成果总览

Signal Temporal Logic (STL) is widely used for specifying complex time-dependent behaviors in cyber-physical systems (CPSs), particularly in safety-critical domains. However, fault diagnosis (FD) and fault tolerant control (FTC) under nested STL (NSTL) specifications remain challenging, especially for nonlinear systems. This paper proposes a collaborative design (CoD) framework that jointly integrates FD and FTC under NSTL constraints to enhance detection accuracy and ensure robust system performance. First, a fault detection observer is developed by constructing fault tolerant feasible sets that can predict whether ongoing system trajectories satisfy NSTL specifications. To address the feasibility issue in real-time control synthesis, we introduce the concept of fault tolerant control with recursive feasibility (FTCRF), enabling the controller to maintain constraint satisfaction and system stability even under faults. A model predictive control scheme guided by control barrier functions (CBFs) ensures safe trajectory tracking within specified bounds. Simulation studies on single integrator and unicycle models demonstrate the effectiveness of the proposed method in accurately detecting faults and maintaining task satisfaction under NSTL constraints. Note to Practitioners—This work is motivated by the need for robust fault management in safety- critical cyber-physical systems such as autonomous vehicles, smart manufacturing systems, and robotic healthcare platforms. In these environments, rapid and accurate fault diagnosis—alongside continued safe operation despite faults—is essential to prevent downtime and maintain system integrity. Our proposed framework provides a practical methodology to jointly design fault detection and control strategies, explicitly accounting for complex temporal task constraints specified using nested Signal Temporal Logic. The key contribution is the collaborative integration of a fault diagnosis observer with a fault tolerant controller based on recursive feasibility, ensuring that the system can continue operating safely even when faults occur. Moreover, the use of STL Trees (STLTs) provides a tractable means of reasoning over complex time- based specifications. Practitioners can adopt this framework to improve system reliability, reduce intervention costs, and ensure compliance with temporal task requirements in dynamic and uncertain environments.

We propose F-DiffWave, a fault-aware wavelet-structured diffusion framework for multi- source domain generalization in rotating machinery diagnosis. By integrating wavelet- domain generation, accelerated diffusion-based synthesis, and contrastive discriminative learning, the method addresses data imbalance and improves generalization under variable operating conditions.

We propose WPANet, an interpretable fault diagnosis framework that combines wavelet priors with algorithm unrolling. By embedding wavelet-guided reconstruction and an unrolled time-enhanced iterative soft-thresholding solver, the method improves physical consistency, robustness, and transparency under time-varying conditions.

We develop CIGNN, an interpretable causal invariant graph neural network for unseen- domain gear fault diagnosis. The framework combines structural causal modeling, causal disentanglement, cross-domain consistency learning, and intervention-based risk minimization to improve robustness and interpretability under distribution shift.

We propose an automated two-stage IGAPSO-based framework for high-precision control of twisted and coiled polymer artificial muscles under parameter variability. The method combines neural-network-assisted offline optimization with rapid online fine-tuning to reduce manual calibration while improving accuracy and robustness.

We propose Dual-BDSTN, a dual spatial-temporal neural network for bottleneck prediction in manufacturing systems. By decoupling temporal and spatial modeling, introducing dynamic graph learning based on material flow, and integrating self-attention and cross- attention mechanisms, the method improves blockage/starvation prediction and bottleneck localization.

We develop an adaptive ridge path regrouping framework for instantaneous frequency extraction in overlapped multicomponent signals. The method combines a variation- constrained time-frequency filter, entropy-guided parameter adaptation, and a refined penalty term to improve robustness and accuracy in complex overlapping regions.

We study control synthesis for uncertain systems subject to nested STL tasks and propose a decomposition-based MPC framework. The method resolves nested specifications into atomic subtasks and combines them with distributed shrinking-horizon MPC to improve tractability and control performance under bounded disturbances.

We introduce BCKG, an event-driven causal knowledge graph framework for bottleneck root cause analysis in manufacturing systems. By combining expert-initialized causal graphs, overlap-duration-based pruning, reverse depth-first search, and a modified PageRank strategy, the method identifies root causes and propagation paths with high accuracy.

We develop an improved transfer path analysis method for multistage gearbox fault tracing. The method combines structural dynamics decoupling, inverse bearing-force identification, and path contribution analysis to enhance fault characteristics and determine dominant propagation paths across different gearbox configurations.

We propose MK-DGNet, a prior-knowledge-guided deep graph network for in-situ quality inspection in laser powder bed fusion. The method fuses melt-pool image information with physically informed knowledge representations, improving defect-related quality monitoring while enhancing interpretability.

We propose a Sparse Temporal Logic Network for interpretable bearing fault diagnosis. The framework combines wavelet-based predicate extraction, sparse and structured neural attention, and temporal logic inference to deliver accurate diagnosis together with formal, human-readable explanations.

We propose BCFNet, a bi-stream cross-mode fusion network for super-elevation monitoring in laser powder bed fusion. By jointly analyzing image and acoustic signals across recoater and laser scenarios, the method improves monitoring accuracy, robustness, and interpretability for defect severity assessment.

We propose a collaborative design framework that jointly integrates fault diagnosis and fault-tolerant control under nested signal temporal logic specifications. The method combines a diagnosis observer with recursively feasible fault-tolerant MPC guided by control barrier functions to maintain task satisfaction under faults.

We propose the adaptive ridge path regrouping algorithm for overlapped multicomponent signals, combining variation-constrained time-frequency filtering, entropy-guided parameter adaptation, and a refined penalty term to improve instantaneous-frequency estimation accuracy and robustness.

We present an interpretable transfer learning framework for unsupervised bearing fault diagnosis from simulation to reality. Built on a Wavelet Temporal Logic Network encoder and a Temporal Logic Discrepancy Metric, the approach aligns structural knowledge across domains while preserving transparent decision logic.

We introduce AD-STLN, an interpretable bearing fault diagnosis framework that combines adaptive dictionary convolution, sparse attention, and weighted signal temporal logic reasoning to produce transparent STL-like diagnostic rules.

We propose a causal-aware neural network for wind turbine gearbox diagnosis that models causal mechanisms, disentangles causal and non-causal subgraphs, and uses causal intervention to improve interpretability and robustness under distribution shift.

We propose LTQU-Net, which integrates differentiable tunable Q-factor wavelet decomposition with deep unfolding to learn interpretable and invariant time-frequency features for robust cross-domain fault diagnosis.

This survey systematically reviews interpretable intelligent fault diagnosis for rotating machinery. It organizes the literature into post-hoc and ante-hoc paradigms, analyzes representative methods and their limitations, and highlights future directions for building more transparent and reliable diagnosis systems.

We develop a nonconvex robust PCA method for bearing fault diagnosis using a gamma-norm rank approximation. The method improves fault feature extraction in noisy vibration signals and outperforms conventional convex RPCA, VMD, and FMD baselines in both simulations and experiments.

This paper proposes an improved transfer path analysis framework for multistage gearbox fault tracing. By combining structural dynamics decoupling, bearing force identification, and dominant-path contribution analysis, the method enhances fault features and enables effective tracing of gear faults under different operating conditions.

We introduce VNCCD, a virtual nonlinear chirp component decoupling method for gearbox fault diagnosis under nonstationary conditions. The method first identifies virtual decoupled transfer paths and then extracts intrinsic nonlinear chirp components, improving vibration quality and strengthening diagnosis under complex structural coupling.

We propose KCFed, a knowledge-based clustering federated learning framework for robotic assembly fault diagnosis. The method addresses both few-shot diagnosis and Non-IID data distribution by clustering similar tasks and accumulating prior knowledge while preserving data privacy across clients.

We present a robotic assembly framework that combines task-attention multimodal fusion with curriculum residual reinforcement learning. The method improves context generalization, robustness to large random errors, and learning stability, achieving high-precision assembly in variable environments.

We propose an object-embodiment-centric imitation and residual reinforcement learning framework for robotic assembly in semi-structured environments. Using only a single demonstration and limited interaction, the method improves sample efficiency and significantly increases assembly success rate and speed.

We study control/physical co-design for LPV systems with spectral temporal logic specifications. By transforming spectral-temporal requirements into mixed-integer matrix inequality constraints, the paper develops an iterative optimization framework and demonstrates its effectiveness on MEMS applications.

We develop a preference-conditioned graph reinforcement learning framework for fast Pareto set approximation in multi-objective flexible job shop scheduling. The method parallelizes preference-based learning and substantially improves both solution quality and computational speed on large-scale scheduling problems.

We develop a neural-symbolic learning architecture for interpretable rolling-bearing diagnosis that combines weighted signal temporal logic, predicate extraction, autoencoding, and timed failure propagation graphs to produce accurate and explainable fault decisions.

We propose a cognitive manipulation framework for robotic assembly in semi-structured environments that integrates skill-graph-guided object detection, coarse operation planning, and classroom-to-real residual reinforcement learning for robust and efficient fine manipulation.

We propose a discrete-event framework for fault-tolerant supervisory control in automated manufacturing systems, using resource sharing and incremental abstraction to characterize the existence of fault-tolerant supervisors and synthesize them efficiently when feasible.

We propose a temporal logic neural network in which neurons act as logic propositions and the whole network can be interpreted as weighted signal temporal logic, enabling accurate and formally interpretable bearing fault diagnosis.

We introduce a human-computer collaboration framework for requirement mining from closed-loop control models, allowing designers and a computer agent to iteratively explore, refine, and optimize signal temporal logic specifications.

We propose a graph reinforcement learning framework with evolution strategies for dynamic job shop scheduling under machine breakdown and stochastic processing times. The approach formulates the scheduling problem as an MDP and learns robust policies that outperform reinforcement-learning-based and traditional baselines.

We propose a knowledge-informed learning and control framework for precise peg-in-hole assembly that combines a partial model, uncertainty-aware reward shaping, and residual reinforcement learning to improve adaptability, safety, and sample efficiency.

This paper presents a data-driven approach to construct spectral timed failure propagation graphs directly from data. By transforming TFPG construction into a spectral-temporal logic inference problem and solving it with a Tree-LSTM guided by supremal language, the method achieves interpretable fault diagnosis without requiring expert-designed models.

We introduce shapelet temporal logic, a formal language that describes temporal relations among discriminative shapelets in sequential data. An incremental inference algorithm with theoretical guarantees is developed to obtain interpretable fault diagnosis rules for rolling element bearing signals.

We formulate fault diagnoser construction as a formal language generation problem and introduce signal spectral logic for interpretable diagnosis in the frequency domain. Adversarial training is used to alleviate the sparse-reward issue and improve the robustness of generated logic formulas under noisy conditions.

We study control synthesis for energy-harvesting MEMS devices under load-based spectral logic specifications, derive satisfaction conditions as LMIs, and develop an iterative semidefinite-programming-based controller synthesis algorithm.

We propose algorithms for checking fault-tolerant properties and synthesizing time- optimal fault-tolerant controllable strings in multi-process systems by combining abstraction, resource sharing, and sequential backtracking.

We present a residual reinforcement learning framework for peg-in-hole assembly with inaccurate trajectories, using task information to balance prior policy and exploration, improve learning efficiency, and reduce collision risk.

We propose real-valued timed failure propagation graphs for continuous-state systems and a systematic refinement framework that combines expert knowledge with data-driven learning. Starting from a partial graph, the method automatically adds discrepancy nodes and causal edges from labeled signals for interpretable fault diagnosis.

We develop an interpretable fault detector for switched nonlinear systems with partially unknown dynamics. The method approximates unknown dynamics using Gaussian processes and introduces a safe temporal logic inference algorithm that guarantees no missed faults during the search while providing probabilistic satisfaction guarantees.

We formalize semantic fault diagnosis for IIoT systems as the task of learning signal temporal logic formulas directly from time-series data. To address combinatorial explosion, we propose an agenda-based, reinforcement-learning-enabled search strategy that discovers scalable and interpretable fault specifications from industrial sensor data.

We introduce a frequency-temporal logic (FTL) framework for bearing fault diagnosis and interpretation. The method combines Bayesian optimization with Bayesian neural networks to infer compact and interpretable logic formulas from vibration data, achieving competitive diagnostic performance while providing human-readable explanations of fault patterns.

We study lossless negative imaginary systems in both continuous-time and discrete-time settings. The paper establishes necessary and sufficient frequency-domain and state- space conditions, develops decomposition results, and derives new relationships between lossless positive real and lossless negative imaginary systems.

We propose a data-driven modeling approach for UAV-based autonomous dispensing of predatory mites in wind. Using outdoor experiments and machine learning, the method characterizes vermiculite distribution as a function of wind conditions and UAV operating parameters, enabling more accurate precision pest management.

We propose a formal interpretation framework that allows a human user to interrogate a cyber-physical system using temporal logic queries. The method formulates interpretation as a temporal logic inference problem and introduces a Gaussian-process-based active learning algorithm to obtain probably approximately correct solutions efficiently.

We develop a data-driven algorithm for constructing approximate abstractions of black- box piecewise affine systems under temporal logic specifications. By integrating system identification, abstraction refinement, and active sampling, the method derives abstractions with bounded error and probability guarantees even when the system dynamics are unknown.

We present a human-in-the-loop semantic parsing framework that automatically translates automobile steering behaviors into signal temporal logic specifications using agenda- based parsing and reinforcement learning.

We develop a data-driven algorithm to construct approximate abstractions for black-box piecewise affine systems by integrating system identification, abstraction, and active sampling under temporal-logic specifications with bounded error and probability guarantees.