Network adaptive pulse synchronization control method and system based on fusion event triggering

Abstract

The invention relates to the technical field of industrial adaptive control, in particular to a network adaptive pulse synchronization control method and system fused with event triggering, and the method comprises the steps: controlling a plurality of controlled units to execute a carrying task of a target load; identifying the controlled unit to obtain state feedback data; detecting the state of the target load in real time to obtain disturbance feedback data; performing performance index analysis on the disturbance feedback data, and judging whether a preset event triggering condition is met or not; when the event triggering condition is met, self-adaptive pulse control is started; the self-adaptive pulse control comprises the following steps: comparing and analyzing the state feedback data and a preset benchmark reference model, and determining control deviation data; based on the control deviation data, the disturbance feedback data and the state feedback data, self-adaptive parameter optimization is carried out, and control pulse parameters including pulse amplitude, pulse width, pulse shape and cooperative application time are generated; according to the invention, based on the control pulse parameters, synchronous control of a plurality of controlled units is realized.

Description

Technical Field

[0001] This invention relates to the field of industrial adaptive control technology, specifically to a network adaptive pulse synchronization control method and system that integrates event triggering. Background Technology

[0002] Control systems composed of multiple mobile robots or actuators collaboratively handle large, heavy, or fragile loads, which has become a key technology for improving efficiency in industrial production. Achieving precise synchronous adjustment of multiple controlled objects is crucial for ensuring a smooth and safe handling process throughout the system. However, traditional synchronous control algorithms face significant challenges when handling special loads, particularly liquid loads.

[0003] During transportation, the starting, stopping, acceleration, deceleration, and steering of the controlled unit inevitably induce dynamic disturbances in the liquid, generating severe nonlinear disturbances that react on the control actuators. This not only changes the dynamic model parameters of the system in real time, such as load distribution and center of mass, but may also lead to liquid overflow, or even cause the entire closed-loop system to become unstable or overturn due to a sharp decline in performance.

[0004] To cope with such disturbances, existing adaptive control systems mostly employ continuous adjustment strategies based on fixed-time sampling, meaning the controller is constantly running and adjusting regardless of whether the system state deviates. In networked control systems, this generates significant communication overhead and computational load. Moreover, for highly nonlinear, sporadic disturbances such as liquid sloshing, this time-based control response may not be timely enough, or it may cause unnecessary system oscillations due to over-adjustment when the disturbance is small.

[0005] Therefore, the field of control systems needs a new control architecture that can leverage the low resource consumption and high efficiency of event-triggered mechanisms to activate high-performance adaptive control algorithms when disturbances occur, optimize in real time and generate precise coordinated actuation commands to proactively respond to liquid sloshing disturbances and ensure the synchronous stability of multiple controlled units under strong disturbances.

[0006] To address this, a network adaptive pulse synchronization control method and system based on fusion event triggering is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a network adaptive pulse synchronization control method and system that integrates event triggering. When the event triggering condition is met, adaptive pulse control is initiated. The adaptive pulse control includes: comparing and analyzing state feedback data with a preset benchmark reference model to determine control deviation data; optimizing adaptive parameters based on control deviation data, disturbance feedback data, and state feedback data to generate control pulse parameters, including pulse amplitude, pulse width, pulse shape, and coordinated application time; and realizing synchronous control of multiple controlled units based on the control pulse parameters.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A network adaptive pulse synchronization control method based on fusion event triggering includes: Control multiple controlled units to perform the target load transportation task; identify the status of the controlled units and obtain status feedback data of the multiple controlled units; detect the status of the target load in real time and obtain disturbance feedback data; A disturbance identification model is constructed to identify the disturbance feedback data. Through performance index analysis, it is determined whether the preset event triggering conditions are met. When the disturbance feedback data meets the event triggering conditions, adaptive pulse control is initiated. The adaptive pulse control includes: acquiring the state feedback data of the multiple controlled units, and performing comparative analysis based on a preset benchmark reference model of the multiple controlled units to determine the control deviation data; An adaptive control model is constructed based on the control deviation data, disturbance feedback data, and state feedback data to perform adaptive parameter optimization and generate control pulse parameters. The control pulse parameters include pulse amplitude, pulse width, pulse shape, and coordinated application time. Multiple controlled units are synchronously controlled based on the control pulse parameters.

[0009] The target load is a liquid load, and the disturbance feedback data is used to characterize the hydrodynamic parameters of the liquid load sloshing state; the disturbance feedback data specifically includes: acquiring and identifying the sloshing amplitude, sloshing frequency, and sloshing energy of the liquid in real time through an Internet of Things sensor installed inside the target load container.

[0010] The disturbance identification model is built on a recurrent neural network to identify the dynamic characteristics of the disturbance feedback data and perform performance index analysis; it includes a runaway risk quantification layer, a runaway magnitude quantification layer, and a performance analysis triggering layer. The runaway risk quantification layer identifies and classifies the sway frequency of the target load to determine the type of runaway disturbance; based on the type of runaway disturbance, it determines the runaway risk factor. The runaway level quantization layer normalizes the sway amplitude and sway energy to obtain normalized sway amplitude data and normalized sway energy data, and then performs identification and judgment by a recurrent neural network to obtain the runaway level factor. The performance analysis triggering layer performs weighted fusion of runaway risk factors and runaway magnitude factors to obtain a comprehensive runaway index; and judges the event triggering conditions based on the comprehensive runaway index according to the preset runaway threshold.

[0011] The status feedback data includes: heterogeneous load data, power data, stability data, and position data of the controlled unit; The heterogeneous load data refers to the load change data of each controlled unit during the transportation process; The power data includes the actual acceleration and actual speed of the controlled unit, as well as the actual output torque and drive current of the drive system; The stable data includes the device tilt angle, device pitch angle, and device yaw rate of the controlled unit; The location data includes high-precision position coordinate data of the controlled unit.

[0012] The benchmark reference model is constructed based on historical transportation data and includes: The baseline collaborative state of multiple controlled units under a collaborative transport target load task is obtained; the baseline collaborative state includes baseline load data, baseline power data, baseline stability data, and baseline position data. The control deviation data is obtained by identifying the data differences between the state feedback data of multiple controlled units and the reference coordinated state; the deviation data includes load deviation data, power deviation data, stability deviation data and position deviation data.

[0013] The process of adaptive parameter optimization and generating control pulse parameters includes: The adaptive control model standardizes and identifies heterogeneous load data in the control deviation data and controlled unit state feedback data to obtain the stable influence weight of each controlled unit. For each controlled unit, the stability influence weight and state feedback data are acquired as controlled unit parameters; all controlled unit parameters and corresponding disturbance feedback data are identified, and the desired state parameters of the controlled unit are output; the control pulse parameters are determined based on the desired parameters of the controlled unit.

[0014] The process of obtaining the stability influence weights for each controlled unit includes: Obtain heterogeneous load data from control deviation data and state feedback data; identify load contribution parameters based on the load deviation data and heterogeneous load data of the controlled unit; identify stability contribution parameters based on the stability deviation data of the controlled unit; identify power contribution parameters based on the power deviation data of the controlled unit; identify position contribution parameters based on the position deviation data of the controlled unit; and obtain stability influence weights by weighting the load contribution parameters, stability contribution parameters, power contribution parameters, and position contribution parameters.

[0015] A network-adaptive pulse synchronization control system based on fusion event triggering, comprising: The data acquisition module controls multiple controlled units to perform target load transport tasks; identifies the status of the controlled units and obtains status feedback data of the multiple controlled units; and detects the status of the target load in real time to obtain disturbance feedback data. The disturbance judgment module constructs a disturbance identification model to identify disturbance feedback data and analyzes performance indicators to determine whether preset event triggering conditions are met. When the disturbance feedback data meets the event triggering conditions, adaptive pulse control is initiated. The pulse control module, wherein the adaptive pulse control includes: acquiring the state feedback data of the plurality of controlled units, and performing comparative analysis based on a preset benchmark reference model of the plurality of controlled units to determine the control deviation data; An adaptive control model is constructed based on the control deviation data, disturbance feedback data, and state feedback data to perform adaptive parameter optimization and generate control pulse parameters. The control pulse parameters include pulse amplitude, pulse width, pulse shape, and coordinated application time. Multiple controlled units are synchronously controlled based on the control pulse parameters.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a disturbance identification model based on a recurrent neural network, which can understand the dynamic characteristics of swaying based on time series analysis, realizing a leap from state monitoring to trend prediction; secondly, its unique three-layer architecture decouples the complex disturbance judgment problem, assesses the probability through the runaway risk quantification layer, assesses the severity through the runaway magnitude quantification layer, and finally makes a comprehensive decision by the performance analysis triggering layer; this makes the event triggering mechanism efficient and intelligent.

[0017] 2. This invention constructs a benchmark reference model with high adaptability and accuracy by combining historical transport data and spatial models. It not only includes ideal kinematic solutions but also incorporates dynamic corrections learned from experience. Furthermore, the benchmark provided is not for a single unit but for the entire formation, ensuring that the subsequently calculated control deviation data truly reflects the coordination misalignment and synchronization errors between units. This deviation calculation based on a cooperative benchmark enables the control system to distinguish between planned maneuvers and unplanned disturbances, thereby avoiding misjudgments of normal maneuvering behavior.

[0018] 3. This invention, by introducing a stability influence weight, can dynamically assess the threat level of each controlled unit to global stability. This weight integrates the unit's deviation and load, enabling the control system to accurately identify the weakest link unit, thereby achieving differentiated and optimal resource allocation. Secondly, by comprehensively considering the weighted parameters of all units and external disturbance feedback data, it ensures that the generated control pulse is not an isolated, locally optimal solution, but rather a system-level, collaborative, globally optimal solution. It can proactively and asymmetrically control different units to produce a synergistic suppression effect, realizing a shift from passive response to proactive optimization. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a network adaptive pulse synchronization control method based on fusion event triggering according to the present invention. Figure 2 This is a schematic diagram of the logic for obtaining the control pulse parameters of the present invention; Figure 3 This is a schematic diagram of the structure of a network adaptive pulse synchronization control system based on fusion event triggering according to the present invention. Detailed Implementation

[0020] 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.

[0021] Example 1: This invention proposes a network adaptive pulse synchronization control method that integrates event triggering, the flow of which is as follows: Figure 1 As shown, it includes: Control multiple controlled units to perform the target load transportation task; identify the status of the controlled units and obtain status feedback data of the multiple controlled units; detect the status of the target load in real time and obtain disturbance feedback data; A disturbance identification model is constructed to identify the disturbance feedback data. Through performance index analysis, it is determined whether the preset event triggering conditions are met. When the disturbance feedback data meets the event triggering conditions, adaptive pulse control is initiated. The adaptive pulse control includes: acquiring the state feedback data of the multiple controlled units, and performing comparative analysis based on a preset benchmark reference model of the multiple controlled units to determine the control deviation data; An adaptive control model is constructed based on the control deviation data, disturbance feedback data, and state feedback data to perform adaptive parameter optimization and generate control pulse parameters. The control pulse parameters include pulse amplitude, pulse width, pulse shape, and coordinated application time. Multiple controlled units are synchronously controlled based on the control pulse parameters.

[0022] The logic for obtaining the control pulse parameters is as follows: Figure 2 As shown.

[0023] The target load is a liquid load, and the disturbance feedback data is used to characterize the hydrodynamic parameters of the liquid load sloshing state; the disturbance feedback data specifically includes: acquiring and identifying the sloshing amplitude, sloshing frequency, and sloshing energy of the liquid in real time through an Internet of Things sensor installed inside the target load container.

[0024] Preferably, a pressure sensor is installed inside the target load container to obtain pressure time-series data. During liquid sloshing, the pressure at different locations changes periodically. By analyzing the pressure time-series data, the sloshing amplitude and frequency of the liquid surface can be calculated. Furthermore, by performing a Fast Fourier Transform on the pressure signal to identify the main frequency components, and combining this with a pressure amplitude relationship model pre-calibrated through computational fluid dynamics simulation, the sloshing amplitude and frequency of the liquid surface can be estimated. After obtaining the sloshing amplitude and frequency, the sloshing energy is calculated based on a preset fluid dynamics model. For example, the sloshing can be simplified to a first-order linear wave, and the sloshing energy can be estimated as being proportional to the square of the sloshing amplitude and the square of the sloshing frequency.

[0025] This invention precisely quantifies disturbances into three key hydrodynamic parameters: sloshing amplitude, sloshing frequency, and sloshing energy. By distinguishing between frequency and energy, it can identify resonant frequencies with high instability risk or large impacts with overturning risk. This provides an extremely reliable basis for event triggering, avoids false control triggering caused by instantaneous amplitude fluctuations, and ensures timely intervention before energy accumulates to a dangerous threshold.

[0026] The disturbance identification model is built on a recurrent neural network to identify the dynamic characteristics of the disturbance feedback data and perform performance index analysis; it includes a runaway risk quantification layer, a runaway magnitude quantification layer, and a performance analysis triggering layer. The runaway risk quantification layer identifies and classifies the sway frequency of the target load to determine the type of runaway disturbance; based on the type of runaway disturbance, it determines the runaway risk factor. The runaway level quantization layer normalizes the sway amplitude and sway energy to obtain normalized sway amplitude data and normalized sway energy data, and then identifies and judges them to obtain the runaway level factor. The performance analysis triggering layer performs weighted fusion of runaway risk factors and runaway magnitude factors to obtain a comprehensive runaway index; and judges the event triggering conditions based on the comprehensive runaway index according to the preset runaway threshold.

[0027] Preferably, liquid sloshing is a typical time-series problem, where the current sloshing state is closely related to its state in previous time periods. Recurrent neural networks have memory capabilities, enabling them to capture the dependencies and dynamic characteristics of data over time; this allows the model not only to see the current sloshing but also to understand whether the sloshing is increasing, decreasing, or about to reach its peak, thus making more accurate judgments.

[0028] Furthermore, the types of runaway disturbances include: high-risk associated disturbances, potential associated disturbances, and safety disturbances; High-risk associated disturbances: refer to the frequencies of swaying or their temporal variation patterns that are currently measured and are highly correlated with frequencies marked as severely out of control in the training data, such as overturning or massive liquid spillage.

[0029] Potentially correlated disturbances: refer to the current swaying frequency being correlated with operating conditions marked as slightly unstable or having a tendency to run out of control in the training data.

[0030] Safety disturbance: refers to the current swaying frequency that appears frequently in the training data, but is almost never associated with any out-of-control or unstable conditions, and is considered a safety signal.

[0031] The process of pattern matching and classification of input sway frequency data by a recurrent neural network model includes: The disturbance identification model acquires real-time time-series data of the swaying frequency. This time-series data is received by a runaway risk quantification layer, which then captures the dynamic characteristics of the swaying frequency based on this data; for example, whether the frequency is rapidly increasing, remaining stable, or oscillating. The output of this layer is a classification result, i.e., the runaway disturbance type. Based on the runaway disturbance type, a runaway risk factor is determined. The runaway risk factor is a normalized numerical value representing the probability of risk posed by the frequency. The runaway risk factor is determined based on the experience of an expert group and obtained through supervised training of the model.

[0032] After obtaining the risk factor and magnitude factor of runaway, they are weighted and fused to obtain a comprehensive runaway index; the weighting and fusion process is carried out in a recurrent neural network (RNN), and the weights are obtained based on supervised training.

[0033] Through the refined risk classification described above, the entire event triggering mechanism is no longer a switch based on a single threshold, but becomes efficient and intelligent; it can avoid control mis-triggering caused by momentary, harmless frequency jitter; at the same time, it maintains extremely high control sensitivity to truly dangerous frequency patterns; ultimately, it achieves the activation of adaptive pulse control only when truly needed, ensuring safety while minimizing unnecessary consumption of computing and communication resources in networked control.

[0034] The training process of the disturbance recognition model is as follows: Through computational fluid dynamics simulation and extensive experiments on a test bench, liquid sloshing under various working conditions is simulated, including different speeds, directions, accelerations and decelerations. At the same time, sensor data is recorded to obtain the corresponding sloshing amplitude, sloshing frequency, sloshing energy, and stability data of the target load, such as whether it is out of control or whether it overflows.

[0035] The collected time-series data is labeled according to the stability data of the target load to obtain data labels.

[0036] The data labels are marked according to the following specific physical quantification standards: High-risk associated perturbation label: Liquid spillage observed.

[0037] Potential associated disturbance label: No high-risk conditions were triggered, but the observed sway frequency entered the range of the inherent sway frequency and lasted for more than 1 second.

[0038] Safety disturbance label 0: Operating conditions that do not meet any of the above conditions.

[0039] The input consists of time-series data of swaying parameters, and the output includes the runaway risk factor, the runaway magnitude factor, and the final comprehensive runaway index. The RNN model is trained using labeled data under supervised learning until it can accurately predict runaway situations under different swaying modes.

[0040] The preset runaway threshold is obtained through statistical analysis of the training dataset of the disturbance recognition model, including: Collect at least 500 sets of experimental or simulation data samples, covering the following typical operating conditions: normal handling without abnormal disturbance; slight disturbance, liquid sloshing amplitude ≤ 2cm; moderate disturbance, liquid sloshing amplitude 2-5cm; severe disturbance, liquid sloshing amplitude approaching the overflow boundary; analyze the runaway threshold based on the collected data. For example, among all samples labeled as runaway, it was found that 95% of the comprehensive runaway index was higher than 0.75, therefore the runaway threshold was set to 0.75. Furthermore, to avoid false triggering caused by transient disturbances, the event triggering condition can be defined as: the comprehensive runaway index continuously exceeding 0.75 for 500 milliseconds.

[0041] This invention constructs a disturbance identification model based on a recurrent neural network, which can understand the dynamic characteristics of swaying based on time series analysis, realizing a leap from state monitoring to trend prediction. Secondly, its unique three-layer architecture decouples the complex disturbance judgment problem. The probability is assessed through the runaway risk quantification layer, the severity is assessed through the runaway magnitude quantification layer, and finally the comprehensive decision is made by the performance analysis triggering layer. This makes the event triggering mechanism efficient and intelligent.

[0042] Compared to traditional single threshold triggering, this invention can accurately identify potential resonance threats of "high risk, low magnitude" and can also reasonably deal with low risk, high magnitude instantaneous impacts, while filtering out low risk, low magnitude harmless fluctuations. This ensures extremely high control sensitivity while minimizing unnecessary consumption of computing and communication resources.

[0043] Furthermore, the acquisition of runaway perturbation types and runaway risk factors not only utilizes the temporal learning capabilities of recurrent neural networks but also incorporates physical models of liquid sloshing, ensuring the classification results have physical interpretability, including: Frequency analysis under physical constraints: The system estimates the physical parameters of the liquid in real time based on the known geometric dimensions of the container, such as width, length, and depth; it estimates the current liquid depth based on the liquid surface height measured in real time by a pressure sensor; and it obtains the dynamic viscosity based on the liquid type. The theoretical natural sloshing frequency is calculated based on the physical parameters of the liquid and the geometry of the container. This includes distinguishing the container type of the target load, such as rectangular and circular. The theoretical natural frequency is calculated in real time when the system starts up or when the liquid depth changes significantly, such as by more than 5 cm. The calculation results are compared with other known operating conditions in the physical parameter library to improve the robustness of the calculation.

[0044] Define the resonant danger range: Based on the theoretical frequency and the known fluid damping ratio, define the resonant danger frequency range. This range indicates that excitation within this frequency range will cause the fluid sloshing to be amplified in a resonant manner.

[0045] Accurate estimation of real-time swaying frequency is achieved through an RNN network model; for example, the input is the raw pressure sensor time series data of the past 3 seconds; the output is the current swaying main frequency and the time derivative of the frequency; wherein, a frequency analysis layer is used, such as through spectral normalization, to force the RNN to learn frequency features. By identifying the similarity between the main frequency of the shaking and the range of the resonant danger frequency, different types of runaway disturbances are distinguished and assigned runaway risk factors.

[0046] The status feedback data includes: heterogeneous load data, power data, stability data, and position data of the controlled unit; The heterogeneous load data refers to the load change data of each controlled unit during the handling process. The power data includes the actual acceleration and actual speed of the controlled unit obtained through accelerometers and wheel speed encoders; and the actual output torque and drive current of the drive system. The stable data includes the tilt angle, pitch angle, and yaw rate of the controlled unit equipment obtained through the inertial measurement unit. The location data includes high-precision position coordinate data of the controlled unit acquired through lidar and differential scanning.

[0047] Preferably, heterogeneous load data: In collaborative handling, due to shaking or uneven road surfaces, the load borne by each unit is dynamically changing and uneven. This data is the real-time change data of the load of each unit.

[0048] Stable data and heterogeneous load data are used to achieve stability; when the liquid sloshes, the impact force will first be reflected in the heterogeneous load data and stable data, such as a sudden increase in the load of the controlled unit, or the controlled unit tilting.

[0049] Position and dynamic data: used to achieve synchronization; the control system must know the speed and position of each unit in order to keep them in formation.

[0050] This invention constructs a multi-dimensional state perception system, comprehensively covering the holographic information required by the control system through heterogeneous load, dynamics, stability, and position, ensuring complete observability of the state. The introduction of heterogeneous load data innovatively quantifies the uneven distribution of disturbances across multiple controlled units, enabling the system to identify which unit experiences the greatest force, which is key to achieving refined, asymmetric control. Secondly, by integrating multiple source sensors such as accelerometers, encoders, lidar, and differential positioning, high accuracy and redundancy of various data are ensured. This provides a solid and reliable data foundation for subsequent benchmark comparison and adaptive control, enabling the system to detect minute state deviations caused by disturbances.

[0051] The benchmark reference model is constructed based on historical transportation data and includes: The reference coordination state of multiple controlled units is obtained under the condition of coordinating the transport of the target load; the reference coordination state includes reference load data, reference power data, reference stability data and reference position data; The control deviation data is obtained by identifying the data differences between the state feedback data of multiple controlled units and the reference coordinated state; the deviation data includes load deviation data, power deviation data, stability deviation data and position deviation data.

[0052] Preferably, the transport path of the controlled unit transporting the target load is obtained, and the spatial model of the transport path is identified; the historical transport data of the controlled unit is projected into the spatial model for identification; the reference model is the optimal cooperative trajectory under a given working condition (such as maximum speed, minimum sway) pre-calculated by trajectory optimization algorithms, such as model predictive control (MPC) or A* algorithm.

[0053] Preferably, in this embodiment, the process of obtaining the benchmark reference model includes: Step 1: Standardize historical data; Collect at least N sets of complete data records for historical handling tasks. Each set of data includes: the spatial coordinate sequence of the handling path, the handling duration, and the sampling frequency; wherein, the spatial coordinate sequence of the handling path is represented as... , Represents key spatial trajectory points, including their three-dimensional spatial position and orientation angle; Obtain a complete set of status feedback data for multiple AGVs at each time t, including four types of data: load, power, stability, and position of the corresponding AGV; perform time-based alignment on all historical data: using shape matching algorithms such as DTW distance, group all handling tasks according to the similarity of the handling path, and align the data within the same path group according to the start time of handling.

[0054] Step 2: Extraction of Cooperative State Baseline; For multiple historical trajectories within the same path group after alignment, the formation baseline coordination state is calculated at each alignment time using the average data, to obtain baseline load data, baseline dynamic data, baseline stability data, and baseline position data.

[0055] Step 3: Spatial model parameterization; The spatial model of the transport path is represented by a parametric B-spline curve: the number of control vertices of the B-spline is determined based on the start and end points of the historical paths and the key turning points in the middle of the paths. For each historical trajectory at time t, its parameter position on the spatial model is calculated. When the state data of multiple historical trajectories at the same parameter position u has been collected, the baseline cooperative state corresponding to that parameter value is extracted at that position, thus obtaining the baseline reference model.

[0056] The baseline reference model of this invention is a parameterized, condition-dependent reference frame. It is a reference trajectory library established based on the family of motion trajectories under undisturbed or low-disturbance conditions in historical handling tasks through parameterization and smoothing.

[0057] The baseline coordinated state is obtained in the time dimension of the controlled unit handling process, that is, the start of handling is the time start point and the end of handling is the time end point. During the time start point and time end point, the baseline load data, baseline power data, baseline stability data and baseline position data of multiple controlled units change with time.

[0058] This invention constructs a highly adaptable and accurate benchmark reference model by combining historical transport data and spatial models. It not only includes ideal kinematic solutions but also incorporates dynamic corrections learned from experience. Furthermore, the benchmark provided is not for a single unit but for the entire formation, ensuring that subsequently calculated control deviation data truly reflects coordination misalignment and synchronization errors between units. This deviation calculation based on a cooperative benchmark enables the control system to distinguish between planned maneuvers and unplanned disturbances, thereby avoiding misjudgments of normal maneuvering behavior.

[0059] The process of adaptive parameter optimization and generating control pulse parameters includes: The adaptive control model standardizes and identifies the heterogeneous load data in the control deviation data and the state feedback data of the controlled unit to obtain the stability influence weight of each controlled unit. For each controlled unit, the stability influence weight and state feedback data are acquired as controlled unit parameters; all controlled unit parameters and corresponding disturbance feedback data are identified, and the desired state parameters of the controlled unit are output; the control pulse parameters are determined based on the desired state parameters of the controlled unit.

[0060] Preferably, in this embodiment, the adaptive control model is constructed through a deep learning-based multi-agent cooperative control network; the network model is pre-trained in a simulation environment through reinforcement learning to achieve adaptive parameter optimization.

[0061] Model construction and input: The adaptive control model is constructed as a Transformer-based encoder-decoder architecture. This architecture is particularly suitable for handling the cooperative relationships between multiple controlled units (agents) and responses to global disturbances.

[0062] The model's input consists of three parts: Global disturbance data: namely, disturbance feedback data, such as liquid sloshing amplitude, frequency, and energy.

[0063] Unit-specific data: For each controlled unit, its state feedback data, control deviation data, and calculated stability influence weights are combined into a feature vector.

[0064] Collaborative identification and state prediction enable the identification of all controlled unit parameters and corresponding disturbance feedback data. Cooperative coding: The feature vectors of all controlled units are input into the Transformer encoder as a set of tokens.

[0065] Attention mechanism: The network's self-attention mechanism or message passing mechanism calculates the complex dependencies between all controlled units, as well as between each unit and the global perturbation feedback data. The stability influence weights adjust the attention allocation in this process, so that the network focuses on the units that pose the greatest threat to system stability.

[0066] State Output: The encoder's output passes through a decoder, such as a multilayer perceptron (MLP), to output the desired state parameters for each controlled unit. These desired state parameters represent the ideal motion state that the unit should achieve under the current disturbance in order to restore system synchronization and stability; for example, the desired instantaneous acceleration, angular velocity, or position correction.

[0067] The output layer of the network model, or a subsequent mapping module, calculates and generates the final control pulse parameters based on the desired state parameters of the controlled unit and the current state feedback data.

[0068] Specifically, a proportional-integral-derivative controller is used to convert the desired state parameters output by the adaptive control model into specific instructions that the driver can execute, i.e., control pulse parameters. Based on the deviation between the desired state parameters of the controlled unit and the current state feedback data, an algorithm that comprehensively considers the current, past, and future deviations is used to calculate a control quantity, which is then used as the output to accurately drive the controlled unit to approach the desired state.

[0069] Preferably, the process of determining the control pulse parameters based on the desired parameters of the controlled unit further includes, when determining the cooperative application time: State delay identification: The input to the adaptive control model includes not only the state feedback data itself, but also the timestamp of the data or network delay information.

[0070] Quantification of terrain disturbance: Real-time analysis of stable data, such as sudden changes in the angular velocity of the IMU, and dynamic data such as instantaneous spikes in the drive current, to identify whether a specific unit is suffering from terrain disturbance.

[0071] Solution of asynchronous timing: When a significant terrain disturbance or data delay is detected in a certain unit, the adaptive control model will simultaneously calculate the optimal pulse application timing when outputting the desired state parameters of the controlled unit, thereby obtaining the cooperative application time.

[0072] This asynchronous application is to wait for the sensor readings of the controlled unit to stabilize or to overcome obstacles, thereby avoiding the application of control pulses in erroneous states and ensuring optimal final cooperative effect. Each controlled unit executes the pulse at its individually specified application time according to the adaptive control model, achieving advanced cooperation with consistent objectives and asynchronous timing.

[0073] The in-depth explanation of the timing of the cooperative application elevates this invention from rigid synchronous control to flexible asynchronous cooperation; it can maintain extremely high control robustness and synchronization accuracy even under real-world conditions of networked control and non-ideal terrain.

[0074] This invention introduces a stability impact weight, enabling dynamic assessment of the threat posed by each controlled unit to global stability. This weight integrates unit deviation and load, allowing the control system to accurately identify bottleneck units, thereby achieving differentiated and optimal resource allocation. Secondly, by comprehensively considering the weighted parameters of all units and external disturbance feedback data, it ensures that the generated control pulses are not isolated, locally optimal solutions, but rather system-level, collaborative, globally optimal solutions. This allows for proactive and asymmetric control of different units to produce a synergistic suppression effect, realizing a shift from passive response to proactive optimization.

[0075] The process of obtaining the stability influence weights for each controlled unit includes: Obtain heterogeneous load data from control deviation data and state feedback data; identify load contribution parameters based on the load deviation data and heterogeneous load data of the controlled unit; identify stability contribution parameters based on the stability deviation data of the controlled unit; identify power contribution parameters based on the power deviation data of the controlled unit; identify position contribution parameters based on the position deviation data of the controlled unit; and identify stability influence weights based on the load contribution parameters, stability contribution parameters, power contribution parameters, and position contribution parameters.

[0076] Preferably, the load contribution parameter includes two parts: one part is the heterogeneous load data in the state feedback data, that is, the load data of the controlled unit in the actual situation; the other part is the load deviation data in the control deviation data, that is, the deviation of the actual load data compared with the benchmark reference model. The load contribution parameter is obtained by normalizing and weighting based on these two parts. The core logic is that the risk of the controlled unit depends not only on the deviation from the benchmark load, but also on how much load it actually bears. The greater the load deviation of the controlled unit and the heavier the actual load it bears (compared to the average value), the higher its load contribution parameter, which means that it contributes the most to system instability.

[0077] The stability contribution parameter is obtained by identifying the stability deviation data of the controlled unit. For example, the stability contribution parameter can be obtained by weighted summation and normalization of stability deviations such as roll angle and pitch angle. The stability contribution parameter is used to quantify the risk of equipment overturning. Roll angle is usually more likely to cause overturning than pitch angle, so it is given a higher weight.

[0078] The dynamic contribution parameters are obtained by identifying the dynamic deviation data of the controlled unit. The dynamic contribution parameters are used to quantify the instability risk caused by uneven acceleration. Abnormal changes in acceleration and torque can cause liquid sloshing or AGV slippage. The greater the deviation from the reference value, the higher the instability risk.

[0079] The position contribution parameters are obtained by identifying the position deviation data of the controlled unit; specifically, this includes calculating the Euclidean distance between the actual position of the controlled unit and the reference position.

[0080] Based on the load contribution parameters, stability contribution parameters, dynamic contribution parameters, and location contribution parameters, stability impact weights are identified. For example, stability impact weights can be obtained by weighted summation of the above contribution parameters, where the weights of each item are preset or obtained through model training, to characterize the comprehensive impact of different contribution parameters on system instability.

[0081] Furthermore, the weight coefficients in the aforementioned stability influence weight calculation process are not fixed values, but are output in real time by the weight prediction network based on the control deviation data, such as a three-layer multilayer perceptron (MLP). This weight prediction network and the adaptive control model undergo end-to-end supervised training, with their common training objective being to minimize the combined runaway index and control deviation data in future time steps.

[0082] This invention provides a diagnostic framework with deep physical insight and interpretability for calculating stability influence weights, innovatively decomposing them into four independent "contribution parameters": load, stability, dynamics, and position. This "factor decomposition"-based diagnostic method enables precise source tracing of control systems, providing a decision-making basis for adaptive control models. This allows them to apply the most effective control impulses for different situations, thereby achieving the most efficient and fundamental stability control.

[0083] Furthermore, during the turning process of the transportation path, even without significant deviation, according to physical principles such as centripetal and centrifugal forces, the controlled units located on the outside of the turning trajectory will inevitably experience a greater lateral overturning moment than the units on the inside, and their instability risk is naturally higher. Therefore, this feedforward information needs to be incorporated into the calculation of the stability influence weights.

[0084] Furthermore, when the event triggering condition is triggered, the adaptive control model will also combine the benchmark reference model to obtain the motion trajectory data of the controlled unit in the future time period and determine whether there is a turning section. When there are no turning sections, the stability impact weight is directly calculated based on the obtained feedback data; When there are curves, the stability impact weights further include a curve contribution parameter to characterize the stability impact of different controlled units on the curve path.

[0085] Specifically, this includes: combining a benchmark reference model and current state feedback data to obtain motion trajectory data of the multiple controlled units; and obtaining the turning path trajectory from the motion trajectory data; Obtain the expected turning angular velocity and turning radius of the formation center, and determine the specific trajectory radius of each controlled unit from the center of rotation. Apply a dynamic model to calculate the expected lateral acceleration that the controlled unit will experience; generate normalized turning contribution parameters based on the expected lateral acceleration. Based on the turning contribution parameters and the four contribution parameters of the implemented feedback, perform identification analysis to obtain the stability influence weights in the case of a turning path.

[0086] This scheme clearly decomposes the stability impact weight into two independent parts: one is the real-time instability contribution based on control deviation data, and the other is the expected risk contribution based on future paths (such as turns) in the benchmark reference model, significantly improving the initiative and predictability of control. Especially in high-risk conditions such as turns, it no longer passively waits for the outer units to generate huge tilt or load deviations before responding, but can identify high-risk units before the risk occurs through a physical model of expected risk contributions, thereby increasing their control weight in advance. This improves system stability and resource allocation efficiency.

[0087] Example 2: This invention also proposes a network adaptive pulse synchronization control system with fusion event triggering, which in this embodiment is used to control multiple autonomous mobile robots (AGVs, as controlled units) to collaboratively transport a large liquid storage tank (as the target load); the structure of the system is as follows. Figure 3 As shown, it specifically includes: The data acquisition module controls multiple controlled units to perform target load transport tasks; identifies the status of the controlled units and obtains status feedback data of the multiple controlled units; and detects the status of the target load in real time and obtains disturbance feedback data.

[0088] The system controls multiple AGVs to perform material handling tasks and collects status and disturbance data in real time. Status feedback data: Each AGV reports its status feedback data in real time through its inertial measurement unit, wheel speed encoder and high-precision differential GPS, including: heterogeneous load data, instantaneous load estimated by pressure sensors; power data, actual speed, acceleration and drive torque; stability data, roll angle, pitch angle and yaw rate; and position data, high-precision coordinates.

[0089] Disturbance feedback number: An IoT pressure sensor array is deployed in the liquid storage tank to acquire pressure time series data in real time, and the sloshing amplitude and sloshing frequency are calculated by fast Fourier transform; then, the sloshing energy is calculated based on a preset fluid dynamics model.

[0090] The disturbance judgment module constructs a disturbance identification model to identify disturbance feedback data and analyzes performance indicators to determine whether preset event triggering conditions are met. When the disturbance feedback data meets the event triggering conditions, adaptive pulse control is initiated.

[0091] In this embodiment, the disturbance identification model is a time-series analysis network built on a gated cyclic unit; the network receives a time-series data sequence of [sway amplitude, sway frequency, sway energy] from the past 3 seconds as input.

[0092] Runaway Risk Quantification Layer: Through a fully connected layer and a Softmax activation function, the swaying frequency pattern analyzed by GRU is classified and matched with the preset runaway disturbance type, and a runaway risk factor in the range of [0,1] is output.

[0093] Runaway magnitude quantization layer: The hidden layer of the GRU, for example, includes 128 units, which captures the dynamic trend of sway amplitude and energy. Its output passes through a fully connected layer and a sigmoid activation function to generate a runaway magnitude factor in the range of [0,1].

[0094] Performance analysis trigger layer: This layer explicitly weights and fuses the two factors to calculate the comprehensive runaway index.

[0095] The corresponding risk weights and magnitude weights were determined through offline simulation training.

[0096] The pulse control module, wherein the adaptive pulse control includes: acquiring the state feedback data of the multiple controlled units, and performing comparative analysis based on a preset benchmark reference model of the multiple controlled units to determine the control deviation data.

[0097] An adaptive control model is constructed based on the control deviation data, disturbance feedback data, and state feedback data to perform adaptive parameter optimization and generate control pulse parameters. The control pulse parameters include pulse amplitude, pulse width, pulse shape, and coordinated application time. Multiple controlled units are synchronously controlled based on the control pulse parameters.

[0098] First, acquire the status feedback data of all AGVs and compare it with the benchmark reference model to determine the control deviation data of each AGV, including load deviation, power deviation, stability deviation, and position deviation.

[0099] An adaptive control model is initiated for parameter optimization. This adaptive control model is a Transformer-based encoder-decoder architecture. The model's hyperparameter settings include: a model dimension of 256, containing 4 encoding layers and 4 decoding layers, with each layer using 8 attention heads.

[0100] The encoder's input is a set of tokens, including: The global perturbation token contains the current amplitude, frequency, and energy of the sway. It is obtained by mapping the global perturbation data to the model dimension through linear projection. Each AGV has a unit-specific token containing the AGV's status feedback data and control deviation data.

[0101] Weight Calculation: Before the data is fed into the Transformer encoder, a weight prediction network is first used to calculate the stability influence weight of each AGV. This weight prediction network is a three-layer multilayer perceptron (MLP). Its input layer receives the specific quantified values ​​of the AGV's contribution parameters, such as: [load deviation data, heterogeneous load data, stability deviation data (tilt angle), stability deviation data (pitch angle), power deviation data (acceleration), power deviation data (torque), position deviation data (X), position deviation data (Y)]. Its output layer outputs a weight in the range of [0,1]. This weight is then incorporated into the AGV's unit-specific token as a key adjustment factor for the Transformer attention mechanism.

[0102] During the offline training phase, the MLP is trained end-to-end together with the Transformer main model. The joint loss function aims to simultaneously minimize the predicted runaway index and the predicted control bias at future time steps.

[0103] Furthermore, the Transformer decoder employs an autoregressive decoding strategy to generate desired state parameters at each time step, and finally outputs its "controlled unit desired state parameters" for each AGV, such as the desired instantaneous corrected acceleration and the desired instantaneous corrected angular velocity. This module includes a pulse generation and modulation unit, which converts the above-mentioned desired state parameters into control pulse parameters that can be executed by the driver.

[0104] Pulse shape determination: For liquid loads, to avoid secondary sloshing, the pulse shape is preset to a trapezoidal pulse with a smooth rising edge and falling edge.

[0105] Specifically, the pulse shape is determined based on the dynamic model of liquid sloshing and the control objective: Common pulse shapes include rectangular pulses, sinusoidal pulses, and trapezoidal pulses; The advantage of rectangular pulses is their fast response; however, their disadvantage is that sudden acceleration changes can cause instantaneous pressure jumps on the liquid surface, triggering high-frequency oscillations in the liquid and potentially exacerbating secondary sloshing.

[0106] The advantages of sinusoidal pulses are smooth acceleration and bounded second derivative; the disadvantages are that the effective control time is short after the control amplitude decreases to a certain level; if it is necessary to extend the control time, the pulse width must be increased, which may lead to excessive container movement.

[0107] The advantages of trapezoidal pulses are that they provide constant acceleration on the rising and falling edges, avoiding sudden acceleration changes, and the plateau segment provides constant speed variation, enabling continuous and effective control of the container. Finally, the ratio of the rising, plateau, and falling phases can be flexibly adjusted to match the sloshing cycle of the liquid and the acceleration response limits of the AGV.

[0108] Pulse amplitude determination: The pulse amplitude is calculated based on a complete multibody dynamics model. The reaction force exerted by the liquid on the container is a nonlinear function, which is determined by the liquid sloshing, the impact force with the container wall, and the liquid viscous damping. Determining the desired pulse amplitude is a multi-objective optimization problem, usually solved by model predictive control (MPC). The goal is to minimize liquid sloshing and restore formation synchronization within a predetermined time window after the pulse is applied. When a pulse event is triggered, the following steps are performed: read the current system state, use the aforementioned multi-objective optimization solver to calculate the optimal pulse amplitude. Considering the physical limitations of the AGV driver, saturate the optimal pulse amplitude and convert it into driver instructions.

[0109] Pulse width determination: The pulse width is calculated based on the total impulse and amplitude limits required by the system. For example, within a fixed control cycle, if the pulse amplitude exceeds the driver's upper limit, the pulse amplitude is reduced to the upper limit, and the pulse width is extended accordingly to ensure that the total corrected impulse remains unchanged.

[0110] Cooperative application timing determination: Solve network latency and terrain interference, such as when the IMU detects severe turbulence, and determine an optimal, possibly asynchronous application time for each AGV to achieve advanced coordination with consistent objectives and asynchronous timing.

[0111] Control execution: The final generated control pulse parameters are sent to the underlying driver of each AGV to achieve precise synchronous control of multiple controlled units.

[0112] 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 of fusion event-triggered network adaptive impulse synchronization control, characterized in that, The method comprises: controlling a plurality of controlled units to perform a target load carrying task; identifying the state of the controlled units to obtain state feedback data of the plurality of controlled units; real-time detection of the state of the target load to obtain disturbance feedback data; constructing a disturbance identification model to identify the disturbance feedback data, and determining whether a preset event trigger condition is met through performance index analysis; when the disturbance feedback data meets the event trigger condition, starting adaptive pulse control; the adaptive pulse control comprises: obtaining the state feedback data of the plurality of controlled units, and comparing and analyzing based on a preset plurality of controlled unit reference model to determine control deviation data; constructing an adaptive control model based on the control deviation data, the disturbance feedback data, and the state feedback data to perform adaptive parameter optimization and generate control pulse parameters; the control pulse parameters include pulse amplitude, pulse width, pulse shape, and cooperative application time; and synchronously controlling the plurality of controlled units based on the control pulse parameters.

2. The fusion event-triggered network adaptive pulse synchronization control method according to claim 1, characterized in that, The target load is a liquid load, and the disturbance feedback data is used to represent the fluid dynamics parameters of the liquid load shaking state; The disturbance feedback data specifically includes: real-time acquisition and identification of the liquid shaking amplitude, shaking frequency, and shaking energy by an Internet of Things sensor installed inside the target load container.

3. The fusion event-triggered network adaptive pulse synchronization control method according to claim 1, characterized in that: The disturbance identification model is constructed based on a recurrent neural network to identify the dynamic characteristics of the disturbance feedback data and perform performance index analysis; including a loss-of-control risk quantification layer, a loss-of-control magnitude quantification layer, and a performance analysis trigger layer; The loss-of-control risk quantification layer identifies and classifies the shaking frequency of the target load to determine the loss-of-control disturbance type; Based on the loss-of-control disturbance type, determine the loss-of-control risk factor; The loss-of-control magnitude quantification layer normalizes the shaking amplitude and shaking energy to obtain shaking amplitude normalized data and shaking energy normalized data, and then performs recurrent neural network identification and judgment to obtain a loss-of-control magnitude factor; The performance analysis trigger layer weights and fuses the loss-of-control risk factor and the loss-of-control magnitude factor to obtain a comprehensive loss-of-control index; and judges the event trigger condition of the comprehensive loss-of-control index based on a preset loss-of-control threshold.

4. The fusion event-triggered network adaptive pulse synchronization control method according to claim 1, wherein, The state feedback data includes: heterogeneous load data, power data, stability data, and position data of the controlled unit; The heterogeneous load data is the load change data of each controlled unit during the carrying process; The power data includes the actual acceleration and actual speed of the controlled unit, and the actual output torque and driving current of the driving system; The stability data includes the device roll angle, device pitch angle, and device yaw angular velocity of the controlled unit; The position data includes high-precision position coordinate data of the controlled unit.

5. The fusion event-triggered network adaptive pulse synchronization control method according to claim 1, wherein, The reference model is constructed based on historical carrying data, which includes: obtaining the reference cooperative state of the plurality of controlled units in the cooperative carrying target load task; the reference cooperative state includes reference load data, reference power data, reference stability data, and reference position data; The state feedback data of the plurality of controlled units is data-differentiated with the reference cooperative state to obtain control deviation data; the deviation data includes load deviation data, power deviation data, stability deviation data, and position deviation data.

6. The fusion event-triggered network adaptive pulse synchronization control method according to claim 1, wherein, The adaptive parameter optimization process includes: The adaptive control model standardizes and identifies the heterogeneous load data in the control deviation data and the controlled unit state feedback data to obtain a stability influence weight of each controlled unit; For each controlled unit, the stability influence weight and the state feedback data are obtained as controlled unit parameters; all controlled unit parameters and corresponding disturbance feedback data are identified to output expected state parameters of the controlled units; and control pulse parameters are determined based on the expected parameters of the controlled units.

7. The fusion event-triggered network adaptive pulse synchronization control method according to claim 6, wherein, The stability influence weight of each controlled unit is obtained by: The heterogeneous load data in the control deviation data and the state feedback data is obtained; the load contribution parameters are identified based on the load deviation data and the heterogeneous load data of the controlled units; the stability contribution parameters are identified based on the stability deviation data of the controlled units; the power contribution parameters are identified based on the power deviation data of the controlled units; the position contribution parameters are identified based on the position deviation data of the controlled units; and the stability influence weight is obtained by weighted identification based on the load contribution parameters, the stability contribution parameters, the power contribution parameters, and the position contribution parameters.

8. A fusion event-triggered network adaptive impulse synchronization control system, characterized in that, It includes: A data acquisition module controls a plurality of controlled units to perform a carrying task of a target load; The state of the controlled units is identified to obtain state feedback data of the plurality of controlled units; The state of the target load is detected in real time to obtain disturbance feedback data; A disturbance judgment module constructs a disturbance identification model to identify the disturbance feedback data, and determines whether a preset event trigger condition is met through performance index analysis; When the disturbance feedback data meets the event trigger condition, adaptive pulse control is started; A pulse control module includes: obtaining the state feedback data of the plurality of controlled units, and comparing and analyzing based on a preset reference model of the plurality of controlled units to determine control deviation data; An adaptive control model is constructed to perform adaptive parameter optimization based on the control deviation data, the disturbance feedback data, and the state feedback data to generate control pulse parameters; the control pulse parameters include pulse amplitude, pulse width, pulse shape, and cooperative application time; and the plurality of controlled units are synchronously controlled based on the control pulse parameters.

Images