A machine learning-based digital filling control method and system

Abstract

The application discloses a kind of digital filling control method and system based on machine learning, including acquisition preprocessing module, for collecting multi-source data and completing preprocessing, generate fusion input data;VMamba prediction module, for outputting prediction distribution parameter and visual risk index;Sequential test module, for executing sliding window log-likelihood accumulation and observation dimension switching, output test determination result;Strategy compensation module, for generating filling control strategy and calculating compensation control amount, output control parameter set;Constraint execution module, for process verification and issue control instruction, realize real-time correction;Trace degradation module, for recording trace data and triggering safety degradation and archiving.The application realizes the real-time prediction and adaptive compensation control of filling amount by improving VMamba model and sequential probability ratio test.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to a digital filling control method and system based on machine learning. Background Technology

[0002] Existing filling equipment mostly employs open-loop control based on set formula parameters or conventional closed-loop control using weighing and liquid level as feedback. Target filling volume control is achieved through valve opening and closing sequence, pump speed curves, and segmented filling strategies. However, in actual continuous production, changes in material viscosity with temperature, fluctuations in feed pressure, pipeline bubbles and foam disturbances, differences in container specifications, and response lags in valves and pumps can cause dynamic changes in flow rate and liquid level, leading to increased filling volume deviation and decreased stability. Traditional methods relying on fixed threshold alarms or manual experience tuning are slow to respond to changes in operating conditions, making it difficult to achieve real-time prediction and adaptive compensation for filling endpoint deviations during the filling process. Consequently, under high-cycle and high-precision requirements, problems such as overfilling, underfilling, overflow, or frequent shutdowns are prone to occur.

[0003] Visual inspection and data acquisition platforms are being gradually introduced into production sites to improve quality monitoring and traceability capabilities. However, existing solutions typically treat visual inspection as an independent post-event judgment step, making it difficult to form a tightly coupled real-time decision-making mechanism with the filling control loop. Furthermore, multi-source data lacks a unified online judgment framework in terms of time alignment, feature fusion, anomaly detection, and strategy switching. Anomaly detection often relies on static thresholds or offline models, which cannot maintain stable false alarm and false negative levels under conditions of drift or changes in sensor noise. Some solutions lack process verification and traceability records for the control strategy generation process, making it difficult to ensure that control commands are executed within safety boundaries and to provide structured archiving of anomaly handling processes. Consequently, it is difficult to simultaneously meet the comprehensive requirements of high-precision control, operational safety, and digital traceability in continuous production scenarios.

[0004] Therefore, how to provide a digital filling control method and system based on machine learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a machine learning-based digital filling control method and system. This invention utilizes multi-source data acquisition, preprocessing, and fusion input to construct a system. It then employs an improved VMamba model to output filling volume prediction distribution parameters and visual risk indicators. A sequential probability ratio test is used to achieve sliding window log-likelihood accumulation and dynamic switching of observation dimensions, thereby generating filling control strategies and compensation control quantities. After process verification, control commands are issued, and full-process traceability and safety degradation archiving are implemented. The improved VMamba model incorporates enhanced directional state modeling, state update sensor gating, and high / low frequency channel structures, with distributed output. The sequential probability ratio test introduces a finite memory sliding window and dynamic switching of observation dimensions, enabling real-time prediction and adaptive compensation control under viscosity, temperature, pressure fluctuations, and equipment hysteresis disturbances. This results in a system with constrained, traceable, and safe and stable continuous production capabilities.

[0006] A machine learning-based digital filling control method according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect multi-source data from the filling process, preprocess the multi-source data and associate it with batch identifiers to generate fused input data;

[0008] S2. Construct an improved VMamba model, introduce directional state modeling to enhance the direction estimation and two-dimensional scanning modeling of the fused input data, perform parallel extraction and fusion of high-frequency edge features and low-frequency structural features based on high- and low-frequency channel structures, use state update sensing gating for state update, and generate predicted distribution parameters and visual risk indicators by outputting in a distribution form.

[0009] S3. Construct a sequential test observation sequence based on the predicted distribution parameters and visual risk indicators. Based on the sequential probability ratio test, introduce a finite memory sliding window to calculate the cumulative log-likelihood ratio of the sequential test observation sequence. Use dynamic switching of observation dimensions to switch the test observation dimensions and obtain the test judgment result.

[0010] S4. Generate a filling control strategy based on the inspection and judgment results, calculate the compensation control quantity, select the target strategy to calculate the filling control parameters, and obtain the set of control parameters.

[0011] S5. Perform process constraint verification on the set of control parameters and generate control commands, and issue control commands for real-time correction based on the verification results;

[0012] S6. Perform digital traceability and security downgrade handling, record traceability data throughout the entire process, and trigger security downgrade strategy when the inspection judgment result meets the abnormal conditions and archive the abnormal handling results.

[0013] Optionally, the modules can be interconnected using the following method: Optionally, the multi-source data includes filling image sequences and data on liquid level, flow rate, pressure, temperature, valve position, and pump speed.

[0014] Optionally, generating the fused input data includes:

[0015] Collect filling image sequences and liquid level, flow rate, pressure, temperature, valve position and pump speed data of the filling station during the filling cycle, write a unified timestamp for the filling image sequences and liquid level, flow rate, pressure, temperature, valve position and pump speed data and establish corresponding sampling records;

[0016] The filling image sequence is denoised, distorted and cropped to obtain images of key filling areas. The liquid level, flow rate, pressure, temperature, valve position and pump speed data are filtered and denoised, outlier removal and missing value filling to obtain sensor preprocessing data.

[0017] The images of key filling areas and the sensor preprocessing data are batch-identified and associated according to the batch number, formula number and container specifications. The images of key filling areas under the same time stamp are spliced ​​with the corresponding sensor preprocessing data to form a fusion input vector, and the fusion input data is obtained by arranging them in the order of the timestamps.

[0018] Optionally, the generation of predicted distribution parameters and visual risk indicators includes:

[0019] An improved VMamba model is constructed, including enhanced directional state modeling, the introduction of sensor gating into state updates, high and low frequency channel structure, and distributed output.

[0020] The system receives fused input data and performs feature encoding. It divides the image sequence of key filling areas in the fused input data into image blocks and maps them into image feature sequences. It aligns the liquid level, flow rate, pressure, temperature, valve position and pump speed data in the fused input data according to the timestamp to form a process signal sequence, and maps the process signal sequence into a gating input sequence.

[0021] Directional state modeling is used to enhance the direction estimation and two-dimensional scan modeling of image feature sequences. Directional features are extracted from the image features of each frame and the main direction is determined. The two-dimensional scan path and scan order are determined according to the main direction. The sequence unfolded according to the scan path is input into the state space unit to complete the state propagation and sequence feature update, and the directional scan feature sequence is obtained.

[0022] A high- and low-frequency channel structure is used to process the directional scanning feature sequence in parallel. The directional scanning feature sequence is decomposed into high-frequency components and low-frequency components, which are then input into the high-frequency state space branch and the low-frequency state space branch respectively to obtain high-frequency output features and low-frequency output features. The high-frequency output features and low-frequency output features are concatenated and linearly mapped to obtain a fused visual feature sequence.

[0023] By introducing sensor gating through state update, the state update of the fused visual feature sequence is performed by gating modulation. The gating input sequence is gating encoded to obtain the gating vector sequence. At each time step, the gating vector sequence is applied to the input projection vector of the state space unit and the state update amount to complete the gating modulation state update, and the gating state output sequence is obtained.

[0024] The output is distributed to generate filling volume prediction distribution parameters and visual risk indicators. The gating state output sequence is input into the distribution output head to obtain the filling volume prediction mean parameter and filling volume prediction variance parameter. The gating state output sequence is input into the risk output head to obtain the visual risk indicator sequence.

[0025] Optionally, obtaining the test result includes:

[0026] Construct a sequential test observation sequence. At each time step, calculate the difference between the filling volume feedback value and the filling volume prediction mean parameter, and normalize it with the square root of the filling volume prediction variance parameter to obtain the standardized deviation sequence. At each time step, calculate the difference between the visual risk indicators of two adjacent time steps to obtain the risk change sequence. Combine the standardized deviation sequence and the risk change sequence to form a sequential test observation sequence.

[0027] Set a first threshold and a second threshold for the sequential probability ratio test. The first threshold is the ratio of the probability of a Type II error to the complementary probability of a Type I error. The second threshold is the ratio of the complementary probability of a Type I error to the probability of a Type II error. The probability of a Type I error is defined as the probability of classifying a normal error as an abnormal error. The probability of a Type II error is defined as the probability of classifying an abnormal error as a normal error.

[0028] The cumulative log-likelihood ratio is calculated using a finite memory sliding window. A sliding window with a length equal to the length of the sliding window is selected. The ratio of the probability density under the abnormal hypothesis to the probability density under the normal hypothesis is calculated for each observation within the sliding window. The natural logarithm of the comparison values ​​is taken and summed to obtain the test statistic. The test statistic is updated with each time step.

[0029] The test results are output by dynamically switching the observation dimensions. The filling process is divided into a pre-filling stage, a slow filling stage, and an end-compensation stage. In the pre-filling stage, the test statistic is updated using the risk change sequence. In the slow filling stage, the test statistic is updated using the standardized deviation sequence. In the end-compensation stage, the test statistic is updated by switching between the risk change sequence and the standardized deviation sequence based on the duration for which the test statistic exceeds the switching threshold. The test statistic is compared with the first threshold and the second threshold, and then the normal judgment result, the abnormal judgment result, and the pending judgment result are output.

[0030] Optionally, the obtained control parameter set includes:

[0031] Receive the inspection and judgment results and generate a candidate set of filling control strategies. The candidate set of filling control strategies includes normal strategy, conservative strategy and degraded strategy. Normal strategy includes the phase sequence and phase parameters of pre-charge phase, slow charge phase and end compensation phase. Conservative strategy includes conservative parameters of pump speed limit, valve opening limit and slow charge switching point. Degraded strategy includes parameters of valve closing, pump stopping, filling pause and alarm flag.

[0032] Select the target strategy based on the test results. When the test results are normal, select the normal strategy; when the test results are pending, select the conservative strategy; when the test results are abnormal, select the downgrade strategy.

[0033] The difference between the predicted average filling volume parameter and the target filling volume is mapped to the valve opening and closing timing correction, pump speed curve correction, pre-charge and slow charge switching point correction, early shutdown compensation, and end-of-line compensation action parameters, and then combined to obtain the control parameter set.

[0034] Optionally, the step of performing process constraint verification on the control parameter set and generating control commands, and issuing control commands for real-time correction based on the verification results, includes:

[0035] Receive the set of control parameters and perform process constraint verification. The process constraint verification includes control quantity amplitude range verification, control quantity change rate verification, maximum end compensation count verification, and abnormal operating condition priority rule verification.

[0036] The control parameter set is modified and control commands are generated. When any verification item is not met, the corresponding control parameter is modified to meet the allowable range of the verification item and the control parameter set is updated. Control commands are generated based on the updated control parameter set. The control commands include valve opening and closing sequence commands, valve opening commands and pump speed curve commands.

[0037] The system issues control commands and makes real-time corrections. It sends control commands to valves, pumps and servo actuators to perform filling actions. During the filling process, it collects filling volume feedback values ​​and calculates the deviation between the filling volume feedback values ​​and the target filling volume. The system then updates the control parameter set based on the deviation.

[0038] Optionally, the digital traceability and security downgrade process includes:

[0039] Generate and record full-process traceability data, which includes batch identification information, visual risk indicators, inspection judgment results, control instructions, and deviations between the filling volume feedback value and the target filling volume, and write the full-process traceability data into the batch traceability file;

[0040] Based on the inspection and judgment results, a safety downgrade is triggered and archived. When the inspection and judgment results are abnormal, a downgrade control command is issued and the downgrade trigger time, trigger stage, control command issuance record and filling result deviation are recorded. When the inspection and judgment results are pending, a conservative control command is issued and the control parameter set correction record and filling result deviation are recorded.

[0041] A machine learning-based digital filling control system according to an embodiment of the present invention includes the following modules:

[0042] The data acquisition and preprocessing module is used to acquire multi-source data, perform preprocessing and batch association, and generate fused input data;

[0043] The VMamba prediction module is used to complete orientation scan modeling and high- and low-frequency feature fusion by fusing input data, and output predicted distribution parameters and visual risk indicators.

[0044] The sequential testing module is used to perform sliding window log-likelihood accumulation and observation dimension switching on the predicted distribution parameters and visual risk indicators, and output the test judgment results.

[0045] The strategy compensation module is used to generate a filling control strategy based on the inspection and judgment results, calculate the compensation control quantity, and output a set of control parameters.

[0046] The constraint execution module is used to perform process verification on the set of control parameters and issue control commands to achieve real-time correction.

[0047] The traceability and degradation module is used to record full-process traceability data and trigger security degradation and archiving.

[0048] The beneficial effects of this invention are:

[0049] This invention collects and fuses multi-source data, including filling image sequences, liquid level, flow rate, pressure, temperature, valve position, and pump speed, into unified fused input data. This enhances the ability to acquire information, align timing, and construct usable features during the filling process. Compared to existing control methods that rely on fixed formula parameters, manual tuning, or single feedback signals, this invention uses an improved VMamba model to output filling volume prediction distribution parameters and visual risk indicators. Furthermore, it strengthens the characterization of complex factors such as liquid surface edges, foam disturbances, and equipment response lag through directional state modeling enhancement, high- and low-frequency channel structures, and state update sensor gating. This enables more stable real-time prediction and adaptive compensation control under conditions of viscosity, temperature, and pressure fluctuations, reducing the risks of overfilling, underfilling, and overflow, and improving filling accuracy and consistency.

[0050] This invention employs a sequential probability ratio test to accumulate the log-likelihood of the observation sequence constructed from predicted distribution parameters and visual risk indicators using a sliding window. By dynamically switching observation dimensions to cover risk characteristics at different filling stages, it improves the timeliness of early anomaly identification, operational condition drift detection, and strategy switching, reducing the impact of false alarms and missed alarms on production line cycle time. Through process constraint verification to generate and issue control commands, and real-time correction of the control parameter set during filling, while simultaneously recording full-process traceability data and triggering safety degradation and archiving when abnormal conditions are met, this invention achieves constrainable control strategy generation, monitorable control execution, and traceable anomaly handling, significantly enhancing the stability, safety, and engineering feasibility of continuous production. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is a flowchart of a machine learning-based digital filling control system proposed in this invention;

[0053] Figure 2 This is a structural block diagram of a machine learning-based digital filling control method proposed in this invention.

[0054] Figure 3 This is a functional diagram of the improved VMamba model of the machine learning-based digital filling control method proposed in this invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0056] refer to Figure 1 A machine learning-based digital filling control system includes the following modules:

[0057] The data acquisition and preprocessing module is used to acquire multi-source data, perform preprocessing and batch association, and generate fused input data;

[0058] The VMamba prediction module is used to complete orientation scan modeling and high- and low-frequency feature fusion by fusing input data, and output predicted distribution parameters and visual risk indicators.

[0059] The sequential testing module is used to perform sliding window log-likelihood accumulation and observation dimension switching on the predicted distribution parameters and visual risk indicators, and output the test judgment results.

[0060] The strategy compensation module is used to generate a filling control strategy based on the inspection and judgment results, calculate the compensation control quantity, and output a set of control parameters.

[0061] The constraint execution module is used to perform process verification on the set of control parameters and issue control commands to achieve real-time correction.

[0062] The traceability and degradation module is used to record full-process traceability data and trigger security degradation and archiving.

[0063] refer to Figure 2 and Figure 3 A machine learning-based digital filling control method includes:

[0064] Collect multi-source data from the filling process, preprocess the multi-source data and associate it with batch identifiers to generate fused input data;

[0065] An improved VMamba model is constructed, and directional state modeling is introduced to enhance the direction estimation and two-dimensional scanning modeling of the fused input data. Based on the high- and low-frequency channel structure, high-frequency edge features and low-frequency structural features are extracted and fused in parallel. State update sensing gating is used for state update. Predictive distribution parameters and visual risk indicators are generated by outputting a distribution.

[0066] Based on the predicted distribution parameters and visual risk indicators, a sequential test observation sequence is constructed. Based on the sequential probability ratio test, a finite memory sliding window is introduced to calculate the cumulative log-likelihood ratio of the sequential test observation sequence. The test observation dimension is switched dynamically to obtain the test judgment result.

[0067] The filling control strategy is generated by verifying the judgment results, the compensation control quantity is calculated, the target strategy is selected to calculate the filling control parameters, and the set of control parameters is obtained.

[0068] Perform process constraint verification on the set of control parameters and generate control commands, and issue control commands for real-time correction based on the verification results;

[0069] Implement digital traceability and security degradation measures, record traceability data throughout the entire process, trigger security degradation strategies when the inspection and judgment results meet abnormal conditions, and archive the abnormal handling results.

[0070] In this embodiment, the multi-source data includes filling image sequences as well as data on liquid level, flow rate, pressure, temperature, valve position, and pump speed.

[0071] In this embodiment, generating the fused input data includes:

[0072] Collect filling image sequences and liquid level, flow rate, pressure, temperature, valve position and pump speed data of the filling station during the filling cycle, write a unified timestamp for the filling image sequences and liquid level, flow rate, pressure, temperature, valve position and pump speed data and establish corresponding sampling records;

[0073] The filling image sequence is denoised, distorted and cropped to obtain images of key filling areas. The liquid level, flow rate, pressure, temperature, valve position and pump speed data are filtered and denoised, outlier removal and missing value filling to obtain sensor preprocessing data.

[0074] The images of key filling areas and the sensor preprocessing data are batch-identified and associated according to the batch number, formula number and container specifications. The images of key filling areas under the same time stamp are spliced ​​with the corresponding sensor preprocessing data to form a fusion input vector, and the fusion input data is obtained by arranging them in the order of the timestamps.

[0075] In this embodiment, the generation of predicted distribution parameters and visual risk indicators includes:

[0076] An improved VMamba model is constructed, including enhanced directional state modeling, the introduction of sensor gating into state updates, high- and low-frequency channel structures, and a distributed output, wherein:

[0077] When constructing the improved VMamba model, a directional state modeling enhancement unit is set in the visual state space backbone. The main direction is estimated and the two-dimensional scanning path and state propagation order are determined for the image sequence of the key filling area to improve the spatiotemporal modeling stability under conditions such as liquid surface tilt and splashing. High and low frequency channel structures are connected in parallel in the backbone to decompose the image features into high frequency edge information and low frequency structural information, which are then processed in the state space branch and fused to enhance the extraction of liquid level line and foam boundary details and suppress background interference.

[0078] In the state update process, sensor gating is introduced to encode process signals such as liquid level, flow rate, pressure, temperature, valve position and pump speed into gating vectors and apply them to the state update quantity. This enables dynamic modulation that changes with operating conditions to adapt to viscosity, pressure fluctuations and equipment lag. At the output end, a distributed output head is adopted, and the predicted mean and predicted variance are generated and a visual risk index is output. This allows the prediction results to carry uncertainty to support subsequent sequential verification and conservative compensation control.

[0079] The system receives fused input data and performs feature encoding. It divides the image sequence of key filling areas in the fused input data into image blocks and maps them to image feature sequences. It aligns the liquid level, flow rate, pressure, temperature, valve position, and pump speed data from the fused input data according to timestamps to form a process signal sequence, and maps this process signal sequence to a gated input sequence. Where:

[0080] The step of dividing the image sequence of the key filling area in the fused input data into image blocks and mapping it into an image feature sequence is as follows: each frame of the key filling area image is divided into grids according to a preset block size to obtain several non-overlapping overlapping image blocks. After pixel normalization and size alignment of each image block, the image block is mapped into a fixed-dimensional vector representation through a linear projection layer, and the image blocks are arranged according to the spatial order of the image blocks in the original image to form the image feature sequence of the current frame. The process is repeated for multiple consecutive frames and stacked in chronological order to obtain an image feature sequence consistent with the timestamp. The preset block size is set to sixteen by sixteen pixels.

[0081] The process signal sequence is mapped to a gated input sequence as follows: First, the liquid level, flow rate, pressure, temperature, valve position, and pump speed data are aligned according to the timestamp. At each time step, the values ​​of each channel are sequentially concatenated into a process signal vector. After unifying the dimensions and normalizing the values ​​of the process signal vector, it is input into the gated encoding to obtain the gated input vector. The gated encoding consists of a fully connected layer and a nonlinear activation layer and outputs a vector that matches the state update dimension. The same mapping is performed at each time step and arranged in chronological order to form a gated input sequence.

[0082] Directional state modeling is used to enhance the direction estimation and two-dimensional scan modeling of image feature sequences. Directional features are extracted from each frame of image features, and the principal direction is determined. The two-dimensional scan path and scan order are then determined according to the principal direction. The sequence expanded along the scan path is input into the state space unit to complete state propagation and sequence feature update, resulting in the directional scan feature sequence, where:

[0083] The step of extracting directional features from each frame of image features and determining the main direction is as follows: the current frame image features are restored into a two-dimensional feature map according to the spatial grid; the feature difference between adjacent positions in the two-dimensional feature map is calculated in the horizontal and vertical directions to obtain the gradient vector field; the cumulative value of the gradient magnitude is statistically calculated in multiple candidate directions to form a directional response; and the directional response with the largest value is selected from the candidate directions as the main direction.

[0084] The determination of the two-dimensional scanning path and scanning order according to the main direction is as follows:

[0085] Three scanning templates are defined. Each scanning template includes a starting boundary, a direction of advancement, and a traversal rule for each row or column. When the main direction is vertical, the scanning is performed row by row from top to bottom. When the main direction is horizontal, the scanning is performed column by column from left to right. When the main direction is diagonal, the scanning is performed diagonally along the main diagonal. The scanning order advances step by step from the starting boundary to the ending boundary. At each advancement, the access order of the current row or column is determined according to the template.

[0086] The process of inputting the sequence expanded according to the scanning path into the state space unit to complete state propagation and sequence feature update is as follows: extract the feature vectors at the corresponding positions in the scanning order and connect them into a one-dimensional sequence. Set the initial state of the state space unit to the zero vector or the last state of the previous frame. Perform input projection on the feature vector of each time step in the one-dimensional sequence to obtain the driving quantity, and input it together with the current state into the state update operator to obtain the next state. Then, project the next state through the output to obtain the updated feature vector. Repeat the update for all time steps of the sequence and backfill according to the original spatial index to obtain the directional scanning feature sequence.

[0087] A high- and low-frequency channel structure is used to process the directional scanning feature sequence in parallel. The directional scanning feature sequence is decomposed into high-frequency components and low-frequency components, which are respectively input into the high-frequency state space branch and the low-frequency state space branch to obtain high-frequency output features and low-frequency output features. The high-frequency output features and low-frequency output features are concatenated and linearly mapped to obtain a fused visual feature sequence, wherein:

[0088] The process of decomposing the directional scanning feature sequence into high-frequency and low-frequency components is as follows: the directional scanning feature sequence is backfilled into a two-dimensional feature map according to the spatial index; a local smoothing operation is performed on the two-dimensional feature map to obtain a low-frequency feature map; the local smoothing operation uses a Gaussian filter of a fixed size; the low-frequency feature map is then subtracted from the two-dimensional feature map corresponding to the directional scanning feature sequence to obtain a high-frequency feature map; the high-frequency feature map and the low-frequency feature map are expanded into a high-frequency component sequence and a low-frequency component sequence respectively according to the original scanning order.

[0089] The process of obtaining high-frequency output features and low-frequency output features by inputting high-frequency state space branches and low-frequency state space branches respectively is as follows: the feature vectors of the high-frequency component sequence are taken sequentially according to the scanning order, the driving vector is obtained by input projection, the current branch state is obtained by inputting the branch state of the previous time step into the state update unit, the high-frequency output feature vector is obtained by output projection, and the high-frequency output feature sequence is formed by splicing them in time order. The low-frequency component sequence is processed in the same recursive way to obtain the low-frequency output feature sequence. The high-frequency state space branch and the low-frequency state space branch use mutually independent input projection, state update and output projection parameters, and output the output feature sequence with the same length as the input sequence respectively.

[0090] By introducing sensor gating through state update, the state update of the fused visual feature sequence is performed by gating modulation. The gating input sequence is gating encoded to obtain the gating vector sequence. At each time step, the gating vector sequence is applied to the input projection vector of the state space unit and the state update amount to complete the gating modulation state update, and the gating state output sequence is obtained.

[0091] The output is distributed to generate filling volume prediction distribution parameters and visual risk indicators. The gated output sequence is input into the distribution output head to obtain the filling volume prediction mean parameter and filling volume prediction variance parameter. The gated output sequence is input into the risk output head to obtain the visual risk indicator sequence, where:

[0092] The distributed output head is specifically composed of a linear mapping layer and outputs two sets of scalar parameters at each time step. One set is used as the mean parameter for filling quantity prediction, and the other set is used as the variance parameter for filling quantity prediction. The variance parameter output is nonnegated by performing an exponential operation on the output of the linear mapping layer to obtain a nonnegative value, thereby obtaining the filling quantity prediction mean parameter sequence and the filling quantity prediction variance parameter sequence corresponding to the time step.

[0093] The risk output head is specifically composed of a linear mapping layer and an activation mapping. It outputs a visual risk score at each time step. The activation mapping maps the output of the linear mapping layer to the interval between zero and one through a logic function to obtain a visual risk index sequence corresponding to the time step.

[0094] In this embodiment, obtaining the test result includes:

[0095] Construct a sequential test observation sequence. At each time step, calculate the difference between the filling volume feedback value and the filling volume prediction mean parameter, and normalize it with the square root of the filling volume prediction variance parameter to obtain the standardized deviation sequence. At each time step, calculate the difference between the visual risk indicators of two adjacent time steps to obtain the risk change sequence. Combine the standardized deviation sequence and the risk change sequence to form a sequential test observation sequence.

[0096] Set a first threshold and a second threshold for the sequential probability ratio test. The first threshold is the ratio of the probability of a Type II error to the complementary probability of a Type I error. The second threshold is the ratio of the complementary probability of a Type I error to the probability of a Type II error. The probability of a Type I error is defined as the probability of classifying a normal error as an abnormal error. The probability of a Type II error is defined as the probability of classifying an abnormal error as a normal error.

[0097] The cumulative log-likelihood ratio is calculated using a finite-memory sliding window. A sliding window of length equal to the window length is selected. For each observation within the sliding window, the ratio of the probability density under the anomaly hypothesis to the probability density under the normal hypothesis is calculated. The natural logarithm of the comparison values ​​is taken and summed to obtain the test statistic, which is updated with each time step. Specifically, the calculation of the ratio of the probability density under the anomaly hypothesis to the probability density under the normal hypothesis for each observation within the sliding window is as follows:

[0098] For the standardized deviation series and the risk change series, respectively set the probability distribution parameters of the normal hypothesis and the probability distribution parameters of the abnormal hypothesis. For each observation value in the sliding window, select the corresponding two sets of distribution parameters according to the observation dimension. Substitute the observation value into the probability density function of the abnormal hypothesis to obtain the probability density value of the abnormal hypothesis. Substitute the observation value into the probability density function of the normal hypothesis to obtain the probability density value of the normal hypothesis. Calculate the ratio of the probability density value of the abnormal hypothesis to the probability density value of the normal hypothesis as the likelihood ratio of the current observation value.

[0099] For standardized deviation sequences, the normal hypothesis is set to a Gaussian distribution with a mean of zero and a variance of one, and the outlier hypothesis is set to a Gaussian distribution with a mean of 1.5 and a variance of one. For risk change sequences, the normal hypothesis is set to a Gaussian distribution with a mean of zero and a variance of 0.25, and the outlier hypothesis is set to a Gaussian distribution with a mean of 0.8 and a variance of 0.25. The corresponding probability density functions for both the normal and outlier hypotheses are Gaussian probability density functions. The independent variables are the observed values, and the parameters are the mean and variance. The mean and variance are obtained from historical normal batch data and updated according to batch identifiers.

[0100] The system employs dynamic switching of observation dimensions to output test results. The filling process is divided into a pre-filling stage, a slow-filling stage, and a final compensation stage. In the pre-filling stage, the test statistic is updated using a risk change sequence. In the slow-filling stage, the test statistic is updated using a standardized deviation sequence. In the final compensation stage, the test statistic is updated by switching between the risk change sequence and the standardized deviation sequence based on the duration for which the test statistic continuously exceeds the switching threshold. The test statistic is then compared with the first and second thresholds to output normal, abnormal, and pending results.

[0101] The filling process is divided into a pre-filling stage, a slow-filling stage, and an end-compensation stage. Specifically, the pre-filling stage is defined as the time period during which the valve opening or pump speed is within the set value range, and the slow-filling stage is defined as the time period during which the valve opening or pump speed switches to the slow-filling set value range and the filling volume does not reach the target filling volume. The end-compensation stage is defined as the time period during which the filling volume reaches the target filling volume minus the end-compensation trigger value and enters the pulse-type replenishment or micro-compensation command mode.

[0102] The duration of the switching threshold is specifically as follows: the duration is three consecutive time steps. When the test statistic is greater than the switching threshold for three consecutive time steps, the test observation dimension is switched from the current sequence to another sequence and the test statistic is updated with the switched observation dimension. When the test statistic is not greater than the switching threshold for three consecutive time steps, the current test observation dimension remains unchanged.

[0103] The step of comparing the test statistic with the first threshold and the second threshold and then outputting a normal judgment result, an abnormal judgment result, and a pending judgment result specifically involves: at each time step, comparing the test statistic with the first threshold and the second threshold; outputting a normal judgment result when the test statistic is less than or equal to the first threshold; outputting an abnormal judgment result when the test statistic is greater than or equal to the second threshold; and outputting a pending judgment result when the test statistic is greater than the first threshold and less than the second threshold.

[0104] In this embodiment, obtaining the control parameter set includes:

[0105] The process involves receiving the inspection and judgment results and generating a candidate set of filling control strategies. This candidate set includes a normal strategy, a conservative strategy, and a degraded strategy. The normal strategy includes the phase sequence and parameters of the pre-charging phase, the slow charging phase, and the end-compensation phase. The conservative strategy includes conservative parameters for the pump speed limit, the valve opening limit, and the slow charging switching point. The degraded strategy includes parameters for closing the valve, stopping the pump, pausing filling, and alarm flags. Specifically, receiving the inspection and judgment results and generating the candidate set of filling control strategies involves:

[0106] Read the current time step's determination type and the corresponding filling stage identifier; read the normal strategy to obtain the stage sequence of the pre-charging stage, slow charging stage, and end compensation stage, as well as the valve opening and closing sequence, pump speed setting, and stage duration parameters for each stage; read the conservative strategy and generate pump speed limit, valve opening limit, and slow charging switching point parameters based on the normal strategy template; read the degradation strategy to generate valve closing command, pump stop command, pause filling command, and alarm flag parameters.

[0107] Select the target strategy based on the test results. When the test results are normal, select the normal strategy; when the test results are pending, select the conservative strategy; when the test results are abnormal, select the downgrade strategy.

[0108] The difference between the predicted average filling volume parameter and the target filling volume is mapped to valve opening and closing timing correction, pump speed curve correction, pre-charge and slow charge switching point correction, early shutdown compensation, and end-of-line compensation action parameters. These are then combined to obtain a set of control parameters. Specifically, the combined set of control parameters is as follows:

[0109] Using the filling control parameters corresponding to the target strategy as the baseline parameter set, the parameters are written into the valve opening and closing sequence parameter group, pump speed curve parameter group, pre-charge and slow charge switching point parameter group, early shutdown compensation parameter group, and end-of-line compensation action parameter group according to parameter type. Then, each correction value generated by the difference between the predicted average filling volume parameter and the target filling volume is superimposed onto the corresponding parameter group according to the parameter key value with the same name to form the updated parameter value. The updated parameter group is aligned with the time axis and the stage consistency is sorted to generate a structured parameter list containing parameter name, value, action stage and effective time. The structured parameter list is used as the control parameter set.

[0110] In this embodiment, the step of performing process constraint verification on the control parameter set and generating control commands, and issuing control commands for real-time correction based on the verification results, includes:

[0111] The process involves receiving a set of control parameters and performing process constraint verification. This verification includes verifying the control quantity amplitude range, the control quantity change rate, the maximum number of end-point compensations, and the priority rule verification for abnormal operating conditions. Specifically, the process constraint verification involves:

[0112] Read the valve opening degree, pump speed, valve opening and closing sequence, stage switching point and end compensation action parameters from the control parameter set, perform rule calculation and comparison for each item, determine whether the valve opening degree parameter falls between 10% and 95%, and determine whether the pump speed parameter falls between zero revolutions per minute and 1,500 revolutions per minute. Based on the time step interval of 0.1 seconds, calculate the valve opening degree change amount of adjacent time steps, divide it by 0.1 seconds to obtain the valve opening degree change rate and determine whether it is not greater than 200 percentage points per second. Calculate the pump speed change amount of adjacent time steps, divide it by 0.1 seconds to obtain the pump speed change rate and determine whether it is not greater than 5,000 revolutions per minute.

[0113] The system checks whether the number of compensations in the end-of-line compensation action parameters is no more than three. At the same time, it reads the strategy type identifier corresponding to the verification result. When the strategy type is a downgraded strategy, only the parameters for closing the valve, stopping the pump, and pausing filling are retained. When the strategy type is a conservative strategy, only the conservative parameters are retained and the normal strategy parameters are set to invalid. The system outputs the verification pass flag and the set of control parameters after verification.

[0114] The control parameter set is modified and control commands are generated. When any verification item is not met, the corresponding control parameter is modified to meet the allowable range of the verification item and the control parameter set is updated. Control commands are generated based on the updated control parameter set. The control commands include valve opening and closing sequence commands, valve opening commands and pump speed curve commands.

[0115] The system issues control commands and makes real-time corrections. It sends control commands to valves, pumps and servo actuators to perform filling actions. During the filling process, it collects filling volume feedback values ​​and calculates the deviation between the filling volume feedback values ​​and the target filling volume. The system then updates the control parameter set based on the deviation.

[0116] In this embodiment, the digital traceability and security downgrade process includes:

[0117] Generate and record full-process traceability data, which includes batch identification information, visual risk indicators, inspection judgment results, control instructions, and deviations between the filling volume feedback value and the target filling volume, and write the full-process traceability data into the batch traceability file;

[0118] Based on the inspection and judgment results, a safety downgrade is triggered and archived. When the inspection and judgment results are abnormal, a downgrade control command is issued and the downgrade trigger time, trigger stage, control command issuance record and filling result deviation are recorded. When the inspection and judgment results are pending, a conservative control command is issued and the control parameter set correction record and filling result deviation are recorded.

[0119] Example 1:

[0120] To verify the feasibility of this invention in practice, it was applied to a high-speed beverage bottling line in an industrial park. The bottling target was 500 ml PET bottled sugary beverages, with a filling volume of 520 grams. During continuous production, the line experienced typical disturbances: material temperature fluctuations between 20 and 32 degrees Celsius caused viscosity changes; feed pressure fluctuations between 0.2 and 0.6 MPa caused flow drift; valves and pumps exhibited response lag; and surface foam and splashing made the visual liquid level boundary unstable. When these factors combined, overfilling, underfilling, overflow, and cycle time fluctuations were likely to occur. Furthermore, traditional fixed threshold alarms often produced false alarms or delayed alarms, affecting the stability and safety of continuous production.

[0121] The system of this invention is deployed in the production line to collect image sequences of key filling areas from cameras, along with data on liquid level, flow rate, pressure, temperature, valve position, and pump speed, and generate fused input data. An improved VMamba model is used to perform orientation estimation and two-dimensional scanning modeling on the fused input data. High-frequency edge and low-frequency structural features are extracted and fused in parallel, while the sensor-gated modulation state is updated, outputting predicted filling volume distribution parameters and visual risk indicators. The predicted distribution parameters and visual risk indicators are constructed into a sequential test observation sequence. A sequential probability ratio test is used for finite-memory sliding window log-likelihood accumulation, and the observation dimensions are dynamically switched during the pre-filling, slow-filling, and end-compensation stages to obtain the test judgment results. Based on the test judgment results, a normal, conservative, or degraded strategy is selected, and compensation control quantities are calculated to generate a set of control parameters. After process verification, control commands are issued and corrected in real time during the filling process. Simultaneously, full-process traceability data is recorded, and safety degradation and archiving are triggered when abnormal conditions are met.

[0122] During periods of high volatility due to combined temperature and pressure disturbances, the system of this invention can identify deviation trends caused by sudden increases in foam and valve / pump lag in advance during the filling process and automatically switch to a conservative or degraded strategy, preventing continuous overflow and batch deviations. Compared with the original on-site control scheme, the fluctuation of filling volume is significantly reduced, the number of reworked bottles caused by overfilling and underfilling is significantly reduced, the production line cycle time is more stable, and the handling of abnormal triggers is more traceable. Furthermore, each strategy switch, control parameter set, and execution command forms a batch file, facilitating quality review and continuous process optimization.

[0123] Table 1 Summary of Comparison Indicators for Filling Accuracy and Stability

[0124]

[0125] As shown in Table 1, in terms of filling accuracy-related indicators, the method of this invention has the most significant advantages in both mean absolute error and three sigma error. The mean absolute error is 0.85g, a decrease of 64.6% compared to 2.40g for threshold + PID, a decrease of 56.4% compared to 1.95g for weighing-only PID, and a decrease of 70.2% compared to 2.85g for level-only closed-loop control. The three sigma error is 2.90g, a decrease of 62.8% compared to 7.8g for threshold + PID, and a decrease of 66.3% compared to 8.60g for level-only closed-loop control. Compared to the two types of vision solutions, this invention is also more stable. Compared to 1.70g for CNN + threshold and 1.45g for ViT + threshold, this invention reduces errors by 50.0% and 41.4% respectively, indicating that prediction and compensation control contribute more directly to error convergence under disturbance conditions.

[0126] From the perspective of quality and risk-related indicators, this invention achieves the lowest or near-lowest levels in over-irrigation rate, under-irrigation rate, and overflow rate: 0.18% for over-irrigation, 0.22% for under-irrigation, and 0.05% for overflow. The overflow rate is 86.8% lower than the threshold + PID (0.38%), 83.9% lower than the weighing-only PID (0.31%), and 89.1% lower than the level-only closed-loop (0.46%). It also represents a 75.0% and 68.8% reduction compared to the CNN + threshold (0.20%) and the ViT + threshold (0.16%), respectively. This indicates that relying solely on visual alarms with fixed thresholds or a single feedback closed loop is insufficient to simultaneously reduce deviations and overflow to lower levels under conditions of overlapping factors such as foaming, pressure fluctuations, and execution lag.

[0127] From the perspective of continuous production stability indicators, this invention demonstrates outstanding performance in cycle time stability, anomaly lead time, and number of downtimes. The standard deviation of cycle time stability is 0.9 bottles / min, a 67.9% reduction compared to 2.8 for threshold + PID, and a 71.9% reduction compared to 3.2 for level-only closed-loop control. The anomaly lead time reaches 3.6 seconds, higher than 2.2 seconds for ViT + threshold and 1.9 seconds for CNN + threshold, and significantly higher than 0.8 seconds for threshold + PID, indicating earlier identification of abnormal trends and more sufficient buffer for strategy switching and degradation. The number of downtimes in eight hours is 0.5, a 78.3% reduction compared to 2.3 for threshold + PID, and a 72.2% reduction compared to 1.8 for weighing-only PID. In summary, this invention achieves superior results in accuracy, safety, and continuity simultaneously, with its advantages being more pronounced in disturbance-sensitive indicators such as overflow, cycle time, and downtime.

[0128] 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 digital filling control method based on machine learning, characterized in that, Includes the following steps: S1. Collect multi-source data from the filling process, preprocess the multi-source data and associate it with batch identifiers to generate fused input data; S2. Construct an improved VMamba model, introduce directional state modeling to enhance the direction estimation and two-dimensional scanning modeling of the fused input data, perform parallel extraction and fusion of high-frequency edge features and low-frequency structural features based on high- and low-frequency channel structures, use state update sensing gating for state update, and generate predicted distribution parameters and visual risk indicators by outputting in a distribution form. S3. Construct a sequential test observation sequence based on the predicted distribution parameters and visual risk indicators. Based on the sequential probability ratio test, introduce a finite memory sliding window to calculate the cumulative log-likelihood ratio of the sequential test observation sequence. Use dynamic switching of observation dimensions to switch the test observation dimensions and obtain the test judgment result. S4. Generate a filling control strategy based on the inspection and judgment results, calculate the compensation control quantity, select the target strategy to calculate the filling control parameters, and obtain the set of control parameters. S5. Perform process constraint verification on the set of control parameters and generate control commands, and issue control commands for real-time correction based on the verification results; S6. Perform digital traceability and security downgrade handling, record traceability data throughout the entire process, and trigger security downgrade strategy when the inspection judgment result meets the abnormal conditions and archive the abnormal handling results.

2. The machine learning-based digital filling control method according to claim 1, characterized in that, The multi-source data includes filling image sequences as well as data on liquid level, flow rate, pressure, temperature, valve position, and pump speed.

3. The machine learning-based digital filling control method according to claim 2, characterized in that, The generation of fused input data includes: Collect filling image sequences and liquid level, flow rate, pressure, temperature, valve position and pump speed data of the filling station during the filling cycle, write a unified timestamp for the filling image sequences and liquid level, flow rate, pressure, temperature, valve position and pump speed data and establish corresponding sampling records; The filling image sequence is denoised, distorted and cropped to obtain images of key filling areas. The liquid level, flow rate, pressure, temperature, valve position and pump speed data are filtered and denoised, outlier removal and missing value filling to obtain sensor preprocessing data. The images of key filling areas and the sensor preprocessing data are batch-identified and associated according to the batch number, formula number and container specifications. The images of key filling areas under the same time stamp are spliced ​​with the corresponding sensor preprocessing data to form a fusion input vector, and the fusion input data is obtained by arranging them in the order of the timestamps.

4. The machine learning-based digital filling control method according to claim 2, characterized in that, The generation of predicted distribution parameters and visual risk indicators includes: An improved VMamba model is constructed, including enhanced directional state modeling, the introduction of sensor gating into state updates, high and low frequency channel structure, and distributed output. The system receives fused input data and performs feature encoding. It divides the image sequence of key filling areas in the fused input data into image blocks and maps them into image feature sequences. It aligns the liquid level, flow rate, pressure, temperature, valve position and pump speed data in the fused input data according to the timestamp to form a process signal sequence, and maps the process signal sequence into a gating input sequence. Directional state modeling is used to enhance the direction estimation and two-dimensional scan modeling of image feature sequences. Directional features are extracted from the image features of each frame and the main direction is determined. The two-dimensional scan path and scan order are determined according to the main direction. The sequence unfolded according to the scan path is input into the state space unit to complete the state propagation and sequence feature update, and the directional scan feature sequence is obtained. A high- and low-frequency channel structure is used to process the directional scanning feature sequence in parallel. The directional scanning feature sequence is decomposed into high-frequency components and low-frequency components, which are then input into the high-frequency state space branch and the low-frequency state space branch respectively to obtain high-frequency output features and low-frequency output features. The high-frequency output features and low-frequency output features are concatenated and linearly mapped to obtain a fused visual feature sequence. By introducing sensor gating through state update, the state update of the fused visual feature sequence is performed by gating modulation. The gating input sequence is gating encoded to obtain the gating vector sequence. At each time step, the gating vector sequence is applied to the input projection vector of the state space unit and the state update amount to complete the gating modulation state update, and the gating state output sequence is obtained. The output is distributed to generate filling volume prediction distribution parameters and visual risk indicators. The gating state output sequence is input into the distribution output head to obtain the filling volume prediction mean parameter and filling volume prediction variance parameter. The gating state output sequence is input into the risk output head to obtain the visual risk indicator sequence.

5. The machine learning-based digital filling control method according to claim 2, characterized in that, The obtained test and judgment results include: Construct a sequential test observation sequence. At each time step, calculate the difference between the filling volume feedback value and the filling volume prediction mean parameter, and normalize it with the square root of the filling volume prediction variance parameter to obtain the standardized deviation sequence. At each time step, calculate the difference between the visual risk indicators of two adjacent time steps to obtain the risk change sequence. Combine the standardized deviation sequence and the risk change sequence to form a sequential test observation sequence. Set a first threshold and a second threshold for the sequential probability ratio test. The first threshold is the ratio of the probability of a Type II error to the complementary probability of a Type I error. The second threshold is the ratio of the complementary probability of a Type I error to the probability of a Type II error. The probability of a Type I error is defined as the probability of classifying a normal error as an abnormal error. The probability of a Type II error is defined as the probability of classifying an abnormal error as a normal error. The cumulative log-likelihood ratio is calculated using a finite memory sliding window. A sliding window with a length equal to the length of the sliding window is selected. The ratio of the probability density under the abnormal hypothesis to the probability density under the normal hypothesis is calculated for each observation within the sliding window. The natural logarithm of the comparison values ​​is taken and summed to obtain the test statistic. The test statistic is updated with each time step. The test results are output by dynamically switching the observation dimensions. The filling process is divided into a pre-filling stage, a slow filling stage, and an end-compensation stage. In the pre-filling stage, the test statistic is updated using the risk change sequence. In the slow filling stage, the test statistic is updated using the standardized deviation sequence. In the end-compensation stage, the test statistic is updated by switching between the risk change sequence and the standardized deviation sequence based on the duration for which the test statistic exceeds the switching threshold. The test statistic is compared with the first threshold and the second threshold, and then the normal judgment result, the abnormal judgment result, and the pending judgment result are output.

6. The machine learning-based digital filling control method according to claim 2, characterized in that, The obtained set of control parameters includes: Receive the inspection and judgment results and generate a candidate set of filling control strategies. The candidate set of filling control strategies includes normal strategy, conservative strategy and degraded strategy. Normal strategy includes the phase sequence and phase parameters of pre-charge phase, slow charge phase and end compensation phase. Conservative strategy includes conservative parameters of pump speed limit, valve opening limit and slow charge switching point. Degraded strategy includes parameters of valve closing, pump stopping, filling pause and alarm flag. Select the target strategy based on the test results. When the test results are normal, select the normal strategy; when the test results are pending, select the conservative strategy; when the test results are abnormal, select the downgrade strategy. The difference between the predicted average filling volume parameter and the target filling volume is mapped to the valve opening and closing timing correction, pump speed curve correction, pre-charge and slow charge switching point correction, early shutdown compensation, and end-of-line compensation action parameters, and then combined to obtain the control parameter set.

7. The machine learning-based digital filling control method according to claim 2, characterized in that, The process of performing process constraint verification on the control parameter set and generating control commands, and issuing control commands for real-time correction based on the verification results, includes: Receive the set of control parameters and perform process constraint verification. The process constraint verification includes control quantity amplitude range verification, control quantity change rate verification, maximum end compensation count verification, and abnormal operating condition priority rule verification. The control parameter set is modified and control commands are generated. When any verification item is not met, the corresponding control parameter is modified to meet the allowable range of the verification item and the control parameter set is updated. Control commands are generated based on the updated control parameter set. The control commands include valve opening and closing sequence commands, valve opening commands and pump speed curve commands. The system issues control commands and makes real-time corrections. It sends control commands to valves, pumps and servo actuators to perform filling actions. During the filling process, it collects filling volume feedback values ​​and calculates the deviation between the filling volume feedback values ​​and the target filling volume. The system then updates the control parameter set based on the deviation.

8. The machine learning-based digital filling control method according to claim 2, characterized in that, The aforementioned digital traceability and security downgrade procedures include: Generate and record full-process traceability data, which includes batch identification information, visual risk indicators, inspection judgment results, control instructions, and deviations between the filling volume feedback value and the target filling volume, and write the full-process traceability data into the batch traceability file; Based on the inspection and judgment results, a safety downgrade is triggered and archived. When the inspection and judgment results are abnormal, a downgrade control command is issued and the downgrade trigger time, trigger stage, control command issuance record and filling result deviation are recorded. When the inspection and judgment results are pending, a conservative control command is issued and the control parameter set correction record and filling result deviation are recorded.

9. A control system for a machine learning-based digital filling control method according to any one of claims 1-8, comprising the following modules: The data acquisition and preprocessing module is used to acquire multi-source data, perform preprocessing and batch association, and generate fused input data; The VMamba prediction module is used to complete orientation scan modeling and high- and low-frequency feature fusion by fusing input data, and output predicted distribution parameters and visual risk indicators. The sequential testing module is used to perform sliding window log-likelihood accumulation and observation dimension switching on the predicted distribution parameters and visual risk indicators, and output the test judgment results. The strategy compensation module is used to generate a filling control strategy based on the inspection and judgment results, calculate the compensation control quantity, and output a set of control parameters. The constraint execution module is used to perform process verification on the set of control parameters and issue control commands to achieve real-time correction. The traceability and degradation module is used to record full-process traceability data and trigger security degradation and archiving.

Images