A material manufacturing process control and optimization method and system based on AI digital twinning

Abstract

The application discloses a material manufacturing process control and optimization method and system based on AI digital twinning, comprising: segmenting and aligning the management of multi-source heterogeneous data of the production line based on process event labels; constructing a multi-modal AI digital twinning model introducing a physical mechanism residual term, and fusing time series and attribute data through a time-varying feature interaction mechanism; establishing a sliding observation window, dynamically calculating a noise suppression factor based on the statistical characteristics of the prediction error to correct the model output, and combining the device health degree to perform gated parameter updating; under the constraints of the confidence domain and the safety boundary, the optimal process setting value is searched by using the Bayesian strategy; after the setting value is kinematically intercepted, it is executed; and the model is evolved without feeling through a shadow mode verification mechanism. The application effectively solves the problems of strong nonlinearity, difficult to suppress working condition drift and lack of safety of AI control in the traditional material manufacturing process, and realizes high-precision closed-loop control and continuous optimization of the manufacturing process.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent manufacturing and advanced process control technology, and in particular to a method and system for controlling and optimizing material manufacturing processes based on AI digital twins. Background Technology

[0002] Materials manufacturing processes (such as the molding and processing of rubber, plastics, and polymers) are characterized by multivariate coupling, large time lags, strong nonlinearity, and significant batch-to-batch variations. Despite the widespread adoption of Supervisory Control and Data Acquisition (SCADA) systems with industrial development, the following technical challenges remain in process control and optimization: Traditional pure mechanism modeling (such as finite element analysis) is computationally intensive and difficult to meet the needs of online real-time control; while general pure data-driven models (such as conventional deep learning) are black box models, lacking physical constraints, and are prone to outputting prediction results that violate physical common sense under out-of-bounds conditions, leading to control failure.

[0003] Existing digital twin technologies often lack effective online synchronization mechanisms, leading to separation between the virtual and real systems during production due to sensor drift, raw material fluctuations, or equipment aging. Conventional PID feedback or simple deviation addition is insufficient to handle complex nonlinear drift, and forcibly updating model parameters during equipment failure (such as sensor damage) can result in the model being contaminated with dirty data, leading to catastrophic amnesia.

[0004] Furthermore, existing AI optimization algorithms often focus on finding theoretically optimal solutions, neglecting the physical limitations of actuators (such as heating power ramp-up rate and valve dead zone). Directly sending algorithm outputs to the control system can easily cause equipment vibration or even damage, making it difficult for AI technology to be truly implemented in core control processes. Therefore, there is an urgent need for a material manufacturing process control and optimization method and system based on AI digital twins. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method and system for controlling and optimizing material manufacturing processes based on AI digital twins.

[0006] To achieve the above objectives, this invention provides a method for controlling and optimizing a material manufacturing process based on AI digital twins, comprising the following steps: S1. Multi-source heterogeneous data governance: Collect real-time time-series data, static attribute data and equipment status data from the material manufacturing production line, and perform segmentation and alignment processing on the real-time time-series data based on process event tags; S2. Physical and Data-Driven Modeling: Construct a multimodal AI digital twin model, use a time-varying feature interaction mechanism to fuse the real-time time series data and the static attribute data, and introduce physical mechanism residual terms to constrain the model output to obtain process prediction values ​​that conform to physical laws. S3. Adaptive synchronization of statistical characteristics: Establish a sliding observation window, dynamically calculate the noise suppression factor based on the statistical variance of the prediction error to correct the model output, and combine the equipment health factor to perform gated elastic updates of the model parameters. S4. Bayesian optimization within the confidence region: Under the dual constraints of model prediction confidence and process safety boundary, a Bayesian optimization strategy is used to search for candidate process setpoints. S5. Kinematic safety interception and execution: Perform amplitude truncation verification and rate limit verification on the candidate process set value, and generate the final instruction to be sent to the control system; S6, Shadow Model Validation and Evolution: Run the shadow model in the background and predict in parallel with the main model, triggering version switching based on the cumulative accuracy evaluation results.

[0007] Furthermore, the construction of the multimodal AI digital twin model in step S2 includes: calculating the dynamic injection weight of the static feature vector at the current time step, and using the dynamic injection weight to perform weighted fusion of the static features; at the same time, constructing a priori calculation module based on material reaction dynamics, multiplying its output by a learnable coefficient, and then superimposing it on the decoding end.

[0008] Furthermore, the adaptive synchronization in step S3 adopts the following strategy: the prediction error variance is mapped to a noise suppression factor using a sigmoid function, and the correction amplitude is suppressed when the variance increases; at the same time, the model parameters are only allowed to be updated at a rate of normalized error rate when the equipment health factor is higher than the safety threshold.

[0009] Furthermore, the safety interception in step S5 includes: determining whether the set value exceeds the physical limit of the device, and determining whether the rate of change of the set value exceeds the response limit of the actuator.

[0010] The present invention also provides a material manufacturing process control and optimization system based on AI digital twin, including a data governance module, a digital twin engine, a decision and security module, and a model operation and maintenance module, for executing the above methods.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: Through time-varying interaction, the model can dynamically perceive the weighted influence of changes in static attributes such as formulation and environment at different process stages (such as heating and holding pressure). Through physical residual constraints, it is ensured that the model's prediction results in data-scarce regions still conform to the laws of material reaction kinetics, thereby solving the problem of unreliable extrapolation of pure data models.

[0012] By dynamically adjusting the noise suppression factor through the variance calculation within the sliding window, the model can distinguish between random noise and system drift, remain stable when the sensor jitters, and quickly follow changes in operating conditions. At the same time, by introducing a device health gating mechanism, parameter updates are forcibly stopped during device failures, effectively preventing abnormal data from contaminating the model and ensuring the health of the digital twin throughout its entire lifecycle.

[0013] Furthermore, this invention constructs a dual defense line of confidence region constraint and kinematic safety interception. In the optimization stage, it automatically eliminates risk areas with high model uncertainty. In the execution stage, it prevents overload or oscillation of the actuator caused by AI output jumps through amplitude truncation and rate limiting, thus solving the key pain point that AI control algorithms are difficult to safely implement in industrial sites.

[0014] By adopting a shadow mode verification mechanism, parallel verification is performed using a background shadow model without interfering with the main model control. Switching is only performed after confirming that the performance is better and there is no risk of divergence, thereby achieving seamless upgrades in the production process and ensuring the continuity and reliability of the control system. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0016] Figure 1 This is a schematic diagram of the overall architecture of a material manufacturing process control and optimization system based on AI digital twin provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the main flow of the material manufacturing process control and optimization method based on AI digital twin provided in the embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the construction principle of a multimodal AI digital twin model driven by both physics and data in this embodiment of the invention. Figure 4 This is a logic diagram of the adaptive synchronization and elastic parameter update strategy for statistical characteristics in this embodiment of the invention; Figure 5 This is a logical diagram of kinematic safety interception and execution control in an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] This invention provides a method and system for controlling and optimizing material manufacturing processes based on AI digital twins. The following four embodiments elaborate on these methods from four dimensions: system architecture and data governance, model building principles, online control strategies, and model evolution mechanisms. This embodiment uses precision rubber injection molding as an example. This process is characterized by typical multivariate coupling, strong nonlinearity, and extremely high requirements for batch stability.

[0019] Example 1: like Figure 1 and Figure 2 As shown, the deployment architecture and execution method of an AI-based digital twin-based material manufacturing process control and optimization method in an industrial setting are as follows: System Deployment Environment: The system described in this embodiment is deployed on an edge computing server within a rubber injection molding production line. The system is connected via an industrial bus (such as OPC UA or Modbus TCP) to the PLC controller, temperature control box, vulcanizer, and an additional high-frequency sensor group (including mold cavity pressure sensor, screw position sensor, and machine vibration sensor) on the production line.

[0020] In actual production, data is often disorganized and asynchronous. This embodiment performs data governance through the following logic: The system first reads a pre-defined variable dictionary, which defines three types of data: Real-time timing data: such as mold temperature (sampling rate 1Hz), injection pressure (sampling rate 100Hz), screw speed (sampling rate 50Hz).

[0021] Static attribute data: such as rubber compound formula number, number of mold cavities, cumulative number of times the mold has been used (reflecting the degree of wear), and current ambient temperature and humidity.

[0022] Equipment status data: such as equipment operating mode, alarm codes, and motor vibration spectrum.

[0023] Because rubber injection molding is a cyclical batch process, the system monitors key process event signals issued by the PLC. For example, when a "mold closing completion signal" is detected, it is marked as... When a "mold opening signal" is detected, it is marked as Based on this, the system performs double alignment, including: Frequency alignment: Based on the timestamp of the highest frequency sensor (such as 100Hz pressure data), linear interpolation is performed on the low frequency sensor (such as 1Hz temperature data) to fill in the missing time point data and generate a unified multi-channel time series tensor.

[0024] And phase alignment: to The absolute time length is mapped to a normalized process time axis from 0% to 100%, so that even if the production cycle of different batches fluctuates slightly (such as fine-tuning of vulcanization time), the data can still maintain alignment in terms of process characteristics, which is convenient for model learning.

[0025] Example 2: like Figure 3 As shown, based on Example 1, the multimodal AI digital twin model (i.e., the main model) constructed in this example includes a temporal encoder, an attribute encoder, a time-varying feature interaction mechanism, and a physical mechanism residual constraint module, specifically: The multimodal input coding model contains two encoder branches, namely: The time-series encoder uses a Long Short-Term Memory (LSTM) network to process real-time time-series data and capture the dynamic trends (dynamic feature vectors) of temperature and pressure changes over time.

[0026] And the attribute encoder: uses a fully connected neural network (Dense Layer) to process static attribute data and extract the potential impact of formulation and environment on the process (static feature vector).

[0027] Because traditional models often directly concatenate static attributes to time-series data, ignoring the changes in attribute weights at different stages, this embodiment introduces a time-varying feature interaction mechanism, specifically: At each time step (e.g., injection stage, holding stage, vulcanization stage), the model dynamically calculates the dynamic injection weights of static features; After performing a linear transformation between the current dynamic feature vector and the global static feature vector, the input is given to the hyperbolic tangent activation function (Tanh) to obtain the interaction score. Then, the score is converted into a weight value between 0 and 1 by the normalized exponential function (Softmax).

[0028] The above steps achieve the following effects: for example, the static attribute "mold wear" has a significant impact on pressure during the "mold closing" and "injection" stages, so the model automatically assigns it a high weight; while during the "pressure holding and vulcanization" stage, its weight automatically decreases. This mechanism greatly enhances the model's ability to perceive process details.

[0029] In addition, this embodiment connects a priori calculation module in parallel at the decoder output. The variable dictionary contains a simplified rubber vulcanization kinetic equation (such as the Arrhenius equation), and calculates the theoretical vulcanization degree curve (mechanism prediction value) based on the current temperature integral.

[0030] The final output of the model is calculated using the following residual summation formula: in, The predicted data output by the decoder. The mechanistic prediction value output by the prior calculation module. These are the residual coefficients, automatically learned by the neural network. This design ensures that the model's predictive trends, while based on data, are always "guided" and corrected by the laws of physics.

[0031] Example 3: like Figure 4 and Figure 5 As shown, based on the physical and data dual-drive model constructed in Example 2, this example details the online adaptive synchronization and safety control process of the system, specifically including: adaptive synchronization based on statistical characteristics, gated elastic parameter update, and kinematic safety interception steps.

[0032] Adaptive synchronization of statistical properties: After the model is deployed, the accuracy issue needs to be addressed, such as... Figure 4 As shown in the upper real-time correction loop, in order to eliminate prediction bias and suppress high-frequency noise, the digital twin engine 12 executes the following logic: Variable definition: Set the actual observation value of the sensor at the current moment. The model's original prediction value is ; Calculate statistics: The system maintains a sliding observation window (e.g., window length). ), calculate the statistical variance of the data within the window in real time ( ) and average deviation ( ); Calculate the noise suppression factor ( The noise suppression factor is calculated using the Sigmoid mapping function, as shown in the following formula: in, This is the sensitivity coefficient. This is the variance threshold; when the data fluctuates drastically (variance)... When it is relatively large, Approaching 1 (representing a high-noise state); when the data is stationary (variance) When it is smaller, Approaching 0 (representing a stable state).

[0033] Deviation compensation calculation: Calculate the final deviation compensation amount ( ): The physical meaning of this formula is that when the noise is high ( )hour, That is, no large-scale correction is performed to prevent misoperation; when the noise is low ( )hour, This means that the deviation is fully compensated.

[0034] Synchronous output: Finally, the original predicted value is added to the deviation compensation amount to obtain the corrected synchronous output value: This value will serve as the basis for subsequent error calculations and control decisions.

[0035] Gated elasticity parameter updates: After obtaining the synchronized output value, in order to ensure the safety of model iteration, the system introduces a strict gating mechanism, specifically: Input calculation: Real-time calculation of the corrected absolute error The equipment health factor (HI) is calculated by combining the real-time collected equipment vibration spectrum data.

[0036] To convert the absolute error into the update step size of the model parameters, the system uses the following formula to calculate the basic update rate. : in, The preset maximum update coefficient (e.g., 0.01). This is the error scaling factor. The function of this formula is: when the error... When the error is small, the update rate increases linearly with the error; when the error is large, the update rate saturates. To prevent excessive parameter updates; Gating logic: The system determines whether the following two conditions are met simultaneously: Equipment health: Preset security threshold; Data valid: absolute error Preset cutoff threshold (to exclude outliers caused by sensor malfunctions).

[0037] The execution branches include: Perform EMA parameter update: If the above conditions are met simultaneously, it indicates that the equipment is operating normally and the data is valid. The system executes the exponential moving average (EMA) parameter update strategy and fine-tunes the model weights to adapt to the drift of operating conditions.

[0038] And stop updating: If " "Too low" (e.g., heating element malfunction) or "Too large error" (e.g., sensor disconnection), such as Figure 4 As shown in the bottom right branch, the system forcibly stops updates (i.e., the update rate). This mechanism constitutes a data firewall, effectively preventing dirty data caused by equipment failure from polluting model parameters and solving the "catastrophic forgetting" problem that easily occurs in traditional online learning.

[0039] Bayesian optimization within the confidence region: Based on ensuring the accuracy of the model, the decision-making and safety module is responsible for finding the optimal process parameters, specifically: Define process soft boundaries (based on historical high-quality batches) Statistical distribution range) and hard boundaries (equipment physical limits); Within the search space defined by the aforementioned boundaries, a surrogate model is constructed using Gaussian process regression (GPR). The GPR model not only outputs the predicted values ​​but also an estimate of the variance of the predictions (i.e., the confidence level). When searching for candidate process settings, a preset uncertainty threshold is set. If a set of parameters causes the model's prediction variance to be too large (indicating that the model is unfamiliar with the region and the risk is high), the sampling point is directly removed to ensure that the optimization is always carried out within the "safety known" confidence region.

[0040] The kinematic safety interception process is as follows: To prevent the instructions output by the AI ​​algorithm from exceeding the physical limits of the actuator, the system performs two levels of verification logic before issuing the instructions: Level 1: Amplitude Truncation Verification. Let the target setpoint given by the optimization algorithm be... The equipment's permissible safety range is The system judges Is it within the range? like If so, then it will be forcibly assigned the value Max; like If the value is 0, then it will be forcibly assigned the value Min; Otherwise, keep The value remains unchanged; the value after this step is denoted as [value]. .

[0041] Level Two: To prevent instruction jumps from overloading the actuator (e.g., excessively rapid increases in heating power), the system introduces a rate limit. Let the execution value at the previous moment be... The instruction currently pending execution is (i.e., the above) ); First, calculate the instruction difference. ; Determine the absolute value of the change Does it exceed the maximum allowable rate of change by the actuator? ): like This indicates that the instruction has changed too quickly, and the system reconstructs the current instruction according to the formula: in, This is a sign function used to determine the direction of increase or decrease in a change. The function of this formula is to force the magnitude of the change in the instruction to be limited. Within this range, smooth "climbing" control is achieved. If Then output directly. As the final safety directive.

[0042] Example 4: Based on the above embodiments, the system's long-term self-iteration process specifically includes the following steps: As production progresses, the system continuously accumulates actual production data with quality labels. When the number of new samples reaches a preset batch size (e.g., 100 models), the cloud training server initiates incremental learning, using the new data to fine-tune the old model and generate candidate model parameters. To avoid production accidents caused by directly updating the model, this embodiment adopts a shadow mode mechanism: the candidate model is loaded into the background process of the edge server as a "shadow model". It accesses real-time production data for inference and prediction, but its output is not sent to the control system, but is only used for recording. The system runs the main model and the shadow model simultaneously, and calculates the performance indicators (such as mean absolute percentage error, MAPE) of both within a set period (such as 24 hours). The system periodically evaluates the performance of the shadow model and triggers a version switch only when all of the following conditions are met: Condition 1: The MAPE of the shadow model is significantly lower than that of the current main model; Condition 2: The shadow model did not trigger a divergence alarm (i.e., the predicted value did not show unreasonable oscillations or drifts); Condition 3: The prediction error of the shadow model on key security indicators is within the allowable range.

[0043] Once confirmed, the system replaces the main model parameters with candidate model parameters during production breaks (such as shift changes), completing the seamless upgrade and evolution of the model.

[0044] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for controlling and optimizing a material manufacturing process based on AI digital twins, characterized in that, Includes the following steps: S1. Multi-source heterogeneous data governance: Collect real-time time-series data, static attribute data and equipment status data from the material manufacturing production line, and perform segmentation and alignment processing on the real-time time-series data based on process event tags; S2. Physical and Data-Driven Modeling: Construct a multimodal AI digital twin model as the main model, use a time-varying feature interaction mechanism to fuse the real-time time series data and the static attribute data, and introduce physical mechanism residual terms to constrain the model output to obtain process prediction values ​​that conform to physical laws. S3. Adaptive synchronization of statistical characteristics: Establish a sliding observation window, dynamically calculate the noise suppression factor based on the statistical variance of the prediction error to correct the model output, and combine the equipment health factor to perform gated elastic updates of the model parameters. S4. Bayesian optimization within the confidence region: Under the dual constraints of model prediction confidence and process safety boundary, a Bayesian optimization strategy is used to search for candidate process setpoints. S5. Kinematic safety interception and execution: Perform amplitude truncation verification and rate limit verification on the candidate process set value, and generate the final instruction to be sent to the control system; S6, Shadow Model Validation and Evolution: Run the shadow model in the background and predict in parallel with the main model, triggering version switching based on the cumulative accuracy evaluation results.

2. The method for controlling and optimizing a material manufacturing process based on AI digital twins according to claim 1, characterized in that, Step S1, which involves segmenting and aligning the real-time time series data, specifically includes: Construct a variable dictionary containing physical quantity definitions, unit systems, and sampling frequencies; Monitoring the material feeding signal, pressure holding signal, and mold opening signal during the production process serves as the anchor point for critical events; Based on the sampling timestamp of the highest frequency sensor, the time series data of the low frequency sensor is interpolated and completed. Based on the key event anchor points, continuous time stream data is segmented into discrete batch process segments, and the batch process segments are mapped onto a unified time axis.

3. The method for controlling and optimizing a material manufacturing process based on AI digital twins according to claim 1, characterized in that, The specific construction and calculation process of the multimodal AI digital twin model in step S2 includes: The dynamic feature vector of the real-time time series data is extracted using a long short-term memory network, and the static feature vector of the static attribute data is extracted using a fully connected layer. Calculate the dynamic injection weight of the static feature vector at the current time step. The dynamic injection weight is obtained by inputting the linearly transformed dynamic feature vector and the static feature vector into the hyperbolic tangent activation function, and then processing them through the normalized exponential function. The static feature vector is weighted using the dynamically injected weights and then superimposed onto the dynamic feature vector to form a fused feature. A priori calculation module based on material reaction dynamics is constructed. The mechanism prediction value output by the priori calculation module is multiplied by a learnable residual coefficient and then superimposed onto the decoding output of the fused feature.

4. The method for controlling and optimizing a material manufacturing process based on AI digital twins according to claim 1, characterized in that, Step S3, which involves dynamically calculating the noise suppression factor based on the statistical variance of the prediction error to correct the model output, specifically includes: Calculate the statistical variance of the difference between the actual observed values ​​and the model predicted values ​​within the sliding observation window; The statistical variance is mapped to a noise suppression factor with a value between zero and one using a sigmoid function, where the larger the statistical variance, the closer the noise suppression factor is to one. Calculate the average deviation within the sliding observation window, and multiply the average deviation by the difference between the average deviation and the noise suppression factor to obtain the deviation compensation amount; The model prediction value is added to the deviation compensation amount to obtain the corrected synchronous output value.

5. The method for controlling and optimizing a material manufacturing process based on AI digital twins according to claim 1, characterized in that, Step S3, which involves gated elastic updates of model parameters based on equipment health factors, specifically includes: Real-time acquisition of equipment vibration spectrum data and calculation of equipment health factor; Calculate the absolute error between the corrected synchronous output value and the actual observation value at the current time, and convert it into the basic update rate; If the device health factor is lower than the preset safety threshold, or the absolute error exceeds the preset anomaly cutoff threshold, the model parameter update rate will be forcibly set to zero. If the device health factor is higher than the safety threshold and the absolute error is within the allowable range, then the model parameters are updated using an exponential moving average based on the base update rate.

6. The method for controlling and optimizing a material manufacturing process based on AI digital twins according to claim 1, characterized in that, Step S4, which involves using a Bayesian optimization strategy to search for candidate process settings, specifically includes: Define process soft boundaries and hard boundaries. The soft boundaries are determined by the statistical distribution range of historical production data, and the hard boundaries are determined by the physical limits of the equipment. Within the search space formed by the soft boundary and the hard boundary, a surrogate model is constructed using Gaussian process regression; Based on the acquisition function, the next sampling point is selected in the search space, and the acquisition function is configured to balance the development of the current optimal value and the exploration of unknown areas; Only sampling points whose prediction variance in the model output is less than a preset uncertainty threshold are retained as candidate process settings.

7. The method for controlling and optimizing a material manufacturing process based on AI digital twins according to claim 1, characterized in that, The specific execution logic for the amplitude truncation check and rate limiting check in step S5 is as follows: Amplitude truncation verification: Compare the candidate process setting value with the maximum and minimum values ​​allowed by the equipment. If the candidate process setting value exceeds the range, it is forcibly assigned the corresponding boundary value. Rate limit verification: Calculate the difference between the candidate process setting value at the current moment and the execution value at the previous moment. If the absolute value of the difference exceeds the maximum rate of change allowed by the actuator, then reconstruct the setting value at the current moment based on the execution value at the previous moment and the maximum rate of change.

8. The method for controlling and optimizing a material manufacturing process based on AI digital twins according to claim 1, characterized in that, The shadow pattern verification and evolution described in step S6 specifically includes: When the accumulated amount of new sample data reaches the preset batch size, incremental learning training is started to generate candidate model parameters. The candidate model parameters are loaded into the shadow model, and real-time data is input for parallel inference without interfering with the control of the main model. The average absolute percentage error between the shadow model and the main model is calculated within a set period. The main model parameters are replaced with the candidate model parameters only when the average absolute percentage error of the shadow model is lower than that of the main model and no divergence alarm is triggered.

9. A material manufacturing process control and optimization system based on AI digital twins for implementing the method as described in any one of claims 1 to 8, characterized in that, include: The data governance module is configured to execute step S1, which completes the segmentation and alignment of the data. The digital twin engine is configured to perform steps S2 and S3, run a multimodal model containing physical residual terms, and perform adaptive correction based on statistical properties. The decision and security module is configured to execute steps S4 and S5, performing Bayesian optimization and kinematic safety interception. The model maintenance module is configured to execute step S6, which manages the version switching between the shadow model and the main model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.

Images