Method for detecting and predicting equipment rule anomaly in digital twinning

Abstract

The invention provides an equipment rule anomaly detection and prediction method in digital twinning. The method comprises the following steps: modeling based on a simulation environment of mechanism and data coupling; defining an action space, a state space and a state transfer function; preprocessing the original data; setting abnormal rule judgment, and constructing a univariate time series data prediction model and a multivariate time series data prediction model according to the abnormal rule judgment; performing model evaluation on each model; packaging the models into an algorithm library; integrating the trained models into a digital twin platform algorithm library to carry out iterative deployment work on the models; and calling the model in a digital twin platform algorithm library, and timely and accurately positioning a key problem sending and processing scheme through real-time abnormal monitoring and prediction of the model on equipment. The real-time monitoring system is established based on digital twinning, data streams from all devices and all point positions of the continuous casting machine can be integrated, the device state can be tracked in real time, potential problems can be detected, and all aspects of the production process can be monitored.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment anomaly detection, and in particular to a method for detecting and predicting equipment rule anomalies in a digital twin. Background Art

[0002] In recent years, intelligent manufacturing has been demonstrated and applied in process industry production, and its importance has gradually gained popularity. In the era of intelligent manufacturing, the requirements for equipment reliability will be higher, and the requirements for equipment maintenance management will also increase accordingly. Manufacturing companies, equipment management information vendors, and predictive maintenance service vendors will all participate in the overall process of equipment maintenance management. In industry, equipment maintenance is a very important link that directly affects the operating time and work efficiency of the equipment. However, the remaining life of the equipment needs to be accurately predicted, and the faults need to be identified and handled in a timely manner to avoid downtime during the production process.

[0003] There are three prediction methods in the current existing technology, namely, fault maintenance, fault maintenance and predictive maintenance. Among them, the fault maintenance strategy will make the maintenance behavior have a lag, and the risk of equipment occurrence is very high. Preventive maintenance strategy, this predictive maintenance intervention will lead to a decrease in the use time of the equipment. Predictive maintenance is a maintenance based on the state, which is to conduct continuous online status monitoring and data analysis of the equipment, diagnose and predict the development trend of equipment failures, formulate predictive maintenance plans in advance and implement inspection and maintenance. The traditional predictive maintenance method is mainly based on physical models. The basic idea is to fit the health status or life of equipment and parts based on fixed physical formulas of physical properties. This method requires researchers to have a deep background knowledge of engineering physics, usually using complex physical formulas, relying on too many physical models and ignoring the laws of the data itself, with low prediction accuracy and low efficiency.

[0004] Digital twins make full use of physical models, sensors, operation history and other data, integrate multi-disciplinary, multi-physical quantity, multi-scale, and multi-probability simulation processes, complete mapping in virtual space, and thus reflect the entire life cycle of the corresponding physical equipment.

[0005] Patent document CN117371220A discloses a method for predicting faults based on digital twins, which relates to the field of fault prediction technology and includes the following steps S: Step S1: data acquisition, the device-side data transmission is transmitted to the gateway device through the communication interface or directly uploaded to the cloud, and the device management system on the cloud performs data analysis, processing and storage after receiving the data; Step S2: data protocol check and analysis, before receiving the data, check the legitimacy of the protocol used and parse the protocol.

[0006] However, the patent document CN117371220A builds a digital twin network of devices, which requires comprehensive consideration of the device configuration information and network working data, and is highly dependent on the accuracy and completeness of the data. If there are errors or omissions in the data, it may affect the results of fault diagnosis, which may be difficult to achieve in some areas.

[0007] This method combines the mechanism model with data-driven, and by establishing a clearer mechanism model, we can clearly understand the working principle and failure mechanism of the equipment, so as to locate the cause of the failure more accurately. It is also more flexible, and by adjusting the coupling strategy, it can flexibly respond to different failure situations and data characteristics.

[0008] Moreover, patent document CN117371220A cannot perform predictive maintenance, and the artificial intelligence module built into the diagnostic system cannot fully diagnose all faults. In this method, a complex expert system has been introduced into the abnormal rule configuration, which can improve the abnormal judgment of the equipment, and different models can be selected in the algorithm model library for verification, and predictive maintenance can be performed in advance. Summary of the invention

[0009] In view of the defects in the prior art, the purpose of the present invention is to provide a method for detecting and predicting equipment rule anomalies in digital twins.

[0010] A method for detecting and predicting device rule anomalies in a digital twin provided by the present invention includes:

[0011] Step S1: Modeling the simulation environment based on the coupling of mechanism and data to obtain a device simulation model;

[0012] Step S2: defining the action space, state space and state transfer function of the device simulation model, and constructing the device simulation model;

[0013] Step S3: obtaining raw data and performing raw data preprocessing;

[0014] Step S4: setting an abnormal rule determination, and constructing a univariate time series data prediction model and a multivariate time series data prediction model according to the abnormal rule determination;

[0015] Step S5: According to the abnormality determination rule, the preprocessed raw data is divided into univariate time series data and multivariate time series data, and the univariate time series data prediction model and the multivariate time series data prediction model are input accordingly to obtain the training results, and the model evaluation and optimization are performed on each model;

[0016] Step S6: Encapsulating the univariate time series data prediction model and the multivariate time series data prediction model into an algorithm library;

[0017] Step S7: Integrate the trained single-variable time series data prediction model, multi-variable time series data prediction model and equipment simulation model into the digital twin platform algorithm library, provide the functions of loading and calling the model through the interface, and complete the iterative deployment of the model;

[0018] Step S8: Call the model in the algorithm library of the digital twin platform, monitor and predict the real-time abnormalities of the equipment through the single-variable time series data prediction model and the multi-variable time series data prediction model, and then use the equipment simulation model to timely and accurately locate the key problems and issue solutions to solve the problems.

[0019] Preferably, step S2 comprises:

[0020] Step S2.1: define the action space, which refers to the set of all possible operations that the system can perform;

[0021] Step S2.2: define a state space, where the state space is a set describing all possible states of the system;

[0022] Step S2.3: The state transfer function is the change of the system state after executing a specific action.

[0023] Preferably, the raw data includes various point data collected by equipment sensors, and the point data includes operating parameters, equipment status and process quality;

[0024] The raw data preprocessing includes data cleaning, data integration, data transformation, data labeling, feature selection, data reduction and data partitioning.

[0025] Preferably, the abnormal rule determination includes upper and lower threshold rules, dual signal deviation threshold rules, jump rules, target value rules, periodic rules, and fluctuation rules;

[0026] Multiple rules can be configured simultaneously for a single warning item.

[0027] Preferably, step S4 comprises:

[0028] Step S4.1: Use the LSTM model to predict future data points and combine it with the Transformer model to build a univariate time series data prediction model;

[0029] Step S4.2: Use the DSTLinear model and the Crossformer model to build a multivariate time series data prediction model.

[0030] Preferably, the LSTM model includes an input gate, a forget gate and an output gate;

[0031] The Transformer model includes a multi-head self-attention module and a position feedforward neural network module.

[0032] Preferably, the DSTLinear model decomposes the trend-seasonal term of the original time series, and selects to extract the time embedding and the space embedding of the trend term respectively;

[0033] The architecture of the Crossformer model includes an encoder, a decoder, and an attention mechanism.

[0034] Preferably, the evaluation method of the model includes determining through set abnormal rules, calculating the prediction error and drawing a 95% confidence interval diagram of the prediction result.

[0035] A system for detecting and predicting device rule anomalies in a digital twin provided by the present invention includes:

[0036] Module M1: Modeling the simulation environment based on the coupling of mechanism and data to obtain the equipment simulation model;

[0037] Module M2: define the action space, state space and state transfer function of the device simulation model, and build the device simulation model;

[0038] Module M3: obtain raw data and perform raw data preprocessing;

[0039] Module M4: setting abnormal rule determination, and constructing a single variable time series data prediction model and a multivariate time series data prediction model according to the abnormal rule determination;

[0040] Module M5: According to the abnormality judgment rules, the preprocessed raw data is divided into univariate time series data and multivariate time series data, and the univariate time series data prediction model and the multivariate time series data prediction model are input accordingly to obtain the training results, and the model evaluation and optimization are performed on each model;

[0041] Module M6: Encapsulating the univariate time series data prediction model and the multivariate time series data prediction model into an algorithm library;

[0042] Module M7: Integrate the trained single-variable time series data prediction model, multivariable time series data prediction model and equipment simulation model into the digital twin platform algorithm library, provide the functions of loading and calling the model through the interface, and complete the iterative deployment of the model;

[0043] Module M8: Call the model in the algorithm library of the digital twin platform, monitor and predict the real-time abnormalities of the equipment through the single-variable time series data prediction model and the multi-variable time series data prediction model, and then use the equipment simulation model to locate the key problems in a timely and accurate manner, and issue solutions to solve the problems.

[0044] Preferably, the module M2 comprises:

[0045] Module M2.1: Define the action space, which is the set of all possible operations that the system can perform;

[0046] Module M2.2: Define the state space, which is the set describing all possible states of the system;

[0047] Module M2.3: The state transfer function is the change of the system state after performing a specific action.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. The present invention establishes a real-time monitoring system based on digital twins, which can integrate data streams from various devices and points of the continuous casting machine, enabling the company to track equipment status in real time, detect potential problems, and monitor all aspects of the production process.

[0050] 2. The present invention uses multi-point data of continuous casting equipment and a variety of deep learning algorithm models to identify abnormal conditions in equipment data, predict future data, help discover potential problems in advance, and further improve the stability of continuous casting equipment.

[0051] 3. The present invention reduces production downtime and maintenance costs, reduces scrap rate, and improves resource utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0053] Figure 1 It is a schematic diagram of the working method of the present invention. DETAILED DESCRIPTION

[0054] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0055] Embodiment 1

[0056] According to a method for detecting and predicting device rule anomalies in a digital twin provided by the present invention, Figure 1 As shown, including:

[0057] Step S1: Modeling the simulation environment based on the mechanism and data coupling to obtain the equipment simulation model. Step S1 includes:

[0058] Step S1.1: Confirm the simulation target according to the dynamic characteristics and key parameters of the equipment to be tested.

[0059] Step S1.2: Establish a mechanism model of each system in the device according to the process principle of the device to be tested.

[0060] Step S1.3: Use machine learning algorithms to build data-driven models.

[0061] Step S1.4: Analyze the connection points between the mechanism model and the data-driven model, and establish a coupling strategy to form a coupling model.

[0062] Taking the equipment continuous casting machine as an example, the simulation environment modeling based on mechanism and data coupling is specifically described as follows: First, the process and main equipment of the continuous casting machine are clarified, the dynamic characteristics and key parameters of the continuous casting machine are analyzed, and the simulation target is confirmed. Then, according to the process principle of the continuous casting machine, the mathematical models of each system are established, such as the crystallizer, the secondary cooling zone, etc. At the same time, the physical phenomena in the continuous casting process, such as solidification and heat transfer, should be considered. Then, the machine learning algorithm is used to establish a data-driven model, and the accuracy and generalization ability of the model are confirmed through data fitting and verification. Then, the correlation points between the mechanism model and the data-driven model are analyzed, such as the physical quantities and time scales of common concern. Design a coupling strategy to form a coupling model.

[0063] Step S2: Define the action space, state space and state transfer function of the device simulation model.

[0064] Step S2.1: Define the action space, which refers to the set of all possible operations that the system can perform. In the continuous casting simulation model, actions usually correspond to parameters or instructions that can be adjusted by the operator or control system. For example:

[0065] Pouring Speed: The pouring speed is adjusted to control the rate at which the liquid metal flows into the mold.

[0066] Cooling intensity: Change the cooling water volume or air flow velocity in the secondary cooling zone to affect the solidification process of the ingot.

[0067] Drawing speed: Adjust the speed of the drawing and straightening machine to control the rate at which the billet is pulled out of the crystallizer.

[0068] Other operations: such as adjusting the vibration frequency of the crystallizer, changing the pouring temperature, etc.

[0069] The action space needs to be precisely defined to ensure that the simulation model can accurately simulate the various operations in the actual production process.

[0070] Step S2.2: Define the state space, which is a set describing all possible states of the system. In the continuous casting simulation model, the state variables usually include various physical quantities related to the continuous casting process, such as temperature, pressure, flow, position, etc. The state space should contain enough information to uniquely identify any state of the system. For example:

[0071] Crystallizer status: including liquid level height and temperature distribution in the crystallizer.

[0072] Secondary cooling zone status: including cooling water flow, air flow velocity, temperature, etc. in the secondary cooling zone.

[0073] Ingot state: including ingot temperature, thickness, surface quality, etc.

[0074] Equipment status: such as the position of the crystallizer vibration device, the speed of the straightening machine, etc.

[0075] The definition of the state space should be as detailed as possible so that it can accurately reflect the actual situation of the continuous casting process.

[0076] Step S2.3: The state transfer function is the change of the system state after performing a specific action. It means that after defining the action space and the state space, it is also necessary to define the state transfer function, which describes how the system state changes after performing a specific action. In the continuous casting simulation model, the state transfer function is usually established based on physical laws and experimental data. The establishment of the state transfer function is based on the combination of selected physical law equations and experimental data. The physical law equations include heat conduction equations and flow dynamics equations. In order to make the model closer to the actual situation, the physical law equations are calibrated with the experimental data.

[0077] Taking a continuous casting machine as an example, the specific description is as follows: the temperature distribution curve under different pouring speeds is measured through experiments; the experimental data is used to adjust the specific values ​​of thermal conductivity (k) and heat source term (Q) in the heat conduction equation; the parameters in the solidification process model are adjusted according to the actual solidification behavior.

[0078] Assume that this embodiment focuses on heat conduction in one dimension, and uses the explicit Euler time integration method to approximately solve the heat conduction equation. The specific calculation steps of the state transfer function are as follows:

[0079] (1) Input current state: state space variables

[0080] (2) Calculate the heat conduction term:

[0081] [\DeltaT_{\text{cond}}=\Deltat\cdot\frac{k}{\rhoc_p}\frac{\partial^2T}{\partialx^2}+\Deltat\cdot\frac{Q}{\rhoc_p}]

[0082] Where \DeltaT_{\text{cond}} represents the temperature change due to heat conduction. \Deltat represents the time step, that is, the time interval considered. k represents the thermal conductivity of the material, the ability of the material to conduct heat. \rho represents the density of the material. c_p represents the specific heat capacity of the material, which represents the amount of heat required to raise the unit temperature of a unit mass of the substance.

[0083] \frac{\partial^2T}{\partialx^2} represents the second-order derivative of the temperature field in the x direction, the rate of change of the temperature change rate, which reflects the diffusion of heat in space. Q is the heat source term, which represents the heat generated per unit volume per unit time. It can be an external heat source or an internal heat source (such as chemical reaction, phase change, etc.). \frac{Q}{\rhoc_p} is the rate of temperature change of unit mass of material due to the heat source.

[0084] (3) Update temperature distribution:

[0085] [T(x,t+\Deltat)=T(x,t)+\DeltaT_{\text{cond}}]

[0086] Where T(x,t) represents the temperature value at position x at time t. T(x,t+\Deltat) represents the updated temperature value at position (x) at time (t+\Deltat).

[0087] (4) Calculation of flow dynamics terms:

[0088] The velocity field (u(x,t+\Deltat)) and the pressure field (p(x,t+\Deltat)) are calculated using the flow dynamics equations.

[0089] (5) Update the liquid metal volume fraction:

[0090] [\DeltaV_l=\Deltat\cdot-\frac{\partialf_s}{\partialT}]

[0091] [V_l(x,t+\Deltat)=V_l(x,t)+\DeltaV_l]

[0092] Among them, \DeltaV_l represents the change in the volume fraction of liquid metal within the time step \Deltat.

[0093] -\frac{\partialf_s}{\partialT}) is the negative derivative of the solid fraction f_s with respect to temperature T. The solid fraction f_s represents the volume fraction of the solid part in the material. The sum of the liquid metal volume fraction V_l and the solid fraction f_s is usually 1 (i.e. (V_l+f_s=1)). The negative sign indicates that as the temperature increases, the solid fraction decreases and the liquid metal volume fraction increases. \frac{\partialf_s}{\partialT} represents the rate of change of the solid fraction for every degree of temperature change.

[0094] (6) Update pouring speed:

[0095] Update the pouring velocity (v(t+\Deltat)) according to the control strategy or external input.

[0096] Through the above steps, we can calculate the new state of the billet according to the current state and action in each time step (\Deltat). This process is repeated until the required simulation time is reached or a specific termination condition is met. In this way, the state transfer function of the continuous casting simulation model is not only based on rigorous physical laws, but also combined with actual experimental data, so that the continuous casting process can be predicted and optimized more accurately.

[0097] By defining the action space, state space, and state transfer function, a complete continuous casting simulation model can be constructed, which can simulate the dynamic behavior of the continuous casting process and be used to optimize the production process and improve product quality. The state of the continuous casting machine simulation environment can include the state and process parameters of each key component of the continuous casting machine, such as the temperature distribution of the crystallizer, the speed of the straightening, the liquid level change of the tundish, etc. The state space should be able to fully describe the current situation of the continuous casting process so that the reinforcement learning model can make decisions. The action space of the continuous casting machine can be defined as adjustable operating parameters, such as the height position of the stopper rod, the temperature of the molten steel in the crystallizer, etc. The action space should contain all possible operations that the model can take so that the model can select the optimal operation strategy.

[0098] Step S3: Preprocessing of raw data. It is necessary to go through reasonable data cleaning, data integration, data transformation, data labeling, feature selection, data reduction and data partitioning. Solve the problems of data duplication, missing, inconsistency and excessive data volume, extract data value and use data effectively.

[0099] Step S4: Design rules for abnormal data judgment, and configure thresholds, weights, boundary conditions, etc. when building the model. At the same time, the results of model training historical data can be compared with the abnormal data marked by rule judgment to determine the accuracy of the model. The rule judgment includes upper and lower threshold rules, dual signal deviation threshold rules, jump rules, target value rules, periodic rules, fluctuation rules, etc. When designing the algorithm, the present invention takes into account the adaptability of different algorithms to different rules as much as possible. At the same time, for a single warning item, multiple rules can be configured at the same time.

[0100] The configuration of anomaly rules requires the combination of expert knowledge and manual annotation of anomaly points in training data, as well as statistical analysis methods, including regression models plus confidence intervals, to find multi-point change relationships. The following introduces four common rules:

[0101] ① Upper and lower threshold rules. For example, the corresponding relationship between the opening of the narrow face foot roller valve in the second cooling water zone 1 and the real-time flow rate 1 of the narrow face foot roller in the second cooling water zone 1 and the real-time pressure of the narrow face foot roller in the second cooling water zone 1. We chose the random forest algorithm to solve this problem, and used two models to learn the mapping relationship between the opening and the upper threshold and the lower threshold respectively. Random forest is an integrated learning method based on decision trees. Its basic principle is to obtain the final prediction value by constructing multiple decision trees and integrating their output results. This is like an integrated classifier with many decision trees, each of which independently classifies the input data, and the final classification result is determined by the majority vote of the outputs of multiple decision trees. Taking the relationship between opening and flow as an example, after model fitting, the opening and flow are not completely one-to-one corresponding. At the same opening, the flow may be high or low. Therefore, the flow corresponding to each opening should have an upper threshold and a lower threshold. Data beyond the upper and lower threshold range is considered to be abnormal data.

[0102] ② Double signal deviation threshold rule. It means that in the experiment, in order to ensure the accuracy of the experimental results, it is necessary to set a clear signal threshold, which can more accurately identify the real amplification signal, avoid the interference of background noise, and ensure the repeatability of the experiment. For example, the correspondence between the set value of the entrance roller gap of the fan segment 10 and the entrance roller gap on the right side of the fan segment 10 and the entrance roller gap on the left side of the fan segment 10, LSTM is used as the regression model to fit the mapping relationship between the set value and the actual values ​​of the left and right roller gaps. The real-time value of the set value is used as the model input to obtain the regression value of the entrance roller gap on the right side of the fan segment 10 and the entrance roller gap on the left side of the fan segment 10. If the actual value of the two entrance roller gaps deviates too much from the regression value, it is judged to be abnormal. Through testing with historical data, the actual value of the roller gap should fluctuate around the set value, and an error greater than 0.5mm is considered to be abnormal data.

[0103] ③ Fluctuation rules. For example, for some point data, such as the crystallizer liquid level, the historical data of the point within t time is obtained every t time, and a sample value within t time is sampled and compared with the previous sample value to obtain the change trend, and then compare the change trends of the most recent n times to see whether the overall trend is maintained (there may be unchanged situations in the middle, that is, the overall trend remains non-decreasing or non-increasing). If there is no such trend, it is judged as abnormal.

[0104] ④ Periodic rules. The count of rule anomalies within a unit time is also used as a rule. If the rule has more than X anomalies within t time, it is determined to be a real anomaly and stored in the database table. When storing, the real-time data of the collected points is no longer stored, but the number of anomalies within t time of the rule is stored instead; if the number of anomalies within t time of the rule is less than X times, it is not determined to be a real anomaly and is not stored.

[0105] The following are some rule configuration examples, as shown in Table 1:

[0106] Table 1

[0107]

[0108] Based on this, we break down and classify complex problems, first considering solving the problem of univariate time series data prediction, including building and training LSTM models and Transformer models.

[0109] Step S4.1: Since the LSTM algorithm is widely used in time series prediction, the present invention needs to consider the problem of past time step information, and LSTM is a powerful choice. By converting the time series data into an appropriate input format, the LSTM model can be trained to predict future data points. LSTM performs well in handling long-term dependencies and sequence modeling. LSTM is a special recurrent neural network (RNN) for processing time series data. Traditional RNNs have the problem of gradient disappearance or gradient explosion, and LSTM solves this problem by introducing a gating mechanism. The LSTM unit includes an input gate, a forget gate, and an output gate, which control the flow and forgetting of information according to the input data and the hidden state of the previous time step, thereby effectively handling long-term dependencies.

[0110] LSTM can judge the one-to-many change relationship in rule anomaly detection. For example, the correspondence between the set value of the entrance roll gap of the sector segment 10 and the entrance roll gap on the right side of the sector segment 10 and the entrance roll gap on the left side of the sector segment 10. According to the user's demand description, the actual value of the roll gap should fluctuate around the set value, with an error of no less than 0.5mm. By using LSTM as a regression model, the mapping relationship between the set value and the actual values ​​of the left and right roll gaps is fitted. The real-time value of the set value is used as the model input to obtain the regression value of the entrance roll gap on the right side of the sector segment 10 and the entrance roll gap on the left side of the sector segment 10. If the actual values ​​of the two entrance roll gaps deviate too much from the regression values, it is judged to be abnormal. The step S4.1 includes:

[0111] Step S4.1.1: Build an LSTM model and use a deep learning framework (such as PyTorch, etc.) to build an LSTM model. The LSTM model usually consists of one or more LSTM layers, and finally a fully connected layer to output the prediction results.

[0112] Step S4.1.2: Train the LSTM model, use the prepared input sequence and output sequence as training data, and use the LSTM model for training. You can set an appropriate loss function and optimizer, and define the number of training iterations.

[0113] Step S4.1.3: Multi-step prediction, after completing the model training, select the initial input to use the trained model for multi-step prediction, such as using the first 300 data as the initial input. It should be noted that in each step of prediction, the prediction result at the current time point is used as the input for the next step in order to achieve continuous prediction.

[0114] Step S4.1.4: Determine the threshold value. For the prediction results, the threshold value can be determined according to specific needs. The factors for selecting the threshold value include error range, business needs, etc.

[0115] Step S4.1.5: Output the results, generate the final output results according to the prediction results of the model and the threshold, for example, reversely transform the prediction results and restore them to the scale of the original data.

[0116] Step S4.2: Because in the field of time series, the Transformer model performs well in sequence modeling and prediction. Due to the parallelization characteristics of the Transformer, it can efficiently process long time series data and also performs well in capturing long-term dependencies in the sequence. Therefore, the Transformer has a wide range of applications in processing large-scale time series data with long-term correlations (such as traffic flow prediction, power load prediction, etc.).

[0117] Transformer is a deep learning model based on self-attention mechanism, which was originally used for natural language processing tasks. Self-attention allows the model to directly focus on all input positions when processing sequence data without processing time steps in a recursive way. Transformer also introduces position encoding to retain the position information of the input sequence. Therefore, Transformer is selected for training.

[0118] Step S4.2.1, build the network structure: The Transformer network contains modules such as multi-head self-attention and position feedforward neural network. It is necessary to set parameters such as the number of layers and the number of self-attention heads of the model.

[0119] Step S4.2.2, training the network: update the parameters of the Transformer network by minimizing the prediction error (e.g., cross entropy loss).

[0120] Step S4.2.3, data prediction: Use the trained Transformer network to predict future changes in time series.

[0121] Step S4.3, the above steps S4.1 and 4.2 solve the problem of univariate time series data prediction. The next step is to consider the problem of multivariate time series data prediction. Considering that there is a strong correlation between the points of the same equipment, the point data of the stopper rod and the crystallizer liquid level are selected as training data when training the multivariate model.

[0122] Later in the training process, it was found that the training results were poor if only different equipment was used to distinguish the point groups for training. Therefore, based on the equipment classification, the points were further classified according to type.

[0123] Similarly, we first consider the algorithm design. DSTLinear is a good choice. DSTLinear is an improvement based on DLinear. Following the design of DLinear, the original time series trend-seasonal term is decomposed. Then, considering that the trend term contains rich spatiotemporal information, we choose to extract time embedding and spatial embedding for the trend term respectively.

[0124] Step S4.3.1, perform trend-seasonal decomposition on the original time series.

[0125] In step S4.3.2, the seasonal term is directly passed through a linear layer to obtain the prediction of the seasonal term.

[0126] Step S4.3.3, processing the trend item, including: using TCN to process the trend item, extracting information of different time spans, that is, obtaining time embedding; using the linear layer to process the trend item, extracting information between different variables, that is, obtaining space embedding; superimposing the time embedding and the space embedding with the original trend item according to the coefficients, and processing them through the linear layer to obtain the prediction of the trend part.

[0127] Step S4.3.4, add the trend part and the season part prediction results to obtain the final prediction result.

[0128] Step S4.4, the attention mechanism of Transformer time series prediction mainly models the correlation of the same variable at different times, but ignores the dependency between different variables. In order to fill this gap, consider using Crossformer to use cross-dimensional dependencies for multivariate time series prediction. The Crossformer model as a whole consists of three modules:

[0129] ①Dimension-Segment-Wise Embedding: The main goal of this module is to divide the time series into patches and obtain patch embeddings as the input of the subsequent model. The time series of each variable is divided into multiple blocks according to a certain window, and each block is mapped through full connection;

[0130] ②Two-Stage Attention Layer: Two-stage attention. The first stage performs attention in the time dimension, and the second stage performs attention between multiple variables, which better captures the spatiotemporal dependency relationship;

[0131] ③Hierarchical Encoder-Decoder: Uses different patch sizes and partitioning methods to form a hierarchical encoding and decoding structure.

[0132] Step S4.4.1, define the architecture of the crossformer model, including encoder, decoder, attention mechanism and other components. Initialize the parameters and weights of the model.

[0133] In step S4.4.2, for each training batch, perform the following operations:

[0134] ① Sample a batch of source language sequences and target language sequences from the training set.

[0135] ② Input the source language sequence into the encoder to obtain the encoded representation.

[0136] ③ Input the encoded representation together with partial input of the target language sequence into the decoder for forward propagation.

[0137] ④Calculate loss functions, such as cross entropy loss.

[0138] ⑤Execute the back propagation algorithm to update the model parameters.

[0139] ⑥ Regularly evaluate the performance of the model on the validation set to monitor overfitting.

[0140] Step S5, according to the abnormality judgment rule, the pre-processed raw data is divided into univariate time series data and multivariate time series data, and the univariate time series data prediction model and the multivariate time series data prediction model are input accordingly to obtain the training results, and each model is evaluated and optimized. In addition to the rule judgment, the model evaluation is not only verified by the designed rule judgment, but also calculates the prediction error (mean square error MSE, root mean square error RMSE) and draws the 95% confidence interval diagram of the prediction result.

[0141] The mean squared error (MSE) is a common indicator for measuring prediction error. In machine learning, MSE is often used to evaluate the difference between the model prediction results and the true value. The specific calculation formula is as follows:

[0142] MSE=1 / nΣ(y_true-y_pred)^2

[0143] Where n is the number of samples, y_true is the true value, and y_pred is the model prediction. By summing this formula and dividing by the number of samples n, we can get the square of the average prediction error.

[0144] Calculate RMSE Root Mean Squared Error (RMSE) is an indicator used to measure the prediction accuracy of a prediction model on continuous data. It measures the root mean square difference between the predicted value and the true value, indicating the average degree of deviation between the predicted value and the true value, and is one of the commonly used performance evaluation indicators in regression tasks.

[0145] The calculation formula of RMSE is as follows:

[0146] RMSE=√((1 / n)*Σ(y_pred-y_true)2)

[0147] Among them, n is the number of samples, y_true is the true value, and y_pred is the model prediction value.

[0148] The confidence interval graph of the prediction results refers to a graph used to display the confidence interval of statistical data, which usually consists of a center line and two boundary lines. The center line represents the mean of the sample, while the boundary lines represent the upper and lower limits of the confidence interval. By observing the position of the boundary lines, we can derive the estimated range of the population parameter. If the boundary line is narrow, it means that the estimate is more accurate; if the boundary line is wide, it means that the estimate is less accurate. A 95% confidence interval means that there is a 95% probability that the population parameter falls within this interval. You can use the code to build a confidence interval graph.

[0149] Step S6, model encapsulation. The single-variable time series data prediction model and the multivariable time series data prediction model are encapsulated into an algorithm library, which is classified layer by layer according to the algorithm purpose, algorithm type, and model point. An external call interface is provided, and the corresponding model is loaded according to the request body content. Anomaly detection or prediction is performed on the incoming data, and the detection results are returned.

[0150] Step S7, iterative deployment. The trained single-variable time series data prediction model, multi-variable time series data prediction model and equipment simulation model are integrated into the digital twin platform algorithm library, and the functions of loading and calling models are provided in the form of interfaces. Subsequent iterations of the model will be based on the algorithm library. It only needs to update and replace the model weight file and modify the loading and calling methods to complete the iterative deployment of the model.

[0151] Step S8, call the model in the algorithm library of the digital twin platform, monitor and predict the real-time abnormality of the equipment through the single variable time series data prediction model and the multivariable time series data prediction model, and then use the equipment simulation model to locate the key problems in a timely and accurate manner, and issue solutions to deal with the problems. It can also conduct historical backtracking, simulate historical data, summarize experience, use it for self-learning and update iteration of the model, and save all available information into the knowledge base for management and utilization.

[0152] Step S9, model optimization. The effect of the model is verified by comparison with the actual continuous casting machine, including the effects of optimizing process parameters, improving product quality, and improving production efficiency. The model is optimized and adjusted according to the actual verification results, including adjusting the neural network structure, adjusting training parameters, filtering low-probability outputs, etc., to further improve the performance and adaptability of the model.

[0153] The present invention aims to utilize digital twin technology to solve the problem of abnormal detection and prediction of multi-point change relationships of equipment through an organic combination of expert rule system and intelligent algorithm model.

[0154] The present invention establishes a real-time monitoring system through digital twin technology, which can integrate data and apply models such as LSTM, transformer, DSTLinear and crossformer to the equipment digital twin system for calling. For time series data of different granularities, we can choose different models for prediction and anomaly detection to meet different business needs. It can effectively detect anomalies and identify trends and patterns that may lead to failures or quality problems in advance, reduce production downtime and maintenance costs, reduce scrap rate, and improve resource utilization efficiency.

[0155] Embodiment 2

[0156] The present invention also provides a system for detecting and predicting device rule anomalies in a digital twin. The system for detecting and predicting device rule anomalies in a digital twin can be implemented by executing the process steps of the method for detecting and predicting device rule anomalies in a digital twin, that is, a person skilled in the art can understand the method for detecting and predicting device rule anomalies in a digital twin as a preferred implementation of the system for detecting and predicting device rule anomalies in a digital twin.

[0157] A system for detecting and predicting device rule anomalies in a digital twin provided by the present invention includes:

[0158] Module M1: Modeling the simulation environment based on the coupling of mechanism and data to obtain the equipment simulation model.

[0159] Module M2: Define the action space, state space and state transfer function of the device simulation model. The module M2 includes: Module M2.1: Define the action space, which refers to the set of all possible operations that can be triggered by the system. Module M2.2: Define the state space, which is a set that describes all possible states of the system. Module M2.3: The state transfer function is the change of the system state after triggering a specific action.

[0160] Module M3: Acquire raw data and perform raw data preprocessing. The raw data includes data from each point collected by the equipment sensor, and the point data includes operating parameters, equipment status and process quality. The raw data preprocessing includes data cleaning, data integration, data transformation, data labeling, feature selection, data reduction and data partitioning.

[0161] Module M4: Set abnormal rule judgment, and build a single variable time series data prediction model and a multivariate time series data prediction model according to the abnormal rule judgment. The abnormal rule judgment includes upper and lower threshold rules, dual signal deviation threshold rules, jump rules, target value rules, periodic rules, and fluctuation rules. Multiple rules can be configured simultaneously for a single warning item. The module M4 includes: Module M4.1: Use the LSTM model to predict future data points, and combine the Transformer model to build a single variable time series data prediction model. Module M4.2: Use the DSTLinear model and the Crossformer model to build a multivariate time series data prediction model. The LSTM model includes an input gate, a forget gate, and an output gate. The Transformer model includes a multi-head self-attention module and a position feedforward neural network module. The DSTLinear model decomposes the original time series trend-seasonal term, and selects to extract time embedding and spatial embedding for the trend term respectively. The architecture of the Crossformer model includes an encoder, a decoder, and an attention mechanism.

[0162] Module M5: According to the abnormality judgment rules, the preprocessed raw data is divided into univariate time series data and multivariate time series data, and the univariate time series data prediction model and the multivariate time series data prediction model are input accordingly to obtain the training results, and each model is evaluated and optimized. The model evaluation method includes judging through the set abnormality rules, calculating the prediction error, and drawing the 95% confidence interval diagram of the prediction result.

[0163] Module M6: Encapsulate the univariate time series data prediction model and the multivariate time series data prediction model into an algorithm library.

[0164] Module M7: Integrate the trained single-variable time series data prediction model, multi-variable time series data prediction model and equipment simulation model into the digital twin platform algorithm library, provide the function of loading and calling the model through the form of an interface, and complete the iterative deployment of the model.

[0165] Module M8: Call the model in the algorithm library of the digital twin platform, monitor and predict the real-time abnormalities of the equipment through the single-variable time series data prediction model and the multi-variable time series data prediction model, and then use the equipment simulation model to locate the key problems in a timely and accurate manner, and issue solutions to solve the problems.

[0166] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, the system and its various devices, modules, and units provided by the present invention can be completely realized in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by performing logic programming on the method step S. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0167] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for detecting and predicting anomalies of equipment rules in a digital twin, characterized in that: include: Step S1: Modeling the simulation environment based on the coupling of mechanism and data to obtain a device simulation model; Step S2: defining the action space, state space and state transfer function of the device simulation model; Step S3: obtaining raw data and performing raw data preprocessing; Step S4: setting an abnormal rule determination, and constructing a univariate time series data prediction model and a multivariate time series data prediction model according to the abnormal rule determination; Step S5: According to the abnormality determination rule, the preprocessed raw data is divided into univariate time series data and multivariate time series data, and the univariate time series data prediction model and the multivariate time series data prediction model are input accordingly to obtain the training results, and the model evaluation and optimization are performed on each model; Step S6: Encapsulating the univariate time series data prediction model and the multivariate time series data prediction model into an algorithm library; Step S7: Integrate the trained single-variable time series data prediction model, multi-variable time series data prediction model and equipment simulation model into the digital twin platform algorithm library, provide the functions of loading and calling the model through the interface, and complete the iterative deployment of the model; Step S8: Call the model in the algorithm library of the digital twin platform, monitor and predict the real-time abnormalities of the equipment through the single-variable time series data prediction model and the multi-variable time series data prediction model, and then use the equipment simulation model to timely and accurately locate the key problems and issue solutions to solve the problems.

2. The method for detecting and predicting equipment rule anomalies in digital twins according to claim 1, characterized in that: The step S2 comprises: Step S2.1: define the action space, which refers to the set of all possible operations that the system can perform; Step S2.2: define a state space, where the state space is a set describing all possible states of the system; Step S2.3: The state transfer function is the change of the system state after executing a specific action.

3. The method for detecting and predicting equipment rule anomalies in digital twins according to claim 1, characterized in that: The raw data includes data of various points collected by equipment sensors, and the data of various points includes operating parameters, equipment status and process quality; The raw data preprocessing includes data cleaning, data integration, data transformation, data labeling, feature selection, data reduction and data partitioning.

4. The method for detecting and predicting equipment rule anomalies in digital twins according to claim 1, characterized in that: The abnormal rule determination includes upper and lower threshold rules, dual signal deviation threshold rules, jump rules, target value rules, periodic rules, and fluctuation rules; Multiple rules can be configured simultaneously for a single warning item.

5. The method for detecting and predicting equipment rule anomalies in digital twins according to claim 1, characterized in that: The step S4 comprises: Step S4.1: Use the LSTM model to predict future data points and combine it with the Transformer model to build a univariate time series data prediction model; Step S4.2: Use the DSTLinear model and the Crossformer model to build a multivariate time series data prediction model.

6. The method for detecting and predicting equipment rule anomalies in digital twins according to claim 5, characterized in that: The LSTM model consists of an input gate, a forget gate, and an output gate; The Transformer model includes a multi-head self-attention module and a position feedforward neural network module.

7. The method for detecting and predicting equipment rule anomalies in digital twins according to claim 5, characterized in that: The DSTLinear model decomposes the trend-seasonal term of the original time series, and selects to extract the time embedding and space embedding of the trend term respectively; The architecture of the Crossformer model includes an encoder, a decoder, and an attention mechanism.

8. The method for detecting and predicting equipment rule anomalies in digital twins according to claim 1, characterized in that: The model is evaluated by setting anomaly rules, calculating the prediction error, and drawing a 95% confidence interval graph of the prediction results.

9. A system for detecting and predicting anomalies of equipment rules in a digital twin, characterized in that: include: Module M1: Modeling the simulation environment based on the coupling of mechanism and data to obtain the equipment simulation model; Module M2: define the action space, state space and state transfer function of the device simulation model, and build the device simulation model; Module M3: obtain raw data and perform raw data preprocessing; Module M4: setting abnormal rule determination, and constructing a single variable time series data prediction model and a multivariate time series data prediction model according to the abnormal rule determination; Module M5: According to the abnormality judgment rules, the preprocessed raw data is divided into univariate time series data and multivariate time series data, and the univariate time series data prediction model and the multivariate time series data prediction model are input accordingly to obtain the training results, and the model evaluation and optimization are performed on each model; Module M6: Encapsulating the univariate time series data prediction model and the multivariate time series data prediction model into an algorithm library; Module M7: Integrate the trained single-variable time series data prediction model, multivariable time series data prediction model and equipment simulation model into the digital twin platform algorithm library, provide the functions of loading and calling the model through the interface, and complete the iterative deployment of the model; Module M8: Call the model in the algorithm library of the digital twin platform, monitor and predict the real-time abnormalities of the equipment through the single-variable time series data prediction model and the multi-variable time series data prediction model, and then use the equipment simulation model to locate the key problems in a timely and accurate manner, and issue solutions to solve the problems.

10. A system for detecting and predicting equipment rule anomalies in a digital twin according to claim 9, characterized in that: The module M2 comprises: Module M2.1: Define the action space, which is the set of all possible operations that the system can perform; Module M2.2: Define the state space, which is the set describing all possible states of the system; Module M2.3: The state transfer function is the change of the system state after performing a specific action.

Images