Predictive Maintenance Vibration Telemetry Inference API Specification

Role: Lead Industrial MLOps and Backend Infrastructure Engineer with deep expertise in time-series inference systems.

Context

• Monitored assets: {{monitored_asset_types}}
• Raw vibration frequency: {{sensor_sampling_rate_khz}} kHz sampling per axis
• Maximum allowable inference latency: {{inference_sla_ms}} ms
• Real-time stream processor: {{stream_processing_framework}}
• Cold archive storage: {{cold_storage_destination}}
• Edge gateway deployment model: {{edge_gateway_spec}}

Task

Author a high-performance backend microservice specification for real-time vibration feature extraction and model inference, enabling immediate anomaly detection and automated work order triggering across {{monitored_asset_types}}.

Method

1. Define the digital signal processing (DSP) worker specifications for Fast Fourier Transform (FFT) and spectral kurtosis calculation.
2. Architect the tumbling and sliding window buffer mechanics within {{stream_processing_framework}}.
3. Design gRPC inference endpoints connecting feature pipelines to serialized machine learning runtimes (e.g., ONNX, Triton).
4. Specify the dynamic thresholding engine that adapts baseline vibrational signatures by operational state.
5. Structure the alert dispatch broker to trigger notifications when anomalies exceed confidence thresholds.
6. Formulate memory-efficient batch ingestion pipelines to offload high-resolution raw signals to {{cold_storage_destination}}.
7. Detail health checking, circuit breaking, and degraded-mode fallback when inference exceeds {{inference_sla_ms}} ms.
8. Model operational metrics, Prometheus scrapers, and OpenTelemetry distributed tracing spans.

Constraints

• MUST guarantee end-to-end telemetry-to-inference latency below {{inference_sla_ms}} ms.
• MUST NOT process unindexed FFT windows larger than available edge worker RAM allocated under {{edge_gateway_spec}}.
• Raw vibration arrays MUST be validated for sensor drift or clipping prior to running feature transforms.
• Failed model inference requests MUST fall back to deterministic statistical thresholding without dropping the stream.

Output format

1. Microservice Topology & Signal Flow (component diagram + data movement description)
2. DSP & Feature Transformation Contract (mathematical transforms, sliding window definitions, buffer size)
3. gRPC Service & Proto Interface (full Protocol Buffer definitions for inference requests and responses)
4. Fallback, Thresholding & Degradation Spec (circuit breaker thresholds and statistical fallback algorithms)
5. Observability & Telemetry Instrumentation (Prometheus metric names, alert rules, and tracing tags)

Self-review

• Does the windowing calculation fit within memory limits for {{sensor_sampling_rate_khz}} kHz sampling rate?
• Is the fallback strategy clearly executable when inference latency violates the {{inference_sla_ms}} ms SLA?
• Are asset-specific quirks for {{monitored_asset_types}} handled in the signal processing contract?