IRVINE supports automotive engineers in detecting and understanding previously unknown errors in manufactured electrical engines by analyzing acoustic correlation signatures. Engineers explore clustered engine signatures, drill down to individual engines, analyze primary and secondary error symptoms through correlation matrices (Hypermatrices) and raw spectrograms, and externalize knowledge through labeling and annotation.
The input consists of acoustic measurements from engine test benches: for each engine, a 3D data structure of loudness values across all combinations of rotational speed (rpm) and frequency orders. First, residual spectrograms are computed by expressing each engine's values as deviations from the ensemble mean in units of standard deviation. Next, acoustic signatures (Hypermatrices) are computed: for each pair of relevant orders (41 orders selected by experts, mapped to 7 sub-components), the outer product of their residual curves is formed, the ensemble mean correlation is subtracted, and the result is reduced via the Frobenius norm. This yields a 41×41 symmetric matrix per engine capturing inter-component correlation deviations. From each Hypermatrix, a 28-dimensional feature vector is extracted (maximum value from each of 28 sub-component region combinations) for clustering.
A Self-Organizing Map is trained on the 28D feature vectors with configurable parameters (grid size, batch size, learning rate). The result is a grid of clusters where each cell aggregates engines with similar signatures. The Cluster View displays mean Hypermatrices per cell with glyphs indicating engine count, labeling progress (pie chart), and anomaly score (red/blue deviation arcs). This gives engineers an immediate structured overview of all engine types.
The engineer explores the SOM grid to identify clusters with interesting signatures (strong correlations, high anomaly scores, unlabeled engines). They select a cluster and drill down to inspect individual engines in the Engine List View (small-multiple Hypermatrices sorted by anomaly). They select an engine of interest and perform detailed analysis: inspecting the full Hypermatrix to identify primary and secondary error symptoms, examining the raw spectrogram to validate hypotheses, and reviewing supporting line charts, scatter plots, and bar charts for specific order pairs. Based on this analysis, they assess whether the engine contains an error. If so, they assign a categorical label (e.g., "B-Bearing/Inner ring") and annotate the error region directly in the spectrogram with a rectangle selection. They then proceed to the next engine or cluster. If the current clustering granularity is insufficient, they retrain the SOM on a subset of clusters or labels with adjusted parameters, obtaining a finer-grained view. This cycle repeats until the engineer has labeled all engines of interest.
All labels and annotations are stored in a knowledge database, accessible to other engineers and stakeholders. The accumulated knowledge supports future analyses (guiding hypothesis formation), enables semi-automatic error detection (threshold-based suggestions from stored annotations), and provides training data for potential machine learning classifiers.