To Örebro University

Testing of complex systems may often require the use of tailored-made solutions, expensive testing equipment, large computing capacity, and manual implementation work due to domain uniqueness. The aforementioned test resources are expensive and time-consuming, which makes them good candidates to optimize. A radio base station (RBS) is a complex system. Upon the arrival of new RBS generations, new testing challenges have been introduced that traditional methods cannot cope with. In order to optimize the test process of RBSs, product quality and production efficiency can be studied.

Despite that AI techniques are valuable tools for monitoring behavioral changes in various applications, there have not been sufficient research efforts spent on the use of intelligent manufacturing in already existing factories and production lines. The concept of intelligent manufacturing involves the whole system development life-cycle, such as design, production, and maintenance. Available literature about optimization and integration of industrial applications using AI techniques has not resulted in common solutions due to the complexity of the real-world applications, which have their own unique characteristics, e.g., multivariate, non-linear, non-stationary, multi-modal, class imbalance; making it challenging to find generalizable solutions. This licentiate thesis aims to bridge the gap between theoretical approaches and the implementation of real industrial applications. 

In this licentiate thesis, two questions are explored, namely how well AI techniques can perform and optimize fault detection and fault prediction on the production of RBSs, as well as how to modify learning algorithms in order to perform transfer learning between different products. These questions are addressed by using different AI techniques for test optimization purposes and are examined in three empirical studies focused on parallel test execution, fault detection and prediction, and automated fault localization. For the parallel test execution study, two different approaches were used to find and cluster semantically similar test cases and propose their execution in parallel. For this purpose, Levenshstein distance and two NLP techniques are compared. The results show that cluster-based test scenarios can be automatically generated from requirement specifications and the execution of semantically similar tests can reduce the number of tests by 95\% in the study case if executed in parallel. 

Study number two investigates the possibility of predicting testing performance outcomes by analyzing anomalies in the test process and classifying them by their compliance with dynamic test limits instead of fixed limits. The performance measures can be modeled using historical data through regression techniques and the classification of the anomalies is learned using support vector machines and convolutional neural networks. The results show good agreement between the actual and predicted learned model, where the root-mean-square error reaches 0.00073. Furthermore, this approach can automatically label the incoming tests according to the dynamic limits, making it possible to predict errors in an early stage of the process. This study contributes to product quality by monitoring the test measurements beyond fixed limits and contributes to making a more efficient testing process by detecting faults before they are measured. Moreover, study two considers the possibility of using transfer learning due to an insufficient number of anomalies in a single product. 

The last study focuses on root cause analysis by analyzing test dependencies between test measurements using two known correlation-based methods and mutual information to find strength associations between measurements. The contributions of this study are twofold. First, test dependencies between measurements can be found using Pearson and Spearman correlation and MI; and their dependencies can be linear or higher order. Second, by clustering the associated tests, redundant tests are found, which could be used to update the test execution sequence and choose to execute only the relevant tests, hence, making a more efficient production process by saving test time.