Algorithms for dynamically adjusting thresholds can be categorized into three types based on implementation complexity and applicable scenarios, covering everything from lightweight engineering implementations to intelligent prediction, adapting to different development capabilities:
1. 3σ Statistical Dynamic Update Method (Most commonly used, lightweight and easy to implement)
This is the most mainstream basic algorithm in industrial scenarios. It has simple logic, low computational load, and is suitable for most factories:
Core logic: Weekly statistics of the historical deviations of two sensors are compiled, the standard deviation σ of the deviation is calculated, and a new threshold is set to 3σ (99.7% of normal fluctuations under a normal distribution will fall within this range; exceeding this range is considered abnormal);
Advantages: Automatic adjustment based entirely on actual operating data, no preset parameters required, adaptable to aging and changes in operating conditions;
Applicable scenarios: Most conventional injection molding production, the first choice for self-developed and modified small and medium-sized factories.
2. Fuzzy Rule Algorithm for Process Matching (Adapts to Multiple Process Switching)
For scenarios with frequent production changes, this algorithm dynamically switches processes based on preset rules using fuzzy logic matching:
Core Logic: Using product precision, raw material corrosivity, and pressure rating as input variables, fuzzy inference automatically calculates corresponding thresholds. For example, for "high glass fiber + high pressure + precision," it infers a result with a 15% relaxation of the threshold.
Advantages: Automatically adapts to different process combinations without requiring separate templates for each product.
Applicable Scenarios: Injection molding plants with multiple product varieties, small batches, and frequent production changes.
3. Machine Learning Predictive Adjustment Algorithm (High-End Intelligent Scenarios)
Based on training models using historical anomaly data, this algorithm achieves predictive dynamic adjustment with the highest accuracy.
Common Implementations: Use classification models (such as logistic regression or lightweight CNN) to distinguish between "true aging deviation" and "operating condition interference deviation," adjusting the threshold only for true deviation. Alternatively, use LSTM to predict the deviation growth trend and pre-amplify the threshold to reserve aging space.
Advantages: Can distinguish complex interference, reducing false alarm rate by more than 30% compared to traditional algorithms.
Disadvantages: Requires accumulating a certain amount of historical anomaly data to train the model, resulting in greater development effort.
Applicable Scenarios: Intelligent high-end injection molding production lines and large-scale batch production scenarios.
Among these, the combination of 3σ statistical dynamic update method + process stage binding is currently the most cost-effective solution, meeting the needs of over 90% of scenarios and requiring no complex development for implementation.
