FD001 Results: +2.3% Improvement

Learning Objectives

By the end of this section, you will:

1. Understand FD001 as the simplest C-MAPSS benchmark dataset
2. Analyze per-seed performance across 5 random seeds
3. Interpret the +2.3% improvement over DKAMFormer
4. Understand seed variance and its implications
5. Recognize best-case performance of 8.69 RMSE

Key Result: On FD001, AMNL achieves 10.43 ± 1.94 RMSE, a +2.3% improvement over DKAMFormer (10.68) and +9.2% over published SOTA (11.49). While this improvement is modest compared to complex datasets, it demonstrates AMNL's competitiveness even on the simplest benchmark.

FD001 Dataset Characteristics

FD001 is the simplest of the four NASA C-MAPSS datasets, designed for controlled evaluation of RUL prediction algorithms.

Dataset Configuration

Why FD001 is the Simplest

With only one operating condition and one fault mode, FD001 presents the most controlled environment for RUL prediction:

• No condition variance: All engines operate at sea level, eliminating the need for condition-invariant features
• Single degradation pattern: High Pressure Compressor (HPC) degradation follows a consistent trajectory
• Baseline benchmark: Algorithms should perform well here before tackling complex datasets

Experimental Setup

We evaluate AMNL with equal task weighting (0.5/0.5) across 5 random seeds for statistical robustness.

Configuration

Baseline Comparisons

Per-Seed Results

Complete results across all 5 random seeds reveal performance variance and best-case potential.

Comprehensive Per-Seed Data

Statistical Summary

Performance Analysis

Understanding why FD001 shows moderate improvement with high variance.

Improvement Summary

High Variance Analysis

The variance in FD001 results is notable:

• Coefficient of Variation: $\sigma / \mu = 1.94 / 10.43 = 18.6\%$
• Range: 13.56 - 8.69 = 4.87 RMSE (47% of mean)
• Outlier Impact: Seed 456 alone contributes 27% of total variance

Statistical Significance

Why FD001 Shows Modest Improvement

Several factors explain why AMNL's improvement is smaller on FD001 compared to complex datasets:

1. Limited multi-task benefit: With single operating condition, the health classification task provides less complementary signal
2. Already saturated: Simple datasets are closer to theoretical limits—less room for improvement
3. DKAMFormer optimization: Previous methods were heavily tuned for FD001 as the primary benchmark
4. Condition-invariance unnecessary: AMNL's strength in learning condition-invariant features is not leveraged

NASA Score Trade-off

AMNL achieves better RMSE but higher NASA Score compared to DKAMFormer:

The higher NASA Score suggests AMNL makes more late predictions (which are penalized exponentially). This trade-off indicates the model prioritizes RMSE accuracy over conservative early predictions.

Summary

FD001 Results Summary:

1. Mean RMSE: 10.43 ± 1.94 (across 5 seeds)
2. Improvement: +2.3% vs DKAMFormer, +9.2% vs SOTA
3. Best single result: 8.69 RMSE (seed 123, +24.4% vs SOTA)
4. Statistical significance: p = 0.1439 (not significant)
5. Variance: High (CV = 18.6%) due to seed 456 outlier

Key Insight: FD001 represents the "easy" case where most methods perform reasonably well. AMNL's true strength emerges on complex multi-condition datasets. The next section examines FD002, where AMNL achieves +37.0% improvement—the largest gain in our evaluation.

FD001 establishes baseline competitiveness. Next, we see how AMNL excels on complex datasets.