LoRaWAN-Signal-Mapping/client_src/js/Worker/AIWrapper.mjs

"use strict";


import brain from 'brain.js';
import haversine from 'haversine-distance';

import {
	normalise_lat, normalise_lng,
	
	normalise_gateway_distance,
	
	unnormalise_rssi
} from '../../../common/Normalisers.mjs';


class AIWrapper {
	get training_mode() {
		return this.index.properties.training_mode;
	}
	
	constructor() {
		this.setup_complete = false;
		
		this.map_bounds = null;
		this.index = null;
		this.Config = null;
		
		this.gateways = new Map();
	}
	
	async setup({ bounds, index, Config }) {
		this.map_bounds = bounds;
		this.index = index;
		this.Config = Config;
		
		console.log("Loading models");
		
		// WebGL isn't available inside WebWorkers yet :-(
		
		for(let gateway of this.index.index) {
			let net = new brain.NeuralNetwork(/*gateway.net_settings*/);
			net.fromJSON(
				gateway.frozen_net
			);
			
			this.gateways.set(
				gateway.id,
				{ ai: net, latitude: gateway.latitude, longitude: gateway.longitude }
			);
		}
		console.log("Model setup complete.");
		this.setup_complete = true;
	}
	
	predict_row(lat) {
		if(!this.setup_complete)
			throw new Error("Error: Can't do predictions until the setup is complete.");
			
		let result = [],
			stats = {
				rssi_min: Infinity,
				rssi_max: -Infinity
			};
		
		for(let lng = this.map_bounds.west; lng < this.map_bounds.east; lng += this.Config.step.lng) {
			let max_predicted_rssi = -Infinity;
			
			for(let [/*gateway_id*/, gateway] of this.gateways) {
				// Generate the input data
				let distance_from_gateway = haversine(
					{ latitude: lat, longitude: lng },
					gateway
				);
				
				let input_data = {
					latitude: normalise_lat(lat),
					longitude: normalise_lng(lng)
				};
				if(this.training_mode !== "unified")
					input_data.distance = normalise_gateway_distance(distance_from_gateway);
				
				// Validate the input data
				if(Number.isNaN(input_data.latitude)
				|| Number.isNaN(input_data.longitude)
				|| (
					this.training_mode !== "unified"
					&& Number.isNaN(input_data.distance))
				) {
					console.error(input_data);
					throw new Error("Error: Invalid neural network input.");
				}
				
				// Run the input through the neural network
				let next_value = gateway.ai.run(input_data);
				if(Number.isNaN(next_value[0])) {
					console.log(next_value);
					throw new Error("Error: Neural network returned NaN");
				}
				
				// Operate on the output
				max_predicted_rssi = Math.max(
					max_predicted_rssi,
					next_value[0]
				);
			}
			
			// Un-normalise the output of the neural nentwork to something sensible
			max_predicted_rssi = unnormalise_rssi(max_predicted_rssi);
			
			// Record the statistics
			if(max_predicted_rssi > stats.rssi_max)
				stats.rssi_max = max_predicted_rssi;
			if(max_predicted_rssi < stats.rssi_min)
				stats.rssi_min = max_predicted_rssi;
			
			
			result.push(max_predicted_rssi);
		}
		
		return { result, stats };
	}
}

export default AIWrapper;