Quality index self-stabilization control method and system for chili sauce pulping

Abstract

The application discloses a quality index self-stabilization control method and system for chili sauce pulping, and belongs to the field of quality control. The method comprises the following steps: acquiring an execution component set of a pulping device, and collecting a real-time working condition data set; performing control stability analysis to acquire a control stability feature set; performing clustering division on the execution component set to acquire multiple groups of execution components, wherein the multiple groups of execution components correspond to multiple stable control units, and further comprise corresponding parameter control constraints; capturing chili material spectrum features, inputting the chili material spectrum features into a quality index self-adaptive model for analysis, and determining multiple self-adaptive execution components; according to the clustering labels, multiple parameter control constraints are retrieved and called, parameter solving is performed on the multiple self-adaptive execution components, and multiple groups of stable control parameters are output. The application solves the technical problems that the existing pulping device lacks stable difference dynamic control, and the pulping quality consistency is poor and the equipment operation reliability is low.

Description

Technical Field

[0001] This invention relates to the field of quality control, and more specifically to a method and system for self-stabilizing quality indicators of chili sauce grinding. Background Technology

[0002] In the industrial-scale grinding and production of chili sauce, chili raw materials from different origins, batches, and maturity levels exhibit significant fluctuations in physicochemical properties such as moisture content, crude fiber ratio, and pre-crushing degree. Existing grinding equipment quality control systems often employ fixed preset parameter adjustment modes. These systems fail to dynamically match optimal control parameters based on real-time material characteristics and do not consider the differences in control stability caused by mechanical wear and load variations in core actuators such as the stator-rotor gap adjustment servo mechanism, the main shaft drive variable frequency motor, and the feeding screw conveyor. Instead, they apply broad adjustment constraints based on hardware ratings to all actuators. This results in the inability to fully unleash the rapid adjustment potential of highly stable actuators, while low-stability actuators are prone to malfunctions such as control trajectory divergence, parameter overshoot, mechanical oscillation, and even stator-rotor collisions due to large parameter adjustments. Furthermore, the lack of quantitative analysis methods for the nonlinear dynamic characteristics of actuators ultimately leads to unstable product uniformity and pulp yield in chili sauce grinding, resulting in a high rate of unplanned equipment downtime and making it difficult to meet the quality control requirements of large-scale standardized production. Summary of the Invention

[0003] This application provides a self-stabilizing control method and system for quality indicators of chili sauce grinding, aiming to solve the technical problem that existing grinding equipment lacks dynamic control of stability differences, resulting in poor consistency of grinding quality and low equipment reliability, thereby improving the quality stability and equipment reliability of chili sauce grinding production.

[0004] In view of the above problems, this application provides a method and system for self-stabilizing control of quality indicators for chili sauce grinding.

[0005] The first aspect disclosed in this application provides a method for self-stabilizing quality indicators of chili sauce grinding, the method comprising: A set of actuators for a grinding equipment is obtained, and a real-time operating condition dataset of the actuator set is collected. Control stability analysis is performed on the real-time operating condition dataset to obtain a set of control stability features. The actuator set is clustered according to the control stability feature set to obtain multiple groups of actuators, wherein each group of actuators corresponds to multiple stable control units, and each stable control unit includes corresponding parameter control constraints. The chili material spectrum characteristics of the grinding equipment are captured, and the chili material spectrum characteristics are input into a quality index adaptive model for analysis to determine multiple adaptive actuators. Multiple parameter control constraints are retrieved and invoked according to the cluster labels to which the multiple adaptive actuators belong, and the parameters of the multiple adaptive actuators are solved under the multiple parameter control constraints to output multiple sets of stable control parameters.

[0006] Another aspect of this application discloses a self-stabilizing control system for quality indicators of chili sauce grinding, the system comprising: The system comprises the following modules: a data acquisition module for acquiring a set of actuators of the grinding equipment and collecting a real-time operating condition dataset of the actuator set; an analysis module for performing control stability analysis on the real-time operating condition dataset and obtaining a set of control stability features; a clustering module for clustering the actuator set according to the set of control stability features to obtain multiple groups of actuators, wherein the multiple groups of actuators correspond to multiple stable control units, and each stable control unit includes corresponding parameter control constraints; a capture module for capturing the chili material spectrum features of the grinding equipment, inputting the chili material spectrum features into a quality index adaptive model for analysis, and determining multiple adaptive actuators; and a control module for retrieving and calling multiple parameter control constraints according to the cluster labels to which the multiple adaptive actuators belong, solving the parameters of the multiple adaptive actuators under the multiple parameter control constraints, and outputting multiple sets of stable control parameters.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: By collecting real-time datasets of all operating conditions of the execution components, quantitative analysis of control stability was completed based on phase space reconstruction and singular value control spectrum entropy. The execution components were clustered into three types of stable control units: high, medium, and low, and matched with differentiated parameter control constraints. At the same time, near-infrared spectroscopy was used to capture the spectral characteristics of chili pepper materials in real time. By integrating gradient boosting tree and long short-term memory network into an adaptive quality index model, adaptive execution components with both high adjustment sensitivity and high operational stability were selected. Under the corresponding parameter constraints, the optimal stable control parameters were solved with the goal of minimizing the difference in predicted grinding quality index, which ultimately improved the quality stability and equipment reliability of chili sauce grinding production.

[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0009] Figure 1 A flowchart illustrating a method for self-stabilizing quality indicators in chili sauce grinding is provided for embodiments of this application. Figure 2 A schematic diagram of the structure of a self-stabilizing control system for quality indicators of chili sauce grinding is provided for embodiments of this application; Figure 3 A comparison chart of the effects of a self-stabilizing control method for quality indicators of chili sauce grinding is provided for the embodiments of this application.

[0010] Figure labeling: Acquisition module 11, Analysis module 12, Clustering module 13, Capture module 14, Control module 15. Detailed Implementation

[0011] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0012] The overall concept of the technical solution provided in this application is as follows: This application provides a self-stabilizing control method and system for the quality index of chili sauce grinding. By collecting real-time operating condition datasets of the execution components and performing control stability analysis based on phase space reconstruction and singular value control spectrum entropy, the execution components are clustered into three types of stable control units: high, medium, and low, and matched with differentiated parameter control constraints. At the same time, near-infrared spectroscopy is used to capture the spectral characteristics of chili materials in real time. By fusing gradient boosting trees and long short-term memory networks into an adaptive quality index model, adaptive execution components with both high adjustment sensitivity and high operational stability are selected. Under the corresponding parameter constraints, the optimal stable control parameters are solved with the objective of minimizing the difference in predicted quality index, thereby achieving self-stabilizing control of the grinding quality index and improving the quality stability and equipment operational reliability of chili sauce grinding production.

[0013] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0014] Example 1, as Figure 1 As shown in the embodiment of this application, a method for self-stabilizing control of quality indicators for chili sauce grinding is provided. The method includes: S100: Obtain the set of actuators of the grinding equipment and collect the real-time operating condition dataset of the set of actuators.

[0015] Specifically, the hardware configuration file of the grinding equipment is first read by the programmable logic controller (PLC). The hardware configuration file is a pre-configured configuration document in the PLC system that stores information about all devices connected to the industrial fieldbus. It contains basic information such as the unique physical address, communication protocol type, hardware model, rated operating parameters, and function definitions of each device. All addressable units connected to the industrial fieldbus are then parsed out. The industrial fieldbus is a dedicated communication network in industrial automation systems that enables high-speed and reliable data transmission between the controller and field devices. It can ensure the real-time issuance of control commands and the synchronous feedback of operating data, and its transmission delay meets the real-time requirements of industrial control.

[0016] Subsequently, non-execution components used only for status indication and audible / visual alarms were eliminated. These components could only output status feedback signals and could not receive control commands to participate in the dynamic adjustment of the grinding process. Then, based on the control chain of the chili sauce grinding process—from the feeding and conveying of chili raw materials, dynamic adjustment of the grinding gap, spindle speed control to the temperature regulation of the cooling system—the links that directly determine key quality indicators such as product particle size, uniformity, pulp yield, and flavor retention were selected. All execution units directly involved in process control were then identified, ultimately forming a structured set of execution components. Here, "structured" means that each execution component is accompanied by an indexable mapping table containing metadata such as a unique device identifier, hardware rated parameters, control interface protocol, and the process link it belongs to. This metadata can then be used to guide the parameters of each subsequent component. Control constraints provide native hardware boundary criteria. Based on this structured set of actuators, a low-latency real-time data communication link is established with the PLC through an edge computing gateway deployed at the equipment site. The edge computing gateway is a computing device deployed close to the equipment in the industrial field, which can complete data preprocessing, filtering, and transmission locally, avoiding network latency and bandwidth consumption caused by cloud transmission, and meeting the millisecond-level real-time requirements of industrial control. A synchronous sampling mechanism is adopted to collect multi-dimensional operating parameters of each actuator at an industrial-grade sampling frequency during the continuous grinding cycle. The synchronous sampling mechanism refers to aligning the sampling clock of all actuators with the global clock of the gateway through the hardware clock synchronization module of the gateway, ensuring that the operating data of all components are collected under the same absolute timestamp. Finally, the data collection process fully covers all operating stages, including equipment startup and preheating, steady-state grinding, dynamic parameter adjustment, switching between different batches of chili materials, and shutdown transition. This results in a real-time operating condition dataset in the form of a multivariate time series. Each data point contains three core fields: timestamp, unique identifier of the executing component, and corresponding operating condition parameter value. The specific parameters collected include the real-time drive current of the stator-rotor gap adjustment servo mechanism, rotor axial displacement, servo motor output torque and encoder feedback position, as well as the output frequency of the main shaft drive variable frequency motor, stator three-phase current, motor winding temperature and main shaft speed, the conveying speed of the feed screw conveyor, feed inlet material pressure and drive motor power, and the flow rate, inlet-outlet pressure difference and pump body temperature of the cooling circulation pump.

[0017] S200: Perform control stability analysis on the real-time operating condition dataset to obtain a set of control stability features.

[0018] Specifically, the real-time operating condition dataset in the form of multivariate time series is first split according to the unique identifier of the actuator to extract the single real-time operating condition sequence corresponding to each actuator. That is, the time-series change sequence of a certain operating condition parameter of a single actuator within a continuous sampling period. The real-time drive current sequence of the stator-rotor gap adjustment servo mechanism is extracted first as the analysis object because this parameter directly maps the dynamic adjustment process of the grinding gap, and its small fluctuations will be directly transmitted to the particle size uniformity of the chili sauce. At the same time, the sequences of rotor axial displacement and servo motor output torque are collected as auxiliary verification.

[0019] Subsequently, autocorrelation analysis and nearest neighbor continuity analysis were used to determine the two phase space reconstruction parameters: time delay parameter and stable feature embedding dimension. The time delay parameter was obtained by calculating the autocorrelation function of a single real-time operating condition sequence. The delay time corresponding to the autocorrelation coefficient decreasing to a preset correlation threshold is the optimal value, ensuring that the reconstructed state vector contains sufficient independent information. The stable feature embedding dimension was determined by nearest neighbor continuity analysis, that is, by gradually increasing the embedding dimension and calculating the distance change rate between adjacent state points. The dimension corresponding to the distance change rate decreasing to a preset ratio threshold is the optimal value, ensuring that all dynamic information in the one-dimensional time series data can be completely mapped to the high-dimensional space. Subsequently, based on the two determined parameters, the single real-time operating condition sequence is sliced ​​using the time delay parameter as the state interval to generate multiple equal-length delayed state vectors. Each vector contains operating condition data from the current time and multiple previous delayed times. These delayed state vectors are then reconstructed into a state evolution matrix by arranging and combining them row by row according to the stability feature embedding dimension. Mapping this matrix to a high-dimensional topological space yields a high-dimensional phase space trajectory matrix. This matrix can visualize the nonlinear dynamic behavior hidden in one-dimensional time series data and intuitively characterize the state evolution law of the actuator in the control process, including trajectory convergence, stability characteristics, periodic fluctuations, periodic characteristics, response speed and recovery capability to external disturbances, and disturbance response characteristics.

[0020] Next, singular value decomposition is performed on the trajectory matrix in the high-dimensional phase space. This involves decomposing the complex trajectory matrix into the product of an orthogonal matrix, a diagonal singular value matrix, and a transposed orthogonal matrix. The non-zero elements in the diagonal matrix are the singular values. Arranging all the singular values ​​in descending order yields the singular value spectrum. The magnitude of the singular values ​​corresponds to the energy proportion in different directions in the phase space. The more concentrated the singular values, the more stable the system state. Conversely, the less concentrated the singular values, the more unstable the system has multiple motion modes.

[0021] Based on this, the probability density distribution of the singular value spectrum is calculated, that is, the frequency of singular values ​​occurring in each singular value interval is statistically analyzed, and then the singular value control spectrum entropy is calculated using the information entropy formula. Spectral entropy is a quantitative indicator that measures the uniformity of the probability density distribution. The lower the spectral entropy value, the more concentrated the singular values ​​are in a few main directions, corresponding to the convergence of the control trajectory of the execution unit and the stable dynamic response, which belongs to a high-stability unit. The higher the spectral entropy value, the more dispersed the singular value distribution, corresponding to the divergence of the control trajectory of the execution unit and its susceptibility to disturbances, which belongs to a low-stability unit.

[0022] Finally, the singular value spectrum, singular value control spectrum entropy, phase space trajectory convergence coefficient, disturbance response recovery time, and other quantitative indicators corresponding to each execution component are integrated to form a set of control stability features. Each feature corresponds to a certain aspect of the stability performance of the execution component.

[0023] S300: Cluster the set of execution components according to the control stability feature set to obtain multiple groups of execution components, wherein the multiple groups of execution components correspond to multiple stable control units, and each stable control unit includes corresponding parameter control constraints.

[0024] Specifically, firstly, a weighted clustering algorithm is used to perform clustering of the execution component set based on the control stability feature set. Clustering refers to an unsupervised learning method that groups samples with similar features into one class based on the distance or similarity of sample points in a multidimensional feature space. Firstly, three cluster centers are initialized according to the three stability levels of high, medium and low preset in this scheme. Each cluster center is a baseline feature vector consistent with the dimension of the control stability features. Its value is obtained by averaging the features of the execution components with labeled stability levels in historical batches. For example, a high stability cluster center corresponds to a feature combination of singular value control spectrum entropy ≤ 0.3, trajectory convergence coefficient ≥ 0.9, and disturbance response recovery time ≤ 100ms.

[0025] Next, the weighted Euclidean distance between the feature vector of each execution component and the three cluster centers is calculated as the intra-class similarity index. The weighted Euclidean distance is the distance calculated after assigning different weights to features of different dimensions. The weight of singular value control spectral entropy is set to 0.5, the weight of trajectory convergence coefficient is set to 0.3, and the weight of disturbance response recovery time is set to 0.2. This is because spectral entropy is the core index that best reflects the overall nonlinear dynamic stability of the system. The smaller the distance, the closer the stability characteristics of the execution component and the cluster center are, and the higher the intra-class similarity. Each execution component is assigned to the cluster with the highest similarity according to the intra-class similarity index. Then, the mean of the feature vectors of all execution components in each cluster is recalculated as the new cluster center. The above assignment and update process is repeated until the change in cluster centers in two consecutive iterations is less than the preset convergence threshold, or the maximum number of iterations is reached. Finally, multiple groups of execution components are output, and the intra-class similarity index of all execution components in the same group is greater than the preset similarity threshold, ensuring that the stability differences of execution components in the group are within an acceptable control range. Finally, each set of actuators is mapped to an independent stable control unit. The stable control unit is a logic control module built to achieve differentiated stable control. Actuators within the same unit share a set of parameter control constraints. Parameter control constraints refer to the dynamic safety boundaries set to ensure that the actuators are always in a stable operating state during dynamic adjustment. It is a further tightening of the hardware boundaries based on the actual operating stability level of the actuators. Specifically, it includes the upper limit of adjustment step size, the upper limit of rated drive current, the upper limit of transient maximum temperature rise rate, and the upper limit of mechanical axial displacement. The upper limit of adjustment step size for high-stability control units can be set to 80%~100% of the rated hardware step size, allowing for larger-scale rapid adjustments to quickly respond to changes in the characteristics of different batches of chili materials. The upper limit of adjustment step size for low-stability control units is only set to 20%~40% of the rated hardware step size, avoiding control trajectory divergence and parameter overshoot through gradual adjustment with small steps. The upper limit of rated drive current is determined by multiplying the historical highest operating current of the actuator by the stability coefficient. The stability coefficient is positively correlated with the stability level of the unit.

[0026] S400: Capture the chili material spectrum characteristics of the grinding equipment, input the chili material spectrum characteristics into the quality index adaptive model for analysis, and determine multiple adaptive execution components.

[0027] Specifically, a near-infrared diffuse reflectance spectroscopy sensor is first installed in the free-fall section of the material at the feed inlet of the grinding equipment. This sensor continuously collects the near-infrared diffuse reflectance spectrum of the chili pepper material at a sampling frequency matched to the rotational speed of the feed screw conveyor. Near-infrared diffuse reflectance refers to the spectrum reflected back when near-infrared light (wavelength 900-1700nm) irradiates the surface of the chili pepper material; some of the light is selectively absorbed by chemical components such as moisture, coarse fiber, capsaicin, and fat within the material. Different components have unique characteristic absorption peaks for specific wavelengths of near-infrared light. The collected raw spectral data is first corrected by multivariate scattering to eliminate the effects of particle size variations. The light scattering interference caused by differences in uniformity and bulk density is eliminated. Then, the electronic noise of the sensor and the interference of ambient light are removed by smoothing filtering. Subsequently, a competitive adaptive reweighting algorithm is used to screen out the 20 feature wavelengths with the highest correlation with grinding quality from the full wavelength spectrum. Redundant information is eliminated, and finally, low-dimensional and highly recognizable chili material spectrum features are obtained. The chili material spectrum features are feature vectors that can indirectly characterize key physicochemical properties such as material moisture content, crude fiber ratio, capsaicin concentration, maturity and pre-crushing degree after feature extraction. These properties directly determine the grinding difficulty of chili and the quality indicators such as particle size, uniformity and flavor of the final product. Subsequently, the real-time acquired chili material spectral features are input into a pre-trained adaptive quality indicator model for analysis. This model is a hybrid deep learning model that integrates gradient boosting trees and long short-term memory networks. Gradient boosting trees are good at mining the nonlinear mapping relationship between spectral features and pulping quality, while long short-term memory networks can capture the temporal change trend of material characteristics in consecutive batches. The model is trained by tens of thousands of sets of material spectral data collected synchronously in historical production and corresponding offline quality detection data, namely average particle size D50, particle size distribution span, uniformity coefficient, and pulp yield. The model is also updated online incrementally based on the actual production data every day to ensure prediction accuracy. Subsequently, the model first predicts in real time, based on the input material spectrum characteristics, multiple refining quality indicators that the current batch of materials will produce under the default control parameters. Then, it aligns these predicted values ​​with the enterprise's preset standard refining quality indicators one by one and calculates the difference to obtain multiple refining quality indicator differences. The refining quality indicator difference is the quantitative deviation between the predicted quality and the target quality. The larger the absolute value, the more serious the deviation from the target quality. The sign indicates the direction of deviation.

[0028] Next, based on the equipment mechanism model and historical control database, a sensitivity coefficient matrix is ​​established between the set of actuators and the differences in multiple grinding quality indicators through partial least squares regression. The sensitivity coefficient refers to the magnitude of the change in the corresponding quality indicator caused by the unit parameter change of the actuator, quantifying the adjustment capability of each actuator to different quality deviations. For example, the sensitivity coefficient of the stator-rotor gap adjustment servo mechanism to particle size D50 can reach 0.85, that is, for every 1μm change in gap, particle size D50 changes by about 0.85μm, while the sensitivity coefficient of the cooling circulation pump flow rate to particle size is only 0.12, indicating that gap adjustment is the most effective means of correcting particle size deviation. According to the sensitivity coefficient matrix, the comprehensive sensitivity coefficient of each actuator to all quality indicator differences is calculated. The set of actuators is initially sorted from high to low according to the comprehensive sensitivity coefficient, and actuators with comprehensive sensitivity coefficients greater than the preset sensitivity threshold are selected as candidate adaptive actuators.

[0029] Subsequently, the cluster label of each candidate actuator is obtained, namely the high, medium, and low stability levels obtained from the previous clustering. A corresponding priority coefficient is assigned to each stability level. The priority coefficient is used to adjust the weight of the comprehensive sensitivity coefficient, reflecting the influence of the actuator's own stability on the adjustment priority. The priority coefficient of the high stability control unit is set to 1.0, medium stability to 0.7, and low stability to 0.4. This is because the actuator with higher stability is less likely to experience control trajectory divergence, parameter overshoot, and mechanical oscillation during dynamic adjustment, and can maintain system stability while ensuring adjustment accuracy. The priority coefficient is multiplied by the original comprehensive sensitivity coefficient of the candidate actuator to obtain the updated comprehensive sensitivity coefficient. The actuators are then sorted from high to low according to the updated comprehensive sensitivity coefficient. Finally, the top 3-5 actuators with the highest comprehensive sensitivity coefficients are selected as the adaptive actuators for this adjustment.

[0030] S500: Retrieve and invoke multiple parameter control constraints according to the cluster labels to which the multiple adaptive execution units belong, solve the parameters of the multiple adaptive execution units under the multiple parameter control constraints, and output multiple sets of stable control parameters.

[0031] Specifically, firstly, based on the unique clustering label attached to each adaptive actuator, the corresponding differential constraint set is accurately retrieved and called from the stable control unit-parameter control constraint mapping library built synchronously during the clustering stage. This mapping library is an indexable database formed by binding each clustered actuator with its exclusive parameter control constraint. The parameter control constraint is a dynamic tightening of the hardware boundary based on the actual operating stability level of the actuator. Specifically, it includes four core dimensions: upper limit of adjustment step size, upper limit of rated drive current, upper limit of transient maximum temperature rise rate, and upper limit of mechanical axial displacement. For example, the upper limit of adjustment step size for a high-stability control unit can be set to 80%-100% of the rated hardware step size, allowing for a large single adjustment to quickly respond to sudden changes in material properties. In contrast, the upper limit of adjustment step size for a low-stability control unit is only 20%-40% of the rated hardware step size, avoiding control trajectory divergence and parameter overshoot through gradual adjustment with small steps.

[0032] Subsequently, a nonlinear programming optimization model with multiple constraints was constructed with the objective of minimizing the weighted sum of squares of the differences between multiple predicted grinding quality indicators output by the adaptive quality indicator model. The weight coefficients were preset according to the production importance of each quality indicator. For example, the weight of the average particle size D50 of chili sauce was set to 0.6, the weight of the particle size distribution span was set to 0.25, and the weight of the pulp yield was set to 0.15, ensuring that core quality indicators were corrected first. The constraints strictly limited the parameter adjustment range of each adaptive execution component to not exceeding the parameter control constraints it retrieved. Simultaneously, a smoothing constraint on parameter changes between adjacent batches was added to avoid parameter abrupt changes impacting the equipment. A sequential quadratic programming algorithm was used to solve the constrained parameters. This algorithm decomposes the complex nonlinear constrained optimization problem into a series of quadratic programming subproblems for iterative solving, possessing both fast convergence speed and high solution efficiency. Its high precision enables it to meet the millisecond-level real-time control requirements of industrial sites. The solution process employs an initial parameter hot-start strategy, using the stable control parameters from the previous grinding cycle as the initial values ​​for this solution. By leveraging the continuity of parameter changes during the grinding process, the number of iterations is significantly reduced, typically requiring only 5-10 iterations to converge to the preset accuracy. After obtaining multiple sets of initial stable control parameters, they are input back into the quality index adaptive model for closed-loop feedback analysis. The final grinding quality under these parameters is simulated and calculated, resulting in an updated predicted grinding quality index difference. If all updated index differences are less than the preset acceptable threshold (e.g., particle size D50 deviation ≤ 2μm), the set of parameters is directly output. If any index difference still exceeds the threshold, the objective function is corrected using the updated index difference, and a second iteration is performed under the original constraints until the quality requirements are met.

[0033] The final output consists of multiple sets of stable control parameters corresponding to multiple adaptive actuators. Each set of parameters includes not only the target operating value of the actuator, but also the adjustment rate and adjustment step size that meet its constraints. For example, the target gap value, maximum single adjustment step size and adjustment frequency of the stator-rotor gap adjustment servo mechanism, the target speed and maximum acceleration rate of the spindle drive frequency converter motor, etc., to ensure that the actuator can complete the adjustment action smoothly and accurately after the parameters are issued.

[0034] Furthermore, in the method provided in the application embodiment, control stability analysis is performed on the real-time operating condition dataset to obtain a control stability feature set. The method includes: obtaining a single real-time operating condition sequence for each execution component in the real-time operating condition dataset, wherein the single real-time operating condition sequence includes at least the real-time drive current of the stator-rotor gap adjustment servo mechanism of the grinding equipment; mapping the single real-time operating condition sequence to a high-dimensional topological space using time delay parameters and stability feature embedding dimensions to obtain a high-dimensional phase space trajectory matrix; performing singular value decomposition on the high-dimensional phase space trajectory matrix to obtain a singular value spectrum; calculating a probability density distribution based on the singular value spectrum; calculating the singular value control spectrum entropy based on the probability density distribution; and outputting a control stability feature set based on the singular value control spectrum entropy.

[0035] Specifically, the real-time operating condition dataset in the form of a multivariate time series is first split according to the unique identifier of the actuator to extract the single real-time operating condition sequence corresponding to each actuator. That is, the one-dimensional time series data set formed by a certain operating condition parameter of a single actuator within a continuous equally spaced sampling period includes the real-time drive current sequence of the stator-rotor gap adjustment servo mechanism of the grinding equipment as the core analysis object. This parameter directly maps the dynamic adjustment process of the grinding gap, and its microsecond-level current fluctuations will be directly transmitted to the particle size uniformity of the chili sauce. It can also most sensitively reflect the hidden unstable factors such as mechanical wear, servo gain mismatch, and load disturbance inside the actuator. At the same time, the sequences of rotor axial displacement, servo motor output torque, and encoder feedback position are collected as auxiliary verification features.

[0036] Subsequently, the time delay parameter was determined through autocorrelation analysis and nearest neighbor continuity analysis. The two hyperparameters for phase space reconstruction are the stable feature embedding dimension *m*, where the time delay parameter is... The optimal delay time is obtained by calculating the autocorrelation function of a single real-time operating condition sequence. When the autocorrelation coefficient drops to a preset correlation threshold, the optimal delay time is determined. This ensures that the reconstructed state vector contains sufficient independent dynamic information, avoiding redundancy between adjacent data points. The stable feature embedding dimension *m* is determined through nearest-neighbor continuity analysis. This involves gradually increasing the embedding dimension and calculating the average distance change rate between adjacent state points in the phase space. The optimal dimension is the one where the distance change rate drops to a preset proportional threshold, ensuring that all nonlinear dynamic information hidden in the one-dimensional time series data can be completely mapped to a high-dimensional space without folding. Based on this determined... The process of performing a high-dimensional topological space mapping with m is specifically as follows: using the time delay parameter A single real-time operating condition sequence with a state interval of length N. Slice the sample to generate N-(m-1) There are three equal-length delay state vectors, each of which has the following form: ; The value of t ranges from (m-1). +1 to N, and then reconstruct the state evolution matrix by arranging and combining all the delayed state vectors in chronological order row by row. This matrix can be directly mapped to the m-dimensional topological space to obtain the high-dimensional phase space trajectory matrix. The mathematical form of this matrix is ​​a matrix with L = N - (m - 1) rows. A two-dimensional real matrix with m columns is used, where each row corresponds to a state point in the m-dimensional phase space. The continuous curve formed by connecting all rows in time sequence is the phase space trajectory of the actuator control process. This trajectory can intuitively and quantitatively characterize the dynamic control behavior of the actuator: the trajectory of a stable actuator will converge to a compact attractor, exhibiting regular periodic or quasi-periodic motion, while the trajectory of an unstable actuator will diverge or exhibit chaotic motion. At the same time, the matrix can also extract the periodic characteristics of the trajectory, corresponding to the inherent oscillation frequency and disturbance response characteristics of the actuator, and the actuator's ability to adjust to changes in external load.

[0037] Next, singular value decomposition is performed on the high-dimensional phase space trajectory matrix A, decomposing it into the product of three matrices: ; in It is an L×m orthogonal left singular matrix. It is an m×m orthogonal right singular matrix. Given an m×m diagonal singular value matrix, the non-zero elements on the diagonal are... These are the singular values. Arranging all singular values ​​in descending order yields the singular value spectrum. The magnitude of the singular values ​​corresponds to the proportion of system energy in different directions in phase space. The more concentrated the singular values ​​are on a few large values, the more singular the system's motion modes and the more stable the control. Subsequently, the probability density distribution of the singular value spectrum is calculated, specifically: first, based on the range of singular values... Divide it into k discrete intervals of equal width, where k is usually between 10 and 20, dynamically adjusted according to the number of singular values, and then count the number of singular values ​​contained in each interval. Calculate the probability of each interval. , Given the total number of singular values, we obtain a probability density distribution consisting of interval endpoints and corresponding probabilities. This distribution can intuitively reflect the concentration of singular values. The probability density distribution of stable execution components will show sharp peaks in a few low-order intervals, while the probability density distribution of unstable execution components is relatively flat, covering multiple intervals.

[0038] Subsequently, the singular value control spectrum entropy is calculated based on the probability density distribution, and its calculation formula is as follows: ; in Representing the logarithm to base 2, the unit of spectral entropy is bits. Its physical meaning is to measure the degree of disorder in the distribution of singular values: when all singular values ​​are exactly equal, that is, when the system energy is uniformly distributed in all directions, =1 / k, the spectral entropy reaches its maximum value. The corresponding system is in a completely unstable chaotic state; when there is only one non-zero singular value, that is, the system energy is completely concentrated in one direction, =1, the rest =0, spectral entropy =0, corresponding to the system being in an ideal stable state. Based on extensive historical data, singular value control spectral entropy ≤0.3 indicates a highly stable actuator, and 0.3 < ≤0.7 indicates moderate stability. A value >0.7 indicates low stability. Finally, the singular value control spectrum entropy, the cumulative energy ratio of the first three singular values, the phase space trajectory convergence coefficient, the disturbance response recovery time, and the standard deviation of the singular value distribution corresponding to each execution component are integrated to form a set of control stability features. This set is a multi-dimensional vector set with the same dimension as the number of extracted stability features, and each element is a quantifiable and comparable numerical indicator.

[0039] Furthermore, in the method provided in the application embodiment, the single real-time operating condition sequence is mapped to a high-dimensional topological space using a time delay parameter and a stable feature embedding dimension. The method for obtaining the time delay parameter and the stable feature embedding dimension includes: obtaining the operating condition features between adjacent sampling points in the single real-time operating condition sequence; performing autocorrelation analysis on the operating condition features of the single real-time operating condition sequence, and using the delay time corresponding to the decrease to a preset correlation threshold as the time delay parameter; performing nearest neighbor continuity analysis on the operating condition features of the single real-time operating condition sequence, and using the dimension corresponding to the decrease to a preset proportion threshold as the stable feature embedding dimension.

[0040] Specifically, the operating characteristics between adjacent sampling points are first calculated from the single real-time operating condition sequence of each extracted actuator. The operating characteristics are the difference between two adjacent sampling points, the first-order rate of change, and the second-order acceleration. These dynamic characteristics can more sensitively capture the instantaneous fluctuations of the operating conditions of the actuator, rather than just reflecting the steady-state operating state. For example, the adjacent difference of the real-time drive current of the stator-rotor gap adjustment servo mechanism can directly reflect the instantaneous load change of the servo motor and the intensity of the adjustment action, which is more suitable for nonlinear dynamic characteristic analysis than the original current value.

[0041] Subsequently, the two parameters required for phase space reconstruction were determined through autocorrelation analysis and nearest-neighbor continuity analysis: the time delay parameter. The choice of these two parameters, along with the stable feature embedding dimension *m*, directly determines the quality of the high-dimensional phase space reconstruction. If the value is too small, the reconstructed state vector will contain a lot of redundant information. If m is too large, the temporal correlation of the sequence will be lost; if m is too small, the dynamic information of the original system will be folded in high-dimensional space; if m is too large, the influence of noise on the analysis results will be amplified. The autocorrelation analysis process is as follows: for a single real-time operating condition sequence of length N... First, calculate the global mean μ of the sequence, then iterate through different delay times. , The value range is 1 to , Typically, 1 / 10 of the sequence length is taken; in this scheme, the sampling frequency is 100Hz. Set to 100, corresponding to a time span of 1 second, according to the autocorrelation function formula: ; Calculate the autocorrelation coefficient R(τ) for each τ. The value of the autocorrelation coefficient ranges from [-1, 1]. The larger the absolute value, the stronger the linear correlation between the two data points at a delay of τ. A positive value indicates a positive correlation, and a negative value indicates a negative correlation.

[0042] Then, the autocorrelation coefficient was plotted as a function of delay time. The changing decay curve, the corresponding value when the curve drops to a preset relevant threshold. The value is the optimal time delay parameter, because when R(τ)≤0.15, x(t) and x(t+ The linear correlation between the states is weak enough to ensure that the reconstructed delayed state vector contains sufficient independent dynamic information, while the causal relationship of the sequence will not be broken due to excessively large τ.

[0043] Next, nearest-neighbor continuity analysis is performed to determine the embedding dimension *m* of the stable features. This analysis is based on the embedding theorem, which states that for any deterministic nonlinear dynamic system, as long as the embedding dimension *m* ≥ 2*d* + 1*, where *d* is the attractor dimension of the original system, the reconstructed phase space is topologically equivalent to the dynamic space of the original system. In practice, the dimension is gradually increased starting from the minimum embedding dimension *m* = 2. For each *m*, preliminary phase space state points are first reconstructed based on the current temporary *τ* value. Then, the *k* nearest neighbors of each state point are calculated, where *k* = 5, which is the average Euclidean distance *d*(m) between the five nearest state points. Finally, the rate of change of distance between adjacent dimensions is calculated. ; This rate of change reflects the degree of influence of increasing the embedding dimension on the relative positions of state points in the phase space. When Δd(m) drops to the preset proportional threshold, it indicates that continuing to increase the embedding dimension will not significantly change the topology of the phase space. All nonlinear dynamic information of the original system has been completely mapped into the space of the current dimension. At this time, m is the optimal stable feature embedding dimension. For example, the attractor dimension of a high-stability actuator is usually low, and the optimal m is generally 3-5, while the attractor dimension of a low-stability actuator is high, and the optimal m may be 6-8.

[0044] Furthermore, in the method provided in the application embodiment, the single real-time operating condition sequence is mapped to a high-dimensional topological space using a time delay parameter and a stable feature embedding dimension. The method includes: performing time-series processing on the single real-time operating condition sequence with the time delay parameter as the state interval to generate multiple delayed state vectors; combining and reconstructing the multiple delayed state vectors according to the stable feature embedding dimension to generate a state evolution matrix; and mapping the state evolution matrix to a high-dimensional topological space to obtain a high-dimensional phase space trajectory matrix, wherein the high-dimensional phase space trajectory matrix is ​​used to characterize the stable characteristics, periodic characteristics, and disturbance response characteristics of the state changes of the actuator during the control process.

[0045] Specifically, firstly, based on the optimal time delay parameter τ and stable feature embedding dimension m determined through autocorrelation analysis and nearest neighbor continuity analysis in the previous steps, a high-dimensional topological space mapping is performed on the single real-time operating condition sequence of each execution component. This is done by first mapping the single real-time operating condition sequence of length N to a fixed state interval using τ as the initial state interval. Time-series slicing is performed to generate multiple delayed state vectors, which are multi-dimensional vectors composed of sampled operating conditions from the current time and multiple previous delayed times, arranged in chronological order. Their purpose is to integrate the scattered historical dynamic information from one-dimensional time-series data into a single vector that can fully characterize the instantaneous state of the nonlinear system. The mathematical form of each delayed state vector is: ; The value of t ranges from (m-1). +1 to N, ensuring that each vector contains m consecutive intervals. The effective sampled values ​​can ultimately generate L=N-(m-1). Each is an independent delayed state vector.

[0046] Subsequently, a combined reconstruction process of the state evolution matrix is ​​performed. All generated delayed state vectors are stacked row by row in strict chronological order. Each row corresponds to the complete system state at a given time, and each column corresponds to the operating condition sample value at a fixed delay time. Finally, a two-dimensional real matrix with L rows and m columns is obtained, namely the state evolution matrix, whose mathematical form is: ; For example, when τ=2, m=3, N=10, L=10-(3-1)×2=6 delayed state vectors can be generated. The corresponding state evolution matrix is ​​a 6-row, 3-column matrix. The first row is [x(5),x(3),x(1)], the second row is [x(6),x(4),x(2)], and so on. This matrix completely records the continuous evolution of the control state of the execution component over time. The small change of each row of elements corresponds to the instantaneous fluctuation of the system state.

[0047] Next, a high-dimensional topological space mapping process is performed, where each row of the state evolution matrix is ​​treated as a coordinate point in an m-dimensional high-dimensional topological space. This high-dimensional topological space is an abstract mathematical space specifically constructed for analyzing the dynamic characteristics of nonlinear systems. Each dimension corresponds to the operating condition value at a delayed moment in the delayed state vector. The Euclidean distance between two points in the space directly corresponds to the degree of difference in the system state at those two moments. The continuous curve formed by connecting all coordinate points in chronological order is the phase space trajectory of the actuator control process. The state evolution matrix itself is a discretized mathematical representation of this continuous trajectory. Therefore, the state evolution matrix is ​​directly output as a high-dimensional phase space trajectory matrix, where each row corresponds to a discrete point in the m-dimensional phase space. Each state point, corresponding to an independent dimension of the phase space, has a numerical distribution that can comprehensively and quantitatively characterize three types of features of the actuator during the control process: stability features are manifested in the Euclidean distance between adjacent row vectors of the matrix gradually converging to a small stable value, corresponding to the phase space trajectory moving around a compact fixed point or limit cycle; periodic features are manifested in the matrix row vectors repeatedly exhibiting highly similar numerical combinations every fixed number of rows, corresponding to the phase space trajectory presenting a periodic closed curve; disturbance response features are manifested in the matrix row vectors exhibiting obvious numerical jumps when external loads change abruptly or control commands are adjusted, and then gradually recovering to the original distribution range, with the jump amplitude and recovery speed corresponding to the degree of disturbance impact and the system's anti-interference capability, respectively.

[0048] Furthermore, in the method provided in the application embodiment, the execution component set is clustered according to the control stability feature set. The method includes: initializing multiple cluster centers; calculating intra-class similarity of the control stability feature set according to the multiple cluster centers to obtain an intra-class similarity index; performing the division of the execution component set according to the intra-class similarity index to output multiple groups of execution components, wherein the intra-class similarity index of the same group of execution components is greater than a preset similarity threshold.

[0049] Specifically, the control stability feature set extracted in the previous steps is first preprocessed using min-max normalization. Features with different dimensions, such as singular value control spectrum entropy, the cumulative energy ratio of the first three singular values, phase space trajectory convergence coefficient, and disturbance response recovery time, are uniformly mapped to the [0,1] interval to eliminate the interference of dimensional differences on similarity calculation. The normalization formula for each feature is as follows: ; For the k-th original feature value of the i-th execution unit, and These are the minimum and maximum values ​​of the k-th feature of all execution components, respectively.

[0050] Subsequently, three cluster centers are initialized based on the three stability levels of high, medium, and low preset in this scheme. The cluster center is a benchmark multidimensional feature vector representing the average stability level of a certain type of execution component. Its initial value is obtained by averaging the features of thousands of execution components whose stability levels have been manually labeled in historical production. For example, the high stability cluster center corresponds to the feature combination with normalized singular value control spectrum entropy ≤ 0.3, trajectory convergence coefficient ≥ 0.9, and disturbance response recovery time ≤ 0.2. This supervised initialization method can significantly accelerate the cluster convergence speed and avoid getting trapped in local optima.

[0051] Next, the intra-class similarity calculation process is performed, using weighted Euclidean distance as the basis for similarity measurement. Unlike ordinary Euclidean distance, which assigns equal weights to all features, weighted Euclidean distance allocates differentiated weights based on the contribution of different features to control stability. In this scheme, the weight of singular value control spectral entropy is set to 0.5, the trajectory convergence coefficient to 0.3, the cumulative energy percentage of the first three singular values ​​to 0.15, and the disturbance response recovery time to 0.05. Specifically, for the normalized feature vector of the i-th execution unit... and the j-th cluster center The weighted Euclidean distance between the two: ; in Let the weight of the k-th feature be used, and then the intra-class similarity index formula be applied: ; Convert the distance to a similarity value between 0 and 1. The closer the value is to 1, the more similar the stability characteristics of the execution component are to the cluster center, and the higher the intra-cluster similarity. Subsequently, each execution component is assigned to the cluster with the highest similarity according to the principle of maximum similarity, completing the first preliminary division. Then, the mean of the feature vectors of all execution components in each cluster is recalculated as the new cluster center. The above similarity calculation and sample allocation process is repeated until the change in the Euclidean distance of all cluster centers in two consecutive iterations is less than the preset convergence threshold 1e-5, or the maximum number of iterations of 100 is reached. After the iteration converges, the intra-cluster similarity index of all execution components and corresponding cluster centers in each cluster is further verified to ensure that they are all greater than the preset similarity threshold of 0.85. For a few execution components with similarity below the threshold, they are reassigned to the cluster with the second highest similarity. If the condition is still not met, they are marked as isolated points and processed separately. Finally, multiple groups of execution components are output, and the execution components in the same group have highly homogeneous control stability characteristics.

[0052] Furthermore, in the method provided in the application embodiments, each stable control unit includes corresponding parameter control constraints, which include the upper limit of the adjustment step size of the corresponding actuator, the upper limit of the rated drive current, the upper limit of the transient maximum temperature rise rate, and the upper limit of the mechanical axial displacement.

[0053] Specifically, after clustering and mapping the actuators into three independent stable control units (high, medium, and low stability), based on the average control stability characteristics of each cluster and the rated hardware parameters of the actuators, a parameter control constraint specific to each unit is generated through a quantitative mapping relationship. The numerical control constraint is a dynamic tightening of the hardware boundary combined with the actual operating stability level of the actuators. It is a safety red line to ensure that the actuators are always in the stable operating range during dynamic adjustment. Specifically, it includes four dimensions: upper limit of adjustment step size, upper limit of rated drive current, upper limit of transient maximum temperature rise rate, and upper limit of mechanical axial displacement. The constraint value of each dimension is positively correlated with the average stability characteristics of the corresponding cluster. That is, the higher the stability of the control unit, the looser the constraint, which can give full play to its adjustment potential. Conversely, the lower the stability, the stricter the constraint, to avoid control trajectory divergence and equipment damage.

[0054] Subsequently, the upper limit of the adjustment step size is calculated. The adjustment step size refers to the maximum parameter change allowed when a single control command is issued by the executing component, which directly determines the correction speed of the quality deviation. Its value is determined by the mapping relationship between the cluster average singular value control spectrum entropy and the hardware rated step size: For high-stability control units, the average spectrum entropy is ≤0.3, and the upper limit of the adjustment step size is set to 80%-100% of the hardware rated step size, allowing for a large single adjustment to quickly respond to the sudden changes in the characteristics of different batches of chili materials; for medium-stability control units, 0.3 < average spectrum entropy ≤0.7, it is set to 40%-80%, using a medium step size to balance adjustment speed and stability; for low-stability control units, the average spectrum entropy is >0.7, it is only set to 20%-40%, using a small step size for gradual adjustment to avoid parameter overshoot and oscillation caused by excessive system inertia; Next, the upper limit of the rated drive current is calculated. This is the maximum instantaneous drive current allowed by the actuator during dynamic adjustment. Its value is the hardware rated drive current multiplied by the average trajectory convergence coefficient of the corresponding cluster. The trajectory convergence coefficient is an indicator of the speed at which the control trajectory of the actuator returns to steady state. The higher the coefficient, the stronger the system's anti-disturbance capability and the higher the allowable upper limit of current: for high-stability units, the average convergence coefficient is ≥0.9, and the upper limit of current is 95% of the rated value; for medium-stability units, the average convergence coefficient is 0.7≤average convergence coefficient<0.9, and the upper limit of current is 80%; for low-stability units, the average convergence coefficient is <0.7, and the upper limit of current is 65%. This constraint can effectively prevent the actuator from overheating of the windings, demagnetization of the permanent magnet, and gearbox impact wear caused by instantaneous current overload.

[0055] The maximum transient temperature rise rate was then calculated. This refers to the maximum temperature rise allowed per unit time during continuous dynamic adjustment of the actuator. Its value is negatively correlated with the average disturbance response recovery time of the cluster: for high-stability units with an average recovery time ≤ 100ms, the upper limit of the temperature rise rate is set at 0.5℃ / s; for medium-stability units with an average recovery time ≤ 300ms and a medium recovery time < 100ms, it is set at 0.3℃ / s; and for low-stability units with an average recovery time > 300ms, it is set at 0.15℃ / s. This is because the adjustment actions of low-stability actuators are more frequent and last longer. Excessive temperature rise can cause thermal deformation of mechanical parts, which in turn affects the control accuracy of the grinding gap and ultimately leads to uneven particle size of the chili sauce.

[0056] Finally, the upper limit of mechanical axial displacement is calculated. This is the maximum allowable displacement of the actuator in the axial direction, which is an insurmountable hard constraint for the stator-rotor gap adjustment servo mechanism. Its value is the maximum axial displacement of the hardware multiplied by a safety factor. The safety factor is positively correlated with the cumulative energy percentage of the average of the top three singular values ​​in the cluster: for high-stability units, the average cumulative energy percentage is ≥95%, and the safety factor is 0.9; for medium-stability units, the average cumulative energy percentage is 85% ≤ average cumulative energy percentage <95%, and the safety factor is 0.8; for low-stability units, the average cumulative energy percentage is <85%, and the safety factor is 0.7. This constraint can prevent stator-rotor collision accidents caused by parameter errors at the algorithm level, protecting the core components of the grinding equipment. After all calculations are completed, the constraint values ​​of the four dimensions are bound to the clustering tags of the corresponding stable control units to build a real-time searchable stable control unit-parameter control constraint mapping library. Stability analysis and constraint updates are re-executed every 100 batches of production to adapt to mechanical wear and performance degradation during long-term operation of the equipment.

[0057] Furthermore, in the method provided in the application embodiment, the chili material spectrum features are input into a quality index adaptive model for analysis to determine multiple adaptive execution components. The method includes: analyzing the chili material spectrum features according to the quality index adaptive model to obtain multiple grinding quality indicators; obtaining multiple grinding quality indicator differences between the multiple grinding quality indicators and multiple standard grinding quality indicators; establishing a sensitivity coefficient between the set of execution components and the differences between the multiple grinding quality indicators; sorting the set of execution components according to the sensitivity coefficient, and selecting execution components with a sensitivity threshold greater than a preset sensitivity threshold as adaptive execution components, and outputting multiple adaptive execution components.

[0058] Specifically, firstly, a 20-dimensional low-dimensional, high-recognition chili material feature vector, selected through multivariate scattering correction to eliminate particle scattering interference, smoothing filtering to remove electronic noise, and a competitive adaptive reweighting algorithm, is used. This feature vector is a quantitative characteristic that can indirectly characterize the core physicochemical properties of chili material, such as moisture content, crude fiber ratio, capsaicin concentration, maturity, and pre-crushing degree. This feature vector is then input into a pre-trained, continuously updated online quality index adaptive model for end-to-end analysis. This model employs a dual-branch hybrid deep learning structure combining temporal feature extraction and nonlinear feature fitting. Specifically, the input layer receives the 20-dimensional material... The first branch is a two-layer long short-term memory network (LSTM) with 64 neurons per layer and a dropout rate of 0.2 to prevent overfitting. This layer captures the temporal trends of material characteristics across consecutive batches, eliminating prediction errors caused by batch-to-batch material fluctuations. The second branch is a gradient boosting tree layer containing 100 decision trees with a maximum depth of 6 and a learning rate of 0.1. This layer is used to mine the complex nonlinear mapping relationship between single-batch spectral features and pulping quality. The fusion layer employs a weighted fusion strategy, combining the temporal features output from the LSTM branch. The static features output from the XGBoost branch are concatenated and fused with each other using a weight ratio of 0.3:0.7. This weight allocation is based on historical data validation, as XGBoost has higher fitting accuracy for static spectral features. The output layer consists of four linear neurons, each corresponding to one of the four core grinding quality indicators. The specific training process of the model is divided into three stages: The first stage is data preparation, collecting production data from the past 12 months, including near-infrared raw spectra, preprocessed material spectrum features, offline physicochemical testing quality indicators, and corresponding execution component operating parameters. Thousands of valid samples are divided into groups at a ratio of 8:1:1. The training, validation, and test sets were used, and all quality metrics were Z-score standardized to eliminate dimensional differences. The second stage was offline pre-training, where the LSTM and XGBoost branches were trained separately until convergence. Then, the underlying parameters of the two branches were frozen, and only the fusion layer and output layer were trained. The mean squared error loss function was used, the Adam optimizer was used, the initial learning rate was 0.001, the batch size was 32, and an early stopping strategy was set, i.e., training was stopped if the validation set loss did not decrease for 5 consecutive epochs. Finally, the determination coefficient R² of the test set reached 0.962, and the mean absolute error was less than 1.2μm; The third stage is online incremental updates. After the daily production ends, 100-200 new samples verified offline are added to the incremental training set. An incremental learning algorithm is used to update only the top-level parameters of the model without retraining the entire network. Each update takes no more than 8 minutes, ensuring the model can adapt to slow drift caused by long-term equipment wear and changes in raw material origin. The model outputs multiple grinding quality indicators for the current batch of materials under default control parameters, including average particle size D50 (the particle size corresponding to 50% of the cumulative particle size distribution, directly determining the smooth texture of the chili sauce), particle size distribution span (reflecting the width of the particle size distribution; a smaller span indicates higher product uniformity), and uniformity coefficient (measuring the consistency of product quality at different discharge ports of the grinder) and pulp yield (the quality of qualified chili sauce produced per unit mass of chili raw material).

[0059] Subsequently, the model-predicted quality indicators are aligned with the enterprise's preset standard refining quality indicators one by one, and the differences are calculated to obtain multiple refining quality indicator differences. The sign of the indicator difference indicates the direction of quality deviation. For example, a positive D50 indicates that the product is too coarse, and a negative D50 indicates that the product is too fine. The absolute value indicates the degree of deviation. This indicator difference is the core optimization target for all subsequent control actions.

[0060] Next, based on the mechanism model of the refining equipment and the historical control database containing tens of thousands of control parameters and quality response data, a partial least squares regression algorithm is used to establish a sensitivity coefficient matrix of the set of actuators and the difference between multiple refining quality indicators. The sensitivity coefficient refers to the magnitude of the change in the corresponding quality indicator caused by the unit parameter change of the actuator, which quantifies the adjustment capability of each actuator to different quality deviations.

[0061] Subsequently, differentiated weights are assigned based on the production importance of each quality indicator. For example, the average particle size D50 has a weight of 0.6, the particle size distribution span is 0.25, the pulp yield is 0.1, and the uniformity coefficient is 0.05. The sensitivity coefficients of multiple single indicators for each actuator are weighted and summed to obtain a comprehensive sensitivity coefficient. The actuator set is sorted from high to low according to the comprehensive sensitivity coefficient, and actuators with a comprehensive sensitivity coefficient greater than a preset sensitivity threshold are selected as adaptive actuators. This threshold is calibrated using historical data and can minimize the number of actuators involved in the adjustment while ensuring the adjustment effect. Finally, 3-5 core actuators are output to avoid control coupling and system oscillation caused by adjusting too many actuators at the same time.

[0062] Furthermore, in the method provided in the application embodiment, multiple adaptive execution components are output. The method further includes: obtaining the cluster label to which the set of execution components belongs; generating a priority coefficient of the set of execution components according to the cluster label; updating the sensitivity coefficient according to the priority coefficient; reordering the set of execution components according to the updated sensitivity coefficient; and obtaining multiple updated adaptive execution components.

[0063] Specifically, firstly, the clustering labels corresponding to all execution components are obtained from the execution component-clustering label mapping table constructed in the clustering stage. These labels are based on the dynamic control capability rating of the execution components obtained by weighted K-means clustering, which is based on quantitative stability characteristics such as singular value control spectrum entropy, trajectory convergence coefficient, and disturbance response recovery time. The rating is divided into three stability levels: high, medium, and low. Each label is directly bound to a stable control unit and its parameter control constraints.

[0064] Subsequently, priority coefficients are generated according to the stability level of the cluster labels. The priority coefficients are weighting factors used to correct the original sensitivity coefficients. Their core function is to incorporate the stability of the execution components into the adaptive adjustment decision system. The optimal coefficient allocation rule is obtained by calibrating through tens of thousands of sets of historical control data: the priority coefficient of the high stability control unit is set to 1.0, the medium stability is set to 0.7, and the low stability is set to 0.4. This rule can control the probability of system oscillation below 0.5% while ensuring the speed of quality adjustment.

[0065] Next, the sensitivity coefficient update calculation process is performed. The original comprehensive sensitivity coefficient of each actuator, which is the sensitivity coefficient that reflects the adjustment effect only and is obtained by weighting multiple quality indicators such as average particle size D50 and particle size distribution span, is multiplied element by element with the corresponding priority coefficient to obtain the updated comprehensive sensitivity coefficient. For example, the original comprehensive sensitivity coefficient of the stator-rotor gap adjustment servo mechanism is 0.82. If it belongs to the high stability control unit, it will still be 0.82 after the update, maintaining the highest priority. The original comprehensive sensitivity coefficient of the feed screw conveyor is 0.55. If it is classified as a low stability control unit due to long-term mechanical wear, it will be 0.55 × 0.4 = 0.22 after the update, which is lower than the preset sensitivity threshold of 0.3.

[0066] Finally, the set of execution components is reordered from high to low according to the updated comprehensive sensitivity coefficient, and execution components with a comprehensive sensitivity coefficient greater than the preset sensitivity threshold of 0.3 are selected again. Finally, 3-5 adaptive execution components with both strong quality adjustment capabilities and high operational stability are output.

[0067] Furthermore, in the method provided in the application embodiment, under the multiple parameter control constraints, the parameters of the multiple adaptive execution components are solved to output multiple sets of stable control parameters. The method includes: solving the parameters of the multiple adaptive execution components to obtain multiple sets of initial stable control parameters; performing feedback analysis on the quality index adaptive model based on the multiple sets of initial stable control parameters to obtain multiple predicted grinding quality index differences; and obtaining multiple sets of stable control parameters with minimizing the multiple predicted grinding quality index differences as the solution objective under the multiple parameter control constraints.

[0068] Specifically, firstly, based on the unique clustering label attached to each adaptive actuator, the corresponding differentiated constraint set is retrieved and called from the stable control unit-parameter control constraint mapping library synchronously built during the clustering stage. The parameter control constraint is a dynamic tightening of the hardware boundary combined with the actual operating stability level of the actuator. It is a safety red line to ensure that the actuator is always in the stable operating range during dynamic adjustment. Specifically, it includes the upper limit of adjustment step size, that is, the maximum parameter change allowed by a single control command of the actuator, which directly determines the quality deviation correction speed; the upper limit of rated drive current, that is, the maximum instantaneous drive current allowed during dynamic adjustment, which prevents winding overheating and permanent magnet demagnetization; the upper limit of transient maximum temperature rise rate, that is, the maximum temperature rise allowed per unit time, which avoids mechanical thermal deformation from affecting control accuracy; and the upper limit of mechanical axial displacement, that is, a hard constraint that cannot be broken in the axial direction, which prevents stator and rotor collisions. The constraints of high-stability control units are more relaxed to release adjustment potential, while low-stability units avoid control trajectory divergence through strict constraints.

[0069] Subsequently, a nonlinear programming optimization model with multiple constraints was constructed, with the objective of minimizing the weighted sum of squares of the differences in multiple predicted pulping quality indicators. The weight coefficients were preset based on the production importance of each quality indicator, with preset weights of 0.6 for average particle size D50, 0.25 for particle size distribution span, and 0.15 for pulp yield. This ensures that core indicators directly affecting product taste are prioritized for correction. The constraints strictly limit the parameter adjustment range of each adaptive actuator to its specific constraints, while also incorporating a smoothing constraint on parameter changes between adjacent batches, meaning that a single adjustment amplitude cannot exceed 10% of the parameter in the previous cycle, preventing sudden parameter changes from impacting mechanical components such as gearboxes and bearings. The parameter solution is performed using a sequential quadratic programming algorithm. This algorithm decomposes the complex nonlinear constraint optimization problem into a series of quadratic programming subproblems and solves them iteratively. The solution process adopts a hot-start strategy for initial parameters, that is, the stable control parameters of the previous grinding cycle are used as the initial values ​​for this solution. By taking advantage of the continuity of parameter changes during the grinding process, the number of iterations is reduced from the usual 20-30 times to 5-10 times. When the iteration converges to the preset accuracy or reaches the maximum number of iterations, the initial stable control parameters are obtained. These are candidate parameters that only satisfy the constraint conditions and minimize the initial objective function. They have not yet been verified by closed-loop effect and may lead to poor actual adjustment effect due to model prediction errors or unknown disturbances in the field.

[0070] Subsequently, a feedback analysis process is performed, inverting multiple sets of initial stable control parameters into the quality index adaptive model to simulate the complete grinding process after parameter adjustment. First, the initial parameters are spliced ​​with the material spectrum characteristics of the current batch of chili peppers as model input. The XGBoost branch of the model fits the influence of parameter changes on the grinding quality of a single batch, and the LSTM branch combines the parameter-quality response time series data of historical batches to correct the prediction bias. Finally, the predicted grinding quality index under this set of parameters is output, namely, average particle size D50, particle size distribution span, uniformity coefficient, and pulp yield.

[0071] Next, the predicted quality indicators are aligned with the enterprise's preset standard grinding quality indicators one by one, and the differences are calculated to obtain multiple updated predicted grinding quality indicator differences. The sign of the indicator difference indicates the direction of quality deviation. For example, a positive D50 indicates that the product is too coarse, and the absolute value indicates the degree of deviation. If all the updated quality indicator differences are less than the preset qualified threshold, the initialization parameters are directly output as the final stable control parameters. If there are still cases where the indicator differences exceed the threshold, the objective function is corrected with the updated indicator differences, and a second iteration is performed under the control constraints of the original parameters. A maximum of 3 second iterations are performed to ensure real-time performance. If the requirements are still not met after 3 iterations, the current optimal solution is output and a minor alarm is triggered to prompt manual review.

[0072] The final output consists of multiple sets of stable control parameters, each corresponding to a specific adaptive actuator. Each set of parameters includes not only the target operating value of the actuator but also the adjustment rate and step size rules that meet its constraints. For example, the target gap value, maximum single adjustment step size, and maximum adjustment frequency of the stator-rotor gap adjustment servo mechanism, and the target speed and maximum acceleration rate of the spindle drive variable frequency motor. This ensures that the actuator can smoothly and accurately complete the adjustment action after the parameters are issued. The control effect of this scheme is compared with that of traditional control methods. Figure 3 As shown.

[0073] In summary, the self-stabilizing control method for quality indicators of chili sauce grinding provided in this application has the following technical effects: By collecting real-time datasets of all operating conditions of the execution components, quantitative analysis of control stability was completed based on phase space reconstruction and singular value control spectrum entropy. The execution components were clustered into three types of stable control units: high, medium, and low, and matched with differentiated parameter control constraints. At the same time, near-infrared spectroscopy was used to capture the spectral characteristics of chili pepper materials in real time. By integrating gradient boosting tree and long short-term memory network into an adaptive quality index model, adaptive execution components with both high adjustment sensitivity and high operational stability were selected. Under the corresponding parameter constraints, the optimal stable control parameters were solved with the goal of minimizing the difference in predicted grinding quality index, which ultimately improved the quality stability and equipment reliability of chili sauce grinding production.

[0074] Example 2, based on the same inventive concept as the self-stabilizing control method for quality indicators of chili sauce grinding in the previous examples, such as... Figure 2 As shown in the embodiment of this application, a self-stabilizing control system for quality indicators of chili sauce grinding is provided. The system includes: The acquisition module 11 is used to acquire the set of actuators of the grinding equipment and acquire the real-time operating condition dataset of the set of actuators; the analysis module 12 is used to perform control stability analysis on the real-time operating condition dataset and acquire a set of control stability features; the clustering module 13 is used to cluster the set of actuators according to the set of control stability features and acquire multiple groups of actuators, wherein the multiple groups of actuators correspond to multiple stable control units, and each stable control unit includes corresponding parameter control constraints; the capture module 14 is used to capture the chili material spectrum features of the grinding equipment, input the chili material spectrum features into the quality index adaptive model for analysis, and determine multiple adaptive actuators; the control module 15 is used to retrieve and call multiple parameter control constraints according to the cluster labels to which the multiple adaptive actuators belong, solve the parameters of the multiple adaptive actuators under the multiple parameter control constraints, and output multiple sets of stable control parameters.

[0075] Furthermore, the analysis module 12 is also used to perform the following steps: obtaining a single real-time operating condition sequence for each execution component in the real-time operating condition dataset, wherein the single real-time operating condition sequence includes at least the real-time drive current of the stator-rotor gap adjustment servo mechanism of the grinding equipment; mapping the single real-time operating condition sequence to a high-dimensional topological space using time delay parameters and stability feature embedding dimensions to obtain a high-dimensional phase space trajectory matrix; performing singular value decomposition on the high-dimensional phase space trajectory matrix to obtain a singular value spectrum; calculating a probability density distribution based on the singular value spectrum; calculating the singular value control spectrum entropy based on the probability density distribution; and outputting a control stability feature set based on the singular value control spectrum entropy.

[0076] Furthermore, the analysis module 12 is also used to perform the following steps: obtaining the operating condition features between adjacent sampling points in the single real-time operating condition sequence; performing autocorrelation analysis on the operating condition features of the single real-time operating condition sequence, and using the delay time corresponding to the decrease to a preset correlation threshold as the time delay parameter; performing nearest neighbor continuity analysis on the operating condition features of the single real-time operating condition sequence, and using the dimension corresponding to the decrease to a preset proportion threshold as the stable feature embedding dimension.

[0077] Furthermore, the analysis module 12 is also used to perform the following steps: performing time-series processing on the single real-time operating condition sequence with the time delay parameter as the state interval to generate multiple delayed state vectors; combining and reconstructing the multiple delayed state vectors according to the stability feature embedding dimension to generate a state evolution matrix; mapping the state evolution matrix to a high-dimensional topological space to obtain a high-dimensional phase space trajectory matrix, wherein the high-dimensional phase space trajectory matrix is ​​used to characterize the stability characteristics, periodic characteristics, and disturbance response characteristics of the state changes of the actuator during the control process.

[0078] Furthermore, the clustering module 13 is also used to perform the following steps: initialize multiple cluster centers, calculate intra-class similarity of the control stability feature set according to the multiple cluster centers, and obtain intra-class similarity index; perform the division of the execution component set according to the intra-class similarity index, and output multiple groups of execution components, wherein the intra-class similarity index of the same group of execution components is greater than a preset similarity threshold.

[0079] Furthermore, the clustering module 13 is also used to perform the following steps: each stable control unit includes corresponding parameter control constraints, the parameter control constraints including the upper limit of the adjustment step size of the corresponding actuator, the upper limit of the rated drive current, the upper limit of the transient maximum temperature rise rate, and the upper limit of the mechanical axial displacement.

[0080] Furthermore, the capture module 14 is also used to perform the following steps: analyze the chili material spectrum characteristics according to the quality index adaptive model to obtain multiple grinding quality indicators; obtain multiple grinding quality indicator differences between the multiple grinding quality indicators and multiple standard grinding quality indicators; establish a sensitivity coefficient between the set of execution components and the multiple grinding quality indicator differences; sort the set of execution components according to the sensitivity coefficient, and take the execution components with a sensitivity threshold greater than a preset sensitivity threshold as adaptive execution components, and output multiple adaptive execution components.

[0081] Furthermore, the capture module 14 is also used to perform the following steps: obtain the cluster label to which the set of execution components belongs, generate the priority coefficient of the set of execution components according to the cluster label; update the sensitivity coefficient according to the priority coefficient, reorder the set of execution components according to the updated sensitivity coefficient, and obtain multiple updated adaptive execution components.

[0082] Furthermore, the control module 15 is also used to perform the following steps: solving the parameters of the plurality of adaptive execution components to obtain a plurality of initial stable control parameters; performing feedback analysis on the quality index adaptive model based on the plurality of initial stable control parameters to obtain a plurality of predicted grinding quality index differences; and obtaining a plurality of stable control parameters with minimizing the plurality of predicted grinding quality index differences as the solution objective under the plurality of parameter control constraints.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for self-stabilizing quality indicators of chili sauce grinding, characterized in that, The method includes: Obtain the set of actuators of the grinding equipment, and collect the real-time operating condition dataset of the set of actuators; Perform control stability analysis on the real-time operating condition dataset to obtain a set of control stability features; The set of execution components is clustered according to the control stability feature set to obtain multiple groups of execution components, wherein the multiple groups of execution components correspond to multiple stable control units, and each stable control unit includes corresponding parameter control constraints; Capture the chili material spectrum characteristics of the grinding equipment, input the chili material spectrum characteristics into the quality index adaptive model for analysis, and determine multiple adaptive execution components; According to the cluster labels to which the multiple adaptive execution units belong, multiple parameter control constraints are retrieved and invoked. Under the multiple parameter control constraints, the parameters of the multiple adaptive execution units are solved, and multiple sets of stable control parameters are output.

2. The method as described in claim 1, characterized in that, The method includes performing control stability analysis on the real-time operating condition dataset to obtain a set of control stability features. Obtain a single real-time operating condition sequence for each execution component in the real-time operating condition dataset, wherein the single real-time operating condition sequence includes at least the real-time drive current of the stator-rotor gap adjustment servo mechanism of the grinding equipment; The single real-time operating condition sequence is mapped to a high-dimensional topological space by using time delay parameters and stable feature embedding dimensions to obtain a high-dimensional phase space trajectory matrix; Singular value decomposition is performed on the high-dimensional phase space trajectory matrix to obtain the singular value spectrum; The probability density distribution is calculated based on the singular value spectrum, the singular value control spectrum entropy is calculated based on the probability density distribution, and the control stability feature set is output based on the singular value control spectrum entropy.

3. The method as described in claim 2, characterized in that, The method for mapping the single real-time operating condition sequence to a high-dimensional topological space using time delay parameters and stable feature embedding dimensions includes: Obtain the operating condition characteristics between adjacent sampling points in the single real-time operating condition sequence; Autocorrelation analysis is performed on the operating characteristics of the single real-time operating condition sequence, and the delay time corresponding to the decrease to a preset correlation threshold is used as the time delay parameter. The working condition features of the single real-time working condition sequence are analyzed by nearest neighbor continuity, and the dimension corresponding to the decrease to the preset proportion threshold is used as the stable feature embedding dimension.

4. The method as described in claim 3, characterized in that, The method for mapping the single real-time operating condition sequence to a high-dimensional topological space using time delay parameters and stable feature embedding dimensions includes: The single real-time operating condition sequence is processed using the time delay parameter as the state interval to generate multiple delayed state vectors. The multiple delayed state vectors are combined and reconstructed based on the embedded dimension of the stable features to generate a state evolution matrix; The state evolution matrix is ​​mapped to a high-dimensional topological space to obtain a high-dimensional phase space trajectory matrix. The high-dimensional phase space trajectory matrix is ​​used to characterize the stability characteristics, periodic characteristics, and disturbance response characteristics of the state changes of the actuator during the control process.

5. The method as described in claim 1, characterized in that, The method for clustering the set of execution components according to the control stability feature set includes: Multiple cluster centers are initialized, and intra-class similarity is calculated for the control stability feature set according to the multiple cluster centers to obtain the intra-class similarity index; The execution component set is divided according to the intra-class similarity index, and multiple groups of execution components are output, wherein the intra-class similarity index of the same group of execution components is greater than a preset similarity threshold.

6. The method as described in claim 5, characterized in that, Each stable control unit includes corresponding parameter control constraints, which include the upper limit of the adjustment step size of the corresponding actuator, the upper limit of the rated drive current, the upper limit of the transient maximum temperature rise rate, and the upper limit of the mechanical axial displacement.

7. The method as described in claim 1, characterized in that, The chili material spectrum characteristics are input into an adaptive quality index model for analysis to determine multiple adaptive execution components. The method includes: The chili material spectrum characteristics are analyzed based on the quality index adaptive model to obtain multiple grinding quality indicators; Obtain the difference between the multiple polishing quality indicators and multiple standard polishing quality indicators; Establish a sensitivity coefficient between the set of execution components and the differences in the multiple polishing quality indicators; The set of execution components is sorted according to the sensitivity coefficient, and the execution components with a sensitivity threshold greater than the preset sensitivity threshold are selected as adaptive execution components, and multiple adaptive execution components are output.

8. The method as described in claim 7, characterized in that, The method also includes outputting multiple adaptive execution components: Obtain the cluster label to which the set of execution components belongs, and generate the priority coefficient of the set of execution components according to the cluster label; The sensitivity coefficient is updated and calculated according to the priority coefficient, and the set of execution components is reordered according to the updated sensitivity coefficient to obtain multiple updated adaptive execution components.

9. The method as described in claim 7, characterized in that, The method involves solving for the parameters of the multiple adaptive actuators under the multiple parameter control constraints, and outputting multiple sets of stable control parameters. The parameters of the multiple adaptive execution components are solved to obtain multiple sets of initial stable control parameters; Based on the multiple sets of initial stable control parameters, feedback analysis is performed on the quality index adaptive model to obtain multiple predicted grinding quality index differences. Under the constraints of the multiple parameters, multiple sets of stable control parameters are obtained with the goal of minimizing the difference between the multiple predicted pulping quality indicators.

10. A self-stabilizing control system for quality indicators in chili sauce grinding, characterized in that, For implementing the self-stabilizing quality control method for chili sauce grinding according to any one of claims 1 to 9, the system comprises: The acquisition module is used to acquire the set of actuators of the grinding equipment and to acquire the real-time operating condition dataset of the set of actuators. The analysis module is used to perform control stability analysis on the real-time operating condition dataset and obtain a set of control stability features. The clustering module is used to cluster the set of execution components according to the control stability feature set to obtain multiple groups of execution components, wherein the multiple groups of execution components correspond to multiple stable control units, and each stable control unit includes corresponding parameter control constraints; The capture module is used to capture the chili material spectrum characteristics of the grinding equipment, input the chili material spectrum characteristics into the quality index adaptive model for analysis, and determine multiple adaptive execution components. The control module is used to retrieve and invoke multiple parameter control constraints according to the cluster labels to which the multiple adaptive execution components belong, solve the parameters of the multiple adaptive execution components under the multiple parameter control constraints, and output multiple sets of stable control parameters.

Images