Hi, I’m Aaron.

Approach. Combines federated class-specific denoising autoencoders, reconstruction-based novelty detection, clustering, and selective classifier expansion. Compares local rehearsal, shared prototypes, and a combination of both.

Results. Evaluated on ultrasonic metal welding across five seeds. The dual-rehearsal variant achieves 88.9% mean accuracy on both old and new classes in the autonomous onboarding experiment. Results concern a controlled manufacturing study; prototype sharing and local rehearsal have different data-disclosure requirements.

Approach. Builds class-specific acceptance regions in a learned embedding, uses clustering to support labeling, and updates selected classifier layers while freezing most parameters.

Results. Reports 96% unknown-fault detection accuracy and up to 98% overall classification accuracy after a five-sample update in the single-new-class experiment. Only 9.2% of model parameters are trained during adaptation. The study assumes a consistent underlying process and sensor configuration.

Approach. An Extractor structures G-code parameters, a Reference agent grounds operating ranges in documentation, Python computes deviations, and a Judge diagnoses defects using the resulting evidence.

Results. The best configuration achieves 87.5% accuracy versus 59.5% for a single-LLM baseline on 200 G-code programs spanning two printers, two materials, and five classes. The reported results cover controlled program-level diagnosis.

Approach. Combines literature curation, structured relationship extraction, iterative evaluation, and supervised fine-tuning for bidirectional QoS–QoE prediction.

Results. Includes 1,026 relationship records and 8,107 derived chat instances. Evaluates continuous values and discrete labels, retaining source evidence and context for each relationship.

Approach. Uses compact trajectory basis functions, image features, process parameters, and similarity-informed training-set selection.

Results. Validated on cookie and apple drying. For apples, the best multimodal configuration reaches RMSE 1.2865 versus 10.1517 for the reported baseline. Here, zero-shot refers to unseen combinations of drying conditions.

Approach. Pairs 16 G-code programs and four printing classes with 30 free-response and 75 multiple-choice questions at three expertise levels. Fourteen domain experts review the evaluation material.

Results. Compares GPT-4o, Claude 3.5 Sonnet, and Llama models to expose gaps between general language competence and manufacturing-specific reasoning. The benchmark and evaluation code are public.

Approach. Converts high-rate sensor signals to spectrogram sequences and uses video compression behavior to identify changes in temporal regularity.

Results. Evaluated on real laboratory vacuum pumps and controlled failure sequences. The study detects eight real pump anomaly events and reduces the size of transmitted data by over 90% in the evaluated setup.

Approach. Combines a shared base model, task-specific upper layers, group-wise aggregation, and regularization. Includes an edge–cloud implementation with Raspberry Pi clients and RabbitMQ.

Results. Reports gains of 5.35–8.08 percentage points over the compared federated baselines on unseen-domain tasks. FTL-TP is the earlier preprint of this work.

Approach. Uses an audio ADC for high-rate acquisition, converts multimodal signals to image-like inputs, and combines CNN features with signal features for classification.

Results. The evaluated system costs as little as $120. The hybrid pipeline averages approximately 385 ms per processing cycle on a Raspberry Pi 4; performance is evaluated across tool and surface-condition tasks.

Approach. Transforms hand coordinates to remove gross movement, normalizes geometric features, and compares support vector machines with convolutional classifiers.

Results. Studies eight activities in eight healthy participants with five-fold cross-validation. The best reported activity-detection CNN reaches 99.88% accuracy and 97.3% macro recall; patient evaluation remains future work.

Approach. Uses genetic search over sensors and signals, time/frequency/wavelet features, dimensionality reduction, and several classifiers. Healthy-participant data support the initial design.

Results. Evaluates 34 activities using nine healthy participants and two participants with Parkinson’s disease. The small patient cohort supports a preliminary adaptation study.

Approach. Uses genetic optimization over wearable IMU configurations and engineered signal features with a nearest-mean classifier.

Results. Evaluates 34 activities in nine healthy participants. The work establishes the sensor-selection and adaptation approach developed further in the 2022 journal article.