In high degree of freedom humanoid robots control, improving task performance and reducing whole-body joint torque load often conflict. This paper proposes a task-prioritized cooperative torque-cost optimization method for upper-lower body dual-policy decomposed reinforcement learning control. To improve training stability when jointly optimizing task performance and the coupled whole-body torque cost under the dual-policy architecture, this paper constructs a global torque-cost critic. Gradient conflicts between the primary task objective and the torque-cost objective are addressed by projecting the torque-cost gradient onto the orthogonal subspace of the task gradient. Furthermore, this paper introduces a ratio constraint to limit the magnitude of the torque-cost gradient correction term. Physics-based simulation experiments on Unitree G1 for standing and walking demonstrate that, relative to a linearly weighted baseline, the proposed method reduces the mean output torque across all joints while preserving task performance, and improves motion smoothness.

Existing fault estimation methods for Markov jump linear systems, particularly the interacting multiple model approaches, heavily rely on pre-defined fault model banks. This reliance renders them ineffective against faults with unknown or time-varying dynamics. Conversely, emerging Bayesian estimation approaches are typically confined to single-mode systems and fail to accommodate stochastic mode switching. To bridge this gap, this paper proposes a novel Bayesian framework for the real-time joint estimation of system states, hidden modes, and mode-dependent faults without requiring prior fault dynamic knowledge. In particular, an adaptive fault update mechanism based on innovation moments is introduced to address the trend discontinuity of fault signals caused by mode transitions, a challenge not addressed by standard recursive Bayesian estimators. Results from a numerical simulation example and a fermentation process simulation case study show improved robustness and estimation accuracy over traditional methods. The adaptive fault update also tracks abrupt fault changes while reducing fluctuations when the fault signal varies slowly.

To tackle the performance decline in zero-shot fault diagnosis under generalized zero-shot learning setting, this paper develops a generative cross-modal alignment network (GCMAN) method. Specifically, the GCMAN reconstructs sample features and attribute vectors through dual-modal encoding and quantifies cross-modal alignment using Barlow matrix. It enhances generalization capability through a sample generation strategy combining features and attribute spaces. Validated via a power transmission system case study, the developed method exhibits outstanding efficacy for zero-shot and generalized zero-shot learning scenarios.

An identification algorithm based on extended basis pursuit de-noising iteration is proposed for feedback nonlinear systems with unknown time-delay in the forward channel and unknown system orders. Firstly, an identification model of the closed-loop system is established based on the input data and the output data of the feedback nonlinear block. Secondly, considering the unknown time-delay and orders in the forward channel by designing polynomials with sufficiently high orders based on the over-parametrization method, a high dimensional sparse vector with a sparsity equal to the number of parameters to be identified is obtained. Then, in order to solve the problem that forward channel output terms and noise terms in the information vector are unmeasurable, the approach taken here is to replace the unknown terms with estimated outputs and estimated residuals, respectively. Finally, based on the over-parameterization model and the basis pursuit de-noising algorithm, an extended basis pursuit de-noising iterative algorithm is proposed to estimate the unknown parameters, orders, and time-delay of the feedback nonlinear systems. A simulation example is given to demonstrate the effectiveness of the proposed algorithm.

To address the problem of simultaneously identifying both seen and unseen classes of faults in zero-shot fault diagnosis (ZSFD), this paper proposes a generative model combining cross-modal reconstruction and semantic alignment (CRSA-GM) for generalized ZSFD. First, a cross-modal alignment module based on the Barlow matrix is designed to learn fault representations with lower redundancy and higher consistency to enhance diagnostic performance. Then, a generation strategy based on the center-consistency assumption is introduced, which integrates fault information from both the feature space and the attribute space to synthesize samples of unseen fault classes. Finally, an incremental learning strategy is adopted, which is designed to actively learn from the synthetic samples to acquire knowledge about unseen fault classes. The cases on the three-phase transmission system and the Tennessee Eastman process demonstrate that the proposed CRSA-GM exhibits remarkable superiority and effectiveness in both zero-shot and generalized zero-shot fault diagnosis.

To build a compact model that meets the requirements of small medical devices, a lightweight self-distillation network (LSDNet) for cervical cell classification is developed in this paper. First, a multi-scale large kernel attention module is proposed to enhance extraction capability of multi-scale fine-grained features. Second, a deep layer selective aggregation structure is designed to selectively combine features at different levels in the network. Finally, the knowledge self-distillation is adopted to compress the network by inducing shallow classifier from different depths of the network, with the accuracy of the deepest classifier improved. The experimental results illustrate that the proposed LSDNet has the fewest parameters and achieves the most advanced classification performance compared to classical networks in recent publications.

State estimation for the Markov jump linear system (MJLS) is a intractable task when the unpredictable measurement loss exists. Although the conventional methods, such as interacting multiple-model method, are widely used in MJLS, their performance still depends on the known transition probability matrix (TPM). In this article, a novel adaptive state estimation method is proposed for MJLS with unknown measurement loss and TPM based on variational Bayesian inference. Specifically, under system state dynamic and measurement loss are independent, the system state, measurement loss probability and TPM are jointly inferred. In particular, when the stochastic measurement loss occurs, a selective learning mechanism is used to the updating of TPM. The efficiency and superiority of the proposed method is verified by a numerical example and a fermenter process compared with the existing methods.

Summary There are two main difficulties for identification of dual‐rate sampled nonlinear systems; one is the unknown nonlinear properties and the other is the dual‐rate sampled data. Considering the above two points, this paper proposes a cycle reservoir with regular jumps (CRJ) network based recursive identification algorithm. Unlike a traditional method, the proposed one does not require a priori knowledge of the nonlinearity and can handle dual‐rate data; therefore, its applicability can be guaranteed. Firstly, the CRJ network is used to describe the nonlinear characteristics of the target systems. However, the update of the CRJ network requires the missing outputs, thus the auxiliary model identification theory is adopted and the missing outputs are replaced by the estimated outputs of the network. In this process, the output weights of the CRJ network are updated at a slow rate, while the estimates of the missing outputs are updated quickly, resulting in an interactive estimation computation process. Then, to improve the identification accuracy, the hyper‐parameters of the CRJ network are optimized by means of the particle swarm optimization algorithm. Finally, simulation examples are presented to demonstrate the effectiveness of the proposed algorithm.

针对含有执行器故障的不确定连续系统,提出了一种基于扩张状态观测器的迭代学习容错控制方法.首先,将实际故障执行器和理想线性执行器之间的偏差合并到系统总扰动中,并利用扩张状态观测器在批次内快速估计和补偿总扰动,保证批次内广义去扰动态模型的不变性;然后,基于广义去扰动态模型设计PD型迭代学习控制律,使系统在批次方向获得满意的跟踪性能,并研究执行器故障情形迭代学习控制系统的收敛性;最后,通过一个二阶控制系统的仿真实验验证了所提控制方法的有效性.

The data driven fault detection models usually require that the training data must be measured under normal operating conditions. However, in the actual industrial processes, it is possible that the collected data set contains outliers even under normal working conditions. In this case, the control limits of the traditional method based on multivariate statistical analysis are often heavily influenced by the outliers, which results in a large number of missed failures. Therefore,in order to ensure that the monitoring model still has good performance even when the training data contains outliers, this paper proposed a fault detection method based on auto-associative kernel regression（AAKR）. First, the training set is whitened on the basis of a robust whitening algorithm that minimizes β divergence to eliminate the influence of correlation between variables on sample similarity measurement. Then, AAKR reconstructs the validation data under normal working conditions to obtain the residuals and establish the correct detection index. In order to avoid the influence of outliers on the reconstructions of faulty test data, a truncation function is constructed to avoid the involvement of outliers similar to the faulty samples in reconstruction. All residual variables involved in Q statistic construction were weighted based on the exponentially weighted moving average（EWMA） to obtain the new monitoring statistic. The proposed method is applied to the Tennessee Eastman（TE） process to verify the effectiveness of the proposed fault detection algorithm.

This paper proposes an identification method of an input-nonlinear system with saturation and dead-zone nonlinearity using the asynchronous input-output data spaced by non-uniform intervals. The piecewise expression of the nonlinear part is simplified as an analytic function with an available switching function. By extending the traditional continuous linear time-invariant processes to non-uniformly sampled input-nonlinear systems, a concise input-output representation model is derived. Based on the key term separation principle and the auxiliary model identification idea, a gradient-based iterative identification algorithm is developed for simultaneously estimating all parameters of the derived model. Through a numerical example, the proposed algorithm shows its superior estimation accuracy compared to the auxiliary model-based forgetting factor stochastic gradient algorithm. Finally, the application to a mathematical model of a two-tank system indicates the effectiveness of the proposed method.

For identification of dual‐rate sampled errors‐in‐variables (EIV) systems with time delays, this paper utilizes the polynomial transformation technique and the redundant rule to derive the augmented dual‐rate model, and proposes a bias compensation‐based least squares (BC‐LS) algorithm. The basic idea is to obtain a biased parameter estimate by using the LS algorithm, then compensate the noise‐induced bias by the estimated noise variances, and finally estimate the time delay according to the unbiased parameter estimate and the given threshold. Considering that proper threshold selection has great influence on the performance of the BC‐LS identification algorithm for high noise level, the generalized orthogonal matching pursuit (GOMP) algorithm is introduced to simultaneously estimate the biased parameters and the time delay, and the bias compensation‐based GOMP (BC‐GOMP) algorithm is further proposed to identify the dual‐rate sampled EIV systems with time delays. The simulation examples demonstrate the effectiveness of the proposed two algorithms, and the BC‐GOMP algorithm performs better especially when the modeling dataset is small.

针对滚动轴承故障诊断方法在低信噪比情况下抗噪性能差以及在不同工况下自适应性不足的问题,提出了一种基于多层训练干扰的卷积神经网络滚动轴承故障诊断方法.首先,通过快速傅里叶变换将原始信号从时域变换到频域并作为卷积神经网络(CNN)的输入,再对多层卷积层进行实时训练干扰,以不同程度随机破坏网络训练,提高模型的抗噪性能和域自适应能力;同时引入批量归一化(BN)方法,改进卷积神经网络结构,改善模型性能.实验结果表明,所提方法不仅在噪声干扰下能达到较高的诊断准确率,同时也能有效地解决工况变化问题.

针对不同相对度的离散线性重复过程,研究有限频域范围的动态迭代学习控制问题.对于零相对度和高相对度的控制对象,结合二维(2D)系统理论,分别设计有限频域的动态迭代学习控制器;然后,运用广义Kalman-Yakubovich-Popov(KYP)引理,以线性矩阵不等式(LMI)的形式给出控制器存在的充分条件以及控制器的增益矩阵;最后,在弹簧阻尼系统和桁架机器人模型的仿真中,与静态迭代学习控制算法进行比较,验证所提算法的优越性和可行性.

利用提升技术可将非均匀采样非线性系统离散化为一个多输入单输出传递函数模型,从而将系统输出表示为非均匀刷新非线性输入和输出回归项的线性参数模型,进一步基于非线性输入的估计或过参数化方法进行辨识.然而,当非线性环节结构未知或不能被可测非均匀输入参数化表示时,上述辨识方法将不再适用.为了解决这个问题,利用核方法将原始非线性数据投影到高维特征空间中使其线性可分,再对投影后的数据应用递推最小二乘算法进行辨识,提出基于核递推最小二乘的非均匀采样非线性系统辨识方法.此外,针对系统含有有色噪声干扰的情况,参考递推增广最小二乘算法的思想,利用估计残差代替不可测噪声,提出核递推增广最小二乘算法.最后,通过仿真例子验证所提算法的有效性.

针对不确定非线性系统的执行器故障问题,提出一种基于扩张状态观测器的自抗扰控制方法.首先,将系统的内部扰动、外部扰动,以及实际故障执行器和理想线性执行器之间的偏差定义为总扰动,并利用扩张状态观测器进行实时估计和补偿,消除总扰动对系统输出的影响;然后,设计误差反馈控制律,使系统获得满意的跟踪性能;最后,两个二阶控制系统的仿真结果表明了所提控制方法的有效性.

This paper proposes the PD-type iterative learning control (ILC) for multiple time-delays systems with polytopic parameter uncertainty. Based on repetitive process framework, the system under study is equivalently converted into a class of uncertain repetitive processes with multiple time-delays. This approach accounts for effective inclusion of both time and trial domain objectives and hence some requirements on transient dynamics and trial-to-trial error convergence are incorporated for robust design procedures. Additionally, this approach can easily avoid the need for computation with very large dimensioned matrices as it is required for the lifting approach. Also, the proposed controller is designed with the generalized Kalman-Yakubovich-Popov lemma to ensure the monotonic trial-to-trial error convergence in finite frequency domain. This allows us to reduce the conservatism inherent to entire frequency range approaches since the reference signal spectrum reside in a known frequency range. Moreover, the sufficient conditions for the convergence of the resulting scheme are expressed by linear matrix inequalities and hence they are amenable to effective algorithmic solution. Finally, numerical simulations of different scenarios are presented to illustrate the effectiveness of the proposed method. In particular, to highlight the potential interest in PD-type ILC the robust tracking performance is compared with the results for P and D types of ILC.

为解决工业过程中机械臂等特殊重复运行系统的输出在有限时间内无需实现全轨迹跟踪,仅需跟踪期望轨迹上某些特殊关键点的控制问题,针对线性时不变离散系统提出一种基于范数最优的点对点迭代学习控制算法.通过输入输出时间序列矩阵模型变换构建综合性多目标点性能指标函数,求解二次型最优解得到优化迭代学习控制律,同时给出模型标称和不确定情形下最大奇异值形式鲁棒控制算法收敛的充分条件,并进一步推广得到输入约束系统优化控制算法的收敛性结果,最后在三轴龙门机器人模型上验证算法的有效性.

A closed-loop iterative learning fault-tolerant control scheme is proposed for batch process with actuator faults, in which the system parameters have uncertainties simultaneously. Firstly, the batch fault-tolerant controller is designed based on the two-dimensional (2D) system theory, and the batch process of iterative learning control is transformed to an equivalent 2D Roesser model. Then the sufficient conditions for the existence of the controller are analyzed in terms of linear matrix inequality (LMI) technique, and the control gain matrices are derived from the convex optimization problems with LMI constrains. Under these conditions of all additive uncertainties on system parameters and admissible failure, the proposed controller can ensure the closed-loop fault-tolerant performance along time. Finally, the simulation on injection molding nozzle pressure control simulation indicate that the proposed method achieve the design objective well, and show excellent fault tolerance performance.

Complex modern industrial processes often exhibit multimodal characteristic because of different manufacturing strategies. Although the conventional auto-associative kernel regression (AAKR) method is suitable for monitoring the nonlinear multimodal processes, different high Pearson correlations between variables from different modes reduce the fault detection accuracy of AAKR model. Moreover, within-mode process data usually present the property of serial correlation, resulting in the auto-correlation and cross-correlation of variables simultaneously existing in the dynamic processes. In order to reduce the influence of correlation and improve the accuracy of fault detection based on AAKR, a novel multimode process monitoring method combining zero-phase component analysis (ZCA) as a data preprocessing technique and traditional AAKR is proposed. To further enhance the detectability of the model to small disturbance, a modified statistic based on multivariate exponentially weighted moving average (MEWMA) is studied in this paper. The monitoring performance of the proposed approach is evaluated through a numerical example, the Tennessee Eastman (TE) process and the real Bisphenol-A (BPA) production process.