What is the smallest anomaly detection model that runs on STM32F4?

A 2-layer autoencoder with 4-8 neurons per layer quantizes to 5-8 KB. This handles single-channel anomaly detection (e.g., vibration amplitude) with good sensitivity. Multi-channel input (vibration + temperature + current) needs 10-15 KB for adequate representation capacity.

How do I set the anomaly threshold on STM32F4?

Run the autoencoder on your validation set of normal data and record the reconstruction errors. Set the threshold at the 95th or 99th percentile — this determines the false positive rate. Lower thresholds catch more anomalies but trigger more false alerts. Tune based on your tolerance for false positives.

Can the STM32F4 run anomaly detection continuously for years?

Yes. The STM32F4 is rated for industrial temperature ranges (-40 to 85°C) with 10+ year production availability. Anomaly detection uses under 10% of the CPU capacity, leaving headroom for watchdog timers, error recovery, and communication. The STM32F4 is designed for long-term industrial deployment with ST's extended availability program.

Hardware Guide

STM32F4 for Anomaly Detection with TensorFlow Lite Micro

The STM32F4 runs autoencoder-based anomaly detection with TFLite Micro using under 20 KB of its 192 KB SRAM. The Cortex-M4F's DSP instructions and FPU handle feature extraction and inference with sub-millisecond latency — a widely used platform for industrial edge AI deployment.

Hardware Specs

Spec	STM32F4
Processor	ARM Cortex-M4F @ 168 MHz
SRAM	192 KB
Flash	1 MB
Key Features	Single-precision FPU, DSP instructions, Widely available ecosystem
Connectivity	USB OTG FS
Price Range	$3 - $10 (chip), $10 - $30 (dev board)

Compatibility: Good

Anomaly detection autoencoders are the lightest ML workload — 10-20 KB models with minimal compute requirements. The STM32F4's 192 KB SRAM provides 6x the 32 KB minimum, leaving 170+ KB for application logic, communication stacks, and sensor management. The Cortex-M4F's DSP instructions handle the small matrix operations in autoencoders efficiently, and CMSIS-NN kernels optimize the dense layers. Inference latency is sub-millisecond for typical anomaly models. The STM32F4 is a proven platform in industrial settings — ST provides 10+ year availability guarantees, and there are thousands of deployed industrial monitoring systems using STM32F4. The ecosystem of industrial communication libraries (Modbus, CAN, PROFINET) is mature. For anomaly detection specifically, the STM32F4 is more than sufficient — you would only need to step up to STM32H7 if combining anomaly detection with heavier workloads like image processing.

Getting Started

1

Configure sensor acquisition on STM32F4

Use STM32CubeMX to configure I2C for accelerometer, ADC for current/temperature sensors, or SPI for high-speed data acquisition. Set up DMA transfers for continuous sampling without CPU interruption.
2

Collect and preprocess training data

Log sensor readings during normal operation to a PC via UART. Collect at least 2000 normal samples. Normalize the data (zero-mean, unit-variance) and compute the normalization parameters — these must be embedded in the firmware for runtime preprocessing.
3

Train a compact autoencoder

Build an autoencoder in TensorFlow with 2-3 layers (input → 8 neurons → 4 neurons → 8 neurons → output). Train on normal data only. Apply int8 quantization — the resulting model should be 8-15 KB.
4

Deploy with threshold-based alerting

Embed the quantized model in firmware. Run inference on each sensor window. Calculate reconstruction error (MSE). Set the anomaly threshold at the 99th percentile of normal-data reconstruction errors. Trigger alerts via GPIO, UART, or CAN bus when errors exceed the threshold consistently.

Alternatives

ESP32 with TFLite Micro

Built-in Wi-Fi enables wireless anomaly reporting. 520 KB SRAM with dual-core for concurrent monitoring. Better for retrofitting existing machinery with wireless connectivity.

STM32L4 with TFLite Micro

Ultra-low-power for battery-operated or energy-harvesting deployments. 128 KB SRAM, 80 MHz. Trade clock speed for extreme power efficiency (< 100 nA shutdown).

Explore More

More STM32F4 guides More Anomaly Detection guides All resources Find the right MCU

FAQ

What is the smallest anomaly detection model that runs on STM32F4?: A 2-layer autoencoder with 4-8 neurons per layer quantizes to 5-8 KB. This handles single-channel anomaly detection (e.g., vibration amplitude) with good sensitivity. Multi-channel input (vibration + temperature + current) needs 10-15 KB for adequate representation capacity.
How do I set the anomaly threshold on STM32F4?: Run the autoencoder on your validation set of normal data and record the reconstruction errors. Set the threshold at the 95th or 99th percentile — this determines the false positive rate. Lower thresholds catch more anomalies but trigger more false alerts. Tune based on your tolerance for false positives.
Can the STM32F4 run anomaly detection continuously for years?: Yes. The STM32F4 is rated for industrial temperature ranges (-40 to 85°C) with 10+ year production availability. Anomaly detection uses under 10% of the CPU capacity, leaving headroom for watchdog timers, error recovery, and communication. The STM32F4 is designed for long-term industrial deployment with ST's extended availability program.