• spare part prediction
• clustering
• machine learning
• information retrieval

This thesis addresses the challenge of predicting the required spare parts for fault repair tasks in the context of a Swiss household appliances company. The primary objective is to gain a comprehensive understanding of the problem, conduct an in-depth exploratory data analysis, preprocess the data, and develop and evaluate various spare part prediction mechanisms. Additionally, the thesis proposes mechanisms to enhance system performance through the utilization of clustering techniques. The dataset used consists of two portions: the initial call taker information containing details about the fault and the model machine, and the subsequent portion filled by the service engineer, includes information about the operations performed and materials used.
The proposed spare part prediction system encompasses multiple approaches, including error code clustering and classification, information retrieval, and a tailored hierarchical filtering approach. These methods aim to identify past maintenance tasks with similar faults and recommend appropriate materials based on the findings
First of all, some key performance indicators are defined in order to understand the capabilities of a specific model, effectively compare different systems and present results to the customer. They are 3: Closed Task Percentage, fraction of tasks closed out of the total number of tasks, considering close a task when the system suggest all the material needed. Average Recall, computed on not closed tasks only, contains information about the number of materials guessed by the system and Average Precision that takes into account also how many materials the system is suggesting.
The first system is based on clustering machine errorcode based on errorcodes that need similar materials to be solved. Then a new repair task will be classified to its appropriate error code using the information available given by the customer. By mapping error code clusters to the most probable materials used in past tasks, this system suggests the required spare parts.
The second system employs information retrieval techniques, comparing different similarity measures and different pipeline comprising of reranking and filtering. So a new task will be compared with previous similar tasks to generate a ranked list of related tasks. Based on these similarities, the system suggests the relevant spare parts.
The third method involves filtering past tasks based on a hierarchy of machine models and types. By considering textual descriptions of errors, this system recommends materials used in tasks that exhibit similarities to the target task.
To further enhance the suggestion model, the thesis proposes clustering materials based on their textual descriptions. This approach not only recommends individual materials but also suggests the type of materials needed, empowering service engineers to choose the most suitable materials for specific machine models from the identified clusters. This improvement solves the detected problem of part replacement following a change of material supplier or a change of product brand.