A method for predicting machining of a discrete manufacturing part based on ASPP-FCN

Abstract

The present application relates to the technical field of machining time prediction, and particularly relates to an intelligent manufacturing machining time prediction method based on ASPP-FCN, comprising the following steps: S1, adopting a group decision method of expert effective information, combining index weight calculation and consistency test under interval number conditions to select 8 work time features with the highest weight; S2, adopting LabeEncoder coding and adding a spatial remapping mechanism to convert features into pseudo images; S3, adopting an improved ASPP module to extract key information of work time features of special vehicle self-made parts; S4, building a fully connected network layer to linearly map features and output prediction results; S5, model training and feasibility verification. The present application establishes a structured spatial relationship between process features and work time consumption by adopting structured pseudo image reconstruction and spatial coding mechanism of discrete process features; based on the ASPP-FCN model, the present application carries out machining time prediction research of special vehicle self-made parts. The present application solves the defects of low efficiency, large error and insufficient generalization ability of existing methods for work time prediction of small-batch customized special vehicle self-made parts.

Description

Technical Field

[0001] This invention belongs to the field of machining time prediction technology, specifically relating to a method for predicting the machining time of discrete manufacturing parts based on ASPP-FCN. Background Technology

[0002] In the current era of rapid development in intelligent manufacturing, enterprises are constantly pursuing a production model that emphasizes high efficiency, high quality, and low cost. In this process, time quota management is not only a crucial basis for production planning and schedule control, but also a core element for enterprises in economic accounting, cost management, and product pricing. Silva et al. pointed out that accurately predicting the operation time at each stage, including machine preparation, actual processing, and finishing, is the core foundation for building an economical machining cost estimation tool. For self-manufactured parts for special vehicles, which are diverse, complex in process, and highly customized, establishing a scientific and reasonable time quota system is crucial for optimizing resource allocation, shortening delivery cycles, and improving quality stability. Therefore, conducting research on machining time prediction for special vehicle parts, constructing a scientific prediction system, and developing an efficient, fast, and accurate time prediction method have become urgent needs.

[0003] Neural networks have demonstrated significant advantages in industrial forecasting by analyzing influencing factors during machining and establishing nonlinear mapping relationships between process parameters and machining time. Michal Matějka et al. addressed the issues of data fragmentation and multi-feature coupling in customized production for SMEs, summarizing the evolution of machine learning in time prediction and highlighting the enormous potential of multi-objective tensor regression and deep sequence models (such as LSTM / CNN) in capturing BOM structure and process dynamics. To overcome the prediction bias of traditional CAM software, Brillinger et al. abandoned traditional analytical models and directly extracted composite features such as toolpath length, feed rate, and machine tool status from NC code, achieving high-precision prediction using ensemble learning algorithms. Chien et al. and Chao Sun et al. both focused on the prediction distortion problem caused by CAM software neglecting the CNC system's interpolation mechanism or machine tool kinematic constraints. The former constructed a bidirectional long short-term memory network (BiLSTM) by extracting G-code features to learn the look-ahead control behavior of the CNC system, while the latter used a feedforward neural network to construct prediction models for each feed axis of the machine tool. Both significantly improved the prediction accuracy of machining cycles for complex toolpaths and thin-walled structural parts. Furthermore, Tong Zhu et al. successfully predicted machining times for different interpolation types in ultra-precision milling by combining feedforward neural networks with the geometric and feed parameters of the NC program. Regarding the improvement and optimization of prediction algorithms, researchers have proposed innovative solutions for different application scenarios. For the special characteristics of aerospace component manufacturing, Liu Juan proposed a feature extraction method based on BERT and K-Means, combined with a GA-BP neural network to improve the accuracy of time quotas; Jia Panpan applied convolutional neural networks to solve the time error problem in small-batch discrete manufacturing. For non-standard shaft parts, Xiang Feng et al. proposed a prediction method combining feature integration, improved clustering, and optimized BP neural networks. To achieve more realistic time assessments, some scholars have further expanded the input feature dimensions of the prediction model. Saric et al. considered discrete features of the production site, such as part complexity, number of processes, operator experience, and equipment anomalies, and comprehensively compared and verified the effectiveness of various neural networks, including MNN, BP, and RBFNN, in actual production time assessment. André Rodrigues et al. achieved good results in predicting the processing time of plastic injection molds by using macroscopic physical features such as workpiece material, number of features, and removal volume as inputs to a neural network. However, their model mainly targets molds with standardized geometric features and struggles to handle the strong multi-scale coupling relationships between complex features. Compared to the workpieces predicted in the above time prediction, self-made parts for special vehicles have higher geometric complexity, involve more time features, and have more complex relationships between features. Summary of the Invention

[0004] The purpose of this invention is to provide a fully connected neural network prediction model based on ASPP (ASPP-FCN) for predicting the working time of small-batch customized special vehicle parts. Traditional methods suffer from drawbacks such as low efficiency, large errors, and insufficient generalization ability. Furthermore, traditional neural networks also face problems such as weak feature interaction and easy degradation of high-dimensional features. This invention maps the working time features into a multi-dimensional feature matrix by using spatial encoding technology; introduces an improved ASPP module to achieve parallel extraction of key features at multiple scales; and finally, the fully connected layer completes the regression prediction of the working time.

[0005] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:

[0006] A smart manufacturing machining time prediction method based on ASPP-FCN includes the following steps:

[0007] S1: Using the group decision-making method with effective expert information, and combining the index weight calculation and consistency test under the interval number condition, the 8 working time features with the highest weight after the decision are selected as the input of the neural network.

[0008] S2: Based on the characteristics of the 8 key features of the input working hours, LabelEncoder is used to process the categorical variables, and a spatial remapping mechanism is added to further convert the encoded and normalized features into pseudo-images that can be extracted by the improved ASPP module.

[0009] S3: Based on the pseudo-image extracted from the time features, an improved ASPP module is used to extract key information of the time features of self-made parts for special vehicles and improve the parameter response of the enhanced key features, and solve the problems of sparsity and scale limitation of time features in the data.

[0010] S4: Using the pseudo-image of enhanced features obtained from the ASPP module, a fully connected network layer is built to linearly map the features and output the prediction results;

[0011] S5: Build an experimental platform to test the effectiveness of the ASPP-FCN time prediction model established in the above steps.

[0012] Furthermore, step S1 specifically includes: summarizing and organizing the characteristic factors of machining process time, obtaining a total of 16 characteristic factors;

[0013] By using a group decision-making method based on effective expert information, and combining the index weight calculation and consistency test under interval number conditions, the eight time features with the highest weights were obtained, including "job type", "processing parameters", "processing surface form", "accuracy level", "raw material", "machine tool type", "tool type" and "raw material type".

[0014] Furthermore, step S2 specifically includes:

[0015] For the eight key features of the work hours obtained, which are non-numerical information and cannot be directly input into the neural network for training, LabelEncoder is used to process the categories. Each category value is mapped to a unique integer, effectively preserving the discriminative power between categories and simplifying feature representation. After encoding, the input features are normalized to accelerate the convergence speed of the neural network and effectively avoid gradient explosion or vanishing problems caused by excessive differences in the dimensions of different features, thereby improving the stability of model training and prediction accuracy.

[0016] Building upon the above, a spatial remapping mechanism is added to further convert the encoded and normalized features into pseudo-images from which the improved ASPP module can extract features. The spatial remapping mechanism consists of two fully connected layers connected in series, mapping the original 8-dimensional machining process key features to a 128-dimensional high-dimensional feature space to enhance feature representation. Subsequently, this 128-dimensional feature vector is rearranged according to channel and spatial structure, transforming it into a 16-channel × 2-row × 4-column feature pseudo-image form, providing structured input for the subsequent parallel execution of multi-scale dilated convolution operations by the ASPP module.

[0017] Furthermore, step S3 specifically includes:

[0018] To fully extract key information on the time characteristics of self-made parts for special vehicles and enhance the parameter response of key features, as well as to address the sparsity and scale limitations of time characteristics in the data, the ASPP module is improved.

[0019] To address training stability and feature representation capabilities, a batch normalization (BN) layer is introduced into the multi-scale dilated convolution branch to mitigate internal covariate shifts, stabilize the training process, accelerate model convergence, and improve overall machining time robustness. To enhance the overall feature representation capability of the time-related pseudo-images, a global average pooling score is added to capture more complete global contextual information.

[0020] Furthermore, step S4 specifically includes:

[0021] The multidimensional tensor of the enhanced feature pseudo-image is flattened into a one-dimensional vector. The flattened vector is linearly mapped to 128 dimensions through the first fully connected layer and then activated by the ReLU function. The nonlinear combination features in the input features are extracted through linear transformation and activation function, preparing for subsequent dimensionality reduction and prediction. The 128-dimensional output of the first layer is further reduced to 64 dimensions through the second fully connected layer and then activated by the ReLU function again to further refine and compress the feature information, retaining the most useful information for prediction while reducing the data dimensionality. Finally, the 64-dimensional features are mapped to the final one-dimensional output dimension (work time) through the third fully connected layer, and the predicted work time value is finally output.

[0022] Furthermore, step S5 specifically includes:

[0023] Based on the actual needs of predicting the working hours of customized parts for special vehicles, and using Windows and NVIDIA RTX4046 GPU hardware, Python and PyTorch 2.4.1 were selected for framework construction, model effectiveness experiments, module ablation experiments, model convergence analysis, and multi-algorithm comparison.

[0024] The beneficial effects of this invention are:

[0025] In predicting the production time of small-batch customized special vehicle parts, traditional manual experience and table lookup methods suffer from low efficiency and large errors. The production time prediction method based on ASPP-FCN proposed in this invention can effectively solve the above problems, with the following core advantages:

[0026] This invention processes categorical variables using LabelEncoder and incorporates a spatial remapping mechanism to further convert the encoded and normalized features into pseudo-images that can be extracted by the improved ASPP module. The improved ASPP module extracts key information about the machining time features of self-made parts for special vehicles and enhances the parameter response of key features, while addressing the sparsity and scale limitations of machining time features in the data. Finally, a fully connected network layer is constructed to linearly map the features and output prediction results. Experiments show that after training cycles, the ASPP-FCN model maintains a goodness of fit of 0.9985, the training mean squared error decreases to 29.6, and the mean absolute error decreases to 4.4, indicating that the model can accurately capture important features during machining and effectively describe the nonlinear mapping between machining time features.

[0027] Model with spatial coding layer enabled vs. without spatial coding layer enabled R 2The MSE decreased from above 0.9900 to below 0.2522; the introduction of ASPP-global reduced the MSE from 120.2 to 83.9, a decrease of 30.2%, indicating that its multi-scale dilated convolution effectively captures the local-global correlation of process features, and that the global pooling branch can effectively capture the main influencing factors of process time for features sensitive to feature scale; the spatial coding layer + SE reduced the MSE by 4.2% compared to the spatial coding layer alone; while using the SE module alone cannot fully learn the meaningful channel weight distribution, and may even introduce interference information. The above experiments show that the combined modules work together to produce a positive synergistic effect.

[0028] This study compares widely used regression neural network models: Fully Connected Feedforward Neural Network (FNN), Convolutional Neural Network (CNN), Residual Network (ResNet), Long Short-Term Memory (LSTM) Network, and Standard-ASPP. The results of both the training and test sets of the neural network models are compared with those of the optimized models. After 160 training epochs, the training MSE of the ASPP-FCN model decreased to 29.9, significantly lower than LSTM's 158.2, CNN's 174.8, ResNet's 112.8, FNN's 105.9, and Standard-ASPP's 101.4. Compared to LSTM, ASPP-FCN achieved approximately 81.09% reduction in training error, and even under the relatively good Standard-ASPP baseline, it still achieved approximately 70.53% error compression. Furthermore, the comparison of MSE with other models further demonstrates the generalization ability of ASPP-FCN. Finally, in terms of the MAE index, the ASPP-FCN model achieved a MAE index of 4.4, which is better than LSTM (5.5), CNN (13.5), ResNet (7.4), FNN (6.2) and Standard-ASPP (8.3). This verifies the superior performance of ASPP-FCN in extracting effective working time features and reducing prediction errors, demonstrating its broad application potential in complex nonlinear regression modeling and improving the accuracy of working time prediction. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 Process card for ball head pin machining process specifications;

[0031] Figure 2The overall structure is based on the ASPP fully connected neural network.

[0032] Figure 3 This is a pseudo-image after spatial encoding;

[0033] Figure 4 This is a diagram of the traditional ASPP module structure.

[0034] Figure 5 Improved ASPP module structure diagram;

[0035] Figure 6 Fully connected layer structure;

[0036] Figure 7 Some special vehicles use self-made parts;

[0037] Figure 8 Work hour dataset content;

[0038] Figure 9 shows the training results of ASPP-FCN; (a) R 2 (b) MSE and (c) MAE;

[0039] Figure 10 For goodness of fit;

[0040] Figure 11 To train MSE;

[0041] Figure 12 To verify MSE;

[0042] Figure 13 To train MAE;

[0043] Figure 14 Test results for 5 neural networks on 1000 test sets;

[0044] Figure 15 Comparison based on 1000 test sets; Detailed Implementation

[0045] 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.

[0046] Example 1

[0047] This embodiment is mainly divided into three parts. First, Label Encoder is used to process the categorical variables, and a spatial remapping mechanism is added to further convert the encoded and normalized features into pseudo-images that can be extracted by the improved ASPP module. Then, batch normalization, global average pooling, and SE attention modules are introduced to improve the ASPP model. Finally, a fully connected network layer is built to map the features and output the prediction results, including:

[0048] Label Encoder encoding

[0049] Label Encoders map each category value to a unique integer, effectively preserving the discriminative power between categories, simplifying feature representation, and making the data structure more suitable for the input layer of neural networks. Especially in embedding layers or spatial remapping mechanisms, Label Encoders provide a solid numerical foundation for subsequent feature mapping and spatial modeling. After encoding, the input features are normalized, but the target label (work time) predicted by the model is not normalized. Normalization not only accelerates the convergence speed of neural networks but also effectively avoids gradient explosion or vanishing problems caused by excessive differences in the dimensions of different features, thereby improving the stability of model training and prediction accuracy.

[0050] Building upon the above, a spatial remapping mechanism is added to further convert the encoded and normalized features into pseudo-images from which the improved ASPP module can extract features. The spatial remapping mechanism consists of two fully connected layers cascaded together. Its function is to map the key features from the original 8-dimensional machining process to a 128-dimensional high-dimensional feature space to enhance feature representation. Then, this 128-dimensional feature vector is rearranged according to channel and spatial structure, transforming it into a 16-channel × 2-row × 4-column feature pseudo-image. This provides structured input for the subsequent parallel execution of multi-scale dilated convolution operations by the ASPP module. The encoded pseudo-image is shown below. Figure 3 As shown. The process of constructing a feature pseudo-image can be formulated as follows: The working process of the spatial coding layer on a given input feature X can be described as follows:

[0051]

[0052]

[0053]

[0054]

[0055] in This represents the input feature matrix after LabelEncoder encoding and Min-Max normalization. B is the batch size of the input data. and These represent the learnable weight matrices of the first and second fully connected layers, respectively. and These represent the learnable bias vectors of the first and second fully connected layers, respectively. and These represent the feature representations of the first and second hidden layers after linear transformation, respectively. This represents the feature matrix after nonlinear activation. To correct the nonlinear activation function of the linear unit, the calculation formula is as follows: , This represents the multi-channel feature pseudo-image generated by the final reshape transformation. That is, it is reconstructed into 16 channels with a spatial dimension of 2 rows and 4 columns.

[0056] Improved ASPP model

[0057] The pseudo-images formed by the spatial coding layer cannot reflect the implicit spatial relationships of the data. To fully extract key information about the time features of self-made parts for special vehicles and improve the parameter response of enhanced key features, and to address the sparsity and scale limitations of time features in the data, this invention improves the standard ASPP model, such as... Figure 5 As shown.

[0058] To address training stability and feature representation capabilities, a batch normalization (BN) layer is introduced into the multi-scale dilated convolution branch to mitigate internal covariate shifts, stabilize the training process, accelerate model convergence, and improve overall robustness during machining. To capture more complete global contextual information and enhance the overall feature representation capability of the work time feature pseudo-images, enabling the model to model the potential correlations between key work time feature factors at different feature interaction scales, global average pooling is introduced. To address the scale adaptation problem of minimal spatial features, larger dilation ratios, such as 6, 12, 18, and 24, are used in the convolution branch to improve the model's sensitivity to scale changes and avoid feature redundancy or loss of key information about work time factors due to insufficient receptive field. To address the issues of insufficient key feature representation and channel redundancy during multi-scale feature fusion, a channel attention recalibration mechanism is introduced in the feature fusion stage of key work time feature factors. An SE attention module is introduced, which adaptively learns the importance weights of each channel through a "compression-activation" operation. Combined with the Sigmoid function, channel attention coefficients are generated to recalibrate features, enhancing the response capability of key feature channels in the work time data while suppressing redundant noise information in the work time features. Finally, 1×1 convolutions are used to compress and fuse the multi-branch output features. While maintaining the original spatial structure of the work time feature pseudo-image, feature relationships at different ranges are fused, and key feature factors of critical work times are strengthened, achieving effective enhancement and unified expression of multi-scale features. For the input feature X, the execution of the ASPP module can be demonstrated as follows:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065] in This represents the pseudo-image of features input to the spatial coding layer. , Indicates the first Local feature maps output by each dilated convolution branch. Indicates the expansion rate dilated convolution kernel weights, , This represents the discrete two-dimensional convolution operator. This indicates a batch normalization operation. This indicates a global average pooling operation. This represents the channel dimensionality reduction used in the global pooling branch. Convolution kernel weights, This indicates that the feature map spatial size is upsampled and restored to its original size using bilinear interpolation. , This represents the extracted global context feature map. This represents the fused feature map obtained by concatenating the convolutional branches at each scale with the global branch along the channel direction. and This represents the weight matrices of the two fully connected layers in the SE channel attention mechanism, one for dimensionality reduction and the other for dimensionality increase. This represents the channel attention weight coefficient vector generated by the SE mechanism. For the Sigmoid function, This represents the feature map after channel attention dynamic recalibration. This represents the channel-wise element-wise multiplication operation between the feature map and the attention weights. This indicates the final product used for dimensionality reduction and feature fusion. The convolution kernel weights, Z represents the enhanced multi-scale feature map output by the improved ASPP module.

[0066] Fully connected output layer

[0067] After obtaining the pseudo-image with enhanced features through the ASPP module, a fully connected network layer is constructed to linearly map the features and output the prediction results. A three-layer fully connected layer (128→64→1) is used to achieve nonlinear reshaping and stepwise compression of multi-scale channel semantics, which is in line with the complex coupling of multiple factors in time prediction. This structure first fully retains multi-source information such as process parameters, processing materials, and machine tool type through a wide representation layer, then gradually compresses to eliminate redundancy and highlight key features, and finally outputs scalar time results.

[0068] The detailed processing of the fully connected network layer, such as Figure 6As shown, the multi-dimensional vector (16 channels, each channel size 2×4) of the enhanced feature pseudo-image is first flattened into a one-dimensional vector (128 dimensions). The first fully connected layer linearly maps the flattened 128-dimensional vector to 128 dimensions and then passes it through the ReLU activation function. This linear transformation and activation function extract non-linear combination features from the input features. The second fully connected layer further reduces the 128-dimensional output of the first layer to 64 dimensions and passes it through the ReLU activation function again to further refine and compress the feature information, retaining the most useful information for prediction while reducing the data dimensionality, which helps reduce the risk of overfitting. Finally, the third fully connected layer maps the 64-dimensional features to the final one-dimensional output dimension, outputting the predicted working time value. The entire fully connected network can be formulated for the input feature X as follows:

[0069]

[0070]

[0071]

[0072] in This represents the one-dimensional feature vector after flattening the pseudo-image output by the ASPP module. , and These represent the learnable weight matrices of the first, second, and output fully connected layers in the network, respectively. , and These represent the learnable bias vectors of the three fully connected layers in the network. and This represents the hidden layer feature representation after processing by the activation function. This represents the target time (i.e., machining time) predicted by the model, with the physical unit being minutes (min). The model output is the absolute time value without normalization.

[0073] Example 2

[0074] To verify the effectiveness of the time prediction model of this invention, based on the actual needs of time prediction for customized parts of special vehicles, this invention uses a Windows environment and NVIDIA RTX 4060 GPU hardware, and selects the Python 2.4.1 framework to build and train a neural network model.

[0075] Experimental platform setup

[0076] The total training cycle of the model was set to 160, and the Adam optimizer with an initial learning rate of 0.001 was used to accelerate convergence and improve stability. At the same time, in order to prevent overfitting and help the model escape local optima, a dynamic learning rate adjustment strategy as shown in Equation (14) was introduced during training. That is, when the validation loss does not improve for 5 consecutive cycles, the learning rate automatically decreases by 0.1. In addition, an early stopping mechanism as shown in Equation (15) was also combined. Once the set patience period is exceeded and the validation loss still does not improve, the training is automatically stopped.

[0077]

[0078] in, It is the first Dynamic learning rate per epoch This represents the learning rate from the previous training cycle. It is the learning rate decay factor, which is usually less than 1.

[0079]

[0080] in This represents the cumulative number of periods in which the loss function on the validation set has not shown a continuous decrease or improvement. This indicates the maximum number of periods the model is allowed to tolerate without improvement before triggering the early stopping mechanism. A logical Boolean signal indicating whether the early stop mechanism is triggered.

[0081] Data sets and metrics

[0082] The sample data used in this invention comes from the Teamcenter product lifecycle management system of a special vehicle manufacturing company. Based on this data, and after rigorous screening, 100 representative self-made parts samples were exported. Some of these self-made parts include... Figure 7 As shown. To ensure the scientific validity and generalization ability of the time prediction model, the sample data was systematically classified according to the geometric topological characteristics and technological process logic of the parts, and divided into six categories of parts, as follows: 25 shaft parts, 2 disc / sleeve parts, 15 fork-shaped parts, 18 box-shaped parts, 11 irregularly shaped parts, and 9 plates and covers. The 3000 machining operations for the 100 parts were divided into a training set of 1300, a validation set of 700, and a test set of 1000, resulting in the time dataset as shown. Figure 8 As shown.

[0083] To evaluate the performance of the predictive model and ensure the scientific validity of the evaluation results and the effectiveness of the decisions, this invention uses four evaluation metrics: root mean squared error, goodness of fit, mean absolute error, and the number of model parameters. The calculation formulas are as follows:

[0084]

[0085]

[0086]

[0087] Where N or This indicates the total number of samples participating in the evaluation. Statement No. The actual machining time data for each part sample is in minutes (min). The model represents the first The predicted machining time for each part sample, in physical units of minutes (min). Indicates all The arithmetic mean of the actual machining time of the sample is expressed in minutes (min). Mean squared error (MSE) measures the average of the sum of squares of the deviations between predicted and actual values. A smaller value indicates better model performance. The physical unit is square minutes (min²). Goodness of fit measures how well the model explains the variance of the data; the closer the value is to the nearest whole number, the better the fit. Mean Absolute Error (MAE) measures the average level of the absolute error between the predicted and actual values. If the MAE of some data points is significantly higher than the overall average, it may indicate that there are anomalies at these points or that the model has a large local prediction error. The physical unit is minutes (min).

[0088] ASPP-FCN Model Efficacy and Ablation Experiments of Key Modules

[0089] The training results of the ASPP-FCN model are shown in Figure 9. The results show that after 160 training cycles, the goodness of fit remains at 0.9985, the training mean squared error decreases to 29.0, the validation mean squared error decreases to 29.6, and the mean absolute error decreases to 4.4. This proves that the ASPP-FCN proposed in this invention can accurately capture important features during machining and can better describe the nonlinear mapping between time features.

[0090] The effectiveness of the modules mentioned in this invention was verified through ablation experiments on key modules. Replacement and removal methods were used for each module, while other modules remained unchanged to ensure the fairness of the experiment. The results are shown in Table 1.

[0091] Table 1. Effectiveness of different ASPP-FCN components on the Teamcenter dataset.

[0092]

[0093] Table 2 shows that when the spatial coding layer is enabled, the model's R² is consistently above 0.9900, while it drops below 0.2522 when disabled, indicating the crucial role of the spatial coding layer in establishing a structured representation of the relationship between process features and time. Comparing the model using only the spatial coding layer with the complete model, introducing ASPP-global reduces the MSE from 120.2 to 83.9, a decrease of 30.2%, demonstrating that its multi-scale dilated convolution effectively captures the local-global correlation of process features, validating that the global pooling branch effectively captures the main influencing factors of time. Using the spatial coding layer + SE reduces the MSE by 4.2% compared to the spatial coding layer alone, proving that the SE channel attention mechanism can optimize feature weight allocation. Since the SE module directly operates on the original pseudo-image, its channel number is limited, its information dimensionality is low, and the feature correlation between channels is weak, making it difficult to form an effective basis for attention allocation. In this case, the SE module cannot fully learn meaningful channel weight distributions and may even introduce invalid or interfering information, leading to a performance decrease rather than an increase. Therefore, the effectiveness of the SE mechanism is based on the high-dimensional multi-channel features output by the spatial encoding and ASPP modules, and there is a dependency between the two. In the final complete model, the spatial encoding layer + ASPP-global + SE achieves the best performance with R²=0.9957 and MSE=83.91. The effectiveness of each component shows a significant hierarchical relationship: spatial encoding layer > ASPP-global > SE attention, and the three components produce a positive synergistic effect when working together.

[0094] Example 3

[0095] This study compares and analyzes the results of the optimized model with those of the widely used regression neural network models, including fully connected feedforward neural networks (FNN), convolutional neural networks (CNN), residual networks (ResNet), long short-term memory networks (LSTM), and the standard Standard-ASPP, from both the training and test set results of the neural network models.

[0096] Model training results

[0097] This invention constructs training and validation sets based on actual production data collected by the Teamcenter system. The system compares the performance differences of six neural networks across three dimensions: model fitting ability, mean squared error (MSE), and parameter size. Training fit goodness of fit is shown below. Figure 10 As shown, training and validation MSE pairs are as follows: Figure 11 and Figure 12 As shown, the mean absolute error (MAE) is... Figure 13 For details on the model parameters, please refer to Table 2.

[0098] Depend on Figure 10 As can be seen, compared to LSTM and standard CNN models, it exhibits better fitting ability, with R² values ​​of 0.9897 and 0.9902, respectively. ResNet and Feedforward Neural Network (FNN) effectively improve the model's expressive power based on the original structure, with R² values ​​increasing to 0.9938 and 0.9946, respectively. The standard ASPP module neural network has a goodness of fit of 0.9926, but the ASPP-FCN model proposed in this invention, by introducing a Spatial Pyramid Pooling (ASPP) structure and a multi-scale perception mechanism, effectively enhances the modeling ability of nonlinear relationships between complex time features, significantly improving the goodness of fit to R²=0.9985, which is significantly better than other models.

[0099] Depend on Figure 11 It can be seen that after 160 training epochs, the training MSE of the ASPP-FCN model decreased to 29.9, far lower than LSTM (158.2), CNN (174.8), ResNet (112.8), FNN (105.9), and Standard-ASPP (101.4). Compared with LSTM, ASPP-FCN achieved a training error reduction of approximately 81.09%, and even under the better Standard-ASPP baseline, it still achieved an error compression of approximately 70.53%. This significant advantage mainly comes from the multi-scale feature fusion strategy and Squeeze-and-Excitation (SE) attention mechanism introduced by ASPP-FCN, which effectively enhances the model's ability to perceive and model key time features, improving overall prediction accuracy and generalization ability. Furthermore, by Figure 12 It can be seen that under the same number of training rounds, ASPP-FCN's MSE on the validation set drops to 29.6, which is significantly lower than LSTM (178.1), CNN (192.8), Standard-ASPP (139.3), ResNet (122.3) and FNN (105.9), showing that it has stronger generalization ability and stability in dealing with complex conditions, further verifying the generalization performance of ASPP-FCN.

[0100] Finally, by Figure 13 As can be seen, the mean absolute error (MAE) metric of ASPP-FCN also demonstrates a significant advantage. After training, its MAE decreased to 4.4, outperforming LSTM (5.5), CNN (13.5), ResNet (7.4), FNN (6.2), and Standard-ASPP (8.3). This result further validates the superior performance of ASPP-FCN in extracting effective work time features and reducing prediction errors, demonstrating its broad application potential in complex nonlinear regression modeling and improving the accuracy of work time prediction.

[0101] Table 2 Number of parameters in neural network models

[0102]

[0103] As shown in Table 2, the ASPP-FCN model proposed in this invention significantly improves feature representation ability by effectively integrating spatial pyramid pooling and spatial encoding mechanisms, using only 43,685 parameters. Ultimately, it achieves the optimal fitting effect of R²=0.9985, fully demonstrating the superior balance and engineering practical value of ASPP-FCN between parameter efficiency and modeling ability.

[0104] Model test results

[0105] The six trained neural network models were evaluated on the test set, and their performance metrics are shown in Table 3. The results show that the ASPP-FCN model performs best overall, with a determination coefficient (R²) of 0.9955, a mean squared error (MSE) of 79.1, and a mean absolute error (MAE) of 4.2. This indicates that the model has significant advantages in both fitting ability and error control. The combination of the ASPP structure and the fully convolutional network (FCN) can effectively extract multi-scale features and enhance the model's ability to express complex processing features, thereby improving prediction accuracy. In comparison, the LSTM model has an R² of 0.9948, MSE of 91.3, and MAE of 9.6. Although it has strong temporal modeling capabilities, its ability to characterize non-temporal features is relatively limited in this task. The FNN model has an R² of 0.9933, MSE of 117.2, and MAE of 10.8, showing some non-linear fitting ability, but lacking in-depth mining of local features. The Standard-ASPP model has an R² of 0.9923, MSE of 134.8, and MAE of 11.6, indicating that relying solely on dilated convolutional structures is still insufficient in feature fusion. The CNN model (R² of 0.9884, MSE of 203.1, and MAE of 8.0) and the ResNet model (R² of 0.9881, MSE of 208.3, and MAE of 14.4) have relatively large overall errors, indicating that their generalization ability in this type of work time prediction problem is relatively weak. Based on the comprehensive indicators, ASPP-FCN has a stronger advantage in feature extraction depth and multi-scale information fusion, thus outperforming other comparative models in overall performance.

[0106] Table 3 Performance metrics of each neural network after testing the test set

[0107]

[0108] The prediction results based on 6 neural network models for 1000 test samples are as follows: Figure 14 As shown, the error comparison analysis is as follows. Figure 15 At an error threshold of 10.0%, ASPP-FCN achieved 612 valid prediction samples, a 6.3% improvement over the second-best model, FNN. When the error threshold was relaxed to 20.0%, ASPP-FCN continued to lead with 747 samples, a 1.2% improvement over the second-best model. At an error threshold of 30.0%, ASPP-FCN still ranked first with 825 valid prediction samples, a 0.9% improvement over the second-best model, ResNet (818). Experimental results show that in the task of predicting the machining time of self-made parts for special vehicles, ASPP-FCN has significant advantages over mainstream neural network models such as LSTM, CNN, ResNet, and FNN in both prediction accuracy and stability. Within the ±20% time error range specified by the enterprise, ASPP-FCN achieved the highest prediction accuracy of 74.7%, demonstrating stronger engineering applicability and robustness.

[0109] In summary, this invention proposes a smart manufacturing machining time prediction method based on ASPP-FCN, which can realize the prediction of machining time for small-batch customized special vehicle self-made parts.

[0110] Among these, the structured "pseudo-image" reconstruction using discrete process features is a necessary prerequisite for achieving high-precision prediction of complex work hours. By employing a spatial coding mechanism to reconstruct the underlying work hour features into a multi-dimensional pseudo-image, a structured spatial relationship between process features and work hour consumption can be successfully established. Ablation experiments confirm that without this dimension of representation transformation, the explained variance (R²) of the prediction model will plummet (to around 0.25), demonstrating the irreplaceable core role of this remapping mechanism in industrial multi-feature regression prediction tasks. The synergistic use of multi-scale spatial feature perception and channel attention mechanisms effectively overcomes the feature coupling bottleneck of high-dimensional sparse manufacturing data. The model achieves a high-precision fit with an MAE of only 4.4 and an R² of 0.9988, achieving an accuracy of 74.7% within the ±20% tolerance error band defined by the enterprise. The ASPP-FCN model proposed in this invention demonstrates good predictive ability and robustness on the overall dataset, effectively meeting the needs of modern manufacturing enterprises for efficient prediction of small-batch, customized complex parts.

[0111] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for predicting machining time of discrete manufacturing parts based on ASPP-FCN, characterized in that, Includes the following steps: S1: Using the group decision-making method with effective expert information, and combining the indicator weights and consistency tests under the interval number condition, the eight working time features with the highest weights are selected as inputs. S2: Based on the characteristics of the 8 key features of the work time, LabelEncoder is used to process the categorical variables, and a spatial remapping mechanism is added to further convert the encoded and normalized features into pseudo-images that can be extracted by the improved ASPP module. S3: Based on the pseudo-image extracted from the time features, an improved ASPP module is used to extract key information of the time features and enhance the parameter response of the key features, and to solve the problems of sparsity and scale limitation of time features in the data. S4: Using the pseudo-image of enhanced features obtained from the ASPP module, a fully connected layer is built to perform linear mapping of the features and output prediction. S5: Build an experimental platform to verify the ASPP-FCN time prediction model established above.

2. The method for predicting machining time of discrete manufacturing parts based on ASPP-FCN according to claim 1, characterized in that: Step S1 specifically includes: using a group decision-making method with effective expert information, and combining the index weights and consistency tests under the interval number condition, selecting the 8 working time features with the highest weights as inputs; The time characteristics of machining processes were summarized and organized, resulting in a total of 16 characteristics. The eight time characteristics with the highest weights were selected through a group decision-making method based on effective expert information.

3. The method for predicting machining time of discrete manufacturing parts based on ASPP-FCN according to claim 1, characterized in that: Step S2 specifically includes: LabelEncoder encoding: mapping each category value to a unique integer, effectively preserving the distinguishability between categories and simplifying feature representation. Feature normalization: Normalize the encoded features to speed up the convergence of the neural network and avoid gradient explosion or gradient vanishing problems caused by excessive differences in the dimensions of different features. Spatial remapping mechanism: It consists of two fully connected layers connected in series, which map the original 8-dimensional machining key feature factors to a 128-dimensional high-dimensional feature space to enhance the feature representation capability. Then, the 128-dimensional feature vector is rearranged according to the channel and spatial structure, and converted into a 16-channel × 2-row × 4-column feature pseudo-image form, which provides structured input for the subsequent ASPP module to perform multi-scale dilated convolution operations in parallel.

4. The method for predicting machining time of discrete manufacturing parts based on ASPP-FCN according to claim 1, characterized in that: Step S3 specifically includes: improving the ASPP-FCN module to fully extract key information on the time features of self-made parts for characteristic vehicles and enhancing the parameter response of key features; Introducing a batch normalization (BN) layer into the multi-scale dilated convolution branch enhances training stability and feature representation capabilities, thereby mitigating internal covariate shifts, stabilizing the training process, accelerating model convergence, and improving overall machining robustness. By introducing global average pooling, the overall feature representation capability of the time feature pseudo-image is enhanced, and more complete global context information is captured.

5. The method for predicting machining time of discrete manufacturing parts based on ASPP-FCN according to claim 1, characterized in that: Step S4 specifically includes: a fully connected neural network: flattening the multidimensional tensor of the enhanced feature pseudo-image into a one-dimensional vector; linearly mapping the flattened vector to 128 dimensions through the first fully connected layer, and then passing it through the ReLU activation function, extracting nonlinear combination features from the input features through linear transformation and activation function, preparing for subsequent dimensionality reduction and prediction; further reducing the 128-dimensional output of the first layer to 64 dimensions through the second fully connected layer, and passing it through the ReLU activation function again to further refine and compress feature information, retaining the most useful information for prediction, while reducing data dimensionality; finally, mapping the 64-dimensional features to a one-dimensional output dimension (work time) through the third fully connected layer, and finally outputting the predicted work time value.

6. The method for predicting machining time of discrete manufacturing parts based on ASPP-FCN according to claim 1, wherein steps S4 and S5 specifically include: The effectiveness of the ASPP-FCN model was validated using mean squared error, goodness of fit, mean absolute error, and the number of model parameters as evaluation metrics. The model was compared with six mainstream models: fully connected feedforward neural network (FNN), convolutional neural network (CNN), residual network (ResNet), long short-term memory network (LSTM), and standard Standard-ASPP. Ablation experiments were conducted to verify the impact of spatial coding layer, ASPP-global, global pooling branch, and SE module on the performance of the ASPP-FCN model.

7. A method for predicting machining time of discrete manufacturing parts based on ASPP-FCN according to any one of claims 1 to 6, characterized in that: This method is applied to predict machining time in the field of small-batch self-made parts manufacturing for special vehicles.

Images