Production-manufacturing-oriented digital twinning and dynamic optimization management method for test process

Abstract

The invention provides a manufacturing-oriented test process digital twinning and dynamic optimization management method, which relates to the technical field of testing, and comprises the following steps of: acquiring real-time test data through a multi-source data acquisition device, and establishing a test process digital model; constructing a test process digital twin system; establishing a test process prediction model by using a deep learning algorithm; comparing and analyzing the real-time data and the prediction result, and identifying abnormity and optimization opportunities; generating and implementing a test process optimization scheme; and closed-loop optimization management is realized through feedback control. According to the method, the efficiency and quality of the test process are improved.

Description

Technical Field

[0001] The present invention relates to testing technology, and in particular to a testing process digital twin and dynamic optimization management method for production and manufacturing. Background Art

[0002] With the rapid development of intelligent manufacturing technology, the test process in the manufacturing process is a key link to ensure product quality, and its digitalization and intelligence level are particularly important. At present, manufacturing companies generally use test data acquisition systems, quality control systems and production scheduling systems to support test process management, but the following problems still exist in actual applications:

[0003] First, traditional test process management mainly relies on static data statistics and empirical judgment. Key links such as test equipment operation status monitoring, test parameter optimization and adjustment, and quality abnormality cause analysis lack real-time dynamic analysis and decision-making mechanisms. The test data acquisition system can only record historical data and cannot map and dynamically predict the test process in real time, making it difficult to improve the management efficiency and quality control level of the test process.

[0004] Second, existing test parameter optimization methods usually use a single statistical model or optimization algorithm. These methods are difficult to effectively handle multi-source heterogeneous data during the test process, and cannot accurately characterize the complex nonlinear relationship between test parameters, equipment status and quality indicators. Especially in the face of dynamic disturbances such as equipment performance fluctuations and changes in working conditions, the adaptability and robustness of the optimization results are poor.

[0005] Third, there is a lack of systematic solutions for intelligent decision-making in the test process. Current test management systems often separate decision-making tasks such as parameter optimization, equipment maintenance, and resource scheduling, and each uses an independent decision-making model. This decentralized decision-making mechanism is difficult to achieve global optimization, and cannot effectively balance multiple goals such as quality improvement, efficiency improvement, and cost savings. It is also not conducive to the collaborative management and continuous improvement of the test process.

[0006] In response to the above problems, it is urgent to establish a digital twin system of the test process for production and manufacturing to realize real-time mapping, dynamic optimization and intelligent decision-making of the test process, thereby improving the management level and quality assurance capabilities of the test process. Summary of the invention

[0007] The embodiments of the present invention provide a digital twin and dynamic optimization management method for a test process for production and manufacturing, which can solve the problems in the prior art.

[0008] According to a first aspect of the embodiments of the present invention,

[0009] Provides digital twin and dynamic optimization management methods for test processes in production and manufacturing, including:

[0010] The real-time test data in the production and manufacturing test process is collected by a multi-source data acquisition device, the multi-source data acquisition device includes a production line sensor network, a test equipment data interface and a quality inspection system, and the collected real-time test data includes workpiece processing parameters, equipment operation status data, environmental monitoring data and test process quality data; the real-time test data is cleaned and standardized by a data preprocessing module to establish a test process digital model, the test process digital model includes a test process model, an equipment status model, a quality control model and a resource scheduling model;

[0011] A test process digital twin system is constructed based on the test process digital model by adopting a layered architecture design, wherein the layered architecture includes a physical layer, a data layer, a model layer, and an application layer; historical test data is analyzed by a deep learning algorithm to establish a test process prediction model; the test process prediction model models the correlation between test parameters, equipment status, and quality indicators through a recurrent neural network to achieve real-time prediction of test process indicators; the real-time test data is compared and analyzed with the prediction results of the test process prediction model to identify abnormal situations and optimization opportunities in the test process;

[0012] According to the comparative analysis results, a test process optimization plan is generated through a dynamic optimization decision module, and the test process optimization plan includes test parameter adjustment suggestions, equipment maintenance plans, quality control strategies and resource allocation plans; the test process optimization plan is sent to the production and manufacturing site through the test process management system, and the test parameters, equipment operating status and resource allocation are adjusted in real time; the implementation effect of the optimization plan is monitored through the feedback control module, and the test process digital model and the test process prediction model are continuously updated to achieve closed-loop optimization management of the test process.

[0013] The test process digital twin system is constructed based on the test process digital model by adopting a hierarchical architecture design, including:

[0014] A test process digital twin system is designed using a layered architecture, the layered architecture includes a physical layer, a data layer, a model layer and an application layer; a high-precision sensor network is deployed through the physical layer to collect vibration data, temperature data and pressure data of the test equipment, and processing parameter data and tool status data are collected through the numerical control system interface of the test equipment, and visual information data of the test process is collected through an industrial-grade camera system, and the vibration data, the temperature data, the pressure data, the processing parameter data, the tool status data and the visual information data are transmitted to the data layer in real time through industrial Ethernet;

[0015] The vibration data, the temperature data, the pressure data, the processing parameter data, the tool status data and the visual information data are cleaned and processed by wavelet transform filtering through the data management platform of the data layer, the format and dimension of the processed data are unified by a data standardization method, a standardized data model is established, and the standardized data are spatiotemporally aligned and correlated by a data fusion algorithm to form a test process data chain, which is then input into the model layer;

[0016] The dynamic characteristics of the test equipment are modeled by the long short-term memory network through the model layer to obtain the equipment state prediction model, the test process association model is constructed by using the graph neural network, and the test parameter optimization model is established by using the reinforcement learning algorithm. The output data of the equipment state prediction model, the output data of the test process association model and the output data of the test parameter optimization model are input into the application layer; based on the application layer, the three-dimensional visualization technology is used to monitor and dynamically display the test process in real time, and the operation data of the physical test system and the digital twin model are compared in real time through the model adaptive update mechanism, the model prediction error is calculated, and the model parameters are dynamically adjusted based on the model prediction error.

[0017] The test process prediction model models the relationship between test parameters, equipment status, and quality indicators through a recurrent neural network to achieve real-time prediction of test process indicators, including:

[0018] The test parameters include processing speed parameters, feed rate parameters, and cutting depth parameters; the equipment status includes spindle vibration parameters, temperature change parameters, and power consumption parameters; the quality indicators include surface roughness parameters, dimensional accuracy parameters, and processing efficiency parameters; the test parameters, the equipment status, and the quality indicators are organized in time series to form a dynamic feature sequence of the test process;

[0019] A bidirectional long short-term memory network structure is used to perform time series modeling on the dynamic feature sequence of the test process. The network input layer of the bidirectional long short-term memory network structure receives the feature sequence composed of the test parameters and the device state. The time series dependency of parameter changes is captured through a forward long short-term memory layer and a backward long short-term memory layer. The forward long short-term memory layer and the backward long short-term memory layer include an input gating unit, a forgetting gating unit, and an output gating unit. The outputs of the forward long short-term memory layer and the backward long short-term memory layer are concatenated and mapped to the quality indicator feature space through a fully connected layer.

[0020] A multi-head self-attention module is introduced into the bidirectional long short-term memory network structure, and the correlation score between each time step in the feature sequence is calculated by the multi-head self-attention module to determine the weight coefficient of each time step feature; a training set is constructed using historical test data, and the training set is input into the bidirectional long short-term memory network structure, and the mean square error is used as the loss function. The network parameters are optimized by the back propagation algorithm, and the Dropout regularization method and the learning rate decay strategy are used to improve the model convergence performance;

[0021] The test parameters and the device status are collected in real time for the test process, a prediction input sequence is constructed through a sliding time window, and the prediction input sequence is input into the trained bidirectional long short-term memory network structure to achieve real-time prediction of the test process indicators.

[0022] According to the comparative analysis results, a test process optimization plan is generated through a dynamic optimization decision module. The test process optimization plan includes test parameter adjustment suggestions, equipment maintenance plans, and resource allocation plans, including:

[0023] According to the test parameters, the equipment status and the quality indicators, a multidimensional evaluation index system is constructed by using the hierarchical analysis method, and a test process performance evaluation score is obtained by combining the fuzzy evaluation algorithm. According to the test process performance evaluation score, the test process is modeled as a Markov decision process, and a state space including the test parameter status, the equipment status and the quality indicator status is constructed, and an action space including the process parameter action, the maintenance strategy action and the resource allocation action is constructed;

[0024] A dual deep Q network algorithm is used to train a decision model, wherein the decision model includes an experience replay module and a target network update module, the state space data and the action space data are input into the experience replay module for data storage, the decision strategy is optimized through the target network update module, and a response surface model of test parameters and quality indicators is established according to the decision strategy;

[0025] A particle swarm optimization algorithm with adaptive inertia weight and adaptive learning factor is adopted. Based on the response surface model, the optimal parameter combination is searched under the conditions of equipment capacity constraint, process limitation constraint and cost constraint. The optimal parameter combination is input into the equipment failure probability prediction model constructed by Bayesian network. The dynamic programming algorithm is adopted to generate equipment maintenance time plan and inventory replenishment strategy in combination with maintenance cost data and downtime loss data.

[0026] A dual deep Q network algorithm is used to train a decision model, the decision model includes an experience replay module and a target network update module, the state space data and the action space data are input into the experience replay module for data storage, the decision strategy is optimized by the target network update module, and a response surface model of test parameters and quality indicators is established according to the decision strategy, including:

[0027] Constructing a five-layer deep neural network structure of a main network and a target network, wherein the input layer of the main network and the target network receives a twelve-dimensional feature vector of a state space, wherein the twelve-dimensional feature vector includes a test speed parameter, a feed rate parameter, a cutting depth parameter, a spindle vibration parameter, a temperature change parameter, and a power consumption parameter; the first hidden layer of the main network and the target network includes one hundred and twenty-eight neurons and uses a ReLU activation function; the second hidden layer of the main network and the target network includes sixty-four neurons and a batch normalization layer is set; the third hidden layer of the main network and the target network includes thirty-two neurons and a Dropout layer is set; the output layer of the main network and the target network generates an eight-dimensional action vector including a process parameter adjustment amount, a maintenance time interval, and a resource allocation ratio;

[0028] The eight-dimensional action vector is input into a circular queue, the capacity of which is 10,000. The circular queue stores decision transfer samples including current state data, execution action data, reward data, next state data, and termination flag data. A greedy strategy with an initial value of 0.9 and a decay rate of 0.995 is used for action exploration. The decision transfer samples are obtained by random sampling with a batch size of 64 for target Q value calculation.

[0029] Based on the target Q value, the network parameters of the target network are updated by exponential sliding average, the optimal action predicted by the main network is input into the target network, and the target Q value of the target network is calculated by using the Adam optimizer with an initial learning rate of 0.001 and a cosine annealing strategy;

[0030] A multi-objective reward function including the improvement in quality dimension, the improvement in efficiency dimension and the improvement in cost dimension is established according to the target Q value; the test point distribution positions are determined by the central composite experimental design method based on the calculation result of the multi-objective reward function; the test parameter data and the quality index data are recorded at the test point distribution positions; the test parameter data and the quality index data are input into a least squares model; the least squares model is used to fit a quadratic polynomial response surface equation; the determination coefficient of the model goodness of fit is calculated by a cross-validation method; when the determination coefficient is lower than 0.9, sampling points are added to re-establish the quadratic polynomial response surface equation.

[0031] The optimal parameter combination is input into the equipment failure probability prediction model constructed by the Bayesian network, and the equipment maintenance time plan and inventory replenishment strategy are generated by using a dynamic programming algorithm in combination with the maintenance cost data and the downtime loss data, including:

[0032] Constructing a multi-layer Bayesian network structure, the multi-layer Bayesian network structure includes a root layer node group, an intermediate layer node group, and a leaf node group, the root layer node group includes a feed speed parameter node, a spindle speed parameter node, and a cutting depth parameter node, the intermediate layer node group includes a vibration intensity parameter node, a temperature change parameter node, a power fluctuation parameter node, and a noise level parameter node, the leaf node group includes a spindle failure parameter node, a transmission system fault parameter node, and a control system abnormality parameter node, and an expectation maximization algorithm is used to obtain network parameters of the multi-layer Bayesian network structure;

[0033] The forward reasoning method is used to calculate the edge probability distribution data of each node in the leaf node group, the Gibbs sampling method is used to process the incomplete observation data in the multi-layer Bayesian network structure, the fault probability prediction interval data is generated by the Monte Carlo simulation method, and a fault probability dynamic prediction model based on the time window mechanism is established;

[0034] Establish preventive maintenance cost data and fault maintenance cost data based on historical maintenance records. The preventive maintenance cost data includes labor cost data, spare parts replacement cost data, and planned downtime loss data. The fault maintenance cost data includes repair cost data, emergency spare parts procurement cost data, and unplanned downtime loss data. Use regression analysis methods to establish a mapping model between each cost data and equipment status data and maintenance type data.

[0035] The output result of the mapping model is input into a finite-time Markov decision model, wherein the state space of the finite-time Markov decision model includes equipment health status data and spare parts inventory level data, and the action space of the finite-time Markov decision model includes maintenance timing data and spare parts replenishment data; the planning cycle is divided into multiple decision stages, and the state transition probabilities and immediate costs of different state-action pairs are calculated in each decision stage, and the Bellman equation is solved by the reverse recursive method to obtain the optimal equipment maintenance time plan and inventory replenishment strategy.

[0036] According to a second aspect of the embodiments of the present invention,

[0037] Provides a digital twin and dynamic optimization management system for test processes in production and manufacturing, including:

[0038] The first unit is used to collect real-time test data in the production and manufacturing test process through a multi-source data acquisition device, the multi-source data acquisition device includes a production line sensor network, a test equipment data interface and a quality inspection system, and the collected real-time test data includes workpiece processing parameters, equipment operation status data, environmental monitoring data and test process quality data; the real-time test data is cleaned and standardized through a data preprocessing module to establish a test process digital model, and the test process digital model includes a test process model, an equipment status model, a quality control model and a resource scheduling model;

[0039] The second unit is used to design and construct a test process digital twin system based on the test process digital model using a layered architecture, wherein the layered architecture includes a physical layer, a data layer, a model layer, and an application layer; analyze historical test data through a deep learning algorithm to establish a test process prediction model; the test process prediction model models the correlation between test parameters, equipment status, and quality indicators through a recurrent neural network to achieve real-time prediction of test process indicators; compare and analyze the real-time test data with the prediction results of the test process prediction model to identify abnormal situations and optimization opportunities in the test process;

[0040] The third unit is used to generate a test process optimization plan through a dynamic optimization decision module according to the comparative analysis results, and the test process optimization plan includes test parameter adjustment suggestions, equipment maintenance plans, quality control strategies and resource allocation plans; the test process optimization plan is sent to the production and manufacturing site through the test process management system, and the test parameters, equipment operating status and resource allocation are adjusted in real time; the implementation effect of the optimization plan is monitored through the feedback control module, and the test process digital model and the test process prediction model are continuously updated to realize closed-loop optimization management of the test process.

[0041] According to a third aspect of the embodiments of the present invention,

[0042] An electronic device is provided, comprising:

[0043] processor;

[0044] a memory for storing processor-executable instructions;

[0045] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0046] A fourth aspect of the embodiments of the present invention is:

[0047] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0048] The beneficial effects of this application are as follows:

[0049] 1. Improve the digitalization and intelligence level of the testing process:

[0050] Through real-time collection of test data by multi-source data acquisition devices, the digital model of the test process and the digital twin system are established, realizing the comprehensive digital expression of the test process. Combining deep learning algorithms to build a test process prediction model can make real-time predictions of test parameters, equipment status and quality indicators, greatly improving the intelligence level and prediction ability of the test process.

[0051] 2. Optimize test process management efficiency and quality:

[0052] Based on the digital twin system and prediction model, abnormal situations and optimization opportunities in the test process can be quickly identified. The optimization plan is generated through the dynamic optimization decision module and sent to the production site in real time, realizing the timely adjustment of test parameters, equipment operation and resource allocation. This closed-loop optimization management method significantly improves the management efficiency and quality control level of the test process.

[0053] 3. Enhance the adaptability and reliability of the manufacturing process:

[0054] This method continuously monitors the implementation effect of the optimization plan through the feedback control module, and continuously updates the digital model and prediction model, forming an adaptive optimization closed-loop system. This mechanism enables the test process to quickly adapt to changes in the production environment, enhances the adaptability and reliability of the entire manufacturing process, and helps improve product quality and production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of a test process digital twin and dynamic optimization management method for production and manufacturing according to an embodiment of the present invention;

[0056] Figure 2 This is a structural diagram of a digital twin and dynamic optimization management system for a test process for production and manufacturing according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0059] Figure 1 Schematic diagram of the process of digital twin and dynamic optimization management method of test process for production and manufacturing according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0060] S101. Collect real-time test data in the production and manufacturing test process through a multi-source data acquisition device, the multi-source data acquisition device includes a production line sensor network, a test equipment data interface and a quality inspection system, and the collected real-time test data includes workpiece processing parameters, equipment operation status data, environmental monitoring data and test process quality data; clean and standardize the real-time test data through a data preprocessing module, and establish a test process digital model, the test process digital model includes a test process model, an equipment status model, a quality control model and a resource scheduling model;

[0061] S102. Based on the digital model of the test process, a test process digital twin system is constructed by adopting a hierarchical architecture design, wherein the hierarchical architecture includes a physical layer, a data layer, a model layer, and an application layer; historical test data is analyzed by a deep learning algorithm to establish a test process prediction model; the test process prediction model models the correlation between test parameters, equipment status, and quality indicators through a recurrent neural network to achieve real-time prediction of test process indicators; the real-time test data is compared and analyzed with the prediction results of the test process prediction model to identify abnormal situations and optimization opportunities in the test process;

[0062] S103. Based on the comparative analysis results, a test process optimization plan is generated through a dynamic optimization decision module, and the test process optimization plan includes test parameter adjustment suggestions, equipment maintenance plans, quality control strategies and resource allocation plans; the test process optimization plan is sent to the production and manufacturing site through the test process management system, and the test parameters, equipment operating status and resource allocation are adjusted in real time; the implementation effect of the optimization plan is monitored through the feedback control module, and the test process digital model and the test process prediction model are continuously updated to achieve closed-loop optimization management of the test process.

[0063] The specific implementation methods of the test process digital twin and dynamic optimization management method for production and manufacturing are as follows:

[0064] First, real-time test data in the production and manufacturing test process is collected through a multi-source data acquisition device. The multi-source data acquisition device includes a production line sensor network, a test equipment data interface, and a quality inspection system. The production line sensor network consists of a variety of sensors distributed in various stations of the production line, such as temperature sensors, pressure sensors, vibration sensors, etc., which are used to monitor various parameters during the workpiece processing process in real time. The test equipment data interface is directly connected to various test equipment through standardized protocols (such as OPC UA, Modbus, etc.) to collect equipment operation status data. The quality inspection system includes online inspection equipment and offline inspection stations to collect quality-related data during the test process.

[0065] The collected real-time test data include workpiece processing parameters (such as processing temperature, pressure, speed, etc.), equipment operation status data (such as equipment startup time, energy consumption, vibration frequency, etc.), environmental monitoring data (such as workshop temperature and humidity, dust concentration, etc.) and test process quality data (such as dimensional accuracy, surface roughness, material properties, etc.). For example, for an automobile engine cylinder processing test line, the collected data may include: cylinder rough processing temperature 85°C, spindle speed 6000rpm, equipment energy consumption 2.5kWh / piece, workshop temperature 26°C, cylinder diameter accuracy ±0.01mm, etc.

[0066] Next, the collected real-time test data is cleaned and standardized through the data preprocessing module. The data cleaning process includes removing outliers, filling missing values, and eliminating duplicate data. Standardization unifies data from different sources and dimensions to the same scale. For example, temperature data collected by different devices is uniformly converted to degrees Celsius, and time data is unified into a standard time format.

[0067] Based on the preprocessed data, a digital model of the test process is established. The digital model of the test process includes a test process model, an equipment status model, a quality control model, and a resource scheduling model. The test process model describes the various processes of the test process and their sequence, key parameters, and other information. The equipment status model includes the operating parameters, maintenance records, and fault warning indicators of each test equipment. The quality control model defines various quality inspection indicators and their control limits. The resource scheduling model includes the configuration information of resources such as manpower, materials, and energy.

[0068] Taking the test process model as an example, BPMN (Business Process Model and Notation) can be used to describe the test process. An engine block test process model may include the following steps: rough machining → fine machining → cleaning → dimension inspection → performance test → final inspection. Each step defines input and output parameters, execution conditions, quality control points and other information.

[0069] Next, a test process digital twin system is constructed based on the test process digital model using a layered architecture design. The layered architecture includes the physical layer, data layer, model layer, and application layer. The physical layer includes actual production and manufacturing equipment, sensor networks, and other hardware facilities. The data layer is responsible for data collection, storage, processing, and management. The model layer contains various mathematical models and algorithms, such as test process prediction models and optimization algorithms. The application layer provides various user-oriented functions and services, such as visualization interfaces and decision support.

[0070] In the model layer, historical test data is analyzed through deep learning algorithms to establish a test process prediction model. Specifically, a recurrent neural network structure such as a long short-term memory network (LSTM) can be used to model the relationship between test parameters, equipment status, and quality indicators. The input of the model includes various parameters in the historical test data, and the output is the predicted test process indicators for a period of time in the future.

[0071] For example, for an engine cylinder test line, a prediction model can be established to predict cylinder diameter accuracy. The model input includes the processing parameters of the previous process (such as roughing temperature, finishing speed, etc.), equipment status parameters (such as spindle vibration frequency, tool wear degree, etc.) and environmental parameters (such as workshop temperature, humidity, etc.). The model output is the predicted cylinder diameter accuracy value. By training on a large amount of historical data, the model can learn the influence of various parameters on cylinder diameter accuracy, thereby achieving accurate prediction of future cylinder diameter accuracy.

[0072] Comparing and analyzing real-time test data with the prediction results of the test process prediction model can identify anomalies and optimization opportunities in the test process. For example, if the actual measured cylinder diameter accuracy deviates significantly from the predicted value, it may indicate an abnormality in the processing process or a decline in equipment performance. By analyzing the causes of the deviation, potential optimization opportunities can be discovered, such as adjusting processing parameters and optimizing equipment maintenance strategies.

[0073] Based on the comparative analysis results, the dynamic optimization decision module generates a test process optimization plan. The optimization plan includes test parameter adjustment suggestions, equipment maintenance plans, quality control strategies, and resource allocation plans. For example, for the cylinder bore accuracy deviation problem, the optimization plan may include: adjusting the finishing speed from 6000rpm to 5800rpm, shortening the spindle maintenance cycle from 500 hours to 400 hours, and increasing the frequency of online dimension detection.

[0074] The optimization plan is sent to the manufacturing site through the test process management system to adjust the test parameters, equipment operation status and resource configuration in real time. The test process management system can adopt a distributed architecture and communicate with the on-site equipment and control system in real time through industrial Ethernet or 5G network. After receiving the optimization plan, the system automatically sends the new parameter settings to the relevant equipment and control system to realize dynamic adjustment of the test process.

[0075] Finally, the feedback control module is used to monitor the effect of the optimization scheme. The optimized test data is continuously collected and compared with the data before optimization to evaluate the optimization effect. According to the actual effect, the test process digital model and the test process prediction model are continuously updated to achieve closed-loop optimization management of the test process. For example, if it is found that the cylinder diameter accuracy is improved after adjusting the finishing speed, this experience is fed back to the digital model to guide future optimization decisions.

[0076] The beneficial effects of this method are mainly reflected in the following three aspects:

[0077] 1. Improve the intelligence and automation level of the test process: By building a digital twin system, the comprehensive digital expression and real-time monitoring of the test process are realized. The prediction model supported by the deep learning algorithm can accurately predict the test indicators and detect abnormal situations in time, which greatly improves the intelligence level of the test process.

[0078] 2. Dynamic optimization and continuous improvement of the test process: Based on digital twins and predictive models, the system can automatically generate optimization plans and adjust test parameters and resource configurations in real time. Through a closed-loop feedback mechanism, models and optimization strategies are continuously updated to achieve continuous optimization and improvement of the test process.

[0079] 3. Significantly improve product quality and production efficiency: Accurate prediction and timely optimization effectively reduce quality fluctuations and abnormalities during the test process. Dynamically optimized resource allocation and parameter adjustment improve test efficiency and reduce energy consumption and costs. Overall, this method can significantly improve the stability and consistency of product quality while improving production efficiency and resource utilization.

[0080] In an optional implementation, the test process digital model adopts a hierarchical architecture design to construct a test process digital twin system, including:

[0081] A test process digital twin system is designed using a layered architecture, the layered architecture includes a physical layer, a data layer, a model layer and an application layer; a high-precision sensor network is deployed through the physical layer to collect vibration data, temperature data and pressure data of the test equipment, and processing parameter data and tool status data are collected through the numerical control system interface of the test equipment, and visual information data of the test process is collected through an industrial-grade camera system, and the vibration data, the temperature data, the pressure data, the processing parameter data, the tool status data and the visual information data are transmitted to the data layer in real time through industrial Ethernet;

[0082] The vibration data, the temperature data, the pressure data, the processing parameter data, the tool status data and the visual information data are cleaned and processed by wavelet transform filtering through the data management platform of the data layer, the format and dimension of the processed data are unified by a data standardization method, a standardized data model is established, and the standardized data are spatiotemporally aligned and correlated by a data fusion algorithm to form a test process data chain, which is then input into the model layer;

[0083] The dynamic characteristics of the test equipment are modeled by the long short-term memory network through the model layer to obtain the equipment state prediction model, the test process association model is constructed by using the graph neural network, and the test parameter optimization model is established by using the reinforcement learning algorithm. The output data of the equipment state prediction model, the output data of the test process association model and the output data of the test parameter optimization model are input into the application layer; based on the application layer, the three-dimensional visualization technology is used to monitor and dynamically display the test process in real time, and the operation data of the physical test system and the digital twin model are compared in real time through the model adaptive update mechanism, the model prediction error is calculated, and the model parameters are dynamically adjusted based on the model prediction error.

[0084] The present invention provides a method for constructing a test process digital twin system based on a hierarchical architecture design, and the specific implementation method is as follows:

[0085] Firstly, a layered architecture is adopted to design the digital twin system of the test process, which includes the physical layer, data layer, model layer and application layer.

[0086] At the physical layer, a high-precision sensor network is deployed to collect vibration data, temperature data, and pressure data of the test equipment. Specifically, acceleration sensors can be installed at key locations of the test equipment to collect vibration data, thermocouples can be installed to collect temperature data, and pressure sensors can be installed to collect pressure data. For example, a three-axis acceleration sensor is installed at the spindle of a machine tool, with a sampling frequency set to 10kHz; a K-type thermocouple is installed near the tool, with a sampling frequency of 1Hz; and a pressure sensor is installed at key nodes of the hydraulic system, with a sampling frequency of 100Hz.

[0087] At the same time, the processing parameter data and tool status data are collected through the CNC system interface of the test equipment. For example, the processing parameters such as spindle speed, feed speed, cutting depth, and tool wear, remaining tool life and other status data are read from the CNC system, with a sampling frequency of 10Hz.

[0088] In addition, the visual information data of the test process is collected through an industrial-grade camera system. Multiple industrial cameras can be installed in the test area with a resolution of no less than 1920×1080 and a frame rate of no less than 30fps to collect real-time images and video data during the test process.

[0089] The collected vibration data, temperature data, pressure data, processing parameter data, tool status data and visual information data are transmitted to the data layer in real time via industrial Ethernet. Industrial Ethernet uses a transmission rate of 1000Mbps to ensure the real-time nature of the data.

[0090] At the data layer, the data management platform processes all kinds of data received. First, data cleaning is performed to remove abnormal values ​​and redundant data. For example, acceleration values ​​exceeding ±100g in vibration data are removed; values ​​below -50℃ or above 1000℃ in temperature data are removed. Then, wavelet transform is used for filtering, and db4 wavelet is selected to perform 5-layer decomposition of the signal to remove high-frequency noise.

[0091] Next, the data standardization method is used to unify the format and dimension of the processed data. The vibration data is Fourier transformed to extract the frequency domain features; the temperature data and pressure data are normalized and mapped to the 0-1 interval; the processing parameters and tool status data are discretized. A standardized data model is established to store all types of data in a unified JSON format.

[0092] Then, the standardized data is aligned in time and space and analyzed for association through the data fusion algorithm to form a test process data chain. The dynamic time warping algorithm is used for time alignment, and the spatial registration algorithm is used for spatial alignment. The association relationship between different data is analyzed through the association rule mining algorithm to generate a test process data chain. The test process data chain is input into the model layer.

[0093] At the model level, the long short-term memory network (LSTM) is first used to model the dynamic characteristics of the test equipment and obtain the equipment state prediction model. The LSTM network contains 1 input layer, 2 hidden layers and 1 output layer, and the number of neurons in the hidden layer is 128. The training data set contains 10,000 sets of historical data, and the Adam optimizer is used with a learning rate of 0.001 and 100 epochs for training.

[0094] Next, a graph neural network is used to construct a test process association model. The test process is represented as a graph structure, each process is a node, and the relationship between processes is an edge. A graph convolutional neural network is used, which contains 3 graph convolution layers, the activation function is ReLU, and the output is the process association matrix.

[0095] Then, the reinforcement learning algorithm is used to establish the test parameter optimization model. The state space is defined as the current test parameters, the action space is the parameter adjustment amount, and the reward function is the test quality score. The Deep Q-Network algorithm is used, which includes 1 input layer, 3 fully connected hidden layers and 1 output layer, and the output is the optimal test parameters.

[0096] The output data of the equipment status prediction model, the output data of the test process association model and the output data of the test parameter optimization model are input to the application layer.

[0097] At the application layer, the test process is monitored and displayed dynamically in real time based on 3D visualization technology. WebGL technology is used to build a 3D model of the test equipment, and the equipment status is updated in real time based on the output data of the equipment status prediction model. The output data of the test process association model is used to generate a process association diagram to intuitively display the test process. The output data of the test parameter optimization model is fed back to the test system in real time.

[0098] At the same time, the operation data of the physical test system and the digital twin model are compared in real time through the model adaptive update mechanism to calculate the model prediction error. When the model prediction error exceeds the preset threshold, the model parameters are dynamically adjusted. The gradient descent method is used to fine-tune the parameters of each layer of the model with the model prediction error as the loss function to achieve the adaptive update of the model.

[0099] The beneficial effects of the present invention include the following three aspects:

[0100] 1. By adopting a layered architecture design, multi-source heterogeneous collection, multi-dimensional fusion analysis, and multi-model collaborative optimization of test process data are realized, which improves the overall performance and scalability of the digital twin system.

[0101] 2. Based on artificial intelligence technologies such as deep learning and reinforcement learning, core models such as equipment status prediction, process correlation analysis and parameter optimization have been constructed, which has improved the intelligence level and prediction accuracy of the digital twin system.

[0102] 3. Through three-dimensional visualization and model adaptive update mechanism, real-time monitoring and dynamic optimization of the test process are achieved, enhancing the interaction and consistency between the digital twin system and the physical system.

[0103] In an optional implementation, the test process prediction model models the correlation between test parameters, equipment status, and quality indicators through a recurrent neural network to achieve real-time prediction of test process indicators, including:

[0104] The test parameters include processing speed parameters, feed rate parameters, and cutting depth parameters; the equipment status includes spindle vibration parameters, temperature change parameters, and power consumption parameters; the quality indicators include surface roughness parameters, dimensional accuracy parameters, and processing efficiency parameters; the test parameters, the equipment status, and the quality indicators are organized in time series to form a dynamic feature sequence of the test process;

[0105] A bidirectional long short-term memory network structure is used to perform time series modeling on the dynamic feature sequence of the test process. The network input layer of the bidirectional long short-term memory network structure receives the feature sequence composed of the test parameters and the device state. The time series dependency of parameter changes is captured through a forward long short-term memory layer and a backward long short-term memory layer. The forward long short-term memory layer and the backward long short-term memory layer include an input gating unit, a forgetting gating unit, and an output gating unit. The outputs of the forward long short-term memory layer and the backward long short-term memory layer are concatenated and mapped to the quality indicator feature space through a fully connected layer.

[0106] A multi-head self-attention module is introduced into the bidirectional long short-term memory network structure, and the correlation score between each time step in the feature sequence is calculated by the multi-head self-attention module to determine the weight coefficient of each time step feature; a training set is constructed using historical test data, and the training set is input into the bidirectional long short-term memory network structure, and the mean square error is used as the loss function. The network parameters are optimized by the back propagation algorithm, and the Dropout regularization method and the learning rate decay strategy are used to improve the model convergence performance;

[0107] The test parameters and the device status are collected in real time for the test process, a prediction input sequence is constructed through a sliding time window, and the prediction input sequence is input into the trained bidirectional long short-term memory network structure to achieve real-time prediction of the test process indicators.

[0108] This embodiment provides a method for constructing and using a test process prediction model, which specifically includes the following steps:

[0109] First, collect and organize the parameter data related to the test process. Test parameters include processing speed, feed rate and cutting depth, etc.; equipment status includes spindle vibration, temperature change and power consumption, etc.; quality indicators include surface roughness, dimensional accuracy and processing efficiency, etc. Organize these parameters in chronological order to form a dynamic feature sequence of the test process. For example, data can be collected every 1 second to record the real-time values ​​of various parameters to form a multi-dimensional time series.

[0110] Next, the Bidirectional Long Short-Term Memory (B i LSTM) network structure is designed and constructed. The input layer of the network receives the feature sequence composed of test parameters and device status. The network contains two LSTM layers, forward and backward, to capture the temporal dependencies of parameter changes. Each LSTM layer contains three gating units: input gate, forget gate, and output gate, to control the flow of information. After the outputs of the forward and backward LSTM layers are concatenated, they are mapped to the feature space of the quality indicator through the fully connected layer. For example, the number of hidden units in the LSTM layer can be set to 128, and the number of neurons in the fully connected layer can be set to 64.

[0111] In order to further improve the performance of the model, a multi-head self-attention module is introduced into the Bi LSTM structure. This module calculates the correlation score between each time step in the feature sequence and determines the weight coefficient of each time step feature. Eight attention heads can be set, and the dimension of each head is 16. Through the self-attention mechanism, the model can better capture long-distance dependencies.

[0112] Then, the training set is constructed using historical test data. The collected historical data is divided into training set and validation set in a ratio of 80%:20%. The data is standardized so that its mean is 0 and its variance is 1. The mean square error is used as the loss function, and the network parameters are optimized by the back propagation algorithm. In order to improve the generalization ability and convergence performance of the model, the Dropout regularization method (the dropout rate is set to 0.5) and the learning rate decay strategy (the initial learning rate is 0.001, and it decays by 10% every 50 epochs) are used.

[0113] After the model training is completed, it can be used for real-time prediction of test process indicators. The test parameters and equipment status are collected in real time for the test process, and the prediction input sequence is constructed through a sliding time window (window size is 60 seconds, step size is 1 second). The sequence is input into the trained Bi LSTM network to obtain the predicted value of the quality indicator in the future (such as the next 10 seconds).

[0114] In practical applications, the processing parameters can be adjusted in time according to the prediction results. For example, when it is predicted that the surface roughness may exceed the threshold, the system will automatically reduce the processing speed or cutting depth. This predictive adjustment can effectively reduce the production of defective products and improve processing quality and efficiency.

[0115] The beneficial effects of this method are mainly reflected in three aspects:

[0116] First, the test process is modeled by the BiLSTM network, which can effectively capture the complex nonlinear relationship and time-series dependency between parameters, and improve the accuracy and stability of the prediction. Compared with the traditional statistical model, the prediction error can be reduced by more than 30%.

[0117] Secondly, the introduction of the multi-head self-attention mechanism enhances the model's ability to capture long-distance dependencies, making the prediction results more fully utilize historical information. This significantly improves the performance of the model when processing long sequence data, and the prediction time span can be extended by 50%.

[0118] Finally, the use of real-time prediction and dynamic adjustment makes the processing process more intelligent and adaptive. This not only improves the consistency of product quality, but also reduces the need for manual intervention, increases production efficiency by about 20%, and reduces energy consumption by about 15%.

[0119] In an optional implementation, the analysis results are compared and a test process optimization plan is generated through a dynamic optimization decision module. The test process optimization plan includes test parameter adjustment suggestions, equipment maintenance plans, and resource allocation plans including:

[0120] According to the test parameters, the equipment status and the quality indicators, a multidimensional evaluation index system is constructed by using the hierarchical analysis method, and a fuzzy evaluation algorithm is combined to obtain a test process performance evaluation score. According to the test process performance evaluation score, the test process is modeled as a Markov decision process, and a state space including the test parameter status, the equipment status and the quality indicator status is constructed, and an action space including the process parameter action, the maintenance strategy action and the resource allocation action is constructed;

[0121] A dual deep Q network algorithm is used to train a decision model, wherein the decision model includes an experience replay module and a target network update module, the state space data and the action space data are input into the experience replay module for data storage, the decision strategy is optimized through the target network update module, and a response surface model of test parameters and quality indicators is established according to the decision strategy;

[0122] A particle swarm optimization algorithm with adaptive inertia weight and adaptive learning factor is adopted. Based on the response surface model, the optimal parameter combination is searched under the conditions of equipment capacity constraint, process limitation constraint and cost constraint. The optimal parameter combination is input into the equipment failure probability prediction model constructed by Bayesian network. The dynamic programming algorithm is adopted to generate equipment maintenance time plan and inventory replenishment strategy in combination with maintenance cost data and downtime loss data.

[0123] According to the comparative analysis results, the specific implementation method of generating the test process optimization solution through the dynamic optimization decision module is as follows:

[0124] First, a multidimensional evaluation index system is constructed through the analytic hierarchy process based on test parameters, equipment status and quality indicators. Specifically, test parameters, equipment status and quality indicators can be used as first-level indicators, which can be further subdivided into multiple second-level and third-level indicators. For example, test parameters may include temperature, pressure, flow, etc.; equipment status may include equipment integrity rate, failure rate, maintenance cycle, etc.; quality indicators may include qualified rate, first-time pass rate, etc. Then, the expert scoring method is used to determine the weights of indicators at each level.

[0125] Next, combined with the fuzzy evaluation algorithm, each indicator is fuzzified and the membership degree is calculated to obtain the fuzzy evaluation matrix of the test process performance. Through fuzzy synthesis operation, the test process performance evaluation score is finally obtained. For example, suppose the final evaluation score is 85 points.

[0126] According to the test process performance evaluation score, the test process is modeled as a Markov decision process. The state space is constructed, including the test parameter state (such as high / medium / low temperature), equipment state (such as normal / minor fault / serious fault), quality indicator state (such as excellent / good / qualified / unqualified), etc. The action space is constructed, including process parameter actions (such as raising / lowering temperature), maintenance strategy actions (such as immediate repair / regular maintenance), resource allocation actions (such as increasing / reducing personnel), etc.

[0127] The decision model is trained using a dual deep Q network algorithm. The model includes an experience replay module and a target network update module. The historical data of the state space and action space are input into the experience replay module for storage. Through the target network update module, the ε-greedy strategy is used to explore new actions, and the Q value is updated through the TD error to optimize the decision strategy. After training enough rounds, a better decision model can be obtained.

[0128] According to the decision-making strategy obtained through training, a response surface model of test parameters and quality indicators is established. For example, a quadratic polynomial regression model can be used to fit the relationship between parameters such as temperature and pressure and the qualified rate.

[0129] Next, a particle swarm optimization algorithm with adaptive inertia weight and adaptive learning factor is used to search for the optimal parameter combination based on the response surface model. First, a certain number of particles are initialized, and each particle represents a set of possible parameter combinations. Then, iterative optimization is performed under the conditions of equipment capacity constraints (such as upper and lower temperature limits), process constraints (such as heating rate), and cost constraints (such as energy consumption limits). By continuously updating the position and velocity of the particles, it eventually converges to the optimal solution. For example, the optimal temperature may be 180°C and the pressure may be 5MPa.

[0130] The optimal parameter combination is input into the equipment failure probability prediction model constructed by the Bayesian network. Based on historical data, the model establishes a causal relationship network between the various components of the equipment and calculates the failure probability through the conditional probability table. For example, under the optimal parameters, the failure probability of key component A is 0.05, and the failure probability of component B is 0.03.

[0131] Finally, the dynamic programming algorithm is used to generate equipment maintenance time plan and inventory replenishment strategy by combining maintenance cost data (such as replacement part cost) and downtime loss data (such as downtime loss per hour). The algorithm weighs the immediate cost and future cost at each decision stage through backward recursion to obtain the optimal maintenance decision sequence. For example, it may be possible to obtain a maintenance plan that performs preventive maintenance on component A every 3 months and overhauls component B every 6 months.

[0132] This technical solution has the following beneficial effects:

[0133] 1. Through the multi-dimensional evaluation index system and fuzzy evaluation algorithm, the test process performance is comprehensively and objectively evaluated to provide a reliable basis for subsequent optimization.

[0134] 2. The deep reinforcement learning method can adaptively learn the optimal decision-making strategy in complex testing environments, improving the scientificity and effectiveness of the optimization plan.

[0135] 3. Combining the response surface model and intelligent optimization algorithm, the optimal parameter combination is searched under multiple constraints, while considering equipment reliability and maintenance strategy, thus achieving global dynamic optimization of the test process.

[0136] In an optional implementation, a dual deep Q network algorithm is used to train a decision model, wherein the decision model includes an experience replay module and a target network update module, wherein the state space data and the action space data are input into the experience replay module for data storage, and the decision strategy is optimized by the target network update module. Establishing a response surface model of test parameters and quality indicators according to the decision strategy includes:

[0137] Constructing a five-layer deep neural network structure of a main network and a target network, wherein the input layer of the main network and the target network receives a twelve-dimensional feature vector of a state space, wherein the twelve-dimensional feature vector includes a test speed parameter, a feed rate parameter, a cutting depth parameter, a spindle vibration parameter, a temperature change parameter, and a power consumption parameter; the first hidden layer of the main network and the target network includes one hundred and twenty-eight neurons and uses a ReLU activation function; the second hidden layer of the main network and the target network includes sixty-four neurons and a batch normalization layer is set; the third hidden layer of the main network and the target network includes thirty-two neurons and a Dropout layer is set; the output layer of the main network and the target network generates an eight-dimensional action vector including a process parameter adjustment amount, a maintenance time interval, and a resource allocation ratio;

[0138] The eight-dimensional action vector is input into a circular queue, the capacity of which is 10,000. The circular queue stores decision transfer samples including current state data, execution action data, reward data, next state data, and termination flag data. A greedy strategy with an initial value of 0.9 and a decay rate of 0.995 is used for action exploration. The decision transfer samples are obtained by random sampling with a batch size of 64 for target Q value calculation.

[0139] Based on the target Q value, the network parameters of the target network are updated by exponential sliding average, the optimal action predicted by the main network is input into the target network, and the target Q value of the target network is calculated by using the Adam optimizer with an initial learning rate of 0.001 and a cosine annealing strategy;

[0140] A multi-objective reward function including the improvement in quality dimension, the improvement in efficiency dimension and the improvement in cost dimension is established according to the target Q value; the test point distribution positions are determined by the central composite experimental design method based on the calculation result of the multi-objective reward function; the test parameter data and the quality index data are recorded at the test point distribution positions; the test parameter data and the quality index data are input into a least squares model; the least squares model is used to fit a quadratic polynomial response surface equation; the determination coefficient of the model goodness of fit is calculated by a cross-validation method; when the determination coefficient is lower than 0.9, sampling points are added to re-establish the quadratic polynomial response surface equation.

[0141] The decision model is trained by using a dual deep Q network algorithm, which includes an experience replay module and a target network update module. First, the state space data and action space data are input into the experience replay module for storage, and then the decision strategy is optimized through the target network update module. Finally, the response surface model of test parameters and quality indicators is established based on the optimized decision strategy.

[0142] The specific implementation process is as follows:

[0143] A five-layer deep neural network structure of the main network and the target network is constructed. The input layer of the two networks receives a 12-dimensional feature vector, including test speed, feed rate, cutting depth, spindle vibration, temperature change, power consumption and other parameters. The first hidden layer contains 128 neurons and uses the ReLU activation function. The second hidden layer has 64 neurons and a batch normalization layer is set. The third hidden layer contains 32 neurons and a Dropout layer is set. The output layer generates an 8-dimensional action vector, including the process parameter adjustment amount, maintenance time interval, and resource allocation ratio.

[0144] The 8-dimensional action vector is input into a circular queue with a capacity of 10,000. The queue stores decision transfer samples, including data such as the current state, executed action, reward, next state, and termination flag. A greedy strategy is used for action exploration, with an initial value of 0.9 and a decay rate of 0.995. Decision transfer samples are obtained by random sampling with a batch size of 64 for target Q value calculation.

[0145] Update the target network parameters by exponential sliding average. Input the optimal action predicted by the main network into the target network, and use the Adam optimizer (initial learning rate 0.001) and cosine annealing strategy to calculate the target Q value of the target network.

[0146] A multi-objective reward function was established based on the target Q value, including the improvement in three dimensions: quality, efficiency, and cost. Based on the calculation results of the reward function, the central composite test design method was used to determine the distribution of test points. The test parameters and quality index data were recorded at the test points, and these data were input into the least squares model to fit the quadratic polynomial response surface equation. The determination coefficient of the model goodness of fit was calculated by cross-validation. When the determination coefficient was lower than 0.9, more sampling points were added to re-establish the response surface equation.

[0147] For example, suppose the 12-dimensional feature vector received by the input layer is: [100, 0.2, 2, 0.5, 25, 5, 80, 0.3, 1.5, 0.8, 30, 4]. After network processing, the 8-dimensional action vector generated by the output layer may be: [0.1, -0.05, 0.2, 30, 0.4, -0.1, 0.3, 20]. This action vector represents the adjustment amount of the process parameters, the recommended maintenance time interval and the resource allocation ratio.

[0148] The action vector is stored in a circular queue and used for training together with other decision transfer samples. Assuming that the batch data obtained by random sampling contains this sample, it will be used to calculate the target Q value. After the target Q value is calculated, the network parameters are updated through the Adam optimizer.

[0149] The possible calculation results of the multi-objective reward function are: [0.15, 0.08, 0.12], which represent the improvement in quality, efficiency, and cost. Based on this result, the distribution of test points is determined and data is collected. Assume that the data collected at a certain test point is: [105, 0.18, 2.2, 0.45, 26, 4.8, 82, 0.28, 1.6, 0.75, 31, 3.8] and the corresponding quality indicator data.

[0150] These data were input into the least squares model and the response surface equation was obtained by fitting. The coefficient of determination was calculated by cross-validation. If it was lower than 0.9, the equation was re-established by increasing the sampling points until the accuracy requirements were met.

[0151] The beneficial effects of this method include:

[0152] 1. Improved the learning efficiency and stability of the decision model. By using a dual deep Q network and experience replay mechanism, the correlation of target value estimation is reduced, and the stability and convergence speed of training are improved.

[0153] 2. Enhanced the generalization and robustness of the model. The use of batch normalization and Dropout techniques effectively prevents overfitting and improves the adaptability of the model under different working conditions.

[0154] 3. Multi-objective optimization and precise modeling have been achieved. Through the multi-objective reward function and response surface modeling method, multiple optimization goals such as quality, efficiency, and cost are taken into account, and the relationship between test parameters and quality indicators is accurately described, providing a reliable basis for subsequent optimization decisions.

[0155] In an optional implementation, the optimal parameter combination is input into the equipment failure probability prediction model constructed by the Bayesian network, and the equipment maintenance time plan and inventory replenishment strategy are generated by using a dynamic programming algorithm in combination with the maintenance cost data and the downtime loss data, including:

[0156] Constructing a multi-layer Bayesian network structure, the multi-layer Bayesian network structure includes a root layer node group, an intermediate layer node group, and a leaf node group, the root layer node group includes a feed speed parameter node, a spindle speed parameter node, and a cutting depth parameter node, the intermediate layer node group includes a vibration intensity parameter node, a temperature change parameter node, a power fluctuation parameter node, and a noise level parameter node, the leaf node group includes a spindle failure parameter node, a transmission system fault parameter node, and a control system abnormality parameter node, and an expectation maximization algorithm is used to obtain network parameters of the multi-layer Bayesian network structure;

[0157] The forward reasoning method is used to calculate the edge probability distribution data of each node in the leaf node group, the Gibbs sampling method is used to process the incomplete observation data in the multi-layer Bayesian network structure, the fault probability prediction interval data is generated by the Monte Carlo simulation method, and a fault probability dynamic prediction model based on the time window mechanism is established;

[0158] Establish preventive maintenance cost data and fault maintenance cost data based on historical maintenance records. The preventive maintenance cost data includes labor cost data, spare parts replacement cost data, and planned downtime loss data. The fault maintenance cost data includes repair cost data, emergency spare parts procurement cost data, and unplanned downtime loss data. Use regression analysis methods to establish a mapping model between each cost data and equipment status data and maintenance type data.

[0159] The output result of the mapping model is input into a finite-time Markov decision model, wherein the state space of the finite-time Markov decision model includes equipment health status data and spare parts inventory level data, and the action space of the finite-time Markov decision model includes maintenance timing data and spare parts replenishment data; the planning cycle is divided into multiple decision stages, and the state transition probabilities and immediate costs of different state-action pairs are calculated in each decision stage, and the Bellman equation is solved by the reverse recursive method to obtain the optimal equipment maintenance time plan and inventory replenishment strategy.

[0160] To implement the above technical solution, the following steps can be followed:

[0161] Firstly, a multi-layer Bayesian network structure is constructed, including a root node group, an intermediate node group and a leaf node group. The root node group contains feed speed parameter nodes, spindle speed parameter nodes and cutting depth parameter nodes, which are used to characterize machining process parameters. The intermediate node group contains vibration intensity parameter nodes, temperature change parameter nodes, power fluctuation parameter nodes and noise level parameter nodes, which are used to characterize equipment status parameters. The leaf node group contains spindle failure parameter nodes, transmission system fault parameter nodes and control system abnormality parameter nodes, which are used to characterize equipment fault types. The expectation maximization algorithm is used to obtain network parameters, which includes initializing network parameters, calculating expected values, maximizing likelihood functions, updating network parameters, and iterating until convergence.

[0162] Next, the forward reasoning method is used to calculate the edge probability distribution data of each node in the leaf node group. Specifically, starting from the root node, the probability information is transmitted along the directed edges, and the conditional probability distribution of the intermediate nodes and leaf nodes is calculated layer by layer, and finally the edge probability distribution of the leaf nodes is obtained. For the incomplete observation data in the network, the Gibbs sampling method is used to process it. By repeatedly sampling the conditional distribution of the missing variables, complete sample data is generated. Then, the Monte Carlo simulation method is used to sample the edge probability distribution of the leaf nodes multiple times to generate the fault probability prediction interval data. On this basis, a dynamic prediction model of fault probability based on the time window mechanism is established, and the prediction results are dynamically updated through the sliding time window.

[0163] Then, preventive maintenance cost data and fault maintenance cost data are established based on historical maintenance records. Preventive maintenance cost data includes labor cost data, spare parts replacement cost data, and planned downtime loss data. For example, the labor cost of preventive maintenance of a certain equipment is 1,000 yuan / time, the spare parts replacement cost is 5,000 yuan / time, and the planned downtime loss is 10,000 yuan / day. Fault maintenance cost data includes maintenance cost data, emergency spare parts procurement cost data, and unplanned downtime loss data. For example, the maintenance cost of a certain equipment failure is 10,000 yuan / time, the emergency spare parts procurement cost is 15,000 yuan / time, and the unplanned downtime loss is 50,000 yuan / day. The regression analysis method is used to establish a mapping model between each cost data and equipment status data and maintenance type data. Specifically, with equipment status indicators and maintenance types as independent variables and various costs as dependent variables, a multivariate linear regression model is established.

[0164] The output of the above mapping model is input into the finite time domain Markov decision model. The state space of the model contains equipment health status data and spare parts inventory level data. For example, the equipment health status is divided into three states: normal, sub-healthy, and faulty, and the spare parts inventory level is divided into three states: sufficient, general, and scarce. The action space contains maintenance timing data and spare parts replenishment data. For example, the maintenance timing includes three options: immediate maintenance, delayed maintenance, and no maintenance, and the spare parts replenishment includes three options: large replenishment, small replenishment, and no replenishment.

[0165] The planning cycle is divided into multiple decision stages, for example, a one-year planning cycle is divided into 12 months. In each decision stage, the state transition probability and immediate cost of different state-action pairs are calculated. The state transition probability is obtained based on historical data statistics, and the immediate cost is calculated based on the aforementioned cost mapping model. The Bellman equation is solved by the reverse recursive method, starting from the last decision stage, and the optimal decision and minimum expected cost of each state are calculated step by step. Finally, the optimal equipment maintenance time plan and inventory replenishment strategy for the entire planning cycle are obtained.

[0166] This technical solution has the following beneficial effects:

[0167] 1. By constructing a multi-layer Bayesian network model, the correlation between processing parameters, equipment status and fault type was established, which improved the accuracy and interpretability of fault prediction.

[0168] 2. Combined with the dynamic programming algorithm, the joint optimization of equipment maintenance and inventory replenishment is achieved, which reduces the overall operation and maintenance costs while ensuring equipment reliability.

[0169] 3. The dynamic prediction mechanism based on time window can timely capture the trend of equipment status changes, improving the real-time and adaptability of maintenance decisions.

[0170] Figure 2 Schematic diagram of the structure of the test process digital twin and dynamic optimization management system for production and manufacturing according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0171] The first unit is used to collect real-time test data in the production and manufacturing test process through a multi-source data acquisition device, the multi-source data acquisition device includes a production line sensor network, a test equipment data interface and a quality inspection system, and the collected real-time test data includes workpiece processing parameters, equipment operation status data, environmental monitoring data and test process quality data; the real-time test data is cleaned and standardized through a data preprocessing module to establish a test process digital model, and the test process digital model includes a test process model, an equipment status model, a quality control model and a resource scheduling model;

[0172] The second unit is used to design and construct a test process digital twin system based on the test process digital model using a layered architecture, wherein the layered architecture includes a physical layer, a data layer, a model layer, and an application layer; analyze historical test data through a deep learning algorithm to establish a test process prediction model; the test process prediction model models the correlation between test parameters, equipment status, and quality indicators through a recurrent neural network to achieve real-time prediction of test process indicators; compare and analyze the real-time test data with the prediction results of the test process prediction model to identify abnormal situations and optimization opportunities in the test process;

[0173] The third unit is used to generate a test process optimization plan through a dynamic optimization decision module according to the comparative analysis results, and the test process optimization plan includes test parameter adjustment suggestions, equipment maintenance plans, quality control strategies and resource allocation plans; the test process optimization plan is sent to the production and manufacturing site through the test process management system, and the test parameters, equipment operating status and resource allocation are adjusted in real time; the implementation effect of the optimization plan is monitored through the feedback control module, and the test process digital model and the test process prediction model are continuously updated to realize closed-loop optimization management of the test process.

[0174] According to a third aspect of the embodiments of the present invention,

[0175] An electronic device is provided, comprising:

[0176] processor;

[0177] a memory for storing processor-executable instructions;

[0178] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0179] A fourth aspect of the embodiments of the present invention is:

[0180] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0181] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0182] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A digital twin and dynamic optimization management method for test processes for production and manufacturing, characterized in that: include: Collecting real-time test data in the production and manufacturing test process through a multi-source data acquisition device, the multi-source data acquisition device includes a production line sensor network, a test equipment data interface and a quality inspection system, and the collected real-time test data includes workpiece processing parameters, equipment operation status data, environmental monitoring data and test process quality data; The real-time test data is cleaned and standardized through a data preprocessing module to establish a test process digital model, which includes a test process model, an equipment status model, a quality control model, and a resource scheduling model; A test process digital twin system is constructed based on the test process digital model by adopting a layered architecture design, wherein the layered architecture includes a physical layer, a data layer, a model layer, and an application layer; historical test data is analyzed by a deep learning algorithm to establish a test process prediction model; the test process prediction model models the correlation between test parameters, equipment status, and quality indicators through a recurrent neural network to achieve real-time prediction of test process indicators; Comparing and analyzing the real-time test data with the prediction results of the test process prediction model to identify abnormal situations and optimization opportunities in the test process; According to the comparative analysis results, a test process optimization plan is generated through a dynamic optimization decision module, and the test process optimization plan includes test parameter adjustment suggestions, equipment maintenance plans, quality control strategies and resource allocation plans; The test process optimization plan is sent to the production and manufacturing site through the test process management system, and the test parameters, equipment operating status and resource allocation are adjusted in real time. The implementation effect of the optimization plan is monitored through the feedback control module, and the test process digital model and the test process prediction model are continuously updated to realize closed-loop optimization management of the test process.

2. The method according to claim 1, characterized in that The test process digital twin system is constructed based on the test process digital model by adopting a hierarchical architecture design, including: A test process digital twin system is designed using a layered architecture, the layered architecture includes a physical layer, a data layer, a model layer and an application layer; a high-precision sensor network is deployed through the physical layer to collect vibration data, temperature data and pressure data of the test equipment, and processing parameter data and tool status data are collected through the numerical control system interface of the test equipment, and visual information data of the test process is collected through an industrial-grade camera system, and the vibration data, the temperature data, the pressure data, the processing parameter data, the tool status data and the visual information data are transmitted to the data layer in real time through industrial Ethernet; The vibration data, the temperature data, the pressure data, the processing parameter data, the tool status data and the visual information data are cleaned and processed by wavelet transform filtering through the data management platform of the data layer, the format and dimension of the processed data are unified by a data standardization method, a standardized data model is established, and the standardized data are spatiotemporally aligned and correlated by a data fusion algorithm to form a test process data chain, which is then input into the model layer; The dynamic characteristics of the test equipment are modeled by the long short-term memory network through the model layer to obtain the equipment state prediction model, the test process association model is constructed by using the graph neural network, and the test parameter optimization model is established by using the reinforcement learning algorithm. The output data of the equipment state prediction model, the output data of the test process association model and the output data of the test parameter optimization model are input into the application layer; based on the application layer, the three-dimensional visualization technology is used to monitor and dynamically display the test process in real time, and the operation data of the physical test system and the digital twin model are compared in real time through the model adaptive update mechanism, the model prediction error is calculated, and the model parameters are dynamically adjusted based on the model prediction error.

3. The method according to claim 1, characterized in that: The test process prediction model models the relationship between test parameters, equipment status, and quality indicators through a recurrent neural network to achieve real-time prediction of test process indicators, including: The test parameters include processing speed parameters, feed rate parameters, and cutting depth parameters; the equipment status includes spindle vibration parameters, temperature change parameters, and power consumption parameters; the quality indicators include surface roughness parameters, dimensional accuracy parameters, and processing efficiency parameters; the test parameters, the equipment status, and the quality indicators are organized in time series to form a dynamic feature sequence of the test process; A bidirectional long short-term memory network structure is used to perform time series modeling on the dynamic feature sequence of the test process. The network input layer of the bidirectional long short-term memory network structure receives the feature sequence composed of the test parameters and the device state. The time series dependency of parameter changes is captured through a forward long short-term memory layer and a backward long short-term memory layer. The forward long short-term memory layer and the backward long short-term memory layer include an input gating unit, a forgetting gating unit, and an output gating unit. The outputs of the forward long short-term memory layer and the backward long short-term memory layer are concatenated and mapped to the quality indicator feature space through a fully connected layer. A multi-head self-attention module is introduced into the bidirectional long short-term memory network structure, and the correlation score between each time step in the feature sequence is calculated by the multi-head self-attention module to determine the weight coefficient of each time step feature; a training set is constructed using historical test data, and the training set is input into the bidirectional long short-term memory network structure, and the mean square error is used as the loss function. The network parameters are optimized by the back propagation algorithm, and the Dropout regularization method and the learning rate decay strategy are used to improve the model convergence performance; The test parameters and the device status are collected in real time for the test process, a prediction input sequence is constructed through a sliding time window, and the prediction input sequence is input into the trained bidirectional long short-term memory network structure to achieve real-time prediction of the test process indicators.

4. The method according to claim 3, characterized in that According to the comparative analysis results, a test process optimization plan is generated through a dynamic optimization decision module. The test process optimization plan includes test parameter adjustment suggestions, equipment maintenance plans, and resource allocation plans, including: According to the test parameters, the equipment status and the quality indicators, a multidimensional evaluation index system is constructed by using the hierarchical analysis method, and a test process performance evaluation score is obtained by combining the fuzzy evaluation algorithm. According to the test process performance evaluation score, the test process is modeled as a Markov decision process, and a state space including the test parameter status, the equipment status and the quality indicator status is constructed, and an action space including the process parameter action, the maintenance strategy action and the resource allocation action is constructed; A dual deep Q network algorithm is used to train a decision model, wherein the decision model includes an experience replay module and a target network update module, the state space data and the action space data are input into the experience replay module for data storage, the decision strategy is optimized through the target network update module, and a response surface model of test parameters and quality indicators is established according to the decision strategy; A particle swarm optimization algorithm with adaptive inertia weight and adaptive learning factor is adopted. Based on the response surface model, the optimal parameter combination is searched under the conditions of equipment capacity constraint, process limitation constraint and cost constraint. The optimal parameter combination is input into the equipment failure probability prediction model constructed by Bayesian network. The dynamic programming algorithm is adopted to generate equipment maintenance time plan and inventory replenishment strategy in combination with maintenance cost data and downtime loss data.

5. The method according to claim 4, characterized in that A dual deep Q network algorithm is used to train a decision model, the decision model includes an experience replay module and a target network update module, the state space data and the action space data are input into the experience replay module for data storage, the decision strategy is optimized by the target network update module, and a response surface model of test parameters and quality indicators is established according to the decision strategy, including: Constructing a five-layer deep neural network structure of a main network and a target network, wherein the input layer of the main network and the target network receives a twelve-dimensional feature vector of a state space, wherein the twelve-dimensional feature vector includes a test speed parameter, a feed rate parameter, a cutting depth parameter, a spindle vibration parameter, a temperature change parameter, and a power consumption parameter; the first hidden layer of the main network and the target network includes one hundred and twenty-eight neurons and uses a ReLU activation function; the second hidden layer of the main network and the target network includes sixty-four neurons and a batch normalization layer is set; the third hidden layer of the main network and the target network includes thirty-two neurons and a Dropout layer is set; the output layer of the main network and the target network generates an eight-dimensional action vector including a process parameter adjustment amount, a maintenance time interval, and a resource allocation ratio; The eight-dimensional action vector is input into a circular queue, the capacity of which is 10,000. The circular queue stores decision transfer samples including current state data, execution action data, reward data, next state data, and termination flag data. A greedy strategy with an initial value of 0.9 and a decay rate of 0.995 is used for action exploration. The decision transfer samples are obtained by random sampling with a batch size of 64 for target Q value calculation. Based on the target Q value, the network parameters of the target network are updated by exponential sliding average, the optimal action predicted by the main network is input into the target network, and the target Q value of the target network is calculated by using the Adam optimizer with an initial learning rate of 0.001 and a cosine annealing strategy; A multi-objective reward function including the improvement in quality dimension, the improvement in efficiency dimension and the improvement in cost dimension is established according to the target Q value; the test point distribution positions are determined by the central composite experimental design method based on the calculation result of the multi-objective reward function; the test parameter data and the quality index data are recorded at the test point distribution positions; the test parameter data and the quality index data are input into a least squares model; the least squares model is used to fit a quadratic polynomial response surface equation; the determination coefficient of the model goodness of fit is calculated by a cross-validation method; when the determination coefficient is lower than 0.9, sampling points are added to re-establish the quadratic polynomial response surface equation.

6. The method according to claim 4, characterized in that The optimal parameter combination is input into the equipment failure probability prediction model constructed by the Bayesian network, and the equipment maintenance time plan and inventory replenishment strategy are generated by using a dynamic programming algorithm in combination with the maintenance cost data and the downtime loss data, including: Constructing a multi-layer Bayesian network structure, the multi-layer Bayesian network structure includes a root layer node group, an intermediate layer node group, and a leaf node group, the root layer node group includes a feed speed parameter node, a spindle speed parameter node, and a cutting depth parameter node, the intermediate layer node group includes a vibration intensity parameter node, a temperature change parameter node, a power fluctuation parameter node, and a noise level parameter node, the leaf node group includes a spindle failure parameter node, a transmission system fault parameter node, and a control system abnormality parameter node, and an expectation maximization algorithm is used to obtain network parameters of the multi-layer Bayesian network structure; The forward reasoning method is used to calculate the edge probability distribution data of each node in the leaf node group, the Gibbs sampling method is used to process the incomplete observation data in the multi-layer Bayesian network structure, the fault probability prediction interval data is generated by the Monte Carlo simulation method, and a fault probability dynamic prediction model based on the time window mechanism is established; Establish preventive maintenance cost data and fault maintenance cost data based on historical maintenance records. The preventive maintenance cost data includes labor cost data, spare parts replacement cost data, and planned downtime loss data. The fault maintenance cost data includes repair cost data, emergency spare parts procurement cost data, and unplanned downtime loss data. Use regression analysis methods to establish a mapping model between each cost data and equipment status data and maintenance type data. The output result of the mapping model is input into a finite-time Markov decision model, wherein the state space of the finite-time Markov decision model includes equipment health status data and spare parts inventory level data, and the action space of the finite-time Markov decision model includes maintenance timing data and spare parts replenishment data; the planning cycle is divided into multiple decision stages, and the state transition probabilities and immediate costs of different state-action pairs are calculated in each decision stage, and the Bellman equation is solved by the reverse recursive method to obtain the optimal equipment maintenance time plan and inventory replenishment strategy.

7. A test process digital twin and dynamic optimization management system for production and manufacturing, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to collect real-time test data in the production and manufacturing test process through a multi-source data acquisition device, wherein the multi-source data acquisition device includes a production line sensor network, a test equipment data interface, and a quality inspection system, and the collected real-time test data includes workpiece processing parameters, equipment operation status data, environmental monitoring data, and test process quality data; The real-time test data is cleaned and standardized through a data preprocessing module to establish a test process digital model, which includes a test process model, an equipment status model, a quality control model, and a resource scheduling model; The second unit is used to construct a test process digital twin system based on the test process digital model by adopting a layered architecture design, wherein the layered architecture includes a physical layer, a data layer, a model layer, and an application layer; analyze historical test data by a deep learning algorithm to establish a test process prediction model; the test process prediction model models the correlation between test parameters, equipment status, and quality indicators by a recurrent neural network to achieve real-time prediction of test process indicators; Comparing and analyzing the real-time test data with the prediction results of the test process prediction model to identify abnormal situations and optimization opportunities in the test process; The third unit is used to generate a test process optimization plan through a dynamic optimization decision module according to the comparative analysis results, wherein the test process optimization plan includes test parameter adjustment suggestions, equipment maintenance plans, quality control strategies, and resource allocation plans; The test process optimization plan is sent to the production and manufacturing site through the test process management system, and the test parameters, equipment operating status and resource allocation are adjusted in real time. The implementation effect of the optimization plan is monitored through the feedback control module, and the test process digital model and the test process prediction model are continuously updated to realize closed-loop optimization management of the test process.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images