Predictive Test Sequencing for Early Manufacturing Failure Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine-learning techniques, such as neural networks, struggle to predict manufacturing test failures in product testing sequences due to the lack of structure and temporal dependencies, leading to inefficient testing processes and resource wastage.
Innovation Solution
A machine-learning-based model is trained to predict failure probabilities of future product tests based on earlier test outcomes, allowing for early termination or optimization of testing sequences, using a computational implementation that receives and processes sequences of test results to adjust testing procedures dynamically.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine-learning techniques are used to monitor production defects, then defect detection capability is improved, but the ability to predict future test failures remains insufficient
Solution Approach 1:
The patent applies preliminary action by training the machine learning model on historical test result sequences before actual testing occurs. The model learns temporal patterns and dependencies from past data, enabling it to predict future test failures before they happen. This preparatory training phase allows the system to anticipate potential failures and optimize test sequences in advance.
Solution Approach 2:
The patent implements feedback by using actual test outcomes to continuously refine and update the machine learning model. The model receives feedback from completed tests in the form of pass/fail results and test value metrics, which are used to adjust predictions for future tests. This closed-loop feedback mechanism improves prediction accuracy over time.
2Reliability
If complete testing sequences are performed on all product units, then quality assurance is improved, but time and resource consumption increase
Solution Approach 1:
The patent applies partial action by performing only the necessary subset of tests rather than completing full testing sequences on all units. The machine learning model predicts which tests are likely to fail based on earlier results, allowing the system to skip redundant tests or terminate sequences early when failure is predicted, thus reducing unnecessary testing while maintaining quality assurance.
Solution Approach 2:
The system performs preliminary prediction using the trained machine learning model before executing the complete test sequence. By predicting future test outcomes based on initial test results, the system can determine in advance which tests can be skipped or terminated early, significantly reducing the actual testing time required while maintaining quality standards.
3Ease of manufacture
If fixed testing sequences are used for all product units, then process simplicity is improved, but adaptability to individual product variations deteriorates
Solution Approach 1:
The patent applies dynamics by making the test sequence adaptive and dynamic rather than fixed. The machine learning model analyzes individual product test results in real-time and dynamically adjusts the remaining test sequence for each product unit. This allows the testing process to adapt to specific product characteristics and predicted failure patterns while maintaining overall process simplicity through automated decision-making.
Solution Approach 2:
The system changes parameters by dynamically modifying test sequence parameters based on machine learning predictions. Instead of using a fixed sequence, the system adjusts which tests are performed, their order, and termination points based on real-time analysis of test results and predicted failure probabilities, enabling adaptability while keeping the underlying process framework simple.
4Loss of information
If machine-learning models learn temporal dependencies in test sequences, then sequence order influence is captured, but computational complexity and inappropriate weighting increase
Solution Approach 1:
The patent extracts only the essential temporal patterns and dependencies from test sequences that are relevant for failure prediction. Rather than using complex models that capture all temporal relationships, the system selectively extracts and utilizes only the critical temporal dependencies needed for accurate prediction, simplifying the model while maintaining effectiveness.
Solution Approach 2:
The system applies local quality by focusing the machine learning model's attention on specific critical temporal relationships in the test sequence rather than treating all time steps equally. The model identifies and weights only the locally important temporal dependencies that most influence failure prediction, reducing overall model complexity while maintaining predictive accuracy.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Example embodiments are disclosed of systems and methods for predicting failure probabilities of future product tests of a testing sequence based on outcomes of prior tests. Predictions are made by a machine-learning-based model (MLM) trained with a set of test-result sequence records (TRSRs) including test values and pass/fail indicators (PRIs) of completed tests. Within training epochs over the set, iterations are carried out over each TRSR. Each iteration involves sub-iterations carried out successively over test results of the TRSR. Each sub-iteration involves (i) inputting to the MLM values of a given test and those of tests earlier in the sequence while masking those later in the sequence, (ii) computing probabilities of test failures for the masked tests found later in the sequence than the given test, and (iii) applying the PFIs of test results later in the sequence than the given test as ground-truths to update parameters of the MLM.