Streaming time-series classification / anomaly detection (UC-1 MVP)

The first fully scoped production use case: a continuous numeric stream (vibration, machine telemetry, grid/sensor channels) delivered as windows, classified or scored for anomalies per window, on CPU with no neuromorphic hardware. This page documents the MVP (P0–P4) delivered by spikeforge/streaming/ against the design in plans/use_case_streaming_timeseries.md.

What is implemented

Phase	Deliverable	Where
P0	Frozen windowing + z-score contract	window_spec.py
P0	Window → spike encode (delta primary, rate fallback)	encoding.py
P1	Deterministic synthetic stream + labels + splits	stream_source.py
P2	sequence_mlp CPU training recipe	recipe.py
P3	Class metrics + one-class anomaly score + threshold rule	recipe.py
P4	.spkf bundle, session, and serving parity	serving.py

Deferred (not MVP): P5 (a /metrics latency gate and serving benchmark) and P6 (container, promotion/rollback, drift monitor). The serving endpoints themselves are PT-W3's spikeforge-serve; UC-1 reuses them.

P0 — windowing and the frozen encode contract

A stream [N, D] becomes fixed-shape windows with a frozen WindowSpec that pins the window length, stride, channel order, and the train-fitted per-channel mean/std. Validation and test windows are normalised with the train statistics, and the same spec is stored in the bundle's preprocessing.json, so a served window can never be normalised differently from a training one.

Windows are encoded by the one encode_windows call site, which resolves the shared EncodeSpec and dispatches through the core SpikeEncoder, reshaping its flat code to [T, B, L, D]. delta coding is deterministic and is the primary contract; rate is the fallback. The optional delta_over_window transform differences the window along its time axis first, so the temporal change — not the absolute level — is what the spike train carries.

P1 — deterministic synthetic dataset

generate_stream stitches class-labelled segments of smooth multi-channel waveforms, injects transient and level-shift anomalies, and windows the stream. The class of segment i is i % classes, labels are a window's centre class, and the anomaly flag is raised when any injected anomaly lands inside the window. Train, validation, and test use independent seeds and the z-score statistics are fitted on train only, so no test information leaks. Everything is seeded: the same StreamSpec reproduces the same windows and labels.

P2/P3 — training, evaluation, anomaly head

The recipe trains the NIR-mappable sequence_mlp trunk on windowed spikes with the shared run temporal loop and surrogate-gradient cross-entropy — CPU only, deterministic, seconds to run. Evaluation reports class accuracy, macro-F1, balanced accuracy, per-class recall, and the additive one-class anomaly score 1 - max softmax whose threshold is the percentile-th quantile of the train score distribution (a rule, not a per-test fit). AUROC is the tie-corrected Mann-Whitney statistic; no scikit-learn dependency is used.

P4 — bundle, serving, and parity

save_checkpoint records the topology spec, the encode spec, and the windowing contract; build_bundle freezes both into a .spkf. predict_batch runs the whole batch through run, while stream_logits drives a stateful InferenceSession one timestep at a time. Because both paths share the same per-step body, the streaming readout equals the closed-loop batch reference exactly (max_abs_diff == 0.0 in the example) — parity by construction, not by tolerance. spikeforge-serve drives that same session, so serve_window reproduces the batch readout too, and reset clears the carried state.

Running it

venv/bin/python examples/11_streaming_timeseries.py

Representative output (the synthetic task is small; numbers move with the config):

windows train/val/test: 95 95 95
window shape: (95, 16, 4)
train anomalies: 34
trained: epochs=25 loss=0.7301 train_accuracy=0.842
test: accuracy=0.884 macro_f1=0.884 balanced_accuracy=0.885
anomaly: auroc=0.680 threshold=0.6240 precision=0.000 recall=0.000
bundle: uc1.spkf coding=delta window_L=16
parity: max_abs_diff=0.00e+00 within_tolerance=True
serve: spikeforge-serve not installed; skipped

Honesty and limits

• This is a runnable MVP slice, not a benchmark. The synthetic dataset is the CI fixture; the plan's public-dataset adapter and the spikeforge-io windowing distribution (W7) are future work, so the helpers live in the core spikeforge.streaming package for now.
• The plan's MVP floor is dataset-specific; the tests assert the metrics are sane (accuracy above chance, AUROC ≥ 0.5) rather than a tuned target.
• The anomaly head is the additive 1 - max softmax score reusing the classifier trunk, exactly as the plan allows; no separate one-class network is trained.
• Only InferenceSession + spikeforge-serve are exercised for P4; P5's Prometheus/OTel metrics and CI latency gate and P6's container/rollback/drift monitor are out of scope.
• spikeforge-serve is a separate distribution; the example degrades honestly when it is not installed.