A mechanism alignment MoE ship-mounted cyclone intelligent control method

Abstract

The application relates to the technical field of automatic control of ore dressing equipment, and discloses a mechanism alignment MoE intelligent control method for a ship-mounted hydrocyclone. The method comprises the following steps: collecting multi-source time sequence data, constructing a first comparison table of the multi-source time sequence data and a learnable fuzzy membership function, and generating fuzzy features; constructing a second comparison table of the multi-source time sequence data and a hydrocyclone adjustment rule, generating rule activation strength, and obtaining mechanism constraint features according to the fuzzy features combined with a multilayer perception machine; obtaining attention features based on the multi-source time sequence data, the fuzzy features and the mechanism constraint features combined with a gating mechanism and a multi-head attention; performing time sequence modeling again to obtain time sequence enhancement features; splicing the time sequence enhancement features and the rule activation strength to obtain a first prediction result; obtaining a second prediction result based on the first prediction result combined with the rule activation strength weighted; and adjusting the ship-mounted hydrocyclone based on the second prediction result. The method solves the problem that the existing ship-mounted hydrocyclone cannot be accurately controlled.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology for mineral processing, and in particular to a shipborne hydrocyclone collaborative control technology based on multi-source time-series data and hydrocyclone separation mechanism. Background Technology

[0002] Current control methods for shipborne hydrocyclones primarily rely on traditional control strategies, which generally suffer from significant technical limitations. On one hand, traditional fuzzy control methods typically employ fixed fuzzy rules and simple membership functions, failing to adaptively adjust the inference logic based on sea state changes and operational fluctuations. Feature extraction and information fusion methods are relatively simplistic, making it difficult to fully exploit the coupling patterns behind multi-source time-series data, resulting in weak system anti-disturbance capabilities and dynamic response performance. On the other hand, existing control schemes often employ a single model structure, enabling limited adjustment only under specific stable operating conditions. They cannot adaptively cover diverse scenarios such as calm sea states, moderately undulating sea states, and strongly disturbed sea states, leading to control failures and response lags during operational condition switching. It is worth noting that the Mixture of Experts (MoE) model, which has emerged in recent years, can dynamically select the most suitable expert for inference based on the characteristics of the input data by introducing multiple expert sub-networks and learnable gating mechanisms. It has shown significant advantages in multi-task and multi-scenario adaptive modeling. Its "divide and conquer" architecture is naturally suited to the needs of shipborne hydrocyclones for switching between multiple sea states, and the gating network can integrate multi-source time-series signals to achieve soft partitioning of operating conditions, thereby improving the model's response flexibility to highly random and highly disturbed environments. However, existing shipborne control research has not yet organically integrated the MoE architecture with hydrocyclone separation mechanisms and fuzzy inference systems, and related applications are still lacking.

[0003] In summary, existing control technologies are ill-suited to the operational characteristics of shipborne hydrocyclones in complex marine environments. They exhibit significant shortcomings in control accuracy, dynamic response, robustness, and adaptability to various operating conditions, failing to meet the requirements for high-precision, high-robustness, and high-stability closed-loop control under extreme marine conditions. This, to some extent, restricts the level of intelligence and large-scale engineering application of shipborne mineral processing equipment. Summary of the Invention

[0004] This invention provides a smart control method for shipborne hydrocyclones with mechanism-aligned MoE, in order to solve the problem of the inability to accurately control existing shipborne hydrocyclones.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] This invention provides a smart control method for a shipborne hydrocyclone with mechanism-aligned MoE, comprising the following steps: Step 1: Collect multi-source time-series data, including ship attitude disturbances, construct a first lookup table of multi-source time-series data and multiple types of learnable fuzzy membership functions, and generate fuzzy features based on multi-source time-series data and the first lookup table. Step 2: Construct a second lookup table corresponding to multi-source time-series data and cyclone adjustment rules. Based on fuzzy features, combine the first and second lookup tables to generate rule activation intensity. Obtain mechanistic constraint features based on fuzzy features and multilayer perceptron. Step 3: Based on multi-source time series data, fuzzy features, and mechanistic constraint features, gating mechanisms are combined to generate gated fusion features. Based on the gated fusion features, fuzzy features, and mechanistic constraint features, multi-head attention is combined to obtain attention features. Temporal modeling is performed on the long-term and local fluctuations of attention features, and the obtained temporal features are spliced ​​to obtain temporal enhancement features. Step 4: After concatenating the temporal enhancement features with the rule activation intensity, combine them with a multilayer perceptron to make a prediction and obtain the first prediction result. Based on the first prediction result, combine the rule weighting of the cyclone adjustment corresponding to the rule activation intensity to obtain the second prediction result. Adjust the shipborne cyclone based on the second prediction result. Sub-objective functions are set based on minimizing control deviation, smoothing control commands, and resisting disturbances, and multi-objective constraint functions are set based on multiple sub-objective functions.

[0007] Furthermore, the multi-source time-series data includes: hull attitude disturbance angle, control relative deviation, control deviation change rate, slurry density, feed pressure, feed flow fluctuation amplitude, and actuator feedback offset.

[0008] Furthermore, the first lookup table divides various physical parameters in the multi-source time series data into multiple value ranges, and sets a level for each value range and the initial value of the learnable fuzzy membership function corresponding to that level. The learnable fuzzy membership functions include triangular learnable fuzzy membership functions, S-shaped learnable fuzzy membership functions, Z-shaped learnable fuzzy membership functions, and Gaussian learnable fuzzy membership functions.

[0009] Through the above design, four learnable fuzzy membership functions—triangular, S-shaped, Z-shaped, and Gaussian—are integrated to achieve adaptive parameter updates, replacing traditional fixed fuzzy rules and fixed membership degrees. This enables adaptive fitting of hull disturbances and nonlinear characteristics under operating conditions, significantly improving fuzzy representation capabilities.

[0010] Furthermore, the generation of the fuzzy features includes: determining the level and learnable fuzzy membership function with initial value adjustment based on the value range of a single type of physical parameter in multi-source time series data; calculating the membership degree based on other learnable fuzzy membership functions with unadjusted initial values ​​and learnable fuzzy membership functions with adjusted initial values; and selecting the level corresponding to the learnable fuzzy membership function with the highest membership degree as the level of that type of physical parameter. Repeatedly calculating the levels of all physical parameters in the original feature constitutes a fuzzy feature.

[0011] Furthermore, in the second reference table, the levels for each physical parameter in the adjustment rules for each hydrocyclone are set to correspond to those in the first reference table; The activation intensity of the rule includes the activation intensity of each hydrocyclone adjustment rule in the second lookup table; The acquisition of the rule activation strength of any hydrocyclone adjustment rule includes: confirming the level of each physical parameter based on multi-source time series data and a second lookup table; calculating the membership degree corresponding to each physical parameter by calling the initial value of the corresponding learnable fuzzy membership function in the first lookup table according to the level; and selecting the minimum positive value among the membership degrees corresponding to all physical parameters of the hydrocyclone adjustment rule as the rule activation strength of the hydrocyclone adjustment rule.

[0012] Furthermore, the process of obtaining the rule activation intensity of any hydrocyclone adjustment rule also includes: setting membership thresholds for each physical parameter and truncating the calculated membership.

[0013] Through the above design, multiple hydrocyclone adjustment rules are embedded, deeply binding the control logic with the hydrocyclone flow field mechanism, multivariable collaborative mechanism, disturbance compensation mechanism, and equipment safety constraint mechanism, so as to achieve strict alignment between control decisions and physical mechanisms.

[0014] Furthermore, in the multi-head attention in step 3, the mechanism constraint feature serves as the query vector in multi-head attention, the fuzzy feature serves as the key vector in multi-head attention, and the gated fusion feature serves as the value vector in multi-head attention.

[0015] Furthermore, the multi-head attention adopts lightweight multi-query attention.

[0016] Through the above design, the mechanism-constrained features, fuzzy features, fused features and the branches of multi-head attention are bound together to achieve accurate alignment between physical semantics and data features, and to avoid the misleading of feature focus by disturbance noise from the source. At the same time, a lightweight multi-query attention is adopted to significantly reduce the amount of computation while ensuring anti-disturbance performance, and it can be directly deployed on a shipborne embedded platform.

[0017] Furthermore, a bidirectional long short-term memory network is used for temporal modeling of attention features over long time. Temporal modeling is performed using a temporal convolutional network to model the local fluctuations of attention features.

[0018] Through the above design, the long-term disturbance hysteresis characteristics and multi-scale wave wave characteristics are captured simultaneously, accurately matching the strong temporal and strong wave characteristics of hull disturbance.

[0019] Furthermore, in step 4, obtaining the first prediction result includes: The temporal augmentation features are concatenated with the rule activation intensity after dimensionality reduction of the operating conditions to obtain temporal augmentation features with physical priors. Then, the temporal augmentation features with physical priors are input into multiple multilayer perceptrons to obtain the control quantities of multiple dimensions of the shipborne hydrocyclone as the first prediction result. The multiple multilayer perceptrons are set up based on expert networks for different working conditions, and each multilayer perceptron performs a single-pull linear mapping for the corresponding working condition. The working condition dimensionality reduction includes selecting multiple rule activation intensities corresponding to multiple multilayer perceptrons.

[0020] Furthermore, it also includes step 5: applying process limits to the second prediction result to obtain the third prediction result, and adjusting the shipborne hydrocyclone based on the third prediction result.

[0021] Furthermore, it also includes step 5: after process limiting the second prediction result, rate smoothing is performed to obtain the third prediction result, and the shipborne hydrocyclone is adjusted based on the third prediction result.

[0022] Furthermore, the sub-objective function for minimizing the control deviation setting is based on the implementation deviation design of minimizing the feed flow rate, feed pressure, and bottom discharge opening of the shipborne hydrocyclone; The sub-objective function for smooth setting of control commands is designed based on the difference in control commands constraining the feed flow rate, feed pressure, and underflow discharge opening at adjacent time points; The sub-objective function of the anti-disturbance setting is designed based on the impact of ship disturbance at any time on feed flow rate, feed pressure and bottom flow discharge opening.

[0023] Through the above design, a comprehensive multi-objective optimization function was constructed that simultaneously considers minimizing control deviation, smoothness of control commands, and anti-disturbance performance. This function unifies the constraints on control accuracy, command stability, equipment safety, and sea state anti-disturbance, achieving multi-objective synergistic optimization.

[0024] Beneficial effects: This invention provides a mechanism-aligned MoE-based intelligent control method for shipborne hydrocyclones. Addressing industry pain points such as flow field distortion caused by wave swaying, strong nonlinear disturbances in operating conditions, and insufficient precision, poor stability, and weak reliability of traditional control methods with pure data models, this method constructs an integrated control system of "mechanism guidance + data-driven + full-link safety constraints," forming a complete technical closed loop from target design to safe output. The comprehensive benefits are as follows: Using a multi-objective function that balances control deviation, command smoothness, and disturbance rejection performance as the top-level constraint, combined with fuzzy reasoning embedded with centrifugal separation mechanism and mechanism-aligned MoE decision, the physical mechanism is integrated throughout the entire control process, overcoming the limitations of traditional single error objective which is incomplete in all aspects.

[0025] To address the strong temporal and multi-scale random disturbances caused by ship swaying, a BiLSTM+TCN dual-backbone temporal unit is constructed to simultaneously capture the lag effect of long-period disturbances and the multi-scale wave fluctuation patterns. Furthermore, a dynamic QKV allocation structure is designed to bind attention Q / K / V to mechanistic constraint features, fuzzy features, and gated fusion features respectively. Physical priors are introduced to achieve precise alignment between mechanistic semantics and data features, overcoming the shortcomings of traditional attention methods, such as lack of physical constraints, susceptibility to disturbance noise, and scattered feature focusing. Finally, a learnable hybrid fuzzy membership degree is used to adaptively fit the nonlinearity of the operating conditions, and this triple mechanism works together to suppress disturbance interference.

[0026] From fuzzy reasoning based on mechanistic principles and attention allocation constrained by mechanistic principles to mechanism-driven MoE (Motion Equivalent Environment) scheduling, the hydrocyclone separation mechanism is deeply integrated into the entire data-driven process. This approach avoids the problems of black-box nature and easy deviation from process boundaries inherent in pure data models, ensuring that the decision-making logic is interpretable and traceable. Furthermore, it possesses stronger adaptive learning capabilities than traditional pure mechanism control, resulting in higher control accuracy under complex sea conditions.

[0027] Employing a lightweight multi-query attention and modular decoupling architecture, it significantly reduces computational load while ensuring performance, allowing direct deployment on shipborne embedded industrial control platforms and adapting to scenarios with limited computing resources at sea. Each functional unit has clear boundaries and standardized interfaces, enabling flexible adaptation and adjustment based on equipment model, mineral type, and sea conditions without requiring overall architecture reconstruction, demonstrating excellent versatility and iterative capabilities. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating a mechanism-aligned MoE-based intelligent control method for shipborne cyclones according to an embodiment of the present invention. Detailed Implementation

[0029] The technical solution of the present invention will be clearly and completely described below. 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.

[0030] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components.

[0031] Please see Figure 1 This application provides a method for intelligent control of a shipborne hydrocyclone with mechanism aligned MoE, comprising the following steps: Step 1: Collect multi-source time-series data, including ship attitude disturbances, construct a first lookup table of multi-source time-series data and multiple types of learnable fuzzy membership functions, and generate fuzzy features based on multi-source time-series data and the first lookup table. Specifically, the collected multi-source time-series data includes: hull attitude disturbance angle, control relative deviation, control deviation change rate, slurry density, feed pressure, feed flow fluctuation amplitude, and actuator feedback offset. For the construction of the first reference table, multiple value intervals are divided for various physical parameters in the multi-source time series data, and a level is set for each value interval and the initial value of the learnable fuzzy membership function corresponding to that level is also set. The types of learnable fuzzy membership functions include: triangular learnable fuzzy membership function, sigmoid learnable fuzzy membership function, z-type learnable fuzzy membership function, and Gaussian learnable fuzzy membership function. The expression for the z-type learnable fuzzy membership function is as follows: ; in, This represents a single type of physical parameter in the input multi-source time series data; This represents the left inflection point of the Z-shaped learnable fuzzy membership function; This represents the right inflection point of a Z-shaped learnable fuzzy membership function.

[0032] The expression for the trigonometric learnable fuzzy membership function is as follows: ; in, Indicates the left boundary; Indicates the right boundary; Indicates the center.

[0033] The expression for the Gaussian learnable fuzzy membership function is as follows: ; in, This represents the mean; It represents the standard deviation.

[0034] The expression for the S-shaped learnable fuzzy membership function is as follows: ; in, This represents the left inflection point of the S-shaped learnable fuzzy membership function; This represents the right inflection point of the S-shaped learnable fuzzy membership function.

[0035] The learnable fuzzy membership functions here are all adaptively updated with the initial values ​​of each membership function through the gradient descent algorithm. This automatically fits the fuzzy characteristics of hull disturbances and hydrocyclone control variables, solving the problem that fixed membership degrees in traditional fuzzy control cannot adapt to shipboard dynamic conditions. It can be understood that the initial values ​​referred to here are the hyperparameters of each learnable fuzzy membership function, namely the inflection points, center values, mean values, and standard deviations of the four types of membership functions. They are adaptively and iteratively updated through the gradient descent algorithm. After calculating the output results of the four types of membership functions, they are merged along the feature dimension to form a complete fuzzy feature. This improves the adaptive representation ability of fuzzy features to hull disturbances and process fluctuations, and adapts to shipboard nonlinear dynamic conditions.

[0036] Please refer to Table 1 for the first reference table constructed: Table 1: First comparison table.

[0037]

[0038] The generation of fuzzy features includes: determining the level and learnable fuzzy membership function with initial value adjustment based on the value range of a single type of physical parameter in multi-source time series data; calculating the membership degree based on other learnable fuzzy membership functions with unadjusted initial values ​​and learnable fuzzy membership functions with adjusted initial values; and selecting the level corresponding to the learnable fuzzy membership function with the highest membership degree as the level of that type of physical parameter. Repeatedly calculating the levels of all physical parameters in the original feature constitutes a fuzzy feature.

[0039] For example, if the ship's attitude disturbance angle is 5°, corresponding to the interval 2-6 of level 2, then the hyperparameters of the trigonometric learnable function are adjusted, and then substituted together with the S-type learnable fuzzy membership function, Z-type learnable fuzzy membership function, and Gaussian learnable fuzzy membership function (without modified hyperparameters) into the angle value of 5 corresponding to the ship's attitude disturbance angle for calculation. Finally, the highest membership value among the four learnable fuzzy membership functions is selected as the level of this type of physical parameter. The levels of control relative deviation, control deviation change rate, slurry density, feed pressure, feed flow fluctuation amplitude, and actuator feedback offset are repeatedly calculated, and finally combined to generate fuzzy features.

[0040] Step 2: Construct a second lookup table corresponding to multi-source time-series data and cyclone adjustment rules. Based on fuzzy features, combine the first and second lookup tables to generate rule activation intensity. Obtain mechanistic constraint features based on fuzzy features and multilayer perceptron. In the second reference table, the levels of each physical parameter in each hydrocyclone adjustment rule are set to correspond to those in the first reference table. The hydrocyclone adjustment rules here are understood to be formulated based on hydrocyclone adjustment strategies that can stably sort hydrocyclones under different operating conditions at sea. These rules can be formulated using statistical methods. In this embodiment, eight hydrocyclone adjustment rules are set, and the operating conditions are divided into stable operating conditions, process fluctuation operating conditions, dynamic disturbance operating conditions, strong disturbance operating conditions, and universal operating conditions. The stable operating condition corresponds to hydrocyclone adjustment rule 1. The process fluctuation conditions correspond to hydrocyclone adjustment rules 2 and 3, which are different situations under the process fluctuation conditions. The dynamic disturbance conditions correspond to hydrocyclone adjustment rule 4, the strong disturbance conditions correspond to hydrocyclone adjustment rule 5, and the full conditions correspond to hydrocyclone adjustment rules 6 to 8. Among them, hydrocyclone adjustment rules 6 to 8 are the basic controls under different critical situations, respectively corresponding to hydrocyclone equipment protection, elimination of system control deviation, and long-term flow field imbalance inside the hydrocyclone.

[0041] The activation intensity of the rules includes the activation intensity of each hydrocyclone adjustment rule in the second reference table; The acquisition of the rule activation strength of any hydrocyclone adjustment rule includes: confirming the level of each physical parameter based on multi-source time-series data and a second lookup table; calculating the membership degree corresponding to each physical parameter by calling the initial value of the corresponding learnable fuzzy membership function in the first lookup table according to the level; and selecting the minimum positive value among the membership degrees corresponding to all physical parameters of the hydrocyclone adjustment rule as the rule activation strength of the hydrocyclone adjustment rule. For details, please refer to Table 2. Table 2: Second comparison table.

[0042]

[0043] The process of obtaining the activation strength of any hydrocyclone adjustment rule also includes: setting a membership threshold for each physical parameter and truncating the calculated membership. A uniform membership threshold of 0.05 is set. If the calculated membership of a single parameter is <0.05, it is directly set to 0; if it is ≥0.05, the original value is retained. For example, if the calculated membership of a parameter is 0.03, it is truncated to 0; if another parameter has a membership of 0.12, it is retained as 0.12. Then, the minimum positive value of the truncated membership of all parameters under the same rule is taken as the activation strength of this rule.

[0044] Step 3: Based on multi-source time series data, fuzzy features, and mechanistic constraint features, gating mechanisms are combined to generate gated fusion features. Based on the gated fusion features, fuzzy features, and mechanistic constraint features, multi-head attention is combined to obtain attention features. Temporal modeling is performed on the long-term and local fluctuations of attention features, and the obtained temporal features are spliced ​​to obtain temporal enhancement features. Specifically, multi-source time-series data, fuzzy features, and mechanistic constraint features are first fused using a gating mechanism to generate gated fusion features. The gated fusion calculation method is as follows: the three types of features are uniformly mapped to 128 dimensions, and the gating weights are output as three-dimensional weights through a single-layer perceptron + sigmoid activation. Gated fusion features = ×Multi-source temporal features+ ×Fuzzy Features+ × Mechanism constraint features; then the mechanism constraint features are used as the query vector in multi-head attention, the fuzzy features are used as the key vector in multi-head attention, and the gated fusion features are used as the value vector in multi-head attention to obtain attention features, where the multi-head attention adopts lightweight multi-query attention.

[0045] The lightweight multi-query attention configuration includes: 4 attention heads, a unified 64-dimensional query / key / value mapping, and a scaling factor. The output increases residual connections and layer normalization, and shared key / value pairs reduce computation.

[0046] After acquiring the attention features, a bidirectional long short-term memory network (BiLSTM) is used for temporal modeling of the long-term attention features. The BiLSTM is a single layer with 64 hidden units. A temporal convolutional network (TCN) is used for temporal modeling of the local fluctuations of the attention features. The TCN has 3 layers, 3 convolutional kernels, dilation coefficients of 1 / 2 / 4, and 64 channels. The temporal features acquired after modeling are concatenated to obtain the temporal enhancement features.

[0047] Step 4: After concatenating the temporal enhancement features with the rule activation intensity, combine them with a multilayer perceptron to make a prediction and obtain the first prediction result. Based on the first prediction result, combine the rule weighting of the cyclone adjustment corresponding to the rule activation intensity to obtain the second prediction result. Adjust the shipborne cyclone based on the second prediction result.

[0048] Specifically, the temporal enhancement features are concatenated with the rule activation intensity after dimensionality reduction of the operating conditions to obtain temporal enhancement features with physical priors. Then, the temporal enhancement features with physical priors are input into three multilayer perceptrons to obtain the control quantities of the shipborne hydrocyclone in multiple dimensions as the first prediction result. In this embodiment, the working condition dimensionality reduction is performed by selecting stable working condition, process fluctuation working condition and strong disturbance working condition respectively. That is, the activation intensity and time-series enhancement features of the corresponding working condition are spliced ​​together. For the process fluctuation working condition, the average value of the rule activation intensity of hydrocyclone adjustment rule 2 and hydrocyclone adjustment rule 3 is spliced ​​together with the time-series enhancement features. Here, the three multilayer perceptrons correspond to the expert network settings for stable operating conditions, process fluctuation operating conditions, and strong disturbance operating conditions. Each multilayer perceptron performs a single-push linear mapping for the corresponding operating condition. The expert networks have a unified structure, with a hidden layer of 64 neurons, ReLU activation, and a single-dimensional linear mapping in the output layer. The activation intensity of the rules after dimensionality reduction under the operating conditions is used as the physical prior and spliced ​​with the temporal enhancement features. After prediction, multiple control quantities of the shipborne hydrocyclone are obtained. Among them, the feed flow rate is predicted by the multilayer sensor under the stable operating conditions, the feed pressure is obtained by the multilayer sensor under the process fluctuation conditions, and the bottom flow opening is obtained by the multilayer sensor under the strong disturbance conditions.

[0049] After obtaining the first prediction result, the second prediction result is obtained by weighted fusion of the first prediction result based on the activation intensity of the corresponding hydrocyclone adjustment rule, i.e., the activation intensity of stable operating conditions, process fluctuation conditions, and strong disturbance conditions. The weighted fusion calculation process includes: setting three outputs. Corresponding activation intensity First, normalize Second prediction output = .

[0050] Step 5: Perform process limiting and rate smoothing on the second prediction result to obtain the third prediction result, and adjust the shipborne hydrocyclone based on the third prediction result.

[0051] The third prediction result includes feed flow rate, feed pressure, and underflow opening.

[0052] Specifically, the second prediction result is first truncated based on a preset process threshold. In this embodiment, process thresholds are set for the control quantities of each shipborne hydrocyclone, and the clip function is used for truncation, limiting each control quantity to the upper and lower limits of the process. Process limit range: feed flow rate 0~30 Feed pressure 0.4~1.2MPa, underflow opening 10%~90%.

[0053] Limiting alone cannot suppress command jumps between sampling periods; large differences between adjacent time points will still cause hydraulic impact. Therefore, a clip function is used to limit the range of incremental changes and constrain the maximum adjustment rate of a single-step command, thereby achieving smooth control command rate. Maximum permissible change per period: flow rate. Pressure ±0.05MPa, opening ±5%, any increments exceeding these limits will be directly cut off.

[0054] In this embodiment, sub-objective functions are set based on minimizing control deviation, smoothing control commands, and resisting disturbances, respectively. A multi-objective constraint function is set based on multiple sub-objective functions, and the multi-objective constraint function is expressed by the following formula: ; in, The value of the overall objective function is [0, 10], and the smaller the value, the better the control performance. To minimize the control deviation, the sub-objective function is... To control the smoothness of the instruction sub-objective function, The sub-objective function is the disturbance rejection performance. The corresponding control deviation sub-target weights, The corresponding control command smoothness sub-objective weights, The corresponding sub-objective weights for anti-disturbance performance; , , Let be the weight coefficient, and satisfy... Preferred , , This is used to balance the priorities of various control objectives; In this embodiment, the multi-objective function is iteratively optimized using gradient descent with a learning rate of 0.001. Iteration stops when the loss change is less than 1e-5 for 20 consecutive rounds, with a maximum of 500 rounds. Early stopping is triggered when the loss decrease is less than 1e-5 for 30 consecutive rounds. The training dataset consists of multi-source time-series samples continuously collected from an offshore ore processing vessel, with a sampling period of 1 second and a total of 100,000 samples.

[0055] The sub-objective function for minimizing control deviation is designed based on minimizing the implementation deviations of the shipborne hydrocyclone's feed flow rate, feed pressure, and bottom flow discharge opening, and is expressed by the following formula: ; in, The total duration of timing sampling represents the total number of control cycles; For time-series sampling time; for Constant deviation in ore flow control, i.e. The difference between the actual and set values ​​of the ore flow rate at all times; for Constantly monitor the deviation in ore supply pressure control, i.e. The difference between the actual and set values ​​of the ore supply pressure at any given time; for The deviation of the underflow discharge opening control at any given time, i.e. The difference between the actual value and the set value of the underflow discharge opening at any given time; The sub-objective function for smooth control command setting is designed based on the control command difference between adjacent time points, constraining feed flow rate, feed pressure, and underflow discharge opening, and is expressed by the following formula: ; in, for Time and The amount of change in the flow control command for the ore at all times; for Time and The amount of change in the pressure control command for the ore at all times; for Time and The change in the control command for the underflow discharge opening at any given moment; The sub-objective function for the disturbance rejection setting is designed based on the impact of ship disturbance at any given time on feed flow rate, feed pressure, and bottom flow discharge opening, and is expressed by the following formula: ; in, The disturbance influence coefficient has a value range of [0.1, 0.3] and is used to quantify the amplification effect of hull disturbance on control deviation. for The data on the ship's attitude disturbance at any given time, namely the attitude angles (roll, pitch, and bow roll) when the ship is swaying, characterizes the degree of disturbance of the cyclone due to the swaying of the ship.

[0056] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for intelligent control of a shipborne hydrocyclone with mechanism-aligned MoE, characterized in that, Includes the following steps: Step 1: Collect multi-source time-series data, including ship attitude disturbances, construct a first lookup table of multi-source time-series data and multiple types of learnable fuzzy membership functions, and generate fuzzy features based on multi-source time-series data and the first lookup table. Step 2: Construct a second lookup table corresponding to multi-source time series data and cyclone adjustment rules. Based on the multi-source time series data, combine the first lookup table and the second lookup table to generate rule activation intensity. Obtain mechanism constraint features based on fuzzy features and multilayer perceptron. Step 3: Based on multi-source time series data, fuzzy features, and mechanistic constraint features, gating mechanisms are combined to generate gated fusion features. Based on the gated fusion features, fuzzy features, and mechanistic constraint features, multi-head attention is combined to obtain attention features. Temporal modeling is performed on the long-term and local fluctuations of attention features, and the obtained temporal features are spliced ​​to obtain temporal enhancement features. Step 4: After concatenating the temporal enhancement features with the rule activation intensity, combine them with a multilayer perceptron to make a prediction and obtain the first prediction result. Based on the first prediction result, combine the rule weighting of the cyclone adjustment corresponding to the rule activation intensity to obtain the second prediction result. Adjust the shipborne cyclone based on the second prediction result. Sub-objective functions are set based on minimizing control deviation, smoothing control commands, and resisting disturbances. Multi-objective constraint functions are set based on multiple sub-objective functions. The intelligent control method of shipborne hydrocyclone with mechanism alignment MoE is constrained according to the multi-objective constraint functions.

2. The intelligent control method for shipborne hydrocyclones with MoE alignment according to claim 1, characterized in that, The multi-source time-series data includes: hull attitude disturbance angle, control relative deviation, control deviation change rate, slurry density, feed pressure, feed flow rate fluctuation amplitude, and actuator feedback offset.

3. The intelligent control method for shipborne hydrocyclones with MoE alignment according to claim 1, characterized in that, The first lookup table divides various physical parameters in multi-source time series data into multiple value ranges, and sets a level for each value range and the initial value of the learnable fuzzy membership function corresponding to that level; The learnable fuzzy membership functions include triangular learnable fuzzy membership functions, S-shaped learnable fuzzy membership functions, Z-shaped learnable fuzzy membership functions, and Gaussian learnable fuzzy membership functions.

4. The intelligent control method for shipborne hydrocyclones with MoE alignment according to claim 1, characterized in that, The generation of the fuzzy features includes: determining the level and learnable fuzzy membership function with initial value adjustment based on the value range of a single type of physical parameter in multi-source time series data; calculating the membership degree based on other learnable fuzzy membership functions with unadjusted initial values ​​and learnable fuzzy membership functions with adjusted initial values; and selecting the level corresponding to the learnable fuzzy membership function with the highest membership degree as the level of that type of physical parameter. The adjustment includes correcting the initial value of the learnable fuzzy membership function according to the value range of the single-type physical parameter; The initial value to be adjusted is the initial value that needs to be corrected according to the value range it is in; the initial value that has been adjusted is the initial value that has been fully corrected; and the initial value that has not been adjusted is the initial value that has not yet been corrected. Repeatedly calculating the levels of all physical parameters in the original feature constitutes a fuzzy feature.

5. The intelligent control method for shipborne hydrocyclones with MoE alignment according to claim 1, characterized in that, In the second reference table, the levels of each physical parameter in the adjustment rules for each hydrocyclone are set to correspond to those in the first reference table. The activation intensity of the rule includes the activation intensity of each hydrocyclone adjustment rule in the second lookup table; The acquisition of the rule activation strength of any hydrocyclone adjustment rule includes: confirming the level of each physical parameter based on multi-source time series data and a second lookup table; calculating the membership degree corresponding to each physical parameter by calling the initial value of the corresponding learnable fuzzy membership function in the first lookup table according to the level; and selecting the minimum positive value among the membership degrees corresponding to all physical parameters of the hydrocyclone adjustment rule as the rule activation strength of the hydrocyclone adjustment rule.

6. The intelligent control method for shipborne hydrocyclones with MoE alignment according to claim 5, characterized in that, The process of obtaining the rule activation strength of any hydrocyclone adjustment rule also includes: setting membership thresholds for each physical parameter and truncating the calculated membership.

7. The intelligent control method for shipborne hydrocyclones with MoE alignment according to claim 1, characterized in that, In the multi-head attention in step 3, the mechanistic constraint feature serves as the query vector, the fuzzy feature serves as the key vector, and the gated fusion feature serves as the value vector.

8. The intelligent control method for shipborne hydrocyclones with MoE alignment according to claim 6, characterized in that, The multi-head attention uses lightweight multi-query attention.

9. The intelligent control method for shipborne hydrocyclones with MoE alignment according to claim 1, characterized in that, A bidirectional long short-term memory network is used for temporal modeling of attention features over long time. Temporal modeling is performed using a temporal convolutional network to model the local fluctuations of attention features.

10. The intelligent control method for shipborne hydrocyclones with MoE alignment according to claim 1, characterized in that, In step 4, obtaining the first prediction result includes: The temporal augmentation features are concatenated with the rule activation intensity after dimensionality reduction of the operating conditions to obtain temporal augmentation features with physical priors. Then, the temporal augmentation features with physical priors are input into multiple multilayer perceptrons to obtain the control quantities of multiple dimensions of the shipborne hydrocyclone as the first prediction result. The multiple multilayer perceptrons are set up based on expert networks for different working conditions, and each multilayer perceptron performs a single-pull linear mapping for the corresponding working condition. The working condition dimensionality reduction includes selecting multiple rule activation intensities corresponding to multiple multilayer perceptrons.

11. The intelligent control method for shipborne hydrocyclones with mechanistic alignment of MoE according to claim 10, characterized in that, It also includes step 5: applying process limits to the second prediction result to obtain the third prediction result, and adjusting the shipborne hydrocyclone based on the third prediction result.

12. The intelligent control method for shipborne hydrocyclones with MoE alignment according to claim 10, characterized in that, Step 5 includes performing process limiting on the second prediction result and then rate smoothing to obtain a third prediction result, and adjusting the shipborne hydrocyclone based on the third prediction result.

13. The intelligent control method for shipborne hydrocyclones with MoE alignment according to claim 1, characterized in that, The sub-objective function for minimizing control deviation is designed based on minimizing the implementation deviation of the shipborne hydrocyclone's feed flow rate, feed pressure, and bottom discharge opening. The sub-objective function for smooth setting of control commands is designed based on the difference in control commands constraining the feed flow rate, feed pressure, and underflow discharge opening at adjacent time points; The sub-objective function of the anti-disturbance setting is designed based on the impact of ship disturbance at any time on feed flow rate, feed pressure and bottom flow discharge opening.

Images