Application of Machine Learning for Predictive and Prognostic Reliability in Flexible Shop Floor

Abstract

Flexible workshops are essential components of modern industry, enabling flexible and efficient production. However, to ensure their proper functioning and prevent unexpected breakdowns, it is crucial to monitor their reliability. Production stoppages caused by unforeseen breakdowns can lead to significant financial losses. This chapter proposes to explore the use of Machine Learning (ML) for predicting the reliability of flexible workshops, thus identifying dates for Preventive Maintenance (PM) interventions and optimizing production management. The objectives of this exploration include the presentation of new predictive model developments and the description of ML models capable of predicting workshop reliability based on real-time data, such as equipment monitoring, production data, and maintenance histories. It also aims to identify optimal times for PM interventions, minimizing production disruptions and optimizing resource utilization. Additionally, the chapter will propose cost optimization models to prevent unplanned breakdowns, extend equipment lifespan, optimize spare parts usage, and maximize productivity by avoiding production interruptions and ensuring the smooth operation of the flexible workshop.

1. Introduction

In today’s economic landscape, characterized by intense globalization and progressively exacting markets, industries are compelled to enhance the efficacy and efficiency of their production lines. This drive aims to bolster their competitiveness and meet the escalating demands of their clientele. The amalgamation of connectivity, data utilization, emerging technologies, inventory streamlining, personalized manufacturing, and regulated production has ushered in the era known as Industry 4.0, which appears to be an unstoppable force shaping the industrial realm.

Industry 4.0 has ushered in a new era of advanced technologies including Artificial Intelligence (AI), Internet of Things (IoT), Cyber-Physical Production Systems (CPPS), and Big Data and Analytics [1, 2, 3]. This technological evolution is transforming processes, expertise, business models, and the maintenance of manufacturing operations. Maintenance, crucial in ensuring equipment functions effectively, has become more vital due to the high value of manufacturing assets and their propensity to degrade over time. The maintenance department’s reliability is critical in boosting productivity. As noted by Parida and Chattopadhyay [4], evaluating maintenance performance is increasingly essential. Several studies emphasize maintenance’s role in enhancing efficiency and economic sustainability in manufacturing [5, 6, 7]. Maintenance is recognized as a key driver for competitive advantage in the industry [8].

Recent advancements in technology, especially in smart resources, have facilitated real-time data collection and integration, aiming to improve reliability, lower repair costs, and predict equipment degradation. Predictive maintenance, in particular, is pivotal in reducing unexpected downtime in manufacturing. Research in this field investigates the impact of Big Data and Machine Learning (ML) techniques on industrial maintenance [9]. AI and ML are extensively used to monitor manufacturing processes and forecast downtimes. Leukel et al. [10] explore ML integration for predicting machine breakdowns, proposing a decision-support framework to minimize downtime in CPPS [11]. Chaudhuri [12] suggests a Hierarchical Modified Fuzzy Support Vector Machine (HMFSVM) for predicting vehicle failure trends.

Kamariotis et al. [13] focus on Prognostic Health Management (PHM) to predict the Remaining Useful Life (RUL) of equipment, introducing a metric for evaluating data-driven prognostic algorithms in Predictive Maintenance (PdM) decisions. Similarly, Khazaelpour and Hashemkhani Zolfani demonstrate the effectiveness of an AI-based predictive maintenance model, the FUCOM-ANN hybrid approach, which outperforms standard Artificial Neuronal Network (ANN) predictions in reliability and cost optimization [14].

The literature also presents various predictive maintenance models suited to different constraints in health assessment contexts. Elkateb et al. [15] describe using AdaBoost ML algorithms for real-time classification of machine stops in knitting machines, achieving a 92% accuracy rate. Additionally, a multi-agent framework, SCMEP, is introduced for integrating production and predictive maintenance scheduling [16]. Zonta et al. [17] propose a model using deep neural networks for optimizing production scheduling. One significant limitation of existing predictive models is the absence of benchmark datasets and standardized evaluation metrics. The presence of such benchmark datasets and standardized evaluation metrics is pivotal to ensuring impartial and consistent comparisons among various predictive maintenance models.

However, the literature also identifies gaps, notably the lack of comprehensive comparisons among various ML algorithms in PdM. Studies often focus on a single algorithm without robust comparisons to others. A holistic evaluation of different algorithms would provide insights into their respective advantages and challenges in various PdM scenarios.

Real-world PdM implementation faces complexities such as varying operating conditions and data inconsistencies. Unfortunately, some studies overlook these challenges in ML model design and evaluation. This study aims to address these real-world challenges to enhance the effectiveness and robustness of PdM models. A significant limitation in current predictive models is the lack of benchmark datasets and standardized evaluation metrics, which are crucial for fair and consistent comparisons across different PdM models.

The rest of this paper is organized as follows. Section 2 introduces the problem statement. In Section 3, we elaborate on the materials and methods, including the proposed method, data analysis and processing procedures, and the developed predictive models using ML techniques. Section 4 presents the results and discussions. Finally, Section 5 offers conclusions and perspectives.

2. Problem statement

This chapter revolves around the operational health of an automated robot responsible for packaging mechanical products in a flexible manufacturing system.

Figure 1 depicts a computer-based simulation model of the dedicated shop floor. This latter is designed to efficiently perform product loads with few settings independently of the production line. However, the factory is grappling with production capacity challenges due to difficulties in activating the conditioning unit. Consequently, the article explores a case study centered on the packaging robot, specifically predicting the degradation of its power transmitters under demanding mechanical and thermal conditions, aiming to identify optimal times for preventive maintenance interventions. Notably, the conditioning robot is equipped with specialized sensors and advanced technology and connected to a local area network, transforming the workshop into a CPPS. These probes measure parameters affected by changes and interruptions in the product packaging process.

Figure 1.

Flexible shop floor facility layer [3].

3. Materials and methods

3.1 Suggested method

This paper delves into a practical case study involving a factory specializing in mechanical products manufacturing. Our main emphasis is on deploying PdM for a robot responsible for packaging final products, particularly forecasting the deterioration of its power transmitter. Our goal is to introduce and compare two predictive maintenance models within the framework of smart era. To accomplish this, we examine and discuss the importance of failure prognosis. Consequently, we develop two predictive models using supervised ML techniques and another one using unsupervised ML. These models forecast the gradual decline of the robot and anticipate its future conditions, aiding maintenance personnel in making well-informed decisions regarding PM interventions. Following this, we will offer the rationale behind the selected predictive models and conduct an in-depth comparative analysis between them.

The initial predictive model relies on a discrete Bayesian filter (DBF), which has been identified as highly suitable for addressing this specific problem [18, 19, 20]. A primary strength of the DBF lies in its adeptness at integrating information from various sources such as processes, configuration variables, and sensor data. Moreover, the DBF effectively manages uncertainties inherent in noisy processes and sensor measurements. Its adaptability to assimilate new information and respond to changing operational conditions allows for swift real-time decision-making within short operating periods. In the context of Industry 4.0, the DBF is well-matched for predictive maintenance, which is crucial for prompt issue identification and suitable corrective measures.

The second model suggested is a predictive strategy based on the Naïve Bayes Filter (NBF), a probabilistic classification model extensively used in machine learning. The NBF exhibits various advantages that render it appropriate for our context. These advantages encompass computational efficiency, adeptness in managing large datasets, resilience to irrelevant features, and ease of implementation. In summary, the Naïve Bayes filter emerges as a practical option for predictive maintenance in our scenario.

Furthermore, an additional model was established utilizing Isolation Forest, an unsupervised learning technique. This approach is particularly effective in identifying anomalies or outliers within a dataset. By leveraging the Isolation Forest method, this model excels in detecting deviations or irregularities that may signify potential issues or anomalies within the operational system.

The primary objectives of this paper are delineated as follows:

• Conducted analysis and structuring of massive data

• Crafting two distinct predictive maintenance models utilizing supervised ML techniques, namely DBF and NBF, and another model based on unsupervised ML: Isolation forest.

The flowchart representing the proposed method in this paper is presented in Figure 2.

Figure 2.

Proposed method’s flowchart.

3.2 Theoretical study

In simpler terms, the suggested PdM models focus on forecasting issues related to the power transmitter of a conditioning robot, as shown in Figure 3.

Figure 3.

Packaging Robot’s power transmitter.

The primary aim is to scrutinize and comprehend the distinctive attributes of these power transmitters, which are detailed in the subsequent subsection. To facilitate analysis, we have compiled a comprehensive list of parameters pertaining to the power transmitter, specifically the belt-pulley system, presented in Table 1. Notably, these values are indicative of a new power transmitter. Understanding these parameters allows us to gain insights into the functionality of the power transmitter, analyze extensive data, and detect any deviations from normal values during the predictive maintenance process.

Parameter	Signification
ND	Large pulley speed
Nd	Small pulley speed
WD	Large pulley angular speed
Wd	Small pulley angular speed
D	Large pulley diameter
d	Small pulley diameter
C	Torque
a	Interaxis
V	Speed
$\theta$D	Large pulley winding angle
$\theta$d	Small pulley winding angle
L	Belt length
T	Stretched strand tension
t	Soft strand tension
T0	Initial tension
f	Friction Coefficient
P	Admissible power

Table 1.

Related parameters of the robot’s power transmitter.

Indeed, the predictive maintenance models developed within the presented research offer insight into the health and performance of the power transmitter in the packaging robot. These models are tailored to oversee and uphold the power transmitter functionality within the packaging robot, utilizing sophisticated algorithms and ML techniques to anticipate and detect potential issues before they escalate into equipment failures. Through the implementation of these models, our objective is to enhance the reliability and efficiency of the packaging robot, enabling us to identify optimal times for PM interventions, minimizing production disruptions, and optimizing resource utilization.

In simpler terms, the packaging robot incorporates a dedicated set of smart and embedded sensors, which are interconnected via a local area network, transforming the workshop into a CPPS. The most important function involves real-time measurement of parameters susceptible to alterations and disruptions in the conditioning process. Throughout the packaging operation, these sensors collect data, which is securely transmitted to a centralized server.

In the subsequent section, our aim is to process the sensor-derived data into a format suitable for storage in a Database. Subsequently, we will employ ML techniques to construct predictive models. These models will evaluate the current degradation status of the power transmitter in the robot and forecast its future performance. Access to real-time degradation updates and future behavior predictions allows us to create a comprehensive maintenance plan, enhancing the reliability and efficiency of our maintenance interventions.

It is crucial to emphasize that our research is specifically dedicated to implementing predictive maintenance strategies for the conditioning robot, with a focal point on its power transmission system.

The following section delves into the specifics of the data examined.

3.3 Process of analyzing data

3.3.1 Corresponding data

In this section, our objective is to analyze a data frame covering the period from 2020 to 2022, including records of 2900 packaging processes conducted during this timeframe. These measurements are stored in files, each containing values for 15 parameters (as outlined in Table 1). Due to the substantial volume of data, with 43 k variables, it is essential to organize this information efficiently. To achieve this, we will use representative descriptors, elaborated upon in the following sections.

By referring to the literature, Ruiz-Sarmiento, and his colleagues have introduced a descriptor as a concise representation of specific data elements that elucidates and retains information essential for addressing a particular problem [18] and will contain information concerning the conditioning robot’s health for our case study.

The forthcoming data analysis will proceed through four distinct steps aimed at identifying pertinent descriptors capable of effectively summarizing data and offering insights for constructing predictive models for the packaging robot’s power transmitter. The entire data analysis process is depicted in Figure 4.

Figure 4.

Flowchart depicting the successive steps of data analysis designed to extract the most relevant descriptors summarizing the raw data [20].

3.3.2 Acquisition of expert knowledge

To ensure a thorough and precise analysis of collected data, thereby avoiding arbitrary or inaccurate methods [21], it is essential to seek input from plant experts. Their expertise is particularly valuable in the conditioning process, where we utilized human elicitation techniques with production floor specialists. Their input was crucial for identifying relevant descriptors and effectively analyzing the collected data.

The following points provide more detail on the information gathered:

• Identification of Key Variables: Of the 15 variables identified in each packaging process, two are related to process progression: the operating range and the conditioning robot’s capacity. The remaining 11 variables result from the configuration of the process and include elements like the speed, angular speed, and linear speed of the small pulley, as well as the winding angle. The last two variables – the coefficient of friction and the tension of the strand – depend on the condition of the power transmitter and could act as indicators of its health. These are integrated into the corresponding descriptor.

• Exploring Potential Interactions: Experts highlighted the importance of examining potential interactions between variables. This exploration forms the fourth step in our proposed data analysis methodology.

• Determining Factors Affecting Asset Deterioration: Plant experts identified key factors indicative of the power transmitter’s deterioration. These are closely linked to configuration variables such as the capacity of the packaging robot and the speed of the small pulley, and are considered crucial for estimating the degradation state of the asset.

3.3.3 Descriptive analysis

In the second phase, our objective is to scrutinize the conduct of variables throughout the packaging procedures of the final products and to pinpoint characteristic descriptors suitable for their delineation. We scrutinized various variables linked to the robot’s transmitter, as highlighted in Table 1.

In a typical lifespan, a newly installed robot’s transmitter exhibits functionality for approximately 40 product-conditioning tasks. Upon meticulous examination of Figure 5, the initiation of power transmitter degradation became evident around process number 35, primarily attributable to substantial variations in input variables. This inference underscores the notable impact of the quantity of packaging processes on the health condition of the power transmitter.

Figure 5.

Graphical representation of critical parameters [20].

3.3.4 Bivariate analysis

In the third phase, our aim is to construct models that establish a relationship between the patterns manifested by the parameters, saved in V, and the deterioration status of the robot’s transmitter, with a particular focus on the belt component. The objective is to discern vectors that precisely portray the condition of the asset.

Through this sub-section, we aim to identify appropriate descriptors for our PdM framework; we employed the coefficient of determination, represented as R². This measure assesses the accuracy of forecasts generated by the linear regression model. In order to assess the predictive performance of our models (DBF and NBF), we conducted a linear regression and examined its residuals, obtaining R².

Following established literature, descriptors exhibiting R² values close to 1 were selected, suggesting a strong correlation among variables and an increased probability of providing precise predictions within PdM frameworks.

Table 2 highlights the corresponding descriptors (V1, V2, and V3).

V1	V2	V3
Small pulley speed	Large pulley speed	Robot’s capacity
Small pulley winding angle	Large pulley winding angle	Packaging process
Small pulley angular speed	Large pulley angular speed
Soft strand tension	Stretched strand tension
Friction coefficient

Table 2.

Potential descriptors.

3.4 DBF based on machine learning techniques

The chosen model for PdM is a DBF. Following the processing and transformation of the data into a useful format, ML techniques are applied to construct the predictive model. DBF serves as a robust mechanism to assimilate uncertain and fluctuating information into a production system. In our case, we focus on a packaging robot fitted with sensors that generate imprecise measurements. Our aim is to derive the most precise estimation possible of the current condition of the power transmitter.

The chosen model for this investigation, the DBF, gauges the degradation status of the power transmitter through discrete value ranges. This model incorporates considerations of confidence in the outcomes, enhancing the accuracy of estimation.

In the following, we will design the PdM mathematical model. Through this model, we note:

• N_s: Number of values or states in which transmitter health/degradation is discretized;

• x: Discrete random variable representing the degradation of the power transmitter, x ∈ {1, …, N_s};

• k: Instant of time, so x_k indicates the degradation’s state of the power transmitter at instant k;

• z_k: Sensor measurements available at time k; in fact it is a reference to the vector of descriptors V;

• u_k: Action taken at the moment k during the spherical bushels’ assembly process;

• c_k: Process configuration parameters that influence sensor measurements.

At each instance, denoted as k, upon completion of a packaging process, we can determine the likelihood of the power transmitter experiencing a specific level of degradation after finalizing the packaging process for 100 items. This computation can be carried out using Eq. (1) as follows (where η denotes a normalization constant):

$$\mathit{Bel}\left(x_k\right)=\mathrm{ꞃ}.P\left(z_k\mathrm{|}{x}_k,c_k,z_{k-w:k-1}\right)\sum _{i=1}^{\mathit{Ns}}P\left(x_k\mathrm{|}{x}_{k-1}\right)\mathit{Bel}\left(x_{k-1}\right)$$E1

3.4.1 Degradation’s estimation

As presented below, Algorithm 1 is utilized for predicting the degradation levels of packaging robot at a specific time denoted k. During the prediction phase (lines 2 to 3), the algorithm anticipates the degradation state at time k based on the preceding belief (at time k − 1) and the action model. Following the prediction, the algorithm advances to the update phase (lines 4 to 9), where it adjusts the prior prediction using sensor measurements from the sensor model. This phase calculates the most probable degradation state aligned with the sensor measurements. Ultimately, after normalization, the belief regarding the degradation state Bel(x_k) can be determined (line 11). This algorithm, employing the Discrete Bayesian Filter (DBF) approach, integrates uncertain and variable information and inspections into the production system. By utilizing a discrete spectrum of values to represent the transmitter’s health, the DBF model computes degradation state while considering confidence in the outcomes.

Algorithm 1: Discrete Bayes Filter for the estimation of the degradation’s state

1. $\text{Procedure}\mathbf{DBF}\mathbf{-}\mathbf{\text{ESTIMATION}}\left(\mathrm{Bel}\left(x_{k-1}\right),u_k,z_{k-w:k-1},c_k,N_s\right)$

   $\mathrm{Bel}(x_{k-1}$) % Belief at the moment x_k-1

   $u_k$ % Control action taken

   $z_{k-w:k-1}$ % Last w measurement (sensors)

   $c_k$ % Configuration settings of the assembly process

   $N_s$ % number of possible discretized degradation’s states

2. $\mathbf{\text{For}}\text{each}{x}_k\mathbf{do}$

3. $\mathbf{\text{For}}\mathrm{i}\text{from}1\text{to}{N}_s\mathbf{do}$

4. $\mathrm{Bel}\left(x_k\right)={\sum }_{\mathrm{i}=1}^{\mathrm{Ns}}\mathrm{P}(x_k\mathrm{|}{u}_k,x_{k-1}^i\ast \mathrm{Bel}\left(x_{k-1}^i\right)$ Prediction phase

5. $\mathbf{End}$

6. $\mathrm{ɳ}=0$

7. $\mathrm{Bel}\left(x_k\right)=\mathrm{P}(z_k\mathrm{|}{x}_k,c_k,z_{k-w:k-1}\ast \mathrm{Bel}\left(x_k\right)$ Update phase

8. $\mathrm{ɳ}=\mathrm{ɳ}+\mathrm{Bel}\left(x_k\right)$

9. $\mathrm{Bel}\left(x_k\right)={\mathrm{ɳ}}^{-1}\ast \mathrm{Bel}\left(x_k\right)$

10. $\mathbf{End}$

11. $\mathbf{\text{Return}}\left(\mathrm{Bel}\left(x_k\right)\right)$

We outline several initial conditions necessary for the development of the PdM:

• Bel (x₁ = 1) = 1: the power transmitters are new;

• Bel (x_i = 1) $\ne$ 1 = 0, whatever i $\ne$ 1 we will not be able to have a first-order degradation state.

4. Results and discussion

4.1 Estimation of the degradation state

Figure 6 provides a visual representation of the degradation progression of a power transmitter after 29 packaging processes, equivalent to 29,000 items being processed.

Figure 6.

Degradation states that the estimation of a power transmitter has performed 29 conditioning processes.

Figure 7 showcases the predicted wear probability attributed by the DBF predictive model concerning the process number. Notably, the graph demonstrates a consistent linear rise in wear probability as the number of processes increases.

Figure 7.

Predicted wear probability of the DBF predictive model in function of process’ number [3].

4.2 Future state prediction

To predict the future state, incorporating additional control actions for packaging processes u_k + 1…u_k + 4 (scenario 1) and u_k + 1…u_k + 15 (scenario 2) is crucial for enhancing the model. Subsequently, the model is executed to generate predictions. Figure 8 depicts the model’s outcomes in two distinct scenarios: the first one is performed after 29 packaging processes with an additional 4 packaging tasks, and the second one follows 29 processes accompanied by 15 additional conditioning tasks. In the first case, maintenance personnel can deduce that the robot’s power transmitter maintains a sufficient state to function after completing 33 tasks (29 + 4). Conversely, case two indicates the necessity of a maintenance task to preempt potential failures.

Figure 8.

Predictive outcomes from the model in two distinct scenarios [3].

Comparing the results between the DBF predictive model and the NBF model rooted in machine learning, we have gauged their matching scores, reflecting their respective abilities to predict outcomes based on input data. The DBF model displays a commendable 62% score, while the NBF model trails slightly behind with a score of 55%.

Based on the outcomes, Figures 9 and 10 derived from the Isolation Forest unsupervised learning algorithm, it is noticeable that after every 20 packaging processes, the degradation state of the system exhibits a variation, fluctuating within a range of 2–27%. This variation pattern suggests a diverse level of degradation at different intervals, implying a non-linear or irregular degradation progression over these specific time intervals.

Figure 9.

Anomaly detection derived from the isolation Forest unsupervised ML technique.

Figure 10.

Percentage of anomalies in each packaging process.

The predictions of the model are either 1 or − 1; where −1 indicates outliers or anomalous values, while 1 signifies normal values.

Conclusively, we affirm that the DBF model exhibits superior predictive capabilities compared to the NBF model. Additionally, in comparison with the Isolation Forest, the DBF model highlights enhanced predictive abilities.

5. Conclusions

In summary, adopting a Machine Learning (ML) based predictive maintenance (PdM) model aligns with Industry 4.0’s objectives, marking a significant step toward its standards. This study utilized a Discrete Bayesian Filter (DBF) to evaluate the degradation status of a power transmitter in a packaging robot. When compared to a traditional ML-based model (NBF) and an Isolation Forest Unsupervised ML technique, the DBF model demonstrated enhanced predictive accuracy, achieving a superior matching score. The selection of descriptors was carried out through a rigorous coefficient of determination R² analysis, and predictions were facilitated by a linear regression model. The DBF model excelled in incorporating uncertain and noisy sensor data, accurately estimating the transmitter’s degradation state. This model is set to support maintenance teams in developing effective strategies, reducing downtime, and increasing productivity in the production system. Its effectiveness in streamlining maintenance planning and cost-efficiency aligns with Industry 4.0’s fundamental goals. A key aim of smart factories is to implement PdM using AI technologies to prevent equipment failures, which is precisely what this study’s PdM model achieves by leveraging historical sensor data and ML algorithms.

Our research necessitates a thorough examination of its implications on the overall efficiency of a flexible shop floor. The findings focus on PdM models for a packaging robot’s power transmitter. On the one hand, PdM models introduce a proactive maintenance approach. By predicting issues in the power transmitter beforehand, maintenance can be scheduled strategically, minimizing unplanned downtimes and optimizing resource use, thereby enhancing the shop floor’s workflow.

However, the potential challenges and limitations must be critically assessed. Predictive models rely on accurate data and consistent system dynamics. Real-world complexities like variable operating conditions and unexpected anomalies could impact model accuracy. Thus, while our research marks a valuable advancement toward efficiency, it also underscores the necessity for continual model refinement and adaptability to changing conditions on the shop floor.

To conclude, our research offers promising methods to improve efficiency on flexible shop floors or other manufacturing systems facility layouts through PdM, but it requires careful consideration of both its advantages and limitations. Recognizing these aspects ensures a comprehensive understanding of how our findings can be effectively implemented and adapted for practical use.

Looking forward, our aim is to compare our highly predictive model (DBF) with a preventive maintenance model based on expert knowledge, highlighting the contrasts in predictive power and practical application.

Conflict of interest

The authors declare no conflict of interest.