Die damage identification method based on improved adaptive depth embedded clustering
By improving the adaptive deep embedding clustering method, constructing a deep neural network encoder and jointly optimizing the cluster centers, the problem of insufficient feature extraction in mold damage identification is solved, and efficient and accurate mold damage identification and adaptive improvement are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2025-12-30
- Publication Date
- 2026-05-26
AI Technical Summary
Existing mold damage identification methods have insufficient feature extraction capabilities and poor generalization when processing high-dimensional, nonlinear, and dynamically fluctuating monitoring data. They are also sensitive to the initial cluster centers and are difficult to adapt to the complex and ever-changing characteristics of data in industrial scenarios.
An improved adaptive deep embedding clustering method is adopted. By constructing a deep neural network encoder, integrating channel attention mechanism and squeezing and excitation module, encoder parameters and cluster centers are jointly optimized. An adaptive clustering mechanism is adopted, combined with data augmentation and optimizer scheduling strategy, to improve feature separability and robustness.
After mapping high-dimensional data to a low-dimensional embedding space, the accuracy and robustness of mold damage identification are significantly improved, the false alarm rate is reduced, data noise and sensor drift can be effectively dealt with, the number of model parameters is optimized, and the model can adapt to the complex changes in industrial scenarios.
Smart Images
Figure CN122087480A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent detection technology for mold damage in manufacturing, specifically a mold damage identification method based on improved adaptive deep embedding clustering. Background Technology
[0002] As a core process equipment in the manufacturing industry, the damage state of molds directly affects product precision and production efficiency. Existing mold damage identification methods mainly rely on vibration analysis, acoustic emission technology, or heuristic threshold assessment. However, when processing high-dimensional, nonlinear, and dynamically fluctuating monitoring data, they often face problems such as insufficient feature extraction capabilities and poor generalization.
[0003] In recent years, deep embedding clustering (DEC) has been introduced into the field of mold damage detection. However, this technology still has significant limitations. Its feature extraction capability is limited, and it is sensitive to the initial cluster centers, easily getting trapped in local optima. Furthermore, it lacks quantitative evaluation capabilities and is difficult to adapt to the complex and variable nature of data in industrial scenarios. Therefore, to address these shortcomings, there is an urgent need to develop a mold damage identification method based on improved adaptive deep embedding clustering to overcome the deficiencies in current practical applications. Summary of the Invention
[0004] The purpose of this invention is to provide a mold damage identification method based on improved adaptive deep embedding clustering to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] A mold damage identification method based on improved adaptive deep embedding clustering includes the following steps:
[0007] (1) Data acquisition and preprocessing: acquire mold vibration signals or surface defect images, scale and normalize the acquired data, and expand the sample diversity through data augmentation to obtain preprocessed data;
[0008] (2) Constructing a deep neural network encoder: The deep neural network encoder integrates a squeezing and excitation module of channel attention mechanism, which is used to map the preprocessed high-dimensional data to a low-dimensional embedding space;
[0009] (3) Joint optimization training: With the goal of minimizing KL divergence and reconstruction loss, the deep neural network encoder parameters and cluster centers are jointly optimized, and an adaptive clustering mechanism is adopted to improve feature separability;
[0010] (4) Model evaluation and parameter determination: The model performance is verified by multi-dimensional evaluation indicators, the optimal model parameters are determined, and mold damage identification is realized.
[0011] As a further aspect of the present invention: in step (1), the data enhancement includes at least one of random rotation and brightness adjustment.
[0012] As a further aspect of the present invention: the deep neural network encoder is composed of multiple convolutional layers stacked together. After each convolutional operation, a batch normalization layer and an activation function are sequentially connected. It also integrates a squeezing and activation module with a channel attention mechanism. The channel features are squeezed through global average pooling, and then the channel weights are learned and activation weighting is completed through a fully connected layer.
[0013] As a further aspect of the present invention: the activation function is the ReLU activation function, and the deep neural network encoder is equipped with a decoder. The decoder and the encoder are designed symmetrically to reconstruct the input data.
[0014] As a further aspect of the present invention: In step (3), an optimizer is used in conjunction with a learning rate scheduling strategy for training. During the training process, the target distribution is updated periodically. If the loss value does not decrease within a preset number of consecutive rounds and the fluctuation range is less than a preset threshold, or if the loss value shows an upward trend, then training is stopped.
[0015] As a further aspect of the present invention: the optimizer is the AdamW optimizer, and the learning rate scheduling strategy is the cosine annealing scheduling strategy.
[0016] As a further aspect of the present invention: in step (3), by jointly optimizing encoder parameters and cluster centers, the model can cope with interference caused by data noise, impulse noise and sensor drift, thereby reducing the rate of accuracy decline.
[0017] As a further aspect of the present invention: the deep neural network encoder has 2.1M parameters and a parameter accuracy ratio of 42.5% / M.
[0018] As a further aspect of the present invention, it also includes an incremental update step:
[0019] When new mold damage data is obtained or a new failure mode occurs, the new data is preprocessed in step (1) and the model is retrained to adapt to the dynamic changes in mold damage state.
[0020] As a further aspect of the present invention: in step (4), the multi-dimensional evaluation index includes at least two of the following: clustering accuracy, normalized mutual information, adjusted RAND index, defect detection rate, and false alarm rate.
[0021] Compared with the prior art, the beneficial effects of the present invention are:
[0022] A deep neural network encoder is constructed to map high-dimensional images to a low-dimensional embedding space. The joint objective of the encoder and cluster centers is optimized to minimize KL divergence and reconstruction loss. An adaptive clustering mechanism is adopted to improve feature separability. On the MNIST dataset, SADEC achieves a clustering accuracy (ACC) of 89.2%, which is higher than DEC (83%) and GNN (87%). Normalized mutual information (NMI) and adjusted RAND index (ARI) are significantly better than the comparison algorithms.
[0023] Data augmentation processing was performed on industrial defect images, including random rotation (±15°) and brightness adjustment (±20%). A convolutional autoencoder combined with the AdamW optimizer and cosine annealing scheduling was used for training to optimize the target distribution update strategy. On the NEU-DET metal defect dataset, the defect detection rate (DR) reached 94.2%, and the false alarm rate (FAR) was reduced to 5.8%, representing a 67.6% and 49.1% reduction in false alarm rate compared to DEC and GNN, respectively.
[0024] By incorporating a channel attention mechanism (SENet) into the model, adaptive parameter adjustment is used to address data noise and distribution shift, and encoder parameters and cluster centers are jointly optimized. Under the interference of Gaussian noise (σ=0.1), impulse noise (density 15%) and sensor drift (offset 20%), the accuracy decrease of SADEC (4.2%~6.3%) is much lower than that of DEC (15.3%~17.9%) and GNN (9.7%~12.7%).
[0025] The design incorporates a network structure with 2-3 stacked convolutional layers, combined with batch normalization and ReLU activation functions, to reasonably control the number of model parameters. SADEC requires only 2.1M parameters, and its parameter accuracy ratio (ACC / params) reaches 42.5% / M, which is significantly higher than DEC (35.6% / M) and GNN (21.7% / M). Attached Figure Description
[0026] Figure 1 This is a flowchart of a mold damage identification method based on improved adaptive deep embedding clustering in an embodiment of the present invention.
[0027] Figure 2 This is a schematic diagram of the deep neural network encoder structure in an embodiment of the present invention.
[0028] Figure 3 This is a comparison chart of the accuracy of different algorithms on the MNIST dataset in this embodiment of the invention;
[0029] The horizontal axis represents the parameter index (corresponding to different hyperparameter combinations), and the vertical axis represents the validation accuracy (%).
[0030] Figure 4This is a graph showing the Wilcoxon signed-rank test results for different algorithms adjusting the Rand Index (ARI) on the MNIST dataset in this embodiment of the invention. Detailed Implementation
[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0032] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.
[0033] Please see Figures 1-4 This invention provides a mold damage identification method based on improved adaptive deep embedding clustering, using mold surface defect detection as an application scenario. The method is implemented on the NEU-DET metal surface defect dataset. It mainly includes three stages: data input and preprocessing, network training, and result verification. These stages are sequentially connected and work synergistically. Figure 1 As shown, this method also includes incremental data updates to the model and retraining the model under new fault modes to adapt to the dynamic changes in the mold damage state, as detailed below:
[0034] I. Data Input and Preprocessing
[0035] First, the publicly available industrial defect dataset NEU-DET was selected as the training and validation data. This dataset contains six typical mold surface defects, including cracks, pits, inclusions, scratches, oxide scale, and patches. There are 300 samples for each type of defect, totaling 1,800 grayscale images, covering common surface damage types of molds. At the same time, mold vibration signals or MNIST handwritten digit images can be collected as input data, and the preprocessing process is the same as that of the NEU-DET dataset.
[0036] The images were then preprocessed: all images were scaled to 64×64 pixels to ensure consistent image size in the input model; the min-max normalization method was used to normalize pixel values to the [0,1] range to eliminate dimensional differences caused by different pixel levels; to improve the model's adaptability to pose changes and lighting fluctuations in real industrial scenarios, random rotation (±15°) and brightness adjustment (±20%) were used for data augmentation to expand sample diversity. This data augmentation method is also applicable to MNIST handwritten digit images.
[0037] Finally, the preprocessed images are converted into tensor format, which is typically processed by deep learning models, thus completing the data preparation.
[0038] The above preprocessing operations ensured the uniformity of the input data distribution and the richness of the samples, providing a high-quality data foundation for subsequent network training and effectively improving the model's robustness to pose and illumination changes in practical applications.
[0039] II. Online Training
[0040] A convolutional autoencoder (CAE) is used to construct the core architecture of the model. The encoder consists of 2-3 stacked convolutional layers. The first layer has a 3×3 kernel, a stride of 1, and 64 output channels. The second layer has a 3×3 kernel, a stride of 2, and 128 output channels. The third layer (optional) has a 3×3 kernel, a stride of 1, and 256 output channels. After each convolutional layer, a batch normalization layer and a ReLU activation function are sequentially applied. After the second layer, a SENet compression and activation module is added (global average pooling generates channel descriptors, which are then learned and weighted through two fully connected layers (compression ratio 16, first layer ReLU, second layer Sigmoid). Then, a 2×2 max pooling layer (stride 2) is applied to reduce the feature dimension. Finally, a fully connected layer outputs a 10-dimensional low-dimensional embedding vector. The decoder structure is symmetrical to the encoder and is used for input image reconstruction. The specific encoder structure is as follows. Figure 2 As shown.
[0041] When building a deep neural network encoder, the network structure (convolutional layer → pooling layer → fully connected layer) must be defined first, then the hyperparameter combinations (learning rate, batch size, number of cluster centers, etc.) must be determined, and parameter indices corresponding to different hyperparameter configurations must be generated. The AdamW optimizer is used during training, with a learning rate of 3e-4, weight decay of 1e-5, batch size of 128, and a total of 300 training epochs. A cosine annealing scheduling strategy is introduced (Tmax=50 represents the number of iterations in the temperature / learning rate cycle, i.e., a cosine cycle is completed every 50 epochs, and the learning rate smoothly decreases from the initial value to near zero within the cycle, and then rises again in the next cycle, thereby improving the model's convergence stability and ability to escape local optima). At the same time, a unique parameter index is assigned to each hyperparameter combination. The hyperparameter combination includes key parameters such as learning rate, batch size, and number of cluster centers. This parameter index serves as the core experimental variable in the subsequent result verification stage, used to clearly distinguish the experimental data corresponding to different hyperparameter configurations.
[0042] The pre-training phase aims to minimize the reconstruction loss, which is calculated using the following formula:
[0043]
[0044] Where N is the number of samples, x i For the original sample, To reconstruct the sample;
[0045] During the formal training phase, the target distribution is updated every 50 rounds. The KL divergence is minimized through backpropagation to achieve joint optimization of encoder parameters and cluster centers until the loss function converges. The quality of the embedding vectors generated in this step directly determines the accuracy of subsequent result verification. The loss changes during training can be used to determine whether training needs to be stopped early (to prevent overfitting). The loss changes are monitored in real time during training. If the loss value does not decrease for 10 consecutive rounds and the fluctuation range is less than 0.001, or if the loss value shows an upward trend, training is stopped early based on the loss changes to further avoid model overfitting.
[0046] The specific calculation method for KL divergence is as follows: First, based on the low-dimensional embedding vector output by the encoder... With cluster center Calculate the predicted distribution (soft assignment probability). The t-distribution kernel function is used to measure the similarity between samples and cluster centers:
[0047]
[0048] In the formula, The number of clusters, is the degree of freedom parameter of the t-distribution (usually taken as 1).
[0049] Subsequently, the target distribution is constructed. To enhance the weighting of high-confidence assignments and suppress the influence of noisy samples:
[0050]
[0051] Finally, the KL divergence loss function is the predicted distribution. With target distribution Divergence between:
[0052]
[0053] By minimizing It can simultaneously drive the encoder to learn more discriminative embedding representations and make the cluster center distribution more closely match the real category structure, thereby improving the accuracy of subsequent verification stages.
[0054] This training scheme effectively improves the feature extraction capability of high-dimensional data through a symmetric autoencoder structure and a joint optimization strategy. At the same time, it avoids the model getting stuck in local optima by updating the target distribution adaptively, thus providing reliable low-dimensional embedded features for subsequent accurate clustering.
[0055] III. Result Verification
[0056] To fully verify the superiority of the method of the present invention, five different hyperparameter combinations were set up. Under each hyperparameter combination, the SADEC method of the present invention, the existing deep embedding clustering (DEC) method, and the GNN method were used for comparative experiments. Each experiment was repeated 10 times, and the mean ± standard deviation was taken as the final result. The evaluation indicators included clustering accuracy (ACC), normalized mutual information (NMI), adjusted RAND index (ARI), defect detection rate (DR), and false alarm rate (FAR).
[0057] The defect detection rate (DR) is defined as the proportion of samples correctly identified as defects out of the total number of true defect samples, and is calculated using the following formula:
[0058]
[0059] In the formula, TP (True Positive) is the number of correctly detected defect samples, and FN (False Negative) is the number of missed defect samples.
[0060] The false alarm rate (FAR) is defined as the proportion of normal samples that are incorrectly identified as defects out of the total number of true normal samples. The formula is:
[0061]
[0062] In the formula, FP (False Positive) represents the number of normal samples that were mistakenly identified as defects, and TN (True Negative) represents the number of normal samples that were correctly identified. These two indicators can be used to comprehensively measure the model's ability to detect defects and its level of control over the risk of misjudgment in real industrial scenarios.
[0063] 1. Clustering accuracy comparison experiment:
[0064] Experiments were conducted on the MNIST handwritten digit dataset using unsupervised clustering. ACC, NMI, and ARI were selected as evaluation metrics. By comparing the metrics of SADEC with DEC and GNN, the effectiveness of the improved adaptive clustering mechanism was clarified. Experimental results show that the SADEC method achieves a clustering accuracy of over 99% under all five hyperparameter configurations, with an error range of only ±0.1% to ±0.3%. In contrast, the highest accuracies of the comparison algorithms DEC and GNN are 97.8% and 98.6%, respectively, with greater error fluctuations. The comparison of validation accuracy of SADEC, DEC, and GNN on the MNIST dataset clearly demonstrates the performance advantage of SADEC.
[0065] 2. Industrial defect detection testing:
[0066] Using the NEU-DET metal surface defect dataset, we simulated the inspection requirements in a real industrial environment and added DR (defect detection rate) and FAR (false alarm rate) as core evaluation indicators (because in industrial scenarios, it is necessary to focus on the ability to identify real defects and the cost of misjudgment). On the NEU-DET dataset, SADEC's defect detection rate (DR) reached 94.2%, and the false alarm rate (FAR) was reduced to 5.8%, which is 67.6% and 49.1% lower than DEC and GNN, respectively.
[0067] 3. Noise robustness test:
[0068] Three typical industrial noises (Gaussian noise σ=0.1, impulse noise density 15%, and sensor drift offset 20%) were added to the NEU-DET dataset, and the decrease in ACC and NMI indices was monitored. The experimental results showed that the accuracy decrease of SADEC was only 4.2% to 6.3%, which was much lower than that of DEC (15.3% to 17.9%) and GNN (9.7% to 12.7%), proving that the channel attention mechanism and joint optimization strategy can effectively filter noise and overcome distribution offset. The SENet module obtains channel statistics through global average pooling and learns channel weights through fully connected layers. It enhances key discrimination channels and suppresses noise-dominated channels, thereby reducing the impact of uniform perturbations of Gaussian noise and local anomalies of impulse noise on features and preserving defect discrimination information. The cluster center update recalculates the cluster centers every 50 rounds based on the current embedding vector distribution, so that the centers can dynamically align with the feature distribution shift caused by sensor drift, maintaining compactness within clusters and separation between clusters, thereby offsetting the misclassification caused by the 20% shift. Therefore, SADEC maintains higher accuracy and robustness under complex working conditions.
[0069] 4. Parameter efficiency analysis:
[0070] The number of parameters for each algorithm was counted, the parameter accuracy ratio (ACC / params) was calculated, and the balance between model complexity and performance was evaluated. SADEC only requires 2.1M (M refers to millions of parameters) of parameters, and the parameter accuracy ratio (ACC / params) reaches 42.5% / M, which is significantly higher than DEC (35.6% / M) and GNN (21.7% / M).
[0071] The Wilcoxon signed-rank test was used to statistically analyze the ARI index. The results showed that the median ARI of SADEC was significantly higher than that of DEC and GNN (p<0.05), with a large effect size, which can exclude the influence of random factors. The test results directly reflect the significant difference between SADEC and the comparison algorithm in the adjusted RAND index (ARI).
[0072] The validation results corresponding to each parameter index are summarized and analyzed, and fed back to the parameter selection strategy during the network training stage. The parameter combination with the highest clustering accuracy, defect detection rate, and lowest false positive rate is prioritized, thus identifying the optimal parameter index and providing a basis for model configuration in subsequent practical applications. Through the above multi-dimensional validation, it is evident that the method of this invention exhibits excellent performance in clustering accuracy, industrial scenario adaptability, anti-interference capability, and parameter efficiency, enabling quantitative, automated, and accurate identification of mold damage states.
[0073] It should be noted that, in this invention, although the specification describes the embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A mold damage identification method based on improved adaptive deep embedding clustering, characterized in that, Includes the following steps: (1) Data acquisition and preprocessing: acquire mold vibration signals or surface defect images, scale and normalize the acquired data, and expand the sample diversity through data augmentation to obtain preprocessed data; (2) Constructing a deep neural network encoder: The deep neural network encoder integrates a squeezing and excitation module of channel attention mechanism, which is used to map the preprocessed high-dimensional data to a low-dimensional embedding space; (3) Joint optimization training: With the goal of minimizing KL divergence and reconstruction loss, the deep neural network encoder parameters and cluster centers are jointly optimized, and an adaptive clustering mechanism is adopted to improve feature separability; (4) Model evaluation and parameter determination: The model performance is verified by multi-dimensional evaluation indicators, the optimal model parameters are determined, and mold damage identification is realized.
2. The mold damage identification method based on improved adaptive deep embedding clustering according to claim 1, characterized in that, In step (1), the data enhancement includes at least one of random rotation and brightness adjustment.
3. The mold damage identification method based on improved adaptive deep embedding clustering according to claim 1, characterized in that, The deep neural network encoder consists of multiple stacked convolutional layers. After each convolutional operation, a batch normalization layer and an activation function are sequentially connected. It also integrates a squeezing and activation module with a channel attention mechanism. The channel features are squeezed through global average pooling, and then the channel weights are learned and activated through a fully connected layer.
4. The mold damage identification method based on improved adaptive deep embedding clustering according to claim 3, characterized in that, The activation function is the ReLU activation function, and the deep neural network encoder is equipped with a decoder. The decoder and encoder are designed symmetrically to reconstruct the input data.
5. The mold damage identification method based on improved adaptive deep embedding clustering according to claim 1, characterized in that, In step (3), an optimizer is used in conjunction with a learning rate scheduling strategy for training. The target distribution is updated periodically during the training process. If the loss value does not decrease within a preset number of consecutive rounds and the fluctuation range is less than a preset threshold, or if the loss value shows an upward trend, then training is stopped.
6. The mold damage identification method based on improved adaptive deep embedding clustering according to claim 5, characterized in that, The optimizer is the AdamW optimizer, and the learning rate scheduling strategy is the cosine annealing scheduling strategy.
7. The mold damage identification method based on improved adaptive deep embedding clustering according to claim 1, characterized in that, In step (3), by jointly optimizing encoder parameters and cluster centers, the model can cope with interference caused by data noise, impulse noise and sensor drift, thereby reducing the rate of accuracy decline.
8. The mold damage identification method based on improved adaptive deep embedding clustering according to claim 1, characterized in that, The deep neural network encoder has 2.1M parameters and a parameter accuracy ratio of 42.5% / M.
9. The mold damage identification method based on improved adaptive deep embedding clustering according to claim 1, characterized in that, It also includes incremental update steps: When new mold damage data is obtained or a new failure mode occurs, the new data is preprocessed in step (1) and the model is retrained to adapt to the dynamic changes in mold damage state.
10. The mold damage identification method based on improved adaptive deep embedding clustering according to claim 1, characterized in that, In step (4), the multi-dimensional evaluation metrics include at least two of the following: clustering accuracy, normalized mutual information, adjusted RAND index, defect detection rate, and false alarm rate.