Intelligent control method for modified starch technological process based on multi-modal data fusion

Abstract

The invention relates to the technical field of modified starch production, in particular to an intelligent control method for a modified starch technological process based on multi-modal data fusion. Comprising the following steps: acquiring multi-modal data; performing multi-mode real-time sensing data fusion processing; carrying out dynamic modeling and prediction on the technological process: carrying out training modeling on the fused multi-modal real-time sensing data and a reaction endpoint viscosity target index by adopting an improved multi-modal time sequence Transform model, and outputting prediction results of the feeding rate and the reaction time; and executing a closed-loop control instruction. Multi-modal real-time sensing data such as the pH value, the conductivity, the infrared spectrum and the real-time temperature are synchronously collected, fusion processing is carried out in combination with three-criterion abnormal value detection, standardized conversion and a feature splicing strategy, multi-dimensional information in the reaction process can be integrated, interference caused by one-sidedness of single-modal data is reduced, and the detection accuracy is improved. The integrity and the accuracy of depicting the modification process of the modified starch are improved.

Description

Technical Field

[0001] This invention relates to the field of modified starch production technology, and more specifically, to an intelligent control method for modified starch production processes based on multimodal data fusion. Background Technology

[0002] Modified starch, a key raw material in industrial production, directly impacts its downstream applications due to its quality stability. In continuous modified starch production lines, the reaction process involves the dynamic interaction of multiple parameters, including pH, temperature, and dosage, with inherent lags in material mixing and chemical reactions. Traditional production control often relies on offline detection or manual adjustment of single parameters (such as temperature or pH), making it difficult to capture subtle fluctuations caused by the coupling of multiple variables in real time. This can easily lead to batch-to-batch inconsistencies in key indicators such as viscosity and reactivity due to sudden temperature changes or dosage deviations, hindering production efficiency and product uniformity. Therefore, achieving precise and dynamic control of the modified starch modification process has become a core requirement for improving product quality stability.

[0003] In existing modified starch production processes, control technologies suffer from the following significant limitations: First, data acquisition is often limited to single-modality or a few static parameters, failing to effectively integrate multi-dimensional information such as infrared spectroscopy and conductivity. This results in a one-sided characterization of the reaction process and difficulty in identifying potential quality risks early on. Second, control strategies are mostly open-loop presets or simple feedback adjustments, lacking modeling of reaction kinetics. They cannot predict the optimal values ​​of key process parameters such as feeding rate and reaction time based on real-time process data, and it is even more difficult to form a closed-loop optimization mechanism of "sensing-prediction-correction." This leads to delayed responses to sudden fluctuations and exacerbates batch-to-batch quality dispersion. Therefore, we propose an intelligent control method for modified starch production processes based on multi-modal data fusion. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent control method for modified starch processing based on multimodal data fusion, so as to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides an intelligent control method for modified starch processing based on multimodal data fusion, comprising the following steps: S100, Multimodal Data Acquisition: Through online sensing devices, multimodal real-time sensing data including but not limited to pH value, conductivity, infrared spectrum data and real-time temperature are synchronously acquired at a preset frequency. At the same time, the target index of viscosity at the reaction endpoint is also connected to ensure the real-time performance of data acquisition and the integrity of multimodal coverage. S200, multimodal real-time sensing data fusion processing: through 3 The system employs outlier detection, standardization transformation, and feature stitching strategies to fuse multimodal real-time sensing data collected by the S100, ensuring the spatiotemporal consistency and feature validity of the multimodal real-time sensing data. S300, Dynamic Modeling and Prediction of Process: An improved multimodal time-series Transformer model is adopted to train and model the fused multimodal real-time sensing data and the viscosity target index at the reaction endpoint, and output the prediction results of feeding rate and reaction time to achieve accurate fitting of the dynamic characteristics of the modified starch modification process. S400 Closed-loop control command execution: Based on the feeding rate and reaction time output by the improved multimodal temporal Transformer model trained and modeled in S300, the feeding rate is adjusted by controlling the frequency of the feeding pump, and the reaction time is adjusted by controlling the running time of the reactor stirring motor, outputting automatic adjustment commands; and a closed-loop control is formed based on real-time feedback multimodal real-time sensing data to improve the stability of modified starch quality.

[0006] As a further improvement to this technical solution, the deployment and data acquisition of the online sensing device in S100 includes the following steps: S110.1, Sensor Equipment Deployment and Positioning: Based on the material flow field distribution and reaction characteristics in the modified starch reaction vessel, pH and conductivity sensors are installed in the turbulent zone in the lower part of the vessel. The infrared spectrometer probe is aligned with the uniformly mixed area of ​​the material in the vessel through the optical window. Temperature sensors are arranged in layers along the axial direction of the vessel. Viscosity detection equipment is integrated into the discharge pipeline at the reaction endpoint. S110.2 Pre-acquisition calibration and self-test: Before data acquisition, perform calibration operations according to the sensor type (e.g., pH sensor is calibrated with standard buffer solution, conductivity meter is calibrated with standard conductivity solution, infrared spectrometer is calibrated with characteristic peak standard material, temperature sensor is calibrated with standard temperature point, and viscosity detection equipment is calibrated with standard viscosity medium), and at the same time complete the self-test of the device communication link, power supply stability and measurement range; S110.3 Multimodal real-time sensing data synchronous acquisition: Based on the real-time clock of the reactor control system, each sensor is triggered to synchronously acquire data at a preset frequency. During the acquisition process, a timestamp (accurate to the millisecond level) is added to each set of data. The acquisition interval of parameters with fast dynamic response (such as temperature) is shorter than that of parameters with slow dynamic response (such as infrared spectral data). S110.4 Real-time verification of acquired data: Real-time validity verification of acquired multimodal real-time sensing data, including whether the data is within the normal measurement range of the sensor, whether the data changes in adjacent acquisition cycles conform to the law of reaction kinetics, marking data with abnormal verification and initiating a supplementary acquisition mechanism (supplemented by interpolation based on the trend of similar historical data).

[0007] As a further improvement to this technical solution, the frequency preset and access target indicators in S100 include the following steps: S120.1 Multimodal data acquisition frequency classification preset: Based on the kinetic characteristics of the modified starch reaction (such as the correlation between reaction rate and temperature and pH value), combined with the dynamic response differences of each parameter (such as the sensitivity of temperature and pH value changes being higher than that of infrared spectral data), the multimodal real-time sensing data is divided into different acquisition frequency levels. The preset frequency of reaction-sensitive parameters is higher than that of non-sensitive parameters. The initial frequency range is determined by statistical analysis of parameter change rates in historical reaction cycles. S120.2 Frequency Dynamic Adaptation Mechanism: During the acquisition process, the change amplitude of multimodal real-time sensing data (such as pH value fluctuation and temperature change rate in adjacent cycles) is monitored in real time. When the change amplitude of a certain parameter exceeds the normal range of the corresponding reaction stage, the acquisition frequency of the parameter is automatically increased (not exceeding the upper limit of the sensor hardware response). After the change amplitude falls back to the normal range, the initial frequency is restored, and the trigger threshold is adjusted based on the historical data distribution of similar reactions. S120.3, Reaction endpoint viscosity target index access: Collect reaction endpoint viscosity data at a preset cycle through a dedicated detection interface (the cycle matches the reaction process, and the collection interval is shortened in the later stage of the reaction). Record the precise timestamp when accessing to ensure that it corresponds to the time dimension of the multimodal real-time sensing data. S120.4, Time-series integration of target indicators and multimodal real-time sensing data: The target viscosity indicator at the reaction endpoint is aligned with the multimodal real-time sensing data collected at the same time through timestamps to form a correlated sample of "multimodal real-time sensing data segment - corresponding viscosity value". During the correlation process, multimodal real-time sensing data that is judged to be abnormal by real-time verification (such as data that exceeds the normal measurement range of the sensor or whose change trend does not conform to the reaction law) is removed. The integration result is used for the sample construction of subsequent data fusion.

[0008] As a further improvement to this technical solution, in S200, 3 The outlier detection criteria involve fusing multimodal real-time sensing data, including the following steps: S210.1, Algorithm Formula Definition: Single-modal real-time sensing data sequence acquired by S100 ,in For the first The raw data at each collection point Given the length of the data sequence; and calculate the mean within the sliding window. and standard deviation The formula is: ; ; in, The sliding window length (representing the number of adjacent data points involved in the calculation). For indexing the data within the window, Index the current data point; when At that time, the judgment This is an outlier; S210.2 Algorithm Parameter Adaptation: The window sliding step size is consistent with the acquisition cycle of the corresponding modal data to ensure a balance between real-time performance and data continuity; the same modal data can be dynamically adjusted at different stages of the reaction (such as the initial, middle, and final stages of the reaction). The value is adjusted based on the data fluctuation characteristics of that period; S210.3, Outlier Handling Procedure: For single-value anomalies (only) abnormal, and (Normal), use linear interpolation replacement: ;in, For corrected data; for multiple consecutive outliers (more than 3 acquisition cycles for this mode), trigger a status re-inspection of the corresponding sensor (including connection stability and probe cleanliness checks). During the re-inspection, use backup sensor channel data and mark the data source attributes.

[0009] As a further improvement to this technical solution, the standardization conversion in S200 includes the following steps: S220.1, Algorithm Formula Definition: For single-modal real-time sensing data after outlier processing... Converted using Z-score standardization The formula is: ; in, This is a reference average (derived from statistical data of the same modality from historical normal production batches). The reference standard deviation is derived from statistical analysis of the same modal data of historical normal production batches. S220.2, Reference Parameter Determination: and The statistical sample consists of historical qualified product data of the same modality, covering different raw material batches and process parameter combinations; the data for each reaction stage are calculated separately based on reaction kinetic characteristics. and This ensures that the data characteristics are standardized and adapted to different reaction stages. S220.3, High-dimensional data adaptation: For high-dimensional infrared spectral data, each wavelength point Implement standardization separately: ; in, , The first Reference mean and standard deviation for each wavelength point; To standardize absorbance; The relative intensity relationship of the spectral characteristic peaks retained after conversion.

[0010] As a further improvement to this technical solution, S200 performs fusion processing on multimodal real-time sensing data through a feature stitching strategy, including the following steps: S230.1 Time Segment Division: The multimodal real-time sensing data after outlier processing and standardization is divided into continuous time segments according to reaction kinetic characteristics (segments include continuously acquired data of temperature, pH value, conductivity and corresponding infrared spectral data, etc.). Each segment covers the same time span (determined based on the data acquisition frequency of each stage to ensure that the segment contains complete reaction characteristic changes). S230.2 Modal Feature Extraction: For numerical data within each time segment, extract the extreme values, mean values, and changes in adjacent data points within the segment to form a numerical feature vector; for high-dimensional infrared spectral data, extract the wavenumber positions of characteristic peaks and their corresponding absorbance values ​​(based on the standardized wavelength points in S220.3) to form a spectral feature vector. S230.3 Cross-modal feature concatenation: The numerical feature vector and the spectral feature vector are concatenated horizontally in the order of "numerical features first, spectral features second". At the same time, each feature is labeled with its corresponding modal source and the time segment identifier, forming a segment-level fusion feature matrix. The matrix dimension is determined by the sum of the number of numerical features and the number of spectral features. S230.4 Feature Validity Screening: By calculating the correlation between each feature in the fusion feature matrix and the viscosity at the reaction endpoint, features with a correlation higher than the baseline value obtained from the statistics of historical normal reaction data are retained, and redundant features with too low correlation are eliminated to form the final fusion feature set used for S300 modeling.

[0011] As a further improvement to this technical solution, in step S300, the improved multimodal temporal Transformer model structure includes the following steps: S310.1, Multimodal Input Encoding: The fused feature set output from S200 (including numerical features and infrared spectral features) and the reaction endpoint viscosity target index from S120.4 are used as model inputs. The numerical features are converted to a dimension of [missing information - likely a specific value]. The embedded vectors and infrared spectral features are extracted using a 1D convolutional layer (with kernel size matching the normalized wavelength point spacing in S220.3) to extract local features and then converted into dimensions. The embedding vector, the viscosity target index is encoded as a dimension The supervision vector; all vectors are appended with a location code based on the S110.3 timestamp, as follows: ; ; in, This is the location-coded value. The timestamp in S110.3 For the dimension index of the embedded vector, and , The total dimension of the embedded vector; S310.2 Cross-modal attention calculation: Introducing modal weight moments in the Transformer encoder ( (Distinguishing between numerical, spectral, and viscous targets), the cross-modal attention calculation formula is as follows: ; in, For modality The query vector, For modality key vector ( ), For modality The value vector, For modality The weight matrix (initialized based on the correlation coefficients between each mode and viscosity in S230.4). This is the normalization function; This represents the matrix transpose operation; Function encapsulation representing cross-attention operations; S310.3, Dual Output Head Mapping: The decoder has two output heads, and the feed rate prediction formula is as follows: The formula for predicting reaction time is: ;in, To predict the feeding rate, To predict reaction time, For decoder output features, , These are the weight matrix and bias term of the feed rate head, respectively. , These are the weight matrix and bias term of the reaction time head, respectively; Within the frequency adjustment range of the feed pump in the S400, Based on the truncation of the reaction kinetics period range in S120.1.

[0012] As a further improvement to this technical solution, the model training and modeling in S300 includes the following steps: S320.1 Training Sample Construction: The sample data is associated with "multimodal real-time sensing data fragments - corresponding viscosity values" integrated by S120.4. The input is the fused feature fragment processed by S200 (same as the filtering result of S230.4), and the label is the actual feeding rate of the corresponding batch. With reaction time ; Based on the reaction stages divided by reaction kinetic characteristics (same as the reaction stage division logic in S220.2), stratified sampling is used, and the sample proportion of each stage is consistent with the distribution of historical production batches. The samples are divided into training set and validation set, where the sample size of the training set is no less than 70% of the total sample size to ensure that the model learns the correlation between multimodal features and process parameters. The validation set is used to independently evaluate the model's generalization ability. S320.2, Loss Function Definition: A weighted joint loss function is used, and the formula is as follows: ; in, MSE loss due to feed rate, ; MAE loss due to reaction time, ; For the sample size, , The first The actual feeding rate and reaction time of each sample , These are the corresponding predicted values; , These are the weighting coefficients, and ; ( , Based on the correlation between the two parameters and viscosity in S120.4, the parameter with a higher correlation has a greater weight. S320.3 Training Parameter Update: The Adam optimizer is used to update the model parameters. The parameter update formula is as follows: ; in, For the first Wheel model parameters, For the updated parameters, The learning rate (dynamically adjusted based on the validation set loss) is used. For first-order momentum estimation, For second-order momentum estimation, To prevent tiny constants with a denominator of zero; After each training round, the validation set loss is calculated. If the loss does not decrease for several consecutive rounds, the learning rate is adjusted until training terminates, and the model parameters with the minimum validation set loss are saved.

[0013] As a further improvement to this technical solution, the dynamic characteristic fitting and prediction output in S300 includes the following steps: S330.1, Real-time Input Adaptation: The fusion feature set processed by S200 within the current reaction cycle (same as the screening result of S230.4) is converted into a model input vector according to the encoding method of S310.1, and the viscosity target index at the reaction endpoint of this cycle is simultaneously connected (same as the connection logic of S120.3); the time dimension of the input vector corresponds one-to-one with the timestamp in S110.3 to ensure temporal consistency; S330.2, Rolling Prediction Execution: A sliding window mechanism is used to generate the prediction sequence, with a window length of... The formula for determining it is: ; in, The minimum lag time for the reaction (determined based on the time-series correlation analysis of multimodal data and viscosity in S120.4). This is the highest sampling frequency. It is a rounding function; The model is based on the most recent The fusion features of each acquisition cycle serve as the input context, and the future features are output every one acquisition cycle. Feed rate prediction sequence for each cycle and reaction time prediction sequences The predicted value ranges are respectively matched with the feed pump adjustment range of S400 and the reaction cycle range of S120.1; S330.3, Fitting Accuracy Verification: The similarity between the predicted sequence and the actual process parameter sequence is calculated using the dynamic time warping algorithm. The formula is as follows: ; in, For the predicted sequence, This is the actual parameter sequence (corresponding to the actual feeding rate or reaction time). For time-ordered paths, The length of the normalized sequence; It is a dynamic time warping algorithm; When the DTW value exceeds the baseline threshold based on historical qualified batch statistics, incremental model training is triggered (the output header parameters in S310.3 are fine-tuned using only the current period data) to ensure the accuracy of dynamic characteristic fitting.

[0014] As a further improvement to this technical solution, the execution of the closed-loop control command in S400 includes the following steps: S410.1 Control Parameter Conversion: Convert the feeding rate and reaction time prediction results output by S300 into control parameters that can be recognized by the corresponding equipment. The feeding rate corresponds to the feeding pump frequency adjustment value (determined based on the rated flow and rate adjustment range of the feeding pump, and the conversion relationship is set through the equipment factory calibration data), and the reaction time corresponds to the running time setting value of the reactor stirring motor (the range matches the reaction kinetic cycle range in S120.1). S410.2 Command Sending and Execution Confirmation: The converted control parameters are sent to the feed pump controller and the stirring motor driver via the industrial control bus. During the sending process, the communication link self-checking mechanism in S110.2 is enabled to ensure the integrity of the command transmission. After receiving the command, the feed pump and the reactor stirring motor return a status response. If a valid response is not received within the set time limit, the backup control channel is activated to resend the command until it is confirmed that the command has been correctly received. S410.3 Real-time feedback data acquisition: After the command is executed, according to the synchronous acquisition rules of S110.3, the actual operating frequency of the feed pump, the working status of the stirring motor, and the real-time sensing data of multi-modal data in the reactor (temperature, pH value, conductivity, infrared spectrum, etc.) are acquired in real time. All feedback data are appended with the timestamp corresponding to the control command to ensure timing consistency. S410.4 Operational Deviation Analysis: After the feedback data is processed and standardized by S200, the deviations between the actual feeding rate and reaction time and the predicted values ​​are compared. At the same time, the characteristic correlation analysis method of S230.4 is used to determine whether the deviation affects the viscosity target index at the reaction endpoint (the deviation threshold is determined based on the fluctuation range of process parameters of historical qualified batches). S410.5 Closed-loop parameter correction: If the deviation exceeds the threshold, based on the current multimodal feedback data and the real-time prediction results of the S300 model, the corrected control parameters are generated. In the next acquisition cycle, the instructions of the feed pump and the reactor stirring motor are updated. By repeatedly adjusting until the deviation falls back to within the threshold, a continuous closed-loop control is formed to ensure the stability of the modified quality.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention simultaneously acquires multimodal real-time sensing data such as pH value, conductivity, infrared spectrum, and real-time temperature, and combines it with 3 By integrating outlier detection, standardization transformation, and feature splicing strategies, multi-dimensional information from the reaction process can be integrated, reducing interference from the one-sidedness of single-modality data and improving the completeness and accuracy of characterizing the modified starch modification process. 2. This invention employs an improved multimodal time-series Transformer model to train and model the fused multimodal data and the target viscosity index at the reaction endpoint. This model can better capture the dynamic coupling characteristics and time-series correlations of multiple variables during the reaction process, improve the reliability of the prediction results for feeding rate and reaction time, and provide an effective basis for adjusting process parameters. 3. Based on the prediction results output by the model, this invention adjusts the process parameters by controlling the frequency of the feeding pump and the running time of the stirring motor in the reactor, and forms a closed-loop control by combining real-time feedback multimodal data. This can respond to parameter fluctuations in the reaction process in a timely manner, reduce batch quality differences caused by lag adjustment, and improve the stability of modified starch quality. 4. This invention uses a dynamic adaptation mechanism for multimodal data acquisition frequency to automatically increase the acquisition frequency when the parameter change exceeds the normal range. This enables the acquisition of more intensive process information during critical reaction stages, enhances the early identification capability of reaction anomalies, and creates conditions for timely adjustment of process parameters. 5. In the data fusion process, this invention filters features based on their effectiveness, retaining features that are highly correlated with the viscosity of the reaction endpoint and eliminating redundant features. This reduces the interference of invalid information on model modeling and improves model training efficiency and prediction accuracy. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the steps of the intelligent control method of the present invention; Figure 2 This is a schematic diagram illustrating the steps of data processing using a feature concatenation strategy in this invention; Figure 3 This is a schematic diagram illustrating the steps of executing closed-loop control commands according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figures 1-3 As shown, this embodiment provides an intelligent control method for modified starch processing based on multimodal data fusion, including: S100, Multimodal Data Acquisition: Through online sensing devices, multimodal real-time sensing data including but not limited to pH value, conductivity, infrared spectrum data and real-time temperature are synchronously acquired at a preset frequency. At the same time, the target index of viscosity at the reaction endpoint is also connected to ensure the real-time performance of data acquisition and the integrity of multimodal coverage. In this step, the deployment and data acquisition of the online sensing devices in S100 includes the following steps: S110.1, Sensor Equipment Deployment and Positioning: Based on the material flow field distribution and reaction characteristics in the modified starch reaction vessel, pH and conductivity sensors are installed in the turbulent zone in the lower part of the vessel. The infrared spectrometer probe is aligned with the uniformly mixed area of ​​the material in the vessel through the optical window. Temperature sensors are arranged in layers along the axial direction of the vessel. Viscosity detection equipment is integrated into the discharge pipeline at the reaction endpoint. As a further explanation of this step, the deployment details of some online sensing devices in this embodiment are as follows: pH and conductivity sensors: Immersion probes are used and installed in the lower part of the turbulent zone of the reactor (this area is determined by material flow simulation during reactor design, ensuring thorough mixing of materials and no obvious dead zones). The distance between the two sensors is ≥10cm to avoid mutual interference. The probe is connected to the reactor body through a sealed interface, and the probe tip is flush with the inner wall of the reactor to reduce material adhesion. Infrared spectrometer detection probe: The optical window is made of quartz material that is resistant to corrosion by the reaction medium and is embedded in the uniformly mixed area of ​​the material on the side wall of the vessel (to ensure that the window is completely covered by the material during detection); a purging device (introducing dry inert gas) is installed on the outside of the window to prevent material from splashing and affecting light transmission; the probe is connected to the spectrometer host via optical fiber, and the length of the optical fiber is determined according to the equipment layout (based on the principle of no significant signal attenuation).

[0019] Temperature sensors: Three groups (upper, middle, and lower layers) are arranged along the axial direction of the vessel. The upper layer is located 5-10 cm below the liquid surface, the middle layer is located at 1 / 2 of the vessel height, and the lower layer is close to the bottom of the vessel (10-15 cm from the bottom). Each group contains two symmetrically distributed sensors, and the average value is taken as the temperature data of that layer to reduce the impact of local temperature fluctuations. Viscosity testing equipment: integrated into the bypass branch of the discharge pipeline at the reaction endpoint (switched via valves, without affecting the material flow in the main pipeline). During testing, the material flows through the viscometer measuring chamber (chemically modified starch; viscosity testing requires heating and gelatinizing a quantitative amount of dry starch before testing, and the sampled material needs to be adjusted for pH and conductivity before testing; simply letting the material flow through the bypass once does not represent the product viscosity). The measuring chamber volume is adapted to the pipeline flow rate (to avoid material stagnation), and the inner wall is treated with anti-sticking material.

[0020] S110.2 Pre-acquisition calibration and self-test: Before data acquisition, perform calibration operations according to the sensor type (e.g., pH sensor is calibrated with standard buffer solution, conductivity meter is calibrated with standard conductivity solution, infrared spectrometer is calibrated with characteristic peak standard material, temperature sensor is calibrated with standard temperature point, and viscosity detection equipment is calibrated with standard viscosity medium), and at the same time complete the self-test of the device communication link, power supply stability and measurement range; As a further explanation of this step, this embodiment details the pre-acquisition calibration operation, specifically including: pH sensor: Use industry-standard buffer solutions (such as pH=4.00, 7.00, 9.18) to calibrate sequentially in the order of "low concentration → high concentration". Record the calibration parameters after the readings stabilize. Conductivity meter: Use a standard conductivity solution with the corresponding measurement range. During calibration, ensure that the solution temperature is consistent with the temperature of the reactants (or correct it using the device's built-in temperature compensation function). Infrared spectrometer: Using industry-recognized standard materials (such as polystyrene film), scan a typical wavenumber range to confirm the position of characteristic peaks and their correlation with the standard spectrum. Figure 1 To; Temperature sensor: Placed in a constant temperature medium at a known temperature (such as boiling water or ice water mixture), the deviation between the calibration reading and the actual temperature is within the allowable range of the equipment; Viscosity testing equipment: Use a standard viscosity medium (such as silicone oil with known viscosity; in the modified starch industry, standard starch is generally used to calibrate viscosity equipment), repeat the measurement multiple times, and the deviation of the result from the standard value should be within the allowable range of the equipment manual.

[0021] Furthermore, this embodiment further refines the self-inspection items before data collection, specifically including: Communication link: The control system sends test signals to each sensor, and the ability to receive feedback normally is considered a normal link; Power supply stability: The power supply voltage of the monitoring sensors is within the nominal operating voltage range of the equipment; Measurement range: Input an analog signal within the range of the device to confirm that it can output the measurement value normally; input an over-range signal to confirm that it can output an "over-range" prompt.

[0022] S110.3 Multimodal real-time sensing data synchronous acquisition: Based on the real-time clock of the reactor control system, each sensor is triggered to synchronously acquire data at a preset frequency. During the acquisition process, a timestamp (accurate to the millisecond level) is added to each set of data. The acquisition interval of parameters with fast dynamic response (such as temperature) is shorter than that of parameters with slow dynamic response (such as infrared spectral data). As a further explanation of this step, the synchronous acquisition mechanism in this embodiment is as follows: Synchronization reference: The real-time clock of the reactor control system is used as the reference (this clock is synchronized with the unified clock of the production line). A synchronization signal is generated every cycle to trigger each sensor to collect data simultaneously. Timestamp generation: The control system uniformly adds a timestamp (including year, month, day, hour, minute, second, and millisecond) to each group of data to ensure that the timestamps of multimodal data are consistent; Acquisition interval setting: Based on the dynamic response characteristics of the parameters, parameters with fast dynamic response (such as temperature, pH value, and conductivity) have short acquisition intervals, while parameters with slow dynamic response (such as infrared spectral data) have long acquisition intervals (the specific interval is determined according to the rate of change of the parameter in the reaction to ensure that key changes can be captured).

[0023] S110.4 Real-time verification of acquired data: Real-time validity verification of acquired multimodal real-time sensing data, including whether the data is within the normal measurement range of the sensor, whether the data changes in adjacent acquisition cycles conform to the law of reaction kinetics, marking data with abnormal verification and initiating a supplementary acquisition mechanism (supplemented by interpolation based on the trend of similar historical data).

[0024] As a further explanation of this step, this embodiment performs real-time validity verification on the collected multimodal real-time sensing data, specifically including: Range verification: The data value is within the sensor's nominal measurement range (e.g., pH value is within the sensor's specified measurable range, conductivity is within the device's measurement range, etc.). Trend verification: The data changes in adjacent collection periods conform to the basic laws of the reaction (e.g., the temperature does not change drastically in a short period of time, and the pH value changes are consistent with the trend of the type and dosage of added reagents).

[0025] As a further explanation of this step, this embodiment specifically includes marking data with verification anomalies and initiating a supplementary data collection mechanism, which includes: For a single abnormal data point: use adjacent valid data points to perform linear interpolation to supplement it; Continuous small amounts of abnormal data: Data from backup sensors with this parameter enabled (if there are redundant sensors). Multiple consecutive abnormal data points: Mark as abnormal data, trigger sensor status check, and make reasonable estimation based on recent data change trends (the estimation result is marked "trend estimation").

[0026] In this step, the frequency preset and access target indicators in S100 include the following steps: S120.1 Multimodal data acquisition frequency classification preset: Based on the kinetic characteristics of the modified starch reaction (such as the correlation between reaction rate and temperature and pH value), combined with the dynamic response differences of each parameter (such as the sensitivity of temperature and pH value changes being higher than that of infrared spectral data), the multimodal real-time sensing data is divided into different acquisition frequency levels. The preset frequency of reaction-sensitive parameters is higher than that of non-sensitive parameters. The initial frequency range is determined by statistical analysis of parameter change rates in historical reaction cycles. As a further explanation of this step, this embodiment divides the multimodal real-time sensing data into different acquisition frequency levels, and the specific criteria for this classification include: Based on historical records of parameter changes in the modified starch reaction during production, the average rate of change for each parameter was statistically analyzed: Reaction-sensitive parameters (temperature, pH, conductivity): These change frequently during critical stages of the reaction, so the initial sampling frequency is set to a high level. Non-sensitive parameters (infrared spectral data): changes are relatively gradual, and the initial acquisition frequency is set to a low level; Furthermore, the sampling frequency of all parameters shall not exceed the maximum sampling capacity of the sensor itself (subject to the device manual).

[0027] S120.2 Frequency Dynamic Adaptation Mechanism: During the acquisition process, the change amplitude of multimodal real-time sensing data (such as pH value fluctuation and temperature change rate in adjacent cycles) is monitored in real time. When the change amplitude of a certain parameter exceeds the normal range of the corresponding reaction stage, the acquisition frequency of the parameter is automatically increased (not exceeding the upper limit of the sensor hardware response). After the change amplitude falls back to the normal range, the initial frequency is restored, and the trigger threshold is adjusted based on the historical data distribution of similar reactions. As a further explanation of this step, the conventional range of the corresponding reaction stage in this embodiment specifically includes: statistically analyzing the historical variation range of each parameter according to the reaction stage (such as the raw material mixing stage, the reaction stage, and the maturation stage), and taking the variation range of the majority of normal batches as the conventional range of that stage.

[0028] Furthermore, the adjustment rule for the sampling frequency in this embodiment is as follows: When the change of a certain parameter exceeds the normal range of the corresponding stage, the sampling frequency of that parameter is increased (the increased frequency shall not exceed the maximum sampling capability of the sensor). Once the parameter variation returns to the normal range and stabilizes for a period of time, the initial frequency is restored.

[0029] S120.3, Reaction endpoint viscosity target index access: Collect reaction endpoint viscosity data at a preset cycle through a dedicated detection interface (the cycle matches the reaction process, and the collection interval is shortened in the later stage of the reaction). Record the precise timestamp when accessing to ensure that it corresponds to the time dimension of the multimodal real-time sensing data. As a further explanation of this step, the dedicated detection interface in this embodiment is specifically: connected to the viscosity detection device through an industrial general communication interface (such as RS485), and automatically uploads data (including viscosity value, measurement time, and device status) after the device completes the measurement.

[0030] Meanwhile, the adjustment of the acquisition period in this embodiment specifically includes: Early stage of reaction: The collection period is relatively long (based on the reaction progress, to ensure that the overall trend of viscosity change can be reflected); Later stage of the reaction: shorten the sampling cycle (because the viscosity change accelerates near the endpoint / the acceptable viscosity change range becomes smaller near the endpoint, requiring more frequent monitoring).

[0031] In addition, after each data acquisition, check whether the data format is correct. If the format is incorrect, request the device to resend. If multiple transmissions fail, switch to the backup viscosity detection channel (if available).

[0032] S120.4, Time-series integration of target indicators and multimodal real-time sensing data: The target viscosity indicator at the reaction endpoint is aligned with the multimodal real-time sensing data collected at the same time through timestamps to form a correlated sample of "multimodal real-time sensing data segment - corresponding viscosity value". During the correlation process, multimodal real-time sensing data that is judged to be abnormal by real-time verification (such as data that exceeds the normal measurement range of the sensor or whose change trend does not conform to the reaction law) is removed. The integration result is used for the sample construction of subsequent data fusion.

[0033] As a further explanation of this step, during the time-series integration process, the timestamp of the viscosity data is used as the benchmark to match the real-time sensing data of multimodal sensors collected within the same time period (time deviation is controlled within a reasonable range to ensure that the data corresponds to the same reaction state); the associated samples are stored in the form of tables or databases, including "sample identifier, viscosity value, corresponding time period, data sequence of each modality (including anomaly markers), data source (main sensor / backup sensor)"; multimodal data that is found to be abnormal after verification by S110.4 is marked with its abnormality type (such as "overrange" or "trend abnormality") during integration, and is not included in the sample set for subsequent data fusion, but is only used to analyze the sensor's working status.

[0034] S200, multimodal real-time sensing data fusion processing: through 3 The system employs outlier detection, standardization transformation, and feature stitching strategies to fuse multimodal real-time sensing data collected by the S100, ensuring the spatiotemporal consistency and feature validity of the multimodal real-time sensing data. In this step, S200 passes through 3 The outlier detection criteria involve fusing multimodal real-time sensing data, including the following steps: S210.1, Algorithm Formula Definition: Single-modal real-time sensing data sequence acquired by S100 ,in For the first The raw data at each collection point Given the length of the data sequence; and calculate the mean within the sliding window. and standard deviation The formula is: ; ; in, The sliding window length (representing the number of adjacent data points involved in the calculation). For indexing the data within the window, Index the current data point; when At that time, the judgment This is an outlier; As a further explanation of this step, the single-modal data sequence in this embodiment... Specific types include: temperature sequences, pH sequences, conductivity sequences, and infrared absorbance sequences. In practical applications, the rationality of outlier determination needs to be confirmed based on the physical meaning of the data type (e.g., temperature data 3). The threshold value needs to be considered in conjunction with the temperature control capability of the reactor to avoid misjudgment due to normal process fluctuations.

[0035] S210.2 Algorithm Parameter Adaptation: The window sliding step size is consistent with the acquisition cycle of the corresponding modal data to ensure a balance between real-time performance and data continuity; the same modal data can be dynamically adjusted at different stages of the reaction (such as the initial, middle, and final stages of the reaction). The value is adjusted based on the data fluctuation characteristics of that period; As a further explanation of this step, the sliding window length in this embodiment... Based on historical response data, the fluctuation characteristics of each stage are statistically analyzed: In the initial stage of the reaction (raw material mixing stage): data fluctuates significantly. Use smaller values ​​(e.g., 5-10 data points) to improve the sensitivity of outlier identification; Mid-reaction phase (main reaction stage): Data is relatively stable. Use larger values ​​(such as 15-20 data points) to reduce misjudgments caused by random fluctuations; Final stage of reaction (maturation stage): Data approaching stability. The value recovers to a moderate level (e.g., 10-15 data points); Meanwhile, adjustments are made through the parameter configuration module of the control system, allowing operators to fine-tune the operation based on actual response stability. Value range (the adjustment range shall not exceed ±5 data points of the initial setting).

[0036] S210.3, Outlier Handling Procedure: For single-value anomalies (only) abnormal, and (Normal), use linear interpolation replacement: ;in, For corrected data; for multiple consecutive outliers (more than 3 acquisition cycles for this mode), trigger a status re-inspection of the corresponding sensor (including connection stability and probe cleanliness checks). During the re-inspection, use backup sensor channel data and mark the data source attributes.

[0037] As a further explanation of this step, the state re-check step in this embodiment is as follows: After the retest is triggered, the control system automatically pauses the data acquisition of the sensor and issues an audible and visual alarm. Operators should check whether the sensor wiring is loose, and whether the probe is scaled or damaged (e.g., whether there are crystals on the pH sensor probe, and whether there are stains on the infrared spectral window). After cleaning or repair, verify the sensor performance through the calibration procedure (same as S110.2). If it passes the test, resume main channel acquisition.

[0038] As a further explanation of this step, this embodiment also sets up a backup sensing channel, specifically including: sensors for key parameters (such as temperature, pH value, conductivity) are all equipped with backup probes, installed adjacent to the main probe (distance ≤ 5cm), and the acquisition frequency is consistent with the main channel; the backup data does not participate in fusion by default, and only automatically switches when the main channel is continuously abnormal, and data continuity is ensured by aligning timestamps during the switching process.

[0039] In this step, the standardization conversion in S200 includes the following steps: S220.1, Algorithm Formula Definition: For single-modal real-time sensing data after outlier processing... Converted using Z-score standardization The formula is: ; in, This is a reference average (derived from statistical data of the same modality from historical normal production batches). The reference standard deviation is derived from statistical analysis of the same modal data of historical normal production batches. S220.2, Reference Parameter Determination: and The statistical sample consists of historical qualified product data of the same modality, covering different raw material batches and process parameter combinations; the data for each reaction stage are calculated separately based on reaction kinetic characteristics. and This ensures that the data characteristics are standardized and adapted to different reaction stages. As a further explanation of this step, the statistical sample in this embodiment is the same modal data of more than 30 batches of qualified products in the past, covering different raw material batches (such as corn starch, potato starch) and the normal fluctuation range of conventional process parameters (such as reaction temperature, reagent addition amount).

[0040] As a further explanation of this step, the reaction stages in this embodiment are divided based on reaction kinetic characteristics, specifically including: Initial stage (0-1 / 3 of total reaction time): raw material mixing and preheating, parameter change rate is rapid; Mid-term (1 / 3-2 / 3 of total reaction time): The main reaction proceeds, and the parameters are relatively stable; Final stage (2 / 3 to end of total reaction time): The reaction is nearing completion, and the parameters gradually stabilize; Moreover, each stage and Calculate using statistical software (such as Excel, Python Pandas library): This represents the arithmetic mean of the data for this period. This is the standard deviation of the data for this stage, calculated by excluding data from known abnormal batches (such as batches that are non-conforming due to equipment failure).

[0041] S220.3, High-dimensional data adaptation: For high-dimensional infrared spectral data, each wavelength point Implement standardization separately: ; in, , The first Reference mean and standard deviation for each wavelength point; To standardize absorbance; The relative intensity relationship of the spectral characteristic peaks retained after conversion.

[0042] As a further explanation of this step, the infrared spectral data contains thousands of wavelengths, but in actual processing, only the characteristic wavenumbers related to the modified starch reaction are retained to reduce redundant calculations.

[0043] like Figure 2As shown, in this step, S200 performs fusion processing on multimodal real-time sensing data through a feature stitching strategy, including the following steps: S230.1, Time-series segment division: The multimodal real-time sensing data after outlier processing and standardization is divided into continuous time-series segments according to reaction kinetic characteristics (segments include continuous acquisition data of temperature, pH value, conductivity and corresponding infrared spectral data, etc.), with each segment covering the same time span (determined based on the data acquisition frequency of each stage to ensure that the segment contains complete reaction characteristic changes). As a further explanation of this step, the time span of the segments in this embodiment is determined based on the acquisition frequency of each modality data to ensure that each segment contains sufficient feature change information: For parameters with fast dynamic response (such as temperature and pH value, with a sampling interval of 1 second): the segment span is set to 10 seconds, containing 10 consecutive data points; For parameters with slow dynamic response (such as infrared spectra, with an acquisition interval of 5 seconds): the segment span is aligned with that of fast parameters (10 seconds), and includes two consecutive spectral data (corresponding to the start and middle of the segment, respectively).

[0044] Meanwhile, segmentation is achieved through a sliding window, with the window sliding step size equal to 1 / 2 of the segment span (e.g., a 10-second segment slides once every 5 seconds), ensuring the continuity of features.

[0045] S230.2 Modal Feature Extraction: For numerical data within each time segment, extract the extreme values, mean values, and changes in adjacent data points within the segment to form a numerical feature vector; for high-dimensional infrared spectral data, extract the wavenumber positions of characteristic peaks and their corresponding absorbance values ​​(based on the standardized wavelength points in S220.3) to form a spectral feature vector. As a further explanation of this step, this embodiment extracts the extreme values, mean values, and changes in adjacent data points within each time segment of numerical data to form a numerical feature vector. The following example is provided in this embodiment. Taking temperature as an example, the following three types of features are extracted within each 10-second segment: Extreme values: the highest and lowest temperatures within a segment; Mean: The arithmetic mean of 10 data points; Change: The temperature difference between the end and beginning of the segment.

[0046] Furthermore, for high-dimensional infrared spectral data, two types of features are extracted within each segment: Characteristic peak location: The actual measured value of the main characteristic wavenumber; Absorbance value: Normalized absorbance corresponding to the position of the characteristic peak .

[0047] S230.3 Cross-modal feature concatenation: The numerical feature vector and the spectral feature vector are concatenated horizontally in the order of "numerical features first, spectral features second". At the same time, each feature is labeled with its corresponding modal source and the time segment identifier, forming a segment-level fusion feature matrix. The matrix dimension is determined by the sum of the number of numerical features and the number of spectral features. As a further explanation of this step, the concatenated feature matrix is ​​a two-dimensional table, with each row corresponding to a time segment, and the columns include: Numerical characteristics: Temperature (extreme values, mean values, changes), pH value (three types of characteristics at the same temperature), conductivity (three types of characteristics at the same temperature). Spectral characteristics: the position of each characteristic wavenumber and its absorbance value; Identifier column: timestamp (start time) corresponding to the segment, modal source (e.g., "temperature-main channel" "spectrum-screen wavenumber").

[0048] S230.4 Feature Validity Screening: By calculating the correlation between each feature in the fusion feature matrix and the viscosity at the reaction endpoint, features with a correlation higher than the baseline value obtained from the statistics of historical normal reaction data are retained, and redundant features with too low correlation are eliminated to form the final fusion feature set used for S300 modeling.

[0049] As a further explanation of this step, the feature validity screening in this embodiment includes the following steps: First, calculate the Pearson correlation coefficient between the fusion characteristics and the reaction endpoint viscosity. For the fusion characteristic matrix... Features ,set up This represents the number of qualified batches in history (representing the total number of production batches included in the statistics). For the first The first batch The Z-score standardized value of each fusion feature (consistent with the standardization rules of S210.2, eliminating dimensional differences). For the first The viscosity measurement at the batch reaction endpoint (determined using industry standard methods) is used to calculate the correlation using the following formula: ; Subsequently, the criteria for determining redundancy features were statistically determined. : From historical normal production data, features that are theoretically not causally related to the modified starch reaction (such as workshop humidity, equipment numbers, etc.) are selected to construct an irrelevant feature set; for each irrelevant feature in the set... ( (where the indices are irrelevant features), repeat the above correlation calculation steps to obtain the correlation set between irrelevant features and the reaction endpoint viscosity. ,That The number of irrelevant features, For the first The correlation calculation results of irrelevant features; Then, using the quantile statistical method, the 95th quantile of set C is taken as the benchmark value θ, expressed by the formula: ,in This is a function used in statistical analysis to calculate quantiles (meaning that only 5% of irrelevant features will have a correlation exceeding this value). Finally, features are selected based on the correlation results and benchmark values. The correlation calculation results for each fused feature are then used. Compared with the benchmark value Compare and retain only those that meet the requirements. Features that eliminate those with a correlation lower than [value missing] The redundant features are used to form a fusion feature set for dynamic modeling of the S300 process.

[0050] S300, Dynamic Modeling and Prediction of Process: An improved multimodal time-series Transformer model is adopted to train and model the fused multimodal real-time sensing data and the viscosity target index at the reaction endpoint, and output the prediction results of feeding rate and reaction time to achieve accurate fitting of the dynamic characteristics of the modified starch modification process. In this step, the improved multimodal temporal Transformer model structure in S300 includes the following steps: S310.1, Multimodal Input Encoding: The fused feature set output from S200 (including numerical features and infrared spectral features) and the reaction endpoint viscosity target index from S120.4 are used as model inputs. The numerical features are converted to a dimension of [missing information - likely a specific value]. The embedded vectors and infrared spectral features are extracted using a 1D convolutional layer (with kernel size matching the normalized wavelength point spacing in S220.3) to extract local features and then converted into dimensions. The embedding vector, the viscosity target index is encoded as a dimension The supervision vector; all vectors are appended with a location code based on the S110.3 timestamp, as follows: ; ; in, This is the location-coded value. The timestamp in S110.3 For the dimension index of the embedded vector, and , The total dimension of the embedded vector; As a further explanation of this step, the embedding vector dimension in this embodiment... Based on the dimensionality of the fused feature set (including the number of numerical features and spectral features) and the model complexity requirements, You can select 64, 128, or 256 (the industry standard range for dimensions). For example, if the fused features include 20 numerical features and 50 spectral features, It can be set to 128 to balance expressive power and computational efficiency; Meanwhile, the kernel size in this embodiment is determined based on the normalized wavelength interval in S220.3. If the wavelength interval is... Interval filtering; the kernel size can be set to 3 or 5 (covering local features of 3-5 adjacent wavelength points); the number of convolution output channels and... Consistency is ensured to guarantee dimensional matching after transformation.

[0051] Furthermore, in this embodiment This refers to the timestamp in S110.3 (with the start time of the current reaction cycle as 0, in seconds). For example, if data is collected 100 seconds after the start of the reaction, then... For multiple batches of data, the calculation is based on the relative time within each batch to avoid the impact of absolute time differences on the time sequence correlation.

[0052] S310.2 Cross-modal attention calculation: Introducing modal weight moments in the Transformer encoder ( (Distinguishing between numerical, spectral, and viscous targets), the cross-modal attention calculation formula is as follows: ; in, For modality The query vector, For modality key vector ( ), For modality The value vector, For modality The weight matrix (initialized based on the correlation coefficients between each mode and viscosity in S230.4). This is the normalization function; This represents the matrix transpose operation; Function encapsulation representing cross-attention operations; As a further explanation of this step, the modal weighting moments in this embodiment... The initialization specifically includes: based on the correlation coefficients between each mode and viscosity in S230.4, the initial values ​​of the weight matrices for numerical features (temperature, pH, etc.) and spectral features are allocated proportionally according to the correlation degree. For example, if the average correlation coefficient between numerical features and viscosity is 0.6, and that of spectral features is 0.4, then... The initial value is set to 0.6 times the base matrix. Set to 0.4 times the reference matrix (the reference matrix is ​​randomly initialized). matrix).

[0053] S310.3, Dual Output Head Mapping: The decoder has two output heads, and the feed rate prediction formula is as follows: The formula for predicting reaction time is: ;in, To predict the feeding rate, To predict reaction time, For decoder output features, , These are the weight matrix and bias term of the feed rate head, respectively. , These are the weight matrix and bias term of the reaction time head, respectively; Within the frequency adjustment range of the feed pump in the S400, Based on the truncation of the reaction kinetics period range in S120.1.

[0054] As a further explanation of this step, regarding the initialization of weights and biases, , , , Xavier initialization is used (to avoid unstable training caused by initial values ​​that are too large or too small), where and for Matrix (used to output single-value predictions) , It is a scalar; at the same time, the range of the prediction results is constrained by the feeding rate. The reaction time must be limited to the hardware adjustment range of the feed pump (e.g., 0-50Hz, specifically set according to the actual equipment parameters). If it exceeds this range, the boundary value should be used. The predicted value is then truncated based on the reaction kinetic cycle range (e.g., 30-120 minutes) in S120.1 to ensure that the predicted value is within the time range allowed by the process.

[0055] In this step, model training and modeling in S300 includes the following steps: S320.1 Training Sample Construction: The sample data is associated with "multimodal real-time sensing data fragments - corresponding viscosity values" integrated by S120.4. The input is the fused feature fragment processed by S200 (same as the filtering result of S230.4), and the label is the actual feeding rate of the corresponding batch. With reaction time ; Based on the reaction stages divided by reaction kinetic characteristics (same as the reaction stage division logic in S220.2), stratified sampling is used, and the sample proportion of each stage is consistent with the distribution of historical production batches. The samples are divided into training set and validation set, where the sample size of the training set is no less than 70% of the total sample size to ensure that the model learns the correlation between multimodal features and process parameters. The validation set is used to independently evaluate the model's generalization ability. As a further explanation of this step, in this embodiment, sampling is performed according to the historical batch distribution of the reaction stage (initial, middle and final stages). For example, if the three stages account for 20%, 60% and 20% of the historical data respectively, then the training set and the validation set are divided in the same proportion to ensure that the characteristics of each stage are evenly reflected in the samples.

[0056] Furthermore, in this embodiment, the total sample size is no less than 30 batches (covering different raw material batches and environmental conditions), of which the training set is ≥21 batches and the validation set is ≥9 batches, to meet the basic learning needs of the model. The samples can be stored in the form of "feature matrix + label", where each row of the feature matrix corresponds to the fusion feature of a time segment, and the label is the actual feeding rate and reaction time corresponding to that segment.

[0057] S320.2, Loss Function Definition: A weighted joint loss function is used, and the formula is as follows: ; in, MSE loss due to feed rate, ; MAE loss due to reaction time, ; For the sample size, , The first The actual feeding rate and reaction time of each sample , These are the corresponding predicted values; , These are the weighting coefficients, and ; ( , Based on the correlation between the two parameters and viscosity in S120.4, the parameter with a higher correlation has a greater weight. As a further explanation of this step, in this embodiment... , The correlation between the two parameters and viscosity in S120.4 is used to determine this; for example, if the fluctuation in the feeding rate accounts for approximately 60% of the impact on viscosity and the reaction time accounts for approximately 40% in historical data, then... The correlation degree is determined by analyzing the linear regression coefficients of parameter deviation and viscosity deviation; the larger the absolute value of the coefficient, the higher the weight.

[0058] S320.3 Training Parameter Update: The Adam optimizer is used to update the model parameters. The parameter update formula is as follows: ; in, For the first Wheel model parameters, For the updated parameters, The learning rate (dynamically adjusted based on the validation set loss) is used. For first-order momentum estimation, For second-order momentum estimation, To prevent tiny constants with a denominator of zero; After each training round, the validation set loss is calculated. If the loss does not decrease for several consecutive rounds, the learning rate is adjusted until training terminates, and the model parameters with the minimum validation set loss are saved.

[0059] As a further explanation of this step, in this embodiment... Based on dynamic adjustment of validation set loss, the initial... Set to 0.001 (Adam optimizer's standard initial value), and calculate the validation set loss after each training round: If losses decrease, maintain constant; If losses do not decrease for three consecutive rounds Halve the value (e.g., 0.0005, 0.00025, etc.) until... Down to (Avoid overfitting).

[0060] In the steps, the dynamic characteristic fitting and prediction output in S300 includes the following steps: S330.1, Real-time Input Adaptation: The fusion feature set processed by S200 within the current reaction cycle (same as the screening result of S230.4) is converted into a model input vector according to the encoding method of S310.1, and the viscosity target index at the reaction endpoint of this cycle is simultaneously connected (same as the connection logic of S120.3); the time dimension of the input vector corresponds one-to-one with the timestamp in S110.3 to ensure temporal consistency; As a further explanation of this step, the update frequency of the model input vector in this embodiment is synchronized with the data acquisition frequency in S110.3. For example, if the temperature data is acquired once every second, the real-time input vector is updated once every second to ensure that the model input matches the reaction process in real time. Meanwhile, when encoding the target viscosity value for the current period into a supervision vector, the same standardization method as historical data is used (subtracting the historical target mean and dividing by the standard deviation) to maintain encoding consistency.

[0061] S330.2, Rolling Prediction Execution: A sliding window mechanism is used to generate the prediction sequence, with a window length of... The formula for determining it is: ; in, The minimum lag time for the reaction (determined based on the time-series correlation analysis of multimodal data and viscosity in S120.4). This is the highest sampling frequency. It is a rounding function; The model is based on the most recent The fusion features of each acquisition cycle serve as the input context, and the future features are output every one acquisition cycle. Feed rate prediction sequence for each cycle and reaction time prediction sequences The predicted value ranges are respectively matched with the feed pump adjustment range of S400 and the reaction cycle range of S120.1; Furthermore, this embodiment provides a window length. The calculation example is as follows: If the minimum lag time of the reaction Historical data shows that the highest sampling frequency is 10 seconds. (1 time / second), then (Rounded up to 10 periods). In practical applications, It can be determined by analyzing the time difference between parameter mutations and viscosity changes in multiple batches of data; At the same time, every acquisition cycle (e.g., 1 second) outputs the predicted values ​​for the next 10 cycles. For example, at the 10th second, the predicted feeding rate and reaction time for 11-20 seconds are output, and at the 11th second, the predicted values ​​for 12-21 seconds are output, ensuring that the rolling updates cover the entire reaction cycle.

[0062] S330.3, Fitting Accuracy Verification: The similarity between the predicted sequence and the actual process parameter sequence is calculated using the dynamic time warping algorithm. The formula is as follows: ; in, For the predicted sequence, This is the actual parameter sequence (corresponding to the actual feeding rate or reaction time). For time-ordered paths, The length of the normalized sequence; It is a dynamic time warping algorithm; When the DTW value exceeds the baseline threshold based on historical qualified batch statistics, incremental model training is triggered (the output header parameters in S310.3 are fine-tuned using only the current period data) to ensure the accuracy of dynamic characteristic fitting.

[0063] As a further explanation of this step, this embodiment provides the following example for constructing the baseline threshold: For example, collect full-cycle data (multimodal time-series features + actual feeding rate / reaction time series) from more than 30 batches of qualified products. First, for each batch of data, calculate the Dynamic Time Warping (DTW) value between the model-predicted sequence and the actual sequence (measures the similarity of the time-series curves; the smaller the value, the better the fit). Then, sort the 30 DTW values ​​and take the 90th percentile as the benchmark (e.g., a statistical threshold of 5.0). When the real-time predicted DTW value exceeds 5.0, it is determined that the fitting accuracy is insufficient, triggering incremental training.

[0064] S400 Closed-loop control command execution: Based on the feeding rate and reaction time output by the improved multimodal temporal Transformer model trained and modeled in S300, the feeding rate is adjusted by controlling the frequency of the feeding pump, and the reaction time is adjusted by controlling the running time of the reactor stirring motor, outputting automatic adjustment commands; and a closed-loop control is formed based on real-time feedback multimodal real-time sensing data to improve the stability of modified starch quality.

[0065] like Figure 3 As shown, in this step, the execution of the closed-loop control command in S400 includes the following steps: S410.1 Control Parameter Conversion: Convert the feeding rate and reaction time prediction results output by S300 into control parameters that can be recognized by the corresponding equipment. The feeding rate corresponds to the frequency adjustment value of the feeding pump, and the reaction time corresponds to the running time setting value of the stirring motor of the reactor. As a further explanation of this step, when converting the predicted feeding rate and reaction time from the model output into controllable parameters of the equipment in this embodiment, the feeding rate and the feeding pump frequency are correlated through a linear mapping: based on the "rated conveying capacity-frequency correspondence" marked on the feeding pump nameplate (e.g., "0-50Hz corresponds to 0-100kg / h material conveying"), the target frequency is calculated according to the following formula: ; The reaction time is directly used as the target running time of the stirring motor and is timed through the corresponding PLC timer module. If the predicted value exceeds the equipment hardware range (e.g., pump frequency exceeds 0-50Hz, reaction time is shorter than the minimum process cycle of 30 minutes), the boundary value (50Hz, 0Hz, or 30 minutes) is taken, and an alarm is triggered to assist in subsequent deviation analysis.

[0066] S410.2 Command Sending and Execution Confirmation: The converted control parameters are sent to the feed pump controller and the stirring motor driver via the industrial control bus. During the sending process, the communication link self-checking mechanism in S110.2 is enabled to ensure the integrity of the command transmission. After receiving the command, the feed pump and the reactor stirring motor return a status response. If a valid response is not received within the set time limit, the backup control channel is activated to resend the command until it is confirmed that the command has been correctly received. As a further explanation of this step, the instruction transmission in this embodiment adopts an industrial control bus redundancy scheme (e.g., ModbusTCP as the main channel and RS485 as the backup channel): After the main channel sends an instruction, a reasonable time limit is set (based on the characteristics of bus communication delay) to wait for the device status response; if the time limit is exceeded, it automatically switches to the backup channel to resend the instruction. After receiving the instruction, the device returns a status code (e.g., "success" or "parameter out of bounds"). If the status code indicates success, record "Instruction executed successfully"; If the status code indicates that the parameter is abnormal, recalculate the valid parameter and send it in combination with the device hardware constraints (such as the upper / lower limit of pump frequency); If there is still no effective response after 3 resentments, an audible and visual alarm will be triggered, and a pop-up window from the central control system will prompt manual troubleshooting of communication link or equipment hardware failure.

[0067] S410.3 Real-time feedback data acquisition: After the command is executed, according to the synchronous acquisition rules of S110.3, the actual operating frequency of the feed pump, the working status of the stirring motor and the real-time sensing data of the multi-modal in the reactor are acquired in real time. All feedback data are appended with the timestamp corresponding to the control command to ensure timing consistency. As a further explanation of this step, after the command is executed, the feed pump can provide feedback on the actual operating frequency through a built-in frequency sensor (such as a Hall sensor), and the stirring motor can provide feedback on the working duration through an encoder. Simultaneously, the multimodal synchronous acquisition rules of S110.3 are reused to acquire data such as temperature, pH, and infrared spectrum inside the reactor. All feedback data are appended with a unique ID consistent with the control command and stored in the InfluxDB real-time database, realizing end-to-end timing correlation of "command-execution-feedback". S410.4 Operational Deviation Analysis: After the feedback data is processed and standardized by S200 outlier handling, the deviations of the actual feeding rate and reaction time from the predicted values ​​are compared. At the same time, the characteristic correlation analysis method of S230.4 is used to determine whether the deviation affects the viscosity target index at the reaction endpoint. As a further explanation of this step, the deviation analysis in this embodiment is based on the statistical regularity of historical qualified production data: by analyzing the correlation between "feeding rate deviation, reaction time deviation" and "reaction endpoint viscosity" in multiple batches of production (such as trend consistency and fluctuation coupling), it is determined whether the current deviation will affect the endpoint viscosity target. If the impact of the deviation on the final viscosity is within the expected fluctuation range of the process (e.g., similar deviations in historical data have not caused product non-compliance), it is judged as "normal fluctuation" and no correction is triggered for the time being. If the deviation exceeds the process tolerance range (e.g., similar deviations in historical data have caused viscosity to deviate from the target), it is judged as an "abnormal deviation that needs to be corrected".

[0068] S410.5 Closed-loop parameter correction: If the deviation exceeds the threshold, based on the current multimodal feedback data and the real-time prediction results of the S300 model, the corrected control parameters are generated. In the next acquisition cycle, the instructions of the feed pump and the reactor stirring motor are updated. By repeatedly adjusting until the deviation falls back to within the threshold, a continuous closed-loop control is formed to ensure the stability of the modified quality.

[0069] As a further explanation of this step, this embodiment adopts a tiered adjustment strategy for "abnormal deviations that need to be corrected": Slight deviation (small deviation amplitude, weak historical correlation): Update the control parameters for the next cycle by proportional adjustment (such as multiplying the deviation value by an empirical coefficient); Moderate deviation (large deviation amplitude, significant correlation effect): Call the S300 real-time prediction model, input the current feedback data (such as actual pump frequency, real-time multimodal data), and re-predict the optimal control parameters; Severe deviation (deviation exceeds hardware / process boundaries): Force the parameters to be adjusted to the hardware-allowed boundary values ​​(e.g., the pump frequency is set to the rated upper limit), and mark the production cycle as "key monitoring" to assist manual inspection.

[0070] Simultaneously, the correction process iterates synchronously with the data acquisition frequency of S110.3 (e.g., 1 second / time) until it is adjusted to: The deviation has fallen back to the "expected fluctuation range of the process"; When the reaction enters the "end point stage" (such as the critical range where the viscosity is close to the target value), the closed-loop correction is terminated, and stable operation is maintained.

[0071] Those skilled in the art will understand that the process of implementing all or part of the steps of the above embodiments can be carried out by hardware or by a program instructing the relevant hardware.

[0072] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent control of modified starch processing based on multimodal data fusion, characterized in that, Includes the following steps: S100, Multimodal Data Acquisition: Through online sensing devices, multimodal real-time sensing data including but not limited to pH value, conductivity, infrared spectrum data and real-time temperature are synchronously acquired at a preset frequency. At the same time, the target index of viscosity at the reaction endpoint is also connected to ensure the real-time performance of data acquisition and the integrity of multimodal coverage. S200, multimodal real-time sensing data fusion processing: through 3 The system employs outlier detection, standardization transformation, and feature stitching strategies to fuse multimodal real-time sensing data collected by the S100, ensuring the spatiotemporal consistency and feature validity of the multimodal real-time sensing data. S300, Dynamic Modeling and Prediction of Process: An improved multimodal time-series Transformer model is adopted to train and model the fused multimodal real-time sensing data and the viscosity target index at the reaction endpoint, and output the prediction results of feeding rate and reaction time to achieve accurate fitting of the dynamic characteristics of the modified starch modification process. S400 Closed-loop control command execution: Based on the feeding rate and reaction time output by the improved multimodal temporal Transformer model trained and modeled in S300, the feeding rate is adjusted by controlling the frequency of the feeding pump, and the reaction time is adjusted by controlling the running time of the reactor stirring motor, outputting automatic adjustment commands; and a closed-loop control is formed based on real-time feedback multimodal real-time sensing data to improve the stability of modified starch quality.

2. The intelligent control method for modified starch processing based on multimodal data fusion according to claim 1, characterized in that, The deployment and data acquisition of the online sensing device in S100 includes the following steps: S110.1, Sensor Equipment Deployment and Positioning: Based on the material flow field distribution and reaction characteristics in the modified starch reaction vessel, pH and conductivity sensors are installed in the turbulent zone in the lower part of the vessel. The infrared spectrometer probe is aligned with the uniformly mixed area of ​​the material in the vessel through the optical window. Temperature sensors are arranged in layers along the axial direction of the vessel. Viscosity detection equipment is integrated into the discharge pipeline at the reaction endpoint. S110.2 Pre-acquisition calibration and self-test: Before data acquisition, perform calibration according to the sensor type, and at the same time complete the self-test of the device communication link, power supply stability and measurement range; S110.3 Multimodal real-time sensing data synchronous acquisition: Based on the real-time clock of the reactor control system, each sensor is triggered to synchronously acquire data at a preset frequency. During the acquisition process, a timestamp is added to each set of data. The acquisition interval for parameters with fast dynamic response is shorter than that for parameters with slow dynamic response. S110.4 Real-time verification of acquired data: Real-time validity verification of acquired multimodal real-time sensing data, including whether the data is within the normal measurement range of the sensor, whether the data changes in adjacent acquisition cycles conform to the law of reaction kinetics, and marking data with abnormal verification and initiating a supplementary acquisition mechanism.

3. The intelligent control method for modified starch processing based on multimodal data fusion according to claim 2, characterized in that, The frequency preset and access target indicators in S100 include the following steps: S120.1 Multimodal data acquisition frequency grading preset: Based on the kinetic characteristics of the modified starch modification reaction and combined with the dynamic response differences of each parameter, the multimodal real-time sensing data is divided into different acquisition frequency levels. The preset frequency of the reaction sensitive parameters is higher than that of the non-sensitive parameters. The initial frequency range is determined by statistical analysis of the parameter change rate in the historical reaction cycle. S120.2 Frequency Dynamic Adaptation Mechanism: During the acquisition process, the change amplitude of multimodal real-time sensing data is monitored in real time. When the change amplitude of a certain parameter exceeds the normal range of the corresponding reaction stage, the acquisition frequency of that parameter is automatically increased. After the change amplitude falls back to the normal range, the initial frequency is restored. The trigger threshold is adjusted based on the historical data distribution of similar reactions. S120.3, Reaction endpoint viscosity target index access: Collect reaction endpoint viscosity data at a preset cycle through a dedicated detection interface, and record the precise timestamp when accessing to ensure that it corresponds to the time dimension of the multimodal real-time sensing data; S120.4, Time-series integration of target indicators and multimodal real-time sensing data: The accessed reaction endpoint viscosity target indicator and the multimodal real-time sensing data collected at the same time are aligned with the timestamp to form a "multimodal real-time sensing data segment - corresponding viscosity value" associated sample. During the association process, multimodal real-time sensing data that is judged to be abnormal by real-time verification is removed. The integration result is used for the construction of samples for subsequent data fusion.

4. The intelligent control method for modified starch processing based on multimodal data fusion according to claim 3, characterized in that, In S200, through 3 The outlier detection criteria involve fusing multimodal real-time sensing data, including the following steps: S210.1, Algorithm Formula Definition: Single-modal real-time sensing data sequence acquired by S100 ,in For the first The raw data at each collection point Given the length of the data sequence; and calculate the mean within the sliding window. and standard deviation The formula is: ； ； in, The length of the sliding window. For indexing the data within the window, Index the current data point; when At that time, the judgment This is an outlier; S210.2 Algorithm Parameter Adaptation: The window sliding step size is consistent with the acquisition period of the corresponding modal data to ensure a balance between real-time performance and data continuity; the same modal data can be dynamically adjusted at different stages of the reaction. The value is adjusted based on the data fluctuation characteristics of that period; S210.3, Outlier Handling Procedure: For single-value anomalies, linear interpolation replacement is used: ;in, For corrected data; for multiple consecutive outliers, trigger a status re-inspection of the corresponding sensor, during which backup sensor channel data is enabled and the data source attribute is marked.

5. The intelligent control method for modified starch processing based on multimodal data fusion according to claim 4, characterized in that, The normalization conversion in S200 includes the following steps: S220.1, Algorithm Formula Definition: For single-modal real-time sensing data after outlier processing... Converted using Z-score standardization The formula is: ； in, This is a reference average. The reference standard deviation; S220.2, Reference Parameter Determination: and The statistical sample consists of historical qualified product data of the same modality, covering different raw material batches and process parameter combinations; the results are calculated separately for each reaction stage. and This ensures that the data characteristics are standardized and adapted to different reaction stages. S220.3, High-dimensional data adaptation: For high-dimensional infrared spectral data, each wavelength point Implement standardization separately: ； in, , The first Reference mean and standard deviation for each wavelength point; To standardize absorbance; The relative intensity relationship of the spectral characteristic peaks retained after conversion.

6. The intelligent control method for modified starch processing based on multimodal data fusion according to claim 5, characterized in that, The S200 performs fusion processing on multimodal real-time sensing data through a feature splicing strategy, including the following steps: S230.1, Time Segment Division: The multimodal real-time sensing data after outlier processing and standardization is divided into continuous time segments according to reaction kinetic characteristics, with each segment covering the same time span; S230.2 Modal Feature Extraction: For numerical data within each time segment, extract the extreme values, mean values, and changes in adjacent data points within the segment to form a numerical feature vector; for high-dimensional data such as infrared spectra, extract the wavenumber positions of characteristic peaks and their corresponding absorbance values ​​to form a spectral feature vector. S230.3 Cross-modal feature concatenation: The numerical feature vector and the spectral feature vector are concatenated horizontally in the order of "numerical features first, spectral features second". At the same time, each feature is labeled with its corresponding modal source and the time segment identifier, forming a segment-level fusion feature matrix. The matrix dimension is determined by the sum of the number of numerical features and the number of spectral features. S230.4 Feature Validity Screening: By calculating the correlation between each feature in the fusion feature matrix and the viscosity at the reaction endpoint, features with a correlation higher than the baseline value obtained from the statistics of historical normal reaction data are retained, and redundant features with too low correlation are eliminated to form the final fusion feature set used for S300 modeling.

7. The intelligent control method for modified starch processing based on multimodal data fusion according to claim 6, characterized in that, In S300, the improved multimodal temporal Transformer model structure includes the following steps: S310.1, Multimodal Input Encoding: The fused feature set output from S200 and the reaction endpoint viscosity target index from S120.4 are used as model inputs, where numerical features are converted to a dimension of [dimensional value missing] through a fully connected layer. The embedding vector, infrared spectral features are converted into dimensions after local features are extracted through a 1D convolutional layer. The embedding vector, the viscosity target index is encoded as a dimension The supervision vector; all vectors are appended with a location code based on the S110.3 timestamp, as follows: ； ； in, This is the location-coded value. The timestamp in S110.3 For the dimension index of the embedded vector, and , The total dimension of the embedded vector; S310.2 Cross-modal attention calculation: Introducing modal weight moments in the Transformer encoder The formula for calculating cross-modal attention is: ； in, For modality The query vector, For modality The key vector and , For modality The value vector, For modality The weight matrix, This is the normalization function; This represents the matrix transpose operation; Function encapsulation representing cross-attention operations; S310.3, Dual Output Head Mapping: The decoder has two output heads, and the feed rate prediction formula is as follows: The formula for predicting reaction time is: ;in, To predict the feeding rate, To predict reaction time, For decoder output features, , These are the weight matrix and bias term of the feed rate head, respectively. , These are the weight matrix and bias term of the reaction time head, respectively; Within the frequency adjustment range of the feed pump in the S400, Based on the truncation of the reaction kinetics period range in S120.

1.

8. The intelligent control method for modified starch processing based on multimodal data fusion according to claim 7, characterized in that, The model training and modeling in S300 includes the following steps: S320.1 Training Sample Construction: The sample data is associated with "multimodal real-time sensing data fragments and corresponding viscosity values" integrated by S120.

4. The input is the fused feature fragment processed by S200, and the label is the actual feeding rate of the corresponding batch. With reaction time ; Stratified sampling based on reaction kinetic characteristics, with the sample proportion of each stage consistent with the distribution of historical production batches, is used to divide the samples into training and validation sets. The training set contains no less than 70% of the total samples to ensure that the model learns the correlation between multimodal features and process parameters. The validation set is used to independently evaluate the model's generalization ability. S320.2, Loss Function Definition: A weighted joint loss function is used, and the formula is as follows: ； in, Represents the loss function. MSE loss due to feed rate, ; MAE loss due to reaction time, ; For the sample size, , The first The actual feeding rate and reaction time of each sample , These are the corresponding predicted values; , These are the weighting coefficients, and ; S320.3 Training Parameter Update: The Adam optimizer is used to update the model parameters. The parameter update formula is as follows: ； in, For the first Wheel model parameters, For the updated parameters, For learning rate, For first-order momentum estimation, For second-order momentum estimation, To prevent tiny constants with a denominator of zero; After each training round, the validation set loss is calculated. If the loss does not decrease for several consecutive rounds, the learning rate is adjusted until training terminates, and the model parameters with the minimum validation set loss are saved.

9. The intelligent control method for modified starch processing based on multimodal data fusion according to claim 8, characterized in that, The dynamic characteristic fitting and prediction output in S300 includes the following steps: S330.1, Real-time Input Adaptation: The fusion feature set processed by S200 within the current reaction cycle is converted into a model input vector according to the encoding method of S310.1, and the viscosity target index at the reaction endpoint of this cycle is simultaneously connected; the time dimension of the input vector corresponds one-to-one with the timestamp in S110.3 to ensure temporal consistency; S330.2, Rolling Prediction Execution: A sliding window mechanism is used to generate the prediction sequence, with a window length of... The formula for determining it is: ； in, For the minimum lag time of the reaction, This is the highest sampling frequency. It is a rounding function; The model is based on the most recent The fusion features of each acquisition cycle serve as the input context, and the future features are output every one acquisition cycle. Feed rate prediction sequence for each cycle and reaction time prediction sequences The predicted value ranges are respectively matched with the feed pump adjustment range of S400 and the reaction cycle range of S120.1; S330.3, Fitting Accuracy Verification: The similarity between the predicted sequence and the actual process parameter sequence is calculated using the dynamic time warping algorithm. The formula is as follows: ； in, For the predicted sequence, For the actual parameter sequence, For time-ordered paths, The length of the normalized sequence; It is a dynamic time warping algorithm; When the DTW value exceeds the baseline threshold based on historical qualified batch statistics, incremental training of the model is triggered to ensure the accuracy of dynamic characteristic fitting.

10. The intelligent control method for modified starch processing based on multimodal data fusion according to claim 1, characterized in that, The execution of the closed-loop control command in S400 includes the following steps: S410.1 Control Parameter Conversion: Convert the feeding rate and reaction time prediction results output by S300 into control parameters that can be recognized by the corresponding equipment. The feeding rate corresponds to the frequency adjustment value of the feeding pump, and the reaction time corresponds to the running time setting value of the stirring motor of the reactor. S410.2 Command Sending and Execution Confirmation: The converted control parameters are sent to the feed pump controller and the stirring motor driver via the industrial control bus. During the sending process, the communication link self-checking mechanism in S110.2 is enabled to ensure the integrity of the command transmission. After receiving the command, the feed pump and the reactor stirring motor return a status response. If a valid response is not received within the set time limit, the backup control channel is activated to resend the command until it is confirmed that the command has been correctly received. S410.3 Real-time feedback data acquisition: After the command is executed, according to the synchronous acquisition rules of S110.3, the actual operating frequency of the feed pump, the working status of the stirring motor and the real-time sensing data of the multi-modal in the reactor are acquired in real time. All feedback data are appended with the timestamp corresponding to the control command to ensure timing consistency. S410.4 Operational Deviation Analysis: After the feedback data is processed and standardized by S200 outlier handling, the deviations of the actual feeding rate and reaction time from the predicted values ​​are compared. At the same time, the characteristic correlation analysis method of S230.4 is used to determine whether the deviation affects the viscosity target index at the reaction endpoint. S410.5 Closed-loop parameter correction: If the deviation exceeds the threshold, based on the current multimodal feedback data and the real-time prediction results of the S300 model, the corrected control parameters are generated. In the next acquisition cycle, the instructions of the feed pump and the reactor stirring motor are updated. By repeatedly adjusting until the deviation falls back to within the threshold, a continuous closed-loop control is formed to ensure the stability of the modified quality.

Images