Polishing powder size grading screening control system

Abstract

The invention relates to the technical field of industrial process control, and particularly discloses a polishing powder size grading screening control system which is characterized in that multi-mode signals are synchronously acquired through an optical diffraction spectrum and an acoustic emission sensor, and a standardized size feature sequence is generated through intrinsic mode decomposition and cross-mode feature fusion; adopting phase-space reconstruction and recursive analysis to quantify the dynamic stability of the process, and calculating a time sequence deterministic factor; joint sensitivity analysis and time-frequency domain coherent processing are carried out on the working condition parameters, a process disturbance characteristic spectrum is constructed, and an external disturbance influence value is calculated; carrying out multi-target collaborative decision making based on a deep reinforcement learning network, and generating a confidence score matrix of process parameter adjustment; and finally, executing an adaptive hierarchical control strategy according to the confidence level, and correcting weight parameters through real-time feedback to form closed-loop optimization.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, specifically to a polishing powder particle size classification and screening control system. Background Technology

[0002] Polishing powder, as a key material in precision polishing processes, directly determines the surface quality and processing accuracy of polished parts through its particle size distribution. In modern high-end manufacturing fields such as semiconductors, optical glass, and sapphire substrates, extremely stringent requirements are placed on the particle size uniformity of polishing powder. Traditional polishing powder particle size classification and screening often employs static sieving or centrifugal separation techniques based on a single optical sensor. These methods indirectly reflect particle size information by detecting the light scattering intensity of the particles and use classic PID control algorithms to adjust the parameters of the classification equipment.

[0003] The existing technology has the following shortcomings:

[0004] Traditional control strategies fail to fully consider the coupling effects between multiple operating parameters such as feeder speed, classifier wheel current, and system pressure, as well as their comprehensive impact on particle size distribution. They also lack the ability to quantitatively analyze and actively compensate for process disturbances. While some research has attempted to introduce algorithms such as neural networks into process control with the development of artificial intelligence, most studies are limited to single-objective parameter optimization. They fail to deeply integrate signal reliability assessment, multi-disturbance quantitative analysis, and multi-objective collaborative decision-making, making it difficult to meet the extreme precision and stability requirements of high-end polishing powder production processes. Summary of the Invention

[0005] The purpose of this invention is to provide a polishing powder particle size classification and screening control system to solve the problems mentioned above.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A polishing powder particle size classification and screening control system, comprising:

[0008] The data acquisition and feature extraction module is used to acquire the raw data stream output by the granularity analysis device in real time, preprocess the raw data stream and fuse multi-scale features to generate a standardized granularity feature sequence.

[0009] The dynamic stability assessment module performs nonlinear dynamic analysis on the standardized granularity characteristic sequence, constructs a process dynamic stability index, and calculates the time series deterministic factor.

[0010] The process disturbance quantification module performs joint sensitivity analysis on the operating parameters during the production process, constructs a process disturbance characteristic spectrum, and calculates the influence value of external disturbances.

[0011] The multimodal decision optimization module integrates the temporal deterministic factor and the influence value of external disturbances into a system state feature vector, inputs the system state feature vector into a pre-trained process optimization model for multi-objective collaborative decision-making, and outputs a confidence score matrix for graded process parameters.

[0012] The intelligent control execution module executes an adaptive pre-control strategy based on a confidence score matrix.

[0013] As a further aspect of the present invention: the process of generating the standardized granularity feature sequence specifically includes:

[0014] By synchronously acquiring the multimodal raw signals output by the particle size analysis device through parallel arranged optical diffraction spectrum sensors and acoustic emission sensors, time-aligned optical pulsation sequences and acoustic vibration sequences are obtained.

[0015] Intrinsic mode decomposition was performed on the optical pulsation sequence and the acoustic vibration sequence respectively, the intrinsic mode function components were extracted, and the sample entropy value and energy entropy value of each component were calculated.

[0016] Cross-modal fusion of intrinsic mode function components in the same frequency band from different sensors is performed, and multi-scale fusion features with time-frequency joint characteristics are generated by calculating their mutual information entropy and phase synchronization index.

[0017] The multi-scale fusion features are normalized by maximum and minimum values ​​and then reorganized according to the time window sequence to finally output the standardized granularity feature sequence.

[0018] As a further aspect of the present invention: the process of obtaining the mutual information entropy and phase synchronization index includes:

[0019] The phase synchronization index is calculated through the following steps: performing Hilbert transform on the intrinsic mode function components of the optical and acoustic frequency bands respectively to extract their instantaneous phase sequences; calculating the cosine function value of the difference between the two phase sequences; performing a moving average operation on the cosine function value within a time window, and the absolute value of the average value is the phase synchronization index characterizing the rhythm coordination of the two signals.

[0020] The mutual information entropy is calculated through the following steps: calculating the amplitude probability distribution and joint probability distribution of the two signals respectively; then calculating the individual information entropy and joint information entropy of the two signals respectively based on these probability distributions; and obtaining the mutual information entropy value characterizing the degree of nonlinear dependence between the two signals by subtracting the joint information entropy from the sum of the individual information entropies.

[0021] As a further aspect of the present invention: the nonlinear dynamic analysis specifically includes:

[0022] The delayed coordinate reconstruction method is used to reconstruct the phase space of the standardized granularity feature sequence, mapping the one-dimensional time series into a set of trajectory points in a high-dimensional phase space;

[0023] Calculate the recursive graph of the phase space trajectory point set, and extract the diagonal and vertical structure features in the recursive graph using image processing techniques;

[0024] Based on the diagonal and vertical structural features, the recursion rate and deterministic rate in the dynamic stability index of the calculation process are quantified respectively.

[0025] By weighting and fusing the recursion rate and the determinism rate, a comprehensive process dynamic stability index is generated.

[0026] As a further aspect of the present invention: the process for obtaining the time-series deterministic factor is as follows:

[0027] Calculate the permutation entropy values ​​of the standardized granularity feature sequence at different time scales to form a multi-scale permutation entropy spectrum;

[0028] The fluctuation characteristics and decay rate of the multi-scale permutation entropy spectrum were analyzed, and the fluctuation variance of the entropy value and the estimated value of the maximum Lyapunov exponent were extracted.

[0029] The entropy fluctuation variance is nonlinearly mapped and defuzzified with the estimate of the maximum Lyapunov exponent, and the output is a time-series deterministic factor ranging from zero to one.

[0030] As a further aspect of the present invention: the joint sensitivity analysis of the operating parameters during the production process specifically includes:

[0031] Time series of multi-dimensional operating parameters, including feeder speed, classifier current, system inlet negative pressure and outlet temperature, were collected.

[0032] Calculate the first-order and total-order sensitivity indices of the impact of fluctuations in parameters under various operating conditions on the standardized granularity feature sequence;

[0033] The sensitivity index of each parameter is convolved with its instantaneous fluctuation amplitude to generate a dynamic sensitivity matrix that characterizes the intensity of multi-parameter coupled disturbance.

[0034] Principal component analysis was performed to reduce the dimensionality of the dynamic sensitivity matrix. The first three principal component components were extracted and arranged in time series to construct a visualized process disturbance feature spectrum.

[0035] As a further aspect of the present invention: the process for obtaining the influence value of the external disturbance is as follows:

[0036] Extract the time series of principal component components representing different perturbation modes from the characteristic spectrum of process perturbation;

[0037] Calculate the complex wavelet coherence coefficients of the time series of each principal component and the standardized granularity feature series in the time-frequency domain;

[0038] By integrating the energy of the complex wavelet coherence coefficients in the main frequency band, the frequency domain energy influence weight of each perturbation mode on the granularity quality is obtained.

[0039] The frequency domain energy influence weights are multiplied by the variance contribution rates of the corresponding principal components and then summed. The final result is quantified as an external disturbance influence value in the range of zero to one.

[0040] As a further aspect of the present invention: the fusion process of the system state feature vector is as follows:

[0041] The standardized granularity feature sequence is reconstructed by using a time-series deterministic factor to generate a first feature vector characterizing the stability of the internal process.

[0042] The principal components of the process disturbance feature spectrum are selectively enhanced using the influence value of external disturbances to generate a second feature vector characterizing the sensitivity to external disturbances.

[0043] The first feature vector and the second feature vector are concatenated using tensors to generate a system state feature vector containing multi-dimensional state information.

[0044] As a further aspect of the present invention: the step of inputting the system state feature vector into a pre-trained process optimization model for multi-objective collaborative decision-making specifically includes:

[0045] The system state feature vector is input into a pre-trained deep reinforcement learning network, which consists of a policy network and a value network connected in parallel.

[0046] The strategy network outputs adjustment strategies for three sets of process parameters: the speed of the classifier wheel, the frequency of the induced draft fan, and the speed of the feeder, along with their corresponding confidence levels.

[0047] The value network predicts the long-term optimization effect of each strategy and provides a stability score.

[0048] The strategy confidence score and stability score are multiplied by a matrix to generate a three-dimensional confidence score matrix.

[0049] The pre-trained deep reinforcement learning network is obtained through the following steps:

[0050] Construct a multimodal training sample set, which includes system state feature vectors, process parameter adjustment actions, and corresponding long-term stability reward signals;

[0051] The policy network is responsible for generating parameter adjustment strategies, while the value network is responsible for evaluating the long-term returns of the strategies.

[0052] Network parameters for alternating optimization of policy networks and value networks;

[0053] When the mean squared error of the value network in evaluating the policy is lower than a preset threshold and the output policy of the policy network converges, the network training is completed and the final model parameters are saved.

[0054] As a further aspect of the present invention: the implementation of the adaptive pre-control strategy specifically includes:

[0055] The confidence score matrix is ​​analyzed to extract three sets of process parameter adjustment schemes for the grader wheel speed, induced draft fan frequency, and feeder speed, and their corresponding confidence scores.

[0056] The process parameter adjustment schemes are sorted according to their confidence scores, and the scheme with the highest confidence score is implemented first.

[0057] Based on the confidence level of the selected adjustment scheme, the response rate and adjustment range of the control system are automatically set. Schemes with high confidence adopt a direct and rapid adjustment strategy, while schemes with low confidence adopt a gradual and exploratory adjustment strategy.

[0058] After the parameters are adjusted, the feedback signal from the particle size analysis device is monitored in real time. By comparing the deviation between the expected and actual results, the weighting parameters of the confidence score matrix are corrected.

[0059] The beneficial effects of this invention are:

[0060] (1) This invention constructs an intelligent system capable of comprehensively perceiving the internal state and external disturbances of a system through multimodal sensing fusion, nonlinear dynamics analysis, and deep reinforcement learning decision-making. This system no longer relies on a single, easily distorted granular signal, but comprehensively utilizes multiple features and algorithms to achieve deep insight and accurate prediction of the production process state, thereby outputting control strategies with higher confidence, effectively avoiding control failures caused by sampling distortion or operating condition fluctuations, and ensuring the stability and consistency of product quality.

[0061] (2) This invention quantifies process disturbances and integrates them into the decision-making process. The intelligent execution module adaptively adjusts the control rate and amplitude based on the strategy confidence level. The system possesses complete closed-loop optimization capabilities, enabling online correction of the decision model parameters based on the actual deviation of the control effect. This allows the system to not only cope with slow changes in equipment status but also quickly suppress sudden disturbances from external conditions such as raw materials and the environment, enhancing the robustness and intelligence of the production process and reducing reliance on human experience. Attached Figure Description

[0062] The invention will now be further described with reference to the accompanying drawings.

[0063] Figure 1 This is a flowchart of the system of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Please see Figure 1 As shown, the present invention is a polishing powder particle size classification and screening control system, comprising:

[0066] The data acquisition and feature extraction module is used to acquire the raw data stream output by the granularity analysis device in real time, preprocess the raw data stream and fuse multi-scale features to generate a standardized granularity feature sequence.

[0067] The dynamic stability assessment module performs nonlinear dynamic analysis on the standardized granularity characteristic sequence, constructs a process dynamic stability index, and calculates the time series deterministic factor.

[0068] The process disturbance quantification module performs joint sensitivity analysis on the operating parameters during the production process, constructs a process disturbance characteristic spectrum, and calculates the influence value of external disturbances.

[0069] The multimodal decision optimization module integrates the temporal deterministic factor and the influence value of external disturbances into a system state feature vector, inputs the system state feature vector into a pre-trained process optimization model for multi-objective collaborative decision-making, and outputs a confidence score matrix for graded process parameters.

[0070] The intelligent control execution module executes an adaptive pre-control strategy based on a confidence score matrix.

[0071] In the data acquisition and feature extraction module, parallel optical diffraction and acoustic emission sensors synchronously acquire the multimodal raw signals output by the particle size analysis device. The optical diffraction sensor monitors in real time the changes in optical intensity caused by polishing powder particles passing through the detection area, forming an optical pulsation sequence; the acoustic emission sensor synchronously acquires the acoustic wave signals generated by particle collisions and friction, forming an acoustic vibration sequence. The two sensors use a unified clock signal for synchronous acquisition, ensuring that the optical pulsation sequence and acoustic vibration sequence are time-aligned, providing a time-aligned data foundation for subsequent cross-modal feature fusion.

[0072] Intrinsic mode decomposition (EMD) was performed on both optical pulsation sequences and acoustic vibration sequences. This process decomposes the non-stationary signal sequences into a series of intrinsic mode function (EMF) components, each representing an oscillation mode within a specific frequency range. For each EMF component obtained from the decomposition, its sample entropy and energy entropy values ​​were calculated. The sample entropy value quantifies the complexity of the signal component, while the energy entropy value characterizes the uniformity of the signal energy distribution in the frequency domain. These entropy calculations provide a quantitative basis for subsequent feature fusion.

[0073] Cross-modal fusion is performed on intrinsic mode function (EMF) components from different sensors operating in the same frequency band. This process is achieved by calculating mutual information entropy and phase synchronization index. The phase synchronization index is calculated by first performing Hilbert transforms on the EMF components of the optical and acoustic signals in the same frequency band to extract their instantaneous phase sequences; then, calculating the cosine function value of the difference between the two phase sequences; finally, performing a moving average of this cosine function value over a time window, and the absolute value of the average is the phase synchronization index characterizing the rhythmic coordination of the two signals. The mutual information entropy is calculated by separately calculating the amplitude probability distribution and the joint probability distribution of the two signals. Based on these probability distributions, the individual information entropy and the joint information entropy of the two signals are calculated. The mutual information entropy value, characterizing the degree of nonlinear dependence between the two signals, is obtained by subtracting the joint information entropy from the sum of the individual information entropies. These two indices together constitute a multi-scale fusion feature with time-frequency joint characteristics.

[0074] The multi-scale fused features are normalized using a minimum-maximum value method, mapping each feature value to the range of zero to one, thus eliminating the influence of different feature scales. The normalized features are then reorganized according to a time window sequence, with each time window containing a fixed number of feature vectors, ultimately forming a standardized granularity feature sequence. This sequence serves as input to subsequent modules, providing the system with a standardized, multi-scale feature representation. The entire process ensures the consistency, comparability, and processability of the feature data, providing a reliable data foundation for subsequent stability assessment and decision optimization.

[0075] In the dynamic stability assessment module, a delayed coordinate reconstruction method is used to reconstruct the phase space of the standardized granular feature sequence. This method maps a one-dimensional time series to a high-dimensional phase space by selecting appropriate delay times and embedding dimensions, thus reconstructing the dynamic characteristics of the system. The delay time is selected based on the time interval corresponding to the decay of the sequence's autocorrelation function to a specific proportion of its initial value. The embedding dimension is determined using the spurious nearest neighbor method to find the minimum dimension that can fully unfold the attractor structure. Through phase space reconstruction, a set of trajectory points reflecting the dynamic characteristics of the system is obtained, laying the foundation for subsequent nonlinear analysis.

[0076] Next, the recursion graph of the phase space trajectory point set is calculated. A recursion graph is a tool for visualizing the dynamic characteristics of a nonlinear system, representing the recursive relationships between points in the phase space trajectory in two dimensions. During the calculation, an appropriate distance threshold is set; when the distance between two trajectory points in the high-dimensional phase space is less than this threshold, the corresponding position on the recursion graph is marked as a recursive point. The distance threshold is typically taken as a specific proportion of the phase space diameter to ensure that the recursion graph contains sufficient dynamic information without being overly dense. The generation of the recursion graph provides an intuitive quantitative basis for extracting the dynamic features of the system.

[0077] Then, image processing techniques are used to extract diagonal and vertical structure features from the recursion graph. Diagonal structure features primarily reflect the determinism of the system, quantified by statistically analyzing the length distribution of consecutive diagonal segments in the recursion graph. Vertical structure features characterize the system's stationarity, evaluated by counting the number of consecutive recursive points in the vertical direction.

[0078] Based on the extracted diagonal and vertical structural features, the recursion rate and deterministic rate in the dynamic stability indices of the calculation process are quantified respectively. The recursion rate represents the frequency with which the system state reverts to a previous state, calculated as the proportion of recursive points to the total number of points in the recursion graph. The deterministic rate reflects the predictability of the system's dynamic behavior, measured by the proportion of long diagonal segments in the diagonal structural features. These two indices characterize the dynamic stability of the system from different perspectives: the recursion rate focuses on the repeatability of system states, while the deterministic rate focuses on the regularity of system behavior.

[0079] The recursion rate and determinism rate are weighted and fused to generate a comprehensive process dynamic stability index. The weighting coefficients are determined based on the contribution of each index to system stability, configured through historical data analysis or expert experience. The fusion process uses a linear weighting method, combining the two standardized indicators with specific weights to produce a comprehensive evaluation value within the range of zero to one. The closer this value is to one, the better the system's dynamic stability; the closer it is to zero, the worse the system's stability.

[0080] This module calculates the permutation entropy values ​​of standardized granularity feature sequences at different time scales, forming a multi-scale permutation entropy spectrum. Permutation entropy is a metric for measuring the complexity of a time series, calculated by analyzing the probability distribution of permutation patterns of sequence values. Appropriate embedding dimensions and time delays are selected during calculation, and permutation entropy values ​​are calculated separately for subsequences at different time scales. The selection of time scales covers multiple levels from short-term to long-term, thus forming a multi-scale permutation entropy spectrum that comprehensively reflects the dynamic characteristics of the system.

[0081] This study analyzes the fluctuation characteristics and decay rate of the multi-scale permutation entropy spectrum, extracting the entropy value fluctuation variance and the maximum Lyapunov exponent estimate. The entropy value fluctuation variance is obtained by calculating the variance of the permutation entropy values ​​at each scale, reflecting the stability of the system's complexity. The maximum Lyapunov exponent estimate is calculated by analyzing the trend of permutation entropy changes with scale, characterizing the system's sensitivity to initial conditions. These two parameters describe the dynamic behavior characteristics of the system from different perspectives, providing a quantitative basis for assessing the time series determinism.

[0082] Finally, the entropy fluctuation variance and the maximum Lyapunov exponent estimate are nonlinearly mapped and defuzzified to output a time-series deterministic factor ranging from zero to one. The nonlinear mapping employs a fuzzy logic-based inference method to establish a correspondence between input parameters and output determinism. The defuzzification process converts the fuzzy inference results into precise numerical outputs, and the weighted average of the output values ​​is calculated using the centroid method. The time-series deterministic factor comprehensively reflects the regularity and predictability of the time series; the closer the value is to one, the higher the degree of determinism of the time series, while the closer it is to zero, the stronger the randomness.

[0083] In the process disturbance quantification module, time series of multi-dimensional operating parameters, including feeder speed, classifier wheel current, system inlet negative pressure, and outlet temperature, are collected. Data acquisition employs synchronous sampling to ensure that each parameter's time series has a consistent timestamp and sampling interval. The feeder speed signal is acquired via an encoder mounted on the feeder motor, the classifier wheel current signal via a current transmitter, the system inlet negative pressure via a pressure sensor, and the outlet temperature via a thermocouple. All sensor signals undergo signal conditioning and analog-to-digital conversion to form a time-aligned multi-dimensional operating parameter time series, providing a data foundation for subsequent joint sensitivity analysis.

[0084] The first-order and total-order sensitivity indices of the impact of fluctuations in various operating parameters on the standardized granularity feature sequence are calculated. The first-order sensitivity index is calculated using the Sobol method based on variance decomposition. It measures the degree of influence of a single parameter's individual variation on granularity features by analyzing the proportion of its contribution to the output variance. The total-order sensitivity index measures the combined effect of individual parameter variations and their interactions with other parameters. Monte Carlo sampling is used to generate the parameter sample space, and the sensitivity index of each parameter is estimated through multiple sampling and model calculations. This step quantifies the importance of each operating parameter to granularity quality.

[0085] The sensitivity indices of each parameter are convolved with their instantaneous fluctuation amplitudes to generate a dynamic sensitivity matrix characterizing the strength of multi-parameter coupled perturbations. The convolution operation essentially performs a weighted moving average of the sensitivity indices and parameter fluctuations in the time domain, where the sensitivity indices serve as the weighting function and the parameter fluctuation amplitudes as the input function. The size of the convolution kernel is determined based on the dynamic characteristics of the process, typically taking 2 to 3 times the main time constant of the process. Through convolution, the static sensitivity indices are transformed into dynamic sensitivity indicators that vary over time. This indicator considers both the importance of the parameters and their instantaneous fluctuations, thus better reflecting the real-time impact of multi-parameter coupled perturbations.

[0086] Principal component analysis (PCA) is used to reduce the dimensionality of the dynamic sensitivity matrix. The first three principal components are extracted and arranged in time series to construct a visualized process disturbance feature spectrum. PCA finds a new orthogonal basis that can explain the data variance to the greatest extent by performing eigenvalue decomposition on the covariance matrix of the dynamic sensitivity matrix. The first three principal components can typically explain more than 80% of the variance of the original data, and each principal component represents a specific disturbance mode. Arranging these three principal components in time order forms a three-dimensional process disturbance feature spectrum, which can intuitively show the main modes of process disturbance and their evolution over time.

[0087] Principal component time series representing different perturbation modes are extracted from the process perturbation characteristic spectrum. Each principal component time series represents a specific perturbation mode; the first principal component typically represents the most influential primary perturbation mode, while the second and third principal components represent secondary perturbation modes. The extraction process involves projecting the original dynamic sensitivity matrix onto the directions of each principal component, obtaining the time series corresponding to each principal component. These time series reflect the intensity changes of various perturbation modes over time, providing input data for subsequent perturbation impact analysis.

[0088] The complex wavelet coherence coefficients of the time series of each principal component and the normalized granularity feature series in the time-frequency domain are calculated. Complex wavelet coherence analysis uses the Morley wavelet as the mother wavelet, and decomposes the two time series into the time-frequency domain through continuous wavelet transform, calculating their coherence at each time-frequency point. The magnitude of the complex wavelet coherence coefficient represents the degree of correlation between the two signals at a specific time-frequency point, and the phase represents the phase relationship between the two signals. Appropriate wavelet scales and time window lengths are selected during the calculation to ensure sufficient resolution in both the time and frequency domains. This analysis can reveal the correlation characteristics between process disturbances and granularity quality across different time scales and frequency ranges.

[0089] By integrating the energy of the complex wavelet coherence coefficients in the main frequency bands, the frequency domain energy influence weights of each perturbation mode on granularity quality are obtained. The integration frequency bands are determined based on process characteristics and typically include bands reflecting fast perturbations and bands reflecting slow perturbations. For the complex wavelet coherence coefficients corresponding to each principal component, their energy integral values ​​in each frequency band are calculated. These integral values ​​represent the intensity of the perturbation mode's influence on granularity quality in different frequency ranges, and are transformed into frequency domain energy influence weights in the range of zero to one through normalization. The larger the weight value, the more significant the impact of the perturbation on granularity quality within that frequency band.

[0090] The influence weights of frequency domain energy are multiplied by the variance contribution rates of the corresponding principal components and then summed. The final result is quantified as an external disturbance influence value in the range of zero to one. The variance contribution rate represents the proportion of the original data variance that each principal component can explain, obtained from the principal component analysis results. The multiplication operation combines the frequency domain influence with the variance contribution, while the summation operation integrates the overall influence of all disturbance modes. The calculation result is normalized to the range of zero to one using a maximum-minimum normalization method. The closer the value is to one, the greater the impact of the external disturbance on the system; the closer it is to zero, the smaller the impact. The external disturbance influence value provides an important quantitative basis for subsequent control decisions.

[0091] In the multimodal decision optimization module, the fusion process of the system state feature vector first uses a time-series determinism factor to weight and reconstruct the standardized granularity feature sequence. The time-series determinism factor is a value between zero and one, representing the degree of determinism of the time series. The weighted reconstruction process multiplies each data point of the standardized granularity feature sequence by the time-series determinism factor, strengthening data points with high determinism and weakening data points with low determinism. Through this weighting method, a first feature vector representing the stability of the internal process is generated. This vector highlights the characteristic patterns during periods of process stability, providing reliable internal state information for subsequent decisions.

[0092] Subsequently, the principal components of the process disturbance feature spectrum are selectively enhanced using the external disturbance influence value. The external disturbance influence value, also a value in the range of zero to one, reflects the overall impact of the external disturbance on the system. The selective enhancement process weights each principal component of the process disturbance feature spectrum with the external disturbance influence value, assigning higher weights to disturbance patterns with a greater impact on the system. This weighting method generates a second eigenvector characterizing the sensitivity to external disturbances. This vector highlights the system's response characteristics to external disturbances, providing information about the external environment for decision-making.

[0093] Tensor concatenation of the first and second feature vectors generates a system state feature vector containing multi-dimensional state information. Tensor concatenation is performed along the feature dimensions, linking the feature dimensions of the two vectors end-to-end to form a higher-dimensional feature vector. The concatenated system state feature vector simultaneously contains internal process stability information and external disturbance sensitivity information, providing a comprehensive system state description for the process optimization model. The dimension of the feature vector depends on the length of the standardized granularity feature sequence and the number of principal components in the process disturbance feature spectrum.

[0094] The system state feature vector is input into a pre-trained deep reinforcement learning network for multi-objective collaborative decision-making. This network consists of a policy network and a value network connected in parallel. The two networks share the underlying feature extraction layer, but perform different functions at the output layer. The policy network uses a fully connected layer structure, receiving the system state feature vector as input. After nonlinear transformation through multiple hidden layers, it outputs three sets of process parameter adjustment strategies for the stage wheel speed, induced draft fan frequency, and feeder speed. Each strategy includes the direction and magnitude of parameter adjustment and is accompanied by a confidence score indicating the reliability of the strategy.

[0095] The value network also employs a fully connected layer structure, but shares some underlying features with the policy network. The value network receives the system state feature vector and the parameter adjustment strategies output by the policy network as input, predicting the long-term optimization effect of each strategy. The prediction process considers multiple optimization objectives, including product quality stability, energy efficiency, and equipment lifespan, ultimately providing a stability score within the range of zero to one. This score reflects the long-term effect of the strategy execution; a higher score indicates better long-term stability of the strategy.

[0096] A matrix multiplication operation is performed between the strategy confidence scores and stability scores to generate a three-dimensional confidence score matrix. The matrix multiplication is performed between two two-dimensional matrices: one containing the confidence scores of each strategy output by the policy network, and the other containing the stability scores of each strategy output by the value network. The result of the multiplication is a three-dimensional matrix where each element represents the overall confidence score of a specific process parameter under a specific strategy. This matrix provides a quantitative basis for subsequent control execution decisions; strategies with higher scores are more likely to be adopted.

[0097] Pre-trained deep reinforcement learning networks are obtained through a specific training process. First, a multimodal training sample set is constructed, containing system state feature vectors, process parameter adjustment actions, and corresponding long-term stability reward signals. The long-term stability reward signal is a comprehensive indicator, encompassing multiple dimensions such as product quality indicators, energy consumption indicators, and equipment operating status indicators, which are weighted and combined to form a single reward value. Training samples are derived from historical production data or simulation data, covering various operating conditions and disturbances.

[0098] During training, the policy network generates parameters to adjust the policy, seeking the optimal policy through a balance between exploration and exploitation. The value network evaluates the long-term returns of the policy, updating the value function through temporal difference learning. The two networks update their parameters using an alternating optimization approach: first, the policy network parameters are fixed while the value network parameters are updated, then the value network parameters are fixed while the policy network parameters are updated. This alternating optimization method helps improve training stability and convergence speed.

[0099] Training termination conditions include the mean squared error (MSE) of the value network's policy evaluation being below a preset threshold and the policy network's output policy converging. The MSE threshold is typically set to 0.001, and policy convergence is determined by the magnitude of policy changes being less than a specific threshold across multiple consecutive iterations. When these conditions are met, the training process terminates, and the final network parameters are saved for online decision-making. The entire training process ensures that the deep learning network can adapt to various operating conditions and make accurate multi-objective optimization decisions.

[0100] In the intelligent control execution module, the confidence score matrix is ​​parsed to extract three sets of process parameter adjustment schemes and their corresponding confidence scores for the stager wheel speed, induced draft fan frequency, and feeder speed. The confidence score matrix is ​​a three-dimensional data structure containing the score values ​​of different process parameters under different operating modes. The parsing process traverses each dimension of the matrix to identify the highest score value and its index for each process parameter. For each process parameter, the top three adjustment schemes with the highest scores and their corresponding confidence scores are extracted to form a candidate strategy set. These adjustment schemes include specific parameter settings or adjustment ranges, and the confidence scores reflect the reliability of the scheme.

[0101] The process parameter adjustment schemes are sorted according to their confidence scores, with the scheme having the highest confidence score being executed first. The sorting process uses a descending order, arranging the confidence scores from highest to lowest. When multiple schemes have the same confidence score, the scheme that is simple to implement and has minimal impact on system stability is selected, taking into account both implementation difficulty and impact on system stability. After sorting, an ordered strategy execution sequence is generated, ensuring that the system prioritizes the most reliable adjustment scheme. This sorting process ensures that the optimal adjustment strategy is always selected from multiple options.

[0102] Based on the confidence level of the selected adjustment scheme, the response rate and adjustment range of the control system are automatically set. For schemes with a confidence level higher than 0.8, a direct and rapid adjustment strategy is adopted, with the control system performing parameter adjustments with a relatively fast response rate and a large adjustment range. For schemes with a confidence level between 0.5 and 0.8, a medium-speed adjustment strategy is adopted, with the control system's response rate and adjustment range being moderate. For schemes with a confidence level lower than 0.5, a gradual trial-and-error adjustment strategy is adopted, with the control system implementing adjustments step by step with a slower response rate and a smaller adjustment range, and the effect is evaluated after each adjustment. This tiered adjustment strategy ensures the execution efficiency of high-confidence schemes while avoiding system disturbances that may be caused by low-confidence schemes.

[0103] After parameter adjustments, the feedback signal from the granularity analysis device is monitored in real time. By comparing the deviation between the expected and actual results, the weighting parameters of the confidence score matrix are corrected. The monitoring process continues for a complete production cycle, collecting adjusted granularity distribution data and quality indicators. The deviation is calculated using the root mean square error method, comparing the difference between the actual and expected granularity distribution. Based on the magnitude of the deviation, the corresponding weight parameters in the confidence score matrix are dynamically adjusted: when the deviation is less than 5%, the weight parameter corresponding to the strategy is increased; when the deviation is between 5% and 10%, the weight parameter remains unchanged; when the deviation is greater than 10%, the weight parameter corresponding to the strategy is decreased.

[0104] The entire process forms a complete control closed loop, continuously monitoring and providing feedback corrections to optimize the execution effect of the control strategy. The system can adaptively adjust control parameters based on actual operating conditions, improving the accuracy and reliability of control. At the end of each control cycle, the system updates the weight parameters of the confidence score matrix, providing a more accurate basis for decision-making in the next cycle. This adaptive mechanism enables the system to gradually improve control performance and adapt to changes in production processes and fluctuations in external conditions.

[0105] The working principle of this invention is as follows: Multimodal signals are simultaneously acquired using optical diffraction spectroscopy sensors and acoustic emission sensors. A standardized granularity feature sequence is generated through intrinsic mode decomposition and cross-modal fusion. Nonlinear dynamics methods are used to analyze process stability and calculate temporal determinism factors. The impact of external disturbances is quantified through joint sensitivity analysis and time-frequency domain processing. A deep reinforcement learning network is used for multi-objective collaborative decision-making to generate a confidence score matrix for process parameter adjustment. Finally, an adaptive pre-control strategy is executed based on the confidence level, forming a complete intelligent control closed loop. This system achieves multi-dimensional state perception, intelligent decision-making, and precise control of the polishing powder particle size classification process.

[0106] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A polishing powder particle size fractionation screening control system, characterized by, Comprise: A data acquisition and feature extraction module for real-time acquisition of raw data streams output by a particle size analysis device, preprocessing of the raw data streams, and multi-scale feature fusion to generate a standardized particle size feature sequence; A dynamic stability evaluation module for nonlinear dynamics analysis of the standardized particle size feature sequence, construction of process dynamic stability indicators, and calculation of time series determinacy factors; A process disturbance quantification module for joint sensitivity analysis of process parameters in the production process, construction of process disturbance feature spectrum, and calculation of external disturbance influence values; A multi-modal decision optimization module for fusion of the time series determinacy factors and the external disturbance influence values into a system state feature vector, input of the system state feature vector into a pre-trained process optimization model for multi-objective collaborative decision-making, and output of a confidence score matrix for hierarchical process parameters; An intelligent control execution module for execution of an adaptive pre-control strategy based on the confidence score matrix.

2. A polishing powder particle size classification screening control system according to claim 1, wherein, The generation process of the standardized particle size feature sequence specifically comprises: Synchronous acquisition of multi-modal raw signals output by the particle size analysis device through parallel arrangement of optical diffraction spectrum sensors and acoustic emission sensors to obtain time-aligned optical pulsation sequences and acoustic vibration sequences; Eigenmode decomposition of the optical pulsation sequences and the acoustic vibration sequences, respectively, extraction of eigenmode function components, and calculation of sample entropy values and energy entropy values of each component; Cross-modal fusion of eigenmode function components of the same frequency band from different sensors, calculation of mutual information entropy and phase synchronization index, and generation of multi-scale fusion features with time-frequency joint characteristics; Maximum and minimum value normalization of the multi-scale fusion features, reorganization according to time window sequences, and final output of the standardized particle size feature sequence.

3. A polishing powder particle size classification screening control system according to claim 2, wherein, The acquisition process of the mutual information entropy and the phase synchronization index comprises: The calculation of the phase synchronization index is realized by the following steps: Hilbert transform of the optical and acoustic eigenmode function components of the same frequency band, extraction of their instantaneous phase sequences, calculation of the cosine function value of the difference between the two phase sequences, and sliding average operation of the cosine function value within a time window. The absolute value of the obtained average value is the phase synchronization index representing the rhythm coordination of the two signals. The calculation of the mutual information entropy is realized by the following steps: calculation of the amplitude probability distribution and the joint probability distribution of the two signals, respectively; then based on these probability distributions, the individual information entropy and the joint information entropy of the two signals are calculated, respectively; by subtracting the joint information entropy from the sum of the individual information entropies, the mutual information entropy value representing the nonlinear dependence degree of the two signals is obtained.

4. The polishing powder particle size classification screening control system of claim 1, wherein, The nonlinear dynamics analysis specifically comprises: Reconstruction of the phase space of the standardized particle size feature sequence using the delay coordinate reconstruction method to map the one-dimensional time series into a trajectory point set in the high-dimensional phase space; Calculation of the recurrence plot of the phase space trajectory point set, and extraction of the diagonal line structure feature and the vertical structure feature in the recurrence plot through image processing technology; Based on the diagonal structure feature and the vertical structure feature, a recurrence rate and a determinacy rate in a process dynamic stability index are respectively quantitatively calculated; The recurrence rate and the determinacy rate are weighted and fused to generate a comprehensive process dynamic stability index.

5. The polishing powder particle size classification screening control system of claim 1, wherein, The acquisition process of the time sequence determinacy factor is as follows: The permutation entropy values of the standardized granularity feature sequence under different time scales are calculated to form a multi-scale permutation entropy spectrum; The fluctuation characteristics and decay rates of the multi-scale permutation entropy spectrum are analyzed to extract an entropy value fluctuation variance and a maximum Lyapunov exponent estimation value; The entropy value fluctuation variance and the maximum Lyapunov exponent estimation value are nonlinearly mapped and de-fuzzified to output a time sequence determinacy factor in the range of zero to one.

6. The polishing powder particle size classification screening control system of claim 1, wherein, The joint sensitivity analysis of the working condition parameters in the production process specifically includes: A multi-dimensional working condition parameter time sequence including the feeder speed, the classifier wheel current, the system inlet negative pressure and the outlet temperature is collected; First-order and total-order sensitivity indexes of the influence of the fluctuation of each working condition parameter on the standardized granularity feature sequence are calculated; The sensitivity indexes of each parameter are convoluted with the instantaneous fluctuation amplitudes to generate a dynamic sensitivity matrix representing the coupling disturbance intensity of multiple parameters; Principal component analysis dimension reduction is performed on the dynamic sensitivity matrix, and the first three principal component components are extracted and arranged in time sequence to construct a visual process disturbance feature spectrum.

7. The polishing powder particle size classification screening control system of claim 1, wherein, The acquisition process of the external disturbance influence value is as follows: The principal component component time sequences representing different disturbance modes are extracted from the process disturbance feature spectrum; Complex wavelet coherence coefficients of each principal component component time sequence and the standardized granularity feature sequence in the time-frequency domain are calculated; The frequency band energy influence weight of each disturbance mode on the granularity quality is obtained by integrating the energy of the complex wavelet coherence coefficients in the main frequency band; The frequency band energy influence weight is multiplied by the variance contribution rate of the corresponding principal component and accumulated, and the calculation result is finally quantized to an external disturbance influence value in the range of zero to one.

8. The polishing powder particle size fractionating screening control system of claim 1, wherein, The fusion process of the system state feature vector is as follows: The standardized granularity feature sequence is weighted and reconstructed using the time sequence determinacy factor to generate a first feature vector representing the internal process stability; The principal components of the process disturbance feature spectrum are selectively strengthened using the external disturbance influence value to generate a second feature vector representing the external disturbance sensitivity; The first feature vector and the second feature vector are tensor spliced to generate a system state feature vector containing multi-dimensional state information.

9. The polishing powder particle size fractionating screening control system of claim 1, wherein, The system state feature vector is input into a pre-trained process optimization model for multi-objective collaborative decision-making, specifically including: The system state feature vector is input into a pre-trained deep reinforcement learning network, and the deep reinforcement learning network is composed of a policy network and a value network in parallel; The policy network outputs three sets of process parameter adjustment strategies for the classifier wheel speed, the air fan frequency and the feeder speed and their corresponding confidence levels; The value network predicts the long-term optimization effect of each set of strategies and gives a stability score; The strategy confidence level and the stability score are matrix multiplied to generate a three-dimensional confidence score matrix; The pre-trained deep reinforcement learning network is trained by the following steps: A multi-modal training sample set is constructed, which contains system state feature vectors, process parameter adjustment actions and corresponding long-term stability reward signals; The policy network is responsible for generating parameter adjustment strategies, and the value network is responsible for evaluating the long-term benefits of the strategies; The network parameters of the policy network and the value network are alternately optimized; When the mean square error of the value network in evaluating the strategy is lower than a preset threshold and the output strategy of the policy network converges, the network training is completed and the final model parameters are saved.

10. The polishing grit sizing and screening control system of claim 1, wherein, The adaptive pre-control strategy is executed, specifically including: The confidence score matrix is analyzed to extract three sets of process parameter adjustment schemes for the classification rotation speed, the induced draft fan frequency and the feeder speed and their corresponding confidence scores; The process parameter adjustment schemes are sorted according to the order of the confidence scores, and the parameter adjustment scheme with the highest confidence score is executed first; According to the confidence level of the selected adjustment scheme, the response rate and adjustment amplitude of the control system are automatically set, the scheme with high confidence adopts a direct and fast adjustment strategy, and the scheme with low confidence adopts a gradual and exploratory adjustment strategy; After the parameter adjustment is executed, the feedback signal of the particle size analysis device is monitored in real time, and the weight distribution parameters of the confidence score matrix are corrected by comparing the deviation between the expected effect and the actual effect.

Images