Feature selection (FS) can improve prediction, while searching many combinations of selectors, subset sizes, and classifiers adds substantial cost. We predict the dataset-level AUC gain from the best combination in this FS-enabled portfolio over the best all-feature classifier. Among 99 datasets in the prediction sample, 51 show any positive gain and five gain at least 0.04 absolute AUC points. A class-weighted Random Forest using five basic descriptors and 21 Problexity measures achieves leave-one-dataset-out AUROC of 0.722 for any positive gain and 0.708 for gains of at least 0.04. Matched random-subset searches and outer-fold tests provide complementary evidence on search utility. SHAP analysis identifies neighborhood structure, feature overlap, and dataset size as influential predictors. These results support dataset-level prioritization of FS portfolio search and identify directions for strengthening its predictive precision.
We present a bearing-fault diagnostic pipeline that combines a 21-feature sensor-to-text representation, retrieval-augmented generation (RAG), and low-rank adaptation (LoRA) of LLaMA3.23B to produce structured reports containing fault class, severity, an explanation that cites signal features, and a maintenance recommendation. On a file-level CWRU split, the LLM achieves 78.45% macro-F1, compared with 93.15% for SVM and 91.21% for Random Forest. In a strict cross-load experiment, LoRA training, validation, and the retrieval index use only 0-1HP recordings and evaluation uses the complete 2-3HP set of 2308 windows. The same adapter, trained with three retrieved cases per prompt, is evaluated with and without retrieval at inference; the comparison does not involve independently trained models. For seed 42, retrieval raises the LLM macro-F1 from 19.91% to 67.89%; across five further training seeds the mean rises from 45.38±15.25% to 66.01±8.19%, with retrieval better in every seed (paired t-test p=0.015). SVM, Random Forest, and a compact raw-signal InceptionTime network (five seeds) remain more accurate than the LLM, at 87.48%, 89.42%, and 84.04±6.29%. In single-seed checks, macro-F1 varied little for retrieval depths k=1-7 and across three sentence encoders, whereas the contribution of operating-metadata and expected-frequency text was not established. On Paderborn, the fixed CWRU-trained model reaches 21.02% macro-F1, whereas target-domain LoRA fine-tuning with a target-specific FAISS index yields 77.33%, below Random Forest (80.39%); this is target-domain adaptation, not cross-dataset generalization, and the Paderborn split is recording-disjoint but not bearing-disjoint. Readability indices characterize linguistic complexity only and are not evidence of human interpretability.
Transformer-based remaining useful life (RUL) models apply the same feed-forward transformation to all timestep representations, despite variation across engines, operating conditions, and degradation stages. We propose Flexible-K Mixture of Experts (FK-MoE), which replaces this sublayer with Segment Gating with One-Stage Dispatch (SG-OSD). SG-OSD constructs expert-specific slot queries from consecutive timestep groups, performs global token-to-slot routing with unnormalised ReLU affinities, and reuses the resulting contributions for dispatch and output combination. This allows each token to receive contributions from a variable number of experts without fixed top-k routing. Across four C-MAPSS and seven N-CMAPSS subsets, routing analysis confirms input-dependent expert participation. In paired C-MAPSS experiments, segment conditioning reduces mean RMSE by 1.71 on FD001 and 1.68 on FD004 relative to a non-segment-conditioned Soft MoE, while the complete model achieves lower RMSE than a matched token-level ReLU-routing comparator on all four subsets. These results show that segment-conditioned Flexible-K routing can improve RUL prediction, although the current implementation adapts which experts contribute to each token while still evaluating all expert networks.
The integration of autonomous aerial drones with traditional ground vehicle fleets offers substantial potential for multi-echelon last-mile delivery networks. However, coordinating multi-modal fleets under operational constraints presents an underlying challenge: bridging discrete routing decisions with the continuous spatio-temporal progression of physical transport. While Deep Reinforcement Learning (DRL) provides an adaptive paradigm for dynamic dispatching, evaluating learned policies requires simulation environments that accurately synchronize discrete allocation logic with the continuous operational state progression of logistical entities. To address this need, this paper presents a Python-native, modular simulation and reinforcement learning environment specifically formulated for coordinated truck-drone delivery systems with intermediate micro-hubs. The core contributions are threefold: (1) a hybrid modeling framework that couples discrete event-based dispatching with the continuous spatio-temporal operational simulation of logistical entities; (2) an extensible, event-driven constraint manager that dynamically enforces defined logistical feasibility, including vehicle payload limits, spatial operating boundaries, and inter-vehicle synchronization, while constraining the combinatorial action space through invalid-action masking and a structured node-pair representation; and (3) an empirical evaluation benchmarking a Maskable Proximal Policy Optimization (PPO) agent against classical dispatch heuristics and unmasked learning configurations across varying operational scales. By offering a standardized and reproducible benchmarking platform, this work provides a practical simulation infrastructure for developing and evaluating constraint-aware reinforcement learning algorithms for operational planning in synchronized multi-modal freight logistics.
