In the field of aeroengine manufacturing, turbine blades are critical components that operate under extreme conditions of high temperature, pressure, and rotational speeds. The presence of internal defects such as porosity, inclusions, and cracks in these blades, which often arise during metal casting processes, can compromise engine performance and safety. Traditional methods for detecting these metal casting defects, such as radiographic testing, rely heavily on manual inspection, leading to inefficiencies, high labor costs, and potential human errors. To address these challenges, we explore an unsupervised learning approach for automating defect detection in turbine blade X-ray images. This method leverages generative adversarial networks (GANs) to learn from defect-free samples and identify anomalies without requiring labeled defect data, thereby reducing dependency on extensive annotations and improving detection efficiency.
The core of our approach involves a deep convolutional generative adversarial network (DCGAN) architecture, which includes a generator network, a discriminator network, and an additional encoder network. The generator reconstructs input images, while the discriminator distinguishes between real and reconstructed images. The encoder aids in capturing latent features to enhance reconstruction accuracy. We design multiple loss functions, including reconstruction loss, discrimination loss, encoding loss, and intermediate encoding loss, combined into a weighted objective function for model optimization. By training exclusively on defect-free X-ray images, the model learns to reconstruct normal patterns effectively; when presented with defective images containing metal casting defects, it produces significant reconstruction errors, enabling defect identification. We investigate key parameters such as input image size, encoding length, and loss types to optimize performance.
Our experiments utilize X-ray images of turbine blades collected from industrial radiography films. Each film contains multiple blade images, which are digitized and preprocessed to extract individual blade regions. We apply Otsu’s method for binary thresholding and contour extraction to isolate blade images, followed by uniform cropping into smaller patches for model input. Data augmentation techniques, including random rotations, flips, and brightness adjustments, are employed to enhance model robustness. The training dataset comprises defect-free images, while the test set includes both defective and non-defective samples to evaluate detection accuracy. We use metrics such as area under the curve (AUC), false positive rate, and false negative rate to assess model performance, with a focus on achieving zero missed detections for critical metal casting defects.
The network architecture consists of three main components: a generator with convolutional encoder and decoder layers, an auxiliary encoder for feature extraction, and a discriminator for classification. The generator maps input images to latent vectors and reconstructs them, while the auxiliary encoder ensures feature consistency. The discriminator uses convolutional layers to differentiate between real and generated images. The loss functions are defined as follows: reconstruction loss measures the difference between input and reconstructed images, encoding loss compares latent features, intermediate encoding loss captures mid-level features, and adversarial loss optimizes the discriminator. The total loss is a weighted sum: $$L = w_{\text{con}} L_{\text{con}} + w_{\text{enc}} L_{\text{enc}} + w_{\text{m-enc}} L_{\text{m-enc}} + w_{\text{adv}} L_{\text{adv}}$$ where $w_{\text{con}} = 50$, $w_{\text{enc}} = 1$, $w_{\text{m-enc}} = 1$, and $w_{\text{adv}} = 1$ are weighting coefficients. For reconstruction loss, we compare L1 and L2 norms: $$L_{1,\text{con}}(x, \hat{x}) = \sum_{i=0}^{\hat{m}} \left| x^{(i)} – \hat{x}^{(i)} \right|$$ and $$L_{2,\text{con}}(x, \hat{x}) = \sum_{i=0}^{\hat{m}} \left( x^{(i)} – \hat{x}^{(i)} \right)^2$$ where $\hat{m}$ is the batch size. Similarly, encoding loss is defined as $$L_{1,\text{enc}}(Z, \hat{Z}) = \sum_{i=0}^{\hat{m}} \left| Z^{(i)} – \hat{Z}^{(i)} \right|$$ or $$L_{2,\text{enc}}(Z, \hat{Z}) = \sum_{i=0}^{\hat{m}} \left( Z^{(i)} – \hat{Z}^{(i)} \right)^2$$ for latent features $Z$ and $\hat{Z}$.
Training is performed using the Adam optimizer with parameters $\beta_1 = 0.9$ and $\beta_2 = 0.999$, and an initial learning rate of 0.0002. Models are trained for up to 100 epochs with early stopping based on validation loss. We evaluate different input image sizes (64×64, 96×96, 128×128, and 160×160 pixels) and encoding lengths (400, 600, 800, and 1000) to determine optimal configurations. The batch sizes vary with input size to manage computational resources. The primary goal is to minimize the false negative rate while maintaining a reasonable false positive rate, ensuring that metal casting defects are reliably detected in production environments.
The results demonstrate that the model with an input size of 128×128 pixels and encoding length of 600 achieves the highest AUC of 0.911 when using L2 reconstruction loss. This configuration effectively balances feature capture and computational efficiency, as smaller input sizes lead to poorer detection due to limited defect information, while larger sizes increase complexity without significant gains. The table below summarizes the performance of various models under zero missed detection conditions, highlighting the trade-off between false positive rates and model parameters. For instance, the model with 96×96 input size and encoding length 600 achieves a false positive rate of 62.1% at 100% detection rate, which, although high, can still improve manual inspection efficiency by 1.6 times when used as an辅助 tool.
| Model | Input Size (pixels) | Encoding Length | Reconstruction Loss | AUC | False Positive Rate at Zero Missed Detection (%) |
|---|---|---|---|---|---|
| A | 64×64 | 600 | L2 | 0.768 | 69.6 |
| B | 96×96 | 600 | L2 | 0.852 | 62.1 |
| C | 128×128 | 600 | L2 | 0.911 | 68.5 |
| D | 160×160 | 600 | L2 | 0.880 | 72.5 |
| E | 128×128 | 400 | L2 | 0.879 | 64.2 |
| F | 128×128 | 800 | L2 | 0.850 | 81.1 |
| G | 128×128 | 1000 | L2 | 0.865 | 70.3 |
| H | 128×128 | 600 | L1 | 0.879 | 75.9 |
The choice of reconstruction loss significantly impacts model performance. Models trained with L2 loss consistently outperform those with L1 loss, as L2’s squared error penalizes larger deviations more heavily, which is beneficial for detecting subtle metal casting defects. The training loss curves show that L2 loss leads to smoother convergence and lower final loss values, reducing instability during optimization. For example, the total loss for the 128×128 model with L2 loss decreases steadily to around 0.5 after 100 epochs, whereas the L1 loss model exhibits higher volatility and a final loss above 0.7. This underscores the importance of loss selection in unsupervised learning for defect detection.
In terms of computational efficiency, smaller input sizes require less training time per epoch but yield inferior results due to reduced feature resolution. The table below compares training times and AUC values across different input sizes, illustrating that the 128×128 model offers the best trade-off, with an AUC of 0.911 and manageable training duration. This is critical for practical applications where both accuracy and speed are essential for identifying metal casting defects in high-volume production.
| Input Size (pixels) | Training Time per Epoch (seconds) | AUC |
|---|---|---|
| 64×64 | 29.314 | 0.768 |
| 96×96 | 33.362 | 0.852 |
| 128×128 | 79.016 | 0.911 |
| 160×160 | 92.671 | 0.880 |
The discriminator network plays a crucial role in guiding the generator towards realistic reconstructions. It consists of seven convolutional layers with Leaky ReLU activation and batch normalization, culminating in a Softmax output for binary classification. The adversarial loss, defined as $$L_{\text{adv}} = \ln[D(x)] + \ln[1 – D(\hat{x})]$$ where $D(x)$ is the discriminator’s output for real images, encourages the generator to produce images that are indistinguishable from real defect-free samples. This adversarial training process enhances the model’s ability to generalize to unseen defects, as the generator learns to minimize reconstruction errors only for normal patterns, leading to high errors for metal casting defects.
Data augmentation is vital for improving model robustness. By applying random transformations such as rotation (90°, 180°, 270°), flipping, and adjustments in brightness, contrast, and saturation, we increase the diversity of the training dataset. This helps the model learn invariant features and reduces overfitting, which is particularly important for detecting rare metal casting defects. The augmented images are generated on-the-fly during training, ensuring that each epoch presents varied samples to the network.
Evaluation on the test set reveals that the proposed method achieves zero missed detections for all models, meeting the stringent requirement of identifying every defective blade. However, the false positive rates remain high, exceeding 62.1% in the best case. This implies that while the method is highly sensitive to metal casting defects, it may flag many non-defective samples as defective, necessitating secondary manual checks. Despite this, the overall detection efficiency improves, as the model pre-screens images and reduces the workload for human inspectors. In production settings, this can lead to a 1.6-fold increase in inspection throughput, making it a valuable tool for quality control.
The impact of encoding length on model performance is also investigated. Longer encoding vectors capture more detailed features but may lead to overfitting, while shorter vectors might miss critical defect information. The optimal encoding length of 600 provides a balance, as evidenced by the highest AUC in the 128×128 model. The mathematical formulation of the encoding process involves mapping input images to latent vectors $Z$ through the encoder network: $$Z = G_e(x)$$ where $G_e$ is the convolutional encoder. The decoder then reconstructs the image as $$\hat{x} = G_d(Z)$$ where $G_d$ is the deconvolutional decoder. The auxiliary encoder computes $\hat{Z} = \hat{G}_e(\hat{x})$, and the encoding loss ensures consistency between $Z$ and $\hat{Z}$.
In conclusion, our unsupervised learning approach based on DCGAN offers a promising solution for automating defect detection in aeroengine turbine blades. By leveraging defect-free X-ray images, the model learns to identify metal casting defects through reconstruction errors, achieving zero missed detections in testing. The optimal configuration with 128×128 input size, encoding length 600, and L2 reconstruction loss yields an AUC of 0.911, though false positive rates are substantial. Future work will focus on refining the network architecture and incorporating semi-supervised techniques to reduce false positives while maintaining high sensitivity. This method has the potential to significantly enhance inspection processes in aerospace manufacturing, ensuring the reliability and safety of critical components.
The generalizability of this approach extends to other applications involving metal casting defects, such as in automotive or industrial components, where unsupervised learning can reduce annotation costs. The mathematical framework and loss functions developed here provide a foundation for further research in anomaly detection. As manufacturing industries increasingly adopt digital transformation, such AI-driven methods will play a pivotal role in improving quality assurance and operational efficiency.