12 results about "Neural network controller" patented technology

This invention relates to the field of spacecraft orbit control technology, and particularly to a spacecraft approach control method based on the fusion of imitation learning and reinforcement learning. The method includes: S1: generating initial and target states, solving for the optimal fuel transfer trajectory offline using a direct method, and constructing an expert dataset; S2: based on the expert dataset, training a neural network controller through imitation learning to obtain a pre-trained model; S3: loading the weights of the pre-trained model, constructing a composite reward function, and training the policy network using a near-end policy optimization algorithm to obtain an optimized policy network model; S4: acquiring relative state data, optimizing the policy network model, and outputting the control quantity for the optimal fuel consumption trajectory transfer in real time. This invention rapidly masters an approximately optimal policy through imitation learning and surpasses the performance of expert data through reinforcement learning, completing control command solutions in milliseconds. It balances terminal control accuracy and fuel consumption optimization, offering advantages such as high solution efficiency, excellent control accuracy, and strong generalization ability.

This invention provides an integrated method for pruning and controlling large-scale models of unmanned aerial vehicles (UAVs), belonging to the field of model lightweighting technology. The proposed method significantly reduces model size and computational load, greatly improves inference speed and control response performance, and enables the deployment of complex neural network controllers on resource-constrained airborne platforms. Simultaneously, the lightweight model maintains or approaches the control accuracy and stability of the original uncompressed model. This method has wide applicability and strong scalability, covering various scenarios from basic flight control to complex mission control, providing a feasible, efficient, and cost-effective solution for applying large models to airborne controllers of UAVs, and possesses high engineering practical value.

This invention discloses a collaborative training method and system for autonomous driving planning and control, belonging to the field of autonomous driving technology. The method includes: acquiring the current state of the vehicle, the states of surrounding traffic participants, and reference road information; using a Transformer-based planning network to output a control sequence, which is then used to generate a planned trajectory via a vehicle kinematics model; inputting the planned trajectory into a model predictive control-based neural network controller for dynamic tracking optimization to obtain the executed control quantity and predicted vehicle dynamic state; constructing a control loss and backpropagating its gradient to the planning network; and, during training, layering constraints on the planning cost term, using safety and traffic rule-related terms as hard constraints and comfort-related terms as soft constraints, and employing a multi-stage cost scheduling strategy to adjust the constraint strength or weight. This invention achieves collaborative optimization of planning and control, improves the dynamic executability and safety compliance of the trajectory, and enhances training stability and closed-loop robustness.

This invention discloses a precise deagglomeration and dispersal modification control system for non-metallic powders, belonging to the field of intelligent control technology based on deep learning. Specifically, it includes: a non-metallic powder identification module, a non-metallic powder deagglomeration module, an atomization coating module, and a coordination control module. The non-metallic powder identification module deploys hardware units to collect raw powder data; the non-metallic powder deagglomeration module constructs a calculation model to calculate the energy required for dispersing the non-metallic powder and implements precise deagglomeration; the atomization coating module uses reverse reasoning to charge and directionally adsorb droplets, completing the atomization of the modifier and coating of the newly formed surface in the same space where deagglomeration occurs; the coordination control module, through the construction of a time-window preemptive scheduling control strategy and the embedding of a main control board with a built-in feedforward fuzzy neural network controller, triggers the injection of the modifier during the interval of deagglomeration energy release, achieving precise synchronous control of "deagglomeration equals modification".

This invention discloses a collaborative training method and system for autonomous driving planning and control, belonging to the field of autonomous driving technology. The method includes: acquiring the current state of the vehicle, the states of surrounding traffic participants, and reference road information; using a Transformer-based planning network to output a control sequence, which is then used to generate a planned trajectory via a vehicle kinematics model; inputting the planned trajectory into a model predictive control-based neural network controller for dynamic tracking optimization to obtain the executed control quantity and predicted vehicle dynamic state; constructing a control loss and backpropagating its gradient to the planning network; and, during training, layering constraints on the planning cost term, using safety and traffic rule-related terms as hard constraints and comfort-related terms as soft constraints, and employing a multi-stage cost scheduling strategy to adjust the constraint strength or weight. This invention achieves collaborative optimization of planning and control, improves the dynamic executability and safety compliance of the trajectory, and enhances training stability and closed-loop robustness.

This invention proposes a multi-constraint-oriented robotic arm tracking control method. The method includes: Step 1: Establishing a robotic arm dynamic model and constraint description; Step 2: Constraint transformation based on a general obstacle function; Step 3: Designing a virtual control law and an adaptive neural network controller; Step 4: Stability analysis and fixed-time convergence proof; Step 5: Constraint satisfaction verification. This invention's method can strictly satisfy state constraints, avoiding the risk of system downtime or mechanical damage due to state exceeding limits, and also achieves higher tracking accuracy.

The application belongs to the field of pumped storage and relates to a variable-speed pumped storage power control method based on a BP algorithm. In view of the problems such as nonlinearity and time variability in the power control of a variable-speed pumped storage unit, a variable-speed pumped storage power control method based on a BP algorithm is provided. By establishing a mathematical model of a variable-speed pumped storage system, a 3-5-1 feedforward neural network controller with online learning ability is designed, and the controller uses an error back propagation algorithm to adjust the network weight in real time. The Matlab / Simulink simulation results show that, compared with a traditional PI controller, the BP neural network controller has a power overshoot close to zero, a small steady-state error and a smooth response without oscillation in step response.

This invention provides a four-cylinder outrigger synchronous motion control system, comprising an oil tank, an oil pump motor, an oil pump, a proportional valve, a directional valve, a display and control assembly, a BP neural network PID controller, and four outriggers. The oil pump is connected to the oil pump motor, the proportional valve, the directional valve, the four outriggers, and the oil tank. The four outriggers are equipped with displacement sensors and motion limit proximity switches. The display and control assembly includes a CPU module, an AD module, a DA module, a KI module, and a KO module. This invention provides excellent control over complex systems, exhibiting strong anti-interference capabilities, high robustness, and stable and reliable synchronous control. By establishing a closed-loop system through real-time position feedback from displacement sensors, it enables synchronous movement control of the four outriggers from both local and remote host computers. It features strong anti-interference capabilities, fast response speed, and reliable and stable synchronous operation.

The application belongs to the technical field of motor speed regulation control, and designs a permanent magnet synchronous motor speed regulation control method based on an improved cerebellar model neural network controller, a storage medium and equipment, which comprises the following steps: firstly, a mathematical model of a permanent magnet synchronous motor is established, and then an improved cerebellar model neural network controller is designed and applied to a speed loop in a vector control structure of the permanent magnet synchronous motor; the controller improves the weight adjustment mode and learning rate of the original network, adopts a gradient descent method combined with an inverse relationship of learning times to update the weight of the cerebellar model neural network, replaces a calculation mode of division by a network generalization parameter C for average distribution, and improves the learning efficiency of the network; meanwhile, the value of the network learning rate is dynamically adjusted according to the motor operating state, and the rapidity and stability of the system are improved. Results show that the improved cerebellar model neural network controller effectively improves the response speed of the system and greatly improves the dynamic stability of the permanent magnet synchronous motor.

This invention discloses a robust control method for air spring suspension based on adaptive event triggering. The method first constructs a quarter-vehicle air spring suspension dynamic model considering nonlinear gas flow, actuator saturation, and time-varying time delay, and uses a generalized fuzzy hyperbolic tangent model (GFH) to approximate the nonlinear force of the air spring with arbitrary accuracy. Secondly, an adaptive event triggering mechanism (AET) is designed, which significantly reduces the solenoid valve switching frequency and the burden on the vehicle network communication by dynamically adjusting the trigger threshold. Furthermore, based on Lyapunov stability theory and a robust H∞ control framework, an adaptive event triggering robust H∞ neural network controller (AET-RNNC) is proposed, ensuring the asymptotic stability of the closed-loop system under multiple constraints and meeting the preset H∞ performance index. Finally, the method is verified through AmeSim-Simulink co-simulation and hardware-in-the-loop (HIL) real-time experiments: during the height adjustment process from 80 mm to -50 mm, the steady-state error is less than ±3 mm, the number of solenoid valve triggers is reduced by more than 68% compared to traditional time triggering, and the root mean square value of the vehicle acceleration is reduced by 18.7%. This method significantly improves the adjustment accuracy, ride comfort, and lifespan of key components of electronically controlled air suspension. It is particularly suitable for passenger cars, commercial vehicles, and rail vehicles, and has high engineering practical value and mass production potential.

The application provides a large workpiece-based ultrasonic sensitivity detection method, relates to the technical field of detection, and aims to solve the problems of inconsistent sensitivity in the full-thickness range and insufficient detection precision of deep defects in traditional ultrasonic detection. The method sets multiple groups of depth-adaptive equivalent reflectors on the test block, dynamically adjusts the gain value in combination with the fuzzy neural network controller, realizes sensitivity adaptive compensation in the full-thickness range, and ensures that the detection error is stable within ±1%. Based on the dynamic reflector interval design of the material attenuation coefficient and the acoustic parameter, the reflector distribution is optimized, and the complexity of artificial calibration is significantly reduced. The TCG curve is fitted by the piecewise weighted least square method and simulated to generate a high-precision DAC curve, realizing high-confidence output of defect quantitative evaluation. The sensitivity, accuracy and engineering applicability of ultrasonic detection of large workpieces are improved, and the method is particularly suitable for defect detection of complex curvature workpieces.

This invention discloses a wind power optimization control method based on brain-like computing, comprising the following steps: Step 1, bio-inspired event encoding of multi-source heterogeneous sensor data; Step 2, constructing a hierarchical spiking neural network controller model with spatiotemporal memory; Step 3, spiking reinforcement learning training based on multi-objective reward signals; Step 4, robust decoding and safe constraint execution of spiking decision output; Step 5, lifelong online adaptive fine-tuning based on performance monitoring; Revolutionary performance and efficiency improvement: The event-driven characteristics enable the system to consume almost zero power under steady-state or slowly changing conditions, and the overall control loop energy consumption can be reduced by 1-2 orders of magnitude compared to traditional AI solutions; Millisecond-level real-time response: The spiking processing mechanism is naturally matched to asynchronous sensor data, and the end-to-end latency from perception to decision is extremely low (less than 1ms), which can accurately capture and respond to rapid disturbances such as gusts.