Intelligent equipment opening degree prediction control method and system based on time sequence graph characteristics

Abstract

The invention relates to the technical field of industrial control, in particular to an intelligent equipment opening degree prediction control method and system based on time sequence map characteristics, and the method comprises the steps: collecting the driving current, position feedback and load torque of an execution mechanism in real time, carrying out the phase-space reconstruction through a delay coordinate embedding technology, and generating a dynamic time sequence map; constructing a correlation matrix in real time; extracting a control characteristic vector containing a kinetic energy inertia characteristic item and a potential energy disturbance characteristic item from the map, and calculating a feed-forward compensation gain in combination with the decoupling matrix so as to predict a target control increment and drive an execution mechanism; and synchronously establishing a reward function based on deep reinforcement learning, and optimizing prediction model parameters in real time by using a deep deterministic strategy gradient algorithm. According to the method, the response lag generated by mechanical inertia is effectively counteracted through the nonlinear evolution characteristics of the deep excavation execution mechanism, and active observation and compensation of multi-source external load interference are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, specifically to a method and system for predictive control of the opening degree of intelligent equipment based on time-series graph features. Background Technology

[0002] Traditional control methods are mostly based on conventional PID control or simple linear predictive control. However, as industrial intelligent equipment becomes more complex and integrated, existing control technologies often perform poorly when faced with the following challenges: Traditional intelligent device control is mostly based on linear models, which makes it difficult to effectively capture the nonlinear and time-varying characteristics of actuators under the influence of wear and environmental fluctuations. Furthermore, due to the response lag effect caused by its own mechanical inertia and transmission delay, the system is prone to overshoot or oscillation during rapid adjustment. At the same time, due to the lack of active observation and compensation mechanisms for multi-source external interference, and the failure to deeply explore the time-series coupling characteristics between multi-dimensional parameters such as drive current and torque, the control accuracy and real-time tracking performance of the system under complex working conditions cannot meet the requirements of high-precision scenarios.

[0003] To address this, a predictive control method and system for the opening degree of intelligent devices based on time-series graph features is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predictive control of the opening degree of intelligent devices based on time-series graph features, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A predictive control method for the opening degree of intelligent devices based on time-series graph features includes: The system collects the operating status parameters of the actuators of intelligent devices in real time, reconstructs the phase space of the operating status parameters using delayed coordinate embedding technology, and maps them to generate a dynamic time series graph. By calculating the coupling weights between nodes in the dynamic time series graph, the system constructs the mutual correlation matrix of the intelligent devices in real time. The control feature vector is extracted from the dynamic time series spectrum; the control feature vector includes: a kinetic energy inertia feature term to compensate for response lag and a potential energy disturbance feature term to characterize the intensity of external load disturbance; A predictive control model with an integrated disturbance observer is constructed. A decoupling matrix is ​​generated based on the inter-correlation matrix, and the feedforward compensation gain is calculated in combination with the control feature vector. Based on the decoupling matrix and the feedforward compensation gain, and combined with the target command curve, the target control increment used to offset the hysteresis effect in the current control cycle is predicted. The target control increment is converted into an execution command to drive the actuator, and the measured position feedback signal of the actuator is synchronously transmitted back; a reward function based on deep reinforcement learning is established to optimize the parameters inside the predictive control model in real time.

[0006] Preferably, the dynamic time series map generation process includes: The operating status parameters include drive current, position feedback, and load torque; The operating status parameters are collected in real time and aligned with multidimensional data streams to eliminate time asynchrony deviations in the sensor feedback process, forming a standardized original observation sequence. Based on the fluctuation characteristics of the original observation sequence, the embedding dimension and delay step size of the delay coordinate embedding technique are adaptively determined. The original observation sequence is reconstructed into a phase space trajectory describing the evolution of the actuator's motion state through mapping. Using the evolving phase points in the phase space trajectory as graph nodes, the topological connection weights are calculated based on the Euclidean distance and evolution slope between phase points to generate a dynamic time series graph.

[0007] Preferably, the specific process of reconstructing the original observation sequence into a phase space trajectory describing the evolution of the actuator's motion state includes: calculating the average mutual information of each parameter in the original observation sequence; using the time corresponding to the first local minimum point as the adaptive delay step size; determining the minimum embedding dimension by calculating the global embedding gain of the observation sequence in different dimensions; performing coordinate mapping and time-domain translation on the aligned operating state parameters based on the adaptive delay step size and the minimum embedding dimension; synthesizing a phase space vector describing the evolution of the actuator from the current steady state to the target opening in a multidimensional Euclidean space; and generating a phase space trajectory by tracing the geometric evolution path of the phase space vector within adjacent sampling periods.

[0008] Preferably, the control feature vector is calculated by taking the evolution rate and phase space volume contraction rate of the current node relative to the historical trajectory nodes in the dynamic time-series graph, and mapping the evolution rate to a kinetic inertial feature term characterizing the mechanical motion inertia of the actuator; simultaneously extracting the gradient value of the node coupling strength characterizing the load torque change in the graph, and mapping the gradient value to a potential energy disturbance feature term characterizing the resistance of the external environment to the surface of the actuator; and normalizing and combining the kinetic inertial feature term and the potential energy disturbance feature term to construct a control feature vector that migrates in real time with the operating conditions of the actuator.

[0009] Preferably, the predictive control model of the integrated disturbance observer includes a matrix decoupling layer, a feature feedforward layer, a disturbance observation layer, and a gain prediction layer; The matrix decoupling layer performs orthogonalization processing on the interrelated matrices, extracts the mapping feature basis between the physical parameters of the actuator, and constructs the decoupling matrix. The feature feedforward layer performs parameterized transformation on the kinetic energy inertia feature term and the potential energy disturbance feature term in the control feature vector, and calculates the feedforward compensation gain in real time by combining the output vector of the decoupling matrix. The disturbance observation layer superimposes the target control increment of the previous control cycle with the position feedback in real time to calculate the theoretical expected position at the current moment, and compares the theoretical expected position with the measured position feedback signal transmitted back in real time at the current moment to generate a real-time disturbance quantity. The gain prediction layer retrieves the preset running trajectory as the target command curve, integrates the decoupling matrix, feedforward compensation gain and the real-time disturbance, and performs time-domain rolling calculation within the sliding prediction window to generate the target control increment within the current control cycle.

[0010] Preferably, the specific process of converting the target control increment into an execution command to drive the actuator includes superimposing the target control increment onto the measured position feedback signal at the current moment to synthesize a target position command; using drive mapping logic to convert the target position command into a pulse drive signal and outputting it to the actuator drive end; and extracting the drive current and load torque during the drive process in real time and transmitting them back to the node update sequence of the dynamic time series graph.

[0011] Preferably, the specific process of parameter optimization within the real-time predictive control model is as follows: The predictive control model is used as a reinforcement learning agent, and the operating state parameters of the actuator and the control feature vector are used as environmental state inputs; a multi-objective weighted reward function is constructed, which includes a trajectory tracking reward, a dynamic response reward, and a control smoothing reward. The target control increment is used as the action output to obtain the calculation result of the multi-objective weighted reward function; a deep deterministic policy gradient algorithm is adopted to maximize the cumulative reward and calculate the gradient correction value for the decoupling matrix parameters and feedforward gain weights in the predictive control model, thereby optimizing the parameters inside the predictive control model in real time.

[0012] A predictive control system for the opening degree of intelligent devices based on time-series graph features includes: The graph construction module collects the operating status parameters of the actuators of intelligent devices in real time, reconstructs the phase space of the operating status parameters using delayed coordinate embedding technology, and maps them to generate a dynamic time series graph; by calculating the coupling weights between each node in the dynamic time series graph, the mutual correlation matrix of the intelligent devices is constructed in real time. Control vector module: extracts control feature vectors from the dynamic time series graph; the control feature vectors include: kinetic energy inertia feature term to compensate for response lag and potential energy disturbance feature term to characterize the intensity of external load disturbance; Decoupling Increment Module: Constructs a predictive control model with an integrated disturbance observer, generates a decoupling matrix based on the cross-correlation matrix, and calculates the feedforward compensation gain in combination with the control feature vector; based on the decoupling matrix and the feedforward compensation gain, and in combination with the target command curve, predicts the target control increment used to offset the hysteresis effect in the current control cycle. Predictive control module: converts the target control increment into execution instructions to drive the actuator, and synchronously transmits back the measured position feedback signal of the actuator; establishes a reward function based on deep reinforcement learning, and optimizes the parameters inside the predictive control model in real time.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application utilizes delayed coordinate embedding technology to perform multi-dimensional data stream alignment and phase space reconstruction of drive current, position feedback and load torque, mapping to generate dynamic time series spectrum and constructing interrelated matrices; this mechanism can deeply explore the nonlinear evolution characteristics of actuators under wear and environmental fluctuations, effectively solving the pain point that traditional linear models are difficult to describe the dynamic characteristics of complex working conditions, and realizing accurate characterization and real-time mapping of the evolution path of equipment operating status.

[0014] 2. This application extracts kinetic inertial characteristic terms from the dynamic time series spectrum and calculates the feedforward compensation gain by combining the decoupling matrix. In the predictive control model, it predicts and offsets the lag effect of the current control cycle in advance. This design can effectively offset the response lag of the actuator caused by mechanical inertia and transmission delay, significantly suppress the risk of overshoot and oscillation in the system during rapid adjustment, and ensure the high synchronization of command execution and feedback signal in the time dimension.

[0015] 3. This application constructs a predictive control model with an integrated disturbance observer, compares the theoretical expected position with the measured feedback signal in real time to identify the disturbance amount, and constructs an adaptive law to correct the model by combining the potential energy disturbance characteristic term. This active observation and dynamic compensation mechanism enables the system to adjust parameters in real time for multi-source external load disturbances, thereby enhancing the control robustness and operational stability of intelligent devices in complex variable load scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a predictive control method for the opening degree of intelligent devices based on time-series graph features. Figure 2 This is a schematic diagram of the dynamic time series graph generation process of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent device opening degree prediction and control system based on time-series graph features. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1

[0019] Please see Figure 1 This invention provides a method for predictive control of the opening degree of intelligent devices based on time-series graph features. The technical solution is as follows: A predictive control method for the opening degree of intelligent devices based on time-series graph features includes: The system collects the operating status parameters of the actuators of intelligent devices in real time, reconstructs the phase space of the operating status parameters using delayed coordinate embedding technology, and maps them to generate a dynamic time series graph. By calculating the coupling weights between nodes in the dynamic time series graph, the system constructs the mutual correlation matrix of the intelligent devices in real time. The control feature vector is extracted from the dynamic time series spectrum; the control feature vector includes: a kinetic energy inertia feature term to compensate for response lag and a potential energy disturbance feature term to characterize the intensity of external load disturbance; A predictive control model with an integrated disturbance observer is constructed. A decoupling matrix is ​​generated based on the inter-correlation matrix, and the feedforward compensation gain is calculated in combination with the control feature vector. Based on the decoupling matrix and the feedforward compensation gain, and combined with the target command curve, the target control increment used to offset the hysteresis effect in the current control cycle is predicted. The target control increment is converted into an execution command to drive the actuator, and the measured position feedback signal of the actuator is synchronously transmitted back; an adaptive law with the prediction deviation as input is established to correct the predictive control model and the weight of the decoupling matrix in real time.

[0020] Real-time acquisition of operating status parameters of the actuator of the intelligent device, including drive current, position feedback and load torque; By integrating a multi-channel analog signal acquisition circuit and communication interface into the actuator, a hardware sensor sensing network is constructed. Using current transformers, photoelectric encoders and torque transmitters, drive current, position feedback and load torque are synchronously captured at a microsecond-level sampling frequency. Among them, the drive current is converted into a voltage signal by a Hall current sensor and then converted by an ADC to characterize the magnetic field torque and power output state inside the motor; the position feedback is transmitted by a high-precision absolute encoder to count pulses, which is used to accurately map the current real-time opening position of the actuator in the physical space; the load torque is obtained by a dedicated sensor to directly quantify the load intensity applied by external mechanical resistance and environmental disturbances. The dynamic time series map generation process includes: The operating status parameters are collected in real time and aligned with multidimensional data streams to eliminate time asynchrony deviations in the sensor feedback process, forming a standardized original observation sequence. Based on the fluctuation characteristics of the original observation sequence, the embedding dimension and delay step size of the delay coordinate embedding technique are adaptively determined. The original observation sequence is reconstructed into a phase space trajectory describing the evolution of the actuator's motion state through mapping. Using the evolving phase points in the phase space trajectory as graph nodes, the topological connection weights are calculated based on the Euclidean distance and evolution slope between phase points to generate a dynamic time series graph.

[0021] Specifically, after acquiring the multidimensional raw signal, multidimensional data stream alignment processing is performed. Hardware timestamps are used to mark the three types of parameters to eliminate the time asynchrony deviation caused by communication link differences in the sensor feedback process. By executing time-domain translation and correction algorithms, asynchronous current, position and torque data are aligned under the same time reference, thereby forming a standardized raw observation sequence. The time-domain shift and correction algorithm uses a hardware timer to attach system timestamps to the collected drive current, position feedback and load torque data streams, and monitors the transmission time of each data stream from the sensor to the processor in real time; and calculates the time offset value of each parameter relative to the reference clock based on the preset communication link reference delay. Based on the calculated offset value, the lagging data sequence is indexed forward or backward in the buffer storage area to align the peak value of the drive current fluctuation, the response start point of the position feedback, and the sudden change point of the load torque at the same discrete time point. Linear interpolation is performed at the gap points after translation, and the transient values ​​missing during the translation process are filled in by calculating the slope between adjacent sampling points, thereby eliminating the microsecond-level asynchronous deviation caused by the different lengths of hardware links, and finally synthesizing the standardized original observation sequence. See Figure 2 Based on the fluctuation characteristics of the original observation sequence, the embedding dimension and delay step size of the delay coordinate embedding technique are adaptively determined, and the original observation sequence is reconstructed into a phase space trajectory describing the evolution of the motion state of the actuator through mapping. The aforementioned delayed coordinate embedding technique involves translating the original observation sequence multiple times at fixed time intervals (i.e., delay step size), and using these translated sequences as new coordinate dimensions to synthesize a series of state vectors in a multidimensional space. These vectors evolve continuously over time in the multidimensional space, enabling the construction of a phase space trajectory that is topologically equivalent to the original physical system. The fluctuation characteristics of the original observation sequence are analyzed to determine the key parameters for the delay coordinate embedding. For the delay step size, the average mutual information of each parameter in the original sequence is calculated, and the degree of information overlap of the sequence at different time offsets is observed. In specific implementation, the first local minimum point after the starting point decreases is automatically searched in the generated average mutual information curve. The time corresponding to the first local minimum point is determined as the adaptive delay step size. The embedding dimension is obtained by using global embedding gain. First, starting from a low-dimensional space, the system identifies pairs of points that are close to each other in the current dimension and observes the distance changes of these pairs after moving to a higher dimension. If the distance between two points increases significantly after the dimension increases, they are identified as false nearest neighbors caused by spatial collapse. By increasing the dimension one by one and continuously counting the proportion of false nearest neighbors, when the proportion tends to stabilize or decreases to below a preset threshold of 5%, it indicates that the dynamic trajectory of the actuator has been fully unfolded in the current space, and the corresponding value is locked as the minimum embedding dimension. Using the evolved phase points in the phase space trajectory as graph nodes, the topological connection weights are calculated based on the Euclidean distance and evolution slope between phase points to generate a dynamic time series graph. By utilizing hardware timestamps and linear interpolation algorithms, the asynchronous time deviation of multidimensional sensing data in the transmission link is effectively eliminated, ensuring a strong physical and logical correlation between current, position, and torque. Combined with adaptive phase space reconstruction technology, this scheme can automatically extract the optimal feature parameters based on the fluctuation characteristics of the operating data, transforming the one-dimensional observation sequence into a high-dimensional topological trajectory, thus breaking through the bottleneck of traditional monitoring methods that are difficult to capture complex nonlinear dynamic evolution.

[0022] The specific process of reconstructing the original observation sequence into a phase space trajectory describing the evolution of the actuator's motion state includes: calculating the average mutual information of each parameter in the original observation sequence, using the time corresponding to the first local minimum point as the adaptive delay step size; determining the minimum embedding dimension by calculating the global embedding gain of the observation sequence in different dimensions; performing coordinate mapping and time-domain translation on the aligned operating state parameters based on the adaptive delay step size and the minimum embedding dimension, and synthesizing a phase space vector describing the evolution of the actuator from the current steady state to the target opening in a multi-dimensional Euclidean space; and generating a phase space trajectory by tracing the geometric evolution path of the phase space vector within adjacent sampling periods.

[0023] The average mutual information of each parameter in the original observation sequence is calculated, and the time corresponding to the first local minimum point is used as the adaptive delay step size. Specifically, the degree of information overlap of the original observation sequence at different time intervals is calculated. By comparing the correlation between the sequence at the initial time and the subsequent moving time, the first local minimum point of the mutual information curve is found. The number of samples corresponding to this minimum point is set as the adaptive delay step size. By calculating the global embedding gain of the observation sequence in different dimensions, the minimum embedding dimension is determined by projecting the observation sequence into different dimensional spaces from low to high and observing whether the trajectory overlaps or intersects in the space. When increasing the dimension no longer significantly improves the degree of trajectory unfolding and the proportion of pseudo-nearest neighbor points in the space is reduced to below the preset 5%, the dimension at this time is determined to be the minimum embedding dimension that can completely cover the dynamic characteristics of the actuator. Based on the aforementioned adaptive delay step size and minimum embedding dimension, coordinate mapping and temporal translation are performed on the multidimensional running state parameters. Using the current observation value as a reference, historical sampling points are traced back according to the delay step size, and these data points with time spans are used as components of the coordinate axes for synthesis in multidimensional Euclidean space. In this way, the originally single time series is transformed into a vector in space, constructing a phase space vector. By continuously tracking the geometric evolution path of phase space vectors within adjacent sampling periods, a phase space trajectory is generated. In each control period, new sampling data generates new phase points. By connecting the evolution trajectories of these phase points in space, geometric features reflecting the operating trend of the equipment are extracted.

[0024] By introducing an adaptive delay step size and a minimum embedding dimension for phase space reconstruction, the scattered device parameters are successfully transformed into high-dimensional trajectories containing dynamic evolution laws. This not only effectively eliminates nonlinear interference between multi-source sensor data, but also solves the pain point of difficulty in quantifying the hysteresis effect of actuators from the underlying logic.

[0025] By calculating the coupling weights between nodes in the dynamic time-series graph, the mutual correlation matrix of intelligent devices is constructed in real time. By calculating the reciprocal of the Euclidean distance between any two nodes in the dynamic time series graph as the basic association value, the inverse mapping between spatial distance and association strength is realized. The cosine similarity algorithm is introduced to quantify the evolution slope between nodes. By comparing the tangent directions of two phase points on their respective evolution trajectories, a directional consistency coefficient between 0 and 1 is obtained. The basic association value is multiplied by the directional consistency coefficient to initially quantify the degree of geometric coupling between nodes. The position feedback response phase and drive current fluctuation phase of the node are extracted in real time, and the absolute value of the phase difference between the two is calculated. When the absolute value of the phase difference approaches 0, the node is assigned a synchronization weight coefficient close to 1 by using a negative exponential decay function. As the phase difference increases, the weight coefficient decays rapidly according to an exponential law. The dynamic coupling weight is obtained by multiplying the geometric coupling degree with the synchronization weight coefficient. The coupling weights are filled into a preset two-dimensional data structure to construct an interrelated matrix. Within each control sampling period, all active nodes in the graph are traversed, and the above quantization calculation process is repeated. Discrete mapping techniques are used to update the calculation results to the corresponding coordinate positions in the matrix. To ensure control stability, the matrix is ​​normalized so that the values ​​of all weights within the matrix are distributed within the standard range of 0 to 1.

[0026] The control feature vector is extracted from the dynamic time series spectrum; the control feature vector includes: a kinetic energy inertia feature term to compensate for response lag and a potential energy disturbance feature term to characterize the intensity of external load disturbance; The control feature vector is calculated by taking the evolution rate and phase space volume contraction rate of the current node relative to the historical trajectory nodes in the dynamic time-series graph, and mapping the evolution rate to a kinetic inertial feature term characterizing the mechanical motion inertia of the actuator; simultaneously extracting the gradient value of the node coupling strength characterizing the load torque change in the graph, and mapping the gradient value to a potential energy disturbance feature term characterizing the resistance of the external environment to the surface of the actuator; and normalizing and combining the kinetic inertial feature term and the potential energy disturbance feature term to construct a control feature vector that migrates in real time with the operating conditions of the actuator.

[0027] Specifically, the kinetic energy inertial characteristic term is constructed by calculating the evolution rate of the current node relative to historical trajectory nodes in the dynamic time series graph. In practice, within each control sampling period, the phase point coordinates corresponding to the current moment are extracted, and the historical phase points of the previous 1 to 5 sampling periods are traced back. The Euclidean displacement of the phase point in the multidimensional Euclidean space is calculated. By dividing this displacement by the corresponding sampling time interval, the instantaneous evolution rate of the actuator in the phase space is quantified. The volume shrinkage rate of phase space is calculated synchronously; at least three nearest neighbor nodes are selected around the current phase point to construct a local geometric cavity. The volume shrinkage rate of phase space is obtained by calculating the change ratio of the space volume occupied by the cavity in adjacent control cycles. This value is used to characterize the speed of energy dissipation of the actuator during the evolution process in real time. The evolution rate and the volume shrinkage rate are summed to complete the quantification of the kinetic energy inertia characteristic term. By extracting the gradient value of the node coupling strength representing the load torque change in the spectrum, a potential energy disturbance feature term is constructed; the weight column vector associated with the load torque node in the cross-correlation matrix is ​​extracted in real time, and the difference between the vectors at adjacent sampling times is calculated, i.e., the coupling strength gradient; the magnitude of the coupling strength gradient directly quantifies the instantaneous change intensity of the external load's interference force on the actuator, and the obtained gradient value is mapped to a potential energy disturbance feature term representing the external environment's effect on the actuator's surface resistance, thus transforming the dynamic disturbance that was originally difficult to measure directly into a quantifiable feature component; The extracted kinetic energy inertial feature term and potential energy disturbance feature term are normalized and combined to construct the control feature vector.

[0028] Specifically, the construction of the control feature vector is achieved through explicit mathematical logic: the kinetic energy inertia feature term first extracts the Euclidean displacement of the phase point in the reconstruction space relative to the previous sampling time, divides it by the sampling period to obtain the evolution rate, and combines it with the geometric cavity volume change ratio formed by the nearest neighbor points of the phase point (i.e., the phase space volume contraction rate), and uses a nonlinear model containing hyperbolic tangent function and exponential function to synthesize it, thereby transforming the geometric feature into an inertial compensation gain characterizing mechanical hysteresis; the potential energy disturbance feature term extracts the column vector of the torque node in the inter-correlation matrix, calculates its gradient magnitude in the time dimension, and combines it with the rated torque of the actuator for normalized coupling to quantify the instantaneous resistance fluctuation of the external load; The normalization process involves independently scaling two features linearly using the hardware limit values ​​automatically measured by the actuator during the initialization phase. The kinetic energy value of the actuator at its maximum design speed under no-load conditions is used as the upper limit benchmark for kinetic energy, and the maximum torque value generated by the actuator under stall current is used as the upper limit benchmark for potential energy. The original values ​​of the real-time extracted kinetic energy inertial features are divided by this upper limit benchmark for kinetic energy, and the original values ​​of the potential energy disturbance features are divided by this upper limit benchmark for potential energy, resulting in two dimensionless scalars between 0 and 1. If the real-time values ​​exceed the upper limit benchmarks, the calculation result is assigned a value of 1; otherwise, it is assigned a value of 0.

[0029] The components of the control feature vector are constructed based on the dynamic response ratio of the actuator; the ratio of the rate of change of position feedback to the rate of change of load torque within the current sampling period is calculated in real time, and this ratio is used as the allocation coefficient. The standardized kinetic energy scalar is multiplied by the allocation coefficient to obtain the first component of the two-dimensional vector; the standardized potential energy scalar is multiplied by 1 and the difference between it and the allocation coefficient to obtain the second component of the two-dimensional vector. In a two-dimensional space, a control feature vector describing the current operating condition is synthesized. The first dimension is set as the abscissa of the coordinate axis, and the second dimension is set as the ordinate, constructing a pointing vector from the origin to the coordinate point. The geometric length of this vector is calculated in real time, which is the arithmetic square root of the sum of the squares of the two components. If the geometric length exceeds 1, the vector direction remains unchanged, and the vector magnitude is forcibly corrected to 1 by dividing both components by this geometric length, ensuring that the control feature vector evolves within the unit circle. Control feature vectors are generated by tracing the migration path of vector endpoints in two-dimensional space within consecutive sampling periods.

[0030] By extracting control feature vectors from dynamic time-series graphs, precise quantification of equipment inertia and external disturbances is achieved. The kinetic inertia feature term, constructed using evolution rate and volume shrinkage rate, effectively offsets physical response lag; while the potential energy disturbance feature term, based on coupling strength gradient quantization, enables instantaneous perception of load fluctuations. Through hardware limit normalization and dynamic allocation logic, the vector's weights are shifted in real time according to operating conditions and its evolution is strictly constrained within a unit circle, ensuring smooth switching and trajectory alignment of control commands under varying loads.

[0031] A predictive control model with an integrated disturbance observer is constructed. A decoupling matrix is ​​generated based on the inter-correlation matrix, and the feedforward compensation gain is calculated in combination with the control feature vector. Based on the decoupling matrix and the feedforward compensation gain, and combined with the target command curve, the target control increment used to offset the hysteresis effect in the current control cycle is predicted. The predictive control model of the integrated disturbance observer includes a matrix decoupling layer, a feature feedforward layer, a disturbance observation layer, and a gain prediction layer; The matrix decoupling layer performs orthogonalization processing on the interrelated matrices, extracts the mapping feature basis between the physical parameters of the actuator, and constructs the decoupling matrix. The feature feedforward layer performs parameterized transformation on the kinetic energy inertia feature term and the potential energy disturbance feature term in the control feature vector, and calculates the feedforward compensation gain in real time by combining the output vector of the decoupling matrix. The disturbance observation layer superimposes the target control increment of the previous control cycle with the position feedback in real time to calculate the theoretical expected position at the current moment, and compares the theoretical expected position with the measured position feedback signal transmitted back in real time at the current moment to generate a real-time disturbance quantity. The gain prediction layer retrieves the preset running trajectory as the target command curve, integrates the decoupling matrix, feedforward compensation gain and the real-time disturbance, and performs time-domain rolling calculation within the sliding prediction window to generate the target control increment within the current control cycle.

[0032] The orthogonalization process is based on the cross-correlation matrix of the current period. Using the Schmitt orthogonalization logic, the non-orthogonal vectors in the matrix are transformed into a set of mutually perpendicular unit eigenbases. The transpose matrix is ​​constructed based on the unit eigenbases as column vectors and used as the decoupling operator. By multiplying the measured current, position and torque vectors with the decoupling operator, the cross-correlation between the parameters is eliminated, and the decoupling matrix is ​​constructed. The Schmitt orthogonalization logic extracts a set of eigenvectors representing drive current, position feedback, and load torque from the inter-correlation matrix, rather than directly processing the correlation matrix. By selecting the position feedback vector as the main control reference axis, the orthogonal projection algorithm is used to remove the coupling components in the remaining vectors and perform normalization processing to construct mutually perpendicular unit eigenbases and generate a decoupling transformation matrix. This mathematically cuts off the cross-interference of current and load fluctuations on the position loop, enabling independent control of each physical parameter in orthogonal space. The feature feedforward layer takes the kinetic energy inertia feature term and the potential energy disturbance feature term in the control feature vector and performs parameterization transformation; the kinetic energy inertia feature term is mapped to the gain ratio coefficient used to offset motion lag, and the potential energy disturbance feature term is mapped to the compensation weight coefficient used to offset load resistance. These coefficients are weighted and superimposed with the independent components output by the decoupling matrix to synthesize the feedforward compensation gain in real time. The disturbance observation layer superimposes the target control increment output from the previous control cycle with the measured position feedback signal in the time domain to calculate the theoretical expected position that the actuator should reach at the current moment. It also calculates the difference between the theoretical expected position and the measured position feedback signal transmitted back from the sensor at the current moment in real time to generate a real-time disturbance quantity. The gain prediction layer retrieves the preset running trajectory as the target command curve, integrates the decoupling matrix, feedforward compensation gain and the real-time disturbance, and performs time-domain rolling calculation within the sliding prediction window to generate the target control increment within the current control cycle.

[0033] The preset operating trajectory is a sequence of expected values ​​of the intelligent device opening degree that is stored in the controller memory over time. It integrates the output vector of the decoupling matrix, the feedforward compensation gain, and the real-time disturbance generated by the disturbance observation layer. Using the current measured position as the calculation benchmark, the hysteresis effect is offset in advance by superimposing the feedforward compensation gain, and the disturbance is introduced in real time for dynamic error correction. The target control increment that can drive the actuator and make its motion trajectory consistent with the target command curve in the future window is calculated within the current control cycle.

[0034] By constructing a multi-level predictive control architecture with an integrated disturbance observer, the decoupling control capability under nonlinear conditions is enhanced. The decoupling matrix constructed by orthogonalization eliminates cross-interference between physical parameters such as current and torque from the mathematical level. With the parameterization transformation of the feature feedforward layer, active offsetting of mechanical lag and running resistance is achieved. Through the dynamic comparison of theoretical and measured positions by the disturbance observation layer, external random disturbances are captured and canceled in real time. Finally, under the time-domain rolling calculation of the gain prediction layer, it is ensured that the target control increment can accurately drive the actuator to strictly follow the command curve.

[0035] The target control increment is converted into an execution command to drive the actuator, and the measured position feedback signal of the actuator is synchronously transmitted back. The specific process of converting the target control increment into an execution command to drive the actuator includes superimposing the target control increment onto the measured position feedback signal at the current moment to synthesize a target position command; using drive mapping logic to convert the target position command into a pulse drive signal and outputting it to the actuator drive end; and extracting the drive current and load torque during the drive process in real time and transmitting them back to the node update sequence of the dynamic time series graph.

[0036] The target control increment is superimposed on the measured position feedback signal at the current moment to synthesize the target position command; the measured position feedback signal transmitted back by the position sensor at the current moment is extracted in real time, and the target control increment calculated by the gain prediction layer is numerically superimposed on it to generate the target position command; The target position command is converted into a pulse drive signal using drive mapping logic and output to the actuator drive end; The drive mapping logic calls a pulse conversion function that matches the hardware characteristics of the actuator, calculates the angle that the drive motor needs to rotate based on the difference between the target position command and the current position, quantizes the angle into a corresponding number of digital pulses according to the electronic gear ratio, determines the output frequency of the pulse based on the sampling time of the current control cycle, and sends the pulse drive signal to the drive end of the actuator through the high-speed output port. The drive current and load torque during the driving process are extracted in real time and transmitted back to the node update sequence of the dynamic timing graph. The drive current and load torque data during the driving process are monitored and extracted in real time. The extracted current values ​​(and torque values) are bound to the timestamp of the current moment to form a synchronized data frame. This real-time data is transmitted back and loaded into the node update sequence of the dynamic timing graph for phase space reconstruction in the next control cycle.

[0037] By using a real-time synthesis and feedback mechanism for target position commands, the precise execution of control outputs and the closed-loop evolution of data are ensured; incremental superposition using measured position feedback eliminates the cumulative positioning error at the physical level; and the prediction results are converted into high-precision pulse signals through drive mapping logic, realizing digital control of the displacement and speed of the actuator.

[0038] Establish a reward function based on deep reinforcement learning to optimize the parameters inside the predictive control model in real time.

[0039] The specific process of parameterization within the real-time optimization predictive control model is as follows: The predictive control model is used as a reinforcement learning agent, and the operating state parameters of the actuator and the control feature vector are used as environmental state inputs; a multi-objective weighted reward function is constructed, which includes a trajectory tracking reward, a dynamic response reward, and a control smoothing reward. The target control increment is used as the action output to obtain the calculation result of the multi-objective weighted reward function; a deep deterministic policy gradient algorithm is adopted to maximize the cumulative reward and calculate the gradient correction value for the decoupling matrix parameters and feedforward gain weights in the predictive control model, thereby optimizing the parameters inside the predictive control model in real time.

[0040] Specifically, the predictive control model is configured as an agent under a deep reinforcement learning architecture, and an interactive environment is established between it and the actuator of the intelligent device. The real-time collected operating state parameters such as drive current, position feedback, and load torque, together with the kinetic energy inertial feature and potential energy disturbance feature extracted from the dynamic time series spectrum, are encapsulated as the environmental state input of the agent. A multi-objective weighted reward function is constructed in the system, which consists of three core evaluation dimensions: Tracking bonus: Real-time comparison of the measured position feedback signal of the actuator with the preset target command curve, calculation of the instantaneous deviation between the two, and the negative correlation ratio of the deviation value as the score; Dynamic response reward: Based on the kinetic energy inertia characteristic term in the control characteristic vector, the response speed of the actuator during start-up, shutdown, and frequency conversion phases is evaluated. If the current control output reduces the phase difference between the measured response curve and the target curve, i.e., the hysteresis effect is effectively suppressed, a positive reward is given; Smoothing control bonus: Monitor the fluctuation range of the target control increment within two adjacent control cycles and set penalty weights for behaviors that cause drastic changes in output; Within each control cycle, the agent takes the target control increment calculated by the predictive control model as the action output based on the current state input; automatically obtains the calculation result of the multi-objective weighted reward function based on the action feedback; uses a deep deterministic policy gradient algorithm to store the current state, action and the obtained comprehensive reward value into the experience replay buffer; and calculates the contribution of the current action to improving the long-term cumulative reward by extracting batch samples from the buffer. The calculated gradient information is used to correct the parameters of the neural network within the predictive control model. The optimization process directly affects the decoupling matrix weight coefficients and feedforward gain calculation parameters in the model. Through continuous online iterative updates, the model can automatically adjust the decoupling logic based on the feedback from the reward function, enabling the feedforward compensation gain to more accurately offset the hysteresis caused by kinetic inertia and the deviation caused by potential energy disturbance. By continuously optimizing the network weights, the predictive control model achieves dynamic adaptation to the nonlinear operating characteristics of intelligent devices, completing closed-loop control optimization.

[0041] This invention achieves in-depth analysis of the complex nonlinear evolution of actuators through the deep fusion of delayed coordinate embedding technology and dynamic time-series mapping. Compared to traditional control methods, this invention extracts kinetic inertia and potential energy disturbance features to anticipate response hysteresis and load impacts in mechanical systems, and utilizes a feedforward compensation mechanism to mitigate hysteresis effects in advance. Simultaneously, the introduction of a deep reinforcement learning reward function for parameter self-tuning enables the control model to dynamically evolve with equipment wear and environmental changes, significantly enhancing the actuator's response sensitivity and operational stability, and effectively solving the technical challenge of mismatch between control commands and actual actions in intelligent devices under varying operating conditions.

[0042] Example 2

[0043] See Figure 3 In this embodiment, the intelligent equipment opening prediction control system based on time-series graph features is applied to the DN300 electric regulating valve in the urban gas high-pressure regulating station. The DN300 electric regulating valve is responsible for adjusting the opening in real time according to the downstream flow demand, and faces high pressure differential fluid scouring and frequent operating condition switching. The graph construction module synchronously acquires data at a frequency of 200Hz. At a certain sampling instant, the measured drive motor current is 6.5A, the current valve opening position feedback is 35.0%, and the load torque generated by the upstream gas pressure is 120N·m. The system detects that due to differences in communication protocols, there is a return delay of approximately 10ms in the position signal. The graph construction module aligns these three types of parameters on the same time axis using hardware timestamps. Calculations show that the first local minimum of the average mutual information occurs at 8 sampling points. Based on this, the system sets the adaptive delay step size to 40ms and the embedding dimension to 3D, thereby reconstructing the trajectory describing the valve core motion evolution in three-dimensional phase space and generating a dynamic time series graph. The control vector module performs in-depth analysis of the time series spectrum. The system monitors that the evolution rate of the valve core relative to the previous 5 sampling points in the current control cycle is 0.85 units / second. Combined with the measured phase space volume contraction rate of 0.12, the kinetic energy inertia characteristic term is synthesized to be 0.78 (normalized value). This value directly reflects the response hysteresis caused by the large mass and strong inertia of the large-diameter valve core. At the same time, the gradient value of the coupling strength of the load torque node is extracted to obtain the potential energy disturbance characteristic term of 0.45, which quantifies the instantaneous resistance intensity of the valve core caused by the sudden change in upstream pipeline pressure. The two features are combined to generate a two-dimensional pointing vector that migrates with the operating conditions. The decoupling incremental module addresses parameter coupling in gas-fired scenarios; the matrix decoupling layer utilizes Schmitt orthogonalization logic to eliminate the strong correlation between drive current and load torque, constructing a decoupling matrix; the feature feedforward layer calculates the compensation gain in real time to offset mechanical hysteresis based on a kinetic energy inertia term of 0.78. The disturbance observation layer compares the target opening (36%) of the previous cycle with the current measured feedback (35.2%), identifying a real-time disturbance of -0.8% caused by gas fluid impact; the gain prediction layer performs rolling calculations within a sliding window, ultimately predicting a target control increment of +2.5% required for the valve to quickly conform to the target curve in this cycle. The predictive control module converts the +2.5% increment into a 5kHz pulse drive signal and sends it to the actuator. A reward function based on deep reinforcement learning is established, dividing the pressure regulation process into agent action segments. If, after execution, the downstream pressure fluctuation range remains within ±1% of the target value and the valve action is smooth and oscillating, the reward function outputs a positive score (+10 points); if overshoot occurs, causing the pressure to exceed the limit, a penalty score (-5 points) is output. A deep deterministic policy gradient algorithm is used to fine-tune the decoupling matrix weight coefficients within the model in real time based on the score, optimizing the correction parameter for the corresponding torque disturbance from 1.2 to 1.18.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predictive control of the opening degree of intelligent devices based on time-series graph features, characterized in that, include: The system collects the operating status parameters of the actuators of intelligent devices in real time, reconstructs the phase space of the operating status parameters using delayed coordinate embedding technology, and maps them to generate a dynamic time series graph. By calculating the coupling weights between nodes in the dynamic time series graph, the system constructs the mutual correlation matrix of the intelligent devices in real time. The control feature vector is extracted from the dynamic time series spectrum; the control feature vector includes: a kinetic energy inertia feature term to compensate for response lag and a potential energy disturbance feature term to characterize the intensity of external load disturbance; A predictive control model with an integrated disturbance observer is constructed. A decoupling matrix is ​​generated based on the inter-correlation matrix, and the feedforward compensation gain is calculated in combination with the control feature vector. Based on the decoupling matrix and the feedforward compensation gain, and combined with the target command curve, the target control increment used to offset the hysteresis effect in the current control cycle is predicted. The target control increment is converted into an execution command to drive the actuator, and the measured position feedback signal of the actuator is synchronously transmitted back; a reward function based on deep reinforcement learning is established to optimize the parameters inside the predictive control model in real time.

2. The intelligent device opening degree prediction and control method based on time-series graph features according to claim 1, characterized in that, The dynamic time series map generation process includes: The operating status parameters include drive current, position feedback, and load torque; The operating status parameters are collected in real time and aligned with multidimensional data streams to eliminate time asynchrony deviations in the sensor feedback process, forming a standardized original observation sequence. Based on the fluctuation characteristics of the original observation sequence, the embedding dimension and delay step size of the delay coordinate embedding technique are adaptively determined. The original observation sequence is reconstructed into a phase space trajectory describing the evolution of the actuator's motion state through mapping. Using the evolution phase points in the phase space trajectory as graph nodes, the topological connection weights are calculated based on the Euclidean distance and evolution slope between graph nodes to generate a dynamic time series graph.

3. The intelligent device opening degree prediction and control method based on time-series graph features according to claim 2, characterized in that, The specific process of reconstructing the original observation sequence into a phase space trajectory describing the evolution of the actuator's motion state includes: calculating the average mutual information of each parameter in the original observation sequence, using the time corresponding to the first local minimum point as the adaptive delay step size; determining the minimum embedding dimension by calculating the global embedding gain of the original observation sequence in different dimensions; performing coordinate mapping and time-domain translation on the aligned operating state parameters based on the adaptive delay step size and the minimum embedding dimension, and synthesizing a phase space vector describing the evolution of the actuator from the current steady state to the target opening in a multi-dimensional Euclidean space; and generating a phase space trajectory by tracing the geometric evolution path of the phase space vector within adjacent sampling periods.

4. The intelligent device opening degree prediction and control method based on time-series graph features according to claim 1, characterized in that, The control feature vector is calculated by measuring the evolution rate of the current node relative to the historical trajectory nodes in the dynamic time series graph and the phase space volume contraction rate, and the evolution rate is mapped to a kinetic inertial feature term characterizing the mechanical motion inertia of the actuator. The gradient values ​​of node coupling strength characterizing load torque changes in the dynamic time series spectrum are extracted synchronously, and the gradient values ​​are mapped to potential energy disturbance feature terms characterizing the external environment's effect on the surface resistance of the actuator. The kinetic energy inertia characteristic term and the potential energy disturbance characteristic term are normalized and combined to construct a control characteristic vector that migrates in real time with the operating conditions of the actuator.

5. The intelligent device opening degree prediction and control method based on time-series graph features according to claim 2, characterized in that, The predictive control model of the integrated disturbance observer includes a matrix decoupling layer, a feature feedforward layer, a disturbance observation layer, and a gain prediction layer; The matrix decoupling layer performs orthogonalization processing on the interrelated matrices, extracts the mapping feature basis between the physical parameters of the actuator, and constructs the decoupling matrix. The feature feedforward layer performs parameterized transformation on the kinetic energy inertia feature term and the potential energy disturbance feature term in the control feature vector, and calculates the feedforward compensation gain in real time by combining the output vector of the decoupling matrix. The disturbance observation layer superimposes the target control increment of the previous control cycle with the position feedback in real time to calculate the theoretical expected position at the current moment, and compares the theoretical expected position with the measured position feedback signal transmitted back in real time at the current moment to generate a real-time disturbance quantity. The gain prediction layer retrieves the preset running trajectory as the target command curve, integrates the decoupling matrix, feedforward compensation gain and the real-time disturbance, and performs time-domain rolling calculation within the sliding prediction window to generate the target control increment within the current control cycle.

6. The intelligent device opening degree prediction and control method based on time-series graph features according to claim 1, characterized in that, The specific process of converting the target control increment into an execution command to drive the actuator includes superimposing the target control increment onto the measured position feedback signal at the current moment to synthesize a target position command; and using drive mapping logic to convert the target position command into a pulse drive signal, which is then output to the actuator drive end. The drive current and load torque during the drive process are extracted in real time and transmitted back to the node update sequence of the dynamic time series graph.

7. The intelligent device opening degree prediction and control method based on time-series graph features according to claim 1, characterized in that, The specific process of parameterization within the real-time optimization predictive control model is as follows: The predictive control model is used as a reinforcement learning agent, and the operating state parameters of the actuator and the control feature vector are used as environmental state inputs; a multi-objective weighted reward function is constructed, which includes a trajectory tracking reward, a dynamic response reward, and a control smoothing reward. The target control increment is used as the action output to obtain the calculation result of the multi-objective weighted reward function; a deep deterministic policy gradient algorithm is adopted to maximize the cumulative reward and calculate the gradient correction value for the decoupling matrix parameters and feedforward gain weights in the predictive control model, thereby optimizing the parameters inside the predictive control model in real time.

8. A predictive control system for the opening degree of intelligent devices based on time-series graph features, characterized in that, include: The graph construction module collects the operating status parameters of the actuators of intelligent devices in real time, reconstructs the phase space of the operating status parameters using delayed coordinate embedding technology, and maps them to generate a dynamic time series graph; by calculating the coupling weights between each node in the dynamic time series graph, the mutual correlation matrix of the intelligent devices is constructed in real time. Control vector module: extracts control feature vectors from the dynamic time series graph; the control feature vectors include: kinetic energy inertia feature term to compensate for response lag and potential energy disturbance feature term to characterize the intensity of external load disturbance; Decoupling Increment Module: Constructs a predictive control model with an integrated disturbance observer, generates a decoupling matrix based on the cross-correlation matrix, and calculates the feedforward compensation gain in combination with the control feature vector; based on the decoupling matrix and the feedforward compensation gain, and in combination with the target command curve, predicts the target control increment used to offset the hysteresis effect in the current control cycle. Predictive control module: converts the target control increment into execution instructions to drive the actuator, and synchronously transmits back the measured position feedback signal of the actuator; establishes a reward function based on deep reinforcement learning, and optimizes the parameters inside the predictive control model in real time.

Images