ai-anomaly-detection-vibration

:::warning [UNTESTED] — v0.5 paper-correct Compiles on native_sim/native/64 + cross-compiles to ensemble_e8_dk/ae402fa0e5597le0/rtss_hp. HiL bring-up with a real motor / pump and a customer-trained model is v0.6 work. :::

Predictive-maintenance edge-AI on the E1M-AEN module: read a sliding window of vibration data from the EVK-populated ICM-42670 accelerometer, feed it through a 1D-CNN anomaly model via TFLite Micro, publish a per-window anomaly score.

Source: examples/ai/ai-anomaly-detection-vibration/.

Why this matters

Factory-floor vibration monitoring without a cloud round-trip — bearing-signature drift gets flagged before a motor / pump / spindle fails. The model learns the baseline at commissioning; departures from that baseline score above a threshold.

Window statistics

The per-window RMS and absolute peak come from one <alp/dsp.h> call rather than hand-written loops:

alp_dsp_stats_t st;
if (alp_dsp_stats_f32(win, n, &st) != ALP_OK || st.rms <= 0.0f) {
    /* ... */
}

alp_dsp_stats_f32() runs CMSIS-DSP arm_rms_f32 / arm_absmax_f32 on the M55 and portable C elsewhere. This example uses the stats surface only — it opens no FFT chain.

board.yaml

libraries:
  - name: tflite-micro
    cores: [m55_hp]
  - name: cmsis-dsp
    cores: [m55_hp]

som:
  sku: E1M-AEN801

preset: e1m-evk
supported_boards:
  - e1m-evk
  - e1m-x-evk

pins:
  - { e1m: E1M_I2C0, macro: EVK_I2C_BUS_SENSORS, doc: "Shared sensor + IO-expander + INA236 bus" }

cores:
  a32_cluster:
    os: "off"
  m55_hp:
    app: ./src
    inference:
      default_arena_kib: 128
    peripherals:
      - i2c

chips:
  - icm42670

diagnostics:
  log_level: info

See also

• <alp/inference.h> · <alp/dsp.h>
• <alp/chips/icm42670.h>
• multimodal-fusion-pdm — fuses this vibration signal with current + temperature
• Examples overview