[KongchangAI]
Model动态传感器关系图神经网络

DySR-GNN

A dynamic sensor-relationship graph neural network model for multi-label root cause analysis in automotive hardware-in-the-loop testing. Through a shared temporal encoder and trainable sensor identity representations, it reconstructs a sparse relationship graph for each input window and performs graph message passing to achieve root cause localization.

Timeline (last 90 days)

Nov 13

DySR-GNN is a supervised multi-label root cause identification method that uses dynamic sensor relation graph neural networks.

Unverified50%
Nov 13

Under the recording-disjoint evaluation protocol, DySR-GNN achieves a recording-level Macro-F1 of 0.9524, an MCC of 0.9302, and a strict exact-match accuracy of 0.8519.

Unverified50%
Nov 13

Under the recording-disjoint evaluation protocol, DySR-GNN is compared with CNN-LSTM, static GCN, static GAT, and context-based variants, and achieves leading results on multiple metrics.

Unverified50%
Nov 13

DySR-GNN contains two key modules: a shared temporal encoder (which independently extracts temporal patterns for each sensor signal) and a trainable sensor identity representation (which, after being combined with each sensor's temporal features, reconstructs a sparse relation graph for each input window).

Unverified50%
Nov 13

The relation graph that DySR-GNN reconstructs for each input window is generated in a data-driven manner rather than being a static prior, and is then used for graph message passing and direct root cause location prediction.

Unverified50%
Nov 13

Static GCN or GAT uses a fixed adjacency matrix and cannot reflect relation changes when operating conditions switch; in contrast, the dynamic graph structure of DySR-GNN can capture state-dependent sensor correlations.

Unverified50%
Nov 13

The research team analyzed the relation graphs learned by DySR-GNN and found that sensor variables exhibit both persistently stable correlations and relations that vary with state.

Unverified50%
Nov 13

The value of DySR-GNN may extend to any system involving multivariate time series in which faults propagate between variables, such as in-vehicle networks, powertrain monitoring, and industrial control systems.

Unverified50%
Nov 13

The strict exact-match metric requires that the set of root causes predicted by the model be exactly identical to the ground-truth set, with any over-reporting or under-reporting judged as wrong; 0.8519 means that about 85% of recorded-segment root cause combinations were fully hit.

Unverified50%

All Facts (9)

Unverified

The strict exact-match metric requires that the set of root causes predicted by the model be exactly identical to the ground-truth set, with any over-reporting or under-reporting judged as wrong; 0.8519 means that about 85% of recorded-segment root cause combinations were fully hit.

50%
Unverified

Under the recording-disjoint evaluation protocol, DySR-GNN achieves a recording-level Macro-F1 of 0.9524, an MCC of 0.9302, and a strict exact-match accuracy of 0.8519.

50%
Unverified

Under the recording-disjoint evaluation protocol, DySR-GNN is compared with CNN-LSTM, static GCN, static GAT, and context-based variants, and achieves leading results on multiple metrics.

50%
Unverified

DySR-GNN contains two key modules: a shared temporal encoder (which independently extracts temporal patterns for each sensor signal) and a trainable sensor identity representation (which, after being combined with each sensor's temporal features, reconstructs a sparse relation graph for each input window).

50%
Unverified

The relation graph that DySR-GNN reconstructs for each input window is generated in a data-driven manner rather than being a static prior, and is then used for graph message passing and direct root cause location prediction.

50%
Unverified

Static GCN or GAT uses a fixed adjacency matrix and cannot reflect relation changes when operating conditions switch; in contrast, the dynamic graph structure of DySR-GNN can capture state-dependent sensor correlations.

50%
Unverified

The research team analyzed the relation graphs learned by DySR-GNN and found that sensor variables exhibit both persistently stable correlations and relations that vary with state.

50%
Unverified

DySR-GNN is a supervised multi-label root cause identification method that uses dynamic sensor relation graph neural networks.

50%
Unverified

The value of DySR-GNN may extend to any system involving multivariate time series in which faults propagate between variables, such as in-vehicle networks, powertrain monitoring, and industrial control systems.

50%

Source Articles