Multimodal Environment Prediction With Position-Weighted Data Fusion

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing environmental prediction techniques using machine learning fail to consider the varying degrees of influence based on the positional relationship between pieces of information, especially when events occur far from the observation location, and do not adequately account for the interrelationships between multiple events affecting the natural environment.

Innovation Solution

An environment prediction system that utilizes multimodal information, including image, numerical, and text data, separates and encodes positional information, and integrates these expressions using a trained environmental indicator prediction model to predict environmental indicators, considering the magnitude of influence based on positional relationships.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If future prediction is performed based on multiple types of input information, then prediction accuracy is improved, but the importance evaluation fails to account for varying degrees of influence based on positional relationships

Engineering Contradiction:
Improveprediction accuracyVSAvoidpositional relationship information
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent applies local quality by encoding positional information to give different weights to different input information based on their spatial relationships. The position information processing unit encodes geographical positions and calculates distances between observation locations and target locations, allowing the system to treat nearby information as more important than distant information, thus resolving the contradiction between using multiple information types and accounting for their varying influences.

Inventive Principle:
Principle #3Local quality

2Adaptability or versatility

If events far from the observation location are considered, then the scope of prediction is improved, but the complexity of processing positional relationships increases

Engineering Contradiction:
Improveprediction scopeVSAvoidprocessing complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent uses parameter changes by transforming positional information into encoded representations that capture distance and spatial relationships. The position information processing unit encodes geographical positions and calculates distance parameters, which are then integrated with other input information. This allows the system to handle events from various locations systematically, expanding prediction scope while managing complexity through parameter-based processing.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If interrelationships between multiple events are considered based on positional relationships, then prediction comprehensiveness is improved, but the computational complexity increases

Engineering Contradiction:
Improveprediction comprehensivenessVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSPower

Solution Approach 1:

The patent applies segmentation by dividing the processing into distinct functional units: position information acquisition, position information processing, input information encoding, and information integration. This modular approach allows the system to handle multiple events and their positional relationships systematically, improving prediction comprehensiveness while managing computational complexity through structured processing stages.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4676098A1System and method for predicting environment
Publication Date: 2026.01.07 HITACHI LTD
  • EP4676098A1 patent drawingFigure 1
  • EP4676098A1 patent drawingFigure 2
  • EP4676098A1 patent drawingFigure 3

AI summary

An environment prediction system for predicting an environmental indicator in a predetermined position and/or region includes: an information input unit configured to acquire, as input information, at least two or more types of information from among image information, numerical information, or text information each including geographical position information; an information estimation unit including at least a position information acquisition unit configured to separate or estimate the position information from the input information, a position information processing unit configured to encode the separated or estimated position information and generate a plurality of position information embedded expressions, an input information encoding unit configured to encode the acquired input information for each piece of information and generate a plurality of input information embedded expressions, and an information integration processing unit configured to embed the position information embedded expressions into the plurality of input information embedded expressions respectively to generate a plurality of integrated embedded expressions, the information estimation unit configured to predict the environmental indicator in the predetermined position and/or region by inputting the plurality of integrated embedded expressions to a trained environmental indicator prediction model; and an information output unit configured to output the predicted environmental indicator.