On-line Neural Network Research Articles

Adaptive control of an amplified piezoelectric-driven micropositioning stage based on notch filter structure inspiration and neural network online tuning

Abstract Piezoelectric-driven flexure micro/nanopositioning stages often exhibit a low-damping resonance mode, which can easily excite mechanical resonance during high-speed movement, and significantly impact the control system's stability, control bandwidth, and trajectory tracking accuracy. To mitigate the reliance on the precise modeling of stage dynamics inherent in current resonant controllers, an adaptive control method based on a back propagation (BP) neural network was designed to suppress resonance in real-time. First, a piezoelectric-driven flexure micro/nanopositioning stage system was constructed. Next, a feedback controller similar to a notch filter was designed, with bilinear transformation applied based on the system's inherent parameters to determine the initial values. Finally, the designed adaptive control method was tested through trajectory tracking experiments using a triangular wave signal. The experimental results showed that, when tracking the triangular wave signal, the maximum tracking error was reduced by 72.66% compared to proportional-integral (PI) control alone and by 68.66% compared to proportional integral control combined with a traditional notch filter. The tracking results demonstrate a significant improvement in the stage's stability and tracking accuracy.

Adaptive Inverse Nonlinear Optimal Control Based on Finite-Time Concurrent Learning and Semidefinite Programming.

This article investigates the problem of inverse optimal control (IOC) for a class of nonlinear affine systems. An adaptive IOC approach is proposed to recover the cost functional using only the system state data, which integrates the finite-time concurrent learning (FTCL) technique and the semidefinite programming (SDP) technique. First, an identifier neural network (NN) is employed to approximate the unknown nonlinear control policy, and an FTCL-based update law is proposed to estimate the weights of the identifier NN online, which removes the traditional persistent excitation (PE) condition. Moreover, the finite-time convergence as well as the uniformly ultimately boundness (UUB) of estimation error of the identifier NN weights are analysed according to whether or not there exists the identifier NN approximation error. Then, with the help of a value NN for approximating the value function, an SDP problem with a quadratic objective function can be set up for determining the weighting matrices of the cost functional. Finally, simulation results are presented to validate the proposed method.

Stress-related discrete variable topology optimization with handling non-physical stress concentrations

The accuracy of stress calculation with a fixed mesh significantly affects the stress-based topology optimization, due to potential non-physical stress concentrations in voxel-based topology descriptions. This paper proposes a novel problem-independent machine learning enhanced high-precision stress calculation method (PIML-HPSCM) to address this challenge. As an immersed analysis method, PIML-HPSCM combines the high efficiency of fixed mesh with the accuracy of body-fitted mesh, without complex integration schemes of other immersed methods. PIML-HPSCM utilizes the extended multiscale finite element method to depict the material heterogeneity within high-resolution boundary elements. The accurate stress field can then be calculated conveniently by establishing stress evaluation matrices of high-resolution boundary elements. Moreover, the PIML is independent of problem settings and is applicable for various problems with the same governing equation type. Invoking the offline-trained neural network online can enhance stress calculation efficiency by 10–20 times. The stress-based discrete variable topology optimization, which naturally avoids singular stress phenomenon, is efficiently addressed by the sequential approximate integer programming method with PIML-HPSCM. Results from 2D and 3D examples demonstrate that the stresses calculated by PIML-HPSCM are consistent with those by body-fitted mesh, and optimized designs effectively eliminate stress concentrations of initial designs and have uniform stress distributions.

Online identification and feed-forward compensation of nonlinear friction in servo system based on RBF neural network model

In this paper, an online identification and compensation method of nonlinear friction based on radial basis function (RBF) neural network model is proposed for the influence of nonlinear friction on machining accuracy in the low speed process of servo feed system of CNC machine tools. First, a three-layer single-input-output RBF neural network model is established for describing the nonlinear friction of servo feeding system. Second, the neural network online learning algorithm is improved based on adaptive gain, which improves the stability and accuracy of the algorithm. Finally, experiments were carried out on a three-axis milling machine to compensate the friction in the servo feed system in real time based on the online identification results. The results show that the method can effectively improve the online identification accuracy and convergence rate, and effectively improved the low-speed performance of the servo feed system.

Reinforcement learning-based moving-target enclosing control for an unmanned surface vehicle in multi-obstacle environments

This paper investigates the moving-target enclosing problem in multi-obstacle environments for an unmanned surface vehicle (USV) with complex unknown factors, including target dynamics, vehicle dynamics, and disturbances. A reinforcement learning-based moving-target enclosing control scheme is proposed to ensure collision-free behavior and bolster the enclosing capability. Specifically, an extended state observer is deployed to estimate the target dynamics. Leveraging the estimated data and control obstacle functions, a virtual safety control law is formulated to dynamically harmonize obstacle avoidance and target enclosing control. Then, a novel controller is constructed to track this control law utilizing a Nussbaum-type function in conjunction with actor–critic neural networks (NNs). The actor and critic NNs are employed to approximate unknown dynamics encapsulating vehicle dynamics and disturbances, and value function, respectively. The Nussbaum-type function is embedded to adaptively identify unknown inertia mass, regulating the control input of the USV with the actor NN online. The proposed scheme effectively decouples obstacle avoidance and target enclosing control and only relies on the measurable variables. A rigorous theoretical analysis is further employed to ensure the closed-loop stability of the USV system. Eventually, Simulations are demonstrated to validate the effectiveness and superiority of the proposed scheme for the USV in multi-obstacle environments.

Neural Network With Cloud-Based Training for MTPA, Flux-Weakening, and MTPV Control of IPM Motors and Drives

An interior permanent magnet (IPM) motor is a prime electric motor used in electric vehicles (EVs), robots, and electric drones. In these applications, maximum torque per ampere (MTPA), flux-weakening, and maximum torque per volt (MTPV) techniques play a critical role in the efficient and reliable torque control of an IPM motor. Although several approaches have been proposed and developed for this purpose, each has its specific limitations. The objective of this paper is to develop a neural network (NN) method to determine MTPA, flux-weakening, and MTPV operating points for the most efficient torque control of the motor over its full speed range. The NN is trained offline by using the Levenberg-Marquardt backpropagation algorithm, which avoids the disadvantages associated with online NN training. A cloud computing system is proposed for routine offline NN training, which enables the lifetime adaptivity and learning capabilities of the offline-trained NN and overcomes the computational challenges related to online NN training. In addition, for the proposed NN mechanism, training data are collected and stored in a highly random manner, which makes it much more feasible and efficient to implement lifetime adaptivity than any other methods. The proposed method is evaluated via both simulation and hardware experiments, which shows the great performance of the NN-based MTPA, MTPV, and flux-weakening control for an IPM motor over its full speed range. Overall, the proposed method can achieve a fast and accurate current reference generation with a simple NN structure, for optimal torque control of an IPM motor.

Silicon-Compatible Ferroelectric Tunnel Junctions with a SiO2/Hf0.5Zr0.5O2 Composite Barrier as Low-Voltage and Ultra-High-Speed Memristors.

The big data era requires ultrafast, low-power, and silicon-compatible materials and devices for information storage and processing. Here, ferroelectric tunnel junctions (FTJs) based on SiO2/Hf0.5Zr0.5O2 composite barrier and both conducting electrodes are designed and fabricated on Si substrates. The FTJ achieves the fastest write speed of 500ps under 5V (2 orders of magnitude faster than reported silicon-compatible FTJs) or 10ns speed at a low voltage of 1.5V (the lowest voltage among FTJs at similar speeds), low write current density of 1.3 × 104 A cm-2, 8 discrete states, good retention >105 s at 85°C, and endurance >107. In addition, it provides a large read current (88 A cm-2) at 0.1V, 2 orders of magnitude larger than reported FTJs. Interestingly, in FTJ-based synapses, gradually tunable conductance states (128 states) with high linearity (<1) are obtained by 10ns pulses of <1.2V, and a high accuracy of 91.8% in recognizing fashion product images is achieved by online neural network simulations. These results highlight that silicon-compatible HfO2-based FTJs are promising for high-performance nonvolatile memories and electrical synapses.

Optimized operation scheme of flash-memory-based neural network online training with ultra-high endurance

With the rapid development of machine learning, the demand for high-efficient computing becomes more and more urgent. To break the bottleneck of the traditional Von Neumann architecture, computing-in-memory (CIM) has attracted increasing attention in recent years. In this work, to provide a feasible CIM solution for the large-scale neural networks (NN) requiring continuous weight updating in online training, a flash-based computing-in-memory with high endurance (109 cycles) and ultra-fast programming speed is investigated. On the one hand, the proposed programming scheme of channel hot electron injection (CHEI) and hot hole injection (HHI) demonstrate high linearity, symmetric potentiation, and a depression process, which help to improve the training speed and accuracy. On the other hand, the low-damage programming scheme and memory window (MW) optimizations can suppress cell degradation effectively with improved computing accuracy. Even after 109 cycles, the leakage current (I off) of cells remains sub-10pA, ensuring the large-scale computing ability of memory. Further characterizations are done on read disturb to demonstrate its robust reliabilities. By processing CIFAR-10 tasks, it is evident that ~90% accuracy can be achieved after 109 cycles in both ResNet50 and VGG16 NN. Our results suggest that flash-based CIM has great potential to overcome the limitations of traditional Von Neumann architectures and enable high-performance NN online training, which pave the way for further development of artificial intelligence (AI) accelerators.

Adaptive fuzzy neural network finite‐time command filtered control of n‐link robotic systems with actuator saturation

AbstractThe finite time tracking control of n‐link robotic system is studied for model uncertainties and actuator saturation. Firstly, a smooth function and adaptive fuzzy neural network online learning algorithm are designed to address the actuator saturation and dynamic model uncertainties. Secondly, a new finite‐time command filtered technique is proposed to filter the virtual control signal. The improved error compensation signal can reduce the impact of filtering errors, and the tracking errors of system quickly converge to a smaller compact set within finite time. Finally, adaptive fuzzy neural network finite‐time command filtered control achieves finite‐time stability through Lyapunov stability criterion. Simulation results verify the effectiveness of the proposed control.

Disturbance suppression and NN compensation based trajectory tracking of underactuated AUV

For the trajectory tracking of underactuated autonomous underwater vehicle (AUV) subject to disturbances, a disturbance suppression and neural network (NN) compensation based control strategy is proposed. Lyapunov method is used to guide the overall design of control system. To deal with the underactuation, Light of sight (LOS) guidance is used to establish the relationship between heading angle and cross-track error. To suppress the external disturbance, L2-gain design is employed to guarantee the controller robustness. To improve the tracking accuracy, on-line neural networks with guaranteed stability are designed to identify unknown dynamics including the derivatives of virtual controls and errors induced by input saturation. Numerical simulation is performed to verify the effectiveness of the proposed control strategy. Compared with sliding mode controller (SMC) and disturbance observer approaches, the proposed controller performs better in terms of robustness and the input saturation is alleviated.

Adaptive local Principal Component Analysis improves the clustering of high-dimensional data

In local Principal Component Analysis (PCA), a distribution is approximated by multiple units, each representing a local region by a hyper-ellipsoid obtained through PCA. We present an extension for local PCA which adaptively adjusts both the learning rate of each unit and the potential function which guides the competition between the local units. Our local PCA method is an online neural network method where unit centers and shapes are modified after the presentation of each data point. For several benchmark distributions, we demonstrate that our method improves the overall quality of clustering, especially for high-dimensional distributions where many conventional methods do not perform satisfactorily. Our online method is also well suited for the processing of streaming data: The two adaptive mechanisms lead to a quick reorganization of the clustering when the underlying distribution changes.

Adaptive neural network fixed-time sliding mode control for trajectory tracking of underwater vehicle

The effectiveness of Autonomous Underwater Vehicles (AUVs) in diverse underwater tasks is heavily reliant on their ability to perform accurate trajectory tracking. However, due to uncertainties in AUVs modeling and the complex underwater environment disturbances, designing effective trajectory tracking controllers and disturbance observers for AUVs is still a major challenge. To address these uncertainties and enable faster convergence of tracking errors, a trajectory-tracking controller based on fixed-time sliding mode control (FTSMC) and a Radial Basis function neural network (RBFNN) observer are used in this paper. In most cases, the AUV platform carries limited computational resources. In most cases, AUV platforms carry limited computational resources, which restricts the practical use of online neural network methods, and it is particularly important to reduce the complexity of computational neural networks and enhance the real-time performance of the observer. Therefore, we adopted a fast online weight update strategy based on a single parameter. Considering actuator faults and input saturation, passive fault-tolerant control (PFTC) is used in this scheme to further reduce the computational burden. Furthermore, the Lyapunov method is used to demonstrate the fixed-time stability of the individual signals of the system. Finally, simulation results and theoretical analysis demonstrate the superiority and effectiveness of the proposed method.

Neural network observer based on fuzzy auxiliary sliding-mode-control for nonlinear systems

This research suggests the use of neural network observers based on fuzzy auxiliary sliding mode control for the fault estimation and isolation of quadrotor unmanned aerial vehicle sensors. The fuzzy auxiliary sliding-mode-control-based adaptive approach for neural network observer is used for the approximation, reconstruction, and isolation of unknown faults by utilizing the multi-layer neural network. The neural network weight parameters are updated adaptively by using the fuzzy auxiliary sliding-mode-control approach. In conventional neural networks, the gradient descent approach-based back-propagation procedures are adapted for the training of the neural network. In this research, a new concept of the nonlinear controller such as the fuzzy auxiliary sliding-mode-control method is adapted for neural network online training, in which the fuzzy auxiliary sliding-mode-control is used as the learning approach; the neural network is used as a control process that calculates the stable and dynamic learning rates. By the consideration of unknown faults diagnosis and isolation, the online learning approach used in this research has shown the faults estimation, reconstruction, and isolation abruptly and with high accuracy compared to conventional approaches as well the approach used in literature. Strategies adopted in the literature are not capable of fault detection, estimation, and reconstruction with high accuracy and abruptness compared to the method adopted in this research. The presented approach is validated by using the nonlinear dynamics of quadrotor unmanned aerial vehicles, results show the accuracy, abruptness, and efficiency compared to the algorithms adapted in the literature. It is suggested that the proposed strategy can be integrated into nonlinear systems fault diagnosis, fault isolation, and for increasing the system performance.

Low-epsilon adversarial attack against a neural network online image stream classifier

An adversary intercepts a stream of images between a sender and a receiver neural network classifier. To minimize its footprint, the adversary only attacks a limited number of images within the stream. The adversary is interested in maximizing the number of successfully conducted attacks among all performed attacks. Upon the arrival of each image and before the arrival of the following image, the adversary must irrevocably decide whether it wants to attack the current image or not. The target model is a fixed deep neural network that may use any form of regularization. The adversary has query access to the target model, which can feed images and obtain the loss, which may contain regularization and classification loss terms. Since this paper’s proposed method needs classification loss term alone, it also suggests a novel method in which the adversary estimates the regularization loss term and eliminates it. All images are partitioned into three groups based on their after-attack classification loss and treated according to their group. Moreover, this paper provides some promising test results on various datasets.

Using a Convolutional Neural Network to Convert Black-and-White Photos into a Colorized Format in the Telegram Bot

Artificial neural networks will be used to solve our problem, since they are best suited for making predictions, due to their ability to detect non-obvious connections and the most general patterns of input and output data. A diagram of the architecture of the created neural network has been developed. Lab color coding method considered. The program is implemented in the high-level language Python. To create a neural network, an open software library for machine learning TensorFlow was used. The ReLU function (linear rectifier or semi-linear element) was chosen as the activation function for all layers, except for the output one. A telegram bot has been developed, the main purpose of which is to convert the sent black-and-white images into a color format. With the help of the developed software algorithm, it is also possible to solve the problem of low quality and low resolution of images by processing them with an online neural network. After the received images can be saved

Online evolutionary neural architecture search for multivariate non-stationary time series forecasting

Time series forecasting (TSF) is one of the most important tasks in data science. TSF models are usually pre-trained with historical data and then applied on future unseen datapoints. However, real-world time series data is usually non-stationary and models trained offline usually face problems from data drift. Models trained and designed in an offline fashion can not quickly adapt to changes quickly or be deployed in real-time. To address these issues, this work presents the Online NeuroEvolution-based Neural Architecture Search (ONE-NAS) algorithm, which is a novel neural architecture search method capable of automatically designing and dynamically training recurrent neural networks (RNNs) for online forecasting tasks. Without any pre-training, ONE-NAS utilizes populations of RNNs that are continuously updated with new network structures and weights in response to new multivariate input data. ONE-NAS is tested on real-world, large-scale multivariate wind turbine data as well as the univariate Dow Jones Industrial Average (DJIA) dataset. Results demonstrate that ONE-NAS outperforms traditional statistical time series forecasting methods, including online linear regression, fixed long short-term memory (LSTM) and gated recurrent unit (GRU) models trained online, as well as state-of-the-art, online ARIMA strategies. Additionally, results show that utilizing multiple populations of RNNs which are periodically repopulated provide significant performance improvements, allowing this online neural network architecture design and training to be successful.

A 3D Printed, Bionic Hand Powered by EMG Signals and Controlled by an Online Neural Network.

About 8% of the Ecuadorian population suffers some type of amputation of upper or lower limbs. Due to the high cost of a prosthesis and the fact that the salary of an average worker in the country reached 248 USD in August 2021, they experience a great labor disadvantage and only 17% of them are employed. Thanks to advances in 3D printing and the accessibility of bioelectric sensors, it is now possible to create economically accessible proposals. This work proposes the design of a hand prosthesis that uses electromyography (EMG) signals and neural networks for real-time control. The integrated system has a mechanical and electronic design, and the latter integrates artificial intelligence for control. To train the algorithm, an experimental methodology was developed to record muscle activity in upper extremities associated with specific tasks, using three EMG surface sensors. These data were used to train a five-layer neural network. the trained model was compressed and exported using TensorflowLite. The prosthesis consisted of a gripper and a pivot base, which were designed in Fusion 360 considering the movement restrictions and the maximum loads. It was actuated in real time thanks to the design of an electronic circuit that used an ESP32 development board, which was responsible for recording, processing and classifying the EMG signals associated with a motor intention, and to actuate the hand prosthesis. As a result of this work, a database with 60 electromyographic activity records from three tasks was released. The classification algorithm was able to detect the three muscle tasks with an accuracy of 78.67% and a response time of 80 ms. Finally, the 3D printed prosthesis was able to support a weight of 500 g with a safety factor equal to 15.

Neural network-based adaptive event-triggered control for cyber–physical systems under resource constraints and hybrid cyberattacks

In this paper, the neural network (NN)-based adaptive event-triggered control (AETC) strategy is developed for cyber–physical systems (CPSs) under resource constraints and hybrid cyberattacks. First, by considering the nonperiodic denial-of-service (DoS) attacks that only constrain their frequency and duration and removing the strict assumption on the malicious information injected by deception attacks in existing literature, an unexplored hybrid cyberattack model is proposed. Second, a novel NN-based adaptive event-triggered state-feedback controller is designed, which not only flexibly solves the resource constraints of CPSs while considering the adverse effects of DoS attacks, but also adopts an on-line weight-adjusted NN to approximate and mitigate the malicious information injected by deception attacks. Then, the switched delay system model is established for the closed-loop control system with the designed novel controller. By carefully constructing a switched Lyapunov–Krasovskii functional and using the switched delay system theory and NN technology, the explicit form of the designed novel controller is obtained, which can ensure that the closed-loop control system is exponentially asymptotically stable (EAS). Finally, a numerical example is given to illustrate the effectiveness of the proposed strategy.

Advanced non-linear control based on artificial intelligence tuner for hand tremor suppression

One of the most common symptoms among diverse patients (e.g. Parkinson’s disease, essential tremor) is hand tremor. Deep brain stimulation (DBS) has been a popular method of controlling hand tremor in recent years. Although recent advancements in the DBS system have enabled DBS applications in the context of various neurodegenerative disorders, the deployment of such interface tactics in DBS applications is seriously threatened by un-adaptability concerns and a lack of an accurate model. In response to these issues, the current letter discusses the construction of an intelligent technique for minimizing hand tremor and self-adaptive DBS settings without requiring a proper hand tremor system modeling setup. For this purpose, the hand tremor is suppressed using an intelligent proportional–integral (iPI) as the main controller and the sliding mode (SM) as an observer. Furthermore, as an auxiliary controller, a deep reinforcement learning (DRL) method with the Actor-Critic structure is adaptively incorporated. The feedback parameters of the ultra-local model (ULM) controller will be adaptively modified utilizing the DRL method by online neural network (NN) updating in the proposed strategy. The advantages of the ULM based on DRL structure over state-of-the-art approaches are demonstrated by a complete assessment.

Long short-term memory (LSTM) model-based reinforcement learning for nonlinear mass spring damper system control

The Neural Networks (NN) model which is incorporated in the control system design has been studied, and the results show better performance than the mathematical model approach. However, some studies consider that only offline NN model learning and does not use the online NN model learning directly on the control system. As a result, the controller's performance decreases due to changes in the system environment from time to time. The Reinforcement Learning (RL) method has been investigated intensively, especially Model-based RL (Mb-RL) to predict system dynamics. It has been investigated and performs well in making the system more robust to environmental changes by enabling online learning. This paper proposes online learning of local dynamics using the Mb-RL method by utilizing Long Short-Term Memory (LSTM) model. We consider Model Predictive Control (MPC) scheme as an agent of the Mb-RL method to control the regulatory trajectory objectives with a random shooting policy to search for the minimum objective function. A nonlinear Mass Spring Damper (NMSD) system with parameter-varying linear inertia is used to demonstrate the effectiveness of the proposed method. The simulation results show that the system can effectively control high-oscillating nonlinear systems with good performance.