A deep neural network-based ceramic tile durability evaluation method

Abstract

This invention relates to the field of ceramic tile performance data processing, and discloses a method for evaluating ceramic tile durability based on deep neural networks. First, ceramic tile sample data is collected and labeled. Then, a generative adversarial network based on a threat mechanism is used to augment the data. The augmented ceramic tile performance data is input into a feature extraction model for training. The feature-extracted ceramic tile performance data is then input into a feature dimensionality reduction model for training. Finally, the dimensionality-reduced ceramic tile performance data is input into a classifier model for training. The trained models are then used to evaluate the durability of new ceramic tile samples, and the durability evaluation result is finally output. This method effectively improves the efficiency and accuracy of data processing.

Description

Technical Field

[0001] This invention relates to the field of ceramic tile performance data processing, and specifically to a method for evaluating ceramic tile durability based on deep neural networks. Background Technology

[0002] In the building materials industry, ceramic tiles, as a widely used material, require durability assessment for ensuring building quality and safety. Traditional methods for assessing ceramic tile durability typically rely on laboratory testing and manual inspection. These methods are not only time-consuming and costly, but also often struggle to meet the demands of rapid production and high-standard quality control. Furthermore, traditional methods suffer from inefficiency and insufficient accuracy when processing large amounts of sample data. Therefore, how to quickly and accurately assess the durability of ceramic tiles, especially under varying usage environments and complex load conditions, has become a pressing technical challenge for the industry.

[0003] Chinese invention patent application number CN202210283283.X proposes a method and apparatus for improving tile image recognition through algorithms. The method includes receiving an image to be recognized; performing tile contour recognition on the image to be recognized; removing the background of the image to be recognized based on the tile contour using perspective transformation algorithm, affine transformation algorithm and / or image masking and cropping algorithm to obtain a target tile image; extracting features from the target tile image based on a neural network model; retrieving images in a library using the feature vector of the target tile image; and obtaining several tile images similar to the target tile image based on the image index, wherein the several tile images represent several known tile models.

[0004] By combining performance evaluation techniques from other fields, we can summarize the shortcomings of existing technologies:

[0005] 1. In existing technologies, the limited amount of sample data for tile performance often leads to insufficient model training, making tile performance evaluation models prone to overfitting when faced with new or unseen data.

[0006] 2. In terms of feature extraction, traditional ceramic tile performance evaluation methods usually rely on manually set parameters or simple algorithms, making feature extraction susceptible to local optima and difficult to fully capture the intrinsic features of the data.

[0007] 3. Feature reduction techniques often overlook the inherent structure of complex ceramic tile performance data, leading to the loss of important information during the reduction process.

[0008] 4. During the training of classification models, the lack of effective tile performance data preprocessing and sample labeling strategies often leads to poor classifier performance, especially when dealing with data with complex distribution characteristics. Summary of the Invention

[0009] The purpose of this invention is to provide a method for evaluating the durability of ceramic tiles based on deep neural networks. This method not only expands the data and prevents overfitting of the training model, but also improves the accuracy and efficiency of ceramic tile performance evaluation by improving the neural network.

[0010] A method for evaluating the durability of ceramic tiles based on deep neural networks includes the following steps:

[0011] Data collection and labeling of ceramic tile samples;

[0012] Data augmentation is performed using a threat-based generative adversarial network. The threat-based generative adversarial network includes training of a generator and a discriminator. The generator is used to generate tile performance data, and the discriminator is used to distinguish between real data and tile performance data generated by the generator.

[0013] The expanded ceramic tile performance data will undergo feature extraction and dimensionality reduction.

[0014] The dimensionality-reduced tile performance data is input into the classifier model for training.

[0015] The durability of new tile samples is evaluated using a pre-trained model. The newly collected tile performance data is sequentially passed through a pre-trained feature extraction model, a data dimensionality reduction model, and a classifier model to finally output the durability evaluation results.

[0016] Furthermore, the collection of tile sample data includes the following steps:

[0017] The data storage uses a structured JSON format;

[0018] The attributes of the test data include: pressure value, temperature value, humidity, impact force, frequency of use, duration of use, degree of surface wear, crack length, degree of color change, and material density.

[0019] Furthermore, the tile sample data is labeled as follows:

[0020] The collected data is manually labeled, and the labeling categories include: low durability level, medium durability level, and high durability level.

[0021] Furthermore, the training process for threat-based generative adversarial networks includes the following steps:

[0022] Initialize the network parameters of the generator and discriminator;

[0023] Select data covering key features from the labeled tile performance dataset as benchmark data;

[0024] Input tile performance data in batches, with each batch including real data and fake tile performance data generated by the generator;

[0025] In each batch, the generator produces new samples of tile performance data based on feedback from the discriminator;

[0026] The discriminator evaluates the authenticity of the tile performance data in each batch, calculates the error, and updates its own parameters based on the error.

[0027] The discrimination threshold is adaptively adjusted based on the performance of the discriminator.

[0028] Repeat the above steps until the preset stopping iteration condition is met.

[0029] Furthermore, the generator parameters are initialized using a standard normal distribution. and the parameters of the discriminator :

[0030]

[0031]

[0032] In the formula, This represents the initial parameters of the generator. Indicates the initial parameters of the discriminator. This represents a normal distribution with a mean of 0 and a standard deviation of 1.

[0033] Let the actual ceramic tile performance set be... The method for selecting a subset of data as the training baseline is as follows:

[0034]

[0035] In the formula, This indicates the selected benchmark dataset. Indicates from The operation of randomly selecting k samples;

[0036] In each training batch, the generated data is mixed with real data and input into the discriminator. The mixing operation of generated and real data is defined as follows:

[0037]

[0038]

[0039] In the formula, This represents the mixed dataset. Indicates a data connection operation. The generator outputs fake data, z, which is derived from the prior distribution. The noise vector obtained from sampling, This indicates that it conforms to a specific distribution.

[0040] Furthermore, the generator update, which produces new tile performance data samples, is constrained by the loss function:

[0041]

[0042] In the formula, Represents the loss function of the generator. Indicates the desired operation. Represents the logarithmic function. This represents the discriminant function of the discriminator;

[0043] This indicates that the cross-entropy loss function is used to calculate the difference between the generated tile performance data and the actual tile performance data. The calculation method is as follows:

[0044]

[0045] In the formula, This indicates that the discriminator reacts to the false data output by the generator. The rating, This is used to map the output of the discriminator to the (0, 1) interval, representing the probability of the generated data being genuine.

[0046] Furthermore, the discriminator calculates the error using an improved loss function, which is:

[0047]

[0048] In the formula, This represents the loss function of the discriminator. Indicates from real dataset Data obtained from sampling;

[0049] The discrimination threshold is adaptively adjusted in the following ways:

[0050]

[0051] In the formula, This indicates the updated discrimination threshold. It's the learning rate. This represents the gradient with respect to the discrimination threshold. It is the loss function of the discriminator;

[0052] The gradient is calculated as follows:

[0053] .

[0054] Furthermore, the classifier model is trained using a cluster-guided support vector machine classification algorithm, including the following steps:

[0055] The feature vectors obtained from the data dimensionality reduction model are input into the classifier model, and clustering operations are first performed on the input feature vectors.

[0056] Based on the clustering results, a pre-label is assigned to each data point, and these labels serve as the initial input for training the support vector machine.

[0057] Train a support vector machine using pre-labeled data;

[0058] The support vector machine module uses these pre-labeled tile performance data for supervised learning to achieve classification.

[0059] Furthermore, K-means or Gaussian mixture models are used to identify clusters in the data;

[0060]

[0061] In the formula, This represents the set of input feature vectors. This is the preset number of clusters, representing the number of data groups to be divided into. A cluster, These are the calculated cluster centers. This represents the cluster index to which each feature vector belongs;

[0062] Pre-labeled tags generated based on clustered indexes are represented as follows: , The cluster index is the feature vector obtained from the clustering step.

[0063] A support vector machine is trained using an adaptive adjustment strategy for the penalty parameter C and the kernel parameter gamma. This is achieved by solving the following optimization problem:

[0064]

[0065] Subject to

[0066] In the formula, It's weight. It's a bias. It is a penalty parameter that controls the strength of the penalty for misclassification. It is a slack variable. It's a tag. It is a feature mapping function.

[0067] Furthermore, slack variables To handle cases where the parts are not completely separable, the computation method is derived through the following optimization objective:

[0068]

[0069] In the formula, It is a sample The tag, It is a weight vector. It is a bias term. It is to put the data Functions mapped to higher-dimensional space;

[0070] The C and gamma parameters are dynamically adjusted based on the classification error rate. The adaptive adjustment strategy is as follows:

[0071]

[0072]

[0073] In the formula, and These are the penalty parameters before and after the update, respectively. It's the learning rate. This is the current error rate. It is the base of the natural logarithm. It is the decay rate parameter. It is the number of training iterations. It is the target error rate. and These are the kernel function parameters before and after the update, respectively. It is the attenuation factor;

[0074] Learning rate that adaptively adjusts the penalty parameter The model is dynamically adjusted based on performance feedback. The adjustment method is as follows:

[0075]

[0076] In the formula, and These are the learning rates before and after the update. It is an adjustment factor. This is the desired classification accuracy. It represents the accuracy of the current classifier.

[0077] The advantages of this invention are:

[0078] 1. By expanding the performance data of ceramic tiles through generative adversarial networks, the problem of insufficient training sample quantity is solved, and the model overfitting is effectively prevented. The data expansion is achieved through generative adversarial networks, which improves the model's generalization ability and prediction accuracy for unknown data.

[0079] 2. Feature extraction is performed using a neural network based on adaptive wave optimization. By simulating the propagation and energy distribution characteristics of waves, the weights and biases in the network are iteratively updated. This overcomes the limitation of traditional neural networks being susceptible to local optima in feature extraction, and enhances the accuracy of feature extraction and the robustness of the model.

[0080] 3. Feature dimensionality reduction is performed using an autoencoder neural network based on manifold mapping learning. By optimizing the representation of the feature space through the geometric properties of the Riemannian manifold, the ability to understand and reconstruct the local structure of the data is enhanced, key information is preserved during the dimensionality reduction process, and the efficiency of data processing is improved.

[0081] 4. Durability assessment is performed using a cluster-guided support vector machine classification algorithm. The clustering algorithm is used to pre-label training samples, which improves the accuracy and efficiency of tile performance assessment, achieves efficient data pattern recognition and classification, and reduces the possibility of misclassification. Attached Figure Description

[0082] Figure 1 This is a flowchart of the ceramic tile durability assessment method based on deep neural networks of the present invention;

[0083] Figure 2 This is a schematic diagram illustrating the training and application process of the present invention;

[0084] Figure 3 The training flowchart for the neural network feature extraction algorithm based on adaptive fluctuation optimization is shown below.

[0085] Figure 4 This is a flowchart of the training process for an autoencoder neural network algorithm based on manifold mapping learning. Detailed Implementation

[0086] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0087] This embodiment discloses a method for evaluating the durability of ceramic tiles based on deep neural networks. Please refer to [link / reference]. Figure 1 The steps include:

[0088] Data collection and labeling of ceramic tile samples;

[0089] Data augmentation is performed using a threat-based generative adversarial network. The threat-based generative adversarial network includes training of a generator and a discriminator. The generator is used to generate tile performance data, and the discriminator is used to distinguish between real data and tile performance data generated by the generator.

[0090] The expanded tile performance data is input into the feature extraction model for training.

[0091] The extracted tile performance data is then input into the feature reduction model for training.

[0092] The dimensionality-reduced tile performance data is input into the classifier model for training.

[0093] The durability of new tile samples is evaluated using a pre-trained model. The newly collected tile performance data is sequentially passed through a pre-trained feature extraction model, a data dimensionality reduction model, and a classifier model to finally output the durability evaluation results.

[0094] S1. Data Acquisition and Labeling

[0095] The data acquisition involved in this embodiment is mainly based on the systematic testing of tile samples under laboratory conditions and the monitoring of daily use environment. The test data is collected by high-precision sensors to collect the performance data of the tiles under different environmental and load conditions. The data is stored in a structured JSON format.

[0096] In this embodiment, the attributes of the test data include: ax is the pressure value (in N); ay is the temperature value (in °C); az is the humidity (percentage); aw is the impact force (in Joules); av is the usage frequency (times / day); au is the usage duration (hours); at is the surface wear degree (in mm); as is the crack length (in mm); ar is the degree of color change (percentage); and aq is the material density (in g / cm³).

[0097] It should be noted that this embodiment is only to illustrate one data format and type of the present invention. In practical applications, the data usually has more than 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0098] The collected data is labeled. The labeling method of this invention is manual labeling. In one embodiment, the labeling categories include: low durability level, medium durability level, and high durability level.

[0099] S2, Data Expansion

[0100] The acquisition, annotation, and preprocessing of training data related to tile performance are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor model generalization ability and affect model accuracy. This invention employs a threat-based generative adversarial network (GAN) algorithm for sample generation, thereby expanding the data. The threat-based GAN comprises two parts: a generator (G) and a discriminator (D). The generator is responsible for generating new data that is as close as possible to real tile performance data, while the discriminator attempts to distinguish between real data and the data generated by the generator. These two networks compete with each other during training, thereby improving their respective performance. This invention employs an adaptive threshold adjustment mechanism, making the judgment of data authenticity more refined and improving the ability to capture subtle features.

[0101] The training process of threat-based generative adversarial networks is as follows.

[0102] S201. Initialize the network parameters of the generator and discriminator. In this embodiment, a standard normal distribution is used to initialize the generator parameters. and the parameters of the discriminator :

[0103]

[0104]

[0105] In the formula, This represents the initial parameters of the generator. Indicates the initial parameters of the discriminator. This represents a normal distribution with a mean of 0 and a standard deviation of 1.

[0106] S202. Select a portion of the labeled tile performance dataset as benchmark data to ensure that the generated tile performance data covers key features. Let the actual tile performance set be , and the method for selecting a portion of the data as the training benchmark is as follows:

[0107]

[0108] In the formula, This indicates the selected benchmark dataset. Indicates from The operation of randomly selecting k samples.

[0109] S203. Input tile performance data in batches, with each batch including real data and fake tile performance data generated by the generator.

[0110] In each training batch, the generated data is mixed with real data and input into the discriminator. The mixing operation of generated and real data is defined as follows:

[0111]

[0112]

[0113] In the formula, This represents the mixed dataset. Indicates a data connection operation. The generator outputs fake data, z, which is derived from the prior distribution. The noise vector obtained from sampling, This indicates that it conforms to a specific distribution.

[0114] Preferably, Set it to a Gaussian distribution.

[0115] S204. In each batch, the generator attempts to generate new tile performance data samples based on feedback from the discriminator, with the goal of making it difficult for the discriminator to distinguish between genuine and fake data. The generator's updates are constrained by a loss function, calculated as follows:

[0116]

[0117] In the formula, Represents the loss function of the generator. Indicates the desired operation. Represents the logarithmic function. This represents the discriminant function of the discriminator.

[0118] This indicates that the cross-entropy loss function is used to calculate the difference between the generated tile performance data and the actual tile performance data. The calculation method is as follows:

[0119]

[0120] In the formula, This indicates that the discriminator reacts to the false data output by the generator. The rating, This is used to map the output of the discriminator to the (0, 1) interval, representing the probability of the generated data being genuine.

[0121] S205. The discriminator evaluates the authenticity of the tile performance data in each batch, calculates the error, and updates its own parameters based on the error. The improved loss function is:

[0122]

[0123] In the formula, This represents the loss function of the discriminator. Indicates from real dataset Data obtained from sampling.

[0124] Furthermore, The calculation method can be expressed as:

[0125] .

[0126] S206. Based on the performance of the discriminator, adaptively adjust the discrimination threshold to optimize the classification accuracy of the discriminator:

[0127]

[0128] In the formula, This indicates the updated discrimination threshold. It's the learning rate. This represents the gradient with respect to the discrimination threshold. It is the loss function of the discriminator, preferably, Set to 0.01.

[0129] The gradient is calculated as follows:

[0130] .

[0131] S207. Repeat the above steps until a preset stopping iteration condition is met, indicating that the model training is complete. In one embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000.

[0132] S3, Training of the Feature Extraction Model

[0133] The expanded ceramic tile performance data is input into the feature extraction model for training. This invention employs a neural network feature extraction algorithm based on adaptive wave optimization, inspired by wave phenomena. Waves exhibit periodic changes and energy transfer characteristics, enabling the transmission and transformation of energy forms in different media. The adaptive wave optimization algorithm iteratively updates the weights and biases in the network by simulating the propagation and energy distribution characteristics of waves. Each parameter update is affected not only by the current gradient but also by the cumulative influence of previous parameter update fluctuations, similar to the wave effects caused by superimposed wave sources. This makes the parameter update process smoother and more adaptive, effectively avoiding the oscillation and local minima problems found in traditional gradient descent methods.

[0134] For details, please refer to Figure 3 The training process of the neural network feature extraction algorithm based on adaptive fluctuation optimization is as follows.

[0135] S301. Define a neural network structure. The neural network has two hidden layers. The first hidden layer contains 100 neurons, and the second hidden layer contains 50 neurons. The activation function is the ReLU activation function.

[0136] The initial parameters of the neural network are set, including the weights and biases of the neural network, and the positions, intensities, and initial phases of the wave sources are initialized. The total number of wave sources is set to , then the initialization can be expressed as:

[0137]

[0138]

[0139] In the formula, For the i-th wave source, the position is... ,strength and the first meeting composition, Choose the initial position for the parameters. and The initial value is set randomly. As the initial weights, This is the initial bias.

[0140] S302. Calculate the loss L through forward propagation. Perform forward propagation on the input tile performance data X using parameters, calculate the output Y and compare it with the target value T. The output Y is the tile performance category output by the model, obtained by a preset Softmax function. The target value T is the true tile performance category corresponding to the sample. The loss L is calculated as follows:

[0141]

[0142]

[0143] In the formula, This represents the network structure, and the output layer of this network structure uses a pre-defined Softmax classifier to obtain the predicted label. For cross-entropy loss, is a parameter, and X is the input tile performance data.

[0144] S303. Calculate the gradient of each parameter based on the loss function:

[0145]

[0146] The gradient is calculated using the chain rule. The network output Y passes through the Softmax function, and the gradient calculation method with respect to the weights w is expressed as follows:

[0147]

[0148] In the formula, This is the output of Softmax. It is the linear output before the Softmax layer, where K is the number of categories;

[0149] in, , , , Indicates the target label.

[0150] S304. Based on the direction and magnitude of the gradient, generate wave sources at the corresponding parameter locations. Each wave source... The intensity has been updated to The calculation method is as follows:

[0151]

[0152] In the formula, It's the learning rate. The attenuation coefficient is... Here, is the amplification factor, and t is the number of iterations. Preferably, Set to 0.01.

[0153] S305. Superimpose the fluctuations of all wave sources in the parameter space. The update amount of each parameter is determined by the superposition result of the fluctuations of all wave sources at that point. The parameter update is as follows:

[0154]

[0155] In the formula, k is the wave number. It is angular frequency, which indicates the speed and direction of wave propagation; It is the ReLU activation function. For fluctuation threshold, To control the disturbance amount of wave source fluctuations.

[0156] Wave threshold for each wave source Dynamically adjusted based on layer sensitivity index:

[0157]

[0158] In the formula, It is the initial fluctuation threshold. The layer sensitivity index, It is an adjustment factor;

[0159] No. Layer sensitivity index The norm of the gradient output of this layer is calculated to measure the contribution of this layer to the overall network error. The calculation method is as follows:

[0160]

[0161] In the formula, It is the first The norm of the gradient of the layer parameters, It refers to the number of network layers.

[0162] S306. Based on the loss reduction in this iteration, dynamically adjust the attenuation coefficient of each wave source. and the growth rate coefficient :

[0163]

[0164]

[0165] In the formula, , These are the decay coefficient and the increase coefficient from the previous iteration, respectively. , These are the attenuation coefficient adjustment factor and the amplification coefficient adjustment factor, respectively. It is the change in loss. Preferably, the parameter and Set them to 0.1 and 0.05 respectively.

[0166] S307. Repeat the above steps until a preset stopping iteration condition is met, indicating that the model training is complete. In one embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000.

[0167] S4. Training of the Feature Dimensionality Reduction Model

[0168] The extracted tile performance data is input into a feature reduction model for training. This invention employs an autoencoder neural network based on manifold mapping learning for feature reduction. This autoencoder neural network not only learns the intrinsic structure of the data during the encoding stage but also attempts to reconstruct the input through the decoding process, ensuring that the reduced features retain key information. The autoencoder neural network model consists of one encoder and one decoder. The encoder maps high-dimensional input data to a low-dimensional feature space, while the decoder reconstructs the original input data from this low-dimensional space. This invention utilizes a Riemann factor based on manifold learning to optimize the representation of the feature space. By leveraging the geometric properties of the Riemannian manifold, a better understanding of the local structure of the data can be achieved, resulting in more accurate data reconstruction.

[0169] For details, please refer to Figure 4 The training process of the autoencoder neural network algorithm based on manifold mapping learning is as follows.

[0170] S401. Initialize network parameters. This invention uses a Gaussian distribution to initialize the weights w and biases b of the autoencoder, which can be expressed as:

[0171]

[0172]

[0173] In the formula, Represents the weight of the i-th layer. Indicates standard deviation, ,in, It represents the number of nodes in the input layer.

[0174] S402. During the encoding process, the input data is encoded and mapped to a low-dimensional feature space; during the decoding process, the decoder attempts to reconstruct the original data. Specifically, through the autoencoder, the tile performance data after feature extraction... Mapped to a low-dimensional feature space The method can be expressed as:

[0175]

[0176] The decoder attempts to reconstruct the original input. The method is as follows:

[0177]

[0178] In the formula, It is an activation function. and These are the encoder's weights and biases, respectively. and These are the weights and biases of the decoder.

[0179] In one embodiment, weight The calculation method, which is pre-defined using principal component analysis, can be expressed as follows:

[0180]

[0181] In the formula, From the covariance matrix The previous A matrix composed of eigenvectors It is the corresponding eigenvalue diagonal matrix.

[0182] S403. In the low-dimensional space of the encoding, the Riemann factor R is used to achieve manifold learning. The way the Riemann factor constrains the encoding of the feature z can be expressed as:

[0183]

[0184] In the formula, Indicates the preset manifold center, It is a parameter that controls the compactness of the manifold, preferably, Set as the mean vector of a low-dimensional space. Set it to 0.1.

[0185] S404. The loss function considers not only reconstruction error but also Riemannian manifold constraints. Reconstruction error ensures that the data can be effectively restored, while manifold constraints strengthen the geometric structure of the feature space. The loss function can be calculated as follows:

[0186]

[0187] In the formula, It is the regularization coefficient, preferably set to 0.3.

[0188] Reconstruction error The calculation method is as follows:

[0189]

[0190] In the formula, It is the value of z in the i-th dimension. It is the value of the pre-defined manifold center in the i-th dimension;

[0191] S405. Based on the loss function, the network parameters are adjusted using the backpropagation algorithm to optimize the performance of the encoder and decoder. (Based on weights) For example, its gradient can be calculated as follows:

[0192]

[0193] Furthermore, gradient Calculated using the error backpropagation method, it can be expressed as:

[0194]

[0195] and, Through decoder weights The transpose of can be represented as:

[0196]

[0197] Furthermore, since z is The result obtained by applying an activation function to a linear combination of x and y is then... The calculation method can be expressed as:

[0198]

[0199] In the formula, It is the derivative of the Sigmoid activation function. Represented by vector A diagonal matrix with diagonal elements. yes The transpose of .

[0200] Update parameters using gradient descent:

[0201]

[0202] In the formula, This is the learning rate, preferably set to 0.05, where t represents the number of iterations. For the updated weight parameters, These are the weight parameters before the update.

[0203] S406. Repeat the above steps until a preset stopping iteration condition is met, indicating that the model training is complete. In one embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000.

[0204] S5. Training the classifier model

[0205] The dimensionality-reduced tile performance data is input into the classifier model for training. This invention employs a cluster-guided support vector machine (SVM) classification algorithm. The clustering guidance module first performs clustering operations on the feature data to discover potential patterns and structures in the data, thereby providing pre-labeled training samples for SVM classification. The SVM module uses these pre-labeled tile performance data for supervised learning, thereby achieving efficient classification.

[0206] Specifically, the training process of the cluster-guided support vector machine classification algorithm is as follows.

[0207] S501. Input the feature vector obtained from the data dimensionality reduction model into the classifier model. First, perform a clustering operation on the input feature vector, using a K-means or Gaussian mixture model to identify clusters in the data. This can be represented as:

[0208]

[0209] In the formula, This represents the set of input feature vectors. This is the preset number of clusters, representing the number of data groups to be divided into. A cluster, These are the calculated cluster centers. This indicates the cluster index to which each feature vector belongs.

[0210] S502. Based on the clustering results, assign a pre-label to each data point. These labels serve as the initial input for support vector machine training. The pre-labels generated based on the cluster index are represented as follows: , This is the cluster index of the feature vectors obtained from the clustering step.

[0211] S503. Train a support vector machine using pre-labeled data. The training utilizes an adaptive adjustment strategy for the penalty parameter C and the kernel parameter gamma. This is achieved by solving the following optimization problem:

[0212]

[0213] Subject to

[0214] In the formula, It's weight. It's a bias. It is a penalty parameter that controls the strength of the penalty for misclassification. It is a slack variable. It's a tag. It is a feature mapping function;

[0215] Slack variables To handle cases where the parts are not completely separable, the computation method is derived through the following optimization objective:

[0216]

[0217] In the formula, It is a sample The tag, It is a weight vector. It is a bias term. It is to put the data Functions mapped to higher-dimensional space.

[0218] S504. Dynamically adjust the C and gamma parameters based on the classification error rate to minimize the error and maximize the marginal return. The adaptive adjustment strategy can be expressed as:

[0219]

[0220]

[0221] In the formula, and These are the penalty parameters before and after the update, respectively. It's the learning rate. This is the current error rate. It is the base of the natural logarithm. It is the decay rate parameter. It is the number of training iterations. The target error rate is preferably set to 0.05. and These are the kernel function parameters before and after the update, respectively. This is the attenuation factor, set to 0.95;

[0222] Learning rate that adaptively adjusts the penalty parameter The model is dynamically adjusted based on performance feedback. The adjustment method is as follows:

[0223]

[0224] In the formula, and These are the learning rates before and after the update. This is an adjustment factor, preferably set to 0.3. This is the desired classification accuracy, set to 0.95. It represents the accuracy of the current classifier.

[0225] S6, Tile Durability Assessment

[0226] The durability of new tile samples is evaluated using a trained model. In one embodiment, newly collected tile performance data is sequentially passed through a trained feature extraction model, a data dimensionality reduction model, and a classifier model to ultimately output a durability evaluation result. In this embodiment, the output evaluation results include: low durability level, medium durability level, and high durability level.

[0227] Based on this, the training and application process of the model is as follows: Figure 2 As shown.

[0228] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating the durability of ceramic tiles based on deep neural networks, characterized in that, Including the following steps: Data collection and labeling of ceramic tile samples; Data augmentation is performed using a threat-based generative adversarial network. The threat-based generative adversarial network includes training of a generator and a discriminator. The generator is used to generate tile performance data, and the discriminator is used to distinguish between real data and tile performance data generated by the generator. Feature extraction and dimensionality reduction were performed on the expanded ceramic tile performance data. The dimensionality-reduced tile performance data is input into the classifier model for training. The durability of new tile samples is evaluated using a trained model. The newly collected tile performance data is sequentially passed through a trained feature extraction model, a data dimensionality reduction model, and a classifier model to finally output the durability evaluation results. The training process of the threat-based generative adversarial network includes the following steps: (1) Initialize the network parameters of the generator and discriminator; (2) Select data covering key features from the labeled ceramic tile performance dataset as benchmark data; (3) Input tile performance data in batches, with each batch including real data and fake tile performance data generated by the generator; (4) In each batch, the generator produces new ceramic tile performance data samples based on the feedback from the discriminator; (5) The discriminator evaluates the authenticity of the ceramic tile performance data in each batch, calculates the error, and updates its own parameters based on the error; (6) Adaptively adjust the discrimination threshold based on the performance of the discriminator; Repeat steps (3) to (6) until the preset stopping iteration condition is met; The classifier model is trained using a cluster-guided support vector machine classification algorithm, including the following steps: The feature vectors obtained from the data dimensionality reduction model are input into the classifier model, and clustering operations are first performed on the input feature vectors. Based on the clustering results, a pre-label is assigned to each data point, and these labels serve as the initial input for training the support vector machine. Train a support vector machine using pre-labeled data; The support vector machine module uses these pre-labeled tile performance data for supervised learning to achieve classification; K-means or Gaussian mixture models can be used to identify clusters in the data; In the formula, This represents the set of input feature vectors. This is the preset number of clusters, representing the number of data groups to be divided into. A cluster, These are the calculated cluster centers. This indicates the cluster index to which each feature vector belongs; Pre-labeled tags generated based on clustered indexes are represented as follows: A support vector machine is trained using an adaptive adjustment strategy for the penalty parameter C and the kernel parameter gamma. This is achieved by solving the following optimization problem: Subject to In the formula, It is a weight vector. It is a bias term. It is a penalty parameter that controls the strength of the penalty for misclassification. It is a slack variable. It is a sample The tag, It is to put the data Functions mapped to higher-dimensional space; The slack variable To handle cases where the parts are not completely separable, the computation method is derived through the following optimization objective: ； The C and gamma parameters are dynamically adjusted based on the classification error rate. The adaptive adjustment strategy is as follows: In the formula, and These are the penalty parameters before and after the update, respectively. It's the learning rate. This is the current error rate. It is the base of the natural logarithm. It is the decay rate parameter. It is the number of training iterations. It is the target error rate. and These are the kernel function parameters before and after the update, respectively. It is the attenuation factor; Learning rate that adaptively adjusts the penalty parameter The model is dynamically adjusted based on performance feedback. The adjustment method is as follows: In the formula, and These are the learning rates before and after the update. It is an adjustment factor. This is the desired classification accuracy. It represents the accuracy of the current classifier.

2. The method for evaluating the durability of ceramic tiles based on deep neural networks according to claim 1, characterized in that, The collection of the ceramic tile sample data includes the following steps: The data storage uses a structured JSON format; The attributes of the sample data include: pressure value, temperature value, humidity, impact force, frequency of use, duration of use, degree of surface wear, crack length, degree of color change, and material density.

3. The method for evaluating the durability of ceramic tiles based on deep neural networks according to claim 1, characterized in that, The tile sample data is labeled as follows: The collected data is manually labeled, and the labeling categories include: low durability level, medium durability level, and high durability level.

4. The method for evaluating the durability of ceramic tiles based on deep neural networks according to claim 1, characterized in that, The generator parameters are initialized using a standard normal distribution. and the parameters of the discriminator : In the formula, This represents the initial parameters of the generator. Indicates the initial parameters of the discriminator. This represents a normal distribution with a mean of 0 and a standard deviation of 1. Let the actual ceramic tile performance set be... The method for selecting a subset of data as the training baseline is as follows: In the formula, This indicates the selected benchmark dataset. Indicates from The operation of randomly selecting k samples; In each training batch, the generated data is mixed with real data and input into the discriminator. The mixing operation of generated and real data is defined as follows: In the formula, This represents the mixed dataset. Indicates a data connection operation. The generator outputs fake data, z, which is derived from the prior distribution. The noise vector obtained from sampling.

5. The method for evaluating the durability of ceramic tiles based on deep neural networks according to claim 4, characterized in that, The generator update, which produces new tile performance data samples, is constrained by the loss function: In the formula, Represents the loss function of the generator. Indicates the desired operation. Represents the logarithmic function. This represents the discriminant function of the discriminator; This indicates that the cross-entropy loss function is used to calculate the difference between the generated tile performance data and the actual tile performance data. The calculation method is as follows: In the formula, This indicates that the discriminator reacts to the false data output by the generator. The rating, This is used to map the output of the discriminator to the (0, 1) interval, representing the probability of the generated data being genuine.

6. The method for evaluating the durability of ceramic tiles based on deep neural networks according to claim 4, characterized in that, The discriminator calculates the error using an improved loss function, which is: In the formula, This represents the loss function of the discriminator. Indicates from real dataset Data obtained from sampling; The discrimination threshold is adaptively adjusted in the following ways: In the formula, This indicates the updated discrimination threshold. It's the learning rate. This represents the gradient with respect to the discrimination threshold. It is the loss function of the discriminator; The gradient is calculated as follows: 。

Images