Water bottom sand control system and method based on machine learning and model predictive control

Abstract

The application discloses a kind of water bottom sand adding control system and method based on machine learning and model predictive control.The method comprises the following steps: collecting hydraulic parameters and sediment concentration spatial distribution field data in real time; data is processed; the generated real-time state data set is input into a trained machine learning prediction module, and the future N-step sediment concentration vertical profile is predicted; with the deviation between the predicted profile and the target profile as the target, the optimal sand adding flow sequence is solved by rolling optimization under the constraint condition; control instructions are issued to the PLC control system to drive the actuator to adjust the sand adding flow.The system comprises a sand mixing tank, a multi-point concentration sensor array, a tide gauge and a flowmeter, an edge computing node, an intelligent control unit comprising a machine learning prediction module and a model predictive control module, a spiral sand feeder, a sand adding pump, a water bottom sand adding pipe and a PLC control system.The application can achieve full-automatic and high-precision control of sediment concentration vertical profile in flume test, and the sand adding scheme is accurate and reliable.

Description

Technical Field

[0001] This application relates to the field of water conservancy experiments, specifically to a method and system for underwater sediment control based on machine learning and model predictive control. Background Technology

[0002] In the field of hydraulic engineering research, sediment flume physical model testing is a common research method. By analyzing water flow and sediment movement, and based on river dynamics theory, sediment flume physical model testing simulates scouring and deposition near hydraulic structures in natural rivers in a laboratory setting, providing technical reference for the planning, design, and evaluation of actual engineering projects.

[0003] Simulating the water and sediment environment in tidal experiments requires sand addition, which places high demands on the sand concentration and process. Current technologies mostly employ sand addition devices positioned above the water tank and surface, with water and sand manually mixed and stirred in a mixing tank based on experience, and then added to the vicinity of the test area via suspended spraying. This method has significant errors in controlling water and sediment concentration and involves a large amount of manual operation; furthermore, spraying sand addition easily disturbs the water surface, affecting water flow velocity and thus interfering with the accuracy of the scouring and sedimentation test results.

[0004] Furthermore, existing underwater automatic sand-addition control systems often use a single concentration sensor for point measurement, which can only obtain concentration information within the sand-mixing pool or at a fixed location, and cannot reflect the spatial distribution characteristics of sand content in the test section of the flume. In terms of control strategy, PID feedback control is adopted, which is a post-correction mode, and has problems such as response lag and large overshoot. Moreover, it cannot handle multi-variable coupling constraints under tidal non-steady flow conditions, making it difficult to achieve precise control of the vertical profile of sand content. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for underwater sediment addition control based on machine learning and model predictive control, which aims to solve the technical problems in the prior art, such as large sediment concentration error, cumbersome manual operation, water surface disturbance, and lagging control strategy, which makes it impossible to achieve precise profile control.

[0006] According to a first aspect of the embodiments of this application, a smart underwater sand-addition control system based on machine learning and model predictive control is provided, the system comprising: A sand mixing tank is used to prepare and store a water-sand solution of a predetermined concentration. The sensing unit includes a multi-point concentration sensor array for collecting spatial distribution field data of sand volume, and a tide gauge and a current meter for collecting tidal level data and flow velocity data, respectively. The multi-point concentration sensor array is arranged longitudinally along the test section of the tidal flume and has multiple depth measuring points. An edge computing node, communicatively connected to the sensing unit, is used to preprocess the collected spatial distribution data of sediment content, and then fuse it with hydraulic parameters from multiple sources to generate a real-time status dataset; the hydraulic parameters include tidal level data and flow velocity data. The execution unit includes a mixing pump, a water inlet pump, a spiral sand conveyor, an upstream sand pump, a downstream sand pump, a frequency converter that drives the water inlet pump, the spiral sand conveyor, the upstream sand pump, and the downstream sand pump, as well as an upstream bottom sand pipe and a downstream bottom sand pipe buried at the bottom of the tidal channel. The intelligent control unit, deployed on the edge computing node or host computer, includes: A machine learning prediction module is used to predict the vertical profile of sediment concentration for the next N steps based on the real-time state dataset. The model prediction control module is used to perform rolling optimization under predetermined constraints with the goal of minimizing the deviation between the predicted sediment concentration profile sequence and the preset target sediment concentration profile, and solve for the optimal sediment addition flow sequence in the future control time domain. The PLC control system is connected to the intelligent control unit and the execution unit respectively, and is used to receive the current control command in the optimal sand addition flow sequence and drive the frequency converter to adjust the sand addition flow.

[0007] Optionally, both the upstream and downstream underwater sand-adding pipes are steel pipes, each with an upward-facing outlet hole; the outlet hole of the upstream underwater sand-adding pipe faces the same direction as the ebb tide flow, and the outlet hole of the downstream underwater sand-adding pipe faces the same direction as the rise tide flow.

[0008] Optionally, the multi-point concentration sensor array includes four measurement sections, which are respectively located upstream, middle and downstream of the engineering area; each measurement section has three layers of depth measurement points, namely bottom, middle and surface measurement points.

[0009] Optionally, the PLC control system controls the operating frequency of the frequency converter through analog and digital outputs.

[0010] According to a second aspect of the embodiments of this application, a method for underwater sediment addition control based on machine learning and model predictive control is provided, applied to the underwater sediment addition intelligent control system based on machine learning and model predictive control described in the first aspect, the method comprising: Step S10: Real-time acquisition of hydraulic parameter data and sediment concentration spatial distribution field data; wherein, the hydraulic parameters include at least tidal level data and flow velocity data; the sediment concentration spatial distribution field data is acquired by a multi-point concentration sensor array with multiple depth measuring points arranged longitudinally along the tidal flume test section; Step S20: The collected spatial distribution data of sediment content is preprocessed at the edge computing node, and then multi-source data is fused with the hydraulic parameters to generate a real-time status dataset; Step S30: Input the real-time state dataset into a trained machine learning prediction module to predict the vertical profile of sediment concentration in the next N steps and obtain the predicted sediment concentration profile sequence. Step S40: The model prediction control module aims to minimize the deviation between the predicted sediment concentration profile sequence and the preset target sediment concentration profile, and performs rolling optimization under predetermined constraints to solve for the optimal sediment addition flow sequence in the future control time domain. Step S50: Send the current control command in the optimal sand addition flow sequence to the PLC control system to drive the actuator to adjust the upstream and downstream sand addition flow, thereby realizing closed-loop control of the vertical profile of sand content.

[0011] Optionally, the multi-point concentration sensor array is arranged with four measurement sections along the longitudinal direction of the tidal flume test section. Each measurement section is equipped with at least three layers of depth measurement points: bottom layer, middle layer, and surface layer, so as to form a real-time perception of the three-dimensional spatial distribution field of sediment content.

[0012] Optionally, the collected spatial distribution data of sediment concentration is preprocessed, and then fused with the hydraulic parameters from multiple sources to generate a real-time status dataset, including: Spatial domain moving mean filtering is applied to the spatial distribution field data of sediment content to suppress measurement noise, automatic compensation is performed for concentration sensor drift, and outlier removal is performed based on the spatial consistency of multi-section data. The preprocessed concentration data, tidal data, and flow velocity data are aligned and fused on the time axis to generate a unified time series dataset as a real-time status dataset.

[0013] Optionally, the machine learning prediction module is a long short-term memory network model; the input features of the long short-term memory network model include: spatial distribution data of sediment concentration and hydraulic parameters of the past T steps; the output of the long short-term memory network model is the predicted value of sediment concentration at each depth measuring point of each measurement section in the next N steps.

[0014] Optionally, the training method for the machine learning prediction module includes: During the offline training phase, network parameters were trained using historical water tank test data through supervised learning. During the online fine-tuning phase, a sliding window approach was used during the experiment. After each tidal cycle, the model was incrementally updated using real-time sensor data to adapt to changes in current water and sediment conditions.

[0015] Optionally, the objective function for the rolling optimization of the model prediction control module is: ; in The objective function for rolling optimization is... For the prediction step number, To predict the total number of steps in the time domain, For the first Predict the sediment concentration profile vector step by step. For the first Step target sediment content profile vector For the first The vector of changes in the flow rate of the step-gaza dam. For the first Step-by-step flow vector of each sand pipe The square of the Euclidean norm. Penalize weights for changes in traffic; The predetermined constraints include: upper and lower limits of flow rate for each sand-adding pipe, concentration range of sand-mixing pool, and the constraint of matching sand-adding timing with tidal phase.

[0016] The technical solutions provided by the embodiments of this application may include the following beneficial effects: By collecting real-time data on the spatial distribution of sand content from an array of multi-point concentration sensors, a three-dimensional, array-based perception of sand content in the engineering area across multiple longitudinal sections and multiple vertical levels was achieved. This overcomes the shortcomings of single-point measurement, which provides limited information and fails to reflect the true vertical distribution of sand content, thus laying a data foundation for high-precision profile control.

[0017] By inputting the preprocessed real-time state dataset into the machine learning prediction module, the vertical profile of sediment concentration in the next N steps is predicted, giving the control system the ability to predict the evolution trend of sediment concentration in advance. This transforms the control mode from the ex-post correction of traditional PID to ex-ante active regulation, fundamentally solving the problems of slow control response and large dynamic deviation in large time-delay systems, and significantly improving the control timeliness and accuracy under tidal non-steady flow conditions.

[0018] By using the model predictive control module to minimize the deviation between the predicted sediment concentration profile and the target profile, rolling optimization is performed under constraints to solve for the optimal sediment addition flow sequence in the future control time domain. This allows for a comprehensive balance between control accuracy and cost within a finite time domain, and automatically satisfies various constraints such as upper and lower limits of flow, sediment concentration in the mixing pool, and tidal phase matching. This not only achieves precise tracking control of the vertical sediment concentration profile but also effectively suppresses drastic fluctuations in sediment addition flow, ensuring the stability of the sediment addition process and the long-term reliable operation of the actuators, thereby improving the overall control quality of the system under time-varying and multi-constraint conditions.

[0019] By sending the current control command to the PLC control system to drive the actuator to adjust the upstream and downstream sand addition flow, a closed-loop control is formed. This, combined with the method of adding sand from the bottom of the water, avoids the disturbance to the water surface caused by traditional suspended spraying sand addition, maintains the natural state of the free surface of the water flow and the stability of the boundary conditions, thereby significantly improving the simulation accuracy and reliability of the tidal scouring and sedimentation physical model test.

[0020] In summary, this application achieves fully automatic, high-precision, and low-disturbance closed-loop control of the vertical profile of sediment content in tidal flume experiments through the synergistic effect of multi-dimensional perception, intelligent prediction, and constraint optimization control.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0023] Figure 1 This is an exemplary embodiment illustrating the overall structure of an intelligent underwater sand-adding control system based on machine learning and model predictive control.

[0024] Figure 2 This is a schematic diagram of a PLC control system and a sand mixing tank according to an exemplary embodiment.

[0025] Figure 3 This is a top view of a sand tank according to an exemplary embodiment.

[0026] Figure 4 This is a schematic diagram illustrating a method for adding sand to the bottom of a sand tank according to an exemplary embodiment.

[0027] Figure 5 This is a side view of the upstream and downstream sand supply pipes in a test water tank, according to an exemplary embodiment.

[0028] Figure 6 This is a schematic diagram of a multi-point concentration sensor array arrangement according to an exemplary embodiment.

[0029] Figure 7 This is a block diagram illustrating the principle of ML prediction and MPC rolling optimization control according to an exemplary embodiment.

[0030] The attached figures are labeled as follows: 1. Sand mixing tank; 2. Concentration sensor array; 3. Tide level gauge; 4. Flow meter; 5. Tidal flume; 6. Agitator pump; 7. Inlet pump; 8. Spiral sand feeder; 9. Upstream sand pump; 10. Downstream sand pump; 11. Frequency converter; 12. Upstream underwater sand adding pipe; 13. Downstream underwater sand adding pipe; 14. Edge computing node; 15. Host computer; 16. Machine learning prediction module; 17. Model prediction control module; 18. PLC control system; 19. Water adding pipe; 20. Electromagnetic flow meter; 21. Sand adding tank; 22. Upstream outlet; 23. Downstream outlet; 24. Upstream electromagnetic flow meter; 25. Downstream electromagnetic flow meter; 26. Concentration sensor. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this application clearer, the application 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 only for explaining this application and are not intended to limit this application.

[0032] It should be noted that when a component is referred to as being "set on" or "connected to" another component, it can be directly set on or connected to the other component, or it can be located in an intermediate component. The terms "upstream," "downstream," "bottom layer," "middle layer," "surface layer," "longitudinal," etc., used herein to indicate orientation or positional relationships are based on the orientation or positional relationships shown by the direction of water flow or water depth in a water tank test, and are used only for the convenience of describing this application and simplifying the description, and are not intended to indicate or imply that the device or component 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 this application.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0034] This application provides a water-based sediment addition control system and method based on machine learning and model predictive control, aiming to solve the technical problems in existing water tank physical model tests, such as large sediment concentration errors, cumbersome manual operation, water surface disturbance caused by the sediment addition process, and the inability of single sensor point measurement and PID post-correction control strategy to accurately control the vertical profile of sediment content.

[0035] refer to Figures 1-7This invention provides an intelligent underwater sand-addition control system based on machine learning and model predictive control, which may include: a sand-mixing tank 1, a sensing unit, an edge computing node 14, an execution unit, an intelligent control unit, and a PLC control system 18. The sand-mixing tank 1 is used to prepare and store a water-sand solution of a predetermined concentration. The sensing unit includes a multi-point concentration sensor array 2 for collecting spatial distribution field data of sand content, and a tide gauge 3 and a flow meter 4 for collecting tidal level data and flow velocity data, respectively. The multi-point concentration sensor array 2 is arranged longitudinally along the test section of the tidal flume 5 and has multiple depth measuring points. The edge computing node 14 is communicatively connected to the sensing unit and is used to preprocess the collected spatial distribution field data of sand content, and then perform multi-source data fusion with hydraulic parameters to generate a real-time status dataset. The hydraulic parameters include tidal level data and flow velocity data. The execution unit includes a stirring pump 6, an inlet pump 7, a spiral sand feeder 8, an upstream sand-adding pump 9, a downstream sand-adding pump 10, and a pump that drives the inlet pump 7. The system includes a spiral sand feeder 8, an upstream sand pump 9, and a downstream sand pump 10, all equipped with frequency converters 11, as well as an upstream underwater sand pumping pipe 12 and a downstream underwater sand pumping pipe 13 buried at the bottom of the tidal flume 5. The intelligent control unit is deployed on the edge computing node 14 or the host computer 15 and includes a machine learning prediction module 16 and a model prediction control module 17. The machine learning prediction module 16 is used to predict the vertical profile of sediment concentration for the next N steps based on the real-time state dataset. The model prediction control module 17 is used to perform rolling optimization under predetermined constraints with the goal of minimizing the deviation between the predicted sediment concentration profile sequence and the preset target sediment concentration profile, and solve for the optimal sand pumping flow sequence in the future control time domain. The PLC control system 18 is connected to the intelligent control unit and the execution unit respectively, and is used to receive the current control command in the optimal sand pumping flow sequence and drive the frequency converter 11 to adjust the sand pumping flow and avoid water surface disturbance, thereby realizing fully automatic and high-precision closed-loop control of the vertical profile of sediment concentration in the engineering area.

[0036] In one embodiment, the sand mixing tank 1 is used to prepare and store a water-sand solution of a certain concentration. Water is added through a water inlet pipe 19, which is connected to a water pump, and the inlet flow rate is measured by an electromagnetic flow meter 20. The sand adding box 21 has an inverted concave structure and is connected to a spiral sand feeder 8. The spiral sand feeder 8 adjusts its speed through a frequency converter 11 to achieve precise addition of dry sand. The stirring pump 6 located at the bottom of the sand mixing tank 1 runs continuously to ensure uniform mixing of the water-sand solution. The concentration sensor 26 in the sand mixing tank 1 monitors the concentration of the sand mixing tank 1 in real time and feeds the signal back to the PLC control system 18 to achieve closed-loop control of the concentration of the sand mixing tank 1.

[0037] In one embodiment, the upstream underwater sand-adding pipe 12 has an upward-facing upstream outlet hole 22, and the downstream underwater sand-adding pipe 13 has an upward-facing downstream outlet hole 23; the upstream outlet hole 22 faces the same direction as the ebb tide flow, and the downstream outlet hole 23 faces the same direction as the rise tide flow. In other embodiments, the outlet hole can be not only a circular hole, but also a narrow slit opened along the pipe axis to produce a more uniform linear outflow. The number and arrangement of the outlet holes can also be multiple rows arranged in an alternating pattern to enhance the initial mixing effect of the water-sand mixture after it exits the pipe.

[0038] Regarding the pre-installation method, the upstream underwater sand-adding pipe 12 is pre-installed upstream of the project area, excavated and buried under the water trough. The burial depth is such that the ground intersects the upstream underwater sand-adding pipe 12 at 45° east of north and 45° west of north of the steel pipe axis. The upstream outlet 22 is located approximately 30° east of north of the axis of the installed upstream underwater sand-adding pipe 12, ensuring that the water-sand mixture is added from the bottom, and the direction of the entering water-sand mixture is consistent with the direction of the ebb tide. The downstream underwater sand-adding pipe 13 is pre-installed downstream of the project area, excavated and buried under the water trough. The burial depth is such that the ground intersects the downstream underwater sand-adding pipe 13 at 45° east of north and 45° west of north of the steel pipe axis. The downstream outlet 23 is located approximately 30° west of north of the axis of the installed steel pipe, ensuring that the water-sand mixture is added from the bottom, and the direction of the entering water-sand mixture is consistent with the direction of the rising tide. By adding sand in stages, with the upstream sand-adding pipe operating during low tide and the downstream sand-adding pipe operating during high tide, precise control of the sand content can be achieved during both high and low tide stages.

[0039] An upstream electromagnetic flow meter 24 is installed on the upstream underwater sand filling pipe 12, and a downstream electromagnetic flow meter 25 is installed on the downstream underwater sand filling pipe 13, to measure the flow rate of the upstream and downstream sand filling pipes respectively.

[0040] like Figures 2 to 5 As shown, the upstream sand-adding pump 9 pressurizes and transports the water-sand solution in the sand-mixing pool 1 to the bottom of the upstream water tank through the upstream underwater sand-adding pipe 12, with the flow rate measured by the upstream electromagnetic flowmeter 24. The downstream sand-adding pump 10 pressurizes and transports the water-sand solution in the sand-mixing pool 1 to the bottom of the downstream water tank through the downstream underwater sand-adding pipe 13, with the flow rate measured by the downstream electromagnetic flowmeter 25. The water-sand solution overflows evenly upwards from the outlet of the underwater sand-adding pipe and enters the water body in the sand-laying project area. Since the sand-adding process is completed entirely underwater, it does not disrupt the water surface morphology, effectively avoiding the water surface disturbance caused by existing suspended spraying methods, ensuring the stability of the water flow boundary conditions, and significantly improving the accuracy and reliability of the scouring and silting test.

[0041] In one embodiment, such as Figure 6As shown, the multi-point concentration sensor array 2 is arranged longitudinally along the experimental section of the tidal flume 5, with four measurement sections: one upstream of the sand-laying project area, two in the middle, and one downstream. Each measurement section has three layers of depth measurement points: bottom, middle, and surface, totaling 12 concentration acquisition channels. The bottom measurement point is used to sense the sediment concentration in the near-bottom high-concentration zone, the surface measurement point is used to sense the sediment concentration in the near-surface low-concentration zone, and the middle measurement point is used to sense the sediment concentration in the intermediate transition zone. This crisscrossing array layout allows for a relatively complete reconstruction of the three-dimensional spatial distribution of sediment concentration within the sand-laying project area. The concentration sensor can be a photoelectric turbidity-concentration composite sensor, which, after indoor calibration, can output the volumetric sediment concentration. The concentration sensor array 2 communicates with the edge computing node 14 via an RS485 bus, with a sampling frequency of no less than 1Hz to ensure the real-time and continuous data acquisition.

[0042] In other embodiments, depending on the actual length of the tidal flume 5, the size of the sand-laying engineering area, and the accuracy requirements of the spatial resolution of the experiment, the number of measurement cross-sections can be increased to 5, 6, or more, and the number of depth measurement points can be increased to 4 or more, in order to obtain a sand concentration distribution field with higher spatial resolution. The specific locations of the cross-sections can also be non-uniformly distributed, for example, denser distribution near the sand-laying engineering area where the expected sand concentration gradient is large, and sparse distribution upstream and downstream of the engineering area, in order to maximize the amount of spatially sensed information with a limited number of sensors. These adjustments to the arrangement scheme all fall within the scope of the multi-point concentration sensor array 2 of this application, the core of which is to form a multi-section, multi-level sensing network that can reflect the three-dimensional spatial distribution of sand concentration, providing sufficient spatial dimension data support for subsequent prediction and control. By increasing the sensor density, the reconstruction accuracy of the sand concentration field and the profile control effect can be further improved.

[0043] In one embodiment, the PLC control system 18 controls the operating frequency of the frequency converter 11 connected to the water inlet pump 7, the spiral sand conveyor 8, the upstream sand pump 9, and the downstream sand pump 10 through analog and digital outputs.

[0044] This invention also provides a method for underwater sediment addition control based on machine learning and model predictive control. This method is applied to the above-mentioned underwater sediment addition control system based on machine learning and model predictive control. The method includes the following steps S10 to S50.

[0045] Step S10: Real-time acquisition of hydraulic parameter data and spatial distribution data of sediment concentration.

[0046] During the scouring and sedimentation test of tidal flume 5, it is necessary to acquire real-time hydrodynamic conditions and sediment distribution within the test section. The hydraulic parameters include at least tidal level data and flow velocity data, collected by a tidal level gauge 3 and a current meter 4, respectively, installed within the test section of tidal flume 5. The tidal level gauge 3 measures the instantaneous elevation change of the water surface in the test section, and the current meter 4 measures the instantaneous flow velocity. By collecting tidal level and flow velocity data, the real-time dynamic conditions of the unsteady tidal flow can be accurately determined, providing crucial hydrodynamic input characteristics for subsequent sediment concentration prediction.

[0047] The spatial distribution data of sediment concentration was acquired by a multi-point concentration sensor array 2 with multiple depth measuring points across multiple measurement sections longitudinally deployed along the experimental section 5 of the tidal flume. Unlike the single-point concentration acquisition method in existing technologies, the multi-point concentration sensor array 2 can simultaneously acquire sediment concentration information at multiple locations along the water flow direction and at multiple levels along the water depth direction, thus forming a real-time perception of the three-dimensional spatial distribution field of sediment concentration. This comprehensive perception capability allows the control system to grasp the complete situation of sediment distribution within the engineering area, rather than relying solely on concentration data from a single point, laying a data foundation for achieving precise profile control.

[0048] Step S20: The collected spatial distribution data of sediment content is preprocessed at the edge computing node 14, and then multi-source data is fused with the hydraulic parameters to generate a real-time status dataset.

[0049] like Figure 7 As shown, the edge computing node 14 is located on the end side close to the test site to process the raw data collected by the sensing unit, so as to reduce data transmission latency and cloud computing burden and improve the real-time response capability of the control system.

[0050] Specifically, firstly, spatial domain moving average filtering is performed on the 12-channel concentration sensor data of the multi-point concentration sensor array 2. By taking the average value of multiple consecutive sampling points, high-frequency random noise introduced during the measurement process is suppressed, thereby improving the signal-to-noise ratio of the data.

[0051] Secondly, automatic compensation is performed for baseline drift that may occur in the concentration sensor due to long-term operation. Corrections are made based on preset drift models or periodic calibration parameters to ensure the long-term stability of the measured values.

[0052] Furthermore, outlier removal is performed based on the spatial consistency of multi-section data. For example, when the concentration value of a certain measuring point shows a significant non-physical abrupt change in spatial terms compared to the data of adjacent measuring points, it is identified as an outlier and removed or corrected by interpolation to ensure data quality.

[0053] Finally, the preprocessed concentration data, tidal data, and flow velocity data are strictly aligned on a unified time axis and fused using algorithms such as timestamp matching interpolation to generate a multi-dimensional time series dataset, which is the real-time status dataset. This dataset provides the subsequent machine learning prediction module 16 and model prediction control module 17 with input information that is uniform in format, reliable in quality, and synchronized in time.

[0054] Step S30: Input the real-time state dataset into a trained machine learning prediction module 16 to predict the vertical profile of sediment concentration in the next N steps, and obtain the predicted sediment concentration profile sequence.

[0055] Existing PID control relies entirely on the feedback deviation at the current moment for adjustment, which is essentially a reactive correction. When dealing with large time-varying systems such as tidal unsteady flows, problems such as control response lag and large overshoot are difficult to avoid. This step introduces a machine learning prediction module 16, which endows the control system with the ability to predict the future evolution trend of sediment concentration, thereby transforming the control mode from reactive correction to proactive prediction and control.

[0056] In this embodiment, the machine learning prediction module 16 adopts a Long Short-Term Memory (LSTM) network structure, which includes an input layer, multiple LSTM hidden layers, a fully connected layer, and an output layer. This structure can effectively capture long-term nonlinear dynamic dependencies in sediment concentration time series. LSTM is a special type of recurrent neural network, adept at processing and predicting long-term dependencies in time series data, and is highly suitable for predicting sediment concentration evolution during tidal processes. The input feature vector of the LSTM network model includes: sediment concentration sequences at multiple points over the past T steps, tidal processes, flow velocity processes, flow sequences from each sediment addition pipe, and concentration in sediment mixing pond 1. These input features comprehensively cover hydrodynamic factors, sediment source factors, and historical state factors that influence sediment concentration changes. By aligning these multi-source heterogeneous information within the same time window and jointly inputting them into the model, the LSTM network can automatically learn the complex coupling mapping relationship between hydrodynamic transport, sediment deposition and diffusion, and artificial sediment addition from the data.

[0057] The output vector of the Long Short-Term Memory (LSTM) network model represents the predicted sediment concentration at each depth point of each measurement section over the next N steps. These predicted values, arranged spatially, constitute a predicted sediment concentration profile. N consecutive predicted profiles form the predicted sediment concentration profile sequence. The number of prediction steps N is determined based on a combination of tidal and control cycles, typically set to 10 to 30 steps, corresponding to one or more tidal cycles, to ensure that the model prediction control module 17 can make optimization decisions over a sufficiently long time domain.

[0058] In terms of model training, this embodiment adopts a two-stage strategy combining offline training and online fine-tuning. During the offline training stage, complete sets of tidal test data accumulated from previous experiments on the flume experimental platform are collected and organized. A training sample set is constructed according to the aforementioned input-output format. Mean squared error is used as the loss function, and supervised learning training of the network parameters is conducted through a time-based backpropagation algorithm and an adaptive moment estimation optimizer until the model's prediction error on the validation set converges to a low level. The offline-trained model serves as the basic prediction model at the start of the experiment. During the online fine-tuning stage, after the experiment officially begins, each time the system completes a complete tidal cycle (e.g., from the start of one ebb tide to the start of the next), incremental training samples are constructed using the real-time sensor data stored in edge computing node 14 within that cycle. A sliding window approach is used, replacing the earliest cycle sample with the latest tidal cycle sample, and incrementally updating the model parameters with a small learning rate. Through this periodic fine-tuning, the model can adapt to the slow changes in water and sediment conditions that may occur during the experiment due to factors such as water temperature changes, sediment particle size sorting, and slow drift in sensor characteristics, maintaining high prediction accuracy at all times. The size of the sliding window can be set to include the most recent K tidal cycles. The value of K is determined based on the actual test duration and the rate of change of conditions, with a typical value of 5 to 10 cycles.

[0059] Through the joint design of the above feature selection, network structure and two-stage training strategy, the machine learning prediction module 16 has the ability to make online high-precision predictions of the evolution of sediment content profile under complex tidal unsteady flow conditions, providing reliable core prediction model support for the model prediction control module 17.

[0060] Step S40: The model prediction control module 17 aims to minimize the deviation between the predicted sediment concentration profile sequence and the preset target sediment concentration profile, and performs rolling optimization under predetermined constraints to solve for the optimal sediment addition flow sequence in the future control time domain.

[0061] After obtaining the predicted results of the sediment concentration evolution over the next N steps in step S30, the model predictive control module 17 is responsible for deciding the optimal control action sequence in this step. The core idea of ​​model predictive control is to predict the future output using the system model in each control cycle, solve a finite-time open-loop optimal control problem online, apply the first element of the obtained control sequence to the system, and repeat this process in the next cycle, thereby achieving rolling optimization and feedback correction.

[0062] The goal of rolling optimization is to find a time-to-time ratio. Beginning, Include Step control sequence , , ..., This enables the system to predict in the time domain. The predicted sediment concentration profile within the area should be as close as possible to the preset target sediment concentration profile, while ensuring that the control actions are not too drastic.

[0063] To achieve this goal, the objective function used for rolling optimization in this embodiment is defined as follows: ; in, The total number of prediction time-domain steps is consistent with step S30. For the first The predicted sediment concentration profile vector is formed by sequentially arranging the predicted sediment concentration values ​​of each measuring point at each depth of each section. The target sediment concentration profile vector is pre-set based on the requirements of the vertical distribution of sediment concentration according to the prototype measured data or engineering design. Its typical feature is that the sediment concentration near the bottom layer is higher than that of the surface layer, so as to conform to the vertical distribution law of sediment concentration in natural tidal channels.

[0064] First item The sum of squared Euclidean distances between the predicted profile and the target profile is used to quantify their deviation, so that the predicted profile tracks the target profile to the greatest extent possible. It is a vector whose elements are the first and second elements. The predicted sediment concentration values ​​at 12 locations across 4 measurement sections with 3 depths each represent the system's predicted future state. This is a target sediment concentration profile vector of the same dimension. This target profile is not a single concentration value, but a distribution curve with a specific gradient along the water depth direction. For example, let the target sediment concentration at the bottom be... The surface target sand content is ,and The intermediate layer is interpolated using linear or exponential laws. This vertical distribution characteristic conforms to the suspension transport pattern of sediment in natural tidal channels. The calculation is the sum of squared Euclidean distances between the predicted profile and the target profile, representing the nth The step predicts the overall deviation between the profile and the target profile. For all The summation step accumulates the overall control accuracy over the entire prediction time domain. This first term is the core performance indicator of the control system, which minimizes the vertical profile of sediment concentration in the engineering area to approximate the preset target.

[0065] Second item Penalties for changes in Gaza traffic flow. For the first The flow vector of each sand supply pipe in this system includes two components: the upstream sand supply pipe flow and the downstream sand supply pipe flow. For the first The vector of changes in sand supply flow rate in each step has the physical meaning of the increase or decrease in the flow rate of each sand supply pipe between adjacent control cycles. The calculation is based on the squared Euclidean distance of the flow rate change vector, which represents the severity of the control action at that step. The penalty weight for flow change is a non-negative, adjustable scalar parameter. It is calculated by summing over all N steps and then multiplying by... This yields the total control cost over the entire prediction time domain. The second term aims to suppress large jumps in control commands, ensuring smooth changes in the Gaza flow. By incorporating this term into the objective function, the Model Predictive Control Module 1725 automatically balances tracking accuracy and control stability during optimization: if... A larger setting will lead to a more gradual generation of Gaza flow sequences in the optimization solution, even if the compensation for profile deviations is slightly slower; if Setting it to a smaller value will lead to more aggressive adjustments to the sand addition flow rate in the optimization solution to quickly eliminate profile deviations, but this may result in frequent actuator movements and fluctuations in the sand addition process. The specific value can be adjusted based on the dynamic response characteristics of the sand pump in actual tests and the requirements for control stability. It is usually determined through numerical simulation and pre-tests.

[0066] The predetermined constraints include the following three categories: First, there are upper and lower limits for the flow rate of each sand-adding pipe. That is, the instantaneous flow rate of each sand-adding pipe cannot exceed the upper limit of the physical conveying capacity of the corresponding sand-adding pump and pipeline, nor can it be lower than the lower limit required to maintain stable conveying.

[0067] Specifically, the first category is the physical constraints on Gaza flow. This includes the upper limit constraint on Gaza flow, i.e. The upper limit is determined by factors such as the rated power of the upstream sand-adding pump 9 and the downstream sand-adding pump 10, the pipe diameter and pressure-bearing capacity of the sand-adding pipe, and the range of the electromagnetic flowmeter 20; as well as the lower limit constraint on the sand-adding flow rate, i.e. The lower limit must ensure that the water-sand solution can still be transported at a sufficient flow rate in the pipeline to prevent sediment settling and clogging. By setting upper and lower flow rate constraints, it is ensured that the optimal flow sequence obtained is physically feasible.

[0068] Second, the concentration range of the sand mixing tank 1 is constrained, that is, the concentration of the water-sand solution in the sand mixing tank 1 must be maintained within a reasonable range that can be stably transported and meets the test requirements.

[0069] Specifically, the second type involves concentration constraints in sand mixing tank 1. Sand mixing tank 1 serves as the supply source of the water-sand solution, and its internal concentration... It needs to be within the allowed range, that is Lower limit Ensure the sand-adding solution has a certain sediment content to avoid ineffective sand addition by conveying almost entirely clear water; upper limit To ensure the solution is pumpable, excessive concentration can lead to pipe blockage or increased pump wear. Concentration constraints in sand mixing tank 1 are achieved by introducing the mass balance equation of sand mixing tank 1 as an internal model of model predictive control module 17 and limiting the range of its state variables.

[0070] Third, there is the constraint of matching the timing of sediment addition with the tidal phase. Specifically, during the ebb tide phase, the upstream sediment addition pipe acts as the main actuator for sediment addition; during the flood tide phase, the downstream sediment addition pipe acts as the main actuator for sediment addition. This constraint ensures that the direction of sediment addition is consistent with the direction of sediment transport by the natural tidal flow, which conforms to the physical laws of sediment movement and effectively reduces the interference of artificial sediment addition on the natural flow field and sediment field.

[0071] Specifically, the third type is the tidal phase matching constraint. In tidal experiments, the direction of water flow reverses periodically with the rise and fall of the tide. To ensure that the artificial sediment addition is consistent with the natural sediment transport direction, this system adopts a phased sediment addition strategy between upstream and downstream. Specifically, this constraint stipulates that during the ebb tide phase, i.e., when the tide level drops and the water flows from upstream to downstream, the flow rate of the downstream sediment addition pipe is constrained to zero or a minimum value, and only the upstream sediment addition pipe is allowed to add sediment; during the flood tide phase, i.e., when the tide level rises and the water flows from downstream to upstream, the flow rate of the upstream sediment addition pipe is constrained to zero or a minimum value, and only the downstream sediment addition pipe is allowed to add sediment. The tidal phase is determined based on real-time tidal data collected by the tide gauge 3. The phase at the current moment is determined by detecting the inflection point of the tide level changing from rising to falling or from falling to rising, supplemented by a water level threshold. By introducing this constraint, the timing of sediment addition is synchronized with the natural tidal sediment transport process, reducing the interference of artificial sediment addition on the natural sediment field, and making the scouring and deposition simulation in the experiment closer to the actual engineering situation.

[0072] Within each control cycle, the model predictive control module 17 calls the quadratic programming solver to solve the constrained quadratic objective function online, calculating the optimal sand-adding flow sequence for the next N control steps in the time domain. This sequence includes the target flow values ​​for each step of the upstream and downstream sand-adding pipes. Only the first element of this sequence, corresponding to the control command for the current control cycle, is sent to the PLC control system 18 for execution. In the next control cycle, the system re-acquires sensor data, updates the status, and re-invokes prediction and optimization, repeating the entire rolling optimization process. This rolling time-domain optimization mechanism gives the system natural robustness to model prediction errors and external disturbances. Combined with feedback correction, even if the prediction of the machine learning prediction module 1624 has a certain deviation, the closed-loop control can still remain stable and gradually converge to the vicinity of the target profile.

[0073] Step S50: Send the current control command in the optimal sand addition flow sequence to the PLC control system 18 to drive the actuator to adjust the upstream and downstream sand addition flow, thereby realizing closed-loop control of the vertical profile of sand content.

[0074] After solving the optimal Gaza flow sequence for the next N steps in step S40, the model predictive control module 17 extracts only the first control action in the sequence, that is, the control command that should be executed immediately at the current moment, and sends it to the PLC control system 18.

[0075] The PLC control system 18, acting as the central execution hub of the entire system, communicates with the edge computing node 14 via the Modbus TCP protocol. Upon receiving control commands, the PLC control system 18 generates corresponding control signals through its analog and digital output modules, precisely adjusting the operating frequencies of the four frequency converters 11 connected to the inlet pump 7, the spiral sand conveyor 8, the upstream sand pump 9, and the downstream sand pump 10. Changes in the frequency converters 11 directly adjust the motor speeds of each pump and the spiral sand conveyor 8, thereby precisely controlling the sand flow rate in each pipeline.

[0076] After the PLC control system 18 completes the control actions of the current control cycle, the next control cycle will restart from step S10, forming a closed-loop control cycle of "sensing-prediction-optimization-execution-re-sensing". Through this continuously operating closed-loop mechanism, the system can perform real-time, dynamic, and precise control of the vertical profile of sediment content.

[0077] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only.

[0078] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A smart underwater sand-addition control system based on machine learning and model predictive control, characterized in that, The system includes: A sand mixing tank is used to prepare and store a water-sand solution of a predetermined concentration. The sensing unit includes a multi-point concentration sensor array for collecting spatial distribution field data of sand volume, and a tide gauge and a current meter for collecting tidal level data and flow velocity data, respectively. The multi-point concentration sensor array is arranged longitudinally along the test section of the tidal flume and has multiple depth measuring points. An edge computing node, communicatively connected to the sensing unit, is used to preprocess the collected spatial distribution data of sediment content, and then fuse it with hydraulic parameters from multiple sources to generate a real-time status dataset; the hydraulic parameters include tidal level data and flow velocity data. The execution unit includes a mixing pump, a water inlet pump, a spiral sand conveyor, an upstream sand pump, a downstream sand pump, a frequency converter that drives the water inlet pump, the spiral sand conveyor, the upstream sand pump, and the downstream sand pump, as well as an upstream bottom sand pipe and a downstream bottom sand pipe buried at the bottom of the tidal channel. The intelligent control unit, deployed on the edge computing node or host computer, includes: A machine learning prediction module is used to predict the vertical profile of sediment concentration for the next N steps based on the real-time state dataset. The model prediction control module is used to perform rolling optimization under predetermined constraints with the goal of minimizing the deviation between the predicted sediment concentration profile sequence and the preset target sediment concentration profile, and solve for the optimal sediment addition flow sequence in the future control time domain. The PLC control system is connected to the intelligent control unit and the execution unit respectively, and is used to receive the current control command in the optimal sand addition flow sequence and drive the frequency converter to adjust the sand addition flow.

2. The system according to claim 1, characterized in that, Both the upstream and downstream underwater sand-adding pipes are steel pipes, each with an upward-facing outlet. The outlet of the upstream underwater sand-adding pipe faces the same direction as the ebb tide, while the outlet of the downstream underwater sand-adding pipe faces the same direction as the rise tide.

3. The system according to claim 1, characterized in that, The multi-point concentration sensor array includes four measurement sections, which are respectively located upstream, middle and downstream of the engineering area; each measurement section has three layers of depth measurement points, namely bottom, middle and surface measurement points.

4. The system according to claim 1, characterized in that, The PLC control system controls the operating frequency of the frequency converter through analog and digital outputs.

5. A method for underwater sediment addition control based on machine learning and model predictive control, applied to the underwater sediment addition intelligent control system based on machine learning and model predictive control as described in any one of claims 1-4, characterized in that, The method includes: Step S10: Real-time acquisition of hydraulic parameter data and sediment concentration spatial distribution field data; wherein, the hydraulic parameters include at least tidal level data and flow velocity data; the sediment concentration spatial distribution field data is acquired by a multi-point concentration sensor array with multiple depth measuring points arranged longitudinally along the tidal flume test section; Step S20: The collected spatial distribution data of sediment content is preprocessed at the edge computing node, and then multi-source data is fused with the hydraulic parameters to generate a real-time status dataset; Step S30: Input the real-time state dataset into a trained machine learning prediction module to predict the vertical profile of sediment concentration in the next N steps and obtain the predicted sediment concentration profile sequence. Step S40: The model prediction control module aims to minimize the deviation between the predicted sediment concentration profile sequence and the preset target sediment concentration profile, and performs rolling optimization under predetermined constraints to solve for the optimal sediment addition flow sequence in the future control time domain. Step S50: Send the current control command in the optimal sand addition flow sequence to the PLC control system to drive the actuator to adjust the upstream and downstream sand addition flow, thereby realizing closed-loop control of the vertical profile of sand content.

6. The intelligent underwater sand-addition control method based on machine learning and model predictive control according to claim 5, characterized in that, The multi-point concentration sensor array is arranged longitudinally along the tidal flume test section with four measurement sections. Each measurement section is equipped with at least three layers of depth measurement points: bottom, middle and surface, to form a real-time perception of the three-dimensional spatial distribution field of sediment content.

7. The intelligent underwater sand-addition control method based on machine learning and model predictive control according to claim 5 or 6, characterized in that, The collected spatial distribution data of sediment concentration is preprocessed, and then multi-source data fusion is performed with the hydraulic parameters to generate a real-time status dataset, including: Spatial domain moving mean filtering is applied to the spatial distribution field data of sediment content to suppress measurement noise, automatic compensation is performed for concentration sensor drift, and outlier removal is performed based on the spatial consistency of multi-section data. The preprocessed concentration data, tidal data, and flow velocity data are aligned and fused on the time axis to generate a unified time series dataset as a real-time status dataset.

8. The intelligent underwater sand-addition control method based on machine learning and model predictive control according to claim 5, characterized in that, The machine learning prediction module is a long short-term memory network model; the input features of the long short-term memory network model include: spatial distribution data of sediment concentration and hydraulic parameters of the past T steps; the output of the long short-term memory network model is the predicted value of sediment concentration at each depth measuring point of each measurement section in the next N steps.

9. The intelligent underwater sand-addition control method based on machine learning and model predictive control according to claim 5, characterized in that, The training methods for the machine learning prediction module include: During the offline training phase, network parameters were trained using historical water tank test data through supervised learning. During the online fine-tuning phase, a sliding window approach was used during the experiment. After each tidal cycle, the model was incrementally updated using real-time sensor data to adapt to changes in current water and sediment conditions.

10. The intelligent underwater sand-addition control method based on machine learning and model predictive control according to claim 5, characterized in that, The objective function for the rolling optimization of the model prediction control module is: ； in The objective function for rolling optimization is... For the prediction step number, To predict the total number of steps in the time domain, Predict the sediment concentration profile vector step by step. No. Step target sediment content profile vector For the first The vector of changes in the flow rate of the step-gaza dam. For the first Step-by-step flow vector of each sand pipe The square of the Euclidean norm. Penalize weights for changes in traffic; The predetermined constraints include: upper and lower limits of flow rate for each sand-adding pipe, concentration range of sand-mixing pool, and the constraint of matching sand-adding timing with tidal phase.

Images