How to optimize the dual-sensor redundancy comparison algorithm?

Shifting from "static rigidity" to "dynamic intelligence" is the fundamental principle behind optimizing the dual-sensor redundancy comparison method. This approach focuses on resolving false alarm issues brought on by variations in operating conditions, signal noise, and sensor aging. Specific optimization solutions in four dimensions are as follows:

1. Implement a "process status awareness"-based dynamic threshold method. False alarms are primarily caused by fixed thresholds, which need to be automatically modified based on the stage of the process:

State binding: The operational cycle is separated into several stages by reading process signals (such as injection start, holding pressure end, and cooling) from the injection molding machine or equipment via PLC;

Setting coefficients that are segmented:

High dynamic stage (e.g., injection/start-up): To prevent false alarms caused by impact interference, the threshold is reduced to 2.0–3.0 times the base value when pressure/temperature changes significantly;

Steady-state stage (such as maintaining temperature or pressure): To guaranty detecting sensitivity, the threshold is tightened to 0.8–1.0 times while the operating parameters are stable;

Transitional stage (e.g., mold opening/closing/shutdown): The threshold is set to 1.5 times or the comparison logic is momentarily disabled in response to mechanical vibration or abrupt temperature fluctuations.

Implementation technique: To change the current effective threshold in real time, create a State_Threshold_Map mapping table in the algorithm.

2. Improve the Filtering Algorithm for Signal Preprocessing

Instantaneous deviations can immediately exceed restrictions due to noise in the original signal, necessitating a combination of filters:

With a window size of three to five points, the median filter successfully eliminates transient outliers such electrical spikes and injection molding shocks while maintaining the actual trend;

Moving Average Filter: This filter smoothes high-frequency random noise and stabilizes the comparison curve by setting the window size to 10–20 points;

First-Order Hysteresis Filter: To further reduce minute jitter, an inertial element is included for slow variables like temperature.

The recommended sequence is to remove the median peak first, then smooth the average, and lastly do comparison computations.

3. Include "Trend Judgment Logic" and "Time Dimension"

Time tolerance and trend analysis must be included because a single sampling variation does not always point to a problem:

Time Delay Confirmation Mechanism: A timer is initiated when the deviation surpasses the threshold. The deviation is only deemed a fault alarm if it surpasses the threshold for an extended duration > T (e.g., 5–10 seconds). The timer is reset as interference if the deviation returns to normal during this time.

Rate of Change Limit: Determine the deviation's rate of change (d(Diff)/dt). If d(Diff)/dt > Limit and the deviation briefly jumps but rapidly recovers, it is considered electrical interference rather than sensor drift.

Trend Warning: If the deviation exhibits a monotonically increasing trend and persists for several cycles, even if the threshold is not surpassed, a "maintenance reminder" is initiated ahead of time, encouraging preventative sensor replacement.

4. Create a Mechanism for Adaptive Baseline Updates

Sensors gradually deteriorate over time, and a fixed baseline eventually results in more false alarms.

Rolling Baseline Correction: During each manufacturing cycle's steady-state phase (such as the conclusion of cooling), note the average deviation (Bias_Avg) between the two sensors.

Low-Pass Baseline Update: Use the Exponentially Weighted Moving Average (EWMA) to update the baseline deviation.

Base new = α⋅Bias current + (1−α)⋅Base old (α is adjusted to 0.01-0.05 to guarantee the baseline is unaffected by individual anomalies and gradually follows aging).

Periodic Reset: Forcibly reset the baseline to zero following each manual calibration or sensor replacement.

Evaluation of Impacts Prior to and Following Optimization

Item for Optimization

Conventional Algorithm

Enhanced Algorithm

The Threshold Approach

A single, fixed threshold

Process-stage-based dynamic thresholds

Ability to Prevent Interference

Weak and susceptible to spikes

Robust, median + average dual filtering

Alarm Reasoning

Only sound an alarm when the limit is immediately exceeded.

Alarm only when the limit is consistently exceeded plus trend analysis

Long-Term Consistency

The number of false alarms rises with age.

Aging and long-term stability are automatically followed by baseline