dependence on welding station; MES system predicts the load change of welding station in the next 30 minutes based on historical data and real-time sensor input, when MES predicts that welding station will run overload, the system actively issues "change direction instruction" to AGV, guiding it to preferentially go to spraying station or stamping station; verify the feasibility of path adjustment through digital twin model, ensure that the new path will not cause congestion of other stations, use market mechanism algorithm such as auction to allocate dynamic task priority, avoid multiple AGVs competing for welding station resources at the same time, AGV real-time perception environment through laser SLAM and vision sensor, combined with RRT algorithm for local path re-planning, avoid sudden obstacles; through dynamic path adjustment and load balancing, welding station task backlog is reduced by 40%, overall production line OEE, equipment comprehensive efficiency is improved from 82% to 95%, production rhythm is increased from original 50 JPH (piece / hour) to 57 JPH, meeting the production capacity demand of M254 model; AGV waiting time around welding station is shortened from an average of 8 minutes to 5 minutes, congestion rate is reduced by 30%, through path optimization, AGV energy consumption is reduced by 12%, which further reduces the operating cost;

[0109] Prediction and control module, based on LSTM neural network analysis device historical data, predict potential failure and trigger preventive maintenance work order, real-time monitoring production quality;

[0110] The hidden layer captures the front and rear dependency relationship through the Bi-LSTM layer, and the Dropout layer prevents overfitting; the output layer uses the classification task Softmax activation function, and through the regression task linear activation function, the code example is as follows:

[0111] ```python

[0112] from keras.models import Sequential

[0113] from keras.layers import LSTM, Bidirectional, Dense, Dropout

[0114] model = Sequential()

[0115] model.add(Bidirectional(LSTM(64, return_sequences=False), input_shape=(60, 6)))

[0116] model.add(Dropout(0.2))

[0117] model.add(Dense(1, activation='sigmoid')) # classification task (fault / normal)

[0118] model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

[0119] ```

[0120] The edge computing node receives sensor data in real time, inputs the LSTM model in a sliding window format, the motor temperature of the AGV is uploaded every second, the sliding window length is 60 seconds, and a prediction sequence is generated; the output fault category is output, such as overheating, bearing wear, and electrical short circuit; the remaining service life or fault probability is output; the fault probability threshold is set according to historical data, and the threshold is dynamically adjusted; the prediction result is pushed to the MES system through the API interface, and a maintenance work order is automatically generated; the fault probability is > 90% after 30 minutes of prediction at the welding station, and the MES system automatically issues a “check the welding gun cooling system” work order; through the Bi-LSTM+Attention mechanism, long-term dependencies in the production process are captured, such as welding current fluctuations causing weld defects, and key abnormal points are identified through attention weights, and the model is regularly updated to adapt to changes in process parameters; Bi-LSTM modeling is performed on quality detection data to identify abnormal patterns, and control limits are set using statistical process control methods;

[0121] According to the historical maintenance records of similar equipment, such as the typical failure time of the welding gun cooling system, the fault probability threshold is set; if the welding gun cooling system usually has a fault probability of 80% to 90% 30 minutes before failure, the fault probability threshold is set to 90%;

[0122] The actual maintenance results, such as whether the fault occurred, are fed back to the training set, the model is retrained, and MLflow or DVC is used to track the model version to ensure traceability; the maintenance work order status (completed / delayed) is fed back to the prediction model to optimize future predictions; the prediction results are synchronized to the digital twin model to verify the effectiveness of the maintenance strategy;

[0123] For example, in the automobile welding station fault prediction and quality monitoring, the welding station has weld quality fluctuations due to equipment aging, and the process parameters need to be predicted in advance and optimized, the welding current, voltage, and weld width sensor data are collected every second, and the Bi-LSTM+Attention model is used to predict the weld defect probability; when the defect probability is > 70%, the welding current is automatically adjusted and a maintenance work order is generated; the weld defect rate is reduced from 5% to 1.2%, and the equipment downtime is reduced by 40%.

[0124] Example 2

[0125] Please refer to Figure 2 Based on example 1, example 2 of the present application also provides a management method for a wire control MES system, including the following specific steps:

[0126] Step one, real-time collection of production line operation data, operation data including sensor data, production line data and process parameter data;

[0127] Step two, by deploying lightweight algorithms on edge nodes, processing the operation data, uploading the processed operation data to the cloud for deep analysis, including trend analysis, bottleneck identification and prediction modeling; combined with the deep analysis results, real-time mapping of physical production environment, simulation of different scheduling strategy response;

[0128] Step three, based on the production line data, using reinforcement learning algorithm to dynamically adjust the task priority, and introducing buffer capacity adjustment mechanism to cope with load fluctuation; receiving the process limiting link information predicted by the MES system, combining with the distributed path planning strategy, dynamically adjusting the transportation path;

[0129] Step four, based on LSTM neural network analysis of device historical data, predicting potential failure and triggering preventive maintenance work order, real-time monitoring of production quality.

[0130] The above examples can be realized by software, hardware, firmware or any combination thereof, in whole or in part. When realized by software, the above examples can be realized in the form of computer program product in whole or in part. Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or combination of computer software and electronic hardware. Whether the functions are realized by hardware or software depends on the specific application and design constraints of the technical solution.

[0131] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, which can be located in one place or distributed on multiple network units. Part or all of the units can be selected to achieve the purpose of the embodiment according to the actual needs.

[0132] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A management system for a by-wire MES system, characterized in that, The system comprises: a data acquisition module, which acquires real-time production line operation data, including sensor data, production line data and process parameter data; a data processing module, which processes the operation data by deploying lightweight algorithms on edge nodes, uploads the processed operation data to the cloud for deep analysis, including trend analysis, bottleneck identification and prediction modeling; combines the deep analysis results to map the physical production environment in real time, simulates different scheduling strategies and responses; a task scheduling and optimization module, which dynamically adjusts task priorities based on production line data using reinforcement learning algorithms, and introduces a buffer capacity adjustment mechanism to cope with load fluctuations; receives process limiting link information predicted by the MES system, and dynamically adjusts the transportation path based on a distributed path planning strategy; a prediction and control module, which analyzes device historical data based on an LSTM neural network, predicts potential failures and triggers preventive maintenance work orders, and monitors production quality in real time; wherein the trend analysis includes: using moving average and wavelet transform to eliminate sensor noise, applying moving average with a large window to the denoised data to extract long-term trends, and using unit root test to determine trend stationarity; combining autocorrelation function and partial autocorrelation function diagrams to identify periodic components, using ARIMA model to predict periodic fluctuations of the differenced stationary sequence, superimposing trend and periodic terms to generate a complete trend curve, and updating model parameters based on real-time MES data; The process of dynamically adjusting the transportation path is: decomposing the global task into multiple subtasks, and distributing the subtasks to the AGV optimal route through a distributed algorithm; the AGV bids for the subtask according to its own state, and preferentially selects the optimal path.

2. A management system for a wire control MES system according to claim 1, characterized in that, The process of real-time acquisition of production line operation data is: By deploying near the production line, sensor data is collected; sensor data includes temperature sensors, pressure sensors and displacement sensors; Production line data includes real-time material location, order urgency and material inventory; Use standardized protocols to transfer data collected by sensors to the MES system for real-time monitoring and adjustment, and save to process parameter data.

3. The management system for a wire-based MES system according to claim 1, wherein, The process of processing operation data is: Deploy low-complexity models on edge nodes to clean and extract features from operation data in real time; use distributed computing to process multiple data streams in parallel, combine LSTM neural network to predict device failure, and dynamically adjust task priorities through reinforcement learning algorithm; Use event-driven architecture to trigger immediate response; Dynamically allocate resources through containerized deployment on edge nodes, and upload critical data to the cloud for deep analysis based on TSN time-sensitive network; Clean the operation data, and unify the operation data from different sensors into a standard format, and structure the unstructured operation data.

4. The management system for a wire-based MES system according to claim 1, wherein, The process of deep analysis also includes: Bottleneck identification: Monitor equipment OEE through Prometheus, combine Apriori algorithm to mine association rules between equipment failure and process parameters; Calculate equipment utilization rate based on sensor data and MES system indicators, if the equipment idle rate exceeds the threshold, it is determined as a bottleneck; Otherwise, it is not a bottleneck; Use RFID to monitor buffer inventory, combined with production line topology to locate the source of accumulation; Prediction modeling: Extract the feature frequency of equipment failure, use random forest to predict equipment failure; Based on the long-term dependence relationship of time series, model the equipment aging trend, quantify the trend direction and duration; Embed the failure prediction model and bottleneck rule into the MES system.

5. The management system for a wire-based MES system according to claim 1, wherein, The process of real-time mapping of the physical production environment is as follows: Deploy a digital twin platform in the cloud, create corresponding virtual entities, use industrial simulation software to build a runnable production line model, define event-driven mechanisms, set state transition logic, and configure task scheduling strategies; Through edge computing nodes, collect and process running data, and upload it to the cloud in real time. When new data is uploaded, the simulation engine updates the model state in real time.

6. A management system for a wire control MES system according to claim 5, characterized in that, The process of dynamically adjusting task priority using reinforcement learning algorithm is as follows: S301: Obtain historical running data, use Plant Simulation to build a digital twin model, and simulate the dynamic behavior of the production line; S302: Balance production efficiency, cost, and quality through discrete action space and continuous action space algorithms combined with entropy weight multi-objective decision method; S303: Initialize the model, randomly generate scheduling strategies, update Q values according to the reward function, use experience replay to store historical state-action-reward data, and use target network to stabilize the training process; When the model reaches the preset performance indicators in the simulation environment, stop training; Verify the robustness of the model in extreme scenarios.

7. A management system for a wire control MES system according to claim 6, characterized in that, The specific process of introducing buffer capacity adjustment mechanism is as follows: Set physical buffers before and after each key station, dynamically adjust the release rhythm of upstream tasks according to the current buffer queue length; If the buffer is close to full, reduce the priority of tasks before the key station; Conversely, speed up the release of tasks.

8. The management system for a wire-based MES system according to claim 1, wherein, The process of predicting potential failures and triggering preventive maintenance work orders is as follows: Edge computing nodes receive sensor data in real time, input LSTM model in sliding window format, generate prediction sequence with sliding window length, output fault category, set fault probability threshold based on historical data, and dynamically adjust the threshold; Push the prediction results to the MES system through the API interface, automatically generate maintenance work orders, capture long-term dependencies in the production process through Bi-LSTM+Attention mechanism, attention weight identifies key abnormal points, and regularly update the model.

9. A management method for a by-wire MES system, characterized by, The process includes the following steps: Step 1: Collect real-time production line running data, including sensor data, production line data, and process parameter data; Step two, process the running data by deploying lightweight algorithms at the edge node, upload the processed running data to the cloud for deep analysis, including trend analysis, bottleneck identification and prediction modeling; combine the deep analysis results, real-time map the physical production environment, simulate different scheduling strategies response; Step three, based on the production line data, use reinforcement learning algorithm to dynamically adjust the task priority, and introduce buffer capacity adjustment mechanism to cope with load fluctuations; receive the process limiting link information predicted by the MES system, combine with the distributed path planning strategy, dynamically adjust the transportation path; Step four, based on LSTM neural network to analyze the historical data of equipment, predict potential failure and trigger preventive maintenance work order, real-time monitoring of production quality; Wherein, the trend analysis includes: using moving average and wavelet transform to eliminate sensor noise, applying large window moving average to the denoising data to extract long-term trend, judging the trend stationarity through unit root test; identify periodic components by combining autocorrelation function and partial autocorrelation function diagram, use ARIMA model to predict periodic fluctuation rule of difference stationary sequence, superimpose trend item and periodic item to generate complete trend curve, and update model parameters based on MES real-time data; The process of dynamically adjusting the transportation path is: decompose the global task into multiple subtasks, assign the subtasks to the AGV optimal route through distributed algorithm; AGV bids for subtasks according to its own state, and preferentially selects the optimal path.

Images