Design scheme joint decision multi-tailed neural network model and training method

Abstract

The application relates to the technical field of design scheme intelligent decision-making, and discloses a design scheme joint decision-making multi-tail neural network model and a training method, which comprises the following steps: generating a target data set based on historical design cases; performing phased joint training on a multi-tail neural network model through the target data set to obtain a scheme decision-making model; and performing joint decision-making on multiple design parameters according to the scheme decision-making model to obtain a target design scheme. The application can effectively solve the technical problems of "small scale, multiple types and uneven distribution" of training samples in scheme selection business, and effectively improve the efficiency and accuracy of scheme decision-making.

Description

Technical Field

[0001] This application relates to the field of intelligent decision-making technology for design schemes, and in particular to a multi-tailed neural network model and training method for joint decision-making of design schemes. Background Technology

[0002] With the rapid advancement of intelligent technology, railway bridge design has been developing towards integration, digitalization, visualization, and intelligence. The intelligent design of the bridge substructure, especially the pile foundation, is a crucial component of intelligent bridge design. Making rapid, accurate, and efficient decisions regarding pile foundation schemes will facilitate the automatic search and optimization of bridge designs. Pile foundations are a common type of foundation in railway bridge engineering, especially in long-span structures that need to withstand large loads. A pile foundation consists of a group of piles that collectively bear the load of the superstructure and transfer these loads to deep soil or rock layers. Commonly used pile diameters include 1.0m, 1.25m, 1.5m, 1.8m, 2.0m, 2.2m, 2.5m, and 3.0m. Commonly used pile group foundations have 8, 9, 10, 11, 12, 15, 16, and 18 piles. As you can see, there are more than 10 different pile diameters and more than 10 different pile quantities. The combinations are numerous, leading to a "combination explosion" problem when making pile foundation scheme decisions, resulting in low efficiency and poor accuracy in pile foundation scheme decisions.

[0003] Therefore, how to effectively improve the efficiency and accuracy of decision-making is an urgent problem that needs to be solved.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main purpose of this application is to provide a multi-tailed neural network model for joint decision-making of design schemes and a training method, aiming to solve the technical problem of how to effectively improve the efficiency and accuracy of scheme decision-making.

[0006] To achieve the above objectives, this application proposes a design scheme for a joint decision-making multi-tailed neural network model and a training method, the method comprising:

[0007] Generate the target dataset based on historical design cases;

[0008] The multi-tailed neural network model is jointly trained in stages using the target dataset to obtain the scheme decision model;

[0009] The target design scheme is obtained by jointly deciding on multiple design parameters based on the scheme decision model.

[0010] In one embodiment, generating the target dataset based on historical design cases includes:

[0011] Feature parameters are extracted from the historical design cases to obtain the initial dataset;

[0012] The initial dataset is augmented and cleaned based on the association parameters to obtain the target dataset.

[0013] In one embodiment, the step of performing phased joint training of the multi-tailed neural network model using the target dataset to obtain the solution decision model includes:

[0014] The target dataset is divided into a training set and a test set according to a preset ratio. The target dataset includes multiple samples, each of which includes different feature parameters and corresponding actual design schemes.

[0015] The multi-tailed neural network model is jointly trained in stages using the training set to obtain the trained multi-tailed neural network model, wherein the multi-tailed neural network model includes a data embedding layer, a common feature extraction layer, and a multi-tailed decision layer.

[0016] The trained multi-tailed neural network model is tested using the test set to obtain the loss value of the test set.

[0017] The parameters of the trained multi-tailed neural network model are updated based on the loss values ​​of the test set to obtain the solution decision model.

[0018] In one embodiment, the step of performing phased joint training on the multi-tailed neural network model using the training set to obtain the trained multi-tailed neural network model includes:

[0019] The feature parameters of each sample in the training set are input into the data embedding layer for embedding processing to obtain the target embedding vector;

[0020] The target embedding vector is input into the common feature extraction layer for feature extraction to obtain the feature vector;

[0021] The feature vector is input into the multi-tailed decision layer for joint decision-making to obtain a predictive design scheme.

[0022] The predicted design scheme is compared with the corresponding actual design scheme to obtain the prediction error;

[0023] The multi-tailed neural network model is trained by backpropagation based on the prediction error, and the model weights are updated to obtain the trained multi-tailed neural network model.

[0024] In one embodiment, the step of inputting the feature parameters into the data embedding layer for embedding processing to obtain the target embedding vector includes:

[0025] Discrete and continuous features are determined based on the aforementioned feature parameters;

[0026] The discrete features are mapped to a continuous vector space to obtain the first embedding matrix;

[0027] The continuous features are automatically discretized, discretized, and aggregated to obtain the second embedding matrix.

[0028] Generate a target embedding vector based on the first embedding matrix and the second embedding matrix.

[0029] In one embodiment, the step of inputting the target embedding vector into the common feature extraction layer for feature extraction to obtain a feature vector includes:

[0030] Obtain the weight matrix and bias vector of a fully connected neural network;

[0031] Based on the weight matrix and bias vector, feature extraction is performed on the target embedding vector to obtain the feature vector.

[0032] In one embodiment, the step of inputting the feature vector into the multi-tailed decision layer for joint decision-making to obtain a predictive design scheme includes:

[0033] Based on the feature vector, a decision is made on the first design parameters to obtain a first decision vector;

[0034] The first decision vector is concatenated with the feature vector to obtain the concatenated vector;

[0035] Based on the splicing vector, a decision is made on the second design parameters to obtain the second decision vector;

[0036] The decision-making process for the third design parameter is performed based on the second decision vector and the common feature vector until the decision vectors for each design parameter are generated.

[0037] The decision vectors generated for each design parameter are summarized to obtain the predicted design scheme.

[0038] Furthermore, to achieve the above objectives, this application also proposes a design scheme for a joint decision-making multi-tailed neural network model and training device, the design scheme comprising:

[0039] The generation module is used to generate target datasets based on historical design cases;

[0040] The training module is used to perform phased joint training of the multi-tailed neural network model using the target dataset to obtain the scheme decision model;

[0041] The decision module is used to make joint decisions on multiple design parameters based on the scheme decision model to obtain the target design scheme.

[0042] Furthermore, to achieve the above objectives, this application also proposes a design scheme for a joint decision-making multi-tailed neural network model and a training device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the design scheme for the joint decision-making multi-tailed neural network model and training method as described above.

[0043] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the joint decision-making multi-tailed neural network model and training method described above.

[0044] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the joint decision-making multi-tailed neural network model and training method described above.

[0045] This application provides a multi-tailed neural network model for joint decision-making of design schemes and a training method. This application first generates a target dataset based on historical design cases, which can improve data balance; then, it performs phased joint training of the multi-tailed neural network model using the target dataset to obtain a scheme decision model, which can effectively improve the model's accuracy and generalization ability; finally, it performs joint decision-making on multiple design parameters based on the scheme decision model to obtain the target design scheme, which can effectively improve the efficiency and accuracy of scheme decision-making.

[0046] In summary, this application uses a target dataset generated from historical design cases to perform phased joint training on a multi-tailed neural network model, and then uses the trained pile foundation scheme decision model to make joint decisions to obtain pile foundation design schemes. This overcomes the technical defects of low efficiency and poor accuracy in design scheme decision-making, and can effectively improve the efficiency and accuracy of scheme decision-making. Attached Figure Description

[0047] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A flowchart illustrating the first embodiment of the joint decision-making multi-tailed neural network model and training method for the design scheme of this application;

[0050] Figure 2 A schematic diagram of a pile foundation for the joint decision-making multi-tailed neural network model and training method of the first embodiment of this application;

[0051] Figure 3 This is a schematic diagram of the data distribution for the main beam type in the joint decision-making multi-tailed neural network model and training method of the first embodiment of this application.

[0052] Figure 4 A distribution diagram of pile foundation types for the joint decision-making multi-tailed neural network model and training method of the first embodiment of this application;

[0053] Figure 5 The network architecture diagram of the pile foundation scheme decision model of the joint decision-making multi-tailed neural network model and training method of the first embodiment of this application is shown.

[0054] Figure 6 A flowchart illustrating the second embodiment of the joint decision-making multi-tailed neural network model and training method for the design scheme of this application;

[0055] Figure 7 This is a schematic diagram of the AutoDis principle of the joint decision-making multi-tailed neural network model and training method designed in the second embodiment of this application.

[0056] Figure 8 This is a schematic diagram of the loss curve of the joint decision-making multi-tailed neural network model and training method of the second embodiment of this application.

[0057] Figure 9 A confusion matrix diagram illustrating the joint decision-making multi-tailed neural network model and training method of the second embodiment of this application;

[0058] Figure 10 This is a schematic diagram of the accuracy and recall curves of the joint decision-making multi-tailed neural network model and training method of the second embodiment of this application as a function of training batch deformation.

[0059] Figure 11 A schematic diagram of the module structure of the joint decision-making multi-tailed neural network model and training device designed for embodiments of this application;

[0060] Figure 12 This is a schematic diagram of the device structure of the hardware operating environment involved in the design scheme of the joint decision-making multi-tailed neural network model and training method of the embodiment of this application.

[0061] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0062] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0063] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0064] The main solution of this application embodiment is: generating a target dataset based on historical design cases; performing phased joint training on a multi-tailed neural network model using the target dataset to obtain a scheme decision model; and making joint decisions on multiple design parameters based on the scheme decision model to obtain a target design scheme.

[0065] A pile group foundation consists of a group of piles that collectively bear the load of the superstructure and transfer these loads to deeper soil or rock layers. Commonly used pile diameters include 1.0m, 1.25m, 1.5m, 1.8m, 2.0m, 2.2m, 2.5m, and 3.0m, and the common number of piles in a pile group foundation includes 8, 9, 10, 11, 12, 15, 16, and 18. As can be seen, there are more than 10 different pile diameters and more than 10 different pile numbers, resulting in numerous combinations and a "combinatorial explosion" problem in pile foundation design decisions, leading to low efficiency and poor accuracy in these decisions. Therefore, how to effectively improve the efficiency and accuracy of foundation design decisions is a pressing issue that needs to be addressed.

[0066] This application uses a target dataset generated based on historical design cases to perform phased joint training on a multi-tailed neural network model, and then uses the trained pile foundation scheme decision model to make joint decisions to obtain pile foundation design schemes. This overcomes the technical defects of low efficiency and poor accuracy in design scheme decision-making, and can effectively improve the efficiency and accuracy of scheme decision-making.

[0067] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of implementing the above functions, a design scheme joint decision-making multi-tailed neural network model, and training equipment, etc. The following description uses a design scheme joint decision-making multi-tailed neural network model and training equipment as an example to illustrate this embodiment and the subsequent embodiments.

[0068] Based on this, embodiments of this application provide a design scheme for a joint decision-making multi-tailed neural network model and a training method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the joint decision-making multi-tailed neural network model and training method designed in this application.

[0069] In this embodiment, the design scheme for the joint decision-making multi-tailed neural network model and training method includes steps S10 to S30:

[0070] Step S10: Generate the target dataset based on historical design cases.

[0071] It should be noted that in this embodiment, various decision-making schemes can be designed, such as pile foundation scheme design and bridge pier scheme design. This embodiment does not impose specific limitations on these; it uses pile foundation scheme design as an example for illustration. Pile foundation scheme design is also known as intelligent pile foundation selection. Intelligent pile foundation selection is only a preferred application example of this invention, and it can be applied to many other fields. By reducing the number of labels in the combined classification problem, the sample size is reduced.

[0072] Understandably, for pile foundation design, historical design cases refer to existing railway bridge pile foundation design cases. A pile group foundation consists of multiple piles that collectively bear the load of the superstructure and transfer these loads to deeper soil or rock layers. Commonly used railway pile diameters include 1.0m, 1.25m, 1.5m, 1.8m, 2.0m, 2.2m, 2.5m, and 3.0m, and commonly used pile numbers include 8, 9, 10, 11, 12, 15, 16, and 18. Common pile group arrangements include row, staggered, and custom arrangements. Pile group foundations are formed by combining pile diameters, arrangement types, and the number of piles to create different types of pile foundation configurations, such as... Figure 2 As shown, Figure 2 This is a schematic diagram of a pile foundation.

[0073] It is understandable that the target dataset consists of multiple samples, and the combination of the feature parameters of each bridge type with the corresponding actual number of piles and the actual pile diameter constitutes a sample.

[0074] It is worth noting that in this embodiment, the design of the pile foundation mainly focuses on the combination of pile diameter and number. The pile diameter and number are influenced by various factors, including design load factors, geological conditions, economic factors, construction technology factors, environmental impact factors, durability and maintenance factors, and regional factors. Design load factors refer to the design loads such as bridge weight, expected traffic load, dynamic load, wind load, and temperature changes, all of which affect the pile foundation design. The greater the load, the larger the number and diameter of piles are typically required. Geological conditions refer to the soil type, layer thickness, bearing capacity, friction angle, cohesion, compressibility, and groundwater level, all of which affect the pile foundation design. For example, in soft soil areas, more piles may be needed to distribute the load, while in hard rock layers, fewer piles may be required. Economic factors refer to the need for designers to find the most cost-effective balance between the number and diameter of piles while meeting safety and functional requirements. Construction technology factors refer to the availability of construction technologies and equipment, which also affect the pile foundation design. For example, if construction equipment limits the pile diameter, the design may need to increase the number of piles to meet the load-bearing requirements. Environmental impact factors refer to the environmental impact of construction activities, which must be considered during the design phase. For example, if there are sensitive ecological environments near the construction area, special designs may be needed to minimize the impact. Durability and maintenance factors refer to the durability of the pile foundation and the long-term maintenance costs, which must also be considered during the design phase. In some cases, using larger diameter piles may reduce long-term maintenance costs. Regional factors refer to the different bridge design codes and standards in different regions, which impose specific requirements on pile foundation design, including the number and diameter of piles.

[0075] In one feasible implementation, step 10 may include: extracting feature parameters from the historical design cases to obtain an initial dataset; and performing data augmentation and data cleaning on the initial dataset based on the association parameters to obtain a target dataset.

[0076] It should be noted that for pile foundation design, in order to fully utilize existing railway bridge pile foundation design cases and achieve new design support for decision-making, effective characterization of design cases is necessary. On the one hand, this provides a basis for the characterization and quantification of bridge cases; on the other hand, effective characterization of design cases can provide data support for the retrieval of similar cases and the inference decision-making of neural network models. In this embodiment, nine feature attributes are selected to characterize the cases. These attributes can be divided into numerical and enumerable types. The numerical type includes the side span at the small mileage (s1), the side span at the large mileage (s2), the pier height (s3), and the average ground slope (s4); the enumerable type includes the seismic intensity (s5), the side beam type at the small mileage (s6), the side beam type at the large mileage (s7), the pier type (s8), and the geological label (s9). The pile group scheme characterization consists of two parts: pile diameter (target1) and the number of piles (target2). It is understood that in this embodiment, data is extracted from existing cases, and the collected dataset contains 1014 records, or 1014 samples. Each sample record describes relevant information about a pile group foundation, namely, characteristic parameters and sample labels. The sample label is the pile type (including quantity and diameter) corresponding to the characteristic parameter, as shown in Table 1. Table 1 is a sample example data table, which includes multiple sample cases. Each case consists of characteristic parameters and sample labels. The characteristic parameters include small mileage side beam type, large mileage side beam type, small mileage side span, large mileage side span, pier type, and pier height. The sample labels include the following information: seismic fortification level, ground slope, and geological label. The label also includes the following pile foundation type: small mileage side beam type (e.g., simply supported beam, continuous beam); large mileage side beam type (e.g., simply supported beam); small mileage side span (e.g., 80m, 32m); large mileage side span (e.g., 32m, 24m); pier type (e.g., hollow pier, solid pier); pier height (e.g., 28m, 21m); seismic fortification level (e.g., C); ground slope (e.g., 0%); geological label (e.g., A); and pile foundation type (e.g., 11 piles, 1.50m diameter, 12 piles, 1.00m diameter).

[0077] Table 1

[0078]

[0079] In the specific implementation, the initial dataset collected from existing cases contained 1014 samples. Among these samples, continuous beam bridges were the most numerous, totaling 455, followed by simply supported beam bridges (328), continuous beam arch bridges (69), T-shaped rigid frame bridges (64), continuous rigid frame bridges (46), low-tower cable-stayed bridges (35), and simply supported tied arch bridges (17). Figure 3 As shown, Figure 3This is a schematic diagram of the data distribution for main beam types. The data distribution is similar to the current design situation. Currently, the main types of bridges using pile foundations are continuous beams and simply supported beams, with fewer bridges having other special structures. The initial dataset contains 37 different pile foundation types, but their distribution is not uniform. For example, bridges with 8-1.0m pile foundations are the most numerous, with 1014 bridges, while other pile foundation types are relatively fewer, such as 129 bridges with 10-1.0m pile foundations and 114 bridges with 10-1.25m pile foundations, etc. Figure 4 As shown, Figure 4 This is a distribution map of pile foundation types.

[0080] Understandably, the distribution of samples in the initial dataset shows that most bridges in the initial dataset use the 8-1.0m (i.e., 8 piles with a diameter of 1.0m) pile foundation type, while other pile foundation types are fewer in number. Therefore, the imbalanced data distribution in the initial dataset will lead to the following problems in subsequent model training: 1. The fact that some pile foundation types (such as 8-1.0m) are far more numerous than others may cause the model to favor recognizing the most common pile foundation types during training. For the few pile foundation types in the dataset, the model may not be able to effectively learn their features, thus affecting the model's generalization ability; 2. Due to the large number of samples for some pile foundation types in the dataset, the model may be biased towards recognizing the most common pile foundation types during training. Overemphasizing certain feature types while neglecting others may lead to the model's inability to accurately identify pile foundation types with limited sample sizes in practical applications. Furthermore, the large sample size of some pile foundation types may cause the model to overfit to these types, resulting in poor performance on the test set. The uneven distribution of data makes it difficult to determine which features contribute most to the identification of different pile foundation types, potentially leading to difficulties in feature selection, as some features may lack sufficient samples to validate their effectiveness on a few pile foundation types. Finally, the distribution of pile foundation types in the dataset may not be sufficient to cover all possible practical application scenarios, meaning that new pile foundation types encountered in actual engineering may not be accurately identified by the model. Therefore, to address the problem of imbalanced data, the initial dataset needs to be augmented and cleaned based on the pile top area and pile group stiffness to obtain the final target dataset for training.

[0081] It is worth noting that for pile foundation design, the associated parameters refer to parameters related to the pile diameter and the number of piles. The pile top area refers to the total area occupied by all pile tops on the pile cap in the pile foundation design. It is usually used to calculate the size of the pile cap and evaluate its bearing capacity. In the case of multi-pile pile caps (i.e., pile group pile caps), the sum of the pile top areas affects the size and shape of the pile cap, as well as the load distribution. The role of the pile cap is to evenly distribute the load of the superstructure onto each pile. Pile group stiffness refers to the overall rigidity exhibited by multiple piles working together in pile foundation design. When a pile group is loaded, the interaction between each pile affects the response of the entire pile group system, including settlement and bearing capacity. The stiffness change caused by the pile group effect is due to the interaction between piles, between piles and soil, and between soils. These interactions may make the stiffness of the pile group different from that of a single pile. Generally, the stiffness of a pile group is smaller than that of a single pile because of the stress dispersion between piles and the deformation coordination of the soil.

[0082] In the specific implementation, the steps for data augmentation and data cleaning of the initial dataset based on the pile top area and pile group stiffness are as follows: For the five special beam types with a relatively small number of cases—continuous beam arch, T-shaped rigid frame, continuous rigid frame, low tower cable-stayed bridge, and simply supported tied arch—design experts are asked to redesign the data to increase the number of design cases; other pile foundation design schemes are added, based on the principle that the change in pile top area and overall stiffness of the pile group foundation should not exceed 10%; the data is then cleaned, and for similar design cases, only one is retained to obtain the final target dataset.

[0083] Step S20: Perform phased joint training on the multi-tailed neural network model using the target dataset to obtain the scheme decision model.

[0084] It should be noted that the multi-tailed neural network model is a deep learning model specifically designed to handle ensemble classification problems. Its unique structure enables it to learn and optimize multiple classification tasks simultaneously and share knowledge and features among these tasks.

[0085] Understandably, for pile foundation design, the scheme decision model is the final model used to predict and evaluate the effectiveness of the pile foundation design scheme. Through training with a multi-tailed neural network model, the pile foundation scheme decision model can automatically extract key features from the data and learn the relationship between different feature parameters and the pile diameter and the number of piles.

[0086] It is worth noting that phased joint training refers to dividing the entire training process into multiple phases. For pile foundation design, the entire training process is divided into two phases to learn and optimize the pile diameter and the number of piles in a targeted manner, and finally obtain the scheme decision model.

[0087] Step S30: Based on the scheme decision model, make joint decisions on multiple design parameters to obtain the target design scheme.

[0088] It should be noted that, according to the scheme decision model, joint decisions are made sequentially on the first design parameter, the second design parameter, and the Nth design parameter, that is, decisions are made on all design parameters in sequence. For the pile foundation scheme design, there are two design parameters in total: the first design parameter is the number of piles, and the second design parameter is the pile diameter. The target design scheme is the pile foundation design scheme. The pile foundation scheme decision model decomposes the combined decision problem of pile diameter and pile number into two problems: pile diameter decision and pile number decision. By classifying pile diameter and pile number separately, the number of decision categories in a single model is reduced. At the same time, to ensure that there is still some connection between the two decisions, the decision models of the two problems share some weights. Furthermore, the decision result of pile diameter will be used as input to the pile number decision model, directly affecting the decision of pile number, thus obtaining the pile foundation design scheme.

[0089] Understandably, the pile foundation scheme decision model is divided into three parts: a data embedding layer, a common feature extraction layer, and a multi-tailed decision layer. Each layer is composed of a fully connected MLP neural network.

[0090] like Figure 5 As shown, Figure 5The network architecture diagram of the pile foundation scheme decision model is shown. The pile foundation scheme decision model includes a data embedding layer, a common feature extraction layer (HIDDEN LAYERS), a multi-tailed decision layer (Design-Making LAYERS), and an output layer (OUTPUT LAYER-A, OUTPUT LAYER-B). In the data embedding layer, the model's input data is divided into continuous data (cont data) and discrete data (dis data). The continuous data (cont data) undergoes automatic discretization, meta embedding, and aggregation using the AutoDis module to obtain an embedding matrix. Numerical data and enumerated data are generated from the discrete data (dis data), and embedding is performed on these data to obtain another embedding matrix. Embedding vectors are then generated from these two embedding matrices and input into the HIDDEN LAYERS layer. The fully connected neural network in the HIDDEN LAYERS layer extracts features, resulting in feature vectors. These feature vectors are then input into the Design-Making LAYERS layer to determine the number of foundation piles. Finally, the output layer (OUTPUT) performs the final step. The LAYER-A layer outputs a decision vector for the number of foundation piles. This decision vector is then concatenated with the feature vector to determine the foundation pile diameter. Finally, the LAYER-B layer outputs the decision vector for the foundation pile diameter.

[0091] In practical implementation, the pile foundation scheme decision model is combined with the automatic hole layout algorithm and scheme evaluation algorithm for railway bridges to obtain the intelligent decision algorithm service for bridge span schemes. By deploying microservices and connecting to the beam professional design system, the overall bridge design scheme can be generated quickly and multiple schemes can be evaluated and compared.

[0092] This embodiment provides a multi-tailed neural network model for joint decision-making of design schemes and a training method. This embodiment first generates a target dataset based on historical design cases, which can improve data balance; then, it performs phased joint training of the multi-tailed neural network model using the target dataset to obtain a scheme decision model, which can effectively improve the model's accuracy and generalization ability; finally, it performs joint decision-making on multiple design parameters based on the scheme decision model to obtain the target design scheme, which can effectively improve the efficiency and accuracy of scheme decision-making.

[0093] In summary, this embodiment uses a target dataset generated based on historical design cases to perform phased joint training on a multi-tailed neural network model, and then uses the trained pile foundation scheme decision model to make joint decisions to obtain pile foundation design schemes. This overcomes the technical defects of low efficiency and poor accuracy in design scheme decision-making, and can effectively improve the efficiency and accuracy of scheme decision-making.

[0094] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 6 Step S20 further includes steps S201-S204:

[0095] Step S201: Divide the target dataset into a training set and a test set according to a preset ratio. The target dataset includes multiple samples, each of which includes different feature parameters and corresponding actual design schemes.

[0096] It should be noted that the preset ratio can be 8:2 or other ratios. This embodiment does not impose specific restrictions on this. This embodiment will use 8:2 as an example for explanation.

[0097] In the specific implementation, 1800 samples are selected from the dataset as the training set and 229 samples as the test set. The Adam optimizer is used with a learning rate of 0.001 to train the multi-tailed neural network model.

[0098] Understandably, for pile foundation design, the target dataset includes multiple samples. Each sample consists of characteristic parameters of different bridge types and the corresponding actual pile foundation type. The actual pile foundation type refers to the combination of the actual number of piles and the actual pile diameter.

[0099] Step S202: The multi-tailed neural network model is jointly trained in stages using the training set to obtain the trained multi-tailed neural network model, wherein the multi-tailed neural network model includes a data embedding layer, a common feature extraction layer, and a multi-tailed decision layer.

[0100] It should be noted that the multi-tailed neural network model consists of three parts: a data embedding layer, a common feature extraction layer, and a multi-tailed decision layer. Each layer is composed of a fully connected MLP neural network. The data embedding layer is used to transform the input feature parameters into numerical vectors that the model can understand. The common feature extraction layer is used to extract the key features of the input data. The multi-tailed decision layer makes decisions based on the extracted features and outputs the pile foundation design scheme.

[0101] In one feasible implementation, step 202 may include: inputting the feature parameters of each sample in the training set into the data embedding layer for embedding processing to obtain a target embedding vector; inputting the target embedding vector into the common feature extraction layer for feature extraction to obtain a feature vector; inputting the feature vector into the multi-tailed decision layer for joint decision-making to obtain a prediction design scheme; comparing the prediction design scheme with the corresponding actual design scheme to obtain a prediction error; and performing backpropagation training on the multi-tailed neural network model and updating the model weights based on the prediction error to obtain a trained multi-tailed neural network model.

[0102] It should be noted that the common feature extraction layer includes one common fully connected neural network, and the multi-tailed decision layer includes multiple independent fully connected neural networks connected to the common fully connected neural network. The (N-1)th decision vector of the multiple independent fully connected neural networks is concatenated with the common feature vector to generate the Nth decision vector.

[0103] Understandably, for pile foundation design, the predicted design scheme is the predicted pile foundation type, and the actual design scheme is the actual pile foundation type. The pile foundation scheme recommendation problem can be abstracted as a classification problem of input data composed of characteristic parameters such as main beam type, main span length, span arrangement, pier height, pier location, and terrain. The classification label is the actual pile foundation type. The common feature extraction layer is used to extract useful information from the input data. It can also be a transformation or remapping of the input features. The common feature extraction layer includes multiple fully connected neural networks. The fully connected layer can serve as a data feature extraction layer. By embedding the attributes of the previous layer into a vector and remapping it to a new vector space, the final decision on the pile foundation scheme is achieved.

[0104] Understandably, due to the numerous and complex influencing variables in the pile foundation scheme, the input parameters include both enumerated and numerical variables, with different units of measurement for the numerical variables. The role of the data embedding layer is to embed the input information into a representation, that is, to uniformly map feature attributes of different types, units, and numerical ranges into a 128-dimensional vector space for subsequent feature learning and extraction processes. The multi-tailed decision layer is used to transform the features extracted by the previous layers into outputs useful for a specific task. That is, after receiving the feature information from the common feature extraction layer, the multi-tailed decision layer first makes a decision on the number of piles. The decision result is then concatenated with the feature information to make a decision on the pile diameter, ultimately obtaining the decision result for the entire pile foundation scheme.

[0105] It's worth noting that the decision regarding the pile diameter is influenced by the decision regarding the number of piles. However, during training, the decision regarding the number of piles is unknown and needs to be replaced by the decision generated by the model itself. This discrepancy leads to errors accumulating during inference, impacting model performance. Conversely, if training is performed entirely using the correct decision regarding the number of piles, the data received for the pile diameter decision will differ from the data received during the actual inference process, affecting the model's robustness and accuracy. To address this issue, a probability value ε(i) that decays during training can be used to select whether to use the actual pile number decision for training the pile diameter. Initially, there is a high probability of using the completely accurate pile number decision, while as training progresses, this probability is gradually reduced, and training uses the data obtained through the model's own inference. A linear decay function is chosen to control the probability value ε(i), i.e., ε(i) = max(ε, k - ci), where ε is the probability of using the minimum true decision scheme, and k and c control the decay rate.

[0106] It is worth noting that the predicted pile foundation type is compared with the corresponding actual pile foundation type, that is, the predicted number of piles is compared with the actual number of piles, and the predicted pile diameter is compared with the actual pile diameter. The cross-entropy loss function is used to measure the difference between the prediction results and the actual results. The differences in quantity and diameter are backpropagated and the model weights are updated. The cross-entropy loss function is as follows:

[0107] L=-∑[y×log(f(x))]

[0108] Where L is the loss value, y is the true result, and f(x) is the predicted result.

[0109] In the specific implementation, during model training, the number of foundation piles is first predicted, the error is backpropagated, and the model weights are updated. Then, the predicted number of foundation piles is concatenated with the feature vector and used as input to predict the diameter of the foundation piles. The error is backpropagated again to update the model weights, thus completing one training cycle.

[0110] In one feasible implementation, the step of inputting the feature parameters into the data embedding layer for embedding processing to obtain a target embedding vector includes: determining discrete features and continuous features based on the feature parameters; mapping the discrete features to a continuous vector space to obtain a first embedding matrix; automatically discretizing, discretely embedding, and aggregating the continuous features to obtain a second embedding matrix; and generating a target embedding vector based on the first embedding matrix and the second embedding matrix.

[0111] It should be noted that the feature parameters of the input model are divided into discrete data and continuous data, namely discrete features and continuous features. Discrete features refer to features that have a finite number of possible values ​​in the feature space, such as the five node attributes: seismic intensity s5, small mileage side beam type s6, large mileage side beam type s7, pier type s8, and geological label s9. Continuous features, on the other hand, refer to features whose feature values ​​change continuously within the real number range, such as the four node attributes: small mileage span s1, large mileage span s2, pier height s3, and average ground slope s4. In order to uniformly process these two types of features, corresponding embedding processing is required.

[0112] Understandably, discrete features, which are usually represented as text or classification labels, cannot be directly input into neural networks for processing. Therefore, it is necessary to map these discrete features into a continuous vector space and transform them into continuous features that the neural network can process.

[0113] In the specific implementation, the input values ​​of the five node attributes in the discrete features—seismic intensity s5, small mileage side beam type s6, large mileage side beam type s7, pier type s8, and geological label s9—are all enumerated integers. Therefore, through the Embedding module, each variable is mapped to a 128-dimensional continuous vector, specifically including: randomly initializing a matrix E, E∈R. V×D In this matrix, R is the set of real numbers, V is the number of enumerated categories, and D is the embedding dimension, which is 128. Each row of the matrix represents the embedding vector of one of the enumerated categories. When an index is input, the Embedding module extracts the corresponding row from matrix E to obtain a high-dimensional vector of variable mapping. Each of the five discrete feature attributes has its own embedding matrix E. i This ultimately yields a 5×128 dimensional embedding matrix, where, during training, this embedding matrix E... i The model will be updated via backpropagation to optimize its performance. The final embedding matrix for discrete features is as follows:

[0114]

[0115] Where A is the final embedding matrix of the discrete features, and E is the embedding matrix corresponding to the five discrete feature attributes.

[0116] Understandably, for continuous features, such as the span of a small mileage measurement (s1), the span of a large mileage measurement (s2), the pier height (s3), and the average ground slope (s4), since they are inherently non-discrete continuous data, the AutoDis module is used to perform end-to-end automatic encoding of continuous numerical values, encoding the continuous data into high-dimensional vectors. The AutoDis module can automatically partition continuous features, providing unique encoding representations for them, mapping high-dimensional sparse attribute features to low-dimensional dense vectors, and providing data augmentation for continuous data in downstream feature learning tasks. The AutoDis module is a deep learning framework for processing embedding learning of continuous numerical features. By introducing three core components—meta-embedding, automatic discretization, and aggregation function—it improves the model's expressive and generalization capabilities.

[0117] like Figure 7 As shown, Figure 7 This is a schematic diagram of the AutoDis principle, with the continuous variable x having a total length of 128. i Automatic discretization is performed to obtain multiple discrete features. Multiple discrete features After performing meta-embedding, aggregation is performed through aggregation functions to obtain the final discrete result of continuous variables.

[0118] It is worth noting that automatic discretization uses a two-layer neural network to automatically discretize a continuous numerical feature into 64 buckets, resulting in a matrix of size (1,64). The data feature is then processed to normalize the distribution, and the mapping result is uniformly normalized using SoftMax to obtain the final discretization result of the continuous variable. This discretization result is similar to one-hot encoding of the discrete variable and is used to select the embedding vector from the embedding matrix.

[0119] Discrete embeddings (Meta-Embeddings) are used to balance model capacity and the number of model parameters. For each feature, a shared set of Meta-Embeddings is defined. Each Meta-Embedding is a 128x64 vector space, and each Meta-Embedding can be considered a subspace of the latent space. These subspaces enhance the model's expressive power for each continuous numerical feature. During training, the Meta-Embedding also belongs to a variable vector space. During computation, the input continuous data undergoes dimensionality increase and upsampling, resulting in a size of (128, 64). Multiplying this by a dot product with the Meta-Embedding yields a matrix of size (128, 64), which is the final embedding matrix for that continuous variable.

[0120] Aggregation refers to the process where, after the input continuous numerical features are processed by the previous two modules, the probability distribution of these continuous numerical features is obtained by dispersing them across 64 discrete embeddings (Meta-Embedding). The results of the processed discrete embeddings (Meta-Embedding) are then weighted and summed with the corresponding probabilities calculated using softmax, resulting in a final discrete vector of size 1x128.

[0121] In the specific implementation, four types of continuous data, namely, the span of small mileage measurement s1, the span of large mileage measurement s2, the pier height s3, and the average ground slope s4, are automatically discretized and mapped into 64-dimensional vectors. These vectors are then aggregated with the 64x128 vector space discrete embedding Meta-Embedding to form a 128-dimensional vector. In the subsequent training process, the vectors in the Meta-Embedding and the weights of the fully connected neural network are trained through the backpropagation algorithm. Finally, AutoDis can discretize a continuous data into a 1x128 vector, thus giving each continuous value an independent embedding vector representation, enabling the network to effectively extract the feature information of continuous values.

[0122] In one feasible implementation, the step of inputting the target embedding vector into the common feature extraction layer for feature extraction to obtain a common feature vector includes: obtaining the weight matrix and bias vector of the fully connected neural network; and performing feature extraction on the target embedding vector based on the weight matrix and bias vector to obtain a common feature vector.

[0123] It should be noted that the common feature extraction layer includes one MLP fully connected neural network, namely the common fully connected neural network, and the multi-tailed decision layer includes multiple MLP fully connected neural networks, namely multiple independent fully connected neural networks connected to the common fully connected neural network. The multiple independent fully connected neural networks are multiple MLP fully connected neural networks with different numbers of nodes. That is, the MLP fully connected neural network can be used as a data feature extraction layer. By embedding the attributes of the previous layer into a vector and remapping it into a new vector space, the final decision on the pile foundation scheme can be realized.

[0124] It is understood that, in this embodiment, the common feature extraction layer may include a 6-layer fully connected MLP neural network with 1024, 512, 256, 256, 128, and 128 nodes respectively. After the feature parameters undergo embedding calculation in the data embedding layer, the data embedding vector A∈R is obtained. 9×128 After flattening the embedding vector, the final input A of the common feature extraction layer is obtained. flatten ∈R 9×1152 The data is then fed into a six-layer MLP fully connected neural network with 1024, 512, 256, 256, 128, and 128 nodes respectively for feature extraction. After calculation at each layer, the final feature extraction output A is obtained. 6 ∈R 1×128 The feature extraction of a fully connected MLP neural network is as follows:

[0125] Z l =W l ·A l-1 +b l

[0126] A l =σ(Z) l )

[0127]

[0128] in, Here is the weight matrix for layer l. Let s be the bias vector of layer l. l and s l-1 These represent the number of nodes in the current layer and the number of nodes in the previous layer, respectively. σ is the activation output of the previous layer, i.e., layer l-1, and σ is the activation function.

[0129] It is worth noting that the weights and parameters of the common feature extraction layer are globally shared parameters. That is, the decisions on the pile diameter and the number of piles will adopt the features extracted by the common feature extraction layer, reflecting the common characteristics of the pile diameter and the number of piles in the business.

[0130] In one feasible implementation, the step of inputting the feature vector into the multi-tailed decision layer for joint decision-making to obtain a predictive design scheme includes: making decisions on a first design parameter based on the feature vector to obtain a first decision vector; concatenating the first decision vector with the feature vector to obtain a concatenated vector; making decisions on a second design parameter based on the concatenated vector to obtain a second decision vector; generating decisions for a third design parameter based on the second decision vector and the common feature vector, until the decision vectors for each design parameter are generated; and summarizing the generated decision vectors for each design parameter to obtain a predictive design scheme.

[0131] It should be noted that decisions are made sequentially on the first design parameter, the second design parameter, and the Nth design parameter based on the feature vector, resulting in decision vectors for all design parameters. For pile foundation design, there are two design parameters: the first design parameter and the second design parameter. The first design parameter is the number of piles, and the first decision vector is the pile number decision vector. The second design parameter is the pile diameter, and the second decision vector is the pile diameter decision vector. Specifically, the pile number decision vector is obtained by deciding on the number of piles based on the feature vector. This pile number decision vector is then concatenated with the feature vector, and the pile diameter decision vector is obtained by deciding on the pile diameter based on the concatenated vector. Finally, the predicted pile foundation type is generated based on the pile number decision vector and the pile diameter decision vector. Feature vector A∈R 1×128 After being input into the multi-tailed decision layer, a four-layer MLP fully connected neural network with 128, 64, 32, and 9 nodes is used to calculate the feature vector, resulting in the final pile number decision vector O. pileNurmber ∈R 1×9 .

[0132] Understandably, the final decision vector for the number of foundation piles is concatenated with the features extracted from the feature layer to obtain the input of the foundation pile diameter decision module, i.e., the concatenated vector A. dia ∈R 1×137 .

[0133] It is worth noting that, according to the concatenation vector A dia ∈R 1×137 A four-layer MLP fully connected neural network with 128, 64, 32, and 8 nodes was used to calculate the spliced ​​vector, resulting in the final pile diameter decision vector O. pileDia ∈R 1×8 .

[0134] Step S203: Perform performance testing on the trained multi-tailed neural network model using the test set to obtain the loss value of the test set.

[0135] It should be noted that by testing the performance of the trained multi-tailed neural network model using a test set and obtaining the loss value of the test set, we can determine how the loss value of the test set changes in different training iterations, and thus determine whether the trained multi-tailed neural network model is overfitting.

[0136] Understandably, in the process of model training, in addition to focusing on the model's performance, it is also necessary to pay attention to the model's generalization ability. Generalization ability refers to the model's ability to adapt to new data, that is, the model can maintain a high accuracy on unseen data. In order to improve the model's generalization ability, some common techniques can be used based on the model's performance on the test set, such as regularization, data augmentation, and early stopping strategies, to prevent the model from overfitting.

[0137] Step S204: Update the parameters of the trained multi-tailed neural network model according to the loss value of the test set to obtain the scheme decision model.

[0138] It should be noted that during the entire model training process, at the 500th iteration, the loss value between the model's predicted result and the true label was calculated, and the loss curve was then plotted. The gradient descent of the loss curves on both the training and prediction sets was rapid, and the gradient descent curves on the training and test sets were very close, indicating good convergence. Figure 8 As shown, Figure 8 The diagram shows the curves of the loss function value changing with the training batch. The diagram includes (a) the tail of the pile foundation type, that is, the curve of the loss function value of the pile foundation type changing with the training batch, and (b) the tail of the number of piles, that is, the curve of the loss function value of the number of piles changing with the training batch.

[0139] In the specific implementation, according to the loss value curve, after about 100 iterations of training, the loss value of the test set tends to change gradually, but the loss function of the training set continues to decrease, indicating that the model has entered an overfitting state. Therefore, the model parameters at the time of 100 iterations of training are selected to update the current multi-tailed neural network model to obtain the solution decision model.

[0140] In the specific implementation, after model training, the final result is the confusion matrix, which shows the model's classification performance on the test set. Rows represent the true labels, columns represent the predicted labels, and each cell shows the number of samples corresponding to the true and predicted labels. The numbers on the diagonal represent the number of correctly classified samples, while the numbers off-diagonal represent the number of misclassified samples. The final model accuracy is approximately 90%. Figure 9 As shown, Figure 9 This is a schematic diagram of a confusion matrix.

[0141] It's worth noting that, regarding pile foundation design, analysis of failed prediction cases revealed that the decision-making process is essentially controlled by the pile top area. In most of the failed prediction cases, the model predicted the pile top area to be equal to the actual pile top area, but different pile foundation schemes were chosen. The model-predicted schemes remained feasible. If predictions with a difference of less than 5% between the model's predicted and actual pile top areas are considered correct, the model's accuracy rate can reach 95.6%.

[0142] like Figure 10 As shown, Figure 10 The figure shows the curves of accuracy and recall as a function of the training batch. The figure includes (a) the curve of accuracy as a function of the training batch and (b) the curve of recall as a function of the training batch.

[0143] In this embodiment, the multi-tailed neural network model is jointly trained and tested in stages using training and test sets, respectively. Then, the model parameters are updated based on the loss value of the test set to obtain the scheme decision model, which can effectively improve the prediction accuracy and generalization ability of the pile foundation scheme decision model.

[0144] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the joint decision-making multi-tailed neural network model and training method of this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0145] This application also provides a design scheme for a joint decision-making multi-tailed neural network model and a training device. Please refer to [reference needed]. Figure 11 The design scheme includes a joint decision-making multi-tailed neural network model and training device, comprising:

[0146] Module 10 is used to generate the target dataset based on historical design cases.

[0147] Training module 20 is used to perform phased joint training of the multi-tailed neural network model using the target dataset to obtain the scheme decision model.

[0148] The decision module 30 is used to make joint decisions on multiple design parameters based on the scheme decision model to obtain the target design scheme.

[0149] This embodiment provides a multi-tailed neural network model and training device for joint decision-making of design schemes. This embodiment first generates a target dataset based on historical design cases, which can improve data balance; then, it performs phased joint training of the multi-tailed neural network model using the target dataset to obtain a scheme decision model, which can effectively improve the model's accuracy and generalization ability; finally, it performs joint decision-making on multiple design parameters based on the scheme decision model to obtain the target design scheme, which can effectively improve the efficiency and accuracy of scheme decision-making.

[0150] In summary, this embodiment uses a target dataset generated based on historical design cases to perform phased joint training on a multi-tailed neural network model, and then uses the trained pile foundation scheme decision model to make joint decisions to obtain pile foundation design schemes. This overcomes the technical defects of low efficiency and poor accuracy in design scheme decision-making, and can effectively improve the efficiency and accuracy of scheme decision-making.

[0151] Optionally, the generation module 10 is further configured to extract feature parameters from the historical design cases to obtain an initial dataset; and to perform data augmentation and data cleaning on the initial dataset according to the association parameters to obtain a target dataset.

[0152] Optionally, the training module 20 is further configured to divide the target dataset into a training set and a test set according to a preset ratio, wherein the target dataset includes multiple samples, each sample including different feature parameters and corresponding actual design schemes; perform phased joint training on the multi-tailed neural network model using the training set to obtain a trained multi-tailed neural network model, wherein the multi-tailed neural network model includes a data embedding layer, a common feature extraction layer, and a multi-tailed decision layer; perform performance testing on the trained multi-tailed neural network model using the test set to obtain the loss value of the test set; and update the parameters of the trained multi-tailed neural network model based on the loss value of the test set to obtain a scheme decision model.

[0153] Optionally, the training module 20 is further configured to: input the feature parameters of each sample in the training set into the data embedding layer for embedding processing to obtain a target embedding vector; input the target embedding vector into the common feature extraction layer for feature extraction to obtain a feature vector; input the feature vector into the multi-tailed decision layer for joint decision-making to obtain a prediction design scheme; compare the prediction design scheme with the corresponding actual design scheme to obtain a prediction error; and perform backpropagation training on the multi-tailed neural network model and update the model weights based on the prediction error to obtain the trained multi-tailed neural network model.

[0154] Optionally, the training module 20 is further configured to determine discrete features and continuous features based on the feature parameters; map the discrete features to a continuous vector space to obtain a first embedding matrix; automatically discretize, discretely embed, and aggregate the continuous features to obtain a second embedding matrix; and generate a target embedding vector based on the first embedding matrix and the second embedding matrix.

[0155] Optionally, the training module 20 is further configured to obtain the weight matrix and bias vector of the fully connected neural network; and to perform feature extraction on the target embedding vector based on the weight matrix and bias vector to obtain a feature vector.

[0156] Optionally, the training module 20 is further configured to: make decisions on the first design parameters based on the feature vector to obtain a first decision vector; concatenate the first decision vector with the feature vector to obtain a concatenated vector; make decisions on the second design parameters based on the concatenated vector to obtain a second decision vector; generate decisions for the third design parameters based on the second decision vector and the common feature vector, until the decision vectors for each design parameter are generated; and summarize the generated decision vectors for each design parameter to obtain a predicted design scheme.

[0157] The design scheme of the joint decision-making multi-tailed neural network model and training device provided in this application, employing the design scheme of the joint decision-making multi-tailed neural network model and training method in the above embodiments, can solve the technical problem of how to effectively improve the efficiency and accuracy of scheme decision-making. Compared with the prior art, the beneficial effects of the design scheme of the joint decision-making multi-tailed neural network model and training device provided in this application are the same as the beneficial effects of the design scheme of the joint decision-making multi-tailed neural network model and training method provided in the above embodiments, and other technical features in the design scheme of the joint decision-making multi-tailed neural network model and training device are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0158] This application provides a design scheme for a joint decision-making multi-tailed neural network model and training device. The design scheme for the joint decision-making multi-tailed neural network model and training device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the design scheme for the joint decision-making multi-tailed neural network model and training method in the above embodiment 1.

[0159] The following is for reference. Figure 12This document illustrates a structural schematic diagram of a joint decision-making multi-tailed neural network model and training device suitable for implementing the design schemes of the embodiments of this application. The joint decision-making multi-tailed neural network model and training device in the embodiments of this application can include, but are not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 12 The design scheme of the joint decision-making multi-tailed neural network model and training device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0160] like Figure 12 As shown, the design scheme for the joint decision-making multi-tailed neural network model and training device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 1002 or the program loaded from the storage device 1003 into the random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the design scheme's joint decision-making multi-tailed neural network model and training device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the design scheme's joint decision-making multi-tailed neural network model and training equipment to wirelessly or wiredly communicate with other devices to exchange data. Although the figure shows a design scheme's joint decision-making multi-tailed neural network model and training equipment with various systems, it should be understood that implementing or having all the systems shown is not required. More or fewer systems can be implemented alternatively.

[0161] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0162] The design scheme of the joint decision-making multi-tailed neural network model and training device provided in this application, employing the design scheme of the joint decision-making multi-tailed neural network model and training method in the above embodiments, can solve the technical problem of how to effectively improve the efficiency and accuracy of scheme decision-making. Compared with the prior art, the beneficial effects of the design scheme of the joint decision-making multi-tailed neural network model and training device provided in this application are the same as the beneficial effects of the design scheme of the joint decision-making multi-tailed neural network model and training method provided in the above embodiments, and other technical features in the design scheme of the joint decision-making multi-tailed neural network model and training device are the same as the features disclosed in the previous embodiment method, and will not be repeated here.

[0163] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0164] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0165] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the joint decision-making multi-tailed neural network model and training method of the above embodiments.

[0166] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0167] The aforementioned computer-readable storage medium may be included in the joint decision-making multi-tailed neural network model and training device of the design scheme; or it may exist independently and not be assembled into the joint decision-making multi-tailed neural network model and training device of the design scheme.

[0168] The aforementioned computer-readable storage medium carries one or more programs. When the aforementioned one or more programs are executed by the design scheme joint decision-making multi-tailed neural network model and training device, the design scheme joint decision-making multi-tailed neural network model and training device: generate a target dataset based on historical design cases; perform phased joint training on the multi-tailed neural network model using the target dataset to obtain a scheme decision model; and make joint decisions on multiple design parameters based on the scheme decision model to obtain a target design scheme.

[0169] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, C++, C#, and Python—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0170] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0171] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0172] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., computer programs) for executing the joint decision-making multi-tailed neural network model and training method of the above-described design scheme. This addresses the technical problem of how to effectively improve the efficiency and accuracy of scheme decision-making. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the joint decision-making multi-tailed neural network model and training method provided in the above embodiments, and will not be elaborated upon here.

[0173] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the joint decision-making multi-tailed neural network model and training method as described above.

[0174] The computer program product provided in this application can solve the technical problem of how to effectively improve the efficiency and accuracy of solution decision-making. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the joint decision-making multi-tailed neural network model and training method provided in the above embodiments, and will not be repeated here.

[0175] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A training method for a joint decision-making multi-tailed neural network model, characterized in that, The method includes: Generate the target dataset based on historical design cases; The target dataset is divided into a training set and a test set according to a preset ratio. The target dataset includes multiple samples, each of which includes different feature parameters and corresponding actual design schemes. The multi-tailed neural network model is jointly trained in stages using the training set to obtain the trained multi-tailed neural network model, wherein the multi-tailed neural network model includes a data embedding layer, a common feature extraction layer, and a multi-tailed decision layer. The multi-tailed neural network model is jointly trained in stages using the training set to obtain the trained multi-tailed neural network model, including: The feature parameters of each sample in the training set are input into the data embedding layer for embedding processing to obtain the target embedding vector; The target embedding vector is input into the common feature extraction layer for feature extraction to obtain a common feature vector; The common feature vector is input into the multi-tailed decision layer for joint decision-making to obtain a predictive design scheme. The predicted design scheme is compared with the corresponding actual design scheme to obtain the prediction error; The multi-tailed neural network model is backpropagated and trained based on the prediction error, and the model weights are updated to obtain the trained multi-tailed neural network model. The trained multi-tailed neural network model is tested using the test set to obtain the loss value of the test set. The parameters of the trained multi-tailed neural network model are updated based on the loss values ​​of the test set to obtain the solution decision model; The target design scheme is obtained by jointly deciding on multiple design parameters based on the scheme decision model.

2. The method as described in claim 1, characterized in that, The target dataset generated based on historical design cases includes: Feature parameters are extracted from the historical design cases to obtain the initial dataset; The initial dataset is augmented and cleaned based on the association parameters to obtain the target dataset.

3. The method as described in claim 1, characterized in that, The step of inputting the feature parameters into the data embedding layer for embedding processing to obtain the target embedding vector includes: Discrete and continuous features are determined based on the aforementioned feature parameters; The discrete features are mapped to a continuous vector space to obtain the first embedding matrix; The continuous features are automatically discretized, discretized, and aggregated to obtain a second embedding matrix; Generate a target embedding vector based on the first embedding matrix and the second embedding matrix.

4. The method as described in claim 1, characterized in that, The step of inputting the target embedding vector into the common feature extraction layer for feature extraction to obtain a common feature vector includes: Obtain the weight matrix and bias vector of a fully connected neural network; Based on the weight matrix and bias vector, feature extraction is performed on the target embedding vector to obtain a common feature vector.

5. The method as described in claim 1, characterized in that, The step of inputting the common feature vector into the multi-tailed decision layer for joint decision-making to obtain a predictive design scheme includes: Based on the common feature vector, a decision is made on the first design parameters to obtain the first decision vector; The first decision vector is concatenated with the common feature vector to obtain the concatenated vector; Based on the splicing vector, a decision is made on the second design parameters to obtain the second decision vector; The decision-making process for the third design parameter is performed based on the second decision vector and the common feature vector until the decision vectors for each design parameter are generated. The decision vectors generated for each design parameter are summarized to obtain the predicted design scheme.

6. A training device for a joint decision-making multi-tailed neural network model, characterized in that, The training device for the joint decision-making multi-tailed neural network model of the aforementioned design scheme includes: The generation module is used to generate target datasets based on historical design cases; The training module is used to divide the target dataset into a training set and a test set according to a preset ratio. The target dataset includes multiple samples, each containing different feature parameters and corresponding actual design schemes. The multi-tailed neural network model is jointly trained in stages using the training set to obtain a trained multi-tailed neural network model. This multi-tailed neural network model includes a data embedding layer, a common feature extraction layer, and a multi-tailed decision layer. The multi-tailed neural network model is jointly trained in stages using the training set to obtain the trained multi-tailed neural network model, including: inputting the feature parameters of each sample in the training set into the data embedding layer for embedding processing to obtain the target dataset. Embedded vector; the target embedded vector is input to the common feature extraction layer for feature extraction to obtain a common feature vector; the common feature vector is input to the multi-tailed decision layer for joint decision-making to obtain a predicted design scheme; the predicted design scheme is compared with the corresponding actual design scheme to obtain the prediction error; the multi-tailed neural network model is trained by backpropagation based on the prediction error and the model weights are updated to obtain a trained multi-tailed neural network model; the trained multi-tailed neural network model is tested using the test set to obtain the loss value of the test set; the parameters of the trained multi-tailed neural network model are updated based on the loss value of the test set to obtain a scheme decision model; The decision module is used to make joint decisions on multiple design parameters according to the scheme decision model to obtain the target design scheme. The training program device of the design scheme joint decision multi-tailed neural network model is used to implement the training method of the design scheme joint decision multi-tailed neural network model as described in any one of claims 1 to 5.

7. A training device for a joint decision-making multi-tailed neural network model, characterized in that, The training device for the design scheme joint decision-making multi-tailed neural network model includes: a memory, a processor, and a training program for the design scheme joint decision-making multi-tailed neural network model stored in the memory and capable of running on the processor. The training program for the design scheme joint decision-making multi-tailed neural network model is configured to implement the training method for the design scheme joint decision-making multi-tailed neural network model as described in any one of claims 1 to 5.

8. A storage medium, characterized in that, The storage medium stores a training program for a design scheme joint decision-making multi-tailed neural network model. When the training program for the design scheme joint decision-making multi-tailed neural network model is executed by the processor, it implements the training method for the design scheme joint decision-making multi-tailed neural network model as described in any one of claims 1 to 5.

Images