This section outlines the implementation results of the proposed procedure. The experiment was performed at the Smart Manufacturing Process Innovation Center, an open smart manufacturing technology testbed at the Korea Electronics Technology Institute in Korea. The empirical validation was carried out on a fully automated machining and logistics line within the testbed. This line includes three Computerized Numerical Control machines (CNCs) with robotic arm, two Automated Guided Vehicles (AGVs), and an Automated Storage and Retrieval System (AS/RS). The machining line can produce a variety of products, including chip pans, impellers, engine blocks, and standard reference parts. In this study, the focus is on machining a standard reference part. Information regarding the standard reference part is provided in Fig. 3.
The cycle time for machining each standard reference part is about 2 hours. Under normal conditions, without material shortages or machine failures, it is feasible to produce 70 pieces over a span of 5 days (8 hours/day), considering the standard working time. This quantity is taken as the assumed order volume for the experiment. The experimental results were examined to confirm the applicability and effectiveness of the proposed framework, with the experimental data being applied at each stage of the framework.
4.1 Process discovery stage
In this production line, products go through the stages of receiving, machining, transporting, and shipping. The level of detail and the format of the output data are depicted in Fig. 4.
In the receiving stage, the system assigns a box number for the receiving order using a QR code. After the receiving information is entered and executed, two AGVs transport the raw materials in the boxes to the storage by reading the QR codes on the boxes. The storage shelves the boxes in the order they arrive. During the raw material machining stage, AGVs transport the boxes from the storage to the machining process sequentially. The robotic arm (Worker#A, #B, #C) at the CNC machine reads the barcode on the box, secures the material in the machine, and begins machining as per the order. After machining, the robotic arm places the finished product back into the original box, and the AGV returns it to the storage. In the final shipping stage, when a shipping order is entered into the system at the scheduled time, the AGV loads the specified product to the shipping area. The defined activities for this process are listed in Table 3.
Table 3
Define an Activity for a Machining-Logistics Line
No.
|
Activity
|
Resource
|
Description
|
1
|
Store.order
|
SYSTEM
|
Receiving orders
|
2
|
Supplement_in
|
AGV1, 2
|
Moving raw material to a storage
|
3
|
ST_in_RawPallet
|
SHUTTLE
|
Loading material
|
4
|
cncMachine#.pickRawPallet.moveUnitToAisle
|
AGV1, 2
|
Moving raw materials to the machining line
|
5
|
cncMachine#.set_pick_Item
|
WORKER#
|
Material clamping in the CNC machine
|
6
|
cncMachine#.Start_CNC
|
WORKER#
|
Adjustments for Starting a CNC Machine
|
7
|
cncMachine#.processing
|
CNC#
|
Machining raw materials
|
8
|
cncMachine#.Confirm_End
|
WORKER#
|
Removing the CNC machined product
|
9
|
cncMachine#.send_FinishedPallet
|
AGV1, 2
|
Moving product to a storage
|
10
|
ST_in_FinishedPallet
|
SHUTTLE
|
Loading product
|
11
|
Release.Order
|
SYSTEM
|
Shipping orders
|
12
|
supplement_out
|
AGV1, 2
|
Moving product to a shipping point
|
13
|
Delivery.Confirmation
|
SYSTEM
|
Complete a shipment
|
#: A, B, C
|
Throughout this process, each box number is automatically scanned, and a timestamp is recorded whenever raw materials or finished products are processed, as shown in Table 4.
Table 4
Event log data with data pre-processing
Case_ID
|
Activity
|
Start
|
Complete
|
Resource
|
BOX-0006
|
Store.Order
|
2023-10-16 09:00:00
|
2023-10-16 09:00:10
|
SYSTEM
|
BOX-0006
|
supplement_in
|
2023-10-16 09:00:30
|
2023-10-16 09:01:14
|
AGV1
|
BOX-0006
|
ST_in_RawPallet
|
2023-10-16 09:01:53
|
2023-10-16 09:02:38
|
SHUTTLE
|
BOX-0006
|
cncMachineB.pickRawPallet.moveUnitToAisle
|
2023-10-16 10:17:50
|
2023-10-16 10:17:58
|
AGV1
|
BOX-0006
|
cncMachineB.set_pick_Item
|
2023-10-16 10:18:47
|
2023-10-16 10:19:45
|
WorkerB
|
...
|
...
|
...
|
...
|
...
|
This event log enables the derivation of a process model through process mining. The process model visualizes the flow of raw materials over the observed period. To compare different algorithms and modeling techniques, the process was represented as a Petri net using Alpha, Heuristic, and Inductive miner algorithms in Table 5.
In Table 5, the models represented as Petri nets for each algorithm reveal differences in accuracy through silent transitions. The model using the Inductive miner effectively handles complex process structures and clearly expresses the relationships between activities, providing a detailed reflection of the process structure and clearly illustrating various process paths and relationships. While this enhances the model's accuracy and level of detail, it may also increase computation time. Therefore, when selecting an algorithm, it is important to consider the required model accuracy, data complexity, and the feasibility of real-time application. Among the algorithms, the Inductive miner, which offers the highest accuracy, was used to derive the model using BPMN, Petri-net, and Process tree modeling techniques, as shown in Fig. 5.
BPMN is a standardized graphical language used to represent processes, utilizing various BPMN symbols such as flows, activities, events, gateways, and connections to illustrate parallel and optional flows [42]. Petri nets are the most fundamental process model in process mining, using a graph structure composed of places and transitions to model the states and events of a process, making them useful for representing parallel executions and detailed control flows. They are particularly useful for modeling and analyzing complex systems and for mathematically representing system behavior [43]. Process trees depict processes in a hierarchical manner, simplifying the representation of complex processes while ensuring soundness, thus avoiding issues like deadlock and safeness [44]. Other process representation techniques include Dependency Graph, C-NET, and WF-NET, among others, and the appropriate method should be selected based on the process's characteristics and goals [45].
4.2 Analysis stage
Before the analysis, a Conformance Check is conducted to derive the fitness value between the extracted process model and the actual event log. The Token-based replay conformance checking method uses tokens to replay the process model and track their usage. As shown in Fig. 6, tokens have four states: p (produced), c (consumed), m (missing), and r (remaining). The fitness value is calculated using the formula \(\:Fitness=\frac{1}{2}\left(1-\frac{m}{c}\right)+\frac{1}{2}\left(1-\frac{r}{p}\right)\). The resulting value ranges from 0 to 1, with values closer to 1 indicating a better fit of the model. However, this method does not take into account the frequency information of the event log and cannot be applied in cases with silent transitions or duplicate transitions.
The Alignments-based conformance checking method evaluates fitness based on the trace of the process model that most closely matches the trace of the event log ("Closest matching path"). The optimal alignment for each trace is derived and used to calculate fitness with the formula \(\:Fitness\left(L,M\right)=1-\frac{cost(L,\:\:M)}{{move}_{L}\left(L\right)+\left|L\right|*{move}_{M}\left(M\right)}\). An optimal alignment has a cost of 0, indicating a perfect fit. When the alignment is not optimal, a cost is incurred. Fitness is then calculated based on how this cost compares to the cost of the worst fitting trace. The closer the cost of the optimal alignment is to that of the worst fitting trace, the lower the fitness value. Despite taking a long time to compute and requiring significant memory, alignment-based conformance checking is the most realistic and accurate fitness verification method. It determines fitness by identifying the process model behavior that most closely matches the event log behavior [46].
Figure 7 shows the results of Alignments and TBR (Token-based replay) conformance checking using PM4Py, with a fitness value of "1.0," indicating that the process model perfectly matches the log data. This means that the process model accurately reproduces the log data, making it suitable for analysis and application based on this model [47].
4.2.1 Process analysis
(1) Process model discovery
In the process model discovery analysis, a process model like the one shown in Fig. 5 was derived. Figure 8 provides a more detailed analysis, showing a model annotated with the frequency and duration of each activity, created using BupaR [48].
In Fig. 8, the nodes represent activities, which are types of unit tasks, and the colors of the nodes indicate relative frequency. The edges represent the connections between tasks, indicating precedence and succession relationships, as well as the average waiting times between nodes. Based on this, a global and local analysis of the process model is performed. This process model reveals three main issues. First, out of the total 70 raw materials, only 64 were processed, failing to meet the order allocation. Second, among CNC#A, #B, and #C, CNC#A had the lowest production output. Third, there were significant waiting times from storage to the CNC machines. However, the long waiting time for the ‘Release.Order’ process in storage is due to batch shipping. These issues discovery form the basis for a more detailed analysis.
(2) Process pattern analysis
Process pattern analysis aims to enhance work efficiency and identify anomalies by understanding high-frequency process patterns. After deriving the process model, the variants, which are process trace patterns, are analyzed from the event log to identify the most frequent patterns and key patterns. This ensures visibility into the main process flow and can be used for corporate decision-making [21]. By deriving a variant list, the commonly used representative variants or abnormal variants can be identified, providing an understanding of the various execution possibilities of the process. In this model, there are 8 variants, and the variant list is visualized as shown in Fig. 9 [48].
Among the derived variants, the top 3 variants, V1 to V3, account for 91.43% of the total traces, encompassing 64 traces. These patterns indicate that production and shipping were completed successfully, aligning with the analysis in the process model discovery. The V4 has three traces, representing 4.29% of the total number of traces. This suggests that there is still raw material in storage. The bottom 3 variants, V5 to V7, depict traces where raw materials were left unfinished in the machining process.
(3) Bottleneck analysis
Bottleneck analysis aims to identify points in the process with long lead times (or throughput times) that cause performance degradation or reduced efficiency. Bottlenecks are determined using the average time for each activity [18]. Metrics such as sojourn time, working time, and waiting time can be utilized. Essentially, by calculating the quartiles between the preceding and succeeding tasks of each activity, the instances with the longest times in each quartile can be identified as bottleneck points.
Table 6
List of top bottlenecks by waiting time
No. | Predecessor | Successor | Event count | Wait time |
---|
mean | min | max | median |
---|
1 | ST_in_RawPallet | cncMachineB.pickRawPallet.moveUnitToAisle | 24 | 2d 5h 46m 42s | 1h 16m | 4d 7h 14m | 2d 4h 45m |
2 | ST_in_RawPallet | cncMachineA.pickRawPallet.moveUnitToAisle | 19 | 2d 5h 46m 25s | 1h 16m | 4d 7h 4m | 2d 4h 26m |
3 | ST_in_RawPallet | cncMachineC.pickRawPallet.moveUnitToAisle | 24 | 2d 5h 23m 35s | 1h 16m | 4d 7h 1m | 2d 4h 19m |
4 | cncMachineA.processing | cncMachineA.Confirm_End | 18 | 5h 36m 46s | 2h 25m | 17h 25m | 2h 26m |
5 | cncMachineB.processing | cncMachineB.Confirm_End | 23 | 4h 24m | 2h | 16h 32m | 2h 2m |
6 | cncMachineC.processing | cncMachineC.Confirm_End | 23 | 4h 18m 28s | 1h 37m | 15h 43m | 1h 39m |
7 | ST_in_FinishedPallet | Release.Order | 52 | 1d 22h 21m 52s | 1m | 4d 3h 59m | 2d 55m 30s |
Table 6 presents a list of significant bottlenecks, highlighting tasks that prolong the work time and consequently extend the overall lead time. Understanding these bottlenecks is simplified when compared with the process model. For instance, bottlenecks in items 4, 5, and 6 contribute to those in items 1, 2, 3, and 7, particularly in the CNC machine processing stage. Moreover, the performance spectrum in Fig. 10, obtained via BupaR, enables the identification of bottlenecks by analyzing lead time quartiles [49]. The longest instances are indicated in red, and the shortest in blue, facilitating the identification of problematic sections within the process.
4.2.2 Time analysis
(1) Time performance analysis
The purpose of time performance analysis is to reduce process time by analyzing the time metrics of events. It involves measuring time metrics such as work time, wait time, and lead time for each case, and extracting the average or median time metrics for each activity and resource. By identifying and estimating the causes of prolonged time intervals, and calculating lead times for each resource, activity, and case as needed, the optimal time for resource combinations can be determined to improve process performance [24]. Additionally, time analysis is performed based on process model discovery, focusing on processing and wait times as performance aspects. By analyzing the minimum, maximum, average, and median values of work time, the specific causes can be identified, and the time performance of resources is evaluated to see if it matches the standard time.
Initially, by organizing and analyzing the work times for each activity in the process by quartiles, it was found that the CNC machine's work time is considerably longer than that of other tasks. The CNC machine constitutes about 95% of the total work time, underscoring its critical importance. Moreover, there is significant variability in work times across different CNC machines. Table 7 presents a comparison between the Predetermined Time Standard (PTS) and the actual work times.
Table 7
Compare to PTS and Working Time
Resource | Activity description | PTS (min) | Real Mean working time (min) | Increase/Decrease Ratio |
---|
AGV | Moving raw material to a storage | 0.5 | 0.73 | + 46% |
SHUTTLE | Loading material | 0.5 | 0.74 | + 48% |
AGV | Moving raw materials to the machining line | 0.33 | 0.39 | + 18.18% |
Worker | Material clamping in the machine | 0.5 | 0.89 | + 78% |
Worker | Adjustments for Starting a CNC Machine | 0.17 | 0.15 | -11.76% |
CNC | Machining raw materials | 120 | 122.14 | + 1.78% |
Worker | Removing the machined product | 0.01 | 0.02 | + 50% |
AGV | Moving product to a storage | 0.33 | 0.71 | + 115.15% |
SHUTTLE | Loading product | 0.5 | 0.74 | + 48% |
1 product completed working time | 122.84 | 126.51 | + 2.99% |
When comparing the production time of a single workpiece to the standard time, it was observed that most actual work times exceeded the standard. The actual work time was about 3% longer than the PTS, indicating the need to adjust either the standard time or the machine settings for each resource. For a more detailed analysis, the mean working time was calculated by separating the lines. Table 8 was created to compare the work speeds of each task and CNC machine.
Table 8
Mean working time by machining line
No. | Activity | Resource | Mean working time (min) |
---|
Line A | Line B | Line C |
---|
1 | cncMachine#.pickRawPallet.moveUnitToAisle | AGV | 0.385 | 0.436 | 0.433 |
2 | cncMachine#.set_pick_Item | Worker | 0.653 | 0.901 | 1.128 |
3 | cncMachine#.Start_CNC | Worker | 0.154 | 0.149 | 0.153 |
4 | cncMachine#.processing | CNC | 168.711 | 121.75 | 98.955 |
5 | cncMachine#.Confirm_End | Worker | 0.023 | 0.017 | 0.020 |
6 | cncMachine#.send_FinishedPallet | AGV | 0.471 | 0.704 | 0.959 |
Total working time | 170.397 | 123.956 | 101.648 |
This revealed time differences between lines that were difficult to detect with only the process model. The most significant discrepancy was between the robotic arm in Line C and the CNC machine in Line A. A comparison using Line C, which had the shortest work time, as a benchmark is shown in Table 9.
Table 9
mean work time for each line compared to line C
No. | Based on line C | Mean wokring time (min, %) |
---|
Line A | Line B |
---|
1 | Each line processing Time difference | 68.75(167%) | 22.31(122%) |
2 | Robot arm processing time difference | 0.47(63%) | 0.23(82%) |
3 | CNC machining time difference | 69.76(171%) | 22.80(123%) |
When reviewing the work times for each line, it was found that the robotic arm Worker#A is 63% faster than Worker#C, and Worker#B is 23% faster. However, the impact of the robotic arms on the overall lead time per line is negligible and can be overlooked. CNC#A's work time is 171% longer than CNC#C's, and CNC#B's work time is 123% longer. This analysis allows for the identification of specific time improvement areas to enhance process performance.
(2) Delay analysis
Delay analysis is conducted to identify delay patterns and reduce process lead time. By creating a chart with activities and resources on the x-axis and delay time on the y-axis, the delay times due to resource unavailability can be visualized. This allows for the identification of resources that significantly increase delay times [30].
In Fig. 11, the red boxes highlight activities and resources with sharply increased delay times, indicating that the process cannot produce products quickly enough, thereby signifying bottlenecks. Referring to bottleneck analysis, it becomes clear that the activities within the red boxes correspond to those affected by bottlenecks caused by preceding activities. Although less significant than the main issues, ‘Supplement_in’ and ‘Supplement_out’ can also be considered bottlenecks caused by AGVs or processes due to simultaneous inflow and outflow situations.
(3) Dotted chart analysis
Dotted chart analysis is used to understand the overall time distribution of events for each case. On the chart, the x-axis represents time, and the y-axis represents cases. An arbitrarily marked horizontal red line indicates one case, with the dots on the line representing the activities of that case. The chart can be divided into two types based on the x-axis. The first type, where the x-axis represents the occurrence time of activities, allows for comparing process progress to identify process issues [23]. The second type, where the x-axis represents the elapsed time since the start of the case, helps identify characteristics of delayed activities or cases that took longer than average [16]. Figure 12 is a dotted chart with the x-axis showing the time when activities occurred. This chart is based on actual process time and can identify exceptions like break times.
The case ID are ordered from top to bottom on the y-axis according to their start times. A filled circle marks the start of a task, while an empty circle marks the end of a task. The work schedule spans from 9 AM to 12 PM and from 1 PM to 6 PM. Interestingly, CNC machine operations continue during the break time from 12 PM to 1 PM, and also beyond work hours, which end at 6 PM.
4.2.3 Resource analysis
(1) Resource network analysis
Resource network analysis seeks to understand the relationships between resources and identify key resources with many connections to others. This analysis helps derive the organizational structure of resources, determine their importance, and analyze the network [17]. The resource network can be examined using Social Network Analysis (SNA) to study relationships and interactions between organizations [50]. Figure 13 depicts the resource network of the process.
In Fig. 13, unnecessary networks were omitted and resources performing similar tasks were clustered to reduce complexity. The resources are divided into groups: CNC #A, CNC #B, CNC #C, Storage/Retrieval, and AGV. The AGV group, in particular, serves as a network hub, playing a central role in task transfers between groups. To understand resource connections, centrality, which measures the importance of each resource, is used. The results are displayed in Table 10.
Table 10
Centrality by top 5 resources
resource | Average Clustering Coefficient | Degree centrality | Betweenness centrality | Eigenvector centrality | In Degree | Out Degree | In event | Out event |
---|
AGV#1 | 0.0985 | 0.91 | 0.28 | 0.47 | 5 | 5 | 131 | 132 |
AGV#2 | 0.91 | 0.28 | 0.47 | 5 | 5 | 133 | 133 |
SYSTEM | 0.64 | 0.13 | 0.36 | 4 | 3 | 198 | 197 |
Worker#A | 0.55 | 0.17 | 0.31 | 2 | 2 | 47 | 47 |
Worker#B | 0.55 | 0.17 | 0.31 | 2 | 2 | 47 | 47 |
The Average Clustering Coefficient indicates the ratio of actual connections among a node's neighbors to the maximum possible connections among them in the network. This value ranges from 0 to 1, with higher values showing the formation of tighter clusters. With an average clustering coefficient of 0.0985, the network shows low clustering, indicating independent connections among resources. Degree centrality measures a node's importance based on its number of direct connections. Betweenness centrality shows how often a node lies on the shortest path between two other nodes. Eigenvector centrality assesses a node's importance based on the importance of its neighbors, meaning a node surrounded by important nodes is also important. AGV#1 and AGV#2 have the highest centrality values, highlighting their key roles in task transfer within the process. Thus, the process manager can suggest improvements to reduce the failure rate of AGVs.
(2) Workload analysis
The objective of workload analysis is to identify points of overload by examining the work frequency of resources in the entire event log [26]. The analysis was first conducted by categorizing the resources into CNC and non-CNC groups. Figure 14 illustrates the measurement of work frequency and the time required for each CNC
If the CNCs are of the same type, CNC#C, having the shortest processing time, is the most efficient. It is also crucial to monitor each analysis result for CNC#C closely to prevent overload. Additionally, comparing the planned and actual workloads of resources helps identify those with more actual work than planned. Figure 15 illustrates the work frequency and processing time of resources involved in the movement of products.
AGV#1 and AGV#2 were employed with comparable frequency in Fig. 15. Furthermore, when this analysis is cross-checked with Resource Network analysis, it is evident that Worker#A, Worker#B, and Worker#C exhibit varying usage patterns in relation to the utilization of the CNC machines. This indicates that all resources are being utilized in an appropriate manner. This analysis method can assist in the identification of overloaded resources, thereby improving the efficient operation and utilization of production resources.
(3) Cost analysis
Cost analysis can extend the cost items within the process model to thoroughly review the cost elements of each activity. Besides providing a comprehensive analysis of process-related costs, it can categorize and analyze costs such as material costs, labor costs, and overhead costs [20]. Process mining-based cost analysis can utilize features to track the cost values of each task by category.
4.2.4 Quality analysis
(1) Anomaly detection
Anomaly detection is a method for identifying unusual changes in event logs. Typically, outliers in task times can be identified using a box plot of resource-specific task times. Additionally, anomalies can be identified by creating rules based on control-flow, resource, and time perspectives [22], as well as a real-time event log analysis method that calculates the spectral gap when the manufacturing process workflow is a linear program [33]. To detect unusual changes for each resource, outliers were analyzed using the resource's box plot. Figure 16 shows the analysis results for the processing machines.
Outliers were identified in CNC#A and CNC#C. If task times are significantly longer or shorter than expected, it could indicate problems with equipment efficiency or operations, necessitating root cause analysis and improvements. Figure 17 presents the analysis results for other resources.
Multiple outliers were discovered for AGV#1 and AGV#2. To conduct further analysis, the activities for each AGV were separated to identify outliers in task times.
Figure 18 reveals outliers that commonly occur when each AGV dispatches workpieces. Notably, AGV#1 exhibits a high frequency of outliers when transferring to CNC#B and CNC#C. This indicates inefficient operations on specific routes, potentially due to obstacles, route optimization issues, or logistical overload.
(2) Yield monitoring
Yield monitoring aims to track production volume over time, identify defect rates in the process, and detect defects early to maximize efficiency. This method generates graphs that compare the number of products entering and leaving the process [17]. A lower output quantity compared to the input quantity indicates a process problem. The quantities at the beginning and end are used to calculate the process defect rate.
(3) Quality report
The goal of the quality report is to ensure the production and delivery of high-quality products. While improving product quality depends on the final product's quality, improving process quality focuses on the steps used in production. By measuring the defect rate for each event and the first-pass yield, and incorporating this data into the event log, process quality can be assessed [20]. Additionally, by including data on defective products rejected during process inspections and accepted products that have completed all processes in the event log, the distribution of defect event occurrences can be analyzed to evaluate product quality [31].
4.3 Diagnose and improve
Table 11 lists the results from the analysis of event log data from four perspectives. This analysis helps determine which parts of the process need improvement and what specific areas should be addressed.
Table 11
Diagnose and improve summary list
No. | Activity | Resource | process analysis | Time analysis | Resouce analysis | Quality analysis |
Process model discovery | Process pattern analysis | Bottleneck analysis | Time performance analysis | Dotted chart analysis | Delay analysis | Resource network analysis | Workload analysis | Quality report |
1 | supplement_in | AGV | Create a process model (fitness: 1.0) | Create a list of 7 variants | - | - | - | Bottleneck caused by AGVs | central role | - | - |
2 | ST_in_RawPallet | Shuttle | Bottleneck | - | - | - | - | Outliers detection |
3 | cncMachineA.pickRawPallet.moveUnitToAisle | AGV | - | - | Bottleneck caused by Process | central role | - | - |
4 | cncMachineA.processing | CNC | Bottleneck | Slower than PTS | Processing during breaks and after work | Bottleneck caused by CNC | - | - | Outliers detection |
5 | cncMachineA.send_FinishedPallet | AGV | - | - | - | - | central role | - | - |
6 | cncMachineB.pickRawPallet.moveUnitToAisle | AGV | - | - | Bottleneck caused by Process | central role | - | - |
7 | cncMachineB.processing | CNC | Bottleneck | Slower than PTS | Processing during breaks and after work | Bottleneck caused by CNC | - | - | Outliers detection |
8 | cncMachineB.send_FinishedPallet | AGV | - | - | - | - | central role | - | - |
9 | cncMachineC.pickRawPallet.moveUnitToAisle | AGV | - | - | Bottleneck caused by Process | central role | - | - |
10 | cncMachineC.set_pick_Item | Robot | - | Slower than PTS | - | - | - | - |
11 | cncMachineC.processing | CNC | Bottleneck | Faster than PTS | Processing during breaks and after work | - | - | Overload management | Outliers detection |
12 | cncMachineC.send_FinishedPallet | AGV | - | | - | - | central role | - | - |
13 | ST_in_FinishedPallet | Shuttle | Bottleneck | - | Bottleneck caused by Process | - | - | - |
14 | supplement_out | AGV | - | - | Bottleneck caused by AGVs | central role | - | Outliers detection |
4.3.1 Process diagnose and improve
According to the process analysis, 64 out of 70 cases were successfully completed following the main variants (V1, V2, V3). Among the 4 cases that ended abnormally, 3 (V5, V6, V7) had materials left inside the processing machines, and 1 (V4) had materials left in the logistics line. As a result, only 64 out of 70 products were completed. By categorizing the processes into receiving, processing, and shipping, and examining the “Case_ID” flow of each process, it was found that batch order picking, and FIFO methods caused bottlenecks in the receiving and shipping processes. To improve this, the sequence of the process is adjusted. Figure 19 shows the adjusted process model.
When the CNC is idle, incoming raw materials are immediately moved to the CNC to reduce idle time. When the CNC is operating, raw materials are placed in storage. By setting these conditions, the first three raw materials are processed immediately without passing through the logistics line, while the remaining raw materials are stored in the storage. This reduces the CNC's idle time by approximately 77 minutes and allows for the shipment of a total of 67 products. Additionally, five bottleneck points were identified, with waiting times ranging from 30 minutes to 2 days. Time analysis was used to identify the causes of these bottlenecks.
4.3.2 Time diagnose and improve
Time analysis showed that actual working times generally exceeded PTS, leading to a 3.2% difference in lead time. This discrepancy results from differences between the standard and actual working times of robot arms and CNC machines. Environmental conditions and equipment status cause these variations, which can impact long-term production planning. To address this issue, it is necessary to adjust production plans to reflect actual working times or to configure equipment to meet standard times. Here are three proposed improvement measures.
Adjustment of CNC Machine Operating Time: The average working times of CNC #A and #B differ by 45% and 22%, respectively, compared to CNC #C. These differences stem from the characteristics of the machining centers on lines A, B, and C. By adjusting CNC #A to suit the specific workpiece, its working time can be reduced to 2 hours, enabling the production of a total of 69 workpieces.
Improvement of Storage Lines: Bottlenecks in the storage lines occur when raw materials are waiting in the storage and when they are moved from the storage to the CNC machines. This is due to the long working times of the machining centers. To address this, the first proposed solution is to change the process sequence, as suggested in process improvement. Secondly, if waiting time increases due to an influx of materials and workpieces, optimizing AGV routes and increasing the number of AGVs can reduce bottlenecks by improving work processing speed. This can be tested through process simulation to assess the improvement results.
Adjustment of Robot Arm Operating Time: Compared to the robot arm on line #C, robot arm #A has a working time difference of 71%, and robot arm #B has a difference of 33%. Adjusting the operating times of the robot arms is necessary to improve efficiency.
4.3.3 Resource diagnose and improve
Analyzing the relationships and importance of all resources reveals that the average clustering coefficient is 0.1, indicating low interdependence between process stages. This means there is a low risk of problems in one stage spreading to others. Each process stage can operate independently, making it relatively easy to change processes or add new ones, demonstrating good process flexibility [51]. AGVs, which have high centrality, are crucial in the network. They are essential for managing the flow of information and resources and maintaining network stability. Therefore, process managers need to propose measures for the efficient operation of AGVs and strategies to reduce their failure rates to anticipate and prepare for future process issues. Additionally, among the CNCs, #C is the most efficient device, capable of handling more tasks than other CNCs. However, appropriate management strategies and preventive measures are needed to prevent it from becoming overloaded.
4.3.4 Quality diagnose and improve
An analysis of outliers for each resource was conducted, and two improvement measures for the machining centers and AGVs are proposed. However, in this study, the anomalies of AGVs are not related to product quality. Nonetheless, in lines handling perishable products, this could pose a problem.
Machining Center Condition Analysis: In Fig. 16, outliers were identified in CNC#A and CNC#C during processes with Case_IDs BOX-0070 (CNC#C) and BOX-0071 (CNC#A). For CNC#C, the working time was about 1.8 times longer than the average, suggesting a potential mechanical defect. CNC#A had a case where the activity ended significantly faster than the average working time, which could result in uneven product machining and lower quality. It is necessary to investigate issues such as tool wear, raw material quality, and CNC control system problems.
AGV Condition Analysis: In Fig. 18, many outliers in the AGVs occurred when transporting materials to CNC B and C lines, indicating potential issues with AGV path teaching or mechanical problems. Additionally, the outliers in AGV#2 were due to the manager's command to immediately transport finished products to the shipping location, leading to longer average transport distances and extended working times.