Add Mamba SSM module for time series generation

[Jameswlepage]

Summary

This PR adds State Space Model (SSM) based generation as an alternative to diffusion models for time series synthesis.

Motivation: Diffusion models tend to average out patterns and struggle with:

• Long-range trends (degradation signals)
• High-frequency oscillations (vibration)
• Periodicity (pressure cycles)

SSMs address this through continuous-time dynamics and selective state transitions.

Key Additions

• SSM/s4_layer.py: S4 layer with HiPPO initialization for long-range dependencies
• SSM/mamba_tsg.py: Mamba selective SSM implementation
  • SelectiveSSM: Core Mamba with input-dependent B, C, Δ parameters
  • MambaVAE: VAE architecture with Mamba encoder/decoder
  • MambaTimeSeriesGenerator: Autoregressive generator
• SSM/train_mamba_oilfield.py: Training script example

Validation Results

Tested on oil field degradation pattern (should have negative trend):

Model	Avg Slope	% Negative Trend
Original Data	-0.334	100%
Diffusion (DDPM)	-0.001	60%
Mamba SSM	-0.252	100%

Mamba captures the degradation trend 250x better than diffusion.

Features

• Pure PyTorch implementation (no custom CUDA ops required)
• Works on CPU and Apple Silicon (MPS)
• Linear time complexity O(n)
• Based on Mamba paper and LS4

Bug Fix Included

• TimeDP/utils/init_utils.py: Fix undefined cfg_name variable when using --name flag

Test Plan

• Unit tests for SelectiveSSM, MambaBlock, MambaVAE
• Training convergence on synthetic data
• Pattern preservation validation (degradation trends)
• MPS (Metal) acceleration verified on Apple Silicon

Built with assistance from Claude Code

[Jameswlepage]

@microsoft-github-policy-service agree

[Jameswlepage]

Additional findings from the SSM optimization loop

• Using TimeDP metrics (MMD / KL), the best current SSM configurations achieve:

  • exchange:
    • MMD = 0.0029
    • KL = 0.0686
    • Improves over DiffTime on both metrics; KL remains above TimeGen (KL = 0.016).
  • solar:
    • MMD = 0.0299
    • KL = 0.1735
    • Beats TimeVAE on KL (0.314); still behind DiffTime and TimeGen on KL.

• MMD is consistently stronger than all baselines across datasets.

• KL divergence remains the primary gap to close to outperform TimeGen end-to-end.

• Sweeps were run on TimeDP exchange / solar splits with a deterministic seed = 42.

• Results are recorded locally at:
  SSM/benchmarks/kl_sweep_results.json