Research System Architecture Specification

Multivariate Spatiotemporal Fusion of Microenvironmental Pollutants: A Hybrid Neural-Causal Framework

End-to-End System Architecture, Hardware Telemetry Protocol, Dynamic Database Persistence, Web Portal Engine, and PyTorch Neural-Causal Subclinical Health Deterioration Forecasting Pipeline.

IoT Edge Sensor ESP32 / ESP8266 Nodes

Backend Router PHP 5.6+ / SQLite WAL

AI Deep Learning Conv1D + Bi-LSTM + Attention

Causal Analytics Granger Causality (PCMCI)

Interactive Multi-Layer Specifications

Comprehensive 5-Tier Layered Architecture

Physical sensing node to deep neural-causal risk forecasting model pipeline.

Production Active Architecture

Edge Sensor Node

ESP32/ESP8266 microcontrollers execute analog signal sampling from MQ-135, MQ-7, and DHT22 sensors. Data is calibrated, normalized, serialized into JSON format, and transmitted over HTTP POST with failover retry handling.

Dynamic Application Core

Modular PHP Front Controller with custom routing (index.php) and SQLite Write-Ahead Logging (WAL) database. Features dynamic parameter registration: adding a sensor automatically updates the API schema, table columns, dynamic UI cards, and export engines.

Hybrid Neural-Causal AI

PyTorch temporal deep neural model combines 1D CNN for local feature extraction, Bi-LSTM for temporal sequence modeling, Multi-Head Self-Attention for long-range dependencies, and Granger Causality testing to forecast Subclinical Health Deterioration Index (SRI).

End-to-End Operational Workflow Timeline

Sequential execution lifecycle from hardware signal sampling to predictive health risk output.

Phase 1: Hardware Signal Acquisition Interval: 30s Loop

Microcontroller Sensor Sampling & Calibration

ESP32 reads analog voltages on pins 34 (MQ-135 CO₂/Air), 35 (MQ-7 CO), and digital signal on pin 32 (DHT22 Temperature).

float co2_ppm = map(analogRead(MQ135_PIN), 0, 4095, 380, 1200);
float co_ppm = (analogRead(MQ7_PIN) / 4095.0) * 15.0;
float aqi_val = (co2_ppm * 0.08) + (co_ppm * 4.5);

Phase 2: Ingestion & Verification HTTP POST Request

REST API Ingestion & Database Persistence

ESP32 sends HTTP POST payload to /api/sensor_data.php. PHP verifies incoming parameter keys against sensor_config schema, appends Asia/Kolkata timestamp (dd-mm-yyyy hh:mm am), and executes SQL transaction in database.sqlite.

{"device_id": "ESP32_NODE_01", "co2": 445.2, "co": 1.45, "aqi": 42.1, "temp": 26.8}

Phase 3: Web Visualization & Admin User UI Interaction

Dynamic Dashboard, Dynamic Columns & Analytics

Researchers log into web portal (/dashboard, /sensors). Dynamic sensor cards render live parameters with custom icons. Tabular view formats time-series logs with single-click Excel CSV / PDF export capabilities. Adding or removing a sensor in /config dynamically adds/removes table columns automatically.

Phase 4: Causal Machine Learning PyTorch & Statsmodels

Multivariate Spatiotemporal Fusion & Health Deterioration Forecasting

Jupyter Notebook (hybrid_neural_causal_framework.ipynb) fetches batch records via GET /api/sensor_data?limit=200. Computes Granger Causality directional influence matrix, feeds multivariate sliding windows into PyTorch Conv1D-BiLSTM Attention network, and outputs Subclinical Health Deterioration Index (SRI 0-100).

Data Flow & Database Entity Relationship (ER) Schema

SQLite relational database architecture, dynamic JSON schema structure, and REST API payload specs.

Table: users

SQLite

PK id INTEGER (AUTO)

name TEXT NOT NULL

email TEXT UNIQUE

contact TEXT NOT NULL

password TEXT (BCRYPT)

created_at DATETIME

Table: sensor_config

Dynamic Schema

PK id INTEGER (AUTO)

parameter_key TEXT UNIQUE

name TEXT NOT NULL

unit TEXT NOT NULL

icon TEXT NOT NULL

min_val / max_val REAL

display_order INTEGER

Table: sensor_readings

Time Series

PK id INTEGER (AUTO)

timestamp INTEGER (UNIX)

formatted_time TEXT (Kolkata)

data TEXT (JSON Payload)

device_id TEXT

created_at DATETIME

API Payload Standardization Specification (GET /api/sensor_data)

{
  "status": "success",
  "sensors": [
    { "parameter_key": "co2",  "name": "Carbon Dioxide",   "unit": "ppm", "icon": "wind" },
    { "parameter_key": "co",   "name": "Carbon Monoxide",  "unit": "ppm", "icon": "flame" },
    { "parameter_key": "aqi",  "name": "Air Quality Index","unit": "AQI", "icon": "activity" },
    { "parameter_key": "temp", "name": "Temperature",      "unit": "°C",  "icon": "thermometer" }
  ],
  "count": 50,
  "timezone": "Asia/Kolkata",
  "data": [
    {
      "id": 102,
      "timestamp": 1785489000,
      "formatted_time": "31-07-2026 06:30 am",
      "device_id": "ESP32_NODE_01",
      "co2": 442.5,
      "co": 1.20,
      "aqi": 40.8,
      "temp": 26.2
    }
  ]
}

Hybrid Neural-Causal Subclinical Health Risk Pipeline

PyTorch Conv1D-BiLSTM-Attention Deep Learning Architecture coupled with Granger Causal Matrix Discovery.

1D Temporal Convolution

nn.Conv1d(in=N, out=32, k=3)

Extracts local spatial-temporal microenvironmental pollutant feature maps across sliding window time intervals.

Bidirectional LSTM

nn.LSTM(hidden=64, bi=True)

Captures forward and backward temporal dynamics of pollutant accumulation and dispersion kinetics.

Attn

Multi-Head Self-Attention

nn.MultiheadAttention(heads=2)

Dynamically weights critical acute pollutant spikes against baseline concentrations across long sequences.

SRI

Subclinical Risk Score Head

FC + Sigmoid * 100.0

Predicts Subclinical Health Deterioration Index (SRI) on a 0-100 risk scale for early warning health intervention.

Multivariate Granger Causality Discovery Engine

Evaluates directional causality between microenvironmental variables ($CO_2, CO, Temp, AQI$) using Vector Autoregression (VAR) F-tests across temporal lags ($lag=1..3$). Quantifies causal strength $S_{i \to j} = 1 - p_{min}$ to establish structural causal relationships prior to neural network training.

Causal Test formulation:
SSR_test = grangercausalitytests(df[['AQI', 'CO2']], maxlag=3)

Target Output Metric:
Subclinical Risk Index (SRI) Forecast Score [0 - 100]