Anomaly Detection For Business Process Using Integrated Digital Twin-Intelligent Body And Dynamic Declarative Constraints

Putri, Maheswari Parisya (2026) Anomaly Detection For Business Process Using Integrated Digital Twin-Intelligent Body And Dynamic Declarative Constraints. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221096-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Process anomaly detection in business processes is critical for maintaining operational compliance, yet most existing methods rely on static rules or fixed threshold strategies that do not adapt to evolving process behavior. While Li et al. (2025) propose a Digital Twin-Intelligent Body (DT-IB) architecture using EWMA-based dynamic thresholding, their approach depends on a single adaptive mechanism with a fixed smoothing parameter and is validated only on a synthetic platform, limiting generalizability to real-world benchmarks. This research proposes an enhanced DT-IB-DDC framework integrating eight specialized DDC types with iterative two-round baseline calibration. Within this framework, the Digital Twin's role is to maintain and detail the case-level statistical state, that is, the baseline profile of normal execution used as diagnostic context, rather than to perform the anomaly detection decision itself; detection is carried out by the Dynamic Declarative Constraints and Intelligent Body components that consume this state. The Digital Twin applies Otsu-like stable cohort selection (40-95% retention) to derive adaptive thresholds per DDC type, producing threshold T = 0.3343 (k = 1.800). The Intelligent Body fuses five complementary components, namely DDC violation scoring, Z-score normalization, Multi-View Association Rule Mining (MV-ARM), business rule filtering, and Isolation Forest, through hierarchical gating that dampens scores for structurally conforming cases. Experiments on the Traffic Fines event log (129,615 cases) show DT-IB achieves F1 = 0.9182, outperforming Li et al.'s (2025) EWMA-based thresholding approach when adopted on the same dataset (F1 = 0.7978) and all baselines including Static DC (F1 = 0.0460), Single-View ARM (F1 = 0.0076), LSTM Autoencoder (F1 = 0.0070), Transformer Autoencoder (F1 = 0.0000), and Naive Rule-Based (F1 = 0.8836). Stratified 5-fold cross-validation yields stable F1 = 0.9185 ± 0.0024, demonstrating robustness without requiring labeled training data or a predefined reference process model.
=====================================================================================================================================
Deteksi anomali proses dalam proses bisnis sangat penting untuk kepatuhan operasional, namun sebagian besar metode yang ada bergantung pada strategi aturan statis atau threshold tetap yang gagal beradaptasi terhadap perubahan perilaku proses. Li et al. (2025) mengusulkan arsitektur Digital Twin-Intelligent Body (DT-IB) berbasis dynamic thresholding EWMA, namun dengan mekanisme smoothing tunggal yang tetap dan hanya divalidasi pada platform sintetis. Penelitian ini mengusulkan kerangka DT-IB-DDC yang ditingkatkan dengan mengintegrasikan delapan tipe Dynamic Declarative Constraint (DDC) terspesialisasi dan kalibrasi baseline dua putaran secara iteratif, menurunkan threshold adaptif (T = 0.3343, k = 1.800) melalui seleksi cohort stabil berbasis Otsu. Dalam kerangka ini, peran Digital Twin adalah memelihara dan mendetailkan status statistik level-kasus, yaitu profil baseline eksekusi normal yang digunakan sebagai konteks diagnostik, bukan melakukan keputusan deteksi anomali itu sendiri; deteksi dilakukan oleh komponen Dynamic Declarative Constraints dan Intelligent Body yang menggunakan status tersebut. Intelligent Body memfusikan lima komponen komplementer, yaitu penilaian pelanggaran DDC, normalisasi Z-score, Multi-View Association Rule Mining, penyaringan aturan bisnis, dan Isolation Forest, melalui hierarchical gating. Pada event log Traffic Fines (129.615 kasus), DT-IB-DDC mencapai F1 = 0.9182, melampaui pendekatan EWMA Li et al. (2025) saat diadopsi pada dataset yang sama (F1 = 0.7978) dan seluruh lima baseline (F1 = 0.0000–0.8836). Validasi silang stratifikasi 5-fold menghasilkan F1 stabil sebesar 0.9185 ± 0.0024, menunjukkan ketangguhan model tanpa memerlukan data pelatihan berlabel.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Anomaly Detection, Business Process, Digital Twin, Dynamic Declarative Constraints, Multi-View Association Rule Mining.
Subjects:	T Technology > T Technology (General) > T58.5 Information technology. IT--Auditing
Divisions:	Faculty of Information and Communication Technology > Informatics
Depositing User:	Maheswari Parisya Putri
Date Deposited:	27 Jul 2026 04:00
Last Modified:	27 Jul 2026 04:00
URI:	http://repository.its.ac.id/id/eprint/137727

Actions (login required)

View Item