Waste household appliance recovery value evaluation method based on BC-TabNet model

Abstract

The invention provides a waste household appliance recovery value evaluation method based on a BC-TabNet model. According to the BC-TabNet model, the feature selection capability of TabNet, the local feature extraction advantage of a one-dimensional convolutional neural network (1D-CNN) and the nonlinear modeling characteristic of a BP neural network are fused, and the collaborative advantage of each module is exerted. The system comprises a data preprocessing module, a model training module and a value evaluation module. The data preprocessing module performs abnormal value processing, feature coding, standardization and principal component analysis on the original data; the model training module is used for training a BC-TabNet model by using the preprocessed data; and the value evaluation module calls the trained model to evaluate the value of the waste household appliance. Various factors such as brands, types, product types, capacities and service lives of the waste household appliances are comprehensively considered, and the method is suitable for various types of waste household appliances.

Description

Technical Field

[0001] This invention relates to the field of waste household appliance recycling value assessment technology, and in particular to a waste household appliance recycling value assessment method based on the BC-TabNet model. Background Technology

[0002] With economic development and rising living standards, the pace of home appliance replacement is accelerating, leading to a continuous increase in the number of discarded appliances. The government strongly advocates a circular economy and is continuously improving the recycling system for discarded home appliances. As a crucial link in this system, the accuracy and efficiency of the valuation of discarded appliances directly impact recycling efficiency. However, currently, most home appliance recycling companies rely solely on the appearance and condition of discarded appliances for valuation, resulting in low accuracy and requiring on-site assessments by professionals. This approach is not only costly and inefficient but also struggles to handle the large-scale data from the recycling market.

[0003] In recent years, artificial intelligence-based methods for assessing the value of discarded household appliances have been gradually proposed. These methods utilize historical data of discarded appliances to establish nonlinear mapping models by mining their attribute characteristics, effectively improving the accuracy of value assessment and recycling efficiency. However, existing methods still have many shortcomings in assessing the recycling value of discarded appliances. First, the research has not fully considered the complexity of the multidimensional attributes of discarded appliances, and its ability to capture complex attributes is insufficient. Second, the models need improvement in computational accuracy, complexity, and adaptability. The value assessment of discarded appliances involves mixed-type data (such as structured text categorical variables and continuous variables), with high data dimensionality and information redundancy, and complex interactions between attributes. Therefore, a value assessment method for discarded appliances that can accurately capture their attribute characteristics is needed to improve the accuracy of value assessment and recycling efficiency. Summary of the Invention

[0004] This invention aims to address the problems existing in current methods for assessing the recycling value of waste household appliances. It proposes a method based on the BC-TabNet model, comprehensively considering appliance attributes such as brand, type, product type, capacity, and service life to fully evaluate the recycling value of waste household appliances. The model integrates TabNet, 1D-CNN, and BP neural networks. Through TabNet's sparse attention mechanism, it automatically selects the most important features for value prediction and assigns them higher weights, thereby achieving automatic feature screening and importance assessment. 1D-CNN captures local feature patterns and sequence information, uncovering hidden price correlations and enhancing feature representation capabilities. Finally, BP neural networks further integrate and process these complex feature information, using neurons and a multi-layer architecture to fit nonlinear mapping relationships, fully capturing appliance attribute features, and thus improving the accuracy of price prediction. This invention enables automated processing, significantly improving assessment efficiency compared to manual assessment methods, and enhancing the accuracy and recycling efficiency of waste household appliances.

[0005] A method for assessing the recycling value of waste household appliances based on the BC-TabNet model is characterized by the following steps: (1) Data preprocessing cleans and transforms the original waste household appliance recycling data, including removing abnormal data, encoding and transforming categorical variables, standardizing data features, and reducing the dimensionality of the data through principal component analysis to obtain a structured data feature representation; (2) The BC-TabNet model is constructed using three modules: TabNet, 1D-CNN, and BP neural network. The preprocessed data is sequentially processed by the three modules of TabNet, 1D-CNN, and BP neural network to complete the model training; (3) The trained BC-TabNet model is called, and the relevant data of the waste household appliances to be assessed are input into the system. The model automatically calculates and outputs the assessment value based on the previously learned feature patterns.

[0006] Through the above process, step (1) effectively improves data quality and lays the foundation for subsequent model training; step (2) the synergistic effect between modules enables the model to fully learn data features; step (3) the trained model is used to achieve efficient and accurate assessment of the recycling value of waste home appliances, improving the accuracy and efficiency of the assessment. Attached Figure Description

[0007] Figure 1 Process framework of this invention Figure 2 Home appliance attribute coding method Figure 3 TabNet structure Figure 4 Experimental results of this invention Detailed Implementation The technical solutions provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description. Figure 1 This is the process framework of the present invention, and the specific steps are as follows: Step 1. Data preprocessing: including outlier handling, feature encoding, standardization, and principal component analysis. Step (11), outlier handling: outlier detection and missing value handling are performed on the original data. Missing values ​​are padded with 0s, and outliers that deviate significantly from most of the data are directly deleted to ensure the integrity and accuracy of the data.

[0008] Step (12), Feature Encoding: Different encoding methods are used according to different home appliance attributes. For nominal variables (brand, type, product type), since they have no inherent order and a large number of categories, one-hot encoding would generate high-dimensional features, increasing the computational burden. Therefore, binary encoding is used to compress the dimension. First, the categories are mapped to integers (e.g., Haier corresponds to 0, Hisense to 1, Xiaomi to 2), and then converted to 4-bit binary (e.g., Haier is encoded as "0000", Hisense as "0001", Xiaomi as "0010"). For variables with sequential relationships (capacity, service life), label encoding is used, mapping the categories to integers (e.g., "121-199L" is encoded as "1", "200-499L" as "2", "500L and above" as "3"), preserving the sequential information. The specific encoding method is as follows: Figure 2 As shown.

[0009] Step (13), Data Standardization: To remove the dimensional differences between different features, Min-Max standardization is used to scale the features to the [0,1] interval. The standardization formula is as follows: in, The original data, Indicates the maximum value. This represents the minimum value.

[0010] Step (14), Principal Component Analysis: High-dimensional data may lead to model overfitting and low computational efficiency. Therefore, PCA is used for dimensionality reduction to screen principal components with a contribution rate greater than 90%. Standardized training set. Include One sample, Each feature, each column vector To represent a feature, and It is expressed as follows: Calculate the mean of the data and standard deviation The standard matrix is ​​obtained. The standardization process for each element is as follows: in, , .

[0011] Solve the correlation coefficient matrix The calculation method is as follows: Solve the correlation matrix The characteristic equation is obtained by solving it. eigenvalues The solution process is as follows: Calculate the cumulative contribution rate of eigenvalues Principal components with a cumulative contribution rate greater than 0.9 were selected for model input, and the calculation process is as follows: Step 2. Model Training Model training includes three modules: TabNet, 1D-CNN, and BP neural network, as detailed below: In step (21), the TabNet module generates the final prediction by weighting the results of multiple decision steps, and its structure is as follows: Figure 3 As shown. Each decision step consists of a Feature transformer, an Attentive transformer, a Mask layer, a Split layer, and a ReLU activation function.

[0012] In step (211), the Feature transformer component extracts and transforms important features from the high-dimensional structured data, providing high-quality input data for subsequent decision-making steps. The Feature transformer includes batch normalization (BN), gated linear units (GLU), and fully connected layers. The GLU enhances feature representation through a gating mechanism, calculated as follows: Step (212) involves the Split layer dividing the vector output by the Feature transformer into two parts. One part is used in the subsequent decision-making process, and the other part is used for the attention mechanism. This ensures a reasonable allocation of feature information and improves the model's prediction accuracy and feature selection ability. The splitting process is as follows: in, Used to calculate the final output of the model. The decision is passed to the Mask layer in the next decision step.

[0013] In step (213), the Attentive transformer generates a sparse and unique mask matrix based on the output of the previous decision step to achieve dynamic feature selection. The attention weights are normalized using Sparsemax, and the mask matrix is ​​calculated as follows: in, Indicates the frequency of feature usage. These are the feature vectors after being partitioned by the Split layer in the previous step. This is a trainable function based on BN and FC layers.

[0014] Step (22): 1D-CNN is responsible for further extracting local features from the high-dimensional structured data output by the TabNet module. 1D-CNN mainly extracts local features through one-dimensional convolutional kernels. The convolutional kernel slides across the input data with a fixed stride, and the kernel is calculated as follows: in, Indicates the number of input feature maps. and They represent the first The first in the layer The weights and biases of each convolutional kernel. For the first The first in the layer A feature sequence, For the first The first in the layer Each feature is mapped. The convolutional features are sequentially passed through BN, ReLU, and Dropout layers to further optimize the feature representation.

[0015] Step (23): The BP neural network consists of an input layer, hidden layers, and an output layer. Through its multi-layered structure and activation functions, it captures the complex nonlinear relationships between the output features of the 1D-CNN modules. The forward propagation process is as follows: in, For the first Layer The first neuron to the second Layer The weights of each neuron, For the first Layer Bias of each neuron.

[0016] Step (3) Value Assessment Module Step (31) After the model training is completed, the learned parameters and weights will be saved so that they can be called repeatedly without having to go through the time-consuming training process again.

[0017] Step (32): When it is necessary to use the model to carry out evaluation work, first load the previously saved model parameters and weights into the system from the storage location to prepare for accurate evaluation.

[0018] Step (33) switches the model from training mode to evaluation mode, enabling the model to accurately evaluate new data in its optimal state. The model will use the learned feature patterns to make reasonable inferences and judgments based on the characteristics of the input data.

[0019] Step (34): Input the relevant data of the waste household appliances to be evaluated into the model that has completed the above preparation work. The model will analyze and calculate the input data based on the knowledge and rules it has learned during the training process, and finally output the evaluation value of the waste household appliances. Figure 4 This paper presents the fitting curves of the predicted values ​​and actual values ​​of the model of this invention compared to five other models (BC-TabNet (without PCA), LSTM, XGBoost, LightGBM, and SVM) on the same dataset. Figure 4 As can be seen, the predicted scatter points of BC-TabNet (PCA) are closely distributed near the ideal fitted curve, indicating high prediction accuracy and good fitting effect, especially stable performance in the small to medium value range. In contrast, BC-TabNet (without PCA) performs poorly, possibly due to feature redundancy caused by the lack of principal component analysis. LSTM predicts more accurately in the high value region, but the scatter point deviation is large in the small to medium value range, reflecting its insufficient adaptability to imbalanced data. XGBoost, LightGBM, and SVM have moderate performance, and their scatter point distribution is relatively scattered.

Claims

1. A method for assessing the recycling value of waste household appliances based on the BC-TabNet model, characterized in that, The method includes: Step (1) Data preprocessing cleans and transforms the original waste household appliance recycling data, including removing abnormal data, encoding and transforming categorical variables, standardizing data features, and reducing the dimensionality of the data through principal component analysis, thereby obtaining a structured data feature representation. Step (2) involves constructing the BC-TabNet model using three modules: TabNet, 1D-CNN, and BP neural network. The preprocessed data is then processed sequentially through these three modules to complete the model training. Step (3): Call the trained BC-TabNet model, input the relevant data of the waste home appliances to be evaluated into the system, and the model automatically calculates and outputs the evaluation value based on the previously learned feature rules.

2. The method according to claim 1, characterized in that: data preprocessing; In step 1, data preprocessing includes outlier handling, feature encoding, data standardization, and principal component analysis, as detailed below: Step (11), abnormal data processing: outlier detection and missing value processing are performed on the original data; missing values ​​are padded with 0, and outliers that deviate significantly from most of the data are directly deleted to ensure the integrity and accuracy of the data; Step (12), Feature Encoding: Different encoding methods are used according to different home appliance attributes; for nominal variables, including brand, type and product type, binary encoding is used to compress dimensions; for variables with sequential relationships, including capacity and service life, label encoding is used to map categories to integers and retain sequence information. Step (13), Data Standardization: To remove the dimensional differences between different features, Min-Max standardization is used to scale the features to the [0,1] interval; the standardization formula is as follows: in, The original data, Indicates the maximum value. This represents the minimum value. Step (14), Principal Component Analysis: High-dimensional data may lead to model overfitting and low computational efficiency, so PCA is used for dimensionality reduction to screen principal components with a contribution rate greater than 90%; the standardized training set Include One sample, Each feature, each column vector To represent a feature, and It is expressed as follows: Calculate the mean of the data and standard deviation The standard matrix is ​​obtained. The standardization process for each element is as follows: in, , ; Solve the correlation coefficient matrix The calculation method is as follows: Solve the correlation matrix The characteristic equation is obtained by solving it. eigenvalues The solution process is as follows: Calculate the cumulative contribution rate of eigenvalues Principal components with a cumulative contribution rate greater than 0.9 were selected for model input, and the calculation process is as follows: 。 3. The method according to claim 1, characterized in that: in step 2, the model structure and model training are as follows: The model consists of three modules: TabNet, 1D-CNN, and BP neural network, as detailed below: in, The TabNet module generates the final prediction by weighting the results of multiple decision steps, where each decision step consists of a Feature transformer, an Attentive transformer, a Mask layer, a Split layer, and a ReLU activation function. The Feature Transformer component extracts and transforms important features from high-dimensional structured data, providing high-quality input data for subsequent decision-making steps. The Feature Transformer includes Batch Normalization (BN), Gated Linear Units (GLUs), and fully connected layers. The GLUs enhance feature representation through a gating mechanism, calculated as follows: The Split layer divides the vector output by the Feature transformer into two parts: one part is used in the subsequent decision-making process, and the other part is used in the attention mechanism. This ensures the reasonable allocation of feature information and improves the model's prediction accuracy and feature selection ability. The splitting process is as follows: in, Used to calculate the final output of the model. The decision is passed to the Mask layer in the next decision step; The Attentive transformer generates a sparse and unique mask matrix based on the output of the previous decision step to achieve dynamic feature selection; the attention weights are normalized using Sparsemax, and the mask matrix is ​​calculated as follows: in, Indicates the frequency of feature usage. These are the feature vectors after being partitioned by the Split layer in the previous step. This is a trainable function based on BN and FC layers. The 1D-CNN is responsible for further extracting local features from the high-dimensional structured data output by the TabNet module. The 1D-CNN mainly extracts local features using one-dimensional convolutional kernels. The convolutional kernels slide on the input data with a fixed stride, and the kernel calculation is as follows: in, Indicates the number of input feature maps. and They represent the first The first in the layer The weights and biases of each convolutional kernel. For the first The first in the layer A feature sequence, For the first The first in the layer Each feature map; the convolutional features are sequentially passed through BN, ReLU, and Dropout layers to further optimize the feature representation; The BP neural network consists of an input layer, hidden layers, and an output layer. Through its multi-layered structure and activation functions, it captures the complex nonlinear relationships between the output features of 1D-CNN modules. The forward propagation process is as follows: in, For the first Layer The first neuron to the second Layer The weights of each neuron, For the first Layer Bias of each neuron.

4. The method according to claim 1, characterized in that: value assessment; Step 3, the value assessment includes saving model weights, loading model parameters, setting the assessment mode, and final price prediction, as detailed below: Step (31) After the model training is completed, the learned parameters and weights will be saved so that they can be called repeatedly without having to go through the time-consuming training process again. Step (32): When it is necessary to use the model to carry out evaluation work, first load the previously saved model parameters and weights into the system from the storage location to prepare for accurate evaluation. Step (33) switches the model from training mode to evaluation mode, enabling the model to accurately evaluate new data in its optimal state. The model will use the learned feature patterns to make reasonable inferences and judgments based on the characteristics of the input data. Step (34): Input the relevant data of the waste household appliances to be evaluated into the model that has completed the above preparation work. The model will analyze and calculate the input data based on the knowledge and rules it has learned during the training process, and finally output the evaluation value of the waste household appliances.

Images