The core of configuring a dynamic threshold adjustment mechanism for a hot runner predictive maintenance system lies in shifting from "static fixed values" to "dynamic boundaries that adapt based on statistical patterns and operating conditions." This approach is designed to address data drift caused by fluctuations in the injection molding process, material changes, and equipment aging.
1. Core Principle: Statistical Dynamic Thresholds Based on Sliding Windows
Dynamic thresholds are not set arbitrarily; rather, they are calculated in real-time based on the statistical distribution of historical data collected during normal operation.
Basic Algorithms: Employ the 3σ Principle (Three Standard Deviations) or the Percentile Method
Upper Threshold = Rolling Mean + K × Rolling Standard Deviation
Lower Threshold= Rolling Mean - K × Standard Deviation
The value of K is typically set between 2.5 and 3.0, and can be adjusted based on the tolerance level for false alarms.
Sliding Window Mechanism:
Instead of utilizing the entire history of data, a specific subset-comprising data from the most recent N production cycles (e.g., the last 50 shots)-is selected to serve as the sample pool.
Upon the completion of each new production cycle, the oldest data point is discarded, the newest data point is added, and the mean and standard deviation are recalculated.
Advantages: This mechanism automatically adapts to baseline drift caused by changes in ambient temperature or the minor aging of heating elements, thereby preventing "false positive" alarms.
Example: If a specific nozzle temperature has recently stabilized within the range of 220 ± 2°C, the dynamic threshold would be set to [214, 226]. If, due to seasonal changes, the overall temperature baseline shifts upward to 222 ± 2°C, the threshold will automatically adjust to [216, 228]-rather than triggering a low-temperature or high-temperature alarm.
2. Layered Modeling for Multiple Operating Conditions: Resolving the "One-Size-Fits-All" Failure Problem
During the injection molding process, the parameter characteristics associated with different stages (injection, holding pressure, cooling) and different materials vary significantly; therefore, it is necessary to establish a system of operating condition tags.
Step 1: Operating Condition Identification and Tagging
Utilize PLC signals (such as screw position or peak injection pressure) to automatically identify whether the system is currently in the "Injection Phase," the "Holding Pressure Phase," or the "Cooling Phase."
Record the specific product model and material batch currently being produced (e.g., ABS-001, PC-002). Step 2: Clustered Modeling
Establish a distinct dynamic threshold model for each unique "Product + Material + Process Stage" combination.
For example: When producing thin-walled parts, injection pressure is high and the threshold range is broad; conversely, when producing thick-walled parts, holding pressure duration is longer, temperature stability requirements are stricter, and the threshold range is narrower.
Step 3: Context Switching
The system monitors operational condition tags in real-time and automatically switches to the corresponding model's specific threshold range.
Note: During the initial stages of mold or material changes-when data is insufficient-the historical thresholds from similar processes may be temporarily adopted. Alternatively, a "Learning Mode" can be enabled to temporarily suppress alarms until a sufficient sample size (e.g., 30 cycles) has been accumulated, at which point the dynamic thresholds are activated.
3. Introduction of Decay Factors and Confidence Management
To prevent sudden noise spikes from causing drastic fluctuations in thresholds, it is necessary to introduce smoothing techniques and confidence assessment.
Exponentially Weighted Moving Average (EWMA):
Assigns higher weights to newly acquired data points and progressively lower weights to older data, thereby smoothing threshold variations and preventing erratic, "sawtooth-like" fluctuations.
Confidence Interval Contraction:
When the sample size is small, the threshold range is expanded (reducing sensitivity); as the sample size increases, the threshold range is gradually narrowed (increasing sensitivity).
Set a minimum sample threshold (e.g., N ≥ 20); if the sample count falls below this threshold, dynamic adjustment is not initiated, and only the global default thresholds are utilized.
4. Anomaly Data Cleaning and Feedback Loop
The premise of dynamic thresholds is that the "training data consists primarily of normal operational data." If fault-related data becomes intermingled, it will lead to threshold distortion (referred to as "contamination").
Offline Cleaning Mechanism:
Prior to calculating new thresholds, obvious outliers (e.g., data points exceeding global safety limits) are first removed.
By cross-referencing maintenance logs, if a specific time period is tagged as having experienced a "fault," the data from that period is excluded from the pool of normal operational samples.
Manual Feedback and Correction:
Operations and maintenance personnel utilize the system to classify triggered alarms, marking them as either a "Confirmed Fault" or a "False Alarm (Ignored)." If marked as a "False Alarm," the system automatically incorporates that data point into the normal sample set, thereby widening the threshold; if marked as a "True Fault," threshold updates are frozen to prevent fault data from contaminating the model.
5. Implementation Roadmap and Parameter Recommendations
Parameter Recommended Initial Value Adjustment Suggestions
Sliding Window Size (N) 50–100 Cycles Use a larger value for high-frequency production; use a smaller value for low-frequency production.
Smoothing Coefficient (α) 0.2 Decrease if noise levels are high; increase if high responsiveness is required.
Multiplier Coefficient (K) 3.0 Increase to 3.5 if false alarms are frequent; decrease to 2.5 if missed detections are frequent.
Minimum Sample Size 20 Cycles If below this value, use a static default threshold.
Update Frequency After each cycle completes Avoid frequent recalculations during the production process.
