Abstract:In the realm of manufacturing processes, equipment failures can result in substantial financial losses and pose significant safety hazards. Consequently, prior research has primarily been focused on preemptively detecting anomalies before they manifest. However, within industrial contexts, the precise interpretation of predictive outcomes holds paramount importance. This has spurred the development of research in Explainable Artificial Intelligence (XAI) to elucidate the inner workings of predictive models. Previous studies have endeavored to furnish explanations for anomaly detection within these models. Nonetheless, rectifying these anomalies typically necessitates the expertise of seasoned professionals. Therefore, our study extends beyond the mere identification of anomaly causes; we also ascertain the specific adjustments required to normalize these deviations. In this paper, we present novel research avenues and introduce three methods to tackle this challenge. Each method has exhibited a remarkable success rate in normalizing detected errors, scoring 97.30%, 97.30%, and 100.0%, respectively. This research not only contributes to the field of anomaly detection but also amplifies the practical applicability of these models in industrial environments. It furnishes actionable insights for error correction, thereby enhancing their utility and efficacy in real-world scenarios.
engineering, electrical & electronic,computer science, information systems,physics, applied
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the financial losses and potential safety hazards caused by equipment failures in the manufacturing process. Specifically, the paper focuses on how to detect anomalies in advance during the manufacturing process, further explain the reasons for these anomalies, and provide specific adjustment suggestions to restore normal conditions. The following are the main objectives and solutions of the paper:
1. **Detect Anomalies in Advance**: By using machine - learning models (such as K - Nearest Neighbors Classifier, Decision Trees Classifier, Random Forest Classifier, Extra Trees Classifier, and Gradient Boosting Classifier), the paper aims to predict product defects that may occur during the manufacturing process.
2. **Explain the Reasons for Anomalies**: The paper not only identifies anomalies but also uses Explainable Artificial Intelligence (XAI) techniques (such as SHapley Additive exPlanation, SHAP) to explain which features lead to the anomalies. This makes the prediction results of the model more transparent and easier for industrial users to understand.
3. **Propose Adjustment Suggestions**: The paper proposes three methods to solve the detected anomalies:
- **Method #1 (SHAP + Feature Mean)**: By calculating SHAP values, determine the features that have the greatest impact on the model prediction, and adjust these features to the average value of normal data.
- **Method #2 (Most Frequent Feature + Feature Mean)**: By analyzing the features with the highest frequency in the node conditions of the tree model, adjust these features to the median value of normal data.
- **Method #3 (Final Node Feature + Final Node Condition)**: Utilize the final node conditions of the tree model to gradually adjust the feature values until the model prediction results return to normal.
4. **Verify the Effectiveness of the Methods**: The paper verifies the effectiveness of the three methods through experiments, achieving normalized success rates of 97.30%, 97.30%, and 100.00% respectively.
In summary, the problem that this paper attempts to solve is how to use advanced anomaly - detection techniques, combined with Explainable Artificial Intelligence methods, not only to detect anomalies in the manufacturing process but also to provide specific adjustment suggestions to restore normal production. This method can not only reduce economic losses but also improve production safety and efficiency.
In terms of formulas, although this paper does not directly involve complex mathematical formulas, the calculation of SHAP values can be expressed in the following form:
\[ \phi_i=\frac{1}{M!} \sum_{\text{coalitions including } i} \left( v(S \cup \{i\}) - v(S) \right) \]
where $\phi_i$ represents the SHAP value of feature $i$, $v(S)$ is the contribution value of feature set $S$, and $M$ is the total number of features.