**2. Study Background**

In this research, CLAAS Hungária Kft (CLH) is adapted as a case study. CLH was established in 1997 in Hungary as a subsidiary company of CLAAS Group. CLAAS group is an international German family-owned business company based in Germany and owns many manufacturing plants worldwide. CLAAS is a world-leading manufacturer of agricultural equipment and machinery such as tractors and combine harvesters. Since establishment, CLH expanded from 350 workers and 8 hectares plant to more than 700 workers working on a 14-hectare plant and became a center of excellence for combine harvester tables and trolley carts production. CLH manufactures supplementary devices such as combine harvester tables, cutting heads, and trolley carts, as shown in Figure 1. These devices are shipped from Hungary either to the mother company that is located in Germany or directly to the end-customers for final assembly with the machine which can be a combine harvester or a tractor. The cost of a single failure is tremendously high, not only due to the machine cost itself but also due to the entailed logistics and the re-work cost.

**Figure 1.** Sample of devices manufactured at the subsidiary company subject of this study.

CLH's sta ff has developed "Quality Checklists" for every product, process or manufacturing phase. These quality checklists are developed based on the FMEA documents and are being used at the quality gates in the shop floor in order to ensure that common failure causes are avoided. Moreover, this process aims to make sure that critical device components are installed and configured at the optimal conformance to design. However, as mentioned earlier, FMEA documents are prepared during the product design phase and can be changed once the serial production is initiated. Meanwhile, further failure modes can be detected at the final assembly phase. Therefore, these quality checklists are demanded to be dynamic, updatable and responsive to real quality issues reported during or after production.

This research activity is focusing on a single device that consists of the combine harvester feeder house as shown in Figure 2. Feeder house is a device that is attached to the combine harvester to facilitate the control of the cutting head and the flow of crops from the cutting head to the combine harvester. The device consists of several complex systems such as mechanical, hydraulic, electrical, and electronic systems. This device is wholly manufactured in the subsidiary company in Hungary and dispatched to be assembled to the combine harvester at the mother company in Germany.

Failures or defects which are observed during assembly or reported by end-users are gathered on a daily basis through the global ERP system of the company. After that, this information is extracted and manually and reviewed by an experienced quality managemen<sup>t</sup> team. This evaluation process aims at analyzing root cause and consequently taking the needed correction actions in order to maintain profitability and high-quality production. The company uses an internally customized FMEA technique to evaluate reported claims by obtaining RPN for every claim according to FMEA documents. The method which is used here aims at generating an RPN value for every claim on a scale from 1 to 300 points, where 300 is the highest priority number.

**Figure 2.** Feeding house attached to the combine harvester body and ready to be attached with a cutting head.

RPN in this CLH is obtained based on three major factors: (severity, occurrence, and impact). Severity, or gravity as named by the company's internal manuals, represents the risk consequences of the claim from customer and company perspectives. It also includes the cost of resolving this issue and the safety impact on the operator. The weight of this factor ranges between 1 to 10 points, where 1 is the lowest severity and 10 is the highest. In the meanwhile, occurrence represents the number of incidents a specific claim has been witnessed in a specific period. The weighting scale of this factor is also 1–10, where 10 is the highest. Impact is weighted by a scale of 3 points from 1 to 3. Impact represents the repair efforts, time, repetition of the same work, and the overall impact of the claim on the reputation and image of the company. The meaning of every scale value from 1 to 10 is elaborated in detail in [3,9]. The evaluation process is summarized in Figure 3 below.

**Figure 3.** Factors affecting claim ranking and the weight of every factor.

Equation (1) shows the multiplication of the three factors values that results an RPN value between 1 and 300 points. An RPN value above 160 points is classified at a very high priority, while, a value between 100 and 160 points is classified as a high priority. Medium priority is noted if the RPN value is in the range of 35–100, while low priority is noted if the RPN value is less than 35:

$$\text{RPN} = \text{Severity} \times \text{Occurrence} \times \text{Impact} \tag{1}$$

According to the RPN value of every claim, the quality team decides the next handling steps. Further steps could be tracing root cause(s) and ensuring the elimination of such cause(s) and/or updating the quality checklists to ensure further failures will not repeat in the future. Time and experience play a crucial role in this regime. It is important to improve the process of evaluating claims and lever the current experience.

The evaluation and ranking process requires highly experienced people who are fully aware of the FMEA documents and its applications. The volume, velocity, and veracity of claims reported, and their processing time is very critical from a quality managemen<sup>t</sup> perspective. It is essential in such a high-value industry to resolve issues as soon as they are reported. Early and fast processing of quality issues is translated to a lower quality cost and will positively enhance the general business performance. Moreover, standardization of the evaluation process and consistency of the process is vital to guarantee consistent RPN results every time.

The accumulated experience, time of processing, consistency of the evaluation process can be attained through the proposed solution in this paper; utilizing automated machine learning to classify and analyze claims data. Machine learning capabilities provide the capacity to analyze several input features (columns) at one dimension, aligned with a large volume of data (rows) at the other dimension. This helps in discovering and analyzing unseen factors, considering that the best quality practices focus on the claim root cause analysis. Additionally, utilizing technology whenever possible is very promising in the industry, because of its availability at any time (24/7) under any conditions and its ability to go deeper in analysis beyond human capacity. Delegating such tasks to machines will let human intelligence focus on higher strategic issues and to reach a higher level of e fficiency and e ffectiveness.

#### **3. The Proposed FMEA Analysis Method**

In this section, it is suggested to utilize supervised machine learning technology to replace human intervention in processing, evaluating, and categorizing claims. The current flow of claims from involved parties is illustrated in Figure 4. Claims from internal company quality product audit (product audit claims) and issues that were detected during assembly (cross-company claims) are pipelined in the company's ERP system and human intervention is important at one point to evaluate claims manually. Based on the evaluation results, quality managemen<sup>t</sup> decides how to deal with every single claim to find the root cause of the problem. This is done either by following the eight disciplines of problem-solving (8D) methodology for critical or high-ranking issues or by just updating the shop floor quality checklists in order to ensure the quality of next produced devices. Otherwise, this reported issue is just as it is an accidental incident and occupies a very low RPN value.

Accordingly, a dataset that contains one-year data of claims is extracted from the ERP system of the company. This data is concerning the selected device only (the feeder house shown in Figure 2). Firstly, to ensure the accuracy of the developed models, the data was re-evaluated and validated manually by experienced quality engineers to obtain the three RPN elements (severity, occurrence, and detectability) and to define the root cause and the source manufacturing process (such as cutting, bending, welding, painting, assembly, etc.) of every claim. The evaluation process depends on the experience of the quality team and based on the internal FMEA procedure for every failure mode. The resulted updated dataset is used for training and develop an ML model that is deployed to predict an RPN value for future failures claims and classify its root cause instantly without further human intervention.

**Figure 4.** The flow of internal quality audit and cross-company claims to quality management.

In this research, Google AutoML is selected for three reasons: first, its effectiveness, as Google AutoML is free to try and utilizes the latest ML technology developed by Google brain team. Second, its ease of use, which is very important to ensure the sustainability of the project results after the research cooperation ends, keeping in mind that people who are working at the partner company are less experienced with coding and modeling. Therefore, such a friendly system will ensure that the partner company can deal with the work after the end of the project with the least knowledge of coding and data processing. Third, its ability to integrate. It is agreed with the partner company to integrate the developed models in the company's ERP system. Google AutoML offers the ability to deploy and integrate the resulted models by application programming interface (API).

Supervised machine learning is used to improve the failure claims processing process. Claims are reported by engineers at the assembly location in Germany or from service centers to the quality managemen<sup>t</sup> office through the company's ERP system. Claims are analyzed, categorized and ranked by the quality managemen<sup>t</sup> team based on FMEA documents. Accordingly, failures are ranked and prioritized based on their importance and critical impact.

The proposed solution aims at developing an automatic claim ranking system to replace human intervention based on developing four machine learning models that can read, analyze, evaluate, and assign relevant ranking values for every processed claim. In order to do so, a dataset of already evaluated claims is used to train the model. Afterward, the model will be deployed to evaluate new claims based on the experience gained by the training data. Figure 5 below elaborates on the process of models' development, its inputs, and outputs. The first step in models' training is to preprocess the input data, feature selection, and data types. The auto machine learning tool results four models that will be able to predict four independent values by which three of them will be multiplied to calculate an RPN value. The fourth decides the source manufacturing process of the same claim.

**Figure 5.** Development of machine learning model.
