Neural Networks Detect Helicopter Rotor Damage

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 30 2024

In a paper published in the journal Sensors, researchers evaluated the performance of a strain-based structural health monitoring system (SHMS) on a composite helicopter rotor blade root using a finite element (FE) model. They trained a damage detection algorithm with FE analyses and incorporated a load recognition step using an artificial neural network (ANN) or a calibration matrix.

The results showed high accuracy in load identification and anomaly detection, even in damaged blades, offering a promising approach for assessing SHMS performance on complex subcomponents under realistic conditions.

Background

Past work focused on monitoring a critical component of a composite helicopter rotor blade root, where complex and costly inspection techniques were required. Researchers investigated an optical fiber-based SHMS to enhance damage detection in this subcomponent.

The challenges included addressing the complexity of adhesive bonding defects and the variability in load conditions, which could significantly impact the effectiveness of damage detection. Additionally, the geometrical complexity of the blade root further complicated the design and implementation of an effective SHMS.

Blade Root Modeling

The study modeled and analyzed a composite helicopter rotor blade root, focusing on its complex structure and potential defects. Using Simulia Abaqus, the component was detailed with cohesive elements to simulate adhesive behavior and contact interactions.

Initially, the team validated a nonlinear model against experimental data, but a linear model was used for further analysis due to computational constraints. Although the linear model had limitations in capturing contact nonlinearity, it was considered a reasonable approximation for assessing SHMS performance.

The study emphasized the importance of a high-fidelity model, validated with fiber Bragg grating (FBG) sensors, to improve SHMS accuracy in detecting load variations and damages in practical scenarios.

Damage Detection Algorithm

The damage detection algorithm for the composite helicopter rotor blade root involved three primary steps: anomaly detection, damage assessment, and damage localization. These tasks required decoupling the strain field's response to both the applied load and the presence of damage.

The study used three algorithm versions to evaluate noise sensitivity and detection performance. Virtual sensors, represented by FE model elements, simulated the strain acquisition process. The strain was calculated by sampling values along the optical fiber path within the blade root's anti-torsional layer. The study used ANNs trained on datasets created through FE analyses to achieve accurate load identification and damage detection.

The damage detection algorithm's versions differed mainly in the load identification process. Versions #1 and #2 utilized an ANN trained on pristine and damaged configurations to estimate the applied load set from the strain data. Version #3 employed a pseudo-inverse approach for load identification, which was sensitive to system conditioning.

Analysts used the identified load set to calculate nominal strain values, which were then compared to the actual strain measured by the virtual sensors. A damage index was derived for each sensor, with an ANN for pattern recognition detecting damage based on these indices. If damage was detected, another ANN estimated the damage's size and location.

The researchers generated the study's virtual dataset by introducing different damage scenarios into the FE model, focusing on critical adhesive joints. The damage scenarios were varied in size and location, creating 360 unique cases. Load conditions, representing operational loads, were applied to the model to simulate realistic scenarios. The approach demonstrated the potential for effective SHMS implementation in detecting and assessing damage in complex subcomponents like the helicopter rotor blade root.

Algorithm Performance Evaluation

The results of the algorithm versions were evaluated for load identification, anomaly detection, damage assessment, and localization under various noise levels (1%, 2%, 4%, and 6%). To ensure reliability, the results were averaged across ten algorithm runs. In an example of damage identification, two blade roots—one pristine and one with 30 mm damage—were compared under the same load set. The team noted that the strain field in the damaged blade root displayed noticeable differences, especially in areas where the damage had been introduced.

ANN 1 accurately identified load sets from strain data, ANN 2 detected anomalies with high precision, and ANN 3 estimated damage size and location, though smaller damages were harder to detect. Receiver operating characteristic (ROC) curves showed larger damages were more easily identified, while noise adversely affected the algorithms' performance, with version #3 showing higher localization errors under noisy conditions. Despite these challenges, the study underscored the potential of ANNs for effective SHMS applications in complex structures.

Conclusion

To sum up, the study evaluated an SHMS on a composite rotor blade root subjected to variable loading using FE modeling and FBG sensors. The results indicated that including only pristine conditions in ANN training was effective for load identification, and ANN-based methods outperformed calibration matrix approaches.

Although the algorithms performed similarly in detecting damage sizes over 15 mm, the strain field showed minimal effects from adhesive layer damage. Future work will focus on validating the algorithm with experimental tests, addressing nonlinearity, and assessing cost-effectiveness.