Intelligent control system and method for full-automatic milk cover machine

Abstract

The invention relates to the technical field of intelligent beverage preparation and control, in particular to an intelligent control system and method for a full-automatic milk cover machine. According to the method, temperature, conductivity, power and acoustic signals are synchronously collected through a multi-mode sensor, a short-time coupling index is proposed, a sensory quality index MQI is constructed by using support vector regression, and quantitative characterization of the taste of the milk cover is realized. Based on Kalman short-term prediction, the system can correct control parameters in advance before a quality degradation trend appears, and a feedforward-prediction-closed loop linkage regulation and control mechanism is constructed. Furthermore, the multi-target balance of quality and energy efficiency is completed by combining constrained digital twinning and Bayesian optimization, and adaptive updating is performed on different milk types through a migration and incremental learning mechanism, so that the defects of lagging regulation and control, too high energy consumption and lack of expansibility of existing equipment are overcome, and the method is suitable for popularization and application. And the consistency, the energy efficiency level and the intelligent degree of finished products are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent beverage preparation and control, in particular to an intelligent control system and method for a full-automatic milk cover machine. BACKGROUND

[0002] With the rapid development of the instant beverage industry, milk cover beverages are widely favored due to their unique taste and quality requirements. In order to ensure the cup efficiency and quality stability, full-automatic milk cover production equipment is of great significance. It not only reduces the dependence on manpower, but also realizes fine control of multiple parameters such as temperature, stirring power and time, thereby improving the standardization level of beverages and user experience.

[0003] However, the existing milk cover machines mostly adopt fixed programs or single-parameter adjustment mode, which is difficult to cope with the dynamic changes of different raw material batches, environmental temperature and equipment running state, resulting in insufficient consistency of product taste. At the same time, some equipment lack real-time sensing and prediction ability, and the regulation strategy is lagging behind, which is easy to cause problems such as high energy consumption, excessive stirring or insufficient heating, affecting the quality of finished products. In addition, the lack of comprehensive utilization of multi-modal data and intelligent optimization mechanism makes the system lack adaptability and expansibility, which is difficult to meet the beverage production needs of multiple categories and multiple scenes. SUMMARY

[0004] The present application provides an intelligent control system and method for a full-automatic milk cover machine, which realizes multi-modal data fusion and dynamic optimization regulation, improves the quality stability, energy efficiency level and self-adaptive ability of the milk cover production process, and overcomes the shortcomings of existing equipment in consistency and intelligence.

[0005] To achieve the above purpose, the present application provides the following technical solutions: An intelligent control system for a full-automatic milk cover machine, comprising: A data acquisition and processing module for acquiring conductivity, temperature, stirring power and acoustic time sequence signals and performing time alignment, temperature correction and normalization processing to obtain multi-channel time sequence data, and calculating a short-time coupling index based on the multi-channel time sequence data; A predictive control module for mapping the short-time coupling index to a milk cover quality index MQI through a regression model, using a state space model combined with Kalman filtering to perform rolling prediction of the short-term evolution of the MQI, and dynamically adjusting the control parameter sequence in advance when the prediction shows that a gelation critical mode will occur within a predetermined lookahead time; A simulation optimization module for searching for an optimal control parameter sequence in a digital twin simulation environment using Bayesian optimization, and using the searched optimal control parameter sequence as the initial value for on-site parameter adjustment; An adaptive learning module is configured to fine-tune the regression model by using transfer learning for new milk sources and update the regression model and the decision threshold after each batch processing by using incremental learning, thereby constructing an adaptive optimized control strategy library.

[0006] As a preferred technical solution of the present application, the step of obtaining multi-channel time series data comprises: The conductivity, temperature, stirring power and acoustic time series signals are collected. The signals of each channel are aligned according to the sampling time stamp, the time offset is determined by using the cross-correlation function, and the interpolation resampling is performed to the same time axis. The temperature correction based on the factory calibration curve is applied to the conductivity signal, and the correction result is juxtaposed with the temperature signal. Each channel signal is denoised and baseline corrected, and the initial steady state interval is used as a reference to complete the normalization processing, thereby obtaining the processed and time-synchronized multi-channel time series data.

[0007] As a preferred technical solution of the present application, the calculation of the short-time coupling index comprises: The multi-channel time series data is segmented according to the short-time window and the window sequence is slidingly intercepted according to the fixed overlap rate. The power-related features, temperature-related features and acoustic energy-related features are calculated in each time window. The short-time correlation quantities of the power-related features, the temperature-related features and the acoustic-related features are calculated in each window. The short-time correlation quantities are linearly combined according to the weights obtained by offline calibration and smoothed, thereby obtaining the short-time coupling index, and a short-time coupling index sequence is generated.

[0008] As a preferred technical solution of the present application, the step of mapping the short-time coupling index to the milk cover quality index MQI by using the regression model comprises: The short-time coupling index sequence is used as the input vector of the regression model. The input vector is placed into the regression model calibrated offline, and the model outputs the normalized MQI value and the corresponding confidence. The regression model is trained by sensory calibration, and the output MQI corresponds to the pre-defined quality level range.

[0009] As a preferred technical solution of the present application, the step of dynamically adjusting the control parameter sequence in advance comprises: The MQI change rate is calculated according to the MQI value, the MQI and its change rate are constructed as a state vector, a linear state space model is established, and a Kalman filter is used to perform rolling prediction on the MQI evolution trajectory in the future predetermined time window. When the prediction results show that the MQI will drop below the gelation critical threshold within the look-ahead time, the parameter dynamic adjustment mechanism is triggered. Based on the predicted MQI decline rate and current milk source characteristics, the corresponding control parameter sequence is matched from the pre-built control strategy library; The parameter settings in the control parameter sequence are executed sequentially in chronological order, and a first-order filter is used to achieve a smooth transition during the parameter change process. After each step of parameter adjustment is executed, the measured MQI value is compared with the predicted value of that step, and the subsequent unexecuted control parameter sequence is corrected according to the magnitude of the deviation.

[0010] As a preferred embodiment of the present invention, the simulation optimization module performs the following steps: An approximate model of the current operating condition is established in the digital twin, and the control parameters currently in operation on site are used as the initial settings of the simulation model; The initial parameter sample set was generated using the Latin hypercube sampling method and simulated one by one in a twin environment. The MQI estimate and energy consumption index of the simulation output were recorded. The surrogate model is updated based on the simulation results and the next set of parameters to be measured is selected according to the acquisition criteria. The process is iterated until the predetermined termination condition is reached. Select the optimal parameter set from the iteration results and write the parameter set into the field parameter tuning initial value library.

[0011] As a preferred embodiment of the present invention, the transfer learning step includes: Several batches of short-term coupling indices and corresponding measured MQI values ​​from new milk sources were collected as fine-tuning samples. Using the existing regression model as the initial model, perform a small number of iterations of parameter adjustment based on fine-tuned samples and verify the predictive performance of the fine-tuned model; When the fine-tuned model meets the pre-defined performance criteria, the current regression model is replaced and the decision threshold set is updated simultaneously.

[0012] As a preferred embodiment of the present invention, the incremental learning step includes: After each batch is completed, the short-term coupling index sequence, measured MQI value and control trajectory of that batch are added to the online sample set; The regression model is updated incrementally according to a predetermined update rule, with the update magnitude limited to maintain model stability. A sliding update strategy is used for the judgment threshold to gradually absorb statistical changes between batches; When the prediction error exceeds the tolerance and persists, offline retraining is triggered and the retraining results are written to the model library.

[0013] This invention also proposes an intelligent control method for a fully automatic milk foam machine, comprising: Conductivity, temperature, stirring power and acoustic timing signals are collected and time-aligned, temperature-corrected and normalized to obtain multi-channel timing data. Short-time coupling index is calculated based on the multi-channel timing data. The short-term coupling index is mapped to the milk foam quality index (MQI) through a regression model. The short-term evolution of the MQI is predicted using a state-space model combined with Kalman filtering. When the prediction shows that a gelation critical mode will appear within a predetermined look-ahead time, the control parameter sequence is dynamically adjusted in advance. In the digital twin simulation environment, Bayesian optimization is used to search for the optimal control parameter sequence, and the optimal control parameter sequence obtained by the search is used as the initial value for on-site parameter tuning. For new milk sources, transfer learning is used to fine-tune the regression model with a small number of samples, and incremental learning is used to update the regression model and the judgment threshold after each batch of processing to build an adaptive optimization control strategy library.

[0014] The beneficial effects of this invention are: 1. This invention constructs a sensory quality metric based on multimodal short-time coupling indexes and maps it to MQI using SVR. At the same time, it introduces Kalman short-term prediction to achieve feedforward-prediction-closed-loop linkage control, enabling the system to correct key control parameters in advance when raw material batches and environmental changes occur. This significantly improves the consistency of the taste of the finished milk foam, reduces quality fluctuations between batches, and reduces rework / scrap rates.

[0015] 2. This invention uses digital twins as a constrained virtual experimental platform, combined with constrained Bayesian optimization for online multi-objective optimization (balancing quality, energy consumption and safety constraints), and safely transfers the optimization results to new milk types through transfer and incremental learning mechanisms, thereby achieving continuous adaptive updates of the parameter library, thus simultaneously improving quality and energy efficiency while ensuring equipment safety and operating efficiency. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the structure of an intelligent control system for a fully automatic milk foam machine according to the present invention; Figure 2 This is a flowchart illustrating an intelligent control method for a fully automatic milk foam machine according to the present invention. Detailed Implementation

[0017] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0018] Example 1: As Figure 1 As shown, the present invention provides an intelligent control system for a fully automatic milk foam machine, comprising: The data acquisition and processing module is used to acquire conductivity, temperature, stirring power and acoustic timing signals and perform time alignment, temperature correction and normalization to obtain multi-channel timing data, and calculate short-time coupling index based on the multi-channel timing data. Furthermore, the step of obtaining multi-channel time-series data includes: Collect electrical conductivity, temperature, stirring power, and acoustic timing signals; Align the signals of each channel according to the sampling timestamp, use the cross-correlation function to determine the time offset, and resample to the same time axis by interpolation; Apply temperature correction based on the factory calibration curve to the conductivity signal, and then put the correction result alongside the temperature signal; Each channel signal is denoised and baseline corrected, and normalization is performed using the initial steady-state range as a reference to obtain processed and time-synchronized multi-channel time-series data.

[0019] Specifically, the data acquisition and processing module achieves synchronous acquisition of multi-channel signals through a sensor array deployed at key locations in the milk foam machine. The conductivity sensor, using a four-electrode configuration, is installed at the bottom of the stirring container to monitor changes in the ion concentration of the milk; the temperature sensor, a platinum resistance sensor, is placed on the container wall to monitor the milk temperature in real time; the stirring power is obtained through the power feedback interface of the motor driver, reflecting changes in mechanical resistance during the stirring process; and the acoustic sensor, using a wideband microphone, is installed near the stirring head to capture the acoustic signals generated during stirring. All sensor signals are synchronously acquired through a unified data acquisition system.

[0020] Due to differences in response time and signal transmission paths among different sensors, the signals from each channel are not perfectly synchronized in time. The system uses the temperature signal as the reference channel because temperature changes are relatively stable and continuous. A cross-correlation function is used to analyze the similarity between each channel's signal and the reference channel, and the time offset τ is determined by finding the peak position of the cross-correlation function. Subsequently, cubic spline interpolation is used to resample the offset channel, unifying all signals to the same time reference and ensuring accurate correspondence of the data from each channel on the time axis.

[0021] Conductivity measurements are significantly affected by temperature and require temperature compensation. The system uses the factory-calibrated temperature correction curve, and the correction formula is as follows: in The standard conductivity is 25℃. For the measured conductivity, For actual measured temperature, The temperature coefficient is determined based on the sensor model and calibration results.

[0022] The acquired signals were denoised and normalized. A moving average filter was used to remove high-frequency noise, with a window length set to 100ms. Baseline correction used the mean of the initial 30-second steady-state interval as the baseline reference. Normalization mapped each channel signal to the [0,1] interval; the normalization formula is as follows: in and These are the minimum and maximum values ​​of the initial steady-state interval, respectively.

[0023] After time alignment, temperature correction, and normalization, multi-channel time-series data with time synchronization is obtained, including corrected conductivity data, temperature data, stirring power data, and acoustic signal data, providing a standardized data foundation for subsequent feature extraction and analysis.

[0024] Furthermore, the calculation of the short-term coupling index includes: The multi-channel time series data is segmented into short time windows and the window sequence is truncated by sliding with a fixed overlap rate. Within each time window, power-related characteristics, temperature-related characteristics, and acoustic energy-related characteristics are calculated separately. Within each window, calculate the short-time correlation between the power-related feature, the temperature-related feature, and the acoustic-related feature; The short-term correlation quantities are linearly combined according to the weights obtained from offline calibration and then smoothed to obtain short-term coupling indices and generate a short-term coupling index sequence.

[0025] Specifically, the acquired multi-channel time-series data is segmented using a sliding window approach. The window length is determined based on the timescale of the milk foam formation process, requiring a balance between temporal resolution and statistical stability. The overlap ratio is set to ensure the continuity of information between adjacent windows and prevent important changes from being segmented by window boundaries.

[0026] Representative characteristics of each channel were calculated within each time window. Power variation characteristics were obtained by calculating the variance and root mean square value of the first-order difference of the power signal, reflecting the fluctuation intensity and trend of stirring resistance. Temperature gradient was obtained by calculating the first derivative of the temperature signal, characterizing the intensity of the heat transfer process. Acoustic energy characteristics were calculated using the root mean square value of the acoustic signal, reflecting the activity level of acoustic events such as bubble formation and collapse during stirring.

[0027] Calculate the Pearson correlation coefficients between power-related and temperature-related characteristics, and between power-related and acoustic-related characteristics. The power-temperature correlation reflects the coupling strength between the mechanical stirring process and the heat transfer process; when the milk viscosity changes, the stirring power and temperature distribution exhibit synergistic changes. The power-acoustic correlation reflects the synergy between mechanical action and acoustic phenomena; changes in the gas-liquid interface during milk foam formation simultaneously affect stirring resistance and acoustic characteristics.

[0028] A weighted linear combination of power-temperature and power-acoustic correlation quantities is performed. Based on the weighting coefficients determined through statistical analysis of a large number of samples during the offline training phase, in a preferred embodiment, the weighting coefficients are 0.6 and 0.4, respectively, obtained using the least squares method. These weighting coefficients reflect the relative importance of different coupling relationships to the quality of the milk foam, and are obtained through regression analysis with the results of human sensory evaluation. The combined results are smoothed using a moving average to reduce the impact of instantaneous fluctuations, ultimately generating a short-time coupling index sequence. This index can quantitatively describe the dynamic characteristics of multi-physics interactions during the milk foam production process.

[0029] The predictive control module is used to map the short-term coupling index to the milk foam quality index (MQI) through a regression model. It uses a state-space model combined with Kalman filtering to perform rolling predictions on the short-term evolution of the MQI. When the prediction shows that a gelation critical mode will appear within a predetermined look-ahead time, the control parameter sequence is dynamically adjusted in advance. Furthermore, the step of mapping the short-term coupling index to the Milk Foam Quality Index (MQI) through a regression model includes: The short-term coupling index sequence is used as the input vector for the regression model; The input vector is fed into an offline-calibrated regression model, which outputs a normalized MQI value and the corresponding confidence score. The regression model is trained with sensory calibration, and the output MQI corresponds to a predefined range of quality levels.

[0030] Specifically, the regression model employs a support vector regression (SVR) architecture, using a short-time coupled index sequence as the input feature vector. The input vector is constructed as follows: ,in This is the short-term coupling indicator at the current moment. The time window length is specified. The model is built through an offline training phase, with training samples derived from actual data from a large number of milk foam production batches. Each sample includes short-term coupling indicators at the corresponding time point and sensory quality scores assessed by professional tasters.

[0031] During training, sensory quality scores were converted into standardized MQI values. Sensory evaluation covered four dimensions: the smoothness, stability, and texture of the milk foam. A weighted average was used to obtain the overall quality score. The weighting coefficients are determined through expert evaluation. The MQI standardization formula is: ; in and These represent the minimum and maximum values ​​of the quality score, respectively. The regression model learns from the input feature vector. Nonlinear mapping relationship to MQI ,Right now .

[0032] The trained regression model receives real-time short-term coupled index sequences and outputs normalized MQI values ​​and their confidence scores. The confidence scores are obtained through variance estimation of the model's predictions and reflect the reliability of the prediction results. The values ​​are divided into four quality levels: excellent, good, qualified, and unqualified, based on threshold values ​​of 0.8, 0.6, and 0.4.

[0033] Furthermore, the step of dynamically adjusting the control parameter sequence in advance includes: The MQI change rate is calculated based on the MQI value, and the MQI and its change rate are constructed into a state vector. A linear state space model is established, and a Kalman filter is used to make rolling predictions of the MQI evolution trajectory within a predetermined future time window. When the prediction results show that the MQI will drop below the gelation critical threshold within the look-ahead time, the parameter dynamic adjustment mechanism is triggered. Based on the predicted MQI decline rate and current milk source characteristics, the corresponding control parameter sequence is matched from the pre-built control strategy library; The parameter settings in the control parameter sequence are executed sequentially in chronological order, and a first-order filter is used to achieve a smooth transition during the parameter change process. After each step of parameter adjustment is executed, the measured MQI value is compared with the predicted value of that step, and the subsequent unexecuted control parameter sequence is corrected according to the magnitude of the deviation.

[0034] Specifically, the temporal evolution of MQI is modeled as a linear state-space model. The state vector is defined as: ; in The state transition equation is: ; ; ; Among them, the rate of change decay coefficient Based on the rheological relaxation properties of milk, determined by formula Calculate the relaxation time. Milk viscosity and elastic modulus ratio Obtain. Control influence coefficient The value is determined based on the linear response relationship between stirring power and the rate of change in milk structure, with a range of 0.05-0.15. This represents the currently input control parameters and process noise. With observation noise Assuming zero-mean random perturbations, the covariance matrix is... and It can be estimated and recorded as the preferred value by historical residual statistical experiments.

[0035] Based on the state-space model, Kalman filtering is used for rolling prediction and estimation. The prediction steps can be expressed as follows: Covariance prediction The update process uses the standard Kalman gain formula and state correction: ; ; The future is derived based on this recursion. MQI predicted sequence In this embodiment, the preferred step count is... Sampling interval Consistent with short time windows to ensure time scale alignment.

[0036] Predictive triggering uses a threshold determination, with the threshold set proportionally based on the initial quality of the current batch, prioritizing optimal relationships. In this embodiment, the proportionality coefficient is taken as... (The optional range of 0.3–0.5 was obtained through mechanistic and statistical verification). When a minimum number of steps exists... And the standard deviation of the regression residuals If the value falls below the confidence threshold, a critical risk of gelation is identified, triggering dynamic parameter adjustments; if If the threshold is exceeded or a sensor anomaly is detected, aggressive adjustments are suppressed and a conservative template is adopted.

[0037] After triggering, the predicted MQI descent rate will be used. ( The estimated control parameter sequence is matched with the milk source type in the pre-built control strategy library; to avoid actuator abrupt changes, the actual release adopts restricted smooth update, and the magnitude of the target difference is first pruned. Then, the target is approximated asymptotically according to the first-order filtering rule: ; in For the smoothing coefficient, this embodiment preferably takes the value of . Maximum change in a single step Set it to approximately 10% of the corresponding range to balance response and safety. Vector control. By pre-set weight vector Projection is equivalent scalar input Used for the next forecast period.

[0038] Each time parameters are sent and new observations are obtained, the system calculates the prediction bias. For subsequent unexecuted template steps, online corrections are performed according to the exponential decay rule: ; in Channel gain coefficient, attenuation coefficient A value of 0.5 is preferred. To ensure closed-loop stability, the channel gain coefficient... The allowable range is between 0.1 and 0.4, preferably 0.2 to 0.3, and adjustments are made through small-scale online verification before actual deployment.

[0039] Regarding parameter calibration The method recommended for obtaining the results is a small-amplitude step experiment: under controlled conditions, a known step change is applied to a certain control channel, the corresponding short-term MQI response is recorded, and linear regression is used for fitting. The proportionality coefficient with respect to input change is used to estimate .matrix Noise covariance confidence threshold The preferred values ​​are obtained through autoregressive analysis and residual statistics of historical batch data, and typical values ​​and calibration steps are listed in the preferred embodiments of the specification for reproducibility.

[0040] To ensure robustness, when the standard deviation of the residuals output by the regression model is... Exceeding the confidence threshold, or observation noise If the sensor indicates an anomaly, or if the deviation still exceeds the limit after several consecutive corrections, the system automatically switches to a conservative parameter template and records the abnormal samples for offline simulation and model retraining. In the event of continuous anomalies, the system triggers the simulation optimization of the digital twin to regenerate more suitable initial parameter values ​​or prompts for manual intervention.

[0041] The simulation optimization module is used to search for the optimal control parameter sequence in the digital twin simulation environment using Bayesian optimization, and to use the optimal control parameter sequence obtained by the search as the initial value for on-site parameter tuning. Furthermore, the simulation optimization module performs the following steps: An approximate model of the current operating condition is established in the digital twin, and the control parameters currently in operation on site are used as the initial settings of the simulation model; The initial parameter sample set was generated using the Latin hypercube sampling method and simulated one by one in a twin environment. The MQI estimate and energy consumption index of the simulation output were recorded. The surrogate model is updated based on the simulation results and the next set of parameters to be measured is selected according to the acquisition criteria. The process is iterated until the predetermined termination condition is reached. Select the optimal parameter set from the iteration results and write the parameter set into the field parameter tuning initial value library.

[0042] Specifically, firstly, an approximate model for the current working conditions is established in the digital twin (denoted as...). The model uses a control parameter vector. As input, the output is the predicted quality index of the milk foam and the estimated energy consumption, i.e. A digital twin consists of two parts: a physical approximation sub-model based on rheology and heat conduction, and a data-driven residual correction term. Together, they improve the simulation's accuracy in fitting the real system. Initial parameters of the twin. The calibration was performed by minimizing the squared error of several recent batches of field data (control trajectory, measured MQI, and energy consumption). The calibration was carried out in the following manner: ; Where the regularization term coefficient To avoid overfitting, the calibration process uses least squares to solve the problem.

[0043] After establishing the twin and initializing with field parameters, the optimization objective function is determined. To ensure quality while considering energy consumption, a single scalar objective function is adopted: ; in This is the energy consumption weight, used to balance mass and energy consumption (the selection range and example values ​​are given in the preferred embodiments of the specification). Simulation optimization is performed to maximize... The objective is to be controlled by the parameter range constraints. and safety / equipment limit constraints.

[0044] To efficiently search for optimal control parameters, a Bayesian optimization process based on a surrogate model is adopted. First, an initial sample set is generated in the parameter space using Latin hypercube sampling. (Preferred) (Depending on the parameter dimensions), in digital twins Simulate each of these candidates one by one and record the corresponding simulation output. And calculate the target value The obtained sample Fitting a Gaussian process surrogate model, GP in any Give the predicted mean with standard deviation .

[0045] Based on the GP agent, an improved acquisition function is introduced as the next candidate selection rule, and its mathematical form is: ; in This represents the maximum value of the target in the currently evaluated sample. To explore incentive items (preferably 0.01, which can be adjusted within the range of 0.005-0.02), and These represent the standard normal cumulative distribution and the density function, respectively. The next parameter to be evaluated is obtained by numerically optimizing the solution to maximize the EI. and in digital twins Perform simulations to obtain new Then Add samples to the sample set and update the GP agent. The above process of "fitting GP → selecting the maximum EI point → evaluating on a twin → updating GP" is iterated until a termination condition is met (e.g., reaching the maximum number of simulation evaluations, the agent gain falling below a threshold, or the improvement from the most recent optimal values ​​being less than the shutdown criterion). During the iteration process, parameter constraints and safety boundary checks are maintained on the sampling points to avoid generating candidates that are unexecutable or exceed device limits.

[0046] When the iteration ends, the objective function value is selected from all simulation samples. Largest parameter set As a candidate optimal parameter, this parameter is written into the field parameter tuning initial value library as the initial value for the next control step. Before field application, a limited number of field validation tests should be conducted (e.g., 1–3 batch validations). For each field test, the measured MQI and energy consumption should be recorded and compared with the twin predictions. If there is a significant deviation between the measured values ​​and the twin predictions, the new field data should be incorporated into the twin's calibration set. Recalibrate and repeat the Bayesian optimization loop as needed to narrow the simulation-reality gap (i.e., form a closed loop of "simulation-validation-calibration-reoptimization"). When validation passes, The initial values ​​are formally written into the field parameter adjustment library for use by the predictive control module.

[0047] The adaptive learning module is used to fine-tune the regression model with a few samples using transfer learning for new milk sources, and to update the regression model and judgment threshold after each batch of processing through incremental learning, thereby constructing an adaptive optimization control strategy library.

[0048] Furthermore, the transfer learning step includes: Several batches of short-term coupling indices and corresponding measured MQI values ​​from new milk sources were collected as fine-tuning samples. Using the existing regression model as the initial model, perform a small number of iterations of parameter adjustment based on fine-tuned samples and verify the predictive performance of the fine-tuned model; When the fine-tuned model meets the pre-defined performance criteria, the current regression model is replaced and the decision threshold set is updated simultaneously.

[0049] Specifically, when the system detects that a new milk source has entered the usage process, it first collects short-term coupling index sequences from several (preferably 3 to 5 batches) trial production batches. Corresponding measured MQI tags As fine-tuning samples, the existing offline-trained regression model is used. The initial model is loaded into the fine-tuning environment. Fine-tuning employs a weighted retraining strategy: new milk source samples are assigned higher weights to increase their influence, while historical samples are retained to maintain model generalization. The number of iterations is limited and regularization is used to prevent overfitting. After each iteration, the prediction performance metrics (e.g., mean squared error (RMSE) or Pearson correlation coefficient r on the validation set) are evaluated through cross-validation. Fine-tuning is considered successful when the model meets preset performance criteria on the validation set (e.g., RMSE improvement exceeds a threshold or the correlation coefficient meets calibration requirements). The fine-tuned model is then... Replace with an online running model, and simultaneously update the corresponding set of judgment thresholds (such as those calculated based on the MQI statistics of the initial batch of new milk source). If the fine-tuning fails to meet the requirements, new samples will be collected again, or the process will revert to conservative parameters and manual calibration to avoid direct replacement that could lead to erroneous actions on-site. Both fine-tuning and replacement steps will record the model version, training sample index, and validation results, and these records will be written to the control strategy library for auditing and backtracking purposes.

[0050] Furthermore, the incremental learning step includes: After each batch is completed, the short-term coupling index sequence, measured MQI value and control trajectory of that batch are added to the online sample set; The regression model is updated incrementally according to a predetermined update rule, with the update magnitude limited to maintain model stability. A sliding update strategy is used for the judgment threshold to gradually absorb statistical changes between batches; When the prediction error exceeds the tolerance and persists, offline retraining is triggered and the retraining results are written to the model library.

[0051] Specifically, during each batch of regular production, the adaptive learning module maintains an online sample set and performs incremental updates. After each batch is completed, the short-term coupled index sequence of that batch is... Actual measurement of MQI tags With the current batch control trajectory The model is added to the online sample set. To avoid frequent full retraining of the master regression model, an incremental learning architecture of "master model + linear correction layer" is adopted: the master model remains the SVR base obtained offline or through fine-tuning. A lightweight linear correction function is maintained on its output. ,in The low-dimensional feature vectors extracted from the main regression model. This is a parameter vector that can be updated online. Each batch uses a small-step gradient update rule. Implement recursive correction: ; in The learning rate is the batch average of the prediction error. Take a smaller positive value to ensure stability (preferably in the range of 0.001–0.01), and for A mild decaying regularization is added to prevent drift. This linear correction layer can absorb small system drifts and slight differences between milk sources with very low computational cost, thereby reducing the number of times offline retraining is triggered. If necessary, recursive least squares can be used instead of the small step gradient method to achieve faster convergence.

[0052] The online update of the threshold set employs an exponentially weighted moving average strategy to gradually absorb inter-batch statistical changes while avoiding oversensitivity to short-term anomalies. (The threshold is used as the basis for this.) For example, using the EMA update rule: ; in The update rate is the reference value calculated based on the initial MQI or calibration index for this batch. Use small values ​​(preferably 0.01–0.1) to ensure smoothness. All threshold updates are historically recorded in the policy library and associated with the corresponding batch data for offline backtracking and evaluation.

[0053] To prevent performance degradation due to accumulated errors and model drift, several triggering conditions are set: when the online prediction error exceeds the tolerance limit for several consecutive batches (e.g., 3 batches) or the parameter changes of the linear correction layer exceed the preset range, the offline retraining process is automatically triggered. Offline retraining takes a new online sample set as input, retrains the main regression model, and recalibrates the template parameters in the policy library; after retraining, cross-validation and a limited number of field validations are performed. If the validation passes, the online main model is updated and the old model version is archived.

[0054] This implementation method ensures rapid adaptation to "small sample fine-tuning" for new milk sources, maintains online adaptability through incremental correction layers and threshold sliding updates, and sets up retraining trigger and version rollback mechanisms to ensure safety and robustness.

[0055] Example 2, This embodiment uses the production of full-fat milk foam in a tea beverage chain store as a specific application scenario to demonstrate the practical application effect of the intelligent control system of the present invention for a fully automatic milk foam machine.

[0056] The tea shop routinely uses Brand A whole milk to make its milk foam, and the system runs stably. One day, due to a temporary shortage by the supplier, they switched to Brand B whole milk, requiring rapid adaptation to the new milk source and optimization of control parameters. For example... Figure 2 As shown, the process of applying the intelligent control system of the present invention for a fully automatic milk foam machine is as follows: During the data acquisition phase, the system automatically detected changes in the signal characteristics of the new milk source. The baseline conductivity of Brand B milk was 4.2 mS / cm, with a protein content of 3.6% and a fat content of 3.8%. In the first three batches of trial production, the system continuously collected data from four sensors. Specifically, the conductivity sensor had a measurement accuracy of ±0.1 mS / cm and a response time of less than 5 seconds; the temperature sensor had an accuracy of ±0.1℃ and a response time of less than 3 seconds; the stirring power measurement accuracy was ±1W; and the acoustic sensor had a frequency response range of 20Hz-20kHz. The sampling frequency was uniformly set to 100Hz, and the time synchronization accuracy of the data acquisition system was ±1ms.

[0057] In the time alignment process, the calculation window length of the cross-correlation function is set to 20% of the signal length, and the search range is ±50 sampling points. Natural boundary conditions are used for the cubic spline interpolation. The moving average filter has an order of 3 and a cutoff frequency of 10Hz.

[0058] In the calculation of short-time coupling indices, the window length was set to 5 seconds and the overlap rate to 50%. Power variation characteristics were calculated using the power signal variance and the root mean square value of the first-order difference, while the temperature gradient was calculated using the five-point difference method. The short-time coupling index sequence shows that the average power-temperature correlation coefficient for brand B milk source is 0.73, and the average power-acoustic correlation coefficient is 0.58, which are significantly different from the values ​​of 0.65 and 0.62 for brand A milk source.

[0059] In the intelligent prediction and control phase, the first batch was started using the original parameters (stirring speed 1200 rpm, heating power 2.8 kW, stirring time 180 s). The regression model adopted the SVR with the RBF kernel function, kernel parameter γ=0.1, penalty parameter C=100, and fault tolerance parameter ε=0.01. The input vector length L was set to 10, including short-term coupling indices for the current and previous 9 time steps. The predicted MQI value began to decrease rapidly at 120 seconds. The Kalman filter parameters were set as follows: attenuation coefficient β=0.85 in the state transition matrix, control influence coefficient b=0.08, look-ahead step N=3, and sampling interval Δt=1 second. The prediction showed that the MQI would drop below the gelation critical threshold of 0.4 at 165 seconds. The system triggered dynamic parameter adjustments at 125 seconds, reducing the stirring speed to 1100 rpm and the heating power to 2.4 kW. The parameter adjustments used a first-order filter for smooth transition, with a smoothing coefficient α = 0.3. The maximum single-step change was 100 rpm for stirring speed and 0.3 kW for heating power. Ultimately, the MQI for this batch stabilized at 0.52, preventing complete failure.

[0060] The simulation optimization module then sought optimal parameters based on the characteristics of Brand B's milk source. The initial sample set generated 12 parameter combinations using Latin hypercube sampling, with the following ranges: stirring speed 800-1500 rpm, heating power 1.0-3.5 kW, and stirring time 120-300 s. After 15 iterations, the optimal parameters were found: stirring speed 1050 rpm, heating power 2.2 kW, and stirring time 210 s. Digital twin prediction showed that this parameter combination had an MQI of 0.81 and energy consumption of 2.1 kWh, representing a 16.7% reduction in energy consumption compared to the original parameter combination.

[0061] The adaptive learning module performs adaptation to the new milk source. Using parameters optimized through simulation, a second batch was produced, achieving an actual MQI of 0.78, with an error of only 3.7% compared to the digital twin prediction. During transfer learning, the weights of the new milk source samples were set to three times that of the historical samples, the number of fine-tuning iterations was limited to 20, and the initial learning rate was set to 0.001. After fine-tuning, the model's RMSE on the validation set decreased from 0.089 to 0.043, and the Pearson correlation coefficient increased from 0.82 to 0.91, meeting the performance criteria. In incremental learning, the feature vector dimension of the linear correction layer was 3, and the learning rate was dynamically adjusted based on the batch error: 0.001 when the error was less than 0.02, 0.005 when the error was between 0.02 and 0.05, and 0.01 when the error was greater than 0.05.

[0062] The third batch, using the finely tuned model and updated decision thresholds, achieved an MQI of 0.83, fully meeting the quality requirements of tea shops. The system successfully completed rapid adaptation to the new milk source within three batches, significantly improving efficiency and reducing raw material waste compared to the traditional manual parameter tuning method which requires 8-10 batches of trial production. The entire adaptation process demonstrated the predictive advantages of the multi-channel coupled indicators, the early intervention capability of state-space predictive control, the efficient parameter search of digital twin simulation optimization, and the rapid adaptation capability of the adaptive learning mechanism to the new milk source.

[0063] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent control system for a fully automatic milk cap machine, characterized in that, Comprise: a data acquisition and processing module for acquiring conductivity, temperature, stirring power and acoustic time sequence signals and performing time alignment, temperature correction and normalization processing to obtain multi-channel time sequence data, and calculating a short-time coupling index based on the multi-channel time sequence data; a predictive control module for mapping the short-time coupling index to a milk cover quality index MQI through a regression model, using a state space model combined with Kalman filtering to perform rolling prediction of the short-term evolution of the MQI, and dynamically adjusting the control parameter sequence in advance when the prediction shows that a gelation critical mode will occur within a predetermined lookahead time; a simulation optimization module for searching for an optimal control parameter sequence in a digital twin simulation environment using Bayesian optimization, and using the optimal control parameter sequence obtained by the search as the initial value for on-site parameter tuning; an adaptive learning module for using transfer learning to fine-tune the regression model for new milk sources, and updating the regression model and decision threshold after each batch processing through incremental learning to build an adaptive optimized control strategy library.

2. The intelligent control system for the fully automatic milk cover machine according to claim 1, wherein, The step of obtaining multi-channel time sequence data comprises: acquiring conductivity, temperature, stirring power and acoustic time sequence signals; aligning each channel signal according to the sampling time stamp, determining the time offset using the cross-correlation function, and resampling to the same time axis by interpolation; applying temperature correction to the conductivity signal based on the factory calibration curve, and arranging the correction results and temperature signals side by side; performing denoising and baseline correction on each channel signal, and completing normalization processing using the initial steady-state interval as a reference to obtain processed and time-synchronized multi-channel time sequence data.

3. The intelligent control system for the fully automatic milk cap machine as claimed in claim 1 wherein, The calculation of the short-time coupling index comprises: segmenting the multi-channel time sequence data according to short-time windows and sliding the window sequence with a fixed overlap rate; calculating power-related features, temperature-related features and acoustic energy-related features in each time window; calculating the short-time correlation of the power-related features and the temperature-related features and the acoustic-related features in each window; linearly combining the short-time correlations according to the weights obtained offline, and performing smoothing processing to obtain the short-time coupling index, and generating a short-time coupling index sequence.

4. The intelligent control system for the fully automatic milk cover machine according to claim 1 and 3, characterized in that, The step of mapping the short-time coupling index to a milk cover quality index MQI through a regression model comprises: using the short-time coupling index sequence as the input vector of the regression model; placing the input vector into the regression model calibrated offline, and the model output is the normalized MQI value and the corresponding confidence; The regression model is trained by sensory calibration, and the output MQI corresponds to the pre-defined quality level range.

5. The intelligent control system for the fully automatic milk cover machine according to claim 1 and 4, wherein, The step of dynamically adjusting the control parameter sequence in advance comprises: calculating the MQI change rate based on the MQI value, constructing the MQI and its change rate into a state vector, establishing a linear state space model, and using a Kalman filter to perform rolling prediction of the MQI evolution trajectory within a future predetermined time window; when the prediction result shows that the MQI will drop below the gelation critical threshold within the lookahead time, triggering the parameter dynamic adjustment mechanism; according to the predicted MQI drop rate and the current milk source characteristics, matching the corresponding control parameter sequence from the pre-built control strategy library; The parameter setting values in the control parameter sequence are executed in time sequence, and the parameter changing process is realized by first-order filtering for smooth transition. After the execution of each step parameter adjustment, the measured MQI value is compared with the predicted value of this step, and the subsequent unexecuted control parameter sequence is modified according to the deviation.

6. The intelligent control system for the fully automatic milk cap machine as claimed in claim 1 wherein, The simulation optimization module executes the following steps: An approximate model of the current working condition is established in digital twinning, and the control parameters of the current running in the field are taken as the initial setting values of the simulation model; The Latin hypercube sampling method is used to generate an initial parameter sample set, which is simulated one by one in the twinning environment, and the simulation output MQI estimate and energy consumption index are recorded; Based on the simulation results, the proxy model is updated, and the next set of parameters to be tested is selected according to the collection criteria, and the iteration is repeated until the predetermined termination condition is reached; The optimal parameter set is selected from the iteration results, and the parameter set is written into the field parameter adjustment initial value library.

7. The intelligent control system for the fully automatic milk cap machine as claimed in claim 1, wherein, The transfer learning step includes: Collecting several batches of short-term coupling indicators and corresponding measured MQI of the new milk source as fine-tuning samples; Using the existing regression model as the initial model, performing a small number of iterations of parameter adjustment based on the fine-tuning samples, and verifying the prediction performance of the fine-tuned model; When the fine-tuned model meets the pre-set performance criteria, replace the existing regression model and update the decision threshold set at the same time.

8. The intelligent control system for a fully automatic milk cap machine as claimed in claim 1, wherein, The incremental learning step includes: After each batch is completed, the short-term coupling indicator sequence, the measured MQI value and the control trajectory of the batch are appended to the online sample set; According to the predetermined updating rule, the regression model is updated incrementally, and the update amplitude is limited to maintain the stability of the model; The sliding update strategy is used for the decision threshold to gradually absorb the statistical changes between batches; When the prediction error exceeds the tolerance and continues to appear, trigger offline retraining and write the retraining results to the model library.

9. A method for intelligent control of a fully automatic milk cover machine, characterized in that, It includes: Collecting conductivity, temperature, stirring power and acoustic time series signals and performing time alignment, temperature correction and normalization processing to obtain multi-channel time series data, and calculating the short-term coupling indicator based on the multi-channel time series data; The short-term coupling indicator is mapped to the milk cover quality index MQI through the regression model, and the state space model combined with Kalman filtering is used to rollingly predict the short-term evolution of the MQI. When the prediction shows that the gelation critical mode will appear within the predetermined lookahead time, the control parameter sequence is dynamically adjusted in advance; In the digital twinning simulation environment, the Bayesian optimization is used to search for the optimal control parameter sequence, and the optimal control parameter sequence obtained by searching is taken as the initial value of the field parameter adjustment; For the new milk source, the regression model is fine-tuned by transfer learning, and the regression model and the decision threshold are updated after each batch is processed through incremental learning, and an adaptive optimization control strategy library is constructed.

Images