Drilling mud solid phase control method based on size grading

Abstract

The invention relates to the technical field of petroleum drilling solid control, in particular to a drilling mud solid phase control method based on particle size classification, which comprises the following steps: establishing multi-mode characteristic characterization including energy fluctuation and coherence modes in a solid phase separation process; constructing a sub-feature topological association network in the energy fluctuation mode, and extracting key nodes to form a first feature subset; calculating a time-frequency characteristic correlation matrix by a coherence mode, and performing singular value decomposition to obtain a main steering quantity to form a second characteristic subset; the two subsets are orthogonally fused to obtain a difference description vector, and a modulation vector is generated through carrier modulation of the real-time vibration frequency of the vibrating screen. And based on a historical trajectory training behavior prediction model, inputting a modulation vector to output a prediction trajectory deviation value of a key granularity section, and iteratively correcting a process parameter set point. The method accurately captures key characteristics of energy fluctuation, enhances components related to working conditions of the vibrating screen, improves trajectory prediction accuracy, and realizes solid-phase hierarchical control precision optimization.

Description

Technical Field

[0001] This invention relates to the field of solids control technology in oil drilling, and more particularly to a method for controlling the solid phase of drilling mud based on particle size classification. Background Technology

[0002] Drilling mud solids control requires precise differentiation of the behavior differences of solids at different particle sizes to optimize process parameters. Existing technologies mostly employ single-particle-size monitoring or simple statistical feature extraction, combined with fixed-parameter models to predict solids trajectories and correct control processes. Existing solutions have shortcomings: multimodal features are not deeply correlated; key feature extraction relies on empirical thresholds, ignoring dynamic interactions between features; feature representation does not incorporate real-time equipment conditions, such as the impact of vibrating screen frequency changes on solids separation, which is not actively adapted, resulting in weak correlation between features and operating conditions, large prediction bias, and limited control accuracy.

[0003] Existing energy fluctuation mode feature processing focuses only on the superposition of single components, failing to analyze the topological relationships between sub-features, and key driving features are easily masked by noise; coherent mode feature fusion does not consider matching with the real-time vibration state of the equipment, and implicit operating condition information is not activated. This invention aims to address: how to accurately capture key feature nodes that play a decisive role in solid-phase separation in energy fluctuation modes through sub-feature topological relationship networks; and how to utilize the real-time vibration frequency modulation feature vector of the vibrating screen to enhance its strong correlation with the equipment's operating state, thereby improving the targeting of subsequent predictions. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a drilling mud solid phase control method based on particle size classification.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a drilling mud solids control method based on particle size classification, comprising: A multimodal feature characterization of the solid-phase separation process across multiple particle size monitoring dimensions is established, wherein the multimodal feature characterization includes at least energy fluctuation mode and coherence mode; For the energy fluctuation mode, a topological correlation network among its internal sub-features is constructed, and feature nodes with key connection attributes in the topological correlation network are extracted to form a first feature subset; for the coherence mode, the correlation matrix of its internal time-frequency features at different granularity monitoring dimensions is calculated, and the dominant feature vector in the correlation matrix is ​​identified by singular value decomposition to form a second feature subset. The first feature subset and the second feature subset are subjected to orthogonality test and feature fusion to obtain a difference description vector describing the differences in the behavior of solid phases with different particle sizes; Using the real-time vibration frequency of the vibrating screen as the modulation signal, the difference description vector is carrier modulated to enhance the feature components that are strongly correlated with the working state of the vibrating screen, thereby generating a modulated description vector. Based on the historical trajectory data of solid particles in the separation equipment, a behavior prediction model is trained. The modulated description vector is input into the behavior prediction model, and a series of predicted trajectory deviation values ​​for key particle size segments are output. Based on the predicted trajectory deviation value, the process parameter set points in the current drilling mud solid phase control process are iteratively corrected.

[0006] As a further aspect of the present invention, the establishment of multimodal characteristic characterization of the solid-phase separation process across multiple particle size monitoring dimensions includes: The energy fluctuation mode is obtained by decoupling the driving current waveform of the vibrating screen from the mud level fluctuation. The coherent modes are generated by envelope detection and phase locking of acoustic signals in a specific frequency band; The three-phase current signal is obtained from the drive motor controller of the vibrating screen. The three-phase current signal is subjected to Clark transformation to obtain the two-phase current signal in the stationary coordinate system. The instantaneous amplitude sequence of the two-phase current signal is calculated and this sequence is defined as the baseband component of the energy fluctuation mode. The fluctuation signal of the mud level height changing with time is obtained from the mud level sensor, and the fluctuation signal is subjected to Hilbert transform to obtain its analytical signal. The instantaneous amplitude of the analytical signal is calculated, and the instantaneous amplitude sequence is cross-correlated with the baseband component of the energy fluctuation mode. The normalized cross-correlation coefficient is calculated, and the normalized cross-correlation coefficient is used as a sub-feature of the energy fluctuation mode, which together with the baseband component constitutes the energy fluctuation mode. Raw acoustic signals are acquired from an acoustic sensor array arranged in the vibrating screen box and mud pit. The raw acoustic signals are then bandpass filtered to extract specific frequency band signals related to solid particle collisions and fluid vortices. An envelope detection algorithm is applied to the specific frequency band signal to extract the signal envelope, and phase-locked loop technology is used to track the phase of the dominant frequency component in the specific frequency band signal to generate two demodulated signals, one in phase and one quadrature. The coherence coefficient between the two demodulated signals is calculated. This coefficient characterizes the phase locking degree between the signals and is defined as the coherent mode.

[0007] As a further aspect of the present invention, the construction of a topological association network among its internal sub-features, and the extraction of feature nodes with key connection attributes from the topological association network to form a first feature subset, includes: Multiple sub-features are extracted from the energy fluctuation mode, including the variance, skewness, peak factor, and normalized cross-correlation coefficient of the baseband component with the mud level fluctuation signal. Calculate the Spearman rank correlation coefficient between any two of the sub-features, treat all sub-features as nodes, and use the absolute value of the Spearman rank correlation coefficient as the weight of the edge between the nodes to construct a fully connected weighted undirected graph. This weighted undirected graph is the topological network of energy fluctuation modes. In the aforementioned topological network, the betweenness centrality of each node is calculated, which is the number of times the node appears on the shortest path in the network, and the eigenvector centrality of each node is calculated, which is the centrality of a node as a function of the centrality of its neighboring nodes. Multiply the betweenness centrality score of each node by the eigenvector centrality score to obtain the overall centrality score of each node; A centrality score threshold is set, nodes with a comprehensive centrality score higher than the centrality score threshold are selected, and the sub-features of the energy fluctuation mode corresponding to the node are extracted and set to form a first feature subset.

[0008] As a further aspect of the present invention, the step of calculating the correlation matrix of its internal time-frequency features at different granularity monitoring dimensions, and identifying the dominant feature vector in the correlation matrix through singular value decomposition to form a second feature subset includes: Multiple time-frequency features are extracted from the coherent modes, including the energy of the signal in a specific frequency band, the standard deviation of the instantaneous frequency, and the entropy value of the signal's marginal spectrum; The extracted time-frequency features are grouped and arranged according to different granularity monitoring dimensions to form a feature matrix, where each row represents a granularity monitoring dimension and each column represents a time-frequency feature. Calculate the covariance matrix of the feature matrix. This covariance matrix is ​​the correlation matrix that describes the correlation of time-frequency features under different granularity monitoring dimensions. Perform singular value decomposition on the correlation matrix to decompose it into the product of three matrices: a left singular vector matrix, a diagonal matrix composed of singular values, and the transpose of a right singular vector matrix. Identify the largest singular value from the diagonal matrix and extract the left singular vector corresponding to the largest singular value. This left singular vector is the dominant feature vector of the correlation matrix. From the dominant feature vector, elements whose absolute values ​​are greater than a preset contribution threshold are selected, and the time-frequency features corresponding to these elements constitute the second feature subset.

[0009] As a further aspect of the present invention, the orthogonality test and feature fusion of the first feature subset and the second feature subset include: Calculate the Pearson correlation coefficient between each feature in the first feature subset and each feature in the second feature subset to form a correlation coefficient matrix; For the correlation coefficient matrix, the proportion of coefficients whose absolute value is less than a preset orthogonality threshold is counted. If the proportion exceeds the preset proportion threshold, it is determined that the first feature subset and the second feature subset satisfy the orthogonality condition as a whole. When the orthogonality condition is met, all features of the first feature subset are directly concatenated with all features of the second feature subset to form an amplified feature vector. Principal component analysis is performed on the amplified feature vector, and the first few principal components whose cumulative contribution rate exceeds the preset contribution rate threshold are retained to obtain a dimension-reduced fused feature vector. This fused feature vector is the difference description vector describing the differences in the behavior of solid phases of different particle sizes.

[0010] As a further aspect of the present invention, the step of using the real-time vibration frequency of the vibrating screen as a modulation signal to perform carrier modulation on the difference description vector includes: The rotational speed signal of the main shaft of the vibrating screen is acquired in real time, and the real-time vibration frequency signal of the vibrating screen is obtained through the conversion relationship between rotational speed and frequency. The real-time vibration frequency signal is normalized so that its amplitude range is between zero and one, thereby generating a frequency modulation coefficient. Feature components related to vibration energy transfer are identified from the difference description vector, and the identification of the feature components is determined based on the magnitude of the covariance between the feature components and the vibration frequency signal in historical data; The frequency modulation coefficient is multiplied one by one with the identified vibration energy-related characteristic components to achieve amplitude modulation of the characteristic components; The modulated vibration energy-related feature components are recombined with the other unmodulated feature components in the difference description vector to form a new vector, which is the modulated description vector, in which the features related to the state of the vibrating screen are selectively enhanced.

[0011] As a further aspect of the present invention, the training behavior prediction model inputs the modulated description vector into the behavior prediction model and outputs a series of predicted trajectory deviation values ​​for key granularity segments, including: The behavior prediction model is used to simulate the expected separation trajectory of solid particles of different sizes under the current process parameters; Using high-speed camera and image analysis technology, the actual motion trajectory data of solid particles in the separation equipment with representative particle size range under different process parameter settings are obtained as training samples; A behavior prediction model incorporating a long short-term memory network and an attention mechanism is constructed. The input of the behavior prediction model is a combination of the current process parameter setpoint and the modulated description vector, and the output is the predicted sequence of key point coordinates of the particle motion trajectory. The behavior prediction model is trained using the training samples, and the model parameters are optimized by minimizing the root mean square error between the predicted trajectory and the actual trajectory. After training is completed, the current process parameter setpoint and the modulated description vector are input into the behavior prediction model. The model will output the corresponding predicted particle motion trajectory for each of the selected key granularity segments. The predicted trajectory output by the model is compared point by point with the ideal separation trajectory of the corresponding granularity segment stored in the knowledge base. The average Euclidean distance between each predicted trajectory and the corresponding ideal trajectory is calculated. This average Euclidean distance is the predicted trajectory deviation value of the corresponding granularity segment. The predicted trajectory deviation values ​​of all key granularity segments constitute a deviation value set.

[0012] As a further aspect of the present invention, the iterative correction of the process parameter setpoints in the current drilling mud solids control process includes: The process parameter setting points include the centrifuge separation factor setting value, the desliming hydrocyclone feed pressure setting value, and the dilution water addition rate setting value, thereby realizing real-time optimized control of the solid phase separation process. A response surface surrogate model is established between process parameters and predicted trajectory deviation values. The surrogate model is constructed using a radial basis function neural network. Its inputs are the separation factor of the centrifuge, the feed pressure of the hydrocyclone of the sludge dewatering unit, and the dilution water addition rate. Its output is the weighted sum of each deviation value in the deviation value set. With the goal of minimizing the weighted sum, within the safe operating range of the process parameters, the response surface surrogate model is optimized and solved using a sequential quadratic programming algorithm to obtain a new set of optimized process parameter setpoints. The current process parameter setpoint is adjusted slightly towards the process parameter optimization setpoint, and the adjustment step size is controlled by a preset learning rate parameter. The solid-state control system is run for a predetermined time window under the adjusted new process parameters, after which data is reacquired and the entire process from establishing multimodal feature characterization to calculating new predicted trajectory deviation values ​​is repeated. Compare the weighted sum of the predicted trajectory deviation values ​​calculated in the current iteration cycle with the value in the previous iteration cycle. If the decrease in the weighted sum of the deviation values ​​is less than the preset convergence threshold, the process parameter optimization process is determined to have converged, and the current setpoint is maintained; otherwise, a new round of iterative optimization is started from the current setpoint.

[0013] As a further aspect of the present invention, the step of using a sequential quadratic programming algorithm to optimize and solve the response surface surrogate model to obtain a new set of optimized process parameter setpoints includes: With the objective function of minimizing the weighted sum of predicted trajectory deviations, and defining the safe operating range of process parameters as a nonlinear constraint, an initial nonlinear programming problem for process parameter optimization is established. At the current process parameter setpoint, a quadratic programming subproblem is constructed by linearly approximating the objective function and the nonlinear constraints. Solving the quadratic programming subproblem yields the search direction and optimal step size for the process parameters in the current iteration step. Along the search direction, the process parameters are updated according to the optimal step size to obtain a set of candidate process parameter optimization setpoints; Calculate whether the updated process parameters meet the safe operating range constraints. If not, adjust the search direction and step size using the active set method and recalculate until the constraints are met. The candidate process parameter optimization setpoint that satisfies the constraints is taken as the output of the current iteration, and this result is the new process parameter optimization setpoint.

[0014] As a further aspect of the present invention, the process of iteratively correcting the process parameter setpoint also includes a mechanism for recording and rolling back the process parameter optimization path: After each iteration of adjusting the process parameter setpoint, the setpoint before adjustment, the setpoint after adjustment, and the corresponding predicted trajectory deviation value are weighted and recorded in an optimization history stack. If, after several consecutive iterations, the weighted average of the predicted trajectory deviation values ​​not only fails to decrease but continues to rise beyond the preset threshold for the number of deteriorations, then the rollback mechanism is triggered. The rollback mechanism finds the record with the smallest weighted sum of predicted trajectory deviation values ​​in the optimization history stack and restores the process parameter setpoint corresponding to the record to the current operating setpoint of the system. After the rollback operation is executed, the system will reduce the optimization step size of the process parameters and restart the iterative optimization process.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: A sub-feature topological association network is constructed for energy fluctuation modes, and key nodes are extracted to form the first feature subset. Sub-features within the energy fluctuation modes are used as network nodes, and a topological network is constructed according to dynamic association strength. Key nodes are selected based on connectivity properties such as degree centrality and betweenness centrality. This scheme can identify features that play a dominant role in the essence of solid-phase separation energy fluctuations, eliminate interference from non-critical fluctuations, and ensure that the first feature subset accurately reflects the core mechanisms of energy transfer and dissipation. This provides high-value input for differential description and avoids model overfitting caused by feature redundancy in conventional techniques.

[0016] The difference description vector is carrier-modulated using the real-time vibration frequency of the vibrating screen as the modulation signal. The instantaneous frequency of the vibrating screen is obtained as the carrier parameter, and the amplitude of the difference description vector is used as the envelope and multiplied by the sinusoidal carrier to enhance the feature components strongly correlated with the working state of the vibrating screen. This scheme enables the modulated vector to dynamically adapt to the vibration characteristics of the screen surface, highlighting the separation behavior of solid phases of specific particle sizes under vibration, suppressing irrelevant frequency noise, and improving the sensitivity of the behavior prediction model to capturing trajectory deviations. This differs from the static representation in conventional techniques where features are disconnected from the operating conditions. Attached Figure Description

[0017] Figure 1 This is a flowchart of the drilling mud solid phase control method based on particle size classification according to the present invention; Figure 2 A flowchart for constructing multimodal feature representations; Figure 3 The flowchart for extracting the second feature subset. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] See Figure 1A multimodal feature representation of the solid-phase separation process across multiple particle size monitoring dimensions is established. This multimodal feature representation includes at least energy fluctuation modes and coherence modes. For the energy fluctuation mode, a topological correlation network among its internal sub-features is constructed, and feature nodes with key connectivity attributes are extracted from this network to form a first feature subset. For the coherence mode, the correlation matrix of its internal time-frequency features across different particle size monitoring dimensions is calculated, and the dominant feature vector in the correlation matrix is ​​identified through singular value decomposition to form a second feature subset. The first and second feature subsets are subjected to orthogonality testing and feature fusion to obtain a difference description vector describing the differences in solid-phase behavior at different particle sizes. Using the real-time vibration frequency of the vibrating screen as the modulation signal, the difference description vector is carrier-modulated to enhance the feature components strongly correlated with the working state of the vibrating screen, generating a modulated description vector. Based on historical trajectory data of solid particles within the separation equipment, a behavior prediction model is trained. The modulated description vector is input into the behavior prediction model, which outputs a series of predicted trajectory deviation values ​​for key particle size segments. Based on the predicted trajectory deviation value, the process parameter set points in the current drilling mud solid phase control process are iteratively corrected.

[0021] See Figure 2 In one embodiment of the present invention, the specific construction of the energy fluctuation mode during the establishment of the multimodal feature characterization involves the joint processing of the vibrating screen drive current and the mud level signal. In a specific implementation, the three-phase current signals of phases A, B, and C are acquired in real time from the three-phase frequency converter controller of the vibrating screen drive motor. These signals are recorded at a sampling frequency of 1000 times per second. A Clarke transform is performed on the acquired three-phase current signals to convert them to a stationary α-β coordinate system, obtaining the α-axis current signal and β-axis current signal in the stationary coordinate system. The instantaneous amplitude sequence of the α-axis current signal and the β-axis current signal is calculated using the square root of the sum of squares method. This instantaneous amplitude sequence is defined as the baseband component of the energy fluctuation mode. Another sub-feature of the energy fluctuation mode originates from the mud level information. Fluctuation signals of the mud level height changing over time are acquired from radar level gauges installed on the mud tank. A Hilbert transform is performed on these signals to obtain the analytic signal of the mud level fluctuation signal. The modulus of this analytic signal is the instantaneous amplitude sequence of the mud level fluctuation signal. The instantaneous amplitude sequence of the mud level fluctuation signal is then cross-correlated with the baseband component of the energy fluctuation mode to calculate the normalized cross-correlation coefficient. This normalized cross-correlation coefficient reflects the correlation strength between the driving current energy fluctuation and the dynamic changes in the mud level. This normalized cross-correlation coefficient, together with the aforementioned baseband component, constitutes the complete energy fluctuation mode.

[0022] The generation of coherent modes is based on specific frequency band analysis and phase synchronization characteristics of acoustic signals. In some embodiments, multiple acoustic sensors are arranged directly below the vibrating screen and on the inner wall of the mud buffer tank. The acoustic sensor array acquires the raw acoustic signal at a sampling rate of 48 kHz per second. A bandpass digital filter is applied to the raw acoustic signal, with the passband range set between 200 Hz and 800 Hz, to extract specific frequency band signals related to solid particle collisions with the screen and fluid vortex shedding. An envelope detection algorithm is applied to the specific frequency band signal, which uses Hilbert transform to calculate the analytical envelope of the signal, thereby extracting the envelope of the specific frequency band signal. Simultaneously, a digital phase-locked loop (PLL) is used to track the phase of a dominant frequency component in the specific frequency band signal. The PLL generates two orthogonal local oscillation signals that are phase-locked with the input signal. The specific frequency band signal is multiplied by these two orthogonal local oscillation signals and low-pass filtered to generate an in-phase demodulated signal and a quadrature demodulated signal. Optionally, the coherence coefficients of the in-phase demodulated signal and the quadrature demodulated signal within a short time window are calculated. The formula for calculating the coherence coefficient is as follows: in: Represents the coherence coefficient. This represents the cross-power spectral density of the in-phase demodulated signal and the quadrature demodulated signal. and These represent the autopower spectral densities of the in-phase demodulated signal and the quadrature demodulated signal, respectively. The coherence coefficient characterizes the degree of phase lock between the in-phase and quadrature demodulated signals; this coherence coefficient is defined as the coherence mode. It can be understood that a higher coherence coefficient indicates a more pronounced phase synchronization phenomenon caused by particle motion in the acoustic signal.

[0023] See Figure 3In one embodiment of the present invention, the extraction of sub-features is a fundamental step in constructing a topological correlation network for energy fluctuation modes. Specifically, four specific sub-features are extracted from energy fluctuation mode data acquired within a continuous ten-minute sampling period. Sub-feature one is the variance of the baseband component, sub-feature two is the skewness of the baseband component, sub-feature three is the peak factor of the baseband component, and sub-feature four is the normalized cross-correlation coefficient between the energy fluctuation mode and the mud level fluctuation signal. The Spearman rank correlation coefficient between any two sub-features is calculated; for example, the Spearman rank correlation coefficient between the variance and skewness of the baseband component, and the Spearman rank correlation coefficient between the variance and peak factor of the baseband component are calculated, until the Spearman rank correlation coefficients for all six pairs of sub-feature combinations are calculated. Four sub-features are defined as four nodes in the network. The edges connecting the nodes represent the relationships between the sub-features. The weight of each edge is assigned the absolute value of the Spearman rank correlation coefficient between the corresponding two sub-features. The larger the absolute value of the Spearman rank correlation coefficient, the stronger the monotonic correlation between the two sub-features. Based on the four nodes and six weighted edges, a fully connected weighted undirected graph is constructed. This fully connected weighted undirected graph is the topological network of energy fluctuation modes.

[0024] The process of extracting feature nodes with key connectivity attributes from a topologically connected network involves calculating centrality metrics. In a topologically connected network, the betweenness centrality of each node is calculated. The betweenness centrality is defined as the proportion of times the node appears on the shortest paths between all other node pairs in the network, relative to the total number of shortest paths. The eigenvector centrality of each node is also calculated. Eigenvector centrality is obtained by solving the principal eigenvectors of the network's adjacency matrix; the value of each element in the principal eigenvector is the eigenvector centrality score of the corresponding node. Multiplying a node's betweenness centrality score by its eigenvector centrality score yields its overall centrality score. A centrality score threshold is set, and nodes with an overall centrality score higher than this threshold are selected. Optionally, when the overall centrality scores of four nodes are 0.12, 0.45, 0.08, and 0.31, and the centrality score threshold is set to 0.30, nodes two and four are considered feature nodes with key connectivity attributes. The sub-features corresponding to nodes two and four are extracted; the set of these two sub-features forms the first feature subset.

[0025] In the process of correlation analysis and dominant eigenvector identification for coherent modes, the extraction of time-frequency features is the first step. In some embodiments, three time-frequency features are extracted from the coherent mode signal: the total energy of the signal within a specific frequency band in an analysis window, the standard deviation of the instantaneous frequency sequence of the signal within the specific frequency band, and the Shannon entropy value of the marginal spectrum of the signal within the specific frequency band. The extracted time-frequency features are grouped and arranged according to five different particle size monitoring dimensions, which correspond to five particle size intervals of solid particles, forming a feature matrix with five rows and three columns. The first row of the feature matrix represents the three time-frequency feature values ​​under the first particle size monitoring dimension, the second row represents the three time-frequency feature values ​​under the second particle size monitoring dimension, and so on. The covariance matrix of this feature matrix is ​​calculated. This covariance matrix is ​​the correlation matrix, which describes the correlation of time-frequency features under different particle size monitoring dimensions. The correlation matrix is ​​a three-row, three-column square matrix.

[0026] The process of performing singular value decomposition (SVD) on the correlation matrix has explicit mathematical steps. SVD decomposes the correlation matrix into the product of three matrices: a left singular vector matrix, a diagonal matrix composed of singular values, and the transpose of a right singular vector matrix. The largest singular value is identified from the diagonal matrix of singular values; assuming the largest singular value is 2.8, the corresponding left singular vector is extracted. This left singular vector is the dominant eigenvector of the correlation matrix. Optionally, the dominant eigenvector of the correlation matrix can be in the form [0.6, -0.3, 0.74]. From the dominant eigenvector, elements whose absolute values ​​are greater than a preset contribution threshold are selected. This preset contribution threshold can be understood as 0.5. When the dominant eigenvector is [0.6, -0.3, 0.74], the absolute values ​​of the first element 0.6 and the third element 0.74 are greater than 0.5. The first element corresponds to the time-frequency feature of the energy of a signal in a specific frequency band, and the third element corresponds to the time-frequency feature of the entropy value of the marginal spectrum of the signal in that specific frequency band. These two time-frequency features constitute the second feature subset. In some embodiments, the mathematical expression of singular value decomposition is: in: Represents the correlation matrix. Represents the left singular vector matrix. This represents a diagonal matrix containing singular values. This represents the transpose of a right singular vector matrix.

[0027] In one embodiment of the present invention, orthogonality testing is the primary operation in the process of orthogonality testing and feature fusion of the first feature subset and the second feature subset. Specifically, it is assumed that the first feature subset contains two features: the baseband component skewness of the energy fluctuation mode and the normalized cross-correlation coefficient with the mud level fluctuation signal. It is assumed that the second feature subset contains two features: the energy of the specific frequency band signal of the coherent mode and the entropy value of the marginal spectrum of the specific frequency band signal. The Pearson correlation coefficient between each feature in the first feature subset and each feature in the second feature subset is calculated, forming a two-row, two-column correlation coefficient matrix. The Pearson correlation coefficients between the baseband component skewness and the specific frequency band signal energy, the baseband component skewness and the marginal spectrum entropy value, the normalized cross-correlation coefficient and the specific frequency band signal energy, and the normalized cross-correlation coefficient and the marginal spectrum entropy value are calculated. In the correlation coefficient matrix, the proportion of coefficients with absolute values ​​less than a preset orthogonality threshold is calculated out of all coefficients. The preset orthogonality threshold is set to 0.2. Optionally, when the calculated correlation coefficient matrix is ​​[[0.1,0.05],[0.15,0.3]], the proportion of coefficients with absolute values ​​less than 0.2 is three-quarters. If the preset proportion threshold is 0.7, then the current proportion exceeds the preset proportion threshold, and it is determined that the first feature subset and the second feature subset satisfy the orthogonality condition as a whole. When the orthogonality condition is satisfied, the two features of the first feature subset and the two features of the second feature subset are directly concatenated to form an amplified feature vector containing four features.

[0028] Principal component analysis (PCA) is performed on the amplified feature vector to achieve dimensionality reduction and fusion. In practice, PCA is performed on the amplified feature vector containing four features. PCA transforms the original four features into four uncorrelated new features, i.e., four principal components. The variance contribution rate and cumulative contribution rate of each principal component are calculated. A cumulative contribution rate threshold is set, for example, 0.9, and the top few principal components whose cumulative contribution rates exceed the threshold are retained. Assuming the cumulative contribution rate of the first three principal components has reached 0.92, these three principal components are retained, while the fourth principal component with a low contribution rate is discarded. This results in a dimensionality-reduced fused feature vector containing three principal components. This fused feature vector is the difference description vector describing the differences in solid phase behavior at different particle sizes.

[0029] The process of carrier modulation of the difference description vector using the real-time vibration frequency of the vibrating screen as the modulation signal includes the generation of the modulation signal. The pulse signal output from the encoder of the vibrating screen spindle is acquired in real time. The rotational speed signal of the vibrating screen spindle is obtained by calculating the number of pulses per unit time. The real-time vibration frequency signal of the vibrating screen is then obtained through the conversion relationship between rotational speed and vibration frequency. The real-time vibration frequency signal is then normalized using the following formula: in: Represents the frequency modulation coefficient. Represents the real-time vibration frequency signal. and These represent the minimum and maximum values ​​of the preset safe operating range of the vibration frequency, respectively. This normalization process ensures that the amplitude of the frequency modulation coefficient ranges between zero and one.

[0030] The characteristic components related to vibration energy transfer in the differential description vector are identified and modulated. These characteristic components are identified based on the magnitude of their covariance with the vibration frequency signal in historical data. The components with the largest absolute values ​​of their covariance with the vibration frequency signal are selected as the vibration energy-related characteristic components. Optionally, assuming the differential description vector contains three principal components, the covariance of each principal component with the historical vibration frequency signal is calculated, and the two principal components with the largest absolute values ​​of their covariance are selected as the vibration energy-related characteristic components. The frequency modulation coefficient is multiplied one by one with the identified vibration energy-related characteristic components to achieve amplitude modulation. The modulated vibration energy-related characteristic components are then recombine with the other unmodulated characteristic components in the differential description vector to form a new vector, which is the modulated description vector. The features related to the vibrating screen state in the modulated description vector are selectively enhanced.

[0031] The steps of training the behavior prediction model and using the model output to predict trajectory deviation values ​​involve model construction and data preparation. In some embodiments, the behavior prediction model is used to simulate the expected separation trajectory of solid particles of different sizes under current process parameters. High-speed camera systems are used to capture particle motion within the separation equipment at a rate of 1000 frames per second, and image analysis techniques are used to extract the actual trajectory coordinates of solid particles within the separation equipment representing a specific particle size range. Trajectory data is obtained under different combinations of centrifuge separation factors, desliming device feed pressure, and dilution water addition rates to form training samples. A behavior prediction model is constructed, comprising a two-layer long short-term memory network and a single-layer attention mechanism. The input to the behavior prediction model is a combination vector of the current centrifuge separation factor setpoint, desliming device hydrocyclone feed pressure setpoint, dilution water addition rate setpoint, and a modulated descriptive vector. The output is a sequence of coordinates of five consecutive key points on the predicted particle trajectory. The behavior prediction model is trained using the training samples, and the model parameters are optimized by minimizing the root mean square error between the predicted trajectory key point coordinates and the actual trajectory key point coordinates.

[0032] Model application and deviation calculation form the basis for control decisions. After training, the current process parameter setpoint and modulated description vector are input into the behavior prediction model. The model will output the predicted particle motion trajectory for each of the five selected key granularity segments. The knowledge base pre-stores the coordinate sequence of the ideal separation trajectory for each key granularity segment under ideal separation conditions. The predicted trajectory of each granularity segment output by the model is compared point-by-point with the corresponding ideal separation trajectory in the knowledge base. The average Euclidean distance between all corresponding coordinate points of each predicted trajectory and the corresponding ideal trajectory is calculated. This average Euclidean distance is the predicted trajectory deviation value for the corresponding granularity segment. It can be understood that the predicted trajectory deviation values ​​of all five key granularity segments constitute a deviation value set containing five deviation values.

[0033] In one embodiment of the present invention, the process of iteratively correcting the process parameter setpoints in the current drilling mud solid phase control process based on the predicted trajectory deviation value begins with establishing a surrogate model. In specific implementation, the process parameter setpoints include the centrifuge separation factor setpoint, the desliming hydrocyclone feed pressure setpoint, and the dilution water addition rate setpoint. These three setpoints are used as inputs to achieve real-time optimization control of the solid phase separation process. A response surface surrogate model is established between the process parameters and the predicted trajectory deviation value. The surrogate model is constructed using a radial basis function neural network (RBN). The inputs to the RBN are the centrifuge separation factor, the desliming hydrocyclone feed pressure, and the dilution water addition rate. The output of the RBN is the weighted sum of the deviation values ​​in the deviation value set. The formula for calculating the weighted sum of the deviation values ​​in the deviation value set is: in: Represents the weighted sum of deviation values. Representing the The predicted trajectory deviation value for each granularity segment. Representing the The weighting coefficients corresponding to each granularity segment Represents the total number of particle size segments. Weighted sum based on minimizing deviation values. With the goal of optimizing the response surface surrogate model within the safe operating range of process parameters, a new set of optimized process parameter setpoints is obtained by using a sequential quadratic programming algorithm.

[0034] The steps for optimizing the response surface surrogate model using a sequential quadratic programming algorithm during the iterative correction process include constructing sub-problems. In specific implementation, the goal is to minimize the weighted sum of predicted trajectory deviation values. Using the objective function, the safe operating ranges of the centrifuge's separation factor, the desiccant hydrocyclone's feed pressure, and the dilution water addition rate are defined as nonlinear constraints, establishing an initial nonlinear programming problem for process parameter optimization. At the current process parameter setpoint, a first-order Taylor expansion is performed on the objective function and nonlinear constraints to achieve a linear approximation, constructing a quadratic programming subproblem. Solving the quadratic programming subproblem yields the search direction vector and the optimal step size scalar for the process parameters in the current iteration step. Along the search direction vector, the three process parameters—centrifuge separation factor, desiccant hydrocyclone feed pressure, and dilution water addition rate—are updated according to the optimal step size scalar, resulting in a set of candidate process parameter optimization setpoints. It is calculated whether the updated process parameters satisfy their respective preset safe operating range constraints, such as whether the centrifuge separation factor is between 1500 and 2500. If the constraints are not met, the search direction and step size are adjusted using the active set method, and the calculation is repeated until all process parameters satisfy the constraints. The candidate process parameter optimization setpoints that satisfy the constraints are taken as the output of the current iteration; this result is the new process parameter optimization setpoint.

[0035] The adjustment of process parameter setpoints and iterative system operation form a closed loop. After obtaining new optimized process parameter setpoints, the current setpoints for the centrifuge separation factor, the desliming hydrocyclone feed pressure, and the dilution water addition rate are adjusted slightly towards the optimized process parameter setpoints. The adjustment step size is controlled by a preset learning rate parameter, for example, 0.1. The solid-phase control system is run for a predetermined time window under the adjusted new process parameters, for example, five minutes. Afterward, the system re-acquires data such as vibrating screen current, slurry level, and acoustic signals, and repeats the complete process from establishing multimodal feature characterization to calculating new predicted trajectory deviation values. The weighted sum of predicted trajectory deviation values ​​calculated in the current iteration cycle is compared with the weighted sum of predicted trajectory deviation values ​​from the previous iteration cycle. If the decrease in the weighted sum of deviation values ​​is less than a preset convergence threshold, for example, 0.5, the process parameter optimization process is considered to have converged, and the current setpoint is maintained. If the decrease in the weighted sum of the deviation values ​​is greater than or equal to the preset convergence threshold, then a new round of iterative optimization will begin, starting from the current set point.

[0036] The data flow of the process parameter optimization iteration process can be shown in Table 1. Table 1 provides an example of a simplified iteration process, which includes the centrifuge separation factor, the desliming device feed pressure, the dilution water addition rate, and the weighted sum of the deviation values ​​calculated after each iteration.

[0037] Table 1: Examples of Process Parameter Optimization Iterations In one embodiment of the present invention, a recording and rollback mechanism for the process parameter optimization path is introduced during the iterative correction of the process parameter setpoint to enhance system robustness. In specific implementation, after each iteration of adjusting the process parameter setpoint, the setpoint before adjustment, the setpoint after adjustment, and the corresponding weighted sum of the predicted trajectory deviation values ​​are recorded in an optimization history stack. The setpoint before adjustment includes the centrifuge separation factor setpoint, the desliming hydrocyclone feed pressure setpoint, and the dilution water addition rate setpoint. The setpoint after adjustment includes the centrifuge separation factor setpoint, the desliming hydrocyclone feed pressure setpoint, and the dilution water addition rate setpoint. The optimization history stack stores these records using a "last-in, first-out" data structure; for example, the record from the first iteration is pushed to the bottom of the stack, and records from subsequent iterations are pushed to the top. In some embodiments, a record in the optimization history stack includes the following fields: iteration number, the three process parameters before adjustment, the three process parameters after adjustment, and the weighted sum of the predicted trajectory deviation values ​​calculated in this iteration.

[0038] The rollback mechanism is triggered based on the continuous deterioration of performance during optimization. If, after several iterations, the weighted sum of the predicted trajectory deviation values ​​not only fails to decrease but continues to rise beyond a preset deterioration threshold, the rollback mechanism is triggered. The preset deterioration threshold is set to three. For example, if the weighted sums of the predicted trajectory deviation values ​​calculated in three consecutive iterations are 6.81, 6.90, and 7.15, showing a continuous upward trend, and the number of increases exceeds the deterioration threshold, the rollback mechanism is triggered. The rollback mechanism searches the optimization history stack for the record with the smallest weighted sum of predicted trajectory deviation values ​​and restores the corresponding process parameter setpoint to the system's current operating setpoint. For instance, if the third record in the optimization history stack has a weighted sum of predicted trajectory deviation values ​​of 6.82, which is the smallest among all records, the rollback mechanism will restore the centrifuge separation factor setpoint, the desliming hydrocyclone feed pressure setpoint, and the dilution water addition rate setpoint to the adjusted setpoint values ​​stored in the third record.

[0039] After the rollback operation, the system will perform adaptive adjustments to continue the optimization process. Following the rollback, the system will reduce the optimization step size of the process parameters and restart the iterative optimization process. Reducing the optimization step size can be achieved by decreasing the upper limit of the step size defined in the sequential quadratic programming algorithm, or by reducing the learning rate parameter mentioned in the embodiment. For example, the learning rate parameter can be reduced from 0.1 to 0.05. After reducing the step size, the system uses the rolled-back restored process parameter setpoint as a new starting point and restarts the complete closed-loop control process from establishing multimodal feature representations to iteratively correcting the process parameter setpoint. It can be understood that the recording and rollback mechanism constitutes a protection for the optimization search process, and its logic can be summarized as: recording historical states, monitoring performance degradation, locating the historical optimal state, rolling back, and adjusting the search strategy. This logic can be formally expressed using a simple judgment formula: in: This means the rollback mechanism is triggered when the condition is true. Representing the The weighted sum of the predicted trajectory deviation values ​​obtained from each iteration This represents the starting iteration number from which the judgment begins. This represents the preset threshold for the number of deterioration events. The "AND" operation represents a logical AND operation, indicating a continuous sequence. Each iteration must satisfy the subsequent inequality condition. When this condition is satisfied, the rollback mechanism is activated.

[0040] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications 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 protection scope of the present invention.

Claims

1. A drilling mud solids control method based on particle size classification, characterized in that, Including the following steps: A multimodal feature characterization of the solid-phase separation process across multiple particle size monitoring dimensions is established, wherein the multimodal feature characterization includes at least energy fluctuation mode and coherence mode; For the energy fluctuation mode, a topological association network between its internal sub-features is constructed, and feature nodes with key connection attributes in the topological association network are extracted to form a first feature subset; For the coherence mode, the correlation matrix of its internal time-frequency features at different granularity monitoring dimensions is calculated, and the dominant feature vector in the correlation matrix is ​​identified by singular value decomposition to form a second feature subset; The first feature subset and the second feature subset are subjected to orthogonality test and feature fusion to obtain a difference description vector describing the differences in the behavior of solid phases with different particle sizes; Using the real-time vibration frequency of the vibrating screen as the modulation signal, the difference description vector is carrier modulated to enhance the feature components that are strongly correlated with the working state of the vibrating screen, thereby generating a modulated description vector. Based on the historical trajectory data of solid particles in the separation equipment, a behavior prediction model is trained. The modulated description vector is input into the behavior prediction model, and a series of predicted trajectory deviation values ​​for key particle size segments are output. Based on the predicted trajectory deviation value, the process parameter set points in the current drilling mud solid phase control process are iteratively corrected.

2. The drilling mud solids control method based on particle size classification according to claim 1, characterized in that, The establishment of a multimodal characterization of the solid-phase separation process across multiple particle size monitoring dimensions includes: The energy fluctuation mode is obtained by decoupling the driving current waveform of the vibrating screen from the mud level fluctuation. The coherent modes are generated by envelope detection and phase locking of acoustic signals in a specific frequency band; The three-phase current signal is obtained from the drive motor controller of the vibrating screen. The three-phase current signal is subjected to Clark transformation to obtain the two-phase current signal in the stationary coordinate system. The instantaneous amplitude sequence of the two-phase current signal is calculated and this sequence is defined as the baseband component of the energy fluctuation mode. The fluctuation signal of the mud level height changing with time is obtained from the mud level sensor, and the fluctuation signal is subjected to Hilbert transform to obtain its analytical signal. The instantaneous amplitude of the analytical signal is calculated, and the instantaneous amplitude sequence is cross-correlated with the baseband component of the energy fluctuation mode. The normalized cross-correlation coefficient is calculated, and the normalized cross-correlation coefficient is used as a sub-feature of the energy fluctuation mode, which together with the baseband component constitutes the energy fluctuation mode. Raw acoustic signals are acquired from an acoustic sensor array arranged in the vibrating screen box and mud pit. The raw acoustic signals are then bandpass filtered to extract specific frequency band signals related to solid particle collisions and fluid vortices. An envelope detection algorithm is applied to the specific frequency band signal to extract the signal envelope, and phase-locked loop technology is used to track the phase of the dominant frequency component in the specific frequency band signal to generate two demodulated signals, one in phase and one quadrature. The coherence coefficient between the two demodulated signals is calculated. This coefficient characterizes the phase locking degree between the signals and is defined as the coherent mode.

3. The drilling mud solids control method based on particle size classification according to claim 2, characterized in that, The construction of the topological association network among its internal sub-features, and the extraction of feature nodes with key connection attributes from the topological association network to form the first feature subset, includes: Multiple sub-features are extracted from the energy fluctuation mode, including the variance, skewness, peak factor, and normalized cross-correlation coefficient of the baseband component with the mud level fluctuation signal. Calculate the Spearman rank correlation coefficient between any two of the sub-features, treat all sub-features as nodes, and use the absolute value of the Spearman rank correlation coefficient as the weight of the edge between the nodes to construct a fully connected weighted undirected graph. This weighted undirected graph is the topological network of energy fluctuation modes. In the aforementioned topological network, the betweenness centrality of each node is calculated, which is the number of times the node appears on the shortest path in the network, and the eigenvector centrality of each node is calculated, which is the centrality of a node as a function of the centrality of its neighboring nodes. Multiply the betweenness centrality score of each node by the eigenvector centrality score to obtain the overall centrality score of each node; A centrality score threshold is set, nodes with a comprehensive centrality score higher than the centrality score threshold are selected, and the sub-features of the energy fluctuation mode corresponding to the node are extracted and set to form a first feature subset.

4. The drilling mud solids control method based on particle size classification according to claim 2, characterized in that, The calculation of the correlation matrix of its internal time-frequency features at different granularity monitoring dimensions, and the identification of the dominant feature vector in the correlation matrix through singular value decomposition to form the second feature subset includes: Multiple time-frequency features are extracted from the coherent modes, including the energy of the signal in a specific frequency band, the standard deviation of the instantaneous frequency, and the entropy value of the signal's marginal spectrum; The extracted time-frequency features are grouped and arranged according to different granularity monitoring dimensions to form a feature matrix, where each row represents a granularity monitoring dimension and each column represents a time-frequency feature. Calculate the covariance matrix of the feature matrix. This covariance matrix is ​​the correlation matrix that describes the correlation of time-frequency features under different granularity monitoring dimensions. Perform singular value decomposition on the correlation matrix to decompose it into the product of three matrices: a left singular vector matrix, a diagonal matrix composed of singular values, and the transpose of a right singular vector matrix. Identify the largest singular value from the diagonal matrix and extract the left singular vector corresponding to the largest singular value. This left singular vector is the dominant feature vector of the correlation matrix. From the dominant feature vector, elements whose absolute values ​​are greater than a preset contribution threshold are selected, and the time-frequency features corresponding to these elements constitute the second feature subset.

5. The drilling mud solids control method based on particle size classification according to claim 1, characterized in that, The orthogonality test and feature fusion of the first feature subset and the second feature subset include: Calculate the Pearson correlation coefficient between each feature in the first feature subset and each feature in the second feature subset to form a correlation coefficient matrix; For the correlation coefficient matrix, the proportion of coefficients whose absolute value is less than a preset orthogonality threshold is counted. If the proportion exceeds the preset proportion threshold, it is determined that the first feature subset and the second feature subset satisfy the orthogonality condition as a whole. When the orthogonality condition is met, all features of the first feature subset are directly concatenated with all features of the second feature subset to form an amplified feature vector. Principal component analysis is performed on the amplified feature vector, and the first few principal components whose cumulative contribution rate exceeds the preset contribution rate threshold are retained to obtain a dimension-reduced fused feature vector. This fused feature vector is the difference description vector describing the differences in the behavior of solid phases of different particle sizes.

6. The drilling mud solids control method based on particle size classification according to claim 5, characterized in that, The step of using the real-time vibration frequency of the vibrating screen as the modulation signal to perform carrier modulation on the difference description vector includes: The rotational speed signal of the main shaft of the vibrating screen is acquired in real time, and the real-time vibration frequency signal of the vibrating screen is obtained through the conversion relationship between rotational speed and frequency. The real-time vibration frequency signal is normalized so that its amplitude range is between zero and one, thereby generating a frequency modulation coefficient. Feature components related to vibration energy transfer are identified from the difference description vector, and the identification of the feature components is determined based on the magnitude of the covariance between the feature components and the vibration frequency signal in historical data; The frequency modulation coefficient is multiplied one by one with the identified vibration energy-related characteristic components to achieve amplitude modulation of the characteristic components; The modulated vibration energy-related feature components are recombined with the other unmodulated feature components in the difference description vector to form a new vector, which is the modulated description vector, in which the features related to the state of the vibrating screen are selectively enhanced.

7. The drilling mud solids control method based on particle size classification according to claim 6, characterized in that, The training behavior prediction model inputs the modulated description vector into the behavior prediction model and outputs a series of predicted trajectory deviation values ​​for key granularity segments, including: The behavior prediction model is used to simulate the expected separation trajectory of solid particles of different sizes under the current process parameters; Using high-speed camera and image analysis technology, the actual motion trajectory data of solid particles in the separation equipment with representative particle size range under different process parameter settings are obtained as training samples; A behavior prediction model incorporating a long short-term memory network and an attention mechanism is constructed. The input of the behavior prediction model is a combination of the current process parameter setpoint and the modulated description vector, and the output is the predicted sequence of key point coordinates of the particle motion trajectory. The behavior prediction model is trained using the training samples, and the model parameters are optimized by minimizing the root mean square error between the predicted trajectory and the actual trajectory. After training is completed, the current process parameter setpoint and the modulated description vector are input into the behavior prediction model. The model will output the corresponding predicted particle motion trajectory for each of the selected key granularity segments. The predicted trajectory output by the model is compared point by point with the ideal separation trajectory of the corresponding granularity segment stored in the knowledge base. The average Euclidean distance between each predicted trajectory and the corresponding ideal trajectory is calculated. This average Euclidean distance is the predicted trajectory deviation value of the corresponding granularity segment. The predicted trajectory deviation values ​​of all key granularity segments constitute a deviation value set.

8. The drilling mud solids control method based on particle size classification according to claim 7, characterized in that, The iterative correction of the process parameter setpoints in the current drilling mud solids control process includes: The process parameter setting points include the centrifuge separation factor setting value, the desliming hydrocyclone feed pressure setting value, and the dilution water addition rate setting value, thereby realizing real-time optimized control of the solid phase separation process. A response surface surrogate model is established between process parameters and predicted trajectory deviation values. The surrogate model is constructed using a radial basis function neural network. Its inputs are the separation factor of the centrifuge, the feed pressure of the hydrocyclone of the sludge dewatering unit, and the dilution water addition rate. Its output is the weighted sum of each deviation value in the deviation value set. With the goal of minimizing the weighted sum, within the safe operating range of the process parameters, the response surface surrogate model is optimized and solved using a sequential quadratic programming algorithm to obtain a new set of optimized process parameter setpoints. The current process parameter setpoint is adjusted slightly towards the process parameter optimization setpoint, and the adjustment step size is controlled by a preset learning rate parameter. The solid-state control system is run for a predetermined time window under the adjusted new process parameters, after which data is reacquired and the entire process from establishing multimodal feature characterization to calculating new predicted trajectory deviation values ​​is repeated. Compare the weighted sum of the predicted trajectory deviation values ​​calculated in the current iteration cycle with the value in the previous iteration cycle. If the decrease in the weighted sum of the deviation values ​​is less than the preset convergence threshold, the process parameter optimization process is determined to have converged, and the current setpoint is maintained; otherwise, a new round of iterative optimization is started from the current setpoint.

9. The drilling mud solids control method based on particle size classification according to claim 8, characterized in that, The sequential quadratic programming algorithm is used to optimize the response surface surrogate model, resulting in a new set of optimized process parameter setpoints, including: With the objective function of minimizing the weighted sum of predicted trajectory deviations, and defining the safe operating range of process parameters as a nonlinear constraint, an initial nonlinear programming problem for process parameter optimization is established. At the current process parameter setpoint, a quadratic programming subproblem is constructed by linearly approximating the objective function and the nonlinear constraints. Solving the quadratic programming subproblem yields the search direction and optimal step size for the process parameters in the current iteration step. Along the search direction, the process parameters are updated according to the optimal step size to obtain a set of candidate process parameter optimization setpoints; Calculate whether the updated process parameters meet the safe operating range constraints. If not, adjust the search direction and step size using the active set method and recalculate until the constraints are met. The candidate process parameter optimization setpoint that satisfies the constraints is taken as the output of the current iteration, and this result is the new process parameter optimization setpoint.

10. The drilling mud solids control method based on particle size classification according to claim 9, characterized in that, The process of iteratively correcting the process parameter setpoints also includes a mechanism for recording and rolling back the process parameter optimization path: After each iteration of adjusting the process parameter setpoint, the setpoint before adjustment, the setpoint after adjustment, and the corresponding predicted trajectory deviation value are weighted and recorded in an optimization history stack. If, after several consecutive iterations, the weighted average of the predicted trajectory deviation values ​​not only fails to decrease but continues to rise beyond the preset threshold for the number of deteriorations, then the rollback mechanism is triggered. The rollback mechanism finds the record with the smallest weighted sum of predicted trajectory deviation values ​​in the optimization history stack and restores the process parameter setpoint corresponding to the record to the current operating setpoint of the system. After the rollback operation is executed, the system will reduce the optimization step size of the process parameters and restart the iterative optimization process.

Images