Neural network-based concrete pouring process parameter adaptive control system

Abstract

The present application relates to the technical field of intelligent construction and engineering machinery automatic control, in particular to a concrete pouring process parameter adaptive control system based on neural network, comprising: a data space-time alignment unit for collecting high-frequency pumping pressure data and lagging rheological parameter data in the concrete pumping process to obtain a corrected system state vector; a cognitive introspection monitoring unit for calling a shadow monitor to generate a shadow monitoring state vector and calculating a rheological flow shape deviation index; a dynamic calculation scheduling unit for generating a nonlinear model predictive control instruction and a robust rule control instruction, performing smooth switching and weighting processing on the nonlinear model predictive control instruction and the robust rule control instruction based on the rheological flow shape deviation index to obtain a final pumping control instruction; the present application solves the problem of lack of self-reliability evaluation of traditional control models and improves the real-time performance and accuracy of neural network models in working condition prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction and automated control technology for engineering machinery, specifically to an adaptive control system for concrete pouring process parameters based on neural networks. Background Technology

[0002] In the current concrete pouring construction and pumping control environment, the system needs to process high-frequency pumping pressure data from pressure sensor arrays and rheological parameter data fed back by visual analysis or rheometers in real time. These multi-source heterogeneous data have significant differences in sampling frequency and acquisition time.

[0003] Existing control schemes generally rely on preset physical models for parameter adjustment. However, in actual operations, the acquisition of rheological parameters often involves a time lag, leading to a spatiotemporal misalignment between real-time pressure data and rheological data, directly affecting the prediction accuracy of neural networks or physical models. Furthermore, traditional models lack real-time monitoring mechanisms for their own cognitive biases, making it difficult to detect model distortion under abnormal conditions such as sensor failure or concrete segregation. Moreover, highly complex nonlinear model predictive control algorithms, in pursuing optimal solutions, easily overlook the competitive constraint between computation time and concrete physical setting time, resulting in an inability to switch to a safety-safety strategy in a timely manner when the model fails or computation times out, increasing the risk of pipe blockage and equipment damage. Therefore, how to eliminate the spatiotemporal bias of multi-source heterogeneous data, quantitatively evaluate the cognitive reliability of physical models in real time, and thereby achieve adaptive and smooth switching between efficient optimization control and safe and robust control has become an urgent technical problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an adaptive control system for concrete pouring process parameters based on neural networks. Specifically, the technical solution of this invention includes:

[0005] A neural network-based adaptive control system for concrete pouring process parameters includes:

[0006] The data spatiotemporal alignment unit is used to collect high-frequency pumping pressure data and hysteretic rheological parameter data during the concrete pumping process. It uses a physical information neural network and Kalman filter algorithm to perform spatiotemporal deviation correction processing on the high-frequency pumping pressure data and hysteretic rheological parameter data to obtain the corrected system state vector.

[0007] The cognitive introspection monitoring unit is used to call the shadow monitor to generate a shadow monitoring state vector, and calculate the rheological manifold deviation index based on the corrected system state vector and the shadow monitoring state vector.

[0008] The dynamic calculation and scheduling unit is used to generate nonlinear model predictive control commands and robust rule-based control commands. Based on the rheological manifold deviation index, the nonlinear model predictive control commands and robust rule-based control commands are smoothly switched and weighted to obtain the final pumping control commands.

[0009] Optionally, the spatiotemporal deviation correction process performed by the data spatiotemporal alignment unit is as follows:

[0010] A pre-trained physical information neural network model is obtained. High-frequency pumping pressure data and hysteresis rheological parameter data are input into the physical information neural network model to generate prior state estimates including time delay errors.

[0011] The residual between the real-time observation vector and the prior state estimate is calculated, and the Kalman gain matrix is ​​used to calculate the gain of the residual to obtain the correction term.

[0012] The correction term is superimposed on the prior state estimate to obtain the corrected system state vector;

[0013] The corrected system state vector includes the flow velocity values ​​and local equivalent viscosity values ​​for each segment of the pipeline.

[0014] Optionally, the calculation process for the rheological manifold deviation index performed by the cognitive introspection monitoring unit is as follows:

[0015] Calculate the Euclidean distance between the corrected system state vector and the shadow monitoring state vector in the state space, and use the variance of historical state data to perform dimensionless processing on the Euclidean distance to obtain the data consistency component.

[0016] Extract the velocity field data from the corrected system state vector, perform divergence calculation on the velocity field data, and obtain the velocity field divergence value;

[0017] The velocity field divergence values ​​are subjected to time averaging and tolerance normalization to obtain the physical consistency component.

[0018] By using preset weighting coefficients, the data consistency component and the physical consistency component are weighted and summed to obtain the rheological manifold deviation index.

[0019] Optionally, the process of obtaining physically consistent components includes:

[0020] Obtain the duration of the sliding observation time window, and integrate the square of the velocity field divergence value within the duration of the sliding observation time window to obtain the divergence square integral value.

[0021] Obtain the maximum permissible divergence residual, calculate the product of the sliding observation time window duration and the square of the maximum permissible divergence residual, and obtain the normalization factor;

[0022] Dividing the squared divergence integral by the normalization factor yields the physical consistency component.

[0023] Optionally, the instruction generation process executed by the dynamic computation scheduling unit is as follows:

[0024] Construct a constrained optimization problem, solve the constrained optimization problem to calculate the optimal control quantity, and set the optimal control quantity as a nonlinear model predictive control command.

[0025] Obtain the upper limit of the pipeline design safe pressure and the current measured pump inlet pressure, and calculate the difference between the upper limit of the pipeline design safe pressure and the current measured pump inlet pressure;

[0026] Multiply the difference by the proportional gain, and add the result of the multiplication to the minimum flow rate to obtain the minimum control value. Set the minimum control value as the robust rule control instruction.

[0027] Optionally, the smooth switching weighted processing procedure is as follows:

[0028] Obtain the critical threshold of cognitive distortion and calculate the difference between the rheological manifold deviation index and the critical threshold of cognitive distortion;

[0029] The smooth transition factor is calculated using the Sigmoid function based on the difference.

[0030] The robust rule control instructions are weighted using a smooth transition factor to obtain the first weighted component;

[0031] Calculate the complement of the smooth transition factor, and use the complement to weight the predictive control command of the nonlinear model to obtain the second weighted component;

[0032] The first weighted component is added to the second weighted component to obtain the final pumping control command.

[0033] Optionally, a time window detection unit is also included, which is used to perform physical and computational time race constraint analysis.

[0034] Obtain pipeline volume data, minimum flow rate data to prevent segregation at the current viscosity, and concrete initial setting time constant. Calculate the remaining safe time for segregation using the pipeline volume data and minimum flow rate data to prevent segregation, and determine the theoretical maximum residence time by combining the concrete initial setting time constant.

[0035] Obtain the time difference between the current moment and the stirring moment to get the time already consumed. Subtract the consumed time from the theoretical maximum residence time to get the remaining physical safety time.

[0036] Obtain the remaining number of iterations of the optimization algorithm and the average CPU time per iteration. Multiply the remaining number of iterations, the average CPU time per iteration, and the safety redundancy coefficient to obtain the estimated computation time.

[0037] The remaining physical safety time is compared and analyzed with the calculated estimated time.

[0038] Optionally, the time window detection unit is also used to perform the following operations based on the comparison analysis results:

[0039] In response to the physical remaining safety time being less than or equal to the calculated estimated time, the smooth transition factor is forced to be set to the maximum value, and the robust rule control command is output as the final pumping control command.

[0040] In response to the physical remaining safety time being greater than the calculated estimated time, the smooth transition factor calculated by the Sigmoid function is maintained, and the final pumping control command obtained by adding the first weighted component and the second weighted component is output.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. This invention uses a data spatiotemporal alignment unit, a physical information neural network and a Kalman filter algorithm to correct the spatiotemporal deviation between high-frequency pumping pressure data and hysteretic rheological parameter data; it effectively eliminates the spatiotemporal misalignment of data caused by the lag in the acquisition of rheological parameters, and achieves high-precision estimation of the flow velocity and viscosity in the pipeline at the current moment under the premise of data lag, thereby improving the real-time performance and accuracy of the neural network model in predicting operating conditions.

[0043] 2. This invention constructs a cognitive self-reflection monitoring mechanism, which calculates the rheological manifold deviation index from both data consistency and physical consistency dimensions by comparing the output of the physical model with the output of the pure data-driven shadow monitor. This mechanism can quantify the degree of cognitive distortion of the physical model in real time and promptly capture abnormal working conditions that violate the physical conservation laws, such as sensor failure or concrete segregation, thus solving the problem that traditional control models lack their own reliability assessment.

[0044] 3. This invention designs a dynamic computational scheduling strategy to generate nonlinear model predictive control commands and robust rule control commands in parallel, and uses a smooth transition function to perform weighted switching based on the deviation index; it realizes adaptive adjustment that executes efficient optimization control when the model is reliable and automatically degrades to a minimum safety control when the model is distorted, which not only improves the efficiency of pumping operation, but also effectively eliminates the oscillation shock to the hydraulic system caused by the instant of mode switching;

[0045] 4. This invention introduces a time window detection unit to perform competitive constraint analysis between physical setting time and algorithm calculation time. When the remaining physical safety time is insufficient to support complex optimization calculations, the system forcibly switches to robust rule control mode to prevent concrete from hardening in the pipe due to algorithm solution timeout. This ensures that the pump can still obtain the minimum flow rate command in extreme cases such as calculation blockage, significantly reducing the risk of pipe blockage. Attached Figure Description

[0046] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0047] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0049] Example 1:

[0050] Please see Figure 1 An adaptive control system for concrete pouring process parameters based on neural networks includes:

[0051] The data spatiotemporal alignment unit is used to collect high-frequency pumping pressure data and hysteretic rheological parameter data during the concrete pumping process. It uses a physical information neural network and Kalman filter algorithm to perform spatiotemporal deviation correction processing on the high-frequency pumping pressure data and hysteretic rheological parameter data to obtain the corrected system state vector.

[0052] The cognitive introspection monitoring unit is used to call the shadow monitor to generate a shadow monitoring state vector, and calculate the rheological manifold deviation index based on the corrected system state vector and the shadow monitoring state vector.

[0053] The dynamic calculation and scheduling unit is used to generate nonlinear model predictive control commands and robust rule-based control commands. Based on the rheological manifold deviation index, the nonlinear model predictive control commands and robust rule-based control commands are smoothly switched and weighted to obtain the final pumping control commands.

[0054] This embodiment provides a neural network-based adaptive control system for concrete pouring process parameters to address the problems of spatiotemporal misalignment of multi-source heterogeneous data and distortion of physical models. The system includes a data spatiotemporal alignment unit configured to acquire in real-time high-frequency pumping pressure data generated by a pressure sensor array installed on the pump pipe, as well as hysteretic rheological parameter data fed back through a visual analysis system or rheometer. This unit eliminates the time lag of rheological parameters by running a Physical Information Neural Network (PINN) and a Kalman filter algorithm. Output the corrected system state vector The cognitive introspection monitoring unit is configured to run a purely data-driven shadow monitor to generate shadow monitoring state vectors. ;

[0055] The shadow monitor specifically employs a Long Short-Term Memory (LSTM) network architecture. Its input layer receives historical pumping pressure sequences from the past 50 sampling periods, current pumping displacement commands, and hydraulic oil temperature data. Temporal features are extracted through three LSTM hidden layers, and the current shadow monitoring state vector is mapped out by a fully connected layer. The network weights are obtained through offline training based on historical normal operating condition datasets. The system outputs the physical model's data... With the output of the shadow model Perform adversarial comparisons and calculate the rheological manifold deviation index to quantify the cognitive bias of the current physical model. The dynamic computation scheduling unit is configured to generate nonlinear model predictive control commands based on constraint optimization in parallel. and robust rule-based control instructions based on security rules and based on The two types of instructions are smoothly weighted and switched to generate the final pumping control instruction sent to the pump truck's hydraulic system. This architecture realizes a logical closed loop from data perception and state quantification to decision execution. Through a cognitive introspection mechanism, it performs efficient optimization when the model is accurate and automatically degrades to a minimum control when the model is distorted.

[0056] Example 2:

[0057] The spatiotemporal deviation correction process performed by the data spatiotemporal alignment unit is as follows:

[0058] A pre-trained physical information neural network model is obtained. High-frequency pumping pressure data and hysteresis rheological parameter data are input into the physical information neural network model to generate prior state estimates including time delay errors.

[0059] The residual between the real-time observation vector and the prior state estimate is calculated, and the Kalman gain matrix is ​​used to calculate the gain of the residual to obtain the correction term.

[0060] The correction term is superimposed on the prior state estimate to obtain the corrected system state vector;

[0061] The corrected system state vector includes the flow velocity values ​​and local equivalent viscosity values ​​for each segment of the pipeline.

[0062] The data spatiotemporal alignment unit performs spatiotemporal bias correction, aiming to correct prediction biases caused by input lag in neural networks using real-time observation data; and to acquire pre-trained physical information neural network models. The model is trained offline based on the laws of fluid mechanics; where time lag is... The volume of the pipe between the pump inlet and the rheological parameter acquisition point is obtained through real-time calculation. Divide it by the estimated average flow velocity at the current moment to obtain the dynamic lag time. Based on this, the rheological parameters corresponding to the historical moment are extracted from the cache queue; the current moment is... High-frequency pumping pressure data With time lag Hysteresis rheological parameter data Input to In the middle; due to the spatiotemporal asymmetry of the input data, the network output includes prior state estimates containing time delay errors. Extended Kalman filter logic is introduced to calculate the corrected system state vector. :

[0063]

[0064] in, Includes flow velocity values ​​for each section of the pipe. and local equivalent viscosity values ; It contains only the real-time observation vectors actually measured by the sensor at the current moment; The observation matrix is ​​obtained by linearizing the fluid dynamics constitutive equations. Specifically, a plastic fluid constitutive model is used to describe the properties of concrete, and its simplified equation of state is as follows: ,in For pressure, For flow rate, Plastic viscosity, For pipe diameter, For the position of manager; The value of the concrete yield stress, measured by a rheometer in the most recent static test, is input as a known constant into the equation of state; this nonlinear equation at the current operating point... Performing a Taylor expansion at the given point and ignoring higher-order terms yields a linearized observation matrix. Used to map the state space to the observation space; The Kalman gain matrix is ​​obtained by recursively minimizing the posterior error covariance matrix; this scheme utilizes Gain calculation of residuals using real-time pressure data Dynamic compensation is due to the hysteresis of rheological parameters. The error introduced allows for a high-precision estimate of the flow velocity and viscosity within the pipe at the current moment, even with lagging rheological data.

[0065] Example 3:

[0066] The calculation process for the rheological manifold deviation index performed by the cognitive introspection monitoring unit is as follows:

[0067] Calculate the Euclidean distance between the corrected system state vector and the shadow monitoring state vector in the state space, and use the variance of historical state data to perform dimensionless processing on the Euclidean distance to obtain the data consistency component.

[0068] Extract the velocity field data from the corrected system state vector, perform divergence calculation on the velocity field data, and obtain the velocity field divergence value;

[0069] The velocity field divergence values ​​are subjected to time averaging and tolerance normalization to obtain the physical consistency component.

[0070] By using preset weighting coefficients, the data consistency component and the physical consistency component are weighted and summed to obtain the rheological manifold deviation index.

[0071] Cognitive introspection monitoring unit calculates rheological manifold deviation index This calculation quantifies the degree of cognitive distortion in the physical model; it integrates two dimensions: data consistency and physical consistency; and it calculates the corrected system state vector. With shadow monitoring state vector Euclidean distance in state space And utilize data from historical state databases Historical state data variance The distance is dimensionless to obtain the data consistency component; extraction Velocity field data Calculate its divergence The physical consistency component is obtained by performing time averaging and tolerance normalization on the divergence values. Using weighting coefficients Perform a weighted summation:

[0072]

[0073] It is based on a calibration dataset containing historical pipe blockage accident sequences. Backtesting analysis was performed to optimize the alarm accuracy rate (TPR) within a preset time window before the accident occurred. This indicator was obtained by detecting whether the physical model deviated from historical statistical patterns and whether it violated the law of conservation of mass, thereby capturing sensor failures or concrete segregation anomalies.

[0074] Example 4:

[0075] The process of obtaining the physical consistency components includes:

[0076] Obtain the duration of the sliding observation time window, and integrate the square of the velocity field divergence value within the duration of the sliding observation time window to obtain the divergence square integral value.

[0077] Obtain the maximum permissible divergence residual, calculate the product of the sliding observation time window duration and the square of the maximum permissible divergence residual, and obtain the normalization factor;

[0078] Dividing the squared divergence integral by the normalization factor yields the physical consistency component.

[0079] Physical consistency components The calculation aims to assess whether the current flow field satisfies the continuity equation constraints; the duration of the sliding observation time window is set. ; Obtain the physical spacing of sensors arranged along the pipeline or the preset finite element mesh segment length. The spatial derivative of the discrete velocity field is calculated using the central difference method or the backward difference method. Within this window, the velocity field divergence values ​​are... Integrate the square of the product to obtain the squared divergence integral value; set the maximum allowable divergence residual based on the concrete compressibility coefficient. ;calculate and The product of these is used as a normalization factor; the integral of the squared divergence is divided by the normalization factor:

[0080]

[0081] By time window The inner integral smooths out instantaneous numerical fluctuations and introduces... Perform dimensionless transformation to ensure A scalar quantity that reflects the degree of violation of fluid physical properties.

[0082] Example 5:

[0083] The instruction generation process executed by the dynamic computation scheduling unit is as follows:

[0084] Construct a constrained optimization problem, solve the constrained optimization problem to calculate the optimal control quantity, and set the optimal control quantity as a nonlinear model predictive control command.

[0085] Obtain the upper limit of the pipeline design safe pressure and the current measured pump inlet pressure, and calculate the difference between the upper limit of the pipeline design safe pressure and the current measured pump inlet pressure;

[0086] Multiply the difference by the proportional gain, and add the result of the multiplication to the minimum flow rate to obtain the minimum control value. Set the minimum control value as the robust rule control instruction.

[0087] The dynamic computation scheduling unit is constructed as a constrained optimization problem with the objective function of minimizing state tracking error and control increment, while satisfying physical constraints. Specifically, the prediction time domain is defined as... Control time domain is Construct the following quadratic programming objective function. :

[0088]

[0089] in, To predict the future based on the current state Predicted value at time, The preset ideal flow rate and viscosity reference trajectory; To control the increment; and These are the state error weight matrix and the control energy consumption weight matrix, used to balance tracking accuracy and control stability. The constraints include:

[0090] System dynamic constraints: The local linearization model of the neural network in Example 2 is adopted;

[0091] Actuator constraints: and ;

[0092] Safety constraints: Pumping pressure Solving this optimization problem yields the control increment sequence, and solving for the optimal control input at the current time step yields the optimal control input, which is then defined as a nonlinear model predictive control command. Parallel execution of expert-based P control logic: obtaining the upper limit of the pipeline design safety pressure. Compared with the current measured pump inlet pressure Calculate the difference between the two; then multiply this difference by the proportional gain. Multiply and sum the minimum flow rate displacements. The minimum control value is obtained and set as a robust rule control instruction. :

[0093]

[0094] To prevent the calculation results from exceeding the physical limits of the equipment, Perform saturation limiting: if Greater than the maximum displacement of the pump truck Then take If it is less than the minimum displacement Then take ; Minimum flow rate; Suitable for normal operating conditions with high model reliability to improve efficiency. It is suitable for abnormal operating conditions with model distortion to ensure equipment safety and prevent downtime and pipe blockage.

[0095] Example 6:

[0096] The smooth switching weighted processing procedure is as follows:

[0097] Obtain the critical threshold of cognitive distortion and calculate the difference between the rheological manifold deviation index and the critical threshold of cognitive distortion;

[0098] The smooth transition factor is calculated using the Sigmoid function based on the difference.

[0099] The robust rule control instructions are weighted using a smooth transition factor to obtain the first weighted component;

[0100] Calculate the complement of the smooth transition factor, and use the complement to weight the predictive control command of the nonlinear model to obtain the second weighted component;

[0101] The first weighted component is added to the second weighted component to obtain the final pumping control command.

[0102] Using the rheological manifold deviation index Perform a smooth switch; obtain settings under normal operating conditions. Cognitive distortion threshold three times the mean ;calculate and The difference is used to calculate the smooth transition factor using the Sigmoid function. :

[0103]

[0104] in This is used to switch the sensitivity coefficient; its value is typically set between 0.5 and 5.0, and it is used to adjust the steepness of the activation function; utilizing... right The first weighted component is obtained by weighting, and then the complement is used. right The weighted average yields a second weighted component, which is then summed to obtain the final pumping control command. :

[0105]

[0106] This mechanism ensures the continuity of control transfer and eliminates the tremors during mode switching.

[0107] Example 7:

[0108] This system also includes a time window detection unit, which is used to perform physical and computational time race constraint analysis.

[0109] Obtain pipeline volume data, minimum flow rate data to prevent segregation at the current viscosity, and concrete initial setting time constant. Calculate the remaining safe time for segregation using the pipeline volume data and minimum flow rate data to prevent segregation, and determine the theoretical maximum residence time by combining the concrete initial setting time constant.

[0110] Obtain the time difference between the current moment and the stirring moment to get the time already consumed. Subtract the consumed time from the theoretical maximum residence time to get the remaining physical safety time.

[0111] Obtain the remaining number of iterations of the optimization algorithm and the average CPU time per iteration. Multiply the remaining number of iterations, the average CPU time per iteration, and the safety redundancy coefficient to obtain the estimated computation time.

[0112] The remaining physical safety time is compared and analyzed with the calculated estimated time.

[0113] The time window detection unit performs a physical and computational time competition constraint analysis; and acquires pipeline volume data. Minimum flow rate to prevent segregation at the current viscosity Simultaneously, the initial setting time constant of the concrete was obtained. This constant was obtained from concrete mix design experiments;

[0114] Calculate the remaining physical safety time First, calculate the remaining safe setting time to prevent irreversible hardening of the concrete. :

[0115]

[0116] Simultaneously, calculate the remaining safe time for preventing dynamic segregation. :

[0117]

[0118] The smaller of the two values ​​mentioned above will be taken as the final physical remaining safety time. ,Right now ; Obtain the remaining number of iterations for the optimization algorithm Average CPU time per iteration Combined with safety redundancy coefficient Calculate the estimated time consumption :

[0119]

[0120] Will and By comparing and quantifying the time competition between physical condensation and algorithmic solution.

[0121] Example 8:

[0122] The time window detection unit is also used to perform the following operations based on the comparison analysis results:

[0123] In response to the physical remaining safety time being less than or equal to the calculated estimated time, the smooth transition factor is forced to be set to the maximum value, and the robust rule control command is output as the final pumping control command.

[0124] In response to the physical remaining safety time being greater than the calculated estimated time, the smooth transition factor calculated by the Sigmoid function is maintained, and the final pumping control command obtained by adding the first weighted component and the second weighted component is output.

[0125] Based on the comparison results, mandatory intervention will be implemented; in response to In other words, the remaining physical time is insufficient to support MPC calculations, and the system forcibly applies a smooth transition factor. Set to the maximum value of 1, ignore incomplete MPC calculations, and directly output robust rule control instructions. As ; in response to Maintain the result calculated by the Sigmoid function. The output is obtained from weighted logic. This strategy, as a circuit breaker mechanism based on hard constraints, ensures that the pump can still receive the minimum flow rate command even in the event of computational blockage.

[0126] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An adaptive control system for concrete pouring process parameters based on neural networks, characterized in that, include: The data spatiotemporal alignment unit is used to collect high-frequency pumping pressure data and hysteretic rheological parameter data during the concrete pumping process. It uses a physical information neural network and Kalman filter algorithm to perform spatiotemporal deviation correction processing on the high-frequency pumping pressure data and hysteretic rheological parameter data to obtain the corrected system state vector. The cognitive introspection monitoring unit is used to call the shadow monitor to generate a shadow monitoring state vector, and calculate the rheological manifold deviation index based on the corrected system state vector and the shadow monitoring state vector. The dynamic calculation and scheduling unit is used to generate nonlinear model predictive control commands and robust rule-based control commands. Based on the rheological manifold deviation index, the nonlinear model predictive control commands and robust rule-based control commands are smoothly switched and weighted to obtain the final pumping control commands.

2. The adaptive control system for concrete pouring process parameters based on neural networks according to claim 1, characterized in that, The spatiotemporal deviation correction process performed by the data spatiotemporal alignment unit is as follows: A pre-trained physical information neural network model is obtained. High-frequency pumping pressure data and hysteresis rheological parameter data are input into the physical information neural network model to generate prior state estimates including time delay errors. The residual between the real-time observation vector and the prior state estimate is calculated, and the Kalman gain matrix is ​​used to calculate the gain of the residual to obtain the correction term. The correction term is superimposed on the prior state estimate to obtain the corrected system state vector; The corrected system state vector includes the flow velocity values ​​and local equivalent viscosity values ​​for each segment of the pipeline.

3. The adaptive control system for concrete pouring process parameters based on neural networks according to claim 1, characterized in that, The calculation process of the rheological manifold deviation index performed by the cognitive introspection monitoring unit is as follows: Calculate the Euclidean distance between the corrected system state vector and the shadow monitoring state vector in the state space, and use the variance of historical state data to perform dimensionless processing on the Euclidean distance to obtain the data consistency component. Extract the velocity field data from the corrected system state vector, perform divergence calculation on the velocity field data, and obtain the velocity field divergence value; The velocity field divergence values ​​are subjected to time averaging and tolerance normalization to obtain the physical consistency component. By using preset weighting coefficients, the data consistency component and the physical consistency component are weighted and summed to obtain the rheological manifold deviation index.

4. The adaptive control system for concrete pouring process parameters based on neural networks according to claim 3, characterized in that, The process of obtaining the physical consistency component includes: Obtain the duration of the sliding observation time window, and integrate the square of the velocity field divergence value within the duration of the sliding observation time window to obtain the divergence square integral value. Obtain the maximum permissible divergence residual, calculate the product of the sliding observation time window duration and the square of the maximum permissible divergence residual, and obtain the normalization factor; Dividing the squared divergence integral by the normalization factor yields the physical consistency component.

5. The adaptive control system for concrete pouring process parameters based on neural networks according to claim 1, characterized in that, The instruction generation process executed by the dynamic computing scheduling unit is as follows: Construct a constrained optimization problem, solve the constrained optimization problem to calculate the optimal control quantity, and set the optimal control quantity as a nonlinear model predictive control command. Obtain the upper limit of the pipeline design safe pressure and the current measured pump inlet pressure, and calculate the difference between the upper limit of the pipeline design safe pressure and the current measured pump inlet pressure; Multiply the difference by the proportional gain, and add the result of the multiplication to the minimum flow rate to obtain the minimum control value. Set the minimum control value as the robust rule control instruction.

6. The adaptive control system for concrete pouring process parameters based on neural networks according to claim 1, characterized in that, The smooth switching weighted processing procedure is as follows: Obtain the critical threshold of cognitive distortion and calculate the difference between the rheological manifold deviation index and the critical threshold of cognitive distortion; The smooth transition factor is calculated using the Sigmoid function based on the difference. The robust rule control instructions are weighted using a smooth transition factor to obtain the first weighted component; Calculate the complement of the smooth transition factor, and use the complement to weight the predictive control command of the nonlinear model to obtain the second weighted component; The first weighted component is added to the second weighted component to obtain the final pumping control command.

7. The adaptive control system for concrete pouring process parameters based on neural networks according to claim 6, characterized in that, It also includes a time window detection unit, which is used to perform physical and computational time race constraint analysis: Obtain pipeline volume data, minimum flow rate data to prevent segregation at the current viscosity, and concrete initial setting time constant. Calculate the remaining safe time for segregation using the pipeline volume data and minimum flow rate data to prevent segregation, and determine the theoretical maximum residence time by combining the concrete initial setting time constant. Obtain the time difference between the current moment and the stirring moment to get the time already consumed. Subtract the consumed time from the theoretical maximum residence time to get the remaining physical safety time. Obtain the remaining number of iterations of the optimization algorithm and the average CPU time per iteration. Multiply the remaining number of iterations, the average CPU time per iteration, and the safety redundancy coefficient to obtain the estimated computation time. The remaining physical safety time is compared and analyzed with the calculated estimated time.

8. The adaptive control system for concrete pouring process parameters based on neural networks according to claim 7, characterized in that, The time window detection unit is also used to perform the following operations based on the comparison analysis results: In response to the physical remaining safety time being less than or equal to the calculated estimated time, the smooth transition factor is forced to be set to the maximum value, and the robust rule control command is output as the final pumping control command. In response to the physical remaining safety time being greater than the calculated estimated time, the smooth transition factor calculated by the Sigmoid function is maintained, and the final pumping control command obtained by adding the first weighted component and the second weighted component is output.

Images