Data-Driven Process FMEA for Flexible Manufacturing Systems: Framework and Industrial Case Study

Abstract

Flexible automated assembly lines (FAALs) in Industry 4.0 require robust quality management that integrates operational data with systematic risk analysis. However, Process Failure Mode and Effects Analysis (PFMEA) documents are often developed during the design phase and not systematically updated with actual production data, leading to a gap between formal risk assessment and operational reality. This study addresses this gap by developing and validating an integrated data-driven framework that combines classical quality tools (process flow charts, check sheets, cause-and-effect diagrams, and Pareto analysis) with data-driven PFMEA, creating traceable links from operational logs to risk ratings. While individual quality tools are well-established, the core contribution of this work is a structured data transformation pipeline that creates traceable, auditable linkages from raw operational event logs to calibrated PFMEA ratings with quantified uncertainty—a combination not previously demonstrated for flexible assembly systems. The framework was applied to FMS-200, a modular FAAL for bearing units, consisting of eight stations and a common transfer system. Analysis of 186 failure events across 2743 assembly cycles, including 18 product configurations, identified 40 distinct failure modes with risk priority number (RPN) values ranging from 60 to 378, revealing that approximately 90% of the aggregated risk is associated with pneumatic systems. Monte Carlo uncertainty analysis (10,000 iterations) demonstrated robust rank stability, with the top five failure modes maintaining their relative ordering in over 90% of simulations. The framework provides production and quality managers with a systematic methodology to maintain PFMEA relevance through continuous data integration, enabling evidence-based prioritization of improvement actions.

1. Introduction and Background

1.1. Context and Motivation

Flexible Assembly Automated Lines (FAALs) are at the core of modern manufacturing systems, particularly in the context of Industry 4.0, where mass customization, shortened product life cycles, and supply chain uncertainty require simultaneously high flexibility and robust quality. Modular lines configured from autonomous stations and a common transfer system enable rapid reconfiguration and efficient use of equipment, but, at the same time, introduce substantial complexity in the interactions between processes, control logic, and failure modes. Systematic quality and risk management is a critical engineering challenge, especially when the line serves multiple product variants [1,2].

Industrial practice has traditionally relied on a combination of data-based monitoring and established quality management tools. The seven classical quality tools—flow chart, check sheet, control chart, histogram, Pareto diagram, cause-and-effect (Ishikawa) diagram, and scatter diagram—remain widely used because of their simplicity and visual transparency [3]. These “traditional” tools provide immediate operational insights but lack the structured framework for systematic risk assessment and prioritization that Process Failure Mode and Effects Analysis (PFMEA) provides.

1.2. The PFMEA Challenge in Flexible Manufacturing

Failure Mode and Effects Analysis (FMEA), and particularly PFMEA, is a standard method for systematically identifying, prioritizing, and mitigating potential process failures in production and assembly [4,5]. PFMEA provides a structured framework for describing process steps, their functions, possible failure modes, associated effects on the customer and the process, likely root causes, and existing preventive and detective controls. Quantitative assessment through severity (S), occurrence (O), and detection (D) ratings enables risks to be ranked and addressed in a transparent and justified way—a critical requirement for multi-station automated assembly lines.

However, several studies highlight persistent challenges in PFMEA implementation. Documented PFMEA files are often incomplete, use inconsistent terminology, and make limited use of available operational data [6,7]. More critically, PFMEA is frequently developed once during the planning phase or process ramp-up and then updated only sporadically or when major changes occur [8]. This leads to a gap between formal risk analysis and the actual behavior of the line: improvement activities become less effective, while the PFMEA gradually loses practical relevance.

1.3. Evolution of FMEA and Data-Driven Approaches

Within Industry 4.0, there is growing interest in “FMEA 4.0” and hybrid approaches that combine the expert knowledge embedded in FMEA with data-driven diagnostics and prognostics [9,10]. Part of this work focuses on data-oriented PFMEA for assembly processes [11]. In parallel, the concept of Reverse FMEA is gaining traction as a means for periodic auditing and updating of PFMEA based on actual behavior observed in production, rather than maintaining it as a static document [12].

Despite these developments, in many real-world applications, operational data on stops and defects are still analyzed separately from PFMEA documents. Classical quality tools are used for immediate troubleshooting and continuous improvement, while PFMEA remains a parallel, largely static, compliance-oriented document. The separation reduces the effectiveness of both approaches and prevents the creation of a unified, continuously updated risk profile.

FMEA was originally developed in the aerospace industry in the 1960s and uses the Risk Priority Number (RPN = S × O × D) for prioritizing failure modes [13]. However, the RPN approach has been widely criticized: different S, O, and D combinations can yield identical RPNs while representing fundamentally different risk profiles [14,15]. To address these limitations, numerous enhanced methods have been proposed, including fuzzy logic [15], multi-criteria decision-making, grey system theory [14], Bayesian networks [16], and hybrid approaches [17]. The AIAG & VDA guideline introduced Action Priority (AP), replacing RPN-based ranking with rule-based prioritization [5,18].

The integration of operational data into FMEA has emerged as an important research direction. Data-driven frameworks use historical failure records to estimate occurrence ratings and evaluate detection effectiveness [11], reducing reliance on expert-based assessments. In Industry 4.0 environments, hybrid frameworks combine expert knowledge with real-time sensor data and machine learning [9,10]. Uncertainty-aware approaches using Monte Carlo simulation provide probabilistic risk prioritization [14]. More recently, large language models (LLMs) have been explored for automating FMEA authoring [19] and generating risk management recommendations [20,21,22], pointing toward AI-assisted FMEA that aligns with the data-driven framework proposed here.

1.4. FMEA in Flexible Manufacturing and Quality Tool Integration

This study proposes an integrated framework that combines classical quality tools with data-driven PFMEA for FAALs. Previous work on the same FMS-200 flexible assembly automated line [2] demonstrated that classical quality tools—flow charts, check sheets, cause-and-effect diagrams, and Pareto charts—effectively identified that the transfer system and the first stations have dominant contributors to downtime and defects. However, that study did not include a formal PFMEA and did not quantify the risk associated with individual failure modes or with the contribution of each station to the overall risk profile.

Flexible manufacturing systems present specific FMEA challenges due to their modular architecture and complex station interactions [1]. Failure modes occur at the station-level and system-level and are product-specific, and modular FMEA approaches have been proposed to improve scalability [2]. However, operational data integration and quantitative risk aggregation at higher hierarchical levels remain insufficiently addressed.

Classical quality tools—flow charts, check sheets, Pareto diagrams, and cause-and-effect diagrams [3]—are widely used alongside PFMEA for operational problem-solving [2,8]. However, their systematic integration with PFMEA remains limited, particularly regarding structured data flow and traceability. Reverse FMEA [12], which audits and updates FMEA based on operational performance, addresses similar objectives but faces challenges due to limited structured data collection and the absence of clear procedures for translating observations into updated ratings.

1.5. Research Gap, Objectives and Contributions

The present study addresses a methodological gap at the intersection of PFMEA, classical quality tools, and flexible manufacturing. Although many recent studies have enhanced FMEA through fuzzy logic, grey systems, machine learning, and uncertainty-aware ranking methods, most industrial implementations still treat PFMEA as a largely static expert document, while operational failure data, check sheets, Pareto analyses, and root cause investigations are handled separately. This separation weakens traceability, delays PFMEA updates, and limits the usefulness of PFMEA in modular and reconfigurable production systems.

In this paper, the novelty does not lie in introducing new standalone quality tools. Rather, it lies in establishing a traceable integration architecture that transforms operational manufacturing evidence into structured PFMEA knowledge; supports hierarchical aggregation across failure mode, station, and technology levels; and improves prioritization robustness through uncertainty-aware analysis in a flexible manufacturing context. The research is guided by three questions:

• How can outputs of classical quality tools and shop floor evidence be formally translated into PFMEA entries and ratings in flexible assembly systems?
• How can risk be aggregated across different manufacturing levels so that local failure evidence can support system-level decision making?
• How robust are PFMEA priorities under plausible rating uncertainty?

This work contributes to the field in four ways:

• Proposing a five-stage operational framework, linking process modeling, data collection, classical quality tool outputs, PFMEA construction, and uncertainty-aware prioritization;
• Introducing a traceable rating logic that maps occurrence and detection ratings to observed operational evidence and control architecture;
• Enabling hierarchical risk aggregation at the station and technology level;
• Applying an uncertainty-based robustness layer.

When applied to the FMS-200 flexible assembly line, the framework identified strong risk concentration in transfer and pneumatic subsystems, highlighted bolt-handling and positioning failures as dominant critical modes, and showed that top-ranked risks remain stable under bounded rating uncertainty.

2. Research Methodology

2.1. Framework Overview and Design Principles

The proposed framework follows a five-stage process that systematically transforms operational data into actionable PFMEA insights. The design is guided by three principles:

• Traceability: Each PFMEA rating (S, O, D) should be supported, to the greatest extent possible, by either operational evidence or explicitly documented expert judgment, with the basis of assignment preserved throughout the analysis;
• Aggregation: Individual failure events should be aggregated to failure modes, and failure modes should be aggregated to station level and technology level to enable strategic decision-making;
• Uncertainty Awareness: Rating uncertainty should be explicitly quantified and propagated through the analysis to support robust prioritization decisions.

Figure 1 presents the five stages of the proposed framework.

• Stage 1—Process Modeling and Decomposition: The flexible assembly line is decomposed into distinct process steps using process flow charts. Each step is characterized by its primary function, inputs, outputs, and actuating technologies;
• Stage 2—Data Collection and Classification: Operational failure events are systematically collected using check sheets over a defined observation period. Events are classified by process step, failure type, and actuation technology;
• Stage 3—Root Cause Analysis and Prioritization: Cause-and-effect (Ishikawa) diagrams are constructed for dominant failure patterns. Pareto analysis identifies the stations and failure categories contributing to 80% of total events;
• Stage 4—PFMEA Construction with Data-Driven Ratings: For each identified failure mode, a PFMEA entry is created. Severity is assigned through a structured expert-supported procedure, occurrence is derived from observed event intensity, and detection is assigned using a rule-based evaluation of control architecture calibrated, where possible, by escape evidence. RPN is calculated, and AP is assigned using a simplified prioritization logic aligned with AIAG & VDA principles [5];
• Stage 5—Aggregation, Uncertainty Analysis, and Action Prioritization: Failure modes are aggregated by process step and actuation type. Monte Carlo simulation quantifies rating uncertainty and rank stability. Action priorities are established based on combined RPN, AP, and uncertainty metrics.

2.2. Classical Quality Tool Integration

In the present study, a set of classical quality tools was applied in a structured sequence to analyze operational problems in the flexible assembly line and to establish a data foundation for the subsequent PFMEA analysis. The combined use of these tools for problem identification and preliminary analysis has been described in the authors’ previous work [23]. In the current framework, their outputs serve as structured inputs for the PFMEA model.

The integration includes the following tools:

• Process Flow Charts—used to visualize the sequence of operations across the flexible assembly line and to define process step boundaries. Each process step is treated as a unit of analysis in the PFMEA;
• Check Sheets—structured data collection forms used to systematically record failure events during production. Each entry includes timestamp, process step, failure symptom, suspected cause category (6M), downtime duration, scrap/rework outcome, and actuation type;
• Cause-and-Effect (Ishikawa) Diagrams—constructed for the most frequent failure modes identified through Pareto analysis. The 6M structure supports the organization of potential causes and their mapping to PFMEA “Potential Cause” entries;
• Pareto Charts—applied at several analytical levels, including failure events by process step, actuation type, and cause category. The results support prioritization and define the scope of detailed PFMEA analysis.

In the present framework, the transformation from classical quality tool outputs to PFMEA fields follows an explicit mapping logic. Operational event records from check sheets were first normalized by process step and symptom type. Recurrent symptoms were consolidated into PFMEA failure-mode candidates, while Ishikawa-derived 6M branches were used to structure the potential cause field. Pareto outputs were used to prioritize which failure categories received full PFMEA treatment, while low-frequency categories were retained only when associated with elevated severity.

2.3. Data-Driven PFMEA Procedure

The PFMEA was constructed following the structure defined by AIAG & VDA [5] and IEC 60812 [4], with data-supported enhancements for rating assignment.

2.3.1. PFMEA Structure and Fields

For each process step, the following elements are documented: process step, function, potential failure mode, potential effect, severity, potential cause, occurrence, current controls (prevention and detection), detection, RPN, and AP. The AP assignment follows a simplified prioritization logic aligned with the AIAG & VDA guideline [5], as described in Section 2.3.5.

2.3.2. Severity Rating Methodology

Severity ratings are presented in

Table 1. Severity rating scale adapted to flexible assembly context.

S Rating	Description	Example in Assembly Context
9–10	Hazardous or serious safety concern	Assembly fails causing safety hazard to operator or end user
7–8	Very high effect; assembly inoperable	Bearing unit cannot function; customer return likely
5–6	High effect; significant degradation	Reduced bearing life; potential warranty claim
3–4	Moderate effect; minor degradation	Cosmetic defect; rework required
1–2	Low or no effect	Minor variation within tolerance; no impact

and are assigned on the ordinal 1–10 scale defined in the AIAG & VDA guideline [5], adapted to the functional, service, and quality consequences relevant to the studied flexible assembly system.

In the present framework, the final severity rating remains an expert-assigned ordinal score. However, to improve transparency, repeatability, and justification of that assignment, the expert judgment is supported by a proposed semi-quantitative severity–support index ${\mathrm{S}}^{\ast }$. The proposed index decomposes severity into four consequence dimensions relevant to automated assembly: quality impact, safety or service consequence, recoverability, and downstream propagation. It is expressed as:

$${\mathrm{S}}^{\ast }={\mathrm{w}}_{\mathrm{q}}\mathrm{Q}+{\mathrm{w}}_{\mathrm{s}}{\mathrm{S}}_{\mathrm{f}}+{\mathrm{w}}_{\mathrm{r}}\mathrm{R}+{\mathrm{w}}_{\mathrm{d}}{\mathrm{D}}_{\mathrm{t}}$$

(1)

where, in Equation (1), $\mathrm{Q}$ denotes the quality impact class of the failure (e.g., minor deviation, rework, scrap, or functional nonconformance), ${\mathrm{S}}_{\mathrm{f}}$ represents the safety or service consequence, R reflects the recoverability or reworkability of the failure, and ${\mathrm{D}}_{\mathrm{t}}$ captures the extent to which the failure propagates downstream before containment. The coefficients ${\mathrm{w}}_{\mathrm{q}}$, ${\mathrm{w}}_{\mathrm{s}}$, ${\mathrm{w}}_{\mathrm{r}}$, and ${\mathrm{w}}_{\mathrm{d}}$ are non-negative weighting factors satisfying $\sum {\mathrm{w}}_{\mathrm{i}}=1$, allowing the relative importance of the four dimensions to be adjusted to the specific assembly context.

Equation (1) is used as a structured decision aid for assigning the final ordinal severity rating, rather than as a replacement for the standard AIAG & VDA scale [5]. In this sense, $S^{*}$ functions as a supporting aggregation measure that constrains expert judgment within an explicit and auditable reasoning framework.

The composite index ${\mathrm{S}}^{\ast }$ is used to guide, but not mechanically replace, the final AIAG & VDA severity assignment. In the present study, the following interpretation bands were used as decision support: $1.0\le {\mathrm{S}}^{\ast }<1.5$ supports $\mathrm{S}=1–2$; $1.5\le {\mathrm{S}}^{\ast }<2.5$ supports $\mathrm{S}=3–4$; $2.5\le {\mathrm{S}}^{\ast }<3.25$ supports $\mathrm{S}=5–6$; $3.25\le {\mathrm{S}}^{\ast }<3.75$ supports $\mathrm{S}=7–8$; and $3.75\le {\mathrm{S}}^{\ast }\le 4.0$ supports $\mathrm{S}=9–10$. The final selection within each two-point severity band remains an expert decision based on the specific functional context, customer consequences, and consistency with AIAG & VDA guidance. This preserves compatibility with standard PFMEA practice while improving transparency and repeatability of rating assignment.

Each component variable is scored on a bounded ordinal scale (1–4) prior to aggregation, as defined in

Table 2. Scoring scheme for severity–support index component variables.

Score	Q	Sf	R	Dt
1	No relevant quality effect; minor deviation within tolerance	No safety or service impact	Immediate in-station correction possible	Contained at current station
2	Minor nonconformance; local adjustment or minor rework	Minor service degradation; negligible customer impact	Recoverable with minor intervention or local rework	May pass to next station; typically contained internally
3	Significant nonconformance; rework, replacement, or scrap	Functional degradation; warranty complaint probable	Recoverable only through significant disassembly or rework	Propagates through several downstream operations
4	Functional nonconformance affecting performance or acceptance	Serious functional or safety-related consequence	Non-recoverable in process; major intervention required	Can escape line or reach customer if not intercepted

. In the present case study, equal weighting was adopted as the baseline assumption (${\mathrm{w}}_{\mathrm{q}}={\mathrm{w}}_{\mathrm{s}}={\mathrm{w}}_{\mathrm{r}}={\mathrm{w}}_{\mathrm{d}}=0.25$), reflecting the absence of a sufficiently validated basis for differentiated weighting in the studied system. The equal weighting assumption represents a neutral baseline and avoids introducing additional subjectivity without validated calibration data. Future applications may calibrate these weights using structured expert elicitation, analytic hierarchy procedures, or empirical field data.

As an illustrative example, consider bolt cross-threading at FMS-207. The failure was scored as $\mathrm{Q}=3$ (significant nonconformance requiring rework), ${\mathrm{S}}_{\mathrm{f}}=4$ (serious service/safety consequence), $\mathrm{R}=3$ (recovery requires disassembly and rework), and ${\mathrm{D}}_{\mathrm{t}}=4$ (high downstream propagation if not contained). Under equal weights, this yields ${\mathrm{S}}^{\ast }=3.5$, which supports a very high severity classification ($\mathrm{S}=7–8$). Given the service-critical role of the fastening function and the conservative interpretation recommended for safety-relevant assembly failures, the final expert-assigned rating was set to $\mathrm{S}=9$. This example shows how the index supports a structured severity assignment while preserving consistency with engineering judgment and AIAG & VDA logic.

2.3.3. Occurrence Rating Methodology

Occurrence ratings are derived from observed failure frequencies using a consistent production exposure denominator. The occurrence measure is presented in

Table 3. Occurrence rating scale based on occurrence intensity (failures per 1000 production cycles).

O Rating	Occurrence Intensity (λᵢ = Failures per 1000 Cycles)	Assignment Rule
9–10	>50 per 1000 cycles (λ > 0.050)	Failure occurs in majority of production runs; systematic process deficiency
7–8	20–50 per 1000 cycles (0.020 ≤ λ ≤ 0.050)	Frequent failure; recurring pattern requiring priority intervention
5–6	5–20 per 1000 cycles (0.005 ≤ λ < 0.020)	Moderate occurrence; documented recurrence across observation period
3–4	1–5 per 1000 cycles (0.001 ≤ λ < 0.005)	Low occurrence; rare but documented events
1–2	<1 per 1000 cycles (λ < 0.001)	Very low or no observed events; theoretical possibility

and is defined as the number of observed occurrences of a given failure mode per production exposure unit over the observation horizon. To avoid mixing non-equivalent normalization bases, the rating assignment distinguishes clearly between absolute event count, relative share of all observed events, and occurrence intensity per exposure unit. Only the occurrence intensity is used for formal mapping to the O scale, while the absolute event count and the relative share serve as complementary descriptive statistics.

Let ${\mathrm{n}}_{\mathrm{i}}$ denote the number of recorded occurrences of failure mode i during the observation period, and let ${\mathrm{N}}_{\mathrm{e}\mathrm{x}\mathrm{p},\mathrm{i}}$ denote the corresponding production exposure for failure mode i. The occurrence intensity is defined as:

$${\mathsf{\lambda }}_{\mathrm{i}}={\displaystyle \frac{{\mathrm{n}}_{\mathrm{i}}}{{\mathrm{N}}_{\mathrm{e}\mathrm{x}\mathrm{p},\mathrm{i}}}}$$

(2)

For line-wide failure modes, the exposure denominator corresponds to the total number of observed assembly cycles during the study period. For station-specific or variant-dependent failure modes, the denominator corresponds to the effective exposure of the relevant station or product variant, i.e., the number of production cycles in which the corresponding function was active. This distinction is necessary in flexible assembly systems where certain stations may be bypassed or activated only for specific product configurations.

The 1–10 occurrence scale is assigned by mapping intervals of λᵢ from Equation (2) to ordinal classes, with higher classes corresponding to higher recurrence intensity. This formulation improves methodological consistency and makes the occurrence rating reproducible across product variants, routing changes, and future observation windows, provided that the same exposure definition is preserved.

2.3.4. Detection Rating Methodology

Detection ratings reflect the likelihood that the existing control architecture will identify a failure before it propagates downstream or escapes the process. In the present framework, detection is assigned through a semi-quantitative rule set combining recorded escape evidence, control design characteristics, and documented engineering judgment. The D rating is based on four elements: control type (manual inspection, binary sensor, analog sensor, interlocked control, redundant control, or vision-based verification); control directness—whether the control detects the failure mode itself or only a downstream symptom; control independence and redundancy; and observed escape evidence where available.

Unlike occurrence, which is mapped directly from observed event frequency, detection cannot always be derived from complete statistical evidence because escape data are often incomplete at the individual failure-mode level. For this reason, detection was assigned using a structured rule-based procedure rather than a purely data-derived calculation. The D rating is assigned according to the criteria defined in

Table 4. Detection rating scheme based on control architecture, directness of detection, redundancy, and observed escape evidence.

D Rating	Detection Capability	Control Characteristics
9–10	Cannot detect or very remote chance	No sensor; operator visual inspection unreliable
7–8	Remote chance of detection	Single sensor with no redundancy; marginal signal quality
5–6	Moderate chance of detection	Sensor present but no interlock; escape possible if signal marginal
3–4	High chance of detection	Sensor + PLC interlock; occasional escapes due to sensor degradation
1–2	Almost certain detection	Redundant sensors + interlock; error-proofing (poka-yoke); zero escapes observed

.

As an illustrative example, consider the bearing height measurement at FMS-202. The existing control architecture is based on a single linear measurement element that directly measures the target characteristic but operates without redundancy. Historical operational evidence shows that several bearing-related deviations escaped to the next process step. According to Table 4, this reflects poor-to-moderate detection capability and supports assignment of a high detection rating (D = 7), consistent with the rating in Table 4. Adding a redundant displacement sensor and interlocked confirmation logic would be expected to shift the same failure mode into a substantially stronger detection class.

2.3.5. Action Priority Assignment

In addition to RPN, a simplified AP classification is applied, aligned with the principles of the AIAG & VDA guideline [5] but adapted to the specific context of the present case study. The AP categories are:

• High: S ≥ 9 (regardless of O and D), or (S ≥ 7 and O ≥ 6 and D ≥ 6). Immediate action required;
• Medium: S = 5–6 with elevated O or D, or high D with moderate S. Action recommended within planning cycle;
• Low: S ≤ 4 and O ≤ 4 and D ≤ 4. Monitor but no immediate action needed.

2.4. Uncertainty and Sensitivity Analysis

Expert ratings for S, O, and D inherently involve subjectivity and inter-rater variation. To address this, the framework includes Monte Carlo uncertainty analysis.

2.4.1. Monte Carlo Simulation Procedure

Monte Carlo simulations were performed using Python 3.11 with NumPy (v1.26.4) and Matplotlib (v3.8.2).

For each failure mode, the nominal ratings (S, O and D) are treated as the center of a discrete bounded uncertainty set with ±1 rating step. For example, if nominal S = 7, the simulation samples S from {6, 7, 8} with equal probability. A total of 10,000 Monte Carlo iterations were performed. In each iteration, S, O, and D are independently sampled, and RPN is computed. Modes with >90% probability of exceeding a predefined RPN threshold under bounded rating perturbation are classified as “robustly critical” [24,25]. The ±1 step uncertainty assumption is consistent with the inter-rater variation commonly observed in FMEA practice, where a single expert rating may differ by ±1 step from consensus, making this the minimum defensible uncertainty bound. The independence assumption treats S, O, and D as conceptually separate risk dimensions. While correlations may exist in practice—e.g., a failure mode with high occurrence may have lower detection due to more frequent observation—the independent sampling represents a conservative baseline. Accordingly, the results of the Monte Carlo analysis should be interpreted as rank stability indicators under bounded discrete rating perturbation, rather than as exact failure risk probabilities.

2.4.2. Robust Prioritization Rules

Based on uncertainty analysis results, the following decision rules are applied:

• Robust Critical Rule: A failure mode is classified as “robustly critical” if P(RPN > threshold) ≥ 0.9 under ±1 rating uncertainty;
• Leverage Rule: Prioritize actions that move the highest leverage rating first. Prefer controls expected to improve O or D by ≥ 2 steps;
• Escape Risk Rule: Failure modes with high severity and high detection (poor detectability) represent escape risks and receive heightened priority.

2.5. Validation Scope

The proposed framework was validated through application to the FMS-200 flexible assembly line case study. In the present work, validation is limited to methodological applicability, internal consistency of the rating logic, traceability of the transformation from operational evidence to PFMEA outputs, and actionability of the resulting prioritization. The current validation demonstrates applicability and internal consistency but not comparative effectiveness relative to a baseline approach or a second independent case. No comparison with a conventional PFMEA conducted without data integration is provided, and no post-implementation evidence of failure or downtime reduction is available at this stage. External validation through application to additional case studies is proposed as future work.

3. Case Study: FMS-200 Flexible Assembly Line

This section describes the FMS-200 flexible assembly line used as the validation case for the proposed framework.

3.1. Line Architecture and Modular Design

The FMS-200 line is based on equipment from SMC (SMC Corporation, Tokyo, Japan), while each station is controlled by a Siemens SIMATIC S7-1500 PLC (Siemens AG, Munich, Germany). It is designed for the assembly of bearing units. The line shown in Figure 2 consists of eight modular stations (FMS-201 through FMS-208) linked by a common transfer system. Parts are mounted on pallets (carriers) that are transported between stations along a closed conveyor loop. The modular design allows stations to be added, removed, or bypassed depending on the required product configuration, supporting multiple bearing unit variants without major reconfiguration.

The system architecture follows a distributed control philosophy: each station has local PLC control for its actuators and sensors, while a supervisory SCADA system coordinates the overall process sequence and material flow [26].

An important characteristic of the FMS-200 is its inherent multi-variant capability. The assembled bearing unit consists of five component types (base, bearing, shaft, cover, and four bolts), where the shaft and the cover are each available in three variants, and the bearing is available in two variants. This yields 18 distinct product configurations (3 × 3 × 2) that the line can assemble without mechanical reconfiguration, relying only on software-controlled selection and routing logic at each station. The observation period covered production runs across all product variants, ensuring that the collected failure data reflect the variability inherent in flexible multi-variant assembly rather than a single fixed product configuration.

3.2. Station Summary and Process Flow

Table 5. Summary of FMS-200 stations, functions, and representative failure modes.

Station	Primary Function	Key Technologies	Representative Failure Modes
Transfer System	Pallet transport and positioning	Electric motor conveyors, pneumatic stoppers, positioning sensors	Pallet misalignment, stopper jam, sensor drift
FMS-201	Base part loading and orientation verification	Gravity feeder, pneumatic pusher, probe sensor, vacuum gripper	Inverted base not detected, vacuum loss, gripper misalignment
FMS-202	Bearing placement and height measurement	Gravity feeder, rotary actuator, linear encoder, two-finger gripper	Bearing height out-of-tolerance, gripper jamming, feeder empty
FMS-203	Hydraulic pressing (interference fit simulation)	Hydraulic press, force sensor, pneumatic stoppers	Press force deviation, stopper misalignment, sensor calibration drift
FMS-204	Shaft selection and orientation	Pneumatic feeders, vacuum gripper	Shaft orientation error, vacuum leak
FMS-205	Cover placement and orientation	Gravity feeder, pneumatic manipulator, probe sensor, vision verification	Inverted cover, feeder jam, manipulator positioning error, vision system false reject
FMS-206	Bolt feeding and orientation with pallet repositioning	Vibratory feeder, pneumatic orientation unit	Bolt missing, orientation failure, pallet not repositioned
FMS-207	Robotic assembly and bolt tightening	Industrial robot, electric screwdriver, torque monitoring	Bolt cross-threading, torque out-of-spec, robot positioning error
FMS-208	Automatic storage of finished assemblies	Pneumatic ejector, gravity storage rack, presence sensor	Ejector failure, storage position occupied, sensor malfunction

provides a summary of the eight stations and the transfer system.

3.3. Data Collection Procedure

In total, 186 failure events were documented during the observation period of 4 weeks, based on a 5-day working week and 8 h shifts, corresponding to a production volume of approximately 2743 assembly cycles, with a mean cycle time of 3.5 min per unit. Each event was reviewed by production and maintenance personnel to assign the cause category and verify the actuation technology. In addition to failure event data, production volume and cycle time data, maintenance logs, scrap and rework records, and sensor calibration histories were collected. The 186 failure events were collected across production runs involving all product configurations from the 18-variant product family, ensuring that the failure data capture variant-dependent effects such as differential gripping forces for different shaft geometries, height-dependent positioning tolerances for bearing variants, and color-specific vision verification challenges. This multi-variant data collection strengthens the representativeness of the PFMEA analysis beyond what a single-product observation would provide.

3.4. Application of the Framework to FMS-200

The five-stage framework was applied to the FMS-200 operational dataset. Stage 1 involved process flow chart construction, mapping nine process steps. Stage 2 classified the 186 failure events by process step and actuation type. Stage 3 constructed cause-and-effect diagrams for the top three failure categories. Stage 4 developed a PFMEA with 40 distinct failure modes. Stage 5 aggregated risk by station and actuation type and performed a Monte Carlo simulation with 10,000 iterations.

4. Results

This section presents the results of applying the integrated framework to the FMS-200 assembly line.

4.1. Classical Quality Tool Outputs

4.1.1. Pareto Analysis by Process Step

Pareto analysis of the 186 failure events by process step shown in Figure 3 revealed that four entities accounted for 67% of total failures.

• Transfer System: 17 events—pallet positioning, stopper jams, sensor drift. Typical failure of the transfer system is shown in Figure 4.
• FMS-201 (Base loading): 18 events—inverted base detection, vacuum gripper failures.
• FMS-202 (Bearing placement): 18 events—bearing height out-of-size, gripper jamming.
• FMS-206 (Bolt feeding): 17 events—bolt missing, orientation failure, pallet repositioning errors.

Consequently, improvement efforts should primarily target the transfer system and the first two assembly stations (which manage the most complex orientation tasks). The 67% concentration in only four of nine process entities follows the Pareto principle and suggests that targeted interventions on these stations would yield disproportionate improvements in overall line reliability. Notably, FMS-206 (bolt feeding) also emerged as critical despite being late in the sequence, confirming that small-part handling and multi-axis orientation remain inherently challenging operations in pneumatic-based flexible assembly.

4.1.2. Pareto Analysis by Actuation Type

Analysis by actuation technology, illustrated in Figure 5, showed that pneumatic systems dominated the failure profile. Pneumatic cylinders accounted for 135 events (72.6%), followed by electric sensors combined with pneumatic cylinders at 20 events (10.8%) and pneumatic grippers at 12 events (6.5%). The remaining events were associated with electric drives (8; 4.3%), sensing elements (7; 3.8%), and hydraulic systems (4; 2.2%). Overall, pneumatic-related failures accounted for 167 of 186 events (89.8%), which is consistent with the FMS-200 architecture.

Given that pneumatic components drive nearly 90% of failures, enterprise-level improvements (such as air quality control and valve standardization) will outperform localized station interventions. The result also suggests that future system designs should consider reducing reliance on pneumatic actuation in favor of electric servo-driven alternatives for critical positioning and gripping functions, where higher repeatability and lower sensitivity to air quality can be achieved.

4.1.3. Cause-And-Effect Analysis

The constructed Ishikawa diagram shown in Figure 6 confirmed that the machine and method categories dominate.

4.2. PFMEA Findings

4.2.1. PFMEA Structure and Scope

The completed PFMEA comprises 40 distinct failure modes distributed across the nine process steps. Severity ratings range from 5 to 9, occurrence ratings from 4 to 6, and detection ratings from 3 to 7.

4.2.2. RPN Distribution and High-Priority Modes

Across the 186 documented failure events, the resulting RPN values range from 60 to 378, as shown in Figure 7.

The S × O heatmap reveals that the majority of failure modes cluster in the moderate-to-high severity range (S = 5–7), with moderate occurrence (O = 5–6), while a critical outlier exists at S = 9 (bolt tightening at FMS-207). The S × D heatmap shows a concentration of modes in the high-detection region (D = 6–7), indicating that current controls across the FMS-200 line are generally insufficient for reliable failure interception. This dual pattern confirms that improving detection capability represents a high-leverage opportunity across most stations, not only for individual high-RPN modes.

Figure 8 confirms the distribution of RPN values is right-skewed, with a pronounced concentration of failure events in the range of approximately 150–200 and a secondary concentration around higher RPN values. The mean (182.4) exceeds the median (168), indicating the presence of a limited number of high-risk failure modes that disproportionately influence the overall risk profile. The distribution also exhibits a slight bimodal tendency, suggesting the existence of distinct groups of failure modes with different risk characteristics. The presence of RPN values above 300 indicates a small number of high-risk failure modes that require prioritized mitigation actions.

The top-ranked failure modes, as shown in

Table 6. Top 5 failure modes ranked by RPN (condensed view).

Rank	Process Step	Failure Mode	S	O × D	RPN
1	FMS-207	Bolt cross-threading or insufficient torque	9	6 × 7	378
2	FMS-206	Bolt does not present in gripper	7	6 × 7	294
3	FMS-206	Bolt orientation incorrect	7	6 × 7	294
4	Transfer	Pallet positioning error at station	6	6 × 7	252
5	FMS-202	Bearing height out of size	6	6 × 7	252

, reveal a clear pattern: bolt assembly operations (FMS-206 and FMS-207) dominate the highest-risk positions due to the combination of moderate-to-high severity with consistently poor detection (D = 7). All top five modes share the same O × D product (6 × 7 = 42), differing only in severity, which underscores that the detection gap is a systemic issue rather than a station-specific problem. This suggests that improving detection across the line—particularly for pneumatic positioning and fastening operations—would provide the most efficient path to aggregate risk reduction.

4.2.3. Action Priority Distribution

The AP-aligned classification within the proposed framework, shown in Figure 9, resulted in the following distribution: high: 34 failure modes (18.3%)—immediate action required; medium: 63 failure modes (33.9%)—action recommended within planning cycle; low: 89 failure modes (47.8%)—monitoring required but no immediate action needed.

The AP distribution indicates that high-priority failure modes are primarily concentrated in the transfer system, consistent with the RPN-based ranking. However, the AP classification also elevates certain failure modes with high severity (S ≥ 9) that would not be prioritized based on RPN alone, demonstrating the value of the AP-aligned classification logic as a qualitative safeguard against under-prioritizing severe or weakly detectable failure modes. The predominance of low-priority modes (47.8%) suggests that a substantial portion of the identified risks is adequately controlled, allowing improvement efforts to be focused on the 52.2% of failure modes requiring corrective actions.

4.3. Risk Aggregation by Station and Technology

4.3.1. Station-Level Risk Profile

Aggregating RPN values by process step, as represented in Figure 10, provides a station-level risk profile. The transfer system has the highest aggregated risk (Σ RPN = 761), followed by FMS-207 (598, driven by the bolt tightening failure with RPN = 378), FMS-201 (512), and FMS-202 (488).

The station-level aggregation reveals that the transfer system and FMS-207 together account for over 35% of the total aggregated risk, despite representing only two of nine process entities. This concentration indicates that shared infrastructure (pallet transport and positioning) and the final critical assembly operation (bolt tightening) are the primary contributors to overall system risk.

4.3.2. Technology-Level Risk Profile

Aggregating RPN values by actuation technology presented in Figure 11 confirms the dominance of pneumatic systems. Pneumatic cylinders account for the largest share of the aggregated risk (Σ RPN ≈ 2835; 67.8% of total), followed by electro-pneumatic systems combining sensing and actuation functions (Σ RPN ≈ 620; 14.8%). Grippers contribute an additional Σ RPN ≈ 405 (9.7%), while hydraulic systems (Σ RPN ≈ 210; 5.0%) and purely electric systems (e.g., spindles and motors) represent comparatively smaller portions of the total risk. Sensing elements, including photoelectric sensors and encoders, contribute only a minor share of the aggregated RPN.

Pneumatic-related technologies account for 92.3% of the overall risk. Therefore, enhancing pneumatic component reliability and air quality management offers the highest potential for aggregate risk reduction.

4.4. Uncertainty Analysis and Rank Stability

4.4.1. Monte Carlo Simulation Results

Figure 12 shows the results of the Monte Carlo uncertainty analysis for the top 15 failure modes based on 10,000 iterations. For the highest ranked failure mode (FMS-207 bolt tightening, nominal RPN = 378 from S = 9, O = 6, D = 7), the simulated mean RPN (377.9) closely matches the nominal value (378), standard deviation = 52.3, 95% confidence interval [285, 480], P(RPN > 300) = 99.3%, and P(remains #1 rank) = 94.7%.

Across all 15 top failure modes, the Monte Carlo analysis demonstrated that rank positions are highly stable: the top five modes maintained their relative ordering in over 90% of simulation iterations. This rank stability is illustrated in Figure 13 and confirms that the deterministic RPN-based prioritization is robust against reasonable rating uncertainty (±1 step) and that resource allocation decisions based on the current ratings can proceed with confidence. Lower-ranked modes (positions 10–15) showed greater rank interchange probability, suggesting that borderline cases should be monitored for additional data before committing resources.

To assess the robustness of the ±1 step assumption, a supplementary sensitivity analysis was conducted with ±2 rating steps. Under ±2 uncertainty, the highest-ranked mode (FMS-207 bolt tightening) maintained its #1 position in 22.3% of iterations (vs. 42.4% under ±1) and remained within the top 5 in 64.7% of iterations (vs. 83.5% under ±1). Modes ranked 2–8 (nominal RPN = 294) showed P(top 5) ≈ 0.39 under ±2 vs. ≈ 0.44 under ±1. These results confirm that FMS-207 bolt tightening remains the dominant risk contributor even under substantially expanded uncertainty bounds.

Table 7. Sensitivity analysis summary: rank stability under ±1 and ±2 rating perturbation for the top 5 failure modes.

Failure Mode	Nominal RPN	±1: Mean RPN	±1: P(Top 5)	±2: Mean RPN	±2: P(Top 5)
FMS-207 bolt tightening	378	377.9	83.5%	371.8	64.7%
FMS-201 orientation check	294	293.1	44.0%	294.1	39.3%
FMS-202 extractor failure	294	294.3	45.1%	294.4	39.5%
FMS-203 press positioning	294	293.9	45.1%	292.5	38.7%
Transport pallet positioning	294	293.9	44.5%	296.8	39.9%
Interpretation	Rank order stable	Top mode ≥ 83%	Rank order weaker	Top mode ≥ 64%	Rank order stable

summarizes the sensitivity analysis results across the two perturbation scenarios.

4.4.2. Escape Risk Identification

Escape risks, presented in Figure 14, were identified as failure modes combining high severity and poor detection capability using the criterion S ≥ 7 and D ≥ 7.

The identified escape risk modes are primarily associated with bolt assembly operations (FMS-206 and FMS-207) and transfer system positioning. These failure modes combine high severity, indicating potential functional or safety impact, with insufficient detection capability due to reliance on single-sensor controls.

Importantly, these risks are not necessarily captured by RPN ranking alone. Even failure modes with moderate RPN values can pose critical escape risk if detection effectiveness is low. These escape risk modes demand immediate implementation of redundant detection mechanisms (e.g., torque angle monitoring), superseding their standard RPN rank.

5. Discussion

5.1. Mapping Dominant Risk to Physical Mechanisms

The PFMEA results indicate that approximately 90% of aggregated risk is associated with the transfer system and pneumatic positioning and handling functions (pallet buffers, stoppers, pick-and-place axes, and vacuum gripping). This finding is consistent with the FMS-200 hardware architecture described in the technical manual [26], where each station relies on multiple pneumatic cylinders with reed sensors for end-of-stroke confirmation and vacuum pressure switches for part presence confirmation.

Pneumatic systems are sensitive to several physical degradation mechanisms: pressure instability, leakage, valve sticking, seal wear, and contamination. In addition, failure detectability may degrade when sensors are marginally aligned or when PLC logic treats borderline signals as acceptable.

The high RPN for bolt tightening failures (RPN = 378, driven by S = 9 and D = 7) reflects a different mechanism: cross-threading or insufficient torque can occur due to bolt misalignment, worn gripper jaws, or electric screwdriver parameter drift. The high severity (S = 9) is justified because an improperly tightened bearing assembly can fail in service, presenting a safety risk.

Rather than a simple frequency effect, the dominance of pneumatic failures highlights a complex coupling mechanism where pneumatic degradation interacts with mechanical tolerance stack-ups, sensor margins, and control timing. Pressure fluctuation, leakage, valve sticking, and seal wear can alter actuator stroke speed and end position repeatability. Once end position consistency is degraded, binary sensor confirmation may become marginal, which in turn increases the probability of timing mismatch, false part presence confirmation, pallet mispositioning, or incomplete transfer engagement. In bolt-feeding and positioning operations, such coupling can propagate further by misaligning the fastener, degrading insertion quality, or increasing the likelihood of cross-threading or insufficient torque establishment. This chain explains why failures that initially appear to belong to separate categories—pneumatic actuation, sensing, transfer, and fastening—can in fact share a common physical and control-level origin. The coupling also implies that simultaneous degradation in occurrence and detection has a multiplicative effect within the RPN structure: a simultaneous one-step increase in both O and D raises RPN by approximately 40–60% rather than 20–30%, and preventive maintenance should target both failure prevention (reducing O) and detection preservation (preventing D degradation).

5.2. Advantages of the Proposed Framework

The contribution of the proposed framework becomes clearer when compared with both conventional PFMEA practice and the recent enhanced FMEA literature. Compared with conventional PFMEA prepared primarily through expert workshops, the present approach provides stronger traceability from observed manufacturing evidence to PFMEA entries, clearer linkage between quality tool outputs and failure-mode definition, and a multi-level aggregation mechanism that supports station-level and technology-level prioritization. Compared with many advanced FMEA extensions in the literature, the contribution here is less about introducing a new mathematical ranking operator and more about offering a practically deployable integration architecture for flexible manufacturing, with bounded uncertainty robustness checks and explicit operational grounding. The framework therefore complements, rather than competes with, fuzzy, grey, Bayesian, or machine learning-based FMEA variants. Its methodological value lies in structured integration, traceability, and deployability in industrial environments where operational data, classical quality tools, and PFMEA are often present but disconnected.

The main practical advantages of the proposed framework comprise four aspects: it integrates classical quality tools with PFMEA rather than applying them sequentially; it supports risk aggregation at the station and technology level for strategic resource allocation; it incorporates bounded uncertainty analysis through Monte Carlo simulation; and it combines RPN, AP-aligned classification, and escape risk identification into a multi-criteria prioritization logic.

Without data integration, a conventional PFMEA would likely estimate occurrence ratings for the transfer system at O = 3–4 (based on general engineering judgment), ignoring the empirical O = 6 demonstrated here. This difference represents a 50%+ reduction in estimated RPN, illustrating the practical risk of underestimation without data integration.

The proposed approach shares conceptual parallels with recent developments in smart manufacturing process monitoring. An Artificial Intelligence of Things (AIoT) framework for composites manufacturing that integrates sensor arrays, IoT-based data platforms, and AI-based process forecasting is proposed [27]. While their framework addresses a different manufacturing domain and relies on continuous sensor data rather than discrete event logs, both approaches share the principle that structured integration of data acquisition, processing, and analytical tools yields stronger process control than any individual tool in isolation. Future convergence of these approaches—where real-time sensor-based anomaly detection feeds directly into a living PFMEA document—represents a promising direction for Industry 4.0 quality management.

Table 8. Comparative positioning of the proposed framework against conventional PFMEA and FMEA 4.0 approaches with machine learning.

Criterion	Conventional PFMEA	Proposed Framework	FMEA 4.0 (ML-Based)
Rating traceability	Expert judgment only; no audit trail	Data-linked O; documented S, D basis	Sensor-derived; requires training data
Expert effort	High (all ratings subjective)	Moderate (O data-driven; S, D partially)	Low after setup; high initial development
Adaptability	Manual reevaluation	Structured update via check sheets/Pareto	Automatic with retraining; model drift risk
Uncertainty quantification	None (deterministic RPN)	Monte Carlo ±1 step; rank stability	Probabilistic outputs; confidence intervals
Implementation complexity	Low; spreadsheet-based	Moderate; data infrastructure needed	High; ML expertise + sensor integration

summarizes key distinctions between the proposed approach and alternative approaches.

5.3. Action Strategy: From PFMEA Outputs to Engineered Controls

Based on the PFMEA results and uncertainty analysis, a systematic action strategy is developed. Actions are prioritized using a leverage-based rule, favoring control measures capable of improving occurrence or detection ratings by at least two levels, as these provide the most efficient reduction of RPN without requiring design-level changes to severity.

Figure 15 illustrates the projected scenario-based impact of the proposed actions under the assumed rating improvements. The largest relative RPN reductions are observed for FMS-206 (−64%) and FMS-202 (−57%), where improvements are primarily driven by enhanced detection and process stabilization. In contrast, FMS-207—despite having the highest initial RPN—achieves a more moderate reduction (−43%), reflecting the limited potential for reducing severity in safety-critical fastening operations.

The scenario analysis suggests that detection-focused interventions, such as sensor upgrades and vision-based verification, may yield higher percentage reductions than occurrence-focused measures alone. Within the analyzed system and rating structure, these results suggest that improving detection capability is the most effective near-term strategy for mitigating high-risk failure modes.

These findings indicate that the most efficient risk reduction is achieved not through fundamental redesign but through targeted improvements in monitoring, feedback, and control redundancy. This approach enables substantial risk mitigation with minimal disruption to the existing system architecture.

5.3.1. Immediate Actions (High Priority, Target: 0–3 Months)

The following immediate actions target four critical failure modes selected on the basis of a combined criterion including high AP classification, elevated RPN, and escape risk relevance. This combined prioritization logic was adopted because some failure modes warrant immediate intervention not only due to their nominal RPN value but also due to their severity and limited detectability under current control conditions:

• FMS-207 Bolt Tightening (RPN = 378): Implement torque angle signature monitoring. Projected scenario: D 7→4, projected post-action RPN = 216 (−43%);
• FMS-206 Bolt Feeding (RPN = 290): Add vision-based bolt presence confirmation. Projected scenario: D 7→3, O 6→5, projected post-action RPN = 105 (−64%);
• Transfer System Positioning (RPN = 250): Upgrade to high-resolution proximity sensors with analog output. Projected scenario: D 7→4, O 6→5, projected post-action RPN = 120 (−52%);
• FMS-202 Bearing Height Measurement (RPN = 245): Implement redundant laser displacement sensor. Projected scenario: D 7→3, projected post-action RPN = 108 (−57%).

5.3.2. Near-Term Actions (Medium Priority, Target: 3–6 Months)

Near-term actions include pneumatic system preventive maintenance enhancement, which is projected to reduce the RPN of a broad subset of pneumatic-related modes under the assumed control-improvement scenario, as well as FMS-201/202 feeder jam reduction through dimensional verification at feeder input, and PLC interlock enhancement for all sensor inputs.

5.3.3. Long-Term Actions (Low Priority, Target: 6–12 Months)

Long-term actions include digital twin integration for predictive maintenance and establishing a quarterly Reverse FMEA update protocol. Projected long-term effect under the assumed improvement pathway: a substantial reduction in average RPN and in the number of high-priority modes could be achieved under the assumed improvement pathway; however, these values are engineering projections, not experimentally validated outcomes [28,29].

A common concern with single-case validation is whether the results generalize beyond the specific product and line configuration studied. In the present case, this concern is partially mitigated by two factors. First, the FMS-200 line was operated with 18 distinct product configurations during the observation period, varying shaft color, cover color, and bearing height [23]. The 186 failure events therefore reflect multi-variant production conditions, not a single fixed assembly sequence. Second, the modular architecture of FMS-200 allows stations to be added, removed, or bypassed depending on the product variant, and the PFMEA framework accommodates this through modular failure-mode sections: each product variant maintains its own set of variant-specific failure modes (e.g., height-dependent bearing positioning at FMS-202), while shared infrastructure (transfer system, bolt feeding at FMS-206) retains common PFMEA entries. This multi-variant operational context means that the framework has been implicitly stress-tested across product diversity within a single line, although formal cross-line validation remains a priority for future work.

To illustrate concretely, consider two representative product variants: Variant 1 uses the full eight-station sequence, while Variant 2 bypasses FMS-203 (hydraulic press not required for this variant) and uses a different shaft orientation sequence at FMS-204. Under the proposed framework, common failure modes (e.g., transfer system pallet positioning, FMS-206 bolt feeding) retain identical PFMEA entries across both variants, while variant-specific failure modes are activated or deactivated: FMS-203 press-related modes are excluded from the Variant 2 PFMEA, and FMS-204 shaft orientation modes are re-evaluated with variant-specific occurrence data reflecting the different handling geometry. Station-level aggregated RPN is then recomputed for each variant, enabling production managers to identify whether risk concentration shifts across the line when switching between product families.

5.4. Comparison with Conventional Expert-Only PFMEA

To contextualize the contribution of the proposed framework, this subsection provides a structured comparison with a conventional expert-only PFMEA applied to the same FMS-200 case. In a conventional approach, a cross-functional team would assign S, O, and D ratings based on engineering experience and general equipment knowledge, without systematic integration of operational failure logs, Pareto analysis, or Monte Carlo robustness checks.

For the transfer system, a conventional PFMEA would likely estimate occurrence at O = 3–4 based on the general perception that pallet transport is a mature, well-understood function. The data-driven approach yields O = 5–6 based on 17 documented events over 2743 cycles (λ = 6.2 per 1000 cycles), a difference that shifts the transfer system from medium to high priority. Similarly, detection ratings for bolt-feeding operations at FMS-206 would likely be estimated at D = 4–5 under the assumption that PLC interlocks provide adequate coverage, whereas operational escape data reveal D = 7, indicating that existing controls fail to intercept a significant fraction of misaligned or missing bolts. Because RPN is multiplicative, these differences compound in the final ranking, yielding risk estimates 50–100% higher under the data-driven approach than under conventional expert-only assessment for the most critical failure modes.

This comparison is reconstructed analytically rather than derived from a controlled parallel study. A rigorous baseline comparison would require conducting both approaches independently on the same case and comparing their outputs systematically. Nevertheless, the directional differences observed here are consistent with the known tendency of expert-only PFMEA to underestimate occurrence for familiar processes and to overestimate detection capability for sensor-based controls, supporting the practical value of data integration. This comparison should be interpreted as an analytical reconstruction intended to highlight methodological differences, not as a controlled empirical benchmark study.

5.5. Practical Implications

The results carry practical implications across several roles. Station-level risk aggregation gives production managers a direct basis for resource allocation, while the integration of quality tools with PFMEA removes the wall between day-to-day troubleshooting and formal risk documentation. The finding that 90% of risk traces to pneumatic subsystems is relevant for designers of future FAALs. More broadly, the approach shows that data-driven quality management can be adopted incrementally, without replacing existing systems wholesale.

5.6. Limitations and Threats to Validity

This study has several limitations:

• Single-line validation—although the FMS-200 case covers 18 product variants assembled on the same line [23], cross-line external validation through additional case studies on different Flexible manufacturing system (FMS) platforms and industries is needed to establish broader generalizability;
• Data quality and completeness—minor or transient failures may not have been recorded;
• Expert judgment subjectivity—inter-rater reliability was not formally assessed;
• Validation scope—primarily internal validation without controlled experiments;
• Temporal stability—the PFMEA snapshot represents a specific observation period;
• Generalizability—the framework does not address integration with Fault tree analysis (FTA), Reliability block diagram (RBD), or advanced Statistical process control (SPC) methods.

6. Conclusions and Future Work

6.1. Summary of Findings and Contributions

• The proposed framework demonstrates that operational failure data, classical quality tools, and PFMEA can be integrated into a single traceable workflow for flexible assembly systems. This conclusion is supported by application to 186 failure events across 40 failure modes collected over 2743 assembly cycles involving 18 distinct product configurations on the FMS-200 line [23]. The multi-variant nature of the validation data strengthens confidence that the framework is applicable where event-level failure logging and process-step decomposition are available, even under product variability;
• In the studied FMS-200 line, the dominant risk concentration was associated with transfer and pneumatic subsystems (approximately 90% of aggregated RPN), indicating that local actuator and positioning weaknesses can govern system-level risk. The quantitative magnitude is case-specific, but the analytical logic is transferable to other modular assembly systems;
• The framework provides added value beyond conventional PFMEA by linking shop floor evidence to PFMEA entries and by aggregating risk at failure-mode, station, and technology levels. This is especially relevant in flexible manufacturing where isolated failure analysis is insufficient for strategic prioritization;
• Detection-related weaknesses emerged as a major leverage point for risk reduction in the most critical failure modes. Monitoring, interlocking, redundancy, and direct verification controls should be prioritized before more costly design changes whenever the failure mechanism allows this;
• The uncertainty analysis indicates that the most critical failure modes remain high-priority under plausible bounded rating variation (±1 step: top mode retains #1 in 94.7% of iterations; ±2 step: retains top 5 in 64.7%), supporting robust prioritization;
• The present study demonstrates methodological applicability and decision-support relevance, but not full post-implementation effectiveness. Broader effectiveness claims require longitudinal before–after evidence, baseline comparison, and additional case studies across other flexible manufacturing configurations;
• Future work should extend the framework toward multi-case validation, reconfigurable routing scenarios, stronger dependency-aware uncertainty models, and deeper integration with predictive maintenance, digital twins, and complementary reliability methods.

6.2. Practical Implications and Recommendations

For practitioners, the results of this study highlight several key directions for effective PFMEA implementation. First, the establishment of a structured data infrastructure is essential, as data-driven analysis requires consistent and reliable failure data collection. Second, systematic application of root cause analysis—particularly through the 6M framework—supports a deeper understanding of failure mechanisms beyond surface-level symptoms.

In addition, aggregating risk across stations and technologies enables more informed strategic decision-making at the system level, rather than focusing solely on individual failure modes. Incorporating uncertainty analysis further strengthens prioritization by ensuring that decisions remain robust under rating variability. Finally, the integration of Reverse FMEA as a continuous governance process supports the validation of assumptions and helps maintain alignment between design intent and actual system behavior.

6.3. Future Research Directions

Future research directions include the following priorities. First, multi-case validation: although the present study covers multiple product variants on the same FMS line, the framework should next be validated across additional FMS platforms and industrial contexts. Second, post-implementation measurement of actual downtime reduction following corrective actions. Third, refinement of the occurrence rating to use failures per assembly cycle as the primary metric, and systematic escape data collection for detection ratings. Additional directions include digital twin integration, multi-method integration with FTA and RBD, inter-rater reliability studies, and advanced uncertainty methods.

6.4. Concluding Remarks

This study showed that classical quality tools and PFMEA can be linked through a structured data pipeline in flexible assembly, yielding risk ratings that are grounded in operational evidence rather than expert intuition alone. On the FMS-200 line, the approach identified pneumatic subsystems as the dominant risk contributor (92% of aggregated RPN) and confirmed that the top-ranked failure modes remain stable under bounded rating uncertainty. These results provide a starting point for Reverse FMEA protocols and eventual digital twin integration in modular manufacturing.