Mould abnormity management system and method based on machine vision

Abstract

The invention relates to the technical field of mold management, in particular to a mold abnormity management system and method based on machine vision, and the system comprises a mold abnormity detection module, a fault prediction and health management module, an intelligent repair and maintenance scheduling module, and a data analysis and visualization module. The mold anomaly detection module collects data by using a high-definition camera and a thermal infrared imager, and identifies surface defects through a convolutional neural network; the fault prediction and health management module predicts a fault risk and performs early warning by means of time sequence analysis and a machine learning technology; the intelligent repair and maintenance scheduling module optimizes a repair strategy and task allocation by using deep reinforcement learning and a genetic algorithm; and the data analysis and visualization module displays data by means of Grafana and Kibana and realizes intelligent alarm and decision support. According to the invention, accurate detection of mold abnormity, effective prediction of faults, optimization of maintenance scheduling and efficient processing and visual display of data are realized, and the intelligent level of mold management is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mold management, and in particular to a mold abnormality management system and method based on machine vision. Background Art

[0002] In modern manufacturing, molds are key production tools, and their normal operation is crucial to product quality and production efficiency. However, molds are easily affected by a variety of factors during use and may become abnormal. Traditional mold management methods face many challenges in dealing with these problems.

[0003] Traditional mold inspection mainly relies on manual regular inspection, which is highly subjective and inefficient. Manual inspection cannot guarantee that all key parts of the mold can be fully and carefully inspected every time, and it is easy to miss some potential abnormalities. Moreover, the frequency of manual inspection is limited, and it is impossible to monitor the operating status of the mold in real time, resulting in failure to detect abnormalities in the mold in time, causing production interruptions, product quality degradation and other problems. For example, on some large-scale production lines, the mold runs continuously for a long time. If the tiny cracks or wear on the mold surface cannot be detected in time, it may lead to the production of a large number of defective products, increase production costs, and affect the economic benefits of the enterprise.

[0004] In terms of fault prediction, traditional methods lack effective data analysis methods. Often, they can only roughly judge the time when the mold may fail based on experience, and cannot accurately predict the specific time and type of failure. This makes it difficult for companies to formulate reasonable maintenance plans in advance. They can only adopt preventive excessive maintenance or emergency repairs after the failure occurs. The former will increase unnecessary maintenance costs, and the latter may not be able to choose the best maintenance plan due to the tight maintenance time, further affecting production progress.

[0005] In terms of repair and maintenance scheduling, traditional management models usually make arrangements based on simple rules or manual experience, without fully considering the actual health status of the mold, the urgency of the production task, and the reasonable allocation of maintenance resources. This may lead to unreasonable arrangements for maintenance tasks, such as repairing non-critical molds when production tasks are tight, affecting normal production; or delaying maintenance time due to improper allocation of maintenance resources when mold failure is serious, aggravating the degree of equipment damage.

[0006] With the development of industrial automation and intelligence, the manufacturing industry has continuously increased its requirements for production efficiency, product quality and equipment management level. The traditional mold management method can no longer meet the needs of modern production. Therefore, a mold abnormality management system and method based on machine vision is proposed to address the above problems. Summary of the invention

[0007] The purpose of the present invention is to provide a mold abnormality management system and method based on machine vision to solve the problems raised in the above background technology.

[0008] To achieve the above object, the present invention provides the following technical solutions: A mold abnormality management system based on machine vision, comprising: Mold anomaly detection module: It is equipped with a high-definition camera and an infrared thermal imager. The high-definition camera is used to collect mold surface images, and the infrared thermal imager is used to monitor mold temperature changes. The mold anomaly detection module uses a convolutional neural network (CNN) to extract features and identify surface defects based on the collected mold surface images to detect whether there are any abnormalities on the mold surface. Fault prediction and health management module: Relying on time series analysis and machine learning technology, the long short-term memory (LSTM) network is used to analyze mold operation data to predict fault risks. The support vector machine (SVM) and random forest (RF) are used to process mold health data to achieve fault warning and provide a visual display of mold health status. Intelligent repair and maintenance scheduling module: uses deep reinforcement learning (DRL) to optimize the repair scheduling strategy according to the health status of the mold, and uses genetic algorithm (GA) to schedule repair tasks under multi-task and resource constraints, and push equipment status and maintenance suggestions in real time; Data analysis and visualization module: Apache Kafka and Apache Spark are used to process mold data streams, and Grafana and Kibana are used to display relevant data. A data integration platform is set up to aggregate different mold data for real-time analysis, and intelligent alarm push and decision support are implemented based on the analysis results.

[0009] As a preferred solution, the mold anomaly detection module uses the generative adversarial network GAN to generate defect images to enhance data, thereby improving the generalization ability of the convolutional neural network CNN; the generative adversarial network GAN consists of a generator and the discriminator Composition, Generator The goal is to generate defect images that are as realistic as possible, and its loss function can be expressed as: ; Discriminator The goal is to distinguish between real images and images generated by the generator, and its loss function is expressed as: ; Through adversarial training, the generator and the discriminator Continuously optimize and generate defect images that meet the requirements to enhance data for improving the generalization ability of CNN, among which, Indicates input noise The prior distribution of Represents the distribution of real image data.

[0010] As a preferred solution, the fault prediction and health management module combines LSTM, SVM and RF algorithms to improve the accuracy and reliability of fault prediction by using data fusion. The hidden state and cell status The update formula is as follows: ; ; ; ; ; ; in, , , are the activation values ​​of the input gate, forget gate, and output gate, respectively. is the sigmoid function, represents element-wise multiplication, , , , They are the bias items of input gate, forget gate, output gate and candidate cell state respectively. For the moment The input data, For the moment The temporary state obtained by processing the tanh function, For input to the weight matrix of the input gate, Hide the state for the last moment to the weight matrix of the input gate, For the time step The hidden state of For input To the weight matrix of the forget gate, is the hidden state of the previous time step To the weight matrix of the forget gate, For input The weight matrix to the output gate, is the hidden state of the previous time step The weight matrix to the output gate, For input to the weight matrix of the candidate cell state, is the hidden state of the previous time step To the weight matrix of the candidate cell state; perform time series analysis on the mold operation data through LSTM, extract features and predict failure risks; When support vector machine SVM is used for binary classification, its goal is to find a hyperplane To maximize the interval between the two types of data, the optimization objective function is: ; ; in, is the normal vector of the hyperplane, is the offset, is the training sample, is the corresponding category label; combined with mold health data for fault warning; Random Forest RF builds multiple decision trees and combines their results to make predictions. Assuming that the training set have samples, each time from Extraction with Replacement Construct a sub-training set of samples ,in , is the number of decision trees, for decision trees , which selects information gain when a node splits The largest feature is split, and the information gain calculation formula is: ; in, Characterized by Features The number of values ​​of Features The value is A subset of samples, For the dataset The entropy of For subset Entropy; use RF to process mold health data to achieve fault warning, and improve the accuracy and reliability of overall fault prediction by fusing the results of LSTM, SVM and RF.

[0011] As a preferred solution, the intelligent repair and maintenance scheduling module uses the deep reinforcement learning DRL model to dynamically adjust the maintenance strategy and optimize the maintenance scheduling based on the real-time health status of the mold and the predicted failure risk; DRL is built based on the Markov decision process MDP and defines the state space , Action Space , reward function And the state transition probability , the agent learns a strategy by continuously interacting with the environment To maximize long-term cumulative rewards , and its calculation formula is: ; in, is the discount factor, By taking corresponding maintenance scheduling actions according to different mold health states in the mold management environment, the mold health state is the state space Elements, maintenance scheduling actions are action space Elements, dynamically optimize strategies based on reward function feedback , to optimize maintenance scheduling, and optimize maintenance scheduling based on the real-time health status of the mold and predicted failure risks.

[0012] As a preferred solution, the data analysis and visualization module integrates the multi-dimensional data of the mold through the data integration platform, performs real-time processing and intelligent decision support, and pushes abnormal notifications to relevant personnel through the intelligent alarm system; when this module uses Apache Kafka and Apache Spark to process the mold data stream, the core computing logic of Apache Spark is based on the elastic distributed dataset RDD for data processing. , assuming a mapping operation is performed , generate a new RDD , and its conversion relationship can be expressed as: .

[0013] As a preferred solution, the Grafana and Kibana systems of the data analysis and visualization module are used to display the operating status, health data and fault warning information of the mold through dashboards, charts and log analysis; when constructing visualization charts, Grafana constructs bar charts to display the frequency of occurrence of different mold failure types. , and its corresponding frequency is , in the chart, the horizontal axis represents the fault type , the vertical axis represents the frequency , visual display is achieved through graphics drawing functions, while Kibana uses the timestamp of log data to display log analysis. , Event Type , and displays them according to the set time series and classification rules, so that users can view and analyze mold-related data and status changes.

[0014] As a preferred solution, the intelligent repair and maintenance scheduling module optimizes the allocation and scheduling of maintenance tasks by considering multiple resources and task constraints through the genetic algorithm GA. Based on the idea of ​​population evolution, GA first initializes a Population of individuals , individuals are represented by chromosome encoding, assuming that the maintenance task is Item, resources are species, and chromosomes are represented by length The encoding string for individuals in the population , its fitness function It is calculated based on the maintenance task completion time and resource utilization rate, and is expressed as: ;in, According to individual The total time required for the solution to complete the maintenance task, Indicates the corresponding resource utilization, , is the corresponding weight coefficient. Through selection, crossover and mutation, the population is continuously evolved to find the individual with the highest fitness, that is, the optimal maintenance task allocation and scheduling scheme, so as to achieve the optimal scheduling of maintenance tasks under multi-task and resource constraints.

[0015] A mold abnormality management method based on machine vision is implemented by using a mold abnormality management system based on machine vision.

[0016] It can be seen from the technical solution provided by the present invention that the present invention provides a mold abnormality management system and method based on machine vision, which has the following beneficial effects: 1. Improve the accuracy and efficiency of mold anomaly detection Multimodal data fusion detection: The mold surface image is collected through a high-definition camera, and the convolutional neural network (CNN) is used to extract features and identify surface defects. At the same time, the temperature change of the mold is monitored with the help of an infrared thermal imager. Combining visual information with temperature information can comprehensively and accurately determine whether there are abnormalities on the mold surface, effectively improving the accuracy of abnormality detection and reducing misjudgments and missed judgments. Real-time monitoring and rapid response: High-definition cameras and infrared thermal imagers can collect mold data in real time, and the system processes and analyzes the data immediately. Once an abnormality is detected, an alarm is quickly issued to notify relevant personnel to take timely measures. Compared with the traditional regular manual inspection method, the time interval for abnormality discovery is shortened, the safety and stability of the production process are improved, and the risk of production interruption caused by mold failure is reduced, thereby improving production efficiency. 2. Accurate fault prediction and reliable health management Advanced algorithms improve prediction accuracy: Long short-term memory (LSTM) network is used to analyze mold operation data. LSTM can process time series data, capture long-term dependencies in data, and accurately predict fault risks. Support vector machine (SVM) and random forest (RF) are used to process mold health data. Through data fusion, the advantages of multiple algorithms are combined to improve the reliability of fault warning. SVM performs better in processing small samples and nonlinear problems. RF enhances the robustness and generalization ability of the model by building multiple decision trees and integrating their results. Visual display helps decision-making: Provides a visual display of mold health status, presenting complex mold health information to management and maintenance personnel in the form of intuitive charts (such as bar charts showing the frequency of different mold failure types, trend charts showing the changes in key mold parameters over time, etc.) and clear indicator lights (such as green for normal, yellow for warning, and red for failure); enabling them to quickly understand the overall operation status of the mold, discover potential problems in a timely manner, and provide strong support for formulating reasonable maintenance plans and production decisions; 3. Optimize repair and maintenance scheduling strategies State-based dynamic scheduling: Deep reinforcement learning (DRL) is used to optimize maintenance scheduling strategies based on the real-time health status of the mold and predicted failure risks. In a complex mold management environment, the DRL model can dynamically select the optimal maintenance action (such as immediate maintenance, delayed maintenance, replacement of parts, etc.) based on the constantly changing state information (such as the current health status of the mold, the urgency of the production task, the availability of maintenance resources, etc.), thereby achieving reasonable allocation and efficient use of maintenance resources. Task optimization under multiple constraints: Genetic algorithm (GA) is used to schedule maintenance tasks under multiple tasks and resource constraints. Taking into account multiple resource factors such as the number of maintenance personnel, the use of maintenance equipment, the inventory of maintenance materials, and task constraints such as the sequence and time requirements between different maintenance tasks, the GA algorithm can search for the optimal maintenance task allocation plan. It helps to improve maintenance efficiency, shorten maintenance time, reduce maintenance costs, ensure that key mold maintenance work is given priority under limited resource conditions, and ensure the stable operation of the production system. 4. Efficient data processing and comprehensive decision support Powerful data processing capabilities: Apache Kafka and Apache Spark are used to process mold data streams, which can meet the needs of large-scale, high-real-time data collection and processing. Kafka, as a distributed message queue, reliably collects and caches data from data sources to ensure data sequence and integrity. Spark provides powerful distributed computing capabilities to quickly process and analyze real-time data streams, such as calculating mold operation indicators and counting the number of faults. Based on core computing logic such as resilient distributed datasets (RDD), efficient data conversion, cleaning and feature extraction operations are achieved, providing an accurate and timely data basis for subsequent analysis and decision-making. Intelligent alarm and decision support: With the help of Grafana and Kibana to display relevant data, the mold operation status, health data and fault warning information are presented in an intuitive way such as dashboards and charts, which is convenient for users to monitor in real time and conduct in-depth analysis. A data integration platform is set up to aggregate data from different molds for real-time analysis. When an abnormal situation is detected, intelligent alarm push is implemented based on the analysis results to notify relevant personnel to take measures. The comprehensive data and in-depth analysis results provided by the system provide strong data support and scientific basis for managers to make production plans, adjust equipment maintenance strategies, optimize resource allocation and other decisions, helping enterprises to improve the overall operation management level, reduce production costs, and enhance market competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The figure is a schematic diagram of the overall structure of a mold abnormality management system and method based on machine vision of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0019] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0020] like Figure 1 As shown, an embodiment of the present invention provides a mold anomaly management system based on machine vision, including a mold anomaly detection module, a fault prediction and health management module, an intelligent repair and maintenance scheduling module, and a data analysis and visualization module.

[0021] In this embodiment, the mold abnormality detection module is configured with a high-definition camera and an infrared thermal imager. The high-definition camera is used to collect mold surface images, and the infrared thermal imager is used to monitor mold temperature changes. The mold abnormality detection module uses a convolutional neural network (CNN) to extract features and identify surface defects based on the collected mold surface images to detect whether there is an abnormality on the mold surface. Furthermore, the mold anomaly detection module uses a high-definition camera and an infrared thermal imager to monitor the mold, and then uses a convolutional neural network (CNN) to process the collected information to detect whether there is an abnormality on the mold surface. The specific operation steps are as follows: Image and temperature data acquisition: Mold surface image acquisition: Use a high-definition camera to capture mold surface images under different angles, lighting conditions, and other conditions to ensure that comprehensive visual information of the mold surface can be obtained; Mold temperature change monitoring: Use infrared thermal imager to monitor the temperature change of mold in real time during operation and record the temperature data at different times; Data preprocessing: Image data cleaning: De-noise the collected mold surface images to remove salt and pepper noise, Gaussian noise, etc. in the images to improve image quality; at the same time, perform image enhancement operations, such as adjusting contrast and brightness, to make the details of the mold surface clearer and facilitate subsequent feature extraction; Temperature data calibration and normalization: Calibrate the temperature data collected by the infrared thermal imager to ensure the accuracy of temperature measurement; then, normalize the temperature data to make it within a specific value range to facilitate fusion analysis with other data; Using CNN to extract image features: Construct CNN model: Build a convolutional neural network structure suitable for mold surface defect detection, including convolution layer, pooling layer, fully connected layer, etc.; determine the size, number, step size and other parameters of the convolution kernel, as well as the type and parameter settings of the pooling layer, to achieve effective extraction of mold surface image features; Model training and optimization: The preprocessed mold surface image data is divided into training set, validation set and test set; the CNN model is trained using the training set data, and the weights and biases of the model are continuously adjusted through the back-propagation algorithm so that the model can accurately extract the features of the mold surface image; during the training process, appropriate optimization algorithms (such as stochastic gradient descent and its variants) and loss functions (such as cross entropy loss function) are used to optimize model performance, and model selection and tuning are performed based on indicators such as the accuracy of the validation set; Feature extraction: Input the test set or the actual monitored mold surface image into the trained CNN model, and automatically extract the deep features in the image through layer-by-layer processing of convolutional layers and pooling layers, such as the texture, shape, color and other feature information of the mold surface, and convert it into a numerical feature vector that can be understood by the computer, providing a basis for subsequent surface defect identification; Surface defect recognition and anomaly detection: Establish a defect classification model: Based on the extracted mold surface image features, combined with known mold surface defect types and sample data, train a classification model (such as support vector machine, decision tree, etc.) so that the model can determine whether there are defects on the mold surface and the type of defects based on the feature vector; Real-time detection and judgment: During the operation of the mold, the mold surface image is collected in real time and input into the trained CNN model for feature extraction, and then the extracted feature vector is input into the defect classification model for judgment; if the classification model output result shows that there is a defect on the mold surface, the defect type is further determined, and combined with the temperature change data monitored by the infrared thermal imager for comprehensive analysis to determine whether the abnormality may cause more serious problems; for example, if a surface defect appears in a certain area and the temperature in the area rises abnormally, it may indicate a potential serious failure risk; Abnormal alarm and recording: When an abnormality is detected on the mold surface, the alarm system is immediately triggered to notify relevant staff through sound, text message, email, etc. At the same time, the relevant data of the abnormal situation (including images, temperature data, detection time, etc.) is recorded for subsequent fault analysis and tracing.

[0022] In this embodiment, the fault prediction and health management module relies on time series analysis and machine learning technology to analyze mold operation data through long short-term memory LSTM network to predict fault risks, combines support vector machine SVM and random forest RF to process mold health data to achieve fault warning, and provides a visual display of mold health status; Furthermore, the fault prediction and health management module mainly uses a variety of technologies and algorithms to analyze and process mold operation and health data, thereby realizing fault prediction, early warning and health status visualization. The specific operation steps are as follows: Data collection and integration: Operation data collection: collect relevant data of the mold in operation in real time from various sensors of the mold (such as pressure sensor, temperature sensor, displacement sensor, etc.) and the production control system, including pressure value, temperature value, operation speed, number of mold opening and closing, etc., and record the timestamp corresponding to each data point to form time series data; Health data collection: collect data related to the health status of the mold, such as mold maintenance records (including maintenance time, maintenance content, replaced parts, etc.), parts life information, material property data, and mold internal structure status data obtained through regular inspections or non-destructive testing methods; Data integration and cleaning: Integrate operating data and health data to ensure data integrity and consistency; clean the collected data to remove outliers, missing values, and erroneous data; for missing values, interpolation methods (such as linear interpolation, mean interpolation, etc.) can be used to supplement; for outliers, set reasonable thresholds or use methods based on statistical principles (such as the 3σ principle) to identify and correct them; Using LSTM for fault risk prediction: Data preprocessing and feature engineering: Perform feature engineering on the integrated mold operation data and select key features related to fault prediction, such as temperature change trend, pressure fluctuation amplitude, and operating speed stability over a period of time; normalize or standardize these feature data to make them have a unified dimension to facilitate LSTM network analysis; Build LSTM model: Design a suitable LSTM network structure according to the characteristics of mold operation data and the requirements of fault prediction, determine model parameters such as the number of hidden layers, number of neurons, and time step; initialize the weights and biases of the model, and select a suitable activation function (such as tanh or sigmoid function) to calculate the output of neurons; Model training and optimization: Divide the preprocessed running data into training set, validation set and test set; use the training set data to train the LSTM model, calculate the error between the model prediction value and the actual value through the back propagation algorithm, and adjust the model weight and bias according to the error; during the training process, use optimization algorithms (such as the Adam optimization algorithm) to accelerate the training process and improve the convergence of the model; at the same time, according to the performance indicators of the validation set (such as mean square error, accuracy, etc.), perform model tuning to avoid overfitting and select the model with the best performance; Fault risk prediction: The real-time collected mold operation data is input into the trained LSTM model. The model predicts the future operation status of the mold based on the time series pattern learned from historical data, and outputs a fault risk probability value or a fault probability score. Based on the set threshold, it is determined whether the mold has a high fault risk. For example, if the predicted fault risk probability exceeds 0.8, it is considered that the mold may be about to fail. Combining SVM and RF to achieve fault warning: Feature selection and data preparation: For mold health data, select features that are closely related to fault warning, such as the last maintenance time interval, the remaining life ratio of key components, and the number of internal structural defects detected recently; encode and preprocess these feature data to make them suitable for the input requirements of the support vector machine (SVM) and random forest (RF) algorithms; SVM model construction and training: Construct an SVM model for binary classification (normal and faulty), determine the kernel function type (such as linear kernel, radial basis kernel, etc.) and related parameters (such as penalty factor C, kernel function parameter γ, etc.); use the training set of mold health data to train the SVM model, find the optimal hyperplane by optimizing the objective function, separate data of different categories as accurately as possible, and enable the model to determine whether the mold is in a faulty state based on the input health data characteristics; RF model construction and training: Build a random forest model and determine parameters such as the number of decision trees, tree depth, and feature selection strategy; randomly extract multiple sub-training sets from the training set of mold health data with replacement, and each sub-training set is used to train a decision tree; in the process of building the decision tree, select the best features for node splitting based on criteria such as information gain, so that the decision tree can classify and predict the mold health status; finally, the prediction output of the random forest model is obtained by integrating the prediction results of multiple decision trees; Fault warning decision: Input the real-time mold health data into the trained SVM and RF models respectively to obtain the fault warning results of the two models (such as normal or faulty). Fusion strategies such as voting method or weighted average method can be used to make a comprehensive judgment on the results of the two models. For example, if one or more of the two models predicts a fault, a fault warning signal is issued to notify relevant personnel to take timely measures to avoid the occurrence or expansion of the fault. Visual display of mold health status: Data visualization design: Determine the method and content of visualization, such as designing dynamic dashboards, trend charts, bar charts, health status indicators and other visualization elements to display the real-time health status of the mold, failure risk trends, key parameter changes and other information; select appropriate visualization tools and libraries (such as Matplotlib and Seaborn in Python or Echarts in JavaScript, etc.) for development; Real-time data update and display: Establish a data interface to transmit the mold health status data obtained by the fault prediction and health management module to the visual display interface in real time; ensure that the data on the visual interface is always up to date through timed refresh or event triggering mechanism; on the dashboard, intuitively display the current overall health status of the mold through color changes (such as green for normal, yellow for warning, and red for fault); in the trend chart, draw the curve of the change of key parameters of the mold (such as temperature, pressure, etc.) over time, as well as the change trend of the fault risk probability, to help users observe the dynamic change process of the mold status; use a bar chart to show the frequency of different types of faults or the health status distribution of different components, so that users can quickly understand the main problems of the mold; Implementation of interactive functions: Add interactive functions to the visual display interface, such as clicking on chart elements to view detailed data, zooming in and out of charts to view data for a specific time period, switching between different visual views, etc., to facilitate users to deeply analyze mold health status data, identify potential problems in a timely manner and make decisions; for example, users can click on a time point on the trend chart, and a pop-up window will display the detailed parameter values ​​of the mold at that moment, so as to further analyze abnormal situations.

[0023] In this embodiment, the intelligent repair and maintenance scheduling module uses deep reinforcement learning (DRL) to optimize the maintenance scheduling strategy according to the mold health status, and uses the genetic algorithm (GA) to schedule maintenance tasks under multi-task and resource constraints, and push equipment status and maintenance suggestions in real time; Furthermore, the intelligent repair and maintenance scheduling module mainly uses deep reinforcement learning and genetic algorithms to optimize the maintenance scheduling strategy and task allocation, and push equipment status and maintenance suggestions in real time. The specific operation steps are as follows: Maintenance scheduling strategy optimization based on DRL: Environmental modeling: The operating environment of the mold management system is abstractly modeled to define the state space, action space, reward function and state transition probability; the state space includes the health status of the mold (such as normal, minor fault, serious fault, etc.), production task schedule, maintenance resource availability and other information; the action space is the optional maintenance scheduling actions, such as immediate maintenance, delayed maintenance, replacement of parts, etc.; the reward function is designed according to the impact of the maintenance scheduling strategy on production efficiency, cost, equipment life, etc. For example, timely repair of the faulty mold and restoration of normal production can get positive rewards, while negative rewards are given for production interruptions caused by untimely maintenance; the state transition probability describes how the system state transfers to the next state after taking a certain action; DRL model construction and training: Select appropriate deep reinforcement learning algorithms (such as deep Q network DQN, proximal policy optimization algorithm PPO, etc.) to build the model; the input of the model is the current environment state, and the output is the corresponding optimal maintenance scheduling action; through a large number of simulation experiments or historical data for training, the intelligent agent continuously takes actions in the environment, and adjusts the parameters of the model according to reward feedback to learn the optimal maintenance scheduling strategy, that is, the strategy that maximizes long-term cumulative rewards; Real-time strategy application and adjustment: In the actual mold management process, the mold health status and system environment information are obtained in real time and input into the trained DRL model. The model outputs the optimal maintenance scheduling action under the current state; the maintenance scheduling operation is performed according to the action, and the impact of the execution result on the system state is observed in real time; if the actual situation does not match the model's expectations, new data is collected in a timely manner and fed back to the model for online learning and strategy adjustment, so that the model can adapt to the ever-changing mold operation environment; Maintenance task scheduling optimization based on GA: Coding and population initialization: Encode the maintenance task allocation scheme according to the maintenance tasks and resource constraints; for example, use chromosome encoding to allocate each maintenance task to a specific maintenance resource and time slot to form an individual; initialize a population containing multiple individuals, which represent different maintenance task allocation possibilities; Fitness function definition: Design the fitness function according to the characteristics and target requirements of the maintenance task; the fitness function comprehensively considers factors such as the maintenance task completion time, resource utilization, and maintenance cost. For example, individuals with shorter maintenance task completion time, higher resource utilization, and lower maintenance cost can be given higher fitness values; in this way, individuals with high fitness are more likely to be retained and selected during the evolution of the genetic algorithm; Genetic operation execution: genetic operations such as selection, crossover and mutation are performed on individuals in the population; the selection operation selects excellent individuals to enter the next generation population based on the fitness value of the individuals by using methods such as roulette selection and tournament selection; the crossover operation generates new individuals by exchanging some gene fragments of the chromosomes of two individuals, simulating the gene recombination process and increasing the diversity of the population; the mutation operation randomly changes certain gene positions of individual chromosomes, introduces new gene information, and prevents the algorithm from falling into the local optimal solution; by executing these genetic operations multiple times, the population continues to evolve and gradually approaches the optimal maintenance task allocation plan; Determination and application of the optimal solution: After the genetic algorithm iteration reaches the preset termination condition (such as reaching the maximum number of iterations or the fitness of the population no longer increases significantly), the individual with the highest fitness is selected from the final population as the optimal maintenance task allocation solution; this solution is applied to the actual maintenance task scheduling, and maintenance resources and maintenance time are reasonably allocated to ensure that maintenance tasks are completed efficiently under the conditions of meeting multi-task and resource constraints, thereby improving the overall operating efficiency of the mold management system; Equipment status push and maintenance suggestion generation: Real-time status monitoring and data collection: Through various sensors installed on the mold (such as temperature sensors, vibration sensors, pressure sensors, etc.) and integration with the production control system, the mold's operating status data is collected in real time, including parameters such as temperature, vibration amplitude, pressure changes, and operating time. At the same time, relevant data such as the mold's historical maintenance records and usage environment information are collected to provide comprehensive data support for equipment status evaluation and maintenance recommendation generation; Equipment status assessment and analysis: Based on the collected data, the current status of the mold is assessed and analyzed using data analysis algorithms and models. For example, by setting thresholds or establishing fault diagnosis models, it is possible to determine whether the mold is in normal operation, whether there is a potential fault risk, and the severity of the fault. Based on factors such as the mold's operating time, frequency of use, and degree of wear, the remaining service life of components is predicted to provide a basis for preventive maintenance. Generation and push of maintenance recommendations: Generate personalized maintenance recommendations based on equipment status assessment results, combined with mold maintenance manuals and best practices; maintenance recommendations include maintenance items (such as cleaning, lubrication, tightening, replacement of wearing parts, etc.), maintenance intervals, required maintenance resources (manpower, materials, tools, etc.) and maintenance operation steps; push equipment status information and maintenance recommendations to relevant maintenance personnel, equipment management personnel and production operators in real time via SMS, email, system pop-up windows, etc., to ensure that they can understand the health status of the mold in a timely manner and take corresponding maintenance measures to extend the service life of the mold and ensure the continuity and stability of production.

[0024] In this embodiment, the data analysis and visualization module uses Apache Kafka and Apache Spark to process the mold data stream, uses Grafana and Kibana to display the relevant data, and is equipped with a data integration platform to aggregate different mold data for real-time analysis, and implements intelligent alarm push and decision support based on the analysis results; Furthermore, the data analysis and visualization module is mainly responsible for processing mold data flow, displaying data, and providing decision support. The specific operation steps are as follows: Data collection and transmission (Apache Kafka): Data source connection and configuration: Determine the source of mold data, such as sensor data, production system data, maintenance records, etc., and configure the corresponding data source connection method; create topics in Apache Kafka to receive and store mold data from different data sources; each topic can be divided according to data type or business needs, for example, one topic is used to store temperature sensor data, another topic is used to store pressure sensor data, etc.; Data collection and transmission: Write a data collection program or use existing data collection tools to collect mold data from various data sources in real time; encapsulate the collected data in a predetermined format and send it to the corresponding Kafka topic; Kafka, as a distributed message queue, can efficiently process a large amount of real-time data and ensure reliable transmission and sequentiality of data; for example, the sensor data collection program reads the sensor value at a certain time interval and sends a message containing the sensor ID, timestamp and data value to the corresponding Kafka topic; Data caching and management: Kafka caches the received data and manages the data storage according to the configured retention policy (such as retention by time or by data size). This allows you to easily obtain historical data from Kafka if you need to backtrack or reprocess part of the data during subsequent data processing, ensuring data traceability and flexibility. Data Processing and Analysis (Apache Spark): Spark cluster construction and configuration: Build an Apache Spark cluster, configure the number of cluster nodes and resource allocation (such as the number of CPU cores, memory size, etc.) according to the scale and performance requirements of data processing; ensure that Spark and Kafka can communicate normally so as to read data from Kafka for processing; Data reading and conversion (based on RDD or Data Frame): Use Spark APIs (such as Spark Streaming or Spark Structured Streaming) to read mold data streams from Kafka topics in real time; convert the read data into Spark's resilient distributed dataset (RDD) or data frame (Data Frame) format for further processing and analysis; during the conversion process, data cleaning operations can be performed, such as removing invalid data, processing missing values, parsing data formats, etc., to ensure data quality and consistency; for example, for the sensor data read in string format, parse it into a numerical type, check whether the data is within a reasonable range, and remove abnormal values; Data analysis and calculation: Use the rich function libraries and algorithms provided by Spark to perform real-time analysis and calculation on the converted data; this can include statistical analysis (such as calculating the average, standard deviation, maximum, minimum, etc.), feature engineering (such as extracting key features of the data, performing data dimension reduction, etc.), machine learning model training and prediction (such as using clustering algorithms to classify mold operating status, building regression models to predict mold failure time, etc.); define corresponding analysis logic and calculation tasks according to business needs; for example, by analyzing historical temperature data, establish a temperature change trend model to monitor in real time whether the mold temperature deviates from the normal range; Result storage and output: Store the results of data analysis in a suitable storage system, such as Hadoop Distributed File System (HDFS), relational databases (such as MySQL, PostgreSQL), or NoSQL databases (such as MongoDB), for subsequent query and use; at the same time, some real-time analysis results can also be output to message queues or other systems for further processing or display by downstream modules; for example, send the calculated real-time status indicators of molds (such as health scores, failure probability, etc.) to another Kafka topic for use by the intelligent alarm module; Data Integration and Aggregation: Data integration platform construction: Build a data integration platform and select appropriate technologies and tools (such as Apache Sqoop, Apache Flume or a custom data integration framework) to achieve data transmission and integration between different data sources and data processing platforms; configure data integration tasks and define the extraction, transformation, and loading (ETL) process from data sources to data warehouses or data lakes; Data extraction and conversion: extract relevant data from various data sources (including mold design data in relational databases, production planning data in production management systems, maintenance logs in text files, etc.) and convert them according to unified data models and standards; this may involve operations such as data format conversion, data type matching, and field mapping to convert data of different structures and formats into a unified format suitable for aggregation and analysis; for example, converting the date format from one representation to another common date format, and converting values ​​of different units uniformly; Data aggregation and storage: The converted data is aggregated according to predetermined rules and dimensions and stored in a data warehouse or data lake. Data warehouse architecture design such as a star model or a snowflake model can be used to facilitate multi-dimensional data analysis. For example, based on the mold number and time dimension, data such as the production quantity, number of failures, and maintenance costs of different molds in different time periods can be aggregated to form an aggregated data set, providing data support for a comprehensive understanding of the overall situation and historical trends of the molds. Data visualization (Grafana and Kibana): Grafana configuration and dashboard creation: Install and configure Grafana, connect to the data storage system that stores the analysis results (such as the database mentioned above or the data source of the real-time analysis results output); create a dashboard in Grafana, and design the layout and chart type according to the mold data content that needs to be displayed; for example, create a dashboard to display the real-time operation status of the mold, including visualization elements such as temperature curves, pressure bar graphs, and equipment status indicators; Data query and chart configuration: Configure data query statements in the Grafana dashboard to obtain the required data from the data source and map it to the corresponding chart; according to the data characteristics and display requirements, set the chart style, axis labels, data series names and other properties to make the chart clearly and intuitively present the mold data information; for example, for a temperature curve chart, set the horizontal axis to time and the vertical axis to temperature value, and select the appropriate line color and width so that users can clearly observe the temperature change trend over time; Kibana log analysis configuration (optional): If mold-related log data display is involved, install and configure Kibana, and import log data into Kibana; create an index pattern (IndexPattern) in Kibana, define the time field and other key fields of the log data, so as to perform time series analysis and query filtering; by creating visualizations (Visualize) and dashboards, the log data is displayed in the form of tables, bar charts, line charts, etc., to help users analyze event records and error information during mold operation, and gain an in-depth understanding of the mold's operating status and potential problems; for example, display information such as the number of mold startups and the number of fault alarm logs in different time periods to help users find the root cause of the problem; Visualization and interactive operations: Start the Grafana and Kibana services, and users can access the corresponding interfaces through the browser to view the visualization of mold data. Users can perform interactive operations on the dashboard, such as zooming in and out of the chart to view data details for a specific time period, switching between different chart views, and clicking on chart elements to obtain detailed data information. Through these interactive functions, users can explore mold data more deeply, discover potential patterns and problems, and provide strong support for decision-making. Intelligent alarm push and decision support: Alarm rule setting: intelligent alarm rules are set in the data analysis and visualization module according to mold management requirements and business rules; alarm rules can be triggered based on real-time data analysis results, threshold settings, trend changes and other conditions; for example, when the mold temperature exceeds the set safety threshold, the failure probability exceeds a certain proportion, or the production efficiency is lower than the expected level, the corresponding alarm event is triggered; Alarm information generation and push: Once the alarm rule is triggered, the system automatically generates alarm information, including alarm type, alarm time, relevant mold information (such as mold number, location, etc.) and detailed alarm data (such as abnormal temperature value, fault prediction details, etc.); through SMS gateway, mail server or instant messaging tools and other channels, the alarm information is pushed to relevant personnel (such as equipment maintenance personnel, production management personnel, etc.) in real time to ensure that they can receive the alarm in time and take corresponding measures; Decision support data provision: In addition to alarm push, the data analysis and visualization module also provides rich data basis for decision support; by displaying historical data trends, performance comparisons between different molds, maintenance cost analysis and other information, it helps managers to fully understand the overall situation of mold management and evaluate the impact of different decision-making plans on production efficiency, cost control, equipment life, etc.; for example, when deciding whether to replace a mold, managers can refer to the mold's historical failure frequency, maintenance costs, and expected benefits after replacing the new mold, and make scientific and reasonable decisions.

[0025] In this embodiment, the fault prediction and health management module combines LSTM, SVM and RF algorithms to improve the accuracy and reliability of fault prediction by using data fusion. The long short-term memory LSTM network is used at the time The hidden state and cell status The update formula is as follows: ; ; ; ; ; ; in, , , are the activation values ​​of the input gate, forget gate, and output gate, respectively. is the sigmoid function, represents element-wise multiplication, , , , They are the bias items of input gate, forget gate, output gate and candidate cell state respectively. For the moment The input data, For the moment The temporary state obtained by processing the tanh function, For input to the weight matrix of the input gate, Hide the state for the last moment to the weight matrix of the input gate, For the time step The hidden state of For input To the weight matrix of the forget gate, is the hidden state of the previous time step To the weight matrix of the forget gate, For input The weight matrix to the output gate, is the hidden state of the previous time step The weight matrix to the output gate, For input to the weight matrix of the candidate cell state, is the hidden state of the previous time step To the weight matrix of the candidate cell state; perform time series analysis on the mold operation data through LSTM, extract features and predict failure risks; When support vector machine SVM is used for binary classification, its goal is to find a hyperplane To maximize the interval between the two types of data, the optimization objective function is: ; ; in, is the normal vector of the hyperplane, is the offset, is the training sample, is the corresponding category label; combined with mold health data for fault warning; Random Forest RF builds multiple decision trees and combines their results to make predictions. Assuming that the training set have samples, each time from Extraction with Replacement Construct a sub-training set of samples ,in , is the number of decision trees, for decision trees , which selects information gain when a node splits The largest feature is split, and the information gain calculation formula is: ; in, Characterized by Features The number of values ​​of Features The value is A subset of samples, For the dataset The entropy of For subset Entropy; use RF to process mold health data to achieve fault warning, and improve the accuracy and reliability of overall fault prediction by fusing the results of LSTM, SVM and RF.

[0026] In this embodiment, the intelligent repair and maintenance scheduling module uses the deep reinforcement learning DRL model to dynamically adjust the maintenance strategy and optimize the maintenance scheduling based on the real-time health status of the mold and the predicted failure risk; wherein, DRL is constructed based on the Markov decision process MDP and defines the state space , Action Space , reward function And the state transition probability , the agent learns a strategy by continuously interacting with the environment To maximize long-term cumulative rewards , and its calculation formula is: ; in, is the discount factor, By taking corresponding maintenance scheduling actions according to different mold health states in the mold management environment, the mold health state is the state space Elements, maintenance scheduling actions are action space Elements, dynamically optimize strategies based on reward function feedback , to optimize maintenance scheduling, and optimize maintenance scheduling based on the real-time health status of the mold and predicted failure risks.

[0027] In this embodiment, the data analysis and visualization module integrates the multi-dimensional data of the mold through the data integration platform, performs real-time processing and intelligent decision support, and pushes abnormal notifications to relevant personnel through the intelligent alarm system; when this module uses Apache Kafka and Apache Spark to process the mold data stream, the core computing logic of Apache Spark performs data processing based on the elastic distributed dataset RDD. , assuming a mapping operation is performed , generate a new RDD , and its conversion relationship can be expressed as: .

[0028] In this embodiment, the Grafana and Kibana systems of the data analysis and visualization module are used to display the operating status, health data and fault warning information of the mold through dashboards, charts and log analysis; when constructing visualization charts, Grafana constructs a bar chart to display the frequency of occurrence of different mold failure types. , and its corresponding frequency is , in the chart, the horizontal axis represents the fault type , the vertical axis represents the frequency , visual display is achieved through graphics drawing functions, while Kibana uses the timestamp of log data to display log analysis. , Event Type , and displays them according to the set time series and classification rules, so that users can view and analyze mold-related data and status changes.

[0029] The present invention will be further described below in conjunction with specific embodiments: Example 1: Mold abnormality detection module example Image acquisition and preprocessing: Install a high-definition camera on the mold production line and adjust its position and angle to ensure that the key parts of the mold surface can be clearly photographed; set the camera's shooting frequency to once every 5 minutes, and take 3 images from different angles each time; The collected images are preprocessed. First, the images are converted into grayscale images to reduce the amount of data and highlight the texture features of the mold surface. Then, the median filter algorithm is used to remove the salt and pepper noise in the image, and the filter kernel size is 3×3. Next, the image is processed with histogram equalization to enhance the image contrast and make the details of the mold surface more clearly visible. Data augmentation using Generative Adversarial Networks (GANs): A generative adversarial network (GAN) was constructed, in which the generator (G) used a 4-layer fully connected neural network structure, with the number of neurons in each layer being 1024, 512, 256, and 128 respectively, and the last layer outputting a defect image with the same size as the mold surface image; the discriminator (D) used a 3-layer convolutional neural network structure, with convolution kernel sizes of 3×3, 5×5, and 7×7 respectively, and a step size of 2, to distinguish between real images and images generated by the generator; The loss function of the generator (G) is defined as: ,in, It follows the prior distribution The loss function of the discriminator (D) is: ,in, is the real image data, is the distribution of real image data; The collected normal mold surface images are used as the initial data set, and the generator (G) and the discriminator (D) are continuously optimized through adversarial training. After 100 training cycles, the generator (G) can generate more realistic defect images, which are mixed with the original normal images to form an enhanced training data set for subsequent convolutional neural network (CNN) training. Surface defect recognition and anomaly detection based on CNN: Construct a convolutional neural network (CNN) model, which includes 3 convolutional layers, 2 pooling layers, and 2 fully connected layers. The convolutional kernel size of the convolutional layer is 3×3, and the number is 32, 64, and 128 respectively, and the activation function is ReLU. The pooling layer uses maximum pooling, the pooling kernel size is 2×2, and the stride is 2. The number of neurons in the fully connected layer is 256 and 2 (for binary classification: normal and abnormal). The CNN model is trained using the enhanced training data set, the learning rate is set to 0.001, and the Adam optimization algorithm is used. The update formula is: ; in, and are the first-order moment estimates and second-order moment estimates of the gradient, and is the decay rate, is the gradient of the current time step, and are the modified first-order moment estimates and second-order moment estimates, is the learning rate, is a small constant, the training batch size is 32, and the training is performed for 1000 rounds. During the training process, the accuracy of the model on the validation set is evaluated every 100 rounds. When the accuracy no longer improves, the training is stopped and the optimal model is saved. During the operation of the mold, the mold surface image is collected in real time and input into the trained CNN model for prediction after preprocessing; if the prediction result output by the model is abnormal, further analysis is performed in combination with the mold temperature data monitored by the infrared thermal imager; if an area on the mold surface is judged to be abnormal and the temperature in this area is significantly increased (exceeding the normal operating temperature range by more than 10°C), it is determined that the mold has surface defects and may have internal structural problems, triggering the alarm system to notify relevant personnel to conduct inspection and repairs.

[0030] Example 2: Fault prediction and health management module example Data collection and integration: Collect operation data from the mold's temperature sensor, pressure sensor, displacement sensor, etc., with a sampling frequency of once per second; at the same time, obtain the mold maintenance record from the mold maintenance management system, including the time of each maintenance, maintenance content, replaced parts and other information, as well as the initial installation time and expected service life of the parts and components; Integrate the operating data and health data according to the mold number and timestamp to form a unified data set; process the missing values ​​in the data set. For a small number of missing values ​​in the operating data, use the average value of adjacent data points to fill them; for missing values ​​in the health data, make reasonable estimates or supplement them based on maintenance records and component life information; detect and correct abnormal values. For example, for data points that obviously deviate from the normal temperature range (such as exceeding ±50% of the normal temperature), combine the actual working status of the mold and historical data to make a judgment. If it is determined to be abnormal, use the average value of the adjacent normal data points to replace it; Failure risk prediction based on LSTM: Select key features such as temperature, pressure, and displacement from the mold operation data, and divide these feature data into time series. Each time series contains data points from the past 10 minutes (i.e., the time step is 600 and the sampling frequency is 1 time per second). Normalize the feature data so that its value range is between 0 and 1. Use the minimum-maximum normalization method. The formula is: ; in, is the normalized value, is the original value, and are the minimum and maximum values ​​in the feature data; Construct an LSTM network, set the number of hidden layers to 2, and the number of neurons in each hidden layer to 128; the number of neurons in the input layer is determined according to the feature dimension of the time series (for example, if there are 3 features, the number of neurons in the input layer is 3), and the output layer is 1 neuron, which is used to predict the probability of fault risk; select tanh as the activation function, and initialize the weights and biases to random values; Use 70% of the data set as the training set, 20% as the validation set, and 10% as the test set; set the training rounds to 500, the learning rate to 0.005, and use the Adam optimization algorithm for training. The formula is: ,in, is the sample size, is the true value, is the predicted value; during the training process, the mean square error (MSE) on the validation set is calculated every 50 rounds. When the MSE of the validation set does not decrease for 10 consecutive rounds, the training is stopped and the optimal model is saved; The real-time collected mold operation data is subjected to feature extraction and preprocessing in the same way, and then input into the trained LSTM model to obtain the predicted value of the failure risk probability. If the predicted value exceeds the set threshold of 0.7, it is considered that the mold has a high failure risk and a warning signal is triggered. Fault warning combining SVM and RF: Select features such as the time interval between the last maintenance, the remaining life ratio of key components, and the number of recently detected internal structural defects from the mold health data as input features; encode these features, for example, convert the time interval between the last maintenance into a value in days, convert the remaining life ratio of key components into a value between 0 and 1, and classify and encode the number of internal structural defects (such as 0 for no defect, 1 for minor defect, 2 for severe defect, etc.); Construct a support vector machine (SVM) model, select the radial basis function (RBF) as the kernel function, and determine the penalty factor through cross-validation and kernel function parameters ; Use the training set data to train the SVM model so that the model can accurately classify the mold status into normal and faulty categories based on the input features. When the support vector machine SVM is used for binary classification, its goal is to find a hyperplane To maximize the interval between the two types of data, the optimization objective function is: ; ; in, is the normal vector of the hyperplane, is the offset, is the training sample, is the corresponding category label; combined with mold health data for fault warning; A random forest (RF) model was constructed, the number of decision trees was set to 100, the maximum depth of the tree was set to 10, and the number of randomly selected features was the square root of the total number of features. The RF model was trained using the training set data. During the decision tree construction process, the best feature was selected for node splitting based on information gain. The information gain calculation formula is: ; in, Characterized by Features The number of values ​​of Features The value is A subset of samples, For the dataset The entropy of For subset Entropy of The calculation formula for entropy is: ,in, is the number of categories, Yes Category proportion in the dataset; Input the real-time mold health data into the trained SVM and RF models to obtain the prediction results of the two models; use the voting method to merge them. If more than one of the two models predicts a fault, a fault warning signal is issued and the warning information (including mold number, warning time, possible fault cause, etc.) is pushed to relevant personnel; Visual display of mold health status: Use Python's Dash framework to develop a visual display interface; design a dynamic dashboard on the interface, including mold operation status indicator lights (green for normal, yellow for warning, and red for failure), real-time temperature and pressure curve charts, failure risk probability trend charts, and key parts remaining life bar charts and other visual elements; The mold health status data analyzed by the fault prediction and health management module can be obtained in real time through the data interface, and the data on the dashboard is updated every 10 seconds; when the mold status changes (such as from normal to warning or fault), the color of the corresponding indicator light changes immediately, and a prompt box pops up on the interface to display detailed status change information; users can view detailed data at a specific time point by clicking on the chart element, or zoom in and out on the chart to view data trends over a longer time range.

[0031] Embodiment 3: Intelligent repair and maintenance scheduling module embodiment Maintenance scheduling strategy optimization based on DRL: The mold management environment is modeled as a Markov decision process (MDP); the state space is defined as a combination of factors such as the mold health status (normal, minor failure, major failure), the urgency of the production task (high, medium, low), and the availability of maintenance resources (idle, partially occupied, fully occupied). It indicates that, Indicates health status, Indicates the urgency of the production task. represents the availability of maintenance resources; the action space is defined as the maintenance strategy, including immediate maintenance, delayed maintenance, replacement of parts, continued observation, etc. Represents; the reward function is designed as follows: when the production task is completed on time and the mold is kept in good condition, a positive reward is given ; If the production task is interrupted due to mold failure, a negative reward will be given ; For maintenance operations, give corresponding rewards or penalties based on maintenance costs and recovery time; State transition probability Estimates are made based on historical data and actual operating experience; The DRL model is constructed using the Deep Q Network (DQN) algorithm. The input layer of the model receives the state vector, which is processed by two hidden layers (each with 64 neurons and ReLU activation function) to output the Q value of each action. The experience replay buffer is used to store the experience of the agent's interaction with the environment. , the buffer size is 10000; set the learning rate to 0.001, the discount factor , the stochastic gradient descent (SGD) algorithm is used to optimize the model parameters. The update formula of the SGD algorithm is: ,in, are the model parameters at the current time step, is the learning rate, is the loss function About parameters The gradient of When the model is trained, the agent Select Action (use Greedy strategy, The initial value is 0.5, which gradually decreases with training). The reward for observing the feedback from the environment after performing the action and the next state , store the experience in the playback buffer; randomly extract a batch of experience samples from the buffer to calculate the target Q value and update the model parameters; the training process continues for 10,000 training steps, and the model performance is evaluated every 1,000 steps. When the average reward of the model in the verification environment no longer increases, the training is stopped to obtain the optimized maintenance scheduling strategy model; In actual mold management, the mold status information is obtained in real time and converted into a state vector Input the trained DRL model, and the model outputs the optimal maintenance scheduling action ; Perform corresponding maintenance scheduling operations according to the actions, continuously monitor environmental changes, and continuously adjust maintenance strategies based on new status information; Maintenance task scheduling optimization based on GA: Assume that there are 5 molds that need maintenance and there are 3 types of maintenance resources (maintenance personnel, maintenance equipment, and maintenance materials) available. Encode the maintenance task allocation scheme as a chromosome with a chromosome length of 15 (each mold corresponds to 3 genes, representing the allocated maintenance personnel, maintenance equipment, and maintenance materials respectively). For example, chromosome [1,2,1,3,1,2,2,3,2,1,3,1,3,2,1] means that mold 1 is allocated maintenance personnel 1, maintenance equipment 2, and maintenance material 1, and mold 2 is allocated maintenance personnel 3, maintenance equipment 1, and maintenance material 2, etc. Initialize the population and randomly generate 50 individuals (maintenance task allocation scheme); define the fitness function ,in, According to individual The total time required for the solution to complete the maintenance task, Indicates the corresponding resource utilization rate; calculate the fitness value of each individual, and use the roulette wheel selection method to select 30 individuals as parents. The selection probability calculation formula of the roulette wheel selection method is: ,in, Is an individual The probability of being selected, Is an individual The fitness value of is the population size; Perform a crossover operation on the parent individuals, using the two-point crossover method with a crossover probability of 0.8; then perform a mutation operation on the individuals after the crossover, with a mutation probability of 0.05, and the mutation method is to randomly change the value of a gene on the chromosome; a new generation of population is generated through selection, crossover and mutation operations, and this process is repeated until the maximum number of iterations is 100; The individual with the highest fitness is selected from the final population as the optimal maintenance task allocation plan. According to the plan, maintenance personnel are arranged, maintenance equipment is allocated, and maintenance materials are prepared to repair the mold. At the same time, the maintenance plan is adjusted in real time according to the actual use of maintenance resources and the progress of maintenance tasks to ensure that maintenance tasks are carried out efficiently and orderly. Equipment status push and maintenance suggestion generation: Temperature sensors, vibration sensors and pressure sensors are installed on the mold to collect the operating status data of the mold in real time, with a sampling frequency of once per minute; the collected data is sent to the data processing center via wireless transmission; After receiving the data, the data processing center first cleans and preprocesses the data to remove noise and outliers. Then, the operating status of the mold is determined based on historical data and preset thresholds. For example, when the mold temperature exceeds the normal operating temperature range by more than 5°C and the vibration amplitude exceeds the normal range by 30%, it is determined that the mold may be at risk of failure. Generate maintenance recommendations based on the mold's operating time, usage frequency, and current status, combined with the maintenance requirements in the maintenance manual. The recommendations include maintenance items (such as cleaning the mold surface, checking the lubricant level, tightening connectors, etc.), maintenance time (such as immediate maintenance or maintenance after the next production task), and required maintenance resources (such as the required lubricant type and quantity, cleaning tools, etc.). Send equipment status information and maintenance recommendations to equipment maintenance personnel and production managers via the SMS platform to remind them to perform maintenance work in a timely manner.

[0032] Embodiment 4: Data analysis and visualization module embodiment Data collection and transmission (Apache Kafka): Deploy multiple sensors in the mold production workshop, including temperature sensors, pressure sensors, displacement sensors, etc. The sensors are connected to the data acquisition server through a wired network. Install the Apache Kafka client on the data acquisition server, configure the Kafka producer, and send the data collected by the sensors according to different topics. For example, send the temperature sensor data to the "mold_temperature" topic, and send the pressure sensor data to the "mold_pressure" topic. A unique identifier is set for each sensor, and a timestamp is attached when sending data to ensure the accuracy and traceability of the data; the data collection frequency is set to once per second to ensure that the operating status information of the mold can be obtained in real time; the Kafka server adopts a distributed deployment mode and sets up 3 nodes to improve the reliability and throughput of data processing; Configure Kafka's message storage policy and set the data retention time to 7 days so that historical data can be traced back for analysis when needed. At the same time, enable Kafka's compression function to reduce network transmission and storage costs. Data Processing and Analysis (Apache Spark): Build an Apache Spark cluster, including 1 master node and 5 slave nodes, each node is equipped with 8-core CPU and 16GB memory; configure the Spark Streaming application in the Spark cluster to read the mold data stream from Kafka; The read data is processed by first parsing the data into a structured format and extracting key information (such as sensor identifiers, timestamps, data values, etc.). Then, the data is cleaned to remove invalid data (such as data values ​​that are negative or exceed a reasonable range) and duplicate data. For missing data, the sliding average method is used to fill in the missing data, that is, the average value of the previous data points is used to replace the missing value. Perform real-time analysis and calculation based on the data after cleaning; calculate statistical indicators such as the average temperature, maximum pressure, displacement change range, etc. of each mold in different time periods; at the same time, use machine learning algorithms (such as clustering algorithms) to classify the operating status of the mold to determine whether the mold is in normal operating state, potential fault state, or abnormal state; for example, use the K-Means clustering algorithm, take the temperature and pressure data of the mold as features, and classify the operating status into three categories: normal, warning, and fault; The analysis results are stored in the Hadoop Distributed File System (HDFS) and stored in layers according to date and mold number to facilitate subsequent query and analysis. At the same time, some real-time analysis results (such as the status classification results of the current mold) are sent to another Kafka topic for use by the intelligent alarm module. Data Integration and Aggregation: Establish a data integration platform, use Apache Sqoop to extract mold design parameters, production batch information, historical maintenance records and other data from relational databases (such as MySQL), and import them into the Hive data warehouse; at the same time, use Apache Flume to collect relevant data from text files (such as maintenance log files) and store them in HDFS; Create external tables in Hive to integrate data from different data sources; summarize data based on mold number and time dimensions; for example, calculate summary indicators such as monthly production quantity, number of failures, and maintenance costs for each mold; use Hive's SQL query function to convert and process data and unify the data format into a format suitable for analysis; Store the aggregated data in Hive partition tables to improve query performance; regularly update and maintain data to ensure timeliness and accuracy; Data visualization (Grafana and Kibana): Install and configure Grafana, and connect to the Hive database that stores the analysis results; create a dashboard in Grafana and design a visual chart to show the operating status of the mold; for example, create a temperature trend chart with time as the horizontal axis and temperature as the vertical axis to show the temperature changes of different molds in the past 24 hours; create a failure count bar chart, grouped by mold number, to show the number of failures that occurred for each mold in the past month; Configure Grafana's data query statements to obtain the corresponding data from Hive and fill it into the chart; set the style and properties of the chart, such as line color, bar chart color, axis label, etc., to make the chart more beautiful and intuitive; at the same time, add interactive functions to the chart, such as clicking on chart elements to view detailed data, zooming in and out of the chart to view data in different time periods, etc.; Install and configure Kibana, and import mold log data stored in Elasticsearch into Kibana; create index patterns in Kibana, and define time fields and key fields for time series analysis and query filtering; create visualization pages to display key information in mold logs, such as operation records and error messages in different time periods, to help users analyze events and problems during mold operation; Intelligent alarm push and decision support: Set intelligent alarm rules in the data analysis and visualization module; for example, when the mold temperature exceeds 90% of the upper limit of the normal operating temperature and lasts for more than 5 minutes, a high temperature alarm is triggered; when the pressure sensor data fluctuation exceeds 30% of the normal range and lasts for more than 10 minutes, a pressure abnormality alarm is triggered; when the cluster analysis results show that the mold operation status is a fault state, a fault alarm is triggered; When the alarm rule is triggered, the system automatically generates alarm information, including alarm type, alarm time, relevant mold information (such as mold number, location, etc.) and detailed alarm data (such as abnormal temperature value, pressure fluctuation curve, etc.); the alarm information is sent to the mobile phones of equipment maintenance personnel and production management personnel through the SMS gateway, and the alarm information is highlighted on the Grafana dashboard to ensure that relevant personnel can discover and handle problems in a timely manner; Provide decision support for managers; by displaying historical data trends (such as the relationship between mold failure frequency and production efficiency in the past year), performance comparisons between different molds (such as the average failure interval between molds of different models, maintenance cost comparisons, etc.), cost analysis reports (such as the trend of changes in the proportion of maintenance costs to production costs), and other information, help managers evaluate the overall performance of molds and formulate reasonable maintenance strategies and production plans; for example, according to historical data, if it is found that the failure frequency of a certain mold is gradually increasing and the maintenance cost is high, managers can consider arranging to replace the mold in advance to avoid greater impact on production; at the same time, according to the cost analysis report, adjust the maintenance budget allocation and optimize resource utilization efficiency.

[0033] A mold abnormality management method based on machine vision is implemented by using a mold abnormality management system based on machine vision.

[0034] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A mold abnormality management system based on machine vision, characterized by: include: Mold abnormality detection module: equipped with a high-definition camera and an infrared thermal imager. The high-definition camera is used to collect mold surface images, and the infrared thermal imager is used to monitor mold temperature changes. The mold abnormality detection module uses a convolutional neural network (CNN) to extract features and identify surface defects based on the collected mold surface images to detect whether there are abnormalities on the mold surface. Fault prediction and health management module: Relying on time series analysis and machine learning technology, the long short-term memory (LSTM) network is used to analyze mold operation data to predict fault risks. The support vector machine (SVM) and random forest (RF) are used to process mold health data to achieve fault warning and provide a visual display of mold health status. Intelligent repair and maintenance scheduling module: uses deep reinforcement learning (DRL) to optimize the repair scheduling strategy according to the health status of the mold, and uses genetic algorithm (GA) to schedule repair tasks under multi-task and resource constraints, and push equipment status and maintenance suggestions in real time; Data analysis and visualization module: Apache Kafka and Apache Spark are used to process mold data streams, and Grafana and Kibana are used to display relevant data. A data integration platform is set up to aggregate different mold data for real-time analysis, and intelligent alarm push and decision support are implemented based on the analysis results.

2. According to claim 1, a mold abnormality management system based on machine vision is characterized in that: The mold anomaly detection module uses the generative adversarial network GAN to generate defect images to enhance data, thereby improving the generalization ability of the convolutional neural network CNN; the generative adversarial network GAN consists of a generator and the discriminator Composition, Generator The goal is to generate defect images that are as realistic as possible, and its loss function can be expressed as: ； Discriminator The goal is to distinguish between real images and images generated by the generator, and its loss function is expressed as: ； Through adversarial training, the generator and the discriminator Continuously optimize and generate defect images that meet the requirements to enhance data for improving the generalization ability of CNN, among which, Indicates input noise The prior distribution of Represents the distribution of real image data.

3. The mold abnormality management system based on machine vision according to claim 1 is characterized by: The fault prediction and health management module combines LSTM, SVM and RF algorithms to improve the accuracy and reliability of fault prediction by using data fusion. The long short-term memory LSTM network is used at the time The hidden state and cell status The update formula is as follows: ； ； ； ； ； ； in, , , are the activation values ​​of the input gate, forget gate, and output gate, respectively. is the sigmoid function, represents element-wise multiplication, , , , They are the bias items of input gate, forget gate, output gate and candidate cell state respectively. For the moment The input data, For the moment The temporary state obtained by processing the tanh function, For input to the weight matrix of the input gate, Hide the state for the last moment to the weight matrix of the input gate, For the time step The hidden state of For input To the weight matrix of the forget gate, is the hidden state of the previous time step To the weight matrix of the forget gate, For input The weight matrix to the output gate, is the hidden state of the previous time step The weight matrix to the output gate, For input to the weight matrix of the candidate cell state, is the hidden state of the previous time step To the weight matrix of the candidate cell state; perform time series analysis on the mold operation data through LSTM, extract features and predict failure risks; When support vector machine SVM is used for binary classification, its goal is to find a hyperplane To maximize the interval between the two types of data, the optimization objective function is: ； s.t. ; in, is the normal vector of the hyperplane, is the offset, is the training sample, is the corresponding category label; combined with mold health data for fault warning; Random Forest RF builds multiple decision trees and combines their results to make predictions. Assuming that the training set have samples, each time from Extraction with Replacement Construct a sub-training set of samples ,in , is the number of decision trees, for decision trees , which selects information gain when a node splits The largest feature is split, and the information gain calculation formula is: ； in, Characterized by Features The number of values ​​of Features The value is A subset of samples, For the dataset The entropy of For subset Entropy; use RF to process mold health data to achieve fault warning, and improve the accuracy and reliability of overall fault prediction by fusing the results of LSTM, SVM and RF.

4. The mold abnormality management system based on machine vision according to claim 1, characterized in that: The intelligent repair and maintenance scheduling module uses the deep reinforcement learning DRL model to dynamically adjust the maintenance strategy and optimize the maintenance scheduling based on the real-time health status of the mold and the predicted failure risk; wherein, DRL is constructed based on the Markov decision process MDP and defines the state space , Action Space , reward function And the state transition probability , the agent learns a strategy by continuously interacting with the environment To maximize long-term cumulative rewards , and its calculation formula is: ； in, is the discount factor, By taking corresponding maintenance scheduling actions according to different mold health states in the mold management environment, the mold health state is the state space Elements, maintenance scheduling actions are action space Elements, dynamically optimize strategies based on reward function feedback , to optimize maintenance scheduling, and optimize maintenance scheduling based on the real-time health status of the mold and predicted failure risks.

5. The mold abnormality management system based on machine vision according to claim 1, characterized in that: The data analysis and visualization module integrates the multi-dimensional data of the mold through the data integration platform, performs real-time processing and intelligent decision support, and pushes abnormal notifications to relevant personnel through the intelligent alarm system; when this module uses Apache Kafka and Apache Spark to process the mold data stream, the core computing logic of Apache Spark performs data processing based on the elastic distributed dataset RDD. , assuming a mapping operation is performed , generate a new RDD , and its conversion relationship can be expressed as: .

6. The mold abnormality management system based on machine vision according to claim 1, characterized in that: The Grafana and Kibana systems of the data analysis and visualization module are used to display the operation status, health data and fault warning information of the mold through dashboards, charts and log analysis; When building a visualization chart, Grafana builds a bar chart to show the frequency of different mold failure types. , and its corresponding frequency is , the horizontal axis in the chart represents the fault type , the vertical axis represents the frequency , visual display is achieved through graphics drawing functions, while Kibana uses the timestamp of log data to display log analysis. , Event Type , and displays them according to the set time series and classification rules, so that users can view and analyze mold-related data and status changes.

7. The mold abnormality management system based on machine vision according to claim 1 is characterized by: The intelligent repair and maintenance scheduling module optimizes the allocation and scheduling of maintenance tasks based on multiple resource and task constraints through the genetic algorithm GA; GA is based on the idea of ​​population evolution. First, it initializes a Population of individuals , individuals are represented by chromosome encoding, assuming that the maintenance task is Item, resources are species, and chromosomes are represented by length The encoding string for individuals in the population , its fitness function It is calculated based on the maintenance task completion time and resource utilization rate, and is expressed as: ;in, According to individual The total time required for the solution to complete the maintenance task, Indicates the corresponding resource utilization, , is the corresponding weight coefficient. Through selection, crossover and mutation, the population is continuously evolved to find the individual with the highest fitness, that is, the optimal maintenance task allocation and scheduling scheme, so as to achieve the optimal scheduling of maintenance tasks under multi-task and resource constraints.

8. A mold abnormality management method based on machine vision, characterized in that: The mold abnormality management method is performed using the machine vision-based mold abnormality management system described in any one of claims 1-7.

Images