Bambang Riyanto Trilaksono

Deep learning models for bearing fault diagnosis on the Case Western Reserve University (CWRU) dataset routinely achieve near-perfect accuracy. Yet their decisions remain largely opaque at the signal level, so engineers cannot determine where in the raw vibration waveform the model focuses. This study aims to bridge the interpretability gap in deep learning-based bearing fault diagnosis by developing a signal-level explainability framework for the Wide Deep Convolutional Neural Network (WDCNN) on the Case Western Reserve University (CWRU) dataset. SHAP DeepExplainer is applied directly to raw 2,048-point vibration segments to produce per-sample-point attribution maps; Fault Signature Maps (FSMs) are formalized as class-averaged SHAP fingerprints in three variants (signed, absolute, and variance) and validated via discriminability index, severity monotonicity, and split-half stability, complemented by a three-variant ablation study examining how architectural decisions affect both accuracy and explainability. On the 10-class CWRU dataset under a rigorous temporal split (1,240/310/750 samples), WDCNN achieves 99.87% test accuracy with a macro F1-score of 0.997. FSMs demonstrate high reproducibility (split-half stability = 0.940); severity monotonicity ranges from 8.6% to 17.6%; and ablation reveals that removing Batch Normalization increases FSM discriminability by 2.3× at a 3.6% accuracy cost. As the first work applying SHAP DeepExplainer at full 2,048-point raw signal resolution for WDCNN, this study establishes that fault discrimination relies on transient impulse morphology rather than bearing characteristic frequencies, a finding invisible to feature-level XAI, and introduces a multi-resolution diagnostic framework bridging deep learning accuracy with physically interpretable vibration analysis for trustworthy deployment in safety-critical industrial environments.

This study builds an edge AI pipeline for classifying soil nutrient status in shallot cultivation using multivariate IoT sensor data. Data were collected in Tajur, Bogor, Indonesia, from 300 polybags between October 2 and December 22, 2025, covering 10 input features grouped into nitrogen (269,080 samples), phosphorus (286,232 samples), and potassium (338,242 samples), each at low, medium, or high levels. The pipeline includes teacher models MLP, TabTransformer, TabNet knowledge distillation to student models, ONNX conversion, and TensorRT deployment on NVIDIA Jetson Orin NX with FP16, FP32, and INT8 evaluations. Classic tabular models (XGBoost, LightGBM, CatBoost) were benchmarked on servers and edge devices. The transformer-based student model achieved the highest accuracy among neural models, while MLP was most efficient in inference. TabNet excelled for nitrogen (0.99923); TabTransformer consistently performed for phosphorus (0.99884–0.99886) and potassium (0.99746). On Jetson Orin NX, MLP's latency was 0.09354 ms and throughput 7,974.52 samples/second. Boosting models, especially LightGBM, remained strongly competitive across test scenarios. These results confirm the viability of the distillation–ONNX–TensorRT pipeline on edge devices; final model selection should consider trade-offs among accuracy, latency, throughput, and size.

This research proposes an approach to trajectory prediction using the Interacting Multiple Model Kalman Filter (IMM-KF), which forecasts the future positions of multiple sur rounding objects and determines their crossing status - whether they will cross or stay on their paths. The method accommodates diverse object behaviors by integrating constant velocity and constant acceleration models. Building on previous work, this approach enhances the accuracy of trajectory prediction for maneuvering objects, enhancing decision-making capabilities for autonomous trams to improve safety. The proposed method, which integrates a trajectory motion model with the IMM-KF algorithm, has been validated through both simulations and real-world testing on a vehicle platform. The results demonstrate significant improvements in safety and risk assessment.

Formation control has become a popular research topic in recent years. A common challenge in formation control is ensuring that robots can avoid obstacles and maintain a safe distance from one another to prevent collisions while forming a formation. In this research, a distributed formation control approach for a multi-robot system (MRS) with obstacle and collision avoidance is presented. The distributed formation control architecture is based on a consensus algorithm and consists of four layers: consensus tracking, consensus-based formation control, behavior, and physical robot layers. The system was implemented and evaluated through both simulations and experiments. Humanoid robots were used as the platform for these implementations. The result of the simulations and experiments show that the distributed formation control system successfully guided the robots into desired formation while also avoiding obstacles and preventing collisions with other robots.