1 of 1

RI2AP: Robust and Interpretable 2D Anomaly Prediction in Assembly Pipelines

Chathurangi Shyalika¹, Kaushik Roy¹, Renjith Prasad¹, Yuxin Zi¹, Priya Mittal¹, Fadi El Kalach²,

Vignesh Narayanan¹, Ramy Harik², Amit Sheth¹

Artificial Intelligence Institute, University of South Carolina
McNair Center, Department of Mechanical Engineering, University of South Carolina

1. MOTIVATION

2. PROBLEM STATEMENT

3.FUTURE FACTORIES (FF) DATASET [1,2]

6. EXPERIMENTS & RESULTS

Manufacturing sensor data holds the potential to offer valuable insights for optimizing assembly lines, such as improving quality maintenance and enhancing overall pipeline efficiency.
Our primary focus is on one significant aspect, Anomaly Prediction, with a specific application in rocket assembly pipelines.
The repercussions of failing to manage anomalous event effectively carry substantial costs, as demonstrated by prior research indicating financial losses of approximately 20,000

USD per minute.

Additionally, labor costs contribute to the burden, with nearly 44% of manufacturers experiencing unplanned downtime on a bimonthly basis.
The impact is particularly severe when addressing errors in advanced stages of production.

7. CONCLUSION & FURTHER WORK

Acknowledgement

#2119654: RII Track 2 FEC:

Enabling Factory to Factory Networking

for Future Manufacturing

A novel methodology for anomaly prediction in manufacturing assembly pipelines, with demonstrated effectiveness in the context of rocket assembly.
Combining rules can serve improved interpretability.
Domain-expert-friendly anomaly predictions applicable beyond rocket assembly to similar manufacturing pipelines.

Further work:

Implementing a multisensor FA that considers important measurements simultaneously.

Investigating the impact of alternative combining rules, such as tree-structured conditional probability effects, and do-calculus.

Developing higher-level representations of causal phenomena related to anomalies involving exploring connections between sensor measurements and high-level constructs.

5. THE RI2AP METHOD

7. REFERENCES

[1] Harik, R., Kalach, F.E., Samaha, J., Clark, D., Sander, D., Samaha, P., Burns, L., Yousif, I., Gadow, V., Tarekegne, T. and Saha, N., 2024. Analog and Multi-modal Manufacturing Datasets Acquired on the Future Factories Platform. arXiv preprint arXiv:2401.15544.

[2] Harik, R. (2024, January 18). FF: 2023 12 12: Analog Dataset. Kaggle. https://www.kaggle.com/datasets/ramyharik/ff-2023-12-12-analog-dataset

[3] Koller, D. and Friedman, N., 2009. Probabilistic graphical models: principles and techniques. MIT press.

Function Approximation (FA) Methods:

General form for the FA used:

Input: Output:

Model Architectures:
Long Short-term Memory Networks (LSTMs)
Transformer

Decoder-only transformer (Our own)
TimeGPT

Method of Moments: Leverages the idea of taking n^th moment for imbalanced data

Challenges

(i) Anomalous events are extremely rare,

(ii) High fidelity simulation data is scarce (real data is costly

to acquire), and

(iii) The dependence between anomalous events do not readily

lend themselves to traditional ML formulations (different anomaly

labels are usually treated independently and “one-hot” encoded).

Beyond prediction-related challenges, there are also hurdles related to interpreting the result in a domain expert-friendly manner for informing insights into improving pipelines.

SOTA Machine Learning approaches for anomaly prediction and detection have yielded only ∼50% and ∼66% F1-measure scores.

For each series of measurements up to time step t − 1, denoted by the data list,

We first construct a set of 20 different function approximations:

Then, we combine the set of all the 20 outputs from each of the , using a

combining rule denoted as aggr to yield a final value .

Consider an assembly cycle that assembles a rocket from the set of parts.
Each cycle takes place over a sequence of t = 1, 2, ..., T discrete time steps.
At each time step t, a group of 20 sensor measurements are collected.

Anomalies (A) during a cycle can be categorized as None or Types 1-4.

None:

Type 1: Type 2:

Type 3:

Type 4:

Anomaly Encodings:

Task Description: Predict measurements and the token assignment of the anomaly type at time step t (two dimensional prediction) .

4. PROBLEM FORMULATION

Combining rules used: Noisy-OR and Noisy-MAX, introduced in the independence of causal influence framework [3].

6. EXPERIMENTS & RESULTS

Figure 1: Some images from FF Cell: *R01-Robot 1, R02-Robot 2, R03-Robot 3, R04-Robot 4

Figure 2: Abstract illustration of the proposed method

Figure 3: Illustration of RI2AP method. Figure (a) and (b) corresponds to Equations (1) and (2), respectively.

Figure 4: Detailed Illustration of RI2AP

Figure 6: Comparison of F1 Score with the LSTM, Transformers, and the method of moments using Noisy-OR. (*A1:A5, Anomaly types)

Figure 5: Loss/Error comparison of different FA

and combining rule predictions.

Figure 7: Comparison of F1 Score with the LSTM, Transformers,

and the method of moments using Noisy-MAX. (*A1:A5, Anomaly types)

Project Page

Empirical evaluations demonstrate the effectiveness of RI2AP.
Achieves significant performance gain (up to 30 points

on the F1 measure) over SOTA ML baselines.