Practical case: ESP32 anomaly detection with I2S+MQTT

ESP32 Advanced Practical: I2S Anomaly Detection with MQTT, Audio Beep, and LED Feedback

Objective: i2s-anomaly-detection-mqtt
Device family and exact model used: Espressif ESP32-DevKitC V4 (ESP32-WROOM-32) + INMP441 + MAX98357A + WS2812B

This advanced hands-on guide implements real-time acoustic anomaly detection on an ESP32 using an I2S MEMS microphone (INMP441), with anomaly alerts routed over MQTT, visual status via a WS2812B LED, and an audible beep generated through an I2S DAC/amplifier (MAX98357A). The project is reproducible with PlatformIO, includes full source, pin mapping, build/flash commands, and a validation workflow.

Prerequisites

• Comfortable with:
• ESP32 and PlatformIO on VS Code or CLI.
• Basic DSP concepts (FFT, feature extraction, thresholds).
• MQTT brokers and topics.
• Soldering and breadboarding.
• OS with serial driver support:
• Windows: Silicon Labs CP210x USB-to-UART driver (for DevKitC’s CP2102/CP2104).
• macOS: Typically driverless for CP210x, but install from Silicon Labs if needed.
• Linux: Usually driverless; confirm device appears as /dev/ttyUSBx or /dev/tty.SLAB_USBtoUART.
• Installed software:
• PlatformIO Core (CLI) 6.x (tested with 6.1.11) or VS Code + PlatformIO extension.
• Mosquitto clients for validation (mosquitto_pub, mosquitto_sub).
• A working MQTT broker (e.g., Mosquitto on localhost or a network host).
• Solid 5V power if you plan to drive a speaker from MAX98357A (ESP32’s 5V pin can power small loads via USB, but consider a separate supply if needed).

Materials (Exact Models)

• Microcontroller: Espressif ESP32-DevKitC V4 (ESP32-WROOM-32)
• Digital MEMS Microphone (I2S): INMP441
• I2S DAC/Amplifier: MAX98357A (Adafruit MAX98357 I2S Class-D Mono Amp or equivalent breakout)
• Addressable RGB LED: WS2812B (single LED or short strip; you can cut just one pixel from a strip)
• Speaker: 4–8 Ω small speaker for MAX98357A output
• Breadboard + jumpers
• USB cable for ESP32 DevKitC
• Optional level shifting for WS2812B data signal (often works at 3.3 V for a single LED, but best practice is to buffer)

Driver note:
– If your Espressif ESP32-DevKitC V4 enumerates as an unknown device on Windows, install the Silicon Labs CP210x driver from https://www.silabs.com/developers/usb-to-uart-bridge-vcp-drivers.

Setup/Connection

We will use two I2S peripherals in the ESP32:

• I2S0 (RX) for INMP441 microphone capture.
• I2S1 (TX) for audio output to MAX98357A.

And a separate GPIO for WS2812B data.

Recommended pin mapping (tested, doesn’t conflict with flash/boot pins):

• INMP441 (I2S RX, 3.3V device):
• VDD → 3V3
• GND → GND
• L/R → GND (selects Left channel)
• SCK/BCLK → GPIO32
• WS/LRCL → GPIO33
• SD (DOUT) → GPIO34 (input-only pin)
• MAX98357A (I2S TX, powered from 5V):
• VIN → 5V
• GND → GND
• DIN → GPIO26
• BCLK → GPIO27
• LRC → GPIO25
• SD/GAIN pin per board variant (tie high to enable or leave floating as per module; see your board’s datasheet)
• Speaker → + and – outputs
• WS2812B:
• 5V → 5V
• GND → GND (must share common ground with ESP32)
• Data In → GPIO4
• Place a 330–470 Ω resistor in series with the data line if possible; add a 1000 µF capacitor across 5V/GND near the LED strip if driving more than one LED (for a single LED, optional).

Power notes:
– The ESP32 DevKitC can supply limited current from 5V via USB. The MAX98357A plus a small speaker at moderate volume is typically fine. If you experience brownouts, provide an external 5V supply and connect grounds.

I2S details:
– Microphone (INMP441) outputs 24-bit samples, left-justified in I2S frames. We’ll capture as 32-bit and shift to 16-bit for DSP.
– MAX98357A reads standard I2S stereo (we’ll feed identical samples for L/R or generate a mono stream in both channels).

Table: Wiring Summary

Full Code

The project is structured for PlatformIO (Arduino framework). It uses:

• I2S driver from ESP-IDF headers (via Arduino core) for both RX and TX.
• FFT-based features using arduinoFFT.
• MQTT via PubSubClient.
• LED status via Adafruit NeoPixel.

Create a new PlatformIO project and replace the files as shown below.

platformio.ini

Place this in the project root.

; File: platformio.ini
[env:esp32dev]
platform = espressif32 @ 6.5.0
board = esp32dev
framework = arduino
board_build.flash_mode = dio
; optional: larger app partition if you add more features
; board_build.partitions = huge_app.csv

monitor_speed = 115200
monitor_filters = time, esp32_exception_decoder, default

lib_deps =
  arduinoFFT@^1.6.0
  knolleary/PubSubClient@^2.8
  adafruit/Adafruit NeoPixel@^1.12.0

build_flags =
  -DCORE_DEBUG_LEVEL=3
  -DARDUINO_USB_MODE=1
  -DWS2812_PIN=4
  -DMIC_BCLK=32
  -DMIC_LRCL=33
  -DMIC_DOUT=34
  -DSPK_BCLK=27
  -DSPK_LRCL=25
  -DSPK_DIN=26
  -DMQTT_TOPIC_PREFIX=\"device/esp32-anom\"
  -DMQTT_QOS=1

src/main.cpp

// File: src/main.cpp
#include <Arduino.h>
#include <WiFi.h>
#include <PubSubClient.h>
#include <Adafruit_NeoPixel.h>
#include <driver/i2s.h>
#include <arduinoFFT.h>

// ---------------------- User Configuration ----------------------
static const char* WIFI_SSID = "YOUR_WIFI_SSID";
static const char* WIFI_PASS = "YOUR_WIFI_PASSWORD";

// MQTT broker settings
static const char* MQTT_HOST = "192.168.1.100";
static const uint16_t MQTT_PORT = 1883;
static const char* MQTT_USER = "";  // optional
static const char* MQTT_PASS = "";  // optional

// Topics
#ifndef MQTT_TOPIC_PREFIX
#define MQTT_TOPIC_PREFIX "device/esp32-anom"
#endif

// ---------------------- I2S Config ------------------------------
// Microphone (I2S0 RX)
#ifndef MIC_BCLK
#define MIC_BCLK 32
#endif
#ifndef MIC_LRCL
#define MIC_LRCL 33
#endif
#ifndef MIC_DOUT
#define MIC_DOUT 34
#endif

// Speaker (I2S1 TX)
#ifndef SPK_BCLK
#define SPK_BCLK 27
#endif
#ifndef SPK_LRCL
#define SPK_LRCL 25
#endif
#ifndef SPK_DIN
#define SPK_DIN 26
#endif

// ---------------------- LED Config ------------------------------
#ifndef WS2812_PIN
#define WS2812_PIN 4
#endif
#define WS2812_COUNT 1

// ---------------------- DSP/Anomaly Config ----------------------
static const uint32_t SAMPLE_RATE = 16000;    // 16 kHz
static const size_t FFT_N = 512;              // 32 ms frame
static const size_t HOP = FFT_N;              // no overlap for simplicity
static const uint8_t NUM_BANDS = 16;          // coarse spectral bands
static const size_t BASELINE_FRAMES = 300;    // ~9.6 s baseline (300 * 32ms)
static const float ANOMALY_THRESHOLD = 25.0f; // sum |z| across bands

// ---------------------- Globals --------------------------------
WiFiClient espClient;
PubSubClient mqtt(espClient);

Adafruit_NeoPixel pixel(WS2812_COUNT, WS2812_PIN, NEO_GRB + NEO_KHZ800);

// FFT arrays
arduinoFFT FFT = arduinoFFT();
double vReal[FFT_N];
double vImag[FFT_N];

float bandMeans[NUM_BANDS];
float bandStd[NUM_BANDS];
bool baselineDone = false;
size_t baselineCount = 0;

char deviceId[32];

// ---------------------- Utilities -------------------------------
void setLED(uint8_t r, uint8_t g, uint8_t b) {
  pixel.setPixelColor(0, pixel.Color(r, g, b));
  pixel.show();
}

void ledPulseRed() {
  static uint8_t phase = 0;
  uint8_t level = (uint8_t)(127 + 127 * sin(phase * 0.1));
  setLED(level, 0, 0);
  phase++;
}

void ledGreenSolid() {
  setLED(0, 64, 0);
}

void ledBlueBlink() {
  static bool on = false;
  setLED(0, 0, on ? 64 : 0);
  on = !on;
}

void beepAlert(uint16_t freq = 1000, uint16_t ms = 200, uint16_t volume = 2000) {
  // Simple sine wave beep on I2S1 TX
  i2s_config_t i2s_config = {
    .mode = (i2s_mode_t)(I2S_MODE_MASTER | I2S_MODE_TX),
    .sample_rate = (int)SAMPLE_RATE,
    .bits_per_sample = I2S_BITS_PER_SAMPLE_16BIT,
    .channel_format = I2S_CHANNEL_FMT_RIGHT_LEFT,
    .communication_format = I2S_COMM_FORMAT_I2S_MSB,
    .intr_alloc_flags = ESP_INTR_FLAG_LEVEL1,
    .dma_buf_count = 8,
    .dma_buf_len = 256,
    .use_apll = false,
    .tx_desc_auto_clear = true,
    .fixed_mclk = 0
  };

  i2s_pin_config_t pin_config = {
    .bck_io_num = SPK_BCLK,
    .ws_io_num = SPK_LRCL,
    .data_out_num = SPK_DIN,
    .data_in_num = I2S_PIN_NO_CHANGE
  };

  i2s_driver_install(I2S_NUM_1, &i2s_config, 0, NULL);
  i2s_set_pin(I2S_NUM_1, &pin_config);
  i2s_set_clk(I2S_NUM_1, SAMPLE_RATE, I2S_BITS_PER_SAMPLE_16BIT, I2S_CHANNEL_STEREO);

  const uint32_t total_samples = (SAMPLE_RATE * ms) / 1000;
  const float omega = 2.0f * PI * freq / SAMPLE_RATE;

  for (uint32_t n = 0; n < total_samples; n++) {
    int16_t s = (int16_t)(volume * sinf(omega * n));
    uint32_t stereo = ((uint16_t)s << 16) | ((uint16_t)s);
    size_t written;
    i2s_write(I2S_NUM_1, &stereo, sizeof(stereo), &written, portMAX_DELAY);
  }

  i2s_driver_uninstall(I2S_NUM_1);
}

// Compute 16 log-energy bands from FFT magnitude
void computeBands(const double* mag, float* bands, size_t nfft, uint32_t fs) {
  // Frequency resolution
  const double df = (double)fs / (double)nfft;
  // Use pseudo-mel spacing over 0..8kHz (simple linear-to-log mapping)
  // Band edges exponentially spaced
  double fmin = 100.0, fmax = fs / 2.0;
  double logmin = log10(fmin);
  double logmax = log10(fmax);
  for (int b = 0; b < NUM_BANDS; b++) {
    double e0 = pow(10.0, logmin + (logmax - logmin) * b / NUM_BANDS);
    double e1 = pow(10.0, logmin + (logmax - logmin) * (b + 1) / NUM_BANDS);
    int k0 = (int)floor(e0 / df);
    int k1 = (int)ceil(e1 / df);
    if (k0 < 1) k0 = 1; // skip DC
    if (k1 >= (int)(nfft / 2)) k1 = (nfft / 2) - 1;

    double sum = 0.0;
    for (int k = k0; k <= k1; k++) {
      sum += mag[k] * mag[k]; // power
    }
    bands[b] = (sum > 1e-12) ? (float)log10(sum) : -12.0f;
  }
}

float updateBaselineAndScore(const float* bands) {
  if (!baselineDone) {
    // Welford-like online mean/std (keep running stats; std computed post)
    static double m[NUM_BANDS] = {0};
    static double s[NUM_BANDS] = {0};
    baselineCount++;
    for (int b = 0; b < NUM_BANDS; b++) {
      double x = bands[b];
      double delta = x - m[b];
      m[b] += delta / baselineCount;
      s[b] += delta * (x - m[b]);
    }
    if (baselineCount >= BASELINE_FRAMES) {
      for (int b = 0; b < NUM_BANDS; b++) {
        bandMeans[b] = (float)m[b];
        double var = (baselineCount > 1) ? s[b] / (baselineCount - 1) : 1e-3;
        bandStd[b] = (float)sqrt(var + 1e-6);
      }
      baselineDone = true;
    }
    return 0.0f; // no scoring during baseline
  } else {
    float score = 0.0f;
    for (int b = 0; b < NUM_BANDS; b++) {
      float z = (bands[b] - bandMeans[b]) / (bandStd[b] + 1e-6f);
      score += fabsf(z);
    }
    return score;
  }
}

void publishMQTT(const String& topic, const String& payload, bool retain = false) {
  String full = String(MQTT_TOPIC_PREFIX) + "/" + topic;
  mqtt.publish(full.c_str(), payload.c_str(), retain);
}

void connectWiFi() {
  WiFi.mode(WIFI_STA);
  WiFi.begin(WIFI_SSID, WIFI_PASS);
  Serial.printf("[WiFi] Connecting to %s\n", WIFI_SSID);
  uint32_t start = millis();
  while (WiFi.status() != WL_CONNECTED) {
    delay(500);
    Serial.print(".");
    ledBlueBlink();
    if (millis() - start > 30000) {
      Serial.println("\n[WiFi] Timeout, restarting...");
      ESP.restart();
    }
  }
  Serial.printf("\n[WiFi] Connected, IP: %s\n", WiFi.localIP().toString().c_str());
}

void connectMQTT() {
  mqtt.setServer(MQTT_HOST, MQTT_PORT);
  mqtt.setKeepAlive(30);
  while (!mqtt.connected()) {
    String clientId = String("esp32-anom-") + String(deviceId);
    Serial.printf("[MQTT] Connecting to %s as %s ...\n", MQTT_HOST, clientId.c_str());
    if (mqtt.connect(clientId.c_str(), MQTT_USER, MQTT_PASS)) {
      Serial.println("[MQTT] Connected");
      publishMQTT("status", "{\"state\":\"online\"}", true);
      publishMQTT("info", String("{\"device\":\"") + deviceId + "\",\"sr\":" + SAMPLE_RATE + ",\"fft\":" + FFT_N + "}");
    } else {
      Serial.printf("[MQTT] failed rc=%d, retrying in 3s\n", mqtt.state());
      delay(3000);
    }
  }
}

void initI2S_RX() {
  i2s_config_t i2s_config = {
    .mode = (i2s_mode_t)(I2S_MODE_MASTER | I2S_MODE_RX),
    .sample_rate = (int)SAMPLE_RATE,
    .bits_per_sample = I2S_BITS_PER_SAMPLE_32BIT,
    .channel_format = I2S_CHANNEL_FMT_ONLY_LEFT,
    .communication_format = I2S_COMM_FORMAT_I2S_MSB,
    .intr_alloc_flags = ESP_INTR_FLAG_LEVEL1,
    .dma_buf_count = 8,
    .dma_buf_len = 256,
    .use_apll = false,
    .tx_desc_auto_clear = false,
    .fixed_mclk = 0
  };
  i2s_pin_config_t pin_config = {
    .bck_io_num = MIC_BCLK,
    .ws_io_num = MIC_LRCL,
    .data_out_num = I2S_PIN_NO_CHANGE,
    .data_in_num = MIC_DOUT
  };
  i2s_driver_install(I2S_NUM_0, &i2s_config, 0, NULL);
  i2s_set_pin(I2S_NUM_0, &pin_config);
  i2s_set_clk(I2S_NUM_0, SAMPLE_RATE, I2S_BITS_PER_SAMPLE_32BIT, I2S_CHANNEL_MONO);
  Serial.println("[I2S] RX initialized");
}

void setup() {
  Serial.begin(115200);
  delay(200);

  // Unique device id
  uint64_t mac = ESP.getEfuseMac();
  snprintf(deviceId, sizeof(deviceId), "%04X%04X", (uint16_t)(mac >> 32), (uint32_t)(mac & 0xFFFF));

  pixel.begin();
  pixel.clear();
  pixel.show();
  setLED(0, 0, 32); // Blue: boot

  connectWiFi();
  connectMQTT();

  initI2S_RX();

  setLED(32, 32, 0); // Yellow: baseline learning
}

void loop() {
  mqtt.loop();

  // Buffer to read one frame
  static int32_t raw32[FFT_N];
  size_t bytes_read = 0;
  size_t want = sizeof(raw32);
  uint8_t* ptr = (uint8_t*)raw32;
  while (bytes_read < want) {
    size_t n = 0;
    esp_err_t err = i2s_read(I2S_NUM_0, ptr + bytes_read, want - bytes_read, &n, portMAX_DELAY);
    if (err != ESP_OK) {
      Serial.printf("[I2S] read error: %d\n", (int)err);
      break;
    }
    bytes_read += n;
  }

  // Convert to float/double for FFT, with 16-bit scaling
  // INMP441: 24-bit valid left-justified in 32-bit word
  // Shift right 8 to get 24->16 bits, then apply window
  double dc = 0.0;
  for (size_t i = 0; i < FFT_N; i++) {
    int32_t s32 = raw32[i] >> 8; // 24b -> 16b
    int16_t s16 = (int16_t)(s32 & 0xFFFF);
    double x = (double)s16 / 32768.0;
    vReal[i] = x;
    dc += x;
    vImag[i] = 0.0;
  }
  // Remove DC offset
  dc /= FFT_N;
  for (size_t i = 0; i < FFT_N; i++) {
    vReal[i] -= dc;
  }

  // Hann window
  for (size_t i = 0; i < FFT_N; i++) {
    vReal[i] *= 0.5 * (1.0 - cos((2.0 * PI * i) / (FFT_N - 1)));
  }

  // FFT
  FFT.Windowing(vReal, FFT_N, FFT_WIN_TYP_RECTANGLE, FFT_FORWARD); // already windowed manually
  FFT.Compute(vReal, vImag, FFT_N, FFT_FORWARD);
  FFT.ComplexToMagnitude(vReal, vImag, FFT_N);

  // Compute bands
  float bands[NUM_BANDS];
  computeBands(vReal, bands, FFT_N, SAMPLE_RATE);

  // Update baseline or compute score
  float score = updateBaselineAndScore(bands);

  static uint32_t lastPub = 0;
  uint32_t now = millis();

  if (!baselineDone) {
    if (now - lastPub > 1000) {
      String payload = String("{\"state\":\"baseline\",\"frames\":") + baselineCount + "}";
      publishMQTT("status", payload);
      lastPub = now;
    }
    // LED: yellow during baseline
    setLED(32, 32, 0);
  } else {
    // LED and beeper behavior
    if (score > ANOMALY_THRESHOLD) {
      ledPulseRed();
      beepAlert(1000, 150, 2500);
      // Publish anomaly event with summary
      String payload = String("{\"anomaly\":true,\"score\":") + String(score, 2) +
                       ",\"sr\":" + SAMPLE_RATE + ",\"fft\":" + FFT_N + "}";
      publishMQTT("anomaly", payload);
    } else {
      ledGreenSolid();
    }

    // Periodic metrics
    if (now - lastPub > 2000) {
      // publish a compressed band snapshot (first 5 for brevity) and score
      String payload = String("{\"score\":") + String(score, 2) + ",\"bands\":[";
      for (int i = 0; i < NUM_BANDS; i++) {
        payload += String(bands[i], 3);
        if (i < NUM_BANDS - 1) payload += ",";
      }
      payload += "]}";
      publishMQTT("metrics", payload);
      lastPub = now;
    }
  }
}

Notes on code:

• I2S RX: Configured for 32-bit samples; the INMP441’s 24-bit payload is right-shifted to get a 16-bit normalized signal for processing.
• Feature extraction: 16 log-power bands from the magnitude spectrum; baseline mean/std built online; the anomaly score is the sum of absolute z-scores.
• MQTT: Periodically publishes metrics; sends an anomaly message when threshold is exceeded.
• LED: Blue during boot; Yellow during baseline; Green when normal; pulsing Red on anomaly.
• Audio: On anomaly, a short sine beep is generated and sent to MAX98357A via I2S1.

Build/Flash/Run Commands

Use PlatformIO Core (CLI). Ensure your USB serial port is recognized (e.g., COMx on Windows, /dev/ttyUSBx or /dev/tty.SLAB_USBtoUART on Linux/macOS).

Initialize (if creating from scratch):

pio project init --board esp32dev

Build:

pio run

Upload (auto-detects port):

pio run --target upload

Open serial monitor at 115200 baud:

pio device monitor -b 115200

If you have multiple serial devices, specify port:

pio device monitor -b 115200 -p /dev/ttyUSB0

Upgrade PlatformIO platforms if needed:

pio platform update espressif32

Step-by-step Validation

Follow these steps in order. Keep the serial monitor open to observe logs.

1) Power and Basic Bring-up
– Connect the ESP32-DevKitC V4 via USB.
– Verify the CP210x driver is installed if the serial port is missing.
– On reset, the LED should show blue, then yellow while baseline is building.

2) Verify Wi-Fi and MQTT Connectivity
– Ensure WIFI_SSID and WIFI_PASS are correct in main.cpp.
– Ensure MQTT_HOST and MQTT_PORT point to your broker.
– Monitor serial output:
– [WiFi] Connected, IP: …
– [MQTT] Connected
– Use mosquitto_sub to watch topics:

mosquitto_sub -h 192.168.1.100 -t "device/esp32-anom/#" -v

You should see “status”, “info”, and periodic “metrics” messages once baseline completes.

3) Verify Microphone Capture
– During baseline, keep the environment at normal noise levels. Speak or clap and observe that baseline count increases to 300 (~10 seconds).
– After baselineDone is true (shown implicitly by metrics messages), the LED should turn green when quiet.

4) Validate Anomaly Detection
– Create a distinct sound (e.g., jingling keys, whistle, or banging a cup).
– Observe:
– The LED should pulse red briefly during the event.
– You should hear a short beep from the speaker (MAX98357A).
– In mosquitto_sub, you should see an “anomaly” message like:
device/esp32-anom/anomaly {«anomaly»:true,»score»:38.42,»sr»:16000,»fft»:512}

5) Validate Metrics Payloads
– Monitor metrics payload content:
– Confirm “score” keeps low steady values when quiet and higher spikes during events.
– Confirm “bands” contains 16 values (log-power per band).

6) Test MQTT Retained Status
– Restart the ESP32; check that device/esp32-anom/status retains {«state»:»online»} after reconnection.
– Confirm the “info” message with device id, sample rate, FFT size is sent on connect.

7) Validate Speaker Output
– If no beep is heard, verify wiring to MAX98357A:
– DIN → GPIO26, BCLK → GPIO27, LRC → GPIO25, VIN → 5V, GND → GND.
– Try increasing beep volume parameter in beepAlert (e.g., 3000–6000). Beware of clipping and speaker limits.

8) Validate LED Behavior
– During baseline: Yellow.
– Normal: Green solid.
– Anomaly: Pulsing Red during the event.

Troubleshooting

• Serial port missing (Windows):
• Install Silicon Labs CP210x VCP driver.
• Try a different USB cable or port.
• Boot loops or random resets:
• Power supply marginal. Lower speaker volume, use a powered USB hub, or provide a separate 5V supply for MAX98357A and LED. Always share grounds.
• No MQTT messages:
• Confirm broker IP and port.
• Test with mosquitto_pub to ensure broker is reachable:
  mosquitto_pub -h 192.168.1.100 -t test -m hello
• Ensure no firewall blocks port 1883.
• No “anomaly” events:
• The environment may be too consistent; try louder or more distinct acoustic events.
• Lower ANOMALY_THRESHOLD (e.g., 18.0f).
• Extend BASELINE_FRAMES (more stable baseline; also increases baseline time).
• Microphone reads all zeros or garbage:
• Verify INMP441 wiring. L/R must be tied to GND for Left. Ensure SD to GPIO34 (input-only).
• Swap MIC_LRCL and MIC_BCLK if miswired.
• Check that sample rate is compatible (16 kHz generally works; you can try 32 kHz or 48 kHz).
• Distorted or missing beep:
• Confirm MAX98357A SD/EN pin is not held low (some boards have SD pin; tie high).
• Ensure I2S1 sample rate (16 kHz) matches beep generation and that speaker is connected properly.
• WS2812B flicker or no light:
• Ensure a common ground.
• Add a 330–470 Ω resistor in series with the data line.
• Try lowering data line length; ensure 5V power is stable.

Improvements

• Use TLS for MQTT:
• Switch to WiFiClientSecure and configure CA certs; use port 8883.
• Smarter anomaly detection:
• Replace band z-score with Mahalanobis distance using covariance matrix.
• Compute MFCCs (requires DCT and mel filters) and train a baseline GMM or One-Class SVM offline, then run inference on-device.
• Adaptive baseline:
• Slowly adapt means/variance to non-anomalous drift to reduce false positives.
• Overlap and windowing:
• Use 50% overlap to improve time resolution; use Hanning or Blackman windows directly via FFT library for consistency.
• On-device logging:
• Send ring buffers of raw audio around anomaly to MQTT or to an SD card for offline analysis.
• Multi-topic segregation:
• Separate “metrics” and “events” into different QoS and retention policies.
• Power optimization:
• Use light sleep between frames; reduce sample rate when idle, increase upon suspected activity.
• LED UI:
• Map anomaly score to LED color (green to red gradient).
• Calibrated thresholds:
• Implement a calibration mode via MQTT command (e.g., device/esp32-anom/cmd) to restart baseline remotely.

Final Checklist

• Hardware
• Espressif ESP32-DevKitC V4 (ESP32-WROOM-32) connected via USB.
• INMP441 wired: VDD→3V3, GND→GND, L/R→GND, SCK→GPIO32, WS→GPIO33, SD→GPIO34.
• MAX98357A wired: VIN→5V, GND→GND, DIN→GPIO26, BCLK→GPIO27, LRC→GPIO25, speaker attached.
• WS2812B wired: 5V→5V, GND→GND, Data→GPIO4 (optional 330 Ω resistor).
• Software
• PlatformIO installed; project contains platformio.ini and src/main.cpp as provided.
• Wi-Fi and MQTT credentials configured in src/main.cpp.
• Build succeeds via pio run.
• Flash via pio run –target upload; serial monitor at 115200.
• Validation
• Device connects to Wi-Fi and MQTT; status/info messages published.
• Baseline completes (yellow LED transitions to green).
• Making a distinct sound triggers red LED pulse, beep, and MQTT anomaly message.
• Metrics publish periodically and scores make sense.
• Troubleshooting
• If no anomaly triggers, adjust ANOMALY_THRESHOLD and verify sensor wiring.
• If audio issues occur, recheck MAX98357A pin mapping and supply voltage.
• If LED doesn’t light, verify WS2812 power, ground, and data pin.

With this setup, your Espressif ESP32-DevKitC V4 (ESP32-WROOM-32) + INMP441 + MAX98357A + WS2812B operates as a robust, real-time i2s-anomaly-detection-mqtt node, streaming health metrics, announcing anomalies with visual and audible cues, and providing a solid base for more advanced on-device audio analytics.

Carlos Núñez Zorrilla

Electronics & Computer Engineer

Telecommunications Electronics Engineer and Computer Engineer (official degrees in Spain).

Follow me:

Who we are