Intelligent conveying control system for viscous fluid

Abstract

The invention discloses an intelligent viscous fluid conveying control system which comprises a conveying pipeline and a fluid conveying pump arranged on the conveying pipeline, a bent pipe bent upwards is arranged at the joint of an upstream horizontal pipeline and a downstream falling pipeline of the conveying pipeline, and fluid in the upstream horizontal pipeline enters the downstream falling pipeline through the bent pipe; a closed cavity only communicated with the downstream falling pipeline is formed in the bent pipe, the closed cavity is communicated with a vacuum adjusting valve, the vacuum degree in the closed cavity is adjusted through the vacuum adjusting valve, and the falling flow speed and the fluid pressure of fluid in the downstream falling pipeline are controlled. Stable control of viscous fluid in the conveying pipeline is achieved by adjusting the negative pressure in the conveying pipeline, and final control parameters are fed back and output through the conveying control module of a mixed framework integrating the PID controller, the regression prediction model and the reinforcement learning model according to actual fluid parameters in the conveying pipeline. And efficient, safe and accurate regulation and control on the fluid in the whole conveying pipeline are realized.

Description

TECHNICAL FIELD

[0001] The application discloses an intelligent viscous fluid conveying control system and belongs to the technical field of intelligent control of pipeline fluid conveying. BACKGROUND

[0002] In the fields of mining, chemical industry and sewage treatment, there are a large number of technical scenarios of pipeline fluid conveying, especially in the process of long-distance and drop pipeline fluid conveying, the following problems always exist:

[0003] (1) Water hammer effect. The water hammer effect caused by excessive kinetic energy of fluid in the process of pipeline flow, and the large static pressure generated by fluid in the process of long-distance vertical conveying, all can cause pipeline wear and even damage, and in severe cases, can even make the entire conveying system paralyzed.

[0004] (2) Low flow rate regulation efficiency. Traditional methods mostly rely on adjusting a single pump or valve to solve the problem, which is low in efficiency, slow in system response, and has certain safety hazards.

[0005] (3) Lack of self-adaptive ability. The existing control system mostly relies on parameter preset, and is slow in response to the adjustment of the conveying fluid in the pipeline when the working condition changes, and cannot adaptively adjust according to various complex working conditions such as pipeline wear and fluid viscosity change.

[0006] (4) Excessive static pressure. In the vertical pipeline, the static pressure increases with the increase of liquid level height, resulting in excessive static pressure in the downstream, and there is a great safety hazard.

[0007] Therefore, the above fields of pipeline fluid conveying urgently need a safe and efficient intelligent control scheme that can accurately control the flow rate and pressure of fluid and adaptively adjust the fluid flow characteristics in real time according to the working condition. SUMMARY

[0008] The technical problem solved by the application is to provide an intelligent viscous fluid conveying control system in view of the problems of low efficiency, slow response, low safety and low adaptability of the existing viscous fluid conveying system.

[0009] The application adopts the following technical scheme:

[0010] The application discloses a kind of viscous fluid intelligent delivery control system, including delivery pipeline 300 and fluid delivery pump 400 being arranged on delivery pipeline, the delivery pipeline 300 realizes upstream and downstream fluid delivery with height difference, is provided with the elbow pipe 300C that is bent upwards at the junction of upstream horizontal pipeline and downstream falling pipeline, and the upstream horizontal pipeline fluid passes through elbow pipe 300C and enters downstream falling pipeline, forms the closed chamber 300D that only communicates with downstream falling pipeline in elbow pipe, the closed chamber 300D communicates with vacuum regulating valve 500, and the vacuum degree in closed chamber 300D is adjusted by vacuum regulating valve 500, controls the falling flow rate and fluid pressure of fluid in downstream falling pipeline.

[0011] In the viscous fluid intelligent delivery control system of the application, further, the upstream horizontal pipeline of the delivery pipeline 300 is provided with a flow sensor 31, a temperature sensor 32 and a first pressure sensor 33, and the closed chamber 300D is provided with a negative pressure sensor 51; the downstream falling pipeline of the delivery pipeline 300 is provided with a second pressure sensor 34; the upstream horizontal pipeline of the delivery pipeline 300 is provided with a flow control valve 301, and the delivery pipeline 300 is provided with a shut-off valve for cutting off the fluid delivery in the pipeline.

[0012] In the viscous fluid intelligent delivery control system of the application, further, the application further comprises a delivery control module 600, the flow sensor 31, the temperature sensor 32, the first pressure sensor 34, the second pressure sensor 34 and the negative pressure sensor 51 measure and collect the actual fluid parameters in the delivery pipeline and transmit the actual fluid parameters to the delivery control module 600, and the delivery control module 600 controls the flow control valve 301, the fluid delivery pump 400, the vacuum regulating valve 500 and the shut-off valve according to the feedback of the actual fluid parameters.

[0013] In the viscous fluid intelligent delivery control system of the application, further, the upstream horizontal pipeline of the delivery pipeline 300 is further provided with a bursting disc 303.

[0014] In the viscous fluid intelligent delivery control system of the application, further, the delivery control module 600 is integrated with a hybrid decision model, the hybrid decision model adopts a hybrid architecture of a PID controller, a regression prediction model and a reinforcement learning model, the PID controller, the regression prediction model and the reinforcement learning model respectively perform confidence evaluation based on the actual fluid parameters in the delivery pipeline, the actual fluid parameters in the current delivery pipeline are respectively input into the PID controller, the regression prediction model and the reinforcement learning model to output three groups of recommended control parameters, and the three groups of recommended control parameters are dynamically weighted and fused based on the confidence of the PID controller, the regression prediction model and the reinforcement learning model to obtain the final control parameters fed back by the delivery control module 600.

[0015] In the intelligent viscous fluid delivery control system, further, a virtual simulation model synchronously operated with the system is further included, the virtual simulation model is a physical theory model constructed by a fluid dynamics model corresponding to a delivery pipeline of the system, a valve flow characteristic curve, a performance curve of a fluid delivery pump, and physical property parameters of a delivered fluid, a corresponding theoretical fluid parameter of the system is predicted by inputting a control parameter, and a fault category of the system is identified by comparing the theoretical fluid parameter with an actual fluid parameter of the system.

[0016] In the intelligent viscous fluid delivery control system, further, the regression prediction model adopts a gradient boosting decision tree algorithm, paired data of historical fluid parameters and corresponding optimized control parameters of the system are used as training samples for supervised learning training of the model, historical fluid parameters and corresponding optimized control parameters are used as training samples for offline model construction in a first stage, and a time decay weighted method is used for online incremental learning of the offline model, a suggested control parameter of the system is output by the trained regression prediction model for input of the actual fluid parameter in the current delivery pipeline.

[0017] In the intelligent viscous fluid delivery control system, further, the reinforcement learning model adopts a deep deterministic policy gradient algorithm, the theoretical fluid parameter and the corresponding control parameter generated by the virtual simulation model are used for offline training, and a reward function based on flow tracking accuracy, pressure stability and system delivery energy consumption of the fluid is used:

[0018] ,

[0019] Wherein, Rt is an immediate reward calculated by the corresponding control parameter; is a flow tracking reward of the fluid, , is an actual flow of the fluid in the current system, is a set target flow, is a flow proportionality coefficient; is a pressure stability reward of the fluid, , is an actual pressure of the fluid in the current system, is a set optimal pressure value, is a set pressure safety upper limit, , is a pressure proportionality coefficient, respectively; is a delivery energy consumption reward of the system, , is a pump frequency of the fluid delivery pump of the current system, is an energy consumption proportionality coefficient; is a safety event penalty, , , , are respectively , , , reward weight coefficients.

[0020] In the viscous fluid intelligent delivery control system of the present application, further, the delivery control module (600) feedback output final control parameters are weighted and fused by the following formula:

[0021] ,

[0022] wherein, is the recommended control parameter output by the PID controller, is the recommended control parameter output by the regression prediction model, is the recommended control parameter output by the reinforcement learning model, is the real-time weight coefficient of the PID controller, is the real-time weight coefficient of the regression prediction model output, is the real-time weight coefficient of the reinforcement learning model, if the fluid parameters of the system are in a severe transient process, increase while reducing and ; if the fluid parameters of the system are in a steady state or slow change process, the real-time weight coefficients are calculated according to the confidence of the PID controller, the regression prediction model and the reinforcement learning model by the following formula:

[0023] ,

[0024] wherein, is the calculated real-time weight coefficient of the PID controller, the regression prediction model or the reinforcement learning model, is the confidence of the corresponding PID controller, regression prediction model or reinforcement learning model, is the confidence of the PID controller, set as a constant basic confidence, is the confidence of the regression prediction model, selected by the Mahalanobis distance or similarity measure between the actual fluid parameters input by the regression prediction model and the model training samples, the closer the Mahalanobis distance or the more matched the similarity distribution, the higher ; is the confidence of the reinforcement learning model, evaluated by the value function of the action output by the policy network of the reinforcement learning model input by the actual fluid parameters, the higher the value, the higher .

[0025] In the viscous fluid intelligent conveying control system, further, the conveying control module 600 is also provided with a safety verification module, the safety verification module compares the final control parameter with the mechanical limit parameter of the feedback control device: if the final control parameter does not exceed the mechanical limit parameter of the device, the conveying control module 600 outputs according to the final control parameter; if the final control parameter exceeds the mechanical limit parameter of the device, the conveying control module 600 outputs according to the recommended control parameter of the PID controller.

[0026] The technical scheme has the following beneficial effects:

[0027] The pipeline safety is effectively protected. The present application forms a closed chamber in the conveying pipeline through the elbow pipe, changes the vacuum degree in the closed chamber through the vacuum regulating valve, dynamically regulates and controls the falling speed of the viscous fluid in the pipeline, gently starts and stops the viscous fluid in the conveying pipeline, balances or controls the static pressure of the fluid in the downstream pipeline, avoids the rapid change of the kinetic energy of the viscous fluid, and can also control the pressure of the downstream viscous fluid, so that it is within a safe range, thereby protecting the safety of the pipeline system.

[0028] The flow is accurately controlled. The intelligent regulation and control of the conveying control module on each component of the conveying pipeline can accurately control the flow of the fluid in the conveying pipeline according to the material use site, and realize efficient conveying and distribution of the fluid in the conveying pipeline. At the same time, the conveying control module can accurately capture the real-time fluid parameter changes by setting a virtual simulation model, and can monitor and predict the state of each device in the system by comparing the differences between the system sensor data and the output of the virtual simulation model, thereby better preventing accidents.

[0029] The learning and adaptive ability is strong. The conveying control module of the system adopts a hybrid decision model with a mixed architecture of PID controller, regression prediction model and reinforcement learning model, multi-prediction fusion, wherein the PID controller serves as a basic control to ensure the basic stability and safety of the system under any conditions. The regression prediction model is based on a machine learning optimization algorithm to find the optimal or suboptimal working point of the system under specific working conditions, and the reinforcement learning model is based on reinforcement learning and online learning, so that the system can accumulate experience from long-term operation and continuously improve the control performance and adapt to slow working condition drift. According to the sensor data, the control parameters more close to the real working condition are predicted and output to regulate and control each component on the conveying pipeline, and the optimal working parameters are autonomously predicted and recommended by combining historical data, physical models, classical control methods based on expert experience and data-driven artificial intelligence methods are deeply integrated to form a perfect fluid intelligent conveying control system.

[0030] High efficiency and energy saving. The system uses the scheduling of the conveying control module to realize the cooperation of each actuator on the conveying pipeline, optimizes the system working condition, reduces the energy consumption of the pump and the wear of the pipeline, reduces the dependence on people, and improves the stability and production efficiency of the production process.

[0031] In summary, the viscous fluid intelligent conveying control system provided by the application realizes stable control of viscous fluid in the conveying pipeline through negative pressure regulation in the conveying pipeline, realizes efficient, safe and accurate regulation and control of the fluid in the entire conveying pipeline by using a mixed decision model conveying control module, and the entire system has the advantages of simple structure, low energy consumption, high automation degree and strong self-adaptive learning ability.

[0032] The application will be further described below in combination with the drawings and specific embodiments BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 It is a schematic diagram of the overall structure of a viscous fluid intelligent conveying control system in embodiments 1 and 2.

[0034] Figure 2 It is a schematic diagram of the formation of a closed chamber in the conveying pipeline in embodiments 1 and 2.

[0035] Figure 3 It is an internal structure block diagram of the conveying control module in embodiment 2.

[0036] Figure 4 It is a working flowchart of a viscous fluid intelligent conveying control system in embodiment 2.

[0037] In the figure, 100 is an upstream storage tank, 101 is a first valve, 200 is a downstream storage tank, 201 is a downstream storage valve, 300 is a conveying pipeline, 300A is an upstream horizontal pipeline, 300B is a downstream falling pipeline, 300C is a bend, 300D is a closed chamber, 301 is a flow control valve, 302 is a second valve, 303 is a rupture disc, 31 is a flow sensor, 32 is a temperature sensor, 33 is a first pressure sensor, 34 is a second pressure sensor, 400 is a conveying pump, 500 is a vacuum regulating valve, 501 is a vacuum hand valve, 51 is a negative pressure sensor, 600 is a conveying control module, 601 is a PID controller, 602 is a regression prediction model, 603 is a reinforcement learning model, and 604 is a safety verification module. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments will be described clearly and completely below with reference to the drawings. The embodiments take the delivery of latex matrix to the downhole as an example. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0039] Embodiment 1

[0040] Referring to Figure 1 , one of the viscous fluid intelligent delivery control systems in the drawings is a specific embodiment of the present application, which is used for delivering the latex matrix fluid in the upstream storage tank 100 on the ground to the downstream storage tank 200 in the downhole, and includes a delivery pipeline 300 and a fluid delivery pump 400 arranged on the delivery pipeline. The delivery pipeline 300 connects the upstream storage tank 100 and the downstream storage tank 200, and includes at least one upstream horizontal pipeline 300A arranged horizontally on the ground and one downstream falling pipeline 300B extending downward along the mine shaft, so as to realize the delivery of the latex matrix fluid from the ground to the downhole. The fluid delivery pump 400 is located on the upstream horizontal pipeline 300A connected to the upstream storage tank 100. An upwardly curved and arched elbow pipe 300C is arranged at the connection between the upstream horizontal pipeline 300A and the downstream falling pipeline 300B, and has a reverse u-shaped passage structure. The lower wall position of the pipeline at the highest position of the upwardly arched elbow pipe 300C is higher than the upper wall position of the upstream horizontal pipeline. After the latex matrix fluid in the upstream horizontal pipeline 300A enters the downstream falling pipeline 300B through the elbow pipe 300C, the fluid forms a gas-tight isolation, and a sealed chamber 300D is formed at the connection between the elbow pipe 300C and the downstream falling pipeline 300B, which is in communication with only the downstream falling pipeline. As shown in Figure 2 , the sealed chamber 300D is in communication with a vacuum regulating valve 500. The negative pressure vacuum degree in the sealed chamber 300D is adjusted by the vacuum regulating valve 500, so as to control the falling flow rate and fluid pressure of the fluid in the downstream falling pipeline. The vacuum regulating valve 500 is provided with a vacuum hand valve 501, which is a safety redundant valve and is generally kept closed. The vacuum hand valve 501 is manually used only when the vacuum regulating valve fails.

[0041] The embodiment controls the negative pressure vacuum degree in the closed chamber 300D through the vacuum regulating valve 500, so as to control the falling speed of the latex matrix fluid in the downstream falling pipeline 300B. When the viscous fluid falls, the negative pressure inside the downstream falling pipeline will form a traction viscous fluid, and the degree of communication between the closed chamber 300D and the atmosphere is adjusted through the vacuum regulating valve, and the control of the upstream and downstream negative pressure balance of the fluid in the conveying pipeline is realized by adjusting the vacuum degree. When the vacuum degree in the closed chamber 300D becomes larger, the falling speed of the viscous fluid will be inhibited; when the vacuum degree in the closed chamber 300D becomes smaller, the inhibiting effect is weakened, and the falling speed of the viscous fluid will be accelerated. The embodiment can also control the static pressure of the viscous fluid in the downstream falling pipeline 300B. It is known that the pressure of the viscous fluid flowing in the full pipeline of the downstream falling pipeline 300B is the gravity water head pressure, and by adjusting the negative pressure vacuum degree in the closed chamber 300D, a negative pressure water head opposite to the gravity water head pressure can be formed in the upstream horizontal pipeline 300A, so as to partially offset the water head pressure in the downstream falling pipeline 300B to realize pressure reduction.

[0042] The upstream horizontal pipeline 300A of the conveying pipeline 300 is provided with a flow sensor 31, a temperature sensor 32 and a first pressure sensor 33 for detecting the fluid flow, fluid temperature and fluid pressure of the upstream fluid. The downstream falling pipeline of the conveying pipeline 300 is provided with a second pressure sensor 34 for detecting the static pressure and impact pressure of the downstream fluid in the pipeline, and the closed chamber in the elbow pipe 300C is provided with a negative pressure sensor 51 for real-time detection of the negative pressure vacuum degree in the closed chamber 300D. The upstream horizontal pipeline 300A of the conveying pipeline 300 is provided with a flow control valve 301 for regulating the flow of the upstream fluid, and the conveying pipeline 300 is provided with a closing valve for cutting off the fluid conveying in the pipeline, which is used to block the upstream fluid and close the conveying pipeline 300.

[0043] The closing valve includes a first valve 101 arranged at the discharge port of the upstream storage tank and a second valve 302 arranged at the inlet of the downstream storage tank. The first valve 101 is a commonly used closing valve for controlling the feeding of the system, and the second valve 302 is a safety redundant valve which is kept open during fluid conveying and does not participate in feedback regulation of conveying. Only when the first valve fails, the second valve replaces the first valve to execute the closing action of the conveying pipeline. A downstream storage valve 201 is also installed at the discharge port of the downstream storage tank 200 for controlling the discharge of the downstream storage tank 200. The first valve 101, the second valve 302, the downstream storage valve 201, the flow control valve 301 and the vacuum regulating valve 500 are all electrically controlled by electric valves. The fluid conveying pump 400 adopts a variable frequency speed regulating screw pump according to the characteristics of the latex matrix fluid, and is feedback connected with the conveying control module 600 to realize variable frequency control.

[0044] The system further comprises a delivery control module 600, the flow sensor 31, the temperature sensor 32, the first pressure sensor 34, the second pressure sensor 34, and the negative pressure sensor 51 measure the actual fluid parameters in the collection delivery pipeline and transmit the actual fluid parameters to the delivery control module 600, the actual fluid parameters including the fluid flow, the fluid temperature, the fluid pressure of the upstream fluid, the vacuum degree of the sealed chamber, the static pressure and the impact pressure of the downstream fluid, and the delivery control module 600 controls the flow control valve 301, the fluid delivery pump 400, the vacuum regulating valve 500, and the closing valve according to the feedback of the actual fluid parameters, and the feedback control parameters include the opening degree of the flow control valve, the pump frequency of the fluid delivery pump, the opening degree of the vacuum regulating valve, and the start-stop of the closing valve.

[0045] In addition, a rupture disc 303 is arranged on the upstream horizontal pipeline of the delivery pipeline 300, and in the case of uncontrolled fluid pressure in the delivery pipeline, emergency pressure relief is performed through the rupture disc 303.

[0046] The delivery control module 600 can adopt a PLC controller, the flow sensor 31, the temperature sensor 32, the first pressure sensor 34, the second pressure sensor 34, and the negative pressure sensor 51 are connected to the signal input end of the PLC control, and the feedback signal is output to the flow control valve 301, the fluid delivery pump 400, the vacuum regulating valve 500, and the closing valve through programming control, the data of the sensors in the delivery pipeline are received by the PLC controller in real time to feedback the pump frequency control of the delivery pump and the opening and closing of the valves, for example, the PLC controller receives real-time data from the negative pressure sensor 51 and the second pressure sensor 34, and based on the comparison difference between the static pressure signal of the downstream fluid fed back by the second pressure sensor 34 and the preset target pressure, an adjustment instruction for the opening degree of the vacuum regulating valve 500 is output to realize safe and efficient delivery of the fluid. The specific PLC programming control belongs to mature automatic control technology, and will not be described here.

[0047] Embodiment 2

[0048] The PLC programming control is a fixed mode automatic control scheme, which cannot adaptively adjust the feedback control parameters according to different actual working conditions, in order to further optimize the fluid delivery control in the delivery pipeline and make it more accurate and intelligent, the delivery control module 600 of the embodiment integrates a hybrid decision model in the PLC controller, the hybrid decision model adopts a hybrid architecture of a PID controller 601, a regression prediction model 602, and a reinforcement learning model 603, as shown in FIG. 6. Figure 3

[0049] ​The embodiment performs confidence evaluation based on the actual fluid parameters in the conveying pipeline by the PID controller 601, the regression prediction model 602 and the reinforcement learning model 603 respectively, and inputs the actual fluid parameters in the current conveying pipeline into the PID controller 601, the regression prediction model 602 and the reinforcement learning model 603 respectively to output three groups of recommended control parameters. Specifically, the PID controller 601 adopts a classic proportional-integral-derivative control algorithm to output recommended control parameters according to the actual fluid parameters , the regression prediction model 602 adopts a gradient boosting decision tree algorithm to predict and output recommended control parameters according to the actual fluid parameters , and the reinforcement learning model 603 adopts a deep deterministic policy gradient algorithm to predict and output recommended control parameters according to the actual fluid parameters Then, the three groups of recommended control parameters are dynamically weighted and fused based on the real-time confidence of the PID controller 601, the regression prediction model 602 and the reinforcement learning model 603 to obtain the final control parameters fed back by the conveying control module 600 .

[0050] The viscous fluid intelligent conveying control system of the embodiment further includes a virtual simulation model running synchronously with the system. The virtual simulation model is a physical theoretical model constructed by a fluid dynamics model corresponding to the conveying pipeline of the system, a valve flow characteristic curve, a performance curve of a fluid conveying pump and physical property parameters of a latex matrix fluid conveyed by the system, and constitutes a digital twin model synchronous with the system, which is used for state simulation, soft measurement and fault prediction. The input control parameters are simulated by the corresponding physical theoretical model of the system to obtain the theoretical fluid parameters corresponding to the system, and the theoretical fluid parameters are compared with the actual fluid parameters identified by each sensor of the system. The specific comparison and identification process is performed by generating a residual sequence representing the state deviation of the system, and analyzing the residual sequence by using a time series anomaly detection algorithm, so as to produce early warning information when the performance of the conveying pipeline equipment of the system is degraded or a fault occurs, adjust the control strategy or trigger a maintenance alarm, and identify the fault category of the system. The virtual simulation model can be used as a "soft sensor" of the system, and can deduce fluid parameters that are difficult to measure directly, such as the real-time flow rate at a specific point in the pipeline, by internal calculation. By analyzing the residual sequence of the actual fluid parameter data measured by the system and the theoretical fluid parameters of the simulation system, and using time series anomaly detection algorithms such as Isolation Forest or Long Short-Term Memory Autoencoder, the abnormal characteristics of the measured parameters of the system are identified, so as to realize early warning of early faults such as vacuum valve leakage and abnormal sensor signals.

[0051] The delivery control module 600 pre-processes the raw data collected by the sensors, such as the upstream pressure, upstream temperature, upstream flow rate, downstream pressure, etc. of the fluid, to improve the quality of the input data of the hybrid decision model. For the raw data, a median filtering method based on a time window is used, and combined with the Relyada criterion for outlier judgment and elimination, to effectively remove short-term peak noise and false outliers in the data. On this basis, further multi-dimensional feature extraction is carried out, including: in the time dimension, the moving average, fluctuation standard deviation and change rate of adjacent moments of key parameters such as downstream pressure, vacuum degree, delivery pump frequency, etc. are calculated to characterize the steady-state and dynamic characteristics of the system; in the frequency dimension, the pressure signal of the fluid is analyzed to extract the energy of the main oscillation frequency component, which is used to identify abnormal fluctuations that may indicate water hammer effect or pipeline resonance; at the same time, according to the flow set value and the historical viscosity record of the fluid, data is automatically labeled with labels such as "high flow-high viscosity working condition", "low flow-starting working condition", etc. to provide sample classification basis for subsequent supervised machine learning. Finally, all the extracted features are normalized to the minimum-maximum value, and scaled to the interval [0, 1] to improve the efficiency and stability of model training.

[0052] The gradient boosting decision tree algorithm is used in the regression prediction model 602, and the model input is the actual fluid parameters including the fluid flow rate, fluid temperature, fluid pressure, vacuum negative pressure of the sealed chamber, static pressure and impact pressure of the downstream fluid, and the output includes the recommended control parameters of the vacuum regulating valve opening and the delivery pump frequency. The paired data of historical fluid parameters and corresponding optimized control parameters of the same type of viscous fluid intelligent delivery control system are used as training samples for supervised learning training of the model. In the first stage, the historical fluid parameters and corresponding optimized control parameters are used as training samples for offline model construction, and in the second stage, the time decay weighted method is used for online incremental learning of the offline model. The trained regression prediction model is used to predict the output of the recommended control parameters of the system based on the input of the actual fluid parameters in the current delivery pipeline.

[0053] Historical operating data of the fluid delivery pipeline is collected, including fluid parameters and corresponding control parameters under various system conditions, to construct training samples for training the regression prediction model. The core function of the regression prediction model is to establish a rapid correlation mapping between fluid parameters under the current system state and the expected control effect, thereby achieving advanced feedforward compensation and operating point optimization. The regression prediction model selects the gradient boosting decision tree algorithm, which has high fitting efficiency and model interpretability for medium-sized industrial data and can well handle the complex nonlinear relationships between data features in various training samples. The model input includes: upstream and downstream pressure, vacuum, flow rate, and temperature data of the fluid collected from various sensors on the system delivery pipeline at the current moment; control parameters of the feedback control actuator, such as pump frequency of the delivery pump, opening degree of each valve, and set values ​​of target flow rate and pressure safety upper limit; in addition, the trend of downstream pressure change in the previous few moments is also included as time-series context information. The model output is the suggested control parameters for the system at the next control moment, mainly including the adjustment amount of the opening degree of each valve and the pump frequency of the delivery pump at the next moment.

[0054] It should be noted that the training method of the gradient boosting decision tree algorithm belongs to well-known machine learning techniques and standard training procedures. The training of the regression prediction model 602 in this embodiment is divided into two stages: initially, preprocessed historical system operation data is used, and training samples are selected from records that "brought the system to its optimal stable state after the actual control action was issued" for offline model construction; later, a time-decay weighted approach is used for online incremental learning, assigning a weight related to its timestamp to each newly generated operation data sample, with the weight function using an exponential decay form: ,in, It is the first The weights of each sample, It is the current time. This represents the sample collection time, and λ is the decay coefficient, a constant greater than 0 chosen to control the rate at which the importance of older data decreases. During model fine-tuning, the gradient boosting decision tree algorithm calculates a weighted loss function based on this weight, ensuring that recent data has a much greater impact on model parameter updates than earlier data. This assigns higher importance to newer data, and periodic fine-tuning allows the regression prediction model to adapt to changes in system operating conditions caused by pipeline wear, equipment wear, etc.

[0055] The reinforcement learning model 603 adopts a deep deterministic policy gradient algorithm suitable for continuous control scenarios, and is trained offline by theoretical fluid parameters and corresponding control parameters generated by the virtual simulation model, so that it can autonomously learn how to cooperatively control the actuators of the system pipeline to achieve long-term optimal performance in the interaction process with the simulation environment, and endow the fluid delivery control system with the ability to improve the strategy by evaluating the consequences of actions in an uncertain environment. Offline training of the deep deterministic policy gradient algorithm in the virtual simulation environment is a common simulation training transfer learning paradigm for reinforcement learning in industrial control. In this embodiment, the state information perceived by the reinforcement learning model is similar to the input features of the regression prediction model, but it focuses more on dynamic indicators such as the instantaneous rate of change of pressure and the cumulative amount of flow deviation. The output of the proposed control parameters is defined as a series of continuous and normalized control instructions, such as the change amount of the vacuum valve opening in the range of plus or minus zero one, or the adjustment amount of the delivery pump frequency in the range of plus or minus five hertz. To achieve the optimization of multiple target parameters of the system pipeline, a reward function based on the flow tracking accuracy of the fluid, the pressure stability of the fluid, and the energy consumption of the system delivery is used as follows:

[0056] ,

[0057] wherein Rt is the immediate reward calculated by the corresponding control parameters; is the flow tracking reward of the fluid, , is the actual flow of the fluid in the current system, which is the fluid flow of the upstream fluid collected by the flow sensor, is the set target flow, is the flow proportionality coefficient, which encourages the actual fluid flow in the system to closely follow the target flow ; is the pressure stability reward of the fluid, , is the actual pressure of the fluid in the current system, which is the static pressure and impact pressure of the downstream fluid collected by the second pressure sensor, is the set pressure optimal value, is the set pressure safety upper limit, , is the pressure proportionality coefficient; the pressure stability reward includes two parts: the first part encourages the actual pressure of the downstream fluid to be stable around the pressure optimal value , and the second part is a quadratic penalty term that gives a sharply increasing penalty when the actual pressure of the downstream fluid exceeds the pressure safety upper limit; is the delivery energy consumption reward of the system, , is the pump frequency of the fluid delivery pump of the current system, is the energy consumption ratio coefficient; the energy consumption reward of the delivery is negatively related to the pump frequency of the delivery pump, encouraging the system to reduce the energy consumption of the delivery pump and make more use of the gravitational potential energy of the fluid in the delivery pipeline; is a safety event penalty, which is set as a large negative constant, such as -100, triggered when the system detects a dramatic pressure fluctuation or triggers a safety interlock; , , , is the reward weight coefficient of each item, used to adjust the relative importance of different optimization objectives.

[0058] The above flow rate ratio coefficient , the pressure ratio coefficient , , the energy consumption ratio coefficient , the reward weight coefficient , , , , etc. can be determined by system simulation, experimental calibration or optimization algorithm. The specific values are determined according to the actual system characteristics by the following methods: based on the historical operation data of the system and the expert experience debugging; through multi-objective optimization algorithms such as genetic algorithm and particle swarm optimization to optimize the parameters in the simulation environment; combined with the actual operation feedback of the system for online adaptive adjustment.

[0059] The training strategy of the reinforcement learning model 603 is first trained offline in a high-fidelity digital twin virtual simulation model based on the physical mechanism of the fluid delivery system. The virtual simulation model is constructed by integrating the pipeline fluid dynamics model, the valve flow characteristic curve, the pump performance curve and the fluid physical property parameters. It can calculate the fluid parameters in the delivery pipeline of the system at the next time, including fluid pressure, flow rate, vacuum degree, etc. according to the control parameters given by the delivery control module 600, and feed them back to the reinforcement learning model. At the same time, it calculates the immediate reward according to the above reward function of the reinforcement learning model, so that the reinforcement learning model can initially master the basic control strategy. After the reinforcement learning model 603 is deployed to the actual system, it will explore within the safety action baseline provided by the PID controller, and store the "state-action-reward-new state" data sequence obtained by exploration in the model memory library. And use the non-busy period of the fluid delivery system for progressive online learning, so as to continuously optimize its decision network.

[0060] After the PID controller 601, the regression prediction model 602 and the reinforcement learning model 603 output the recommended control parameters, the delivery control module 600 performs dynamic weighted fusion through the following formula to coordinate and integrate the recommendations from the three types of controllers to form a safe, reliable and optimized final output final control parameter :

[0061] ,

[0062] wherein, is the recommended control parameter output by the PID controller, is the recommended control parameter output by the regression prediction model, is the recommended control parameter of the reinforcement learning model, is the real-time weight coefficient of the PID controller, is the real-time weight coefficient of the regression prediction model output, is the real-time weight coefficient of the reinforcement learning model.

[0063] The delivery control module 600 judges whether the latex matrix fluid in the delivery pipeline is currently in a severe transient process or a steady state or slow-changing process by monitoring the rate of change or standard deviation of key fluid parameters such as downstream fluid pressure, flow rate, etc. in real time and comparing them with preset thresholds. If the rate of change or standard deviation of the above-mentioned key fluid parameters exceeds the preset threshold, it is determined to be a severe transient process; otherwise, it is determined to be a steady state or slow-changing process. When the fluid parameters of the system are in a severe transient process, the PID controller is given a higher weight while the regression prediction model and the reinforcement learning model are given lower weights and to prioritize the stability and safety of the system; when the fluid parameters of the system are in a steady state or slow-changing process, the system will calculate the matching degree of the current operating state with the historical distribution of each model training sample data in real time and take it as the confidence, and calculate the real-time weight coefficient according to the confidence evaluated by the PID controller, the regression prediction model and the reinforcement learning model by the following formula:

[0064] ,

[0065] wherein, is the real-time weight coefficient of the PID controller, the regression prediction model or the reinforcement learning model calculated, is the confidence of the corresponding PID controller, regression prediction model or reinforcement learning model, is the confidence of the PID controller, which is set as a constant basic confidence, is the confidence of the regression prediction model. The Mahalanobis distance or similarity measure between the actual fluid parameters such as pressure, flow rate, temperature, vacuum degree, etc. input by the regression prediction model and the model training samples are selected, the closer the Mahalanobis distance or the more matched the similarity distribution, the higher the ; is the confidence of the reinforcement learning model, which is evaluated by the value function of the actual fluid parameters input by the reinforcement learning model in the action output of its policy network, the higher the value, the higher the .

[0066] Specifically to 、 、 As follows:

[0067] 、

[0068] 、

[0069] .

[0070] The delivery control module 600 is also provided with a safety verification module 604. The fused final control parameter is subjected to safety rule verification by the safety verification module 604 constructed based on physical laws and operating experience before being sent to the actuator. The safety verification module 604 compares the final control parameter with the mechanical limit parameter of the feedback control device, such as the valve body opening change in the final control parameter and the mechanical limit opening of the corresponding valve body, the pump frequency of the fluid delivery pump and the frequency limit of the frequency conversion screw pump. If the final control parameter does not exceed the mechanical limit parameter of the device, the final control parameter is output; if the final control parameter exceeds the mechanical limit parameter of the device, the recommended control parameter of the PID controller 601 is output.

[0071] The delivery control module 600 is also provided with a human-computer interaction module for information interaction between the operator and the system, including but not limited to the operator setting the system control parameter, the real-time display of the system fluid parameter and the fault signal, etc.

[0072] The hardware arrangement of the delivery pipeline of the viscous fluid intelligent delivery control system in the embodiment is the same as that of embodiment 1, and reference is made to Figure 4 The specific working process and control method are as follows:

[0073] S1, the system is pre-started and self-checked to confirm that the signals of the system sensors, i.e. the flow sensor 31, the temperature sensor 32, the first pressure sensor 33, the second pressure sensor 34, the negative pressure sensor 51 and the actuator first valve 101, the flow control valve 301, the second valve 302, the delivery pump 400, the vacuum regulating valve 500 and the delivery control module 600 are normally connected, and it is confirmed that the valve opening and closing are normal. After the inspection is correct, the delivery control module 600 loads the regression prediction model of the latex matrix fluid delivery, the reinforcement learning model and the historical parameters of the current working condition of the system from the database, starts the virtual simulation model and keeps synchronization with the delivery control module at the same time.

[0074] S2, the delivery control module 600 opens the first valve to start the material delivery, and starts the delivery pump 400 to the preset safe frequency. The flow sensor 31, temperature sensor 32 and first pressure sensor 33 at the outlet of the delivery pump start to monitor the flow rate, temperature and pressure data of the upstream latex matrix fluid in real time. After the fluid flows to the downstream pipeline, the second pressure sensor 34 at the downstream end feeds the pressure data of the downstream latex matrix fluid to the delivery control module 600 in real time, and the PID controller 601 starts to intervene. According to the difference between the target pressure and the measured pressure, the opening of the vacuum regulating valve 500 is preliminarily adjusted to suppress the acceleration of the latex matrix fluid in the downstream pipeline and prevent the impact pressure from being too large. When the flow rate of the latex matrix fluid in the delivery pipeline tends to be stable, the mixed decision model of the delivery control module 600 starts to intervene, and the real-time fluid parameter data monitored by the sensors are preprocessed, and then the processed fluid parameter data are input into the PID controller 601, the regression prediction model 602 and the reinforcement learning model 603 of the mixed decision model to quickly give the recommended control parameters predicted by each model. The final control parameters are finally generated by the mixed decision model through the integrated analysis and weight distribution of the three groups of recommended control parameters of the PID controller 601, the regression prediction model 602 and the reinforcement learning model 603, and are sent to the first valve 101, the flow control valve 301, the second valve 302, the delivery pump 400 and the vacuum regulating valve 500 of the system by the PLC controller.

[0075] S3, assuming that the downhole material demand changes, the delivery control module 600 will quickly respond through the PID controller 601, predict the influence of the flow rate change on the downstream latex matrix fluid pressure through the regression prediction model 602, and finally adjust the opening of the vacuum regulating valve 500 in advance to reduce or increase the vacuum degree to ensure the stability of the delivery. If the vacuum regulating valve 500 fails, the delivery control module will issue an alarm to notify the operator to manually adjust the vacuum degree by using the vacuum hand valve 501. During the adjustment process, the delivery control module 600 will continuously collect real-time fluid parameter data monitored by each sensor and start the online learning process to update the database with new data to enhance the adaptability of the regression prediction model 602. In addition, during the adjustment process, the delivery control module 600 will run the virtual simulation model synchronously, and compare the simulated theoretical fluid parameter data with the measured fluid parameter data monitored by the sensors in real time. Once the difference between the two exceeds the normal range, the system will trigger an alarm and prompt possible faults. If the upstream first pressure sensor 33 monitors that the upstream latex matrix fluid pressure is abnormally close to the safety threshold and the intelligent adjustment effect is very small, the rupture disc 303 will serve as the final safety barrier for physical pressure relief. If the temperature sensor 32 monitors that the temperature at the outlet of the delivery pump is relatively high and abnormally close to the safety threshold, the control delivery control module 600 will execute the shutdown program.

[0076] S4, when the latex matrix fluid delivery task is completed, the delivery control module 600 executes the shutdown program, the mixed decision model will predict the optimal shutdown curve according to the existing fluid parameter data characteristics, and generate control instructions for the vacuum regulating valve 500 to gradually increase the vacuum degree in the closed cavity to slowly inhibit the latex matrix fluid flow rate in the delivery pipeline, while reducing the pump frequency of the delivery pump 400, so that the downstream latex matrix fluid pressure gradually decreases to eliminate the safety risk. Finally, the system closes the first valve 101 to cut off the flow, if the system detects that the first valve 101 is not closed properly, the safety interlock reaction is triggered immediately, the second valve 302 is automatically closed, and the latex matrix fluid is completely cut off in the delivery pipeline.

[0077] S5, the latex matrix fluid delivered to the downhole is stored in the downstream storage tank 200, when the stope needs to be used, the control system opens the downstream storage valve 201 to discharge.

[0078] The above embodiments are only examples for clearly illustrating the present application, and are not limitations on the embodiments. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all the embodiments are not required to be exhausted, and the obvious changes or variations derived therefrom are still within the protection scope of the present application.

Claims

1. A viscous fluid intelligent delivery control system, characterized by: The system includes a delivery pipeline (300) and a fluid delivery pump (400) arranged on the delivery pipeline. The delivery pipeline (300) enables upstream and downstream fluid delivery with a height difference. An upwardly curved bend (300C) is provided at the connection between the upstream horizontal pipeline and the downstream downward pipeline. Fluid inside the upstream horizontal pipeline enters the downstream downward pipeline through the bend (300C), forming a closed chamber (300D) that is only connected to the downstream downward pipeline. The closed chamber (300D) is connected to a vacuum regulating valve (500). The vacuum level in the closed chamber (300D) is adjusted by the vacuum regulating valve (500) to control the downward flow rate and fluid pressure of the fluid in the downstream downward pipeline.

2. A viscous fluid intelligent delivery control system according to claim 1, wherein: A flow sensor (31), a temperature sensor (32) and a first pressure sensor (33) are provided on the upstream horizontal pipeline of the delivery pipeline (300), and a negative pressure sensor (51) is provided in the sealed chamber (300D); a second pressure sensor (34) is provided on the downstream falling pipeline of the delivery pipeline (300). The upstream horizontal pipeline of the conveying pipeline (300) is provided with a flow control valve (301), and the conveying pipeline (300) is provided with a shut-off valve to cut off the fluid conveyance in the pipeline.

3. A viscous fluid intelligent delivery control system according to claim 2, wherein: It also includes a delivery control module (600), in which the flow sensor (31), temperature sensor (32), first pressure sensor (34), second pressure sensor (34), and negative pressure sensor (51) measure and collect the actual fluid parameters in the delivery pipeline and transmit them to the delivery control module (600). The delivery control module (600) controls the flow control valve (301), fluid delivery pump (400), vacuum regulating valve (500), and shut-off valve based on the feedback of the actual fluid parameters.

4. A viscous fluid intelligent delivery control system according to any one of claims 1-3, wherein: A rupture disc (303) is also provided on the upstream horizontal pipeline of the conveying pipeline (300).

5. A viscous fluid intelligent delivery control system in accordance with claim 3, wherein: The delivery control module (600) integrates a hybrid decision model, which adopts a hybrid architecture of PID controller, regression prediction model and reinforcement learning model. The PID controller, regression prediction model and reinforcement learning model evaluate the confidence level based on the actual fluid parameters in the delivery pipeline. The actual fluid parameters in the current delivery pipeline are input into the PID controller, regression prediction model and reinforcement learning model respectively to output three sets of suggested control parameters. Based on the confidence level of the PID controller, regression prediction model and reinforcement learning model, the three sets of suggested control parameters are dynamically weighted and fused to obtain the final control parameters fed back by the delivery control module (600).

6. A viscous fluid intelligent delivery control system as in claim 5, wherein: It also includes a virtual simulation model that runs synchronously with the system. The virtual simulation model is a physical theoretical model constructed by using the fluid dynamics model of the system's delivery pipeline, the valve flow characteristic curve, the performance curve of the fluid delivery pump, and the physical property parameters of the delivered fluid. By inputting control parameters, it predicts the theoretical fluid parameters corresponding to the system and compares the theoretical fluid parameters with the actual fluid parameters of the system to identify the fault type of the system.

7. A viscous fluid intelligent delivery control system as in claim 5, wherein: The regression prediction model employs a gradient boosting decision tree algorithm. It uses paired data of historical fluid parameters and corresponding optimized control parameters as training samples for supervised learning training of the model. In the first stage, offline model construction is carried out using training samples of historical fluid parameters and corresponding optimized control parameters. In the second stage, online incremental learning is performed on the offline model using a time decay weighted method. The trained regression prediction model predicts the suggested control parameters for the system based on the actual fluid parameters in the current delivery pipeline.

8. A viscous fluid intelligent delivery control system as described in claim 6 wherein: The reinforcement learning model employs a deep deterministic policy gradient algorithm, and is trained offline using theoretical fluid parameters and corresponding control parameters generated by the virtual simulation model. It utilizes a reward function based on fluid flow tracking accuracy, pressure stability, and system transport energy consumption. ， Where: Rt is the instantaneous reward calculated based on the corresponding control parameters; tracking the reward for the flow of fluid, , the actual flow of fluid in the current system, the set target flow, a flow proportionality factor; rewarding the pressure stability of the fluid, , is the actual pressure of the fluid in the current system, is the set pressure optimum value, is the set pressure safety upper limit, , are respectively the pressure proportionality coefficients; rewarding the system for energy consumption of delivery, , pump frequency of the fluid delivery pump of the current system, is a coefficient of proportionality for energy consumption; Punishment for security incidents , , , They are respectively , , , The reward weighting coefficient.

9. The intelligent transport control system for viscous fluids according to claim 5, characterized in that: The final control parameters fed back by the conveying control module (600) Weighted fusion is performed using the following formula: ， in, Recommended control parameters output by the PID controller Recommended control parameters for the regression prediction model output. To suggest control parameters for reinforcement learning models, These are the real-time weighting coefficients for the PID controller. For the real-time weight coefficients output by the regression prediction model, To enhance the real-time weight coefficients of the learning model, if the system's fluid parameters are in a severe transient process, improve... At the same time reduce and If the system's fluid parameters are in a steady state or a slowly changing process, the real-time weighting coefficients are calculated using the following formula based on the confidence levels assessed by the PID controller, regression prediction model, and reinforcement learning model: ， in, The real-time weight coefficients for calculating PID controllers, regression prediction models, or reinforcement learning models. This represents the confidence level of the corresponding PID controller, regression prediction model, or reinforcement learning model. The confidence level for the PID controller is set to a constant base confidence level. To determine the confidence level of the regression prediction model, the Mahalanobis distance or similarity metric between the actual fluid parameters input to the model and the training samples is selected. A closer Mahalanobis distance or a more closely matched similarity distribution indicates a higher confidence level. The higher; To enhance the confidence of the reinforcement learning model, the value function of the output action in its policy network is evaluated using the actual fluid parameters input to the reinforcement learning model. A higher value indicates greater confidence. The higher.

10. A viscous fluid intelligent transport control system according to claim 9, characterized in that: The conveying control module (600) is also equipped with a safety verification module, which compares the final control parameters with the mechanical limit parameters of the equipment under feedback control. If the final control parameters do not exceed the mechanical limit parameters of the equipment, the conveying control module (600) outputs according to the final control parameters; If the final control parameters exceed the mechanical limit parameters of the equipment, the conveying control module (600) outputs the recommended control parameters according to the PID controller.

Images