Temporal Computer Vision for More Accurate NDT Inspection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Traditional computer vision algorithms for non-destructive testing (NDT) are limited by the accuracy of predictions made from single images, leading to non-optimized results due to insufficient information utilization.

Innovation Solution

An inspection system employing a camera and sensors to capture images over time, utilizing a first computer vision algorithm for initial predictions and a second algorithm that incorporates motion data and previous image predictions to enhance accuracy, including the use of convolutional neural networks (CNNs) and recurrent neural networks (RNNs) like Long Short Term Memory (LSTM) networks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional computer vision algorithms are used for single image analysis, then the system complexity remains low, but the prediction accuracy is limited

Engineering Contradiction:
Improveprediction accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent transitions from analyzing single static images to analyzing sequences of images across the time dimension. By incorporating temporal information from multiple frames and motion data from sensors, the system creates a multi-dimensional analysis approach that significantly improves prediction accuracy while managing complexity through structured processing pipelines

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The system performs preliminary actions by capturing and preprocessing image sequences and motion data before the actual prediction analysis. Motion parameters are calculated in advance from sensor data, and image sequences are buffered and pre-processed, allowing the main prediction algorithm to work with prepared data rather than raw inputs, thus improving accuracy without proportionally increasing complexity

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If multiple images and motion data are integrated for analysis, then prediction accuracy improves, but the processing complexity increases

Engineering Contradiction:
Improveprediction accuracyVSAvoidprocessing complexity
Core Design Contradiction:
Measurement precisionVSDifficulty of detecting and measuring

Solution Approach 1:

The patent segments the complex analysis task into distinct modular components: image acquisition, motion data acquisition, motion parameter calculation, image sequence processing, and integrated prediction. Each module handles a specific aspect of the data, making the overall complex processing manageable through clear separation of concerns and specialized processing for each data type

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Motion parameters serve as intermediaries that bridge the gap between raw sensor data and the final prediction analysis. The system calculates motion parameters (such as velocity, acceleration, orientation) from raw motion data and uses these derived parameters in the prediction algorithm, creating an intermediate representation that simplifies the integration of multiple data sources

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4327282B1Computer vision predictions for non-destructive testing
Publication Date: 2026.01.14 BAKER HUGHES CO
  • EP4327282B1 patent drawingFigure 1
  • EP4327282B1 patent drawingFigure 2
  • EP4327282B1 patent drawingFigure 3

AI summary

Systems and methods for target inspection are provided. The system includes a camera, at least one sensor, and a controller. The camera acquires images of a target over time and the sensor(s) acquire motion data characterizing camera and target movement. The controller generates, using a first computer vision (CV) algorithm, an initial prediction and confidence level regarding an object of interest for a first image acquired at a first time. The controller also determines, using the motion data, a motion parameter characterizing relative motion between the camera and the target at the first time. The controller additionally receives a weighting based upon a second image acquired at a second time prior to the first time. The controller generates, using a second CV algorithm, a final prediction and confidence level for the first image based upon the first image, the initial prediction and confidence level, the motion parameter, and the weighting.