Customized product configuration design scheme recommendation method based on retrieval and ranking model

Abstract

The application discloses a customized product configuration design scheme recommendation method based on a retrieval and sorting model. First, according to historical customer order data and corresponding bill of materials, a screening demand feature vector and an instance feature vector of a corresponding module instance are determined; then, module classification is performed on all module instances to obtain a module instance set of different modules, and retrieval data sets and sorting data sets of the modules are processed and obtained; then, the retrieval data sets of the modules are respectively input into corresponding module retrieval models, and the sorting data sets are input into a sorting model for training, so that trained retrieval models and sorting models corresponding to the modules are obtained; according to new order data, all new data sample features are determined and then input into the two trained models, and each optimal module instance is selected to manufacture a current customized product. The application reduces the difficulty and cost of complex configuration knowledge representation, maintenance and update, and improves the customized product configuration design efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to a customized product configuration design scheme recommendation method in the field of customized product configuration design, and in particular to a customized product configuration design scheme recommendation method based on a retrieval and ranking model. BACKGROUND

[0002] Product configuration design can quickly form a new product meeting the individualized customization needs of customers by combining the existing pre-defined configurable components (modules) of an enterprise, thereby supporting low-cost, high-efficiency, short-cycle mass customization of products.

[0003] For traditional product configuration design methods, the representation, maintenance and updating of product complex configuration knowledge is a bottleneck restricting the rationality and effectiveness of these methods. Configuration knowledge representation is highly dependent on expert experience and domain knowledge, and is mainly realized manually, which is high in labor cost. The diversification and individualization of customer needs lead to increasingly complex customized products, thereby increasing the difficulty of connection and constraint analysis between product configurable components, and making it difficult to effectively extract and express configuration knowledge. At the same time, the diversification and individualization of customer needs inevitably lead to dynamic changes in product configurable components, product structure and configuration knowledge, thereby limiting the sharing and reuse of complex design knowledge, and resulting in high cost of configuration knowledge maintenance and updating.

[0004] In the product configuration solving process, if all the rules corresponding to the product configurable components do not completely match the customer needs or all the constraints corresponding thereto cannot be satisfied simultaneously, the configurable components will not be instantiated, thereby leading to the problem of customized product module mismatch. For each mismatched component, we can search for the most similar instance by applying an additional similarity solving method, and then perform structural mutation design on the similar instance to obtain a product module instance meeting the customer needs. For complex customized products, we need to combine multiple configuration design methods to perform hierarchical configuration solving, so as to obtain a product configuration design scheme meeting all the needs of customers. The problem of customized product module mismatch affects the rapid response to diversified and individualized customer customization needs, thereby reducing the efficiency of product configuration design. SUMMARY

[0005] In view of the above defects or improvement needs of existing methods, the present application provides a customized product configuration design scheme recommendation method based on a retrieval and ranking model, which recommends a list of customized product module instances ranked from high to low in matching degree with customer needs, in view of individualized customer needs and customized product module information.

[0006] To achieve the above object, the technical scheme of the present application is as follows:

[0007] 1. A customized product configuration design scheme recommendation method based on a retrieval and ranking model

[0008] 1) According to the historical customized product customer order data and the corresponding configuration design bill of materials, all demand features of each customer order data are identified and extracted, and the demand feature vectors of the corresponding customer order data are composed of all demand features of each customer order data; according to the configuration design bill of materials, all module instances corresponding to each customer order data are screened and obtained, and each instance feature of all module instances is extracted, and the instance feature vectors of the corresponding module instances are composed of each instance feature of all module instances;

[0009] 2) The module classification is performed on all module instances, the module instance sets of different modules are obtained, the instance feature vectors of each module instance of each module are combined with the corresponding demand feature vectors to form each data sample feature of the current module, the retrieval data set of the current module is composed of each data sample feature of the current module, the ranking data set of the current module is composed of each data sample feature and the corresponding true recommendation score, the module instance set of each module is traversed and the data set is constructed, and the retrieval data set and the ranking data set of each module are obtained;

[0010] 3) The retrieval data set and the ranking data set of each module are respectively preprocessed to obtain the preprocessed retrieval data set and the preprocessed ranking data set corresponding to each module;

[0011] 4) The preprocessed retrieval data set of each module is input into the retrieval model corresponding to the module for training, each retrieval model outputs the affinity score between each instance feature vector and the corresponding demand feature vector, and the trained retrieval model corresponding to each module is obtained through training;

[0012] 5) The preprocessed ranking data set of each module is input into the ranking model corresponding to the module for training, each ranking model outputs the predicted recommendation score between each instance feature vector and the corresponding demand feature vector, and the trained ranking model corresponding to each module is obtained through training;

[0013] 6) According to the new customer order data, obtain the demand feature vector of the current customer order data and determine the required modules of the current customer order data, then the instance feature vectors of the required module instances corresponding to each module are all module instance instance feature vectors and the demand feature vector of the current customer order data form all new data sample features, then all new data sample features of each module are preprocessed and input into the trained retrieval model of the corresponding module, and the affinity score between all module instances of each module and the current customer order data is output; select multiple module instances of each module in order of affinity score from high to low and record as the module instance candidate set of the corresponding module; then input the new data sample features corresponding to all module instances in the module instance candidate set of each module into the trained ranking model of the corresponding module, output the recommendation score of all module instances in the module instance candidate set of each module, select the module instance with the highest recommendation score in the module instance candidate set of each module as the configuration design scheme of the current customer order data, and manufacture the corresponding customized product according to the current configuration design scheme.

[0014] In step 3), the data augmentation processing is performed on the data sample features with missing feature values in the retrieval data set and the ranking data set, and the data one-hot encoding processing is performed on the data sample features of the division type.

[0015] The retrieval model structure of each module is the same, and includes a first embedding layer, a query network, a candidate network and a first output layer; the input of the retrieval model is input into the first embedding layer, the demand feature vector of each data sample feature in the retrieval data set is input into the query network after the corresponding first embedding layer, the instance feature vector of each data sample feature in the retrieval data set is input into the candidate network after the corresponding first embedding layer, the output of the query network and the output of the candidate network are input into the first output layer, and the first output layer outputs the affinity score between each instance feature vector and the corresponding demand feature vector.

[0016] The structure of the query network and the candidate network is the same, which is a forward neural network containing two hidden layers.

[0017] The ranking model structures of the respective modules are same, and each includes a second embedding layer, an interaction network and a second output layer; input of the ranking model is input to the second embedding layer as input, customer demand features and module instance features of each data sample feature in the ranking data set are respectively input to corresponding second embedding layers, the interaction network includes an adaptive weight interaction block and an adaptive power order interaction block, outputs of each second embedding layer are respectively input to the adaptive weight interaction block and the adaptive power order interaction block, outputs of the adaptive weight interaction block and the adaptive power order interaction block are both input to the second output layer, and the second output layer outputs a predicted recommendation score between each instance feature vector and a corresponding demand feature vector.

[0018] The adaptive weight interaction block includes two layers of multi-head self-attention interaction layers, input of the adaptive weight interaction block is input to a first multi-head self-attention interaction layer, output of the first multi-head self-attention interaction layer is added to the input of the adaptive weight interaction block and then input to a second multi-head self-attention interaction layer, and output of the second multi-head self-attention interaction layer is added to the output of the input of the adaptive weight interaction block to serve as output of the adaptive weight interaction block.

[0019] The adaptive power order interaction block includes one layer of logarithmic transformation layers and two layers of fully connected hidden layers, input of the adaptive power order interaction block is input to the logarithmic transformation layer, the logarithmic transformation layer is connected to a first fully connected hidden layer and a second fully connected hidden layer in sequence, and output of the second fully connected hidden layer serves as output of the adaptive power order interaction block.

[0020] Parameters of the embedding layer of the ranking model of the corresponding module are initialized according to parameters of the embedding layer of the trained retrieval model of each module.

[0021] II. A storage medium storing a computer program, the computer program being executed by a processor to implement the method.

[0022] The computer program is instructions corresponding to the method.

[0023] The present application has the following advantages:

[0024] The method of the present application combines deep learning to extract feature interactions between customer needs, between customized product modules, and between customer needs and customized product modules, thereby reducing the difficulty and cost of complex configuration knowledge representation, maintenance and updating; collaborative filtering is used to identify similar module instances by calculating the affinity score between customer needs and customized product module instance candidates to narrow down the recommendation range and improve recommendation efficiency; module instance candidates are scored based on cross features, and the module instance that best meets the customer's needs is recommended according to the score, so that there is no need to additionally use similar configuration solving or joint multiple configuration solving methods for hierarchical configuration solving to handle the customized product module mismatch problem, thereby improving the efficiency of customized product configuration design. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 Flowchart of the customized product configuration design scheme recommendation method according to the present application;

[0026] Figure 2 Topological structure diagram of the retrieval model and the ranking model constructed according to the present application;

[0027] Figure 3 Schematic diagram of the first layer multi-head self-attention interaction layer extracting low-order and high-order cross features of the adaptive weight interaction block of the ranking model according to the embodiment of the present application;

[0028] Figure 4 Schematic diagram of the logarithmic transformation layer extracting adaptive-order cross features of the adaptive power-order interaction block of the ranking model according to the embodiment of the present application. DETAILED DESCRIPTION

[0029] The present application will be further described in detail below in combination with the drawings and specific embodiments.

[0030] The present application takes the configuration design of a customized elevator product as an embodiment, combines deep learning, and recommends a list of customized elevator product module instances with a matching degree with customer needs from high to low for personalized customer needs and customized elevator product module information, to quickly respond to diversified and personalized customer needs and realize the configuration design of a customized elevator product. Figure 1 is a flowchart of the customized product configuration design scheme recommendation method according to the present application, as Figure 1 shown.

[0031] A customized product configuration design scheme recommendation method based on retrieval and ranking according to the present application includes the following steps:

[0032] 1) According to historical customer order data of customized products and a corresponding configuration design bill of materials, all demand features of each customer order data are identified and extracted, and a demand feature vector of the corresponding customer order data is formed by all demand features of each customer order data; according to the configuration design bill of materials, all module instances corresponding to each customer order data are screened and obtained, each product of a historical customer order is composed of multiple modules, each module has a corresponding module instance, and each instance feature of all module instances is extracted to form an instance feature vector of the corresponding module instance;

[0033] 2) The module classification is performed on all module instances to obtain module instance sets of different modules, each data sample feature of a current module is formed by the instance feature vector of each module instance of the module instance set of each module and the corresponding demand feature vector, the retrieval data set of the current module is formed by each data sample feature of the current module, the sorting data set of the current module is formed by each data sample feature and the corresponding real recommendation score, the module instance set of each module is traversed and the data set is constructed to obtain the retrieval data set and the sorting data set of each module;

[0034] The data samples in the retrieval data set do not need to be assigned sample labels. When the retrieval model is trained, the affinity score between the customer demand and the customized product module instance satisfying the customer demand needs to be maximized, and the affinity score between the customer demand and other customized product module instances not satisfying the customer demand needs to be minimized. Therefore, when the loss function is calculated, for a batch of data samples (the batch size is N B ), the affinity score real value of the data sample is an eye matrix of N B ×N B .

[0035] In the embodiment of the application, 168 customer demand features and product module information features are identified and extracted for the customized elevator product, the retrieval data set contains 159,555 data samples, and the sorting data set contains 267,355 data samples. The customer demand features include rated load, rated speed, rated power, rated current, floor height, super height, lifting height, shaft width, shaft depth, and the like; and the product module information features include module identification, module production manufacturer, module material identification, and the like.

[0036] In the embodiment of the application, the traction machine module of the customized elevator product is taken as an example, the first three data samples in the retrieval data set of the elevator traction machine module are shown in Table 1, and the first three data samples in the sorting data set of the elevator traction machine module are shown in Table 2.

[0037] Table 1: Three data sample examples in the retrieval data set of the elevator traction machine module

[0038]

[0039]

[0040] Table 2: 3 data sample examples in the ranking data set of the elevator hoist module

[0041]

[0042] 3) data preprocessing is performed on the retrieval data set and the ranking data set of each module respectively, and a preprocessed retrieval data set and a preprocessed ranking data set corresponding to each module are obtained;

[0043] In step 3), data augmentation processing is performed on the data sample features with missing feature values in the retrieval data set and the ranking data set, and one-hot encoding processing is performed on the data sample features of the discrete type.

[0044] When the number of missing feature values of a data sample exceeds 5 during data augmentation processing of the data set by the embodiment of the application, the data sample is removed, otherwise the missing features of the data sample are filled with the mean value of the features.

[0045] When one-hot encoding processing is performed on the data set by the embodiment of the application, when the type of the feature value of a data sample is a discrete type feature, the discrete type feature is converted into a sparse binary vector by one-hot encoding.

[0046] 4) the preprocessed retrieval data set of each module is input into the retrieval model of the corresponding module for training, each retrieval model outputs the affinity score between each instance feature vector and the corresponding demand feature vector, and a trained retrieval model corresponding to each module is obtained by training;

[0047] The retrieval model of each module has the same structure and includes a first embedding layer, a query network, a candidate network and a first output layer; the input of the retrieval model is input into the first embedding layer as the input of the first embedding layer, the demand feature vector of each data sample feature in the retrieval data set is input into the query network after passing through the corresponding first embedding layer, and the instance feature vector of each data sample feature in the retrieval data set is input into the candidate network after passing through the corresponding first embedding layer, wherein each feature in the vector has a unique first embedding layer corresponding thereto, the output of the query network and the output of the candidate network are input into the first output layer, and the first output layer outputs the affinity score between each instance feature vector and the corresponding demand feature vector, i.e. the affinity score between the customer demand and each module instance.

[0048] The query network and the candidate network have the same structure to ensure that the feature representation of customer needs and the feature representation of customized product modules have the same dimension. Both are feedforward neural networks containing two hidden layers.

[0049] Figure 2 This is a topological structure diagram of the retrieval model and ranking model constructed in this invention. (See diagram below.) Figure 2 As shown, for the input of the retrieval model, the first j features are all demand features, therefore the demand feature vector can be expressed as x. query = [x1,x2,…,x j The instance feature vector can be expressed as x. candidate =[x j+1 ,x j+2 ,…,x m If the i-th input feature is a categorical feature, then x i It is a sparse binary vector; if the i input features are numerical features, then x i It is a scalar value. m is the total number of input features, which is the sum of the number of required features and the number of instance features.

[0050] To reduce dimensionality, the embedding layer transforms the sparse input feature vector into a low-dimensional, dense feature embedding. Therefore, the feature embedding required by the customer can be expressed as:

[0051] e query =[e1,e2,…,e j ],

[0052]

[0053] in, The feature embedding representing customer needs, e l (1≤l≤j) represents the feature embedding of the l-th customer demand feature, j represents the number of customer demand features, and d represents the feature embedding of the l-th customer demand feature. e Indicates the dimension of feature embedding; These are the parameters for the first and second models, respectively; when x l When it is a categorical feature, d vocab,l x represents the l-th customer demand feature. l The total number of possible values ​​for q l x represents the l-th customer demand feature. l The number of elements in the set whose value is 1.

[0054] The feature embedding of a customized product module instance can be expressed as:

[0055] e candidate =[e j+1 ,e j+2 ,…,e m ],

[0056]

[0057] wherein, denotes the module M p the feature embedding of the corresponding module instance, e k (j+1≤k≤m) denotes the feature embedding of the kth module instance feature x k , (m-j) denotes the number of module instance features corresponding to the module M p . are the third and fourth model parameters, respectively; when x k is a categorical feature, d vocab,k denotes the total number of values of the kth module instance feature x k , q k denotes the number of elements in the kth module instance feature x k with a value of 1.

[0058] The structures of the query network and the candidate network are the same to ensure that the feature expression of the customer demand and the feature expression of the customized product module instance have the same dimension, both being a forward neural network containing two hidden layers, wherein the number of neurons of the hidden layers is {256, 64}, and the activation function of all hidden layers is a ReLU function.

[0059] The embedding layer converts the high-dimensional sparse feature vector into a low-dimensional dense feature vector to reduce the model parameters and improve the model training efficiency. The query network takes the output of the customer demand features after the embedding layer as the input, extracts the association relationship between the customer demands, and outputs the feature expression of the customer demand; and the candidate network takes the output of the module instance features after the embedding layer as the input, extracts the association relationship between the module instance features of the module M p , and outputs the feature expression of the module instance corresponding to the module M p . In the output layer, the outputs of the query network and the candidate network are multiplied to calculate the affinity score of the customer demand and the module instance of the module M p . The higher the affinity score, the higher the matching degree between the module instance and the customer demand.

[0060] In specific implementation, the retrieval dataset is divided into a training set and a test set according to a ratio of 7:3, the training set is used to train the model parameters, and the test set is used to test the model performance;

[0061] The feature expression of the customer demand output by the query network can be represented as:

[0062]

[0063]

[0064] wherein, respectively represent the outputs of the first layer, the second layer hidden layer of the query network, is the output of the query network. are model parameters, representing the weight and bias of the first layer hidden layer of the query network, respectively. are model parameters, representing the weight and bias of the second layer hidden layer of the query network, respectively. l1 l2 respectively represent the number of neurons of the first layer, the second layer hidden layer of the query / candidate network, and ReLU() represents the ReLU function.

[0065] The module M p The feature expression of the module instance of the module M

[0066]

[0067]

[0068] wherein, respectively represent the outputs of the first layer, the second layer hidden layer of the query / candidate network, is the output of the query / candidate network. are model parameters, representing the weight and bias of the first layer hidden layer of the query / candidate network, respectively. are model parameters, representing the weight and bias of the second layer hidden layer of the query / candidate network, respectively.

[0069] The output layer of the retrieval model calculates the affinity score between the feature expression of the customer demand and the feature expression of the module instance of the module M p by multiplying the feature expression of the customer demand and the feature expression of the module instance of the module M p . Therefore, the affinity score between the customer demand and the module instance of the module M p can be expressed as:

[0070]

[0071] wherein, represents the affinity score between the customer demand and the module instance of the module M p , and T represents the transpose.

[0072] When training the retrieval model, it is necessary to maximize the affinity score between the customer demand and the module instance of the module M p that meets the customer demand, and minimize the affinity score between the customer demand and the module instance of the module M p that does not meet the customer demand. Therefore, as Figure 2 ​The output of the retrieval model is a vector containing affinity scores between the customer demand and module instances of different modules M p , and the retrieval model is trained using a categorical cross-entropy loss function, which can be represented as:

[0073]

[0074] wherein, respectively represent the true value and the calculated value of the affinity score between the customer demand corresponding to the s-th retrieval training sample and the module instance of the t-th module M p . When the module instance of the t-th module M p meets the customer demand corresponding to the s-th retrieval training sample, otherwise, calculated by the retrieval model; N B represents the sample batch size in the training process; represents the parameter set of the retrieval model; λ1 is a penalty coefficient. For a batch of data samples, the true value of the affinity score is an eye matrix of N B × N B . Therefore, the data samples of the retrieval data set do not need to be labeled.

[0075] 5) input the preprocessed ranking data set of each module into the ranking model of the corresponding module for training, and each ranking model outputs a predicted recommendation score between each instance feature vector and the corresponding demand feature vector, and a trained ranking model corresponding to each module is obtained through training;

[0076] The ranking model of each module has the same structure, which includes a second embedding layer, an interaction network and a second output layer; the input of the ranking model is input into the second embedding layer as the input of the second embedding layer, and the customer demand feature and the module instance feature of each data sample feature in the ranking data set are input into the corresponding second embedding layer, wherein each feature in the vector has a unique second embedding layer corresponding thereto, the interaction network includes an adaptive weight interaction block and an adaptive power order interaction block, the output of each second embedding layer is input into the adaptive weight interaction block and the adaptive power order interaction block respectively, the output of the adaptive weight interaction block and the output of the adaptive power order interaction block are input into the second output layer, and the second output layer outputs a predicted recommendation score between each instance feature vector and the corresponding demand feature vector.

[0077] The embedding layer converts a high-dimensional sparse feature vector into a low-dimensional dense feature vector to reduce model parameters and improve model training efficiency. The interaction network takes the feature embedding of the customer demand and the module instance as input, extracts the customer demand, the module M pBetween module instances and customer requirements and module M p The relationships between module instances are learned, including low-order, high-order, and adaptive-order cross features. The output layer connects all learned cross features to module M. p The module instance is used to calculate the recommendation score. The higher the recommendation score, the better the module M performs. p The more likely a module instance is to meet customer needs, the higher its recommendation score can be used to evaluate the modules M selected by the retrieval model. p The candidate module instances are sorted and the best candidate module instance is recommended.

[0078] In practice, the sorted dataset is divided into a training set and a test set in a 7:3 ratio. The training set is used to train the model parameters, and the test set is used to test the model performance.

[0079] The parameters of the embedding layer of the ranking model for the corresponding module are initialized based on the parameters of the embedding layer of the pre-trained retrieval model for each module.

[0080] The output of the embedding layer of the ranking model can be represented as:

[0081] e = [e1, e2, ..., e m ],

[0082]

[0083] in, Let m represent the feature embedding of the i-th input, and d represent the number of feature embeddings of the input. e This represents the dimension of each feature embedding. When 1 ≤ i ≤ j, e i This corresponds to the feature embedding of the sparse feature vector of the input customer demand; when j+1≤i≤m, e i This corresponds to the feature embedding of the sparse feature vector of the input module instance. These are all parameters of the fifth and sixth models; when x i When it is a categorical feature, d vocab,i x represents i The total number of possible values ​​for q i Represents a sparse binary vector x i The number of elements in the set whose value is 1.

[0084] The adaptive weight interaction block comprises two multi-head self-attention interaction layers. The input of the adaptive weight interaction block is input to the first multi-head self-attention interaction layer. The output of the first multi-head self-attention interaction layer is added to the input of the adaptive weight interaction block and then input to the second multi-head self-attention interaction layer. The output of the second multi-head self-attention interaction layer after being added to the input of the adaptive weight interaction block is the output of the adaptive weight interaction block. Based on the multi-head self-attention mechanism, the multi-head self-attention interaction layer extracts meaningful high-order cross features. Meanwhile, in the adaptive weight interaction block, a residual connection is used to retain all low-order cross features.

[0085] The adaptive power-order interaction block comprises one logarithmic transformation layer and two fully connected hidden layers. The input of the adaptive power-order interaction block is input to the logarithmic transformation layer. The logarithmic transformation layer is connected to the first fully connected hidden layer and then connected to the second fully connected hidden layer. The output of the second fully connected hidden layer is the output of the adaptive power-order interaction block. The logarithmic transformation layer comprises a plurality of logarithmic neurons. Each logarithmic neuron can extract different adaptive-order cross features.

[0086] For the output e of the embedding layer of the ranking model, the adaptive weight interaction block connects them into a two-dimensional matrix E0. The formal expression is as follows:

[0087]

[0088] As shown in the following formula, for the two-dimensional matrix E0, the output of the first interaction layer can be expressed as: Figure 3

[0089]

[0090]

[0091]

[0092]

[0093] wherein, represents the output of the first interaction layer; represents all learned 2-order cross features, and H1 represents the number of heads in the first interaction layer; represents the learned cross features under the hth head in the first interaction layer; represents a matrix connection operator, and has is a weight matrix, and has softmax(·) is a normalized exponential function, and has are all parameter matrices.​

[0094] Similarly, the output of the second layer interaction layer can be represented as:

[0095]

[0096]

[0097]

[0098]

[0099] wherein, represents the output of the second layer interaction layer; represents the learned cross feature under the hth head in the second layer interaction layer, and H2 represents the number of heads in the second layer interaction layer; is a weight matrix; represents all the learned 2-4 order cross features; are all parameter matrices.

[0100] As shown in Figure 4 , for the input customer demand or the feature embedding e p of the module instance of the module M i (1≤i≤m), the output of the u th logarithmic neuron of the logarithmic transformation layer can be expressed as:

[0101]

[0102] wherein, represents the output of the u th logarithmic neuron of the logarithmic transformation layer, N L represents the number of logarithmic neurons of the logarithmic transformation layer; w iu (1≤i≤m,1≤u≤N L ) is a model parameter, and w iu may be any non-negative real number. Therefore, y u may be an arbitrary order cross feature. The adaptive cross feature means that the order of each feature in the cross feature is determined by the adaptive learned parameter w iu , instead of a specified positive integer. For example: when w1 = w2 = 1, represents the 2-order cross feature of e1 and e2; when w1 = 0.5, w2 = 0.7, represents the 1.2-order cross feature of e1 and e2. Wherein, ⊙ represents the element-wise multiplication operator, and [a b]⊙[c d] = [ac bd].

[0103] The outputs of all the logarithmic neurons connected to the logarithmic transformation layer are connected as the input of the fully connected hidden layer, i.e. y Therefore, the output of the adaptive power order interaction block can be expressed as:

[0104] z1 = ReLU(W1z0 + b1),

[0105] z2 = ReLU(W2z1 + b2),

[0106] in, This represents the output of the second fully connected hidden layer of the adaptive power-order interaction block, i.e., the output of the adaptive power-order interaction block. This represents the output of the first fully connected hidden layer of the adaptive power-order interaction block; Both are model parameters, representing the weights and biases of the first fully connected hidden layer of the adaptive power-order interaction block, respectively, and d n1 This represents the number of neurons in the fully connected hidden layer. Both are model parameters, representing the weights and biases of the second fully connected hidden layer of the adaptive power-order interaction block, respectively, and d n2 This represents the number of neurons in the fully connected hidden layer.

[0107] like Figure 1 As shown, the output layer of the ranking model takes the output of the adaptive weight interaction block and the power-order interaction block as its input, and the computation module M... p The recommendation score of the module instance is used as the output. Therefore, the output of the output layer of the ranking model can be expressed as:

[0108] x e =flatten(E2),

[0109]

[0110] in, This represents a one-dimensional vector after flattening the two-dimensional output matrix E2 of the adaptive weighted interaction layer. flatten(·) is the matrix flattening function. The output of the ranking model is the recommendation score calculated by the ranking model for the customized product module instance. b out Both are model parameters, representing the weights and biases of the output layer of the ranking model, respectively; σ(·) represents the sigmoid function, and σ(x) = 1 / (1+e -x ).

[0111] When training a ranking model, it is necessary to identify positive and negative samples. Positive samples represent the module M corresponding to the input ranking training samples. p The candidate module instances satisfy the customer requirements corresponding to the sample; negative samples represent the module M corresponding to the input sorted training samples. pThe module instance candidate of the module does not satisfy the customer demand corresponding to the sample. Based on this, a binary cross-entropy loss function is used to train the ranking model, and the loss function can be represented as:

[0112]

[0113] Wherein, y a 、 respectively represent the real value and the calculated value of the recommended score of the module instance of the module M p corresponding to the a-th ranking training sample. When the module instance of the module M p corresponding to the a-th ranking training sample satisfies the customer demand corresponding to the sample, y a = 1; otherwise, y a = 0. is calculated by the ranking model. N s represents the number of samples in the ranking training set; is the parameter set of the ranking model; and λ2 is a penalty coefficient.

[0114] 6) According to the new customer order data, the demand feature vector of the current customer order data is obtained, and the required modules of the current customer order data are determined, then the instance feature vectors of all module instances corresponding to the module instance set of each module are combined with the demand feature vector of the current customer to form all new data sample features, and then all new data sample features of each module are preprocessed and input into the trained retrieval model of the corresponding module, and the affinity score between all module instances of each module and the current customer is output. The module instance candidate set of each module is selected in the order of affinity score from high to low and recorded as the module instance candidate set of the corresponding module; then the new data sample features corresponding to all module instances in the module instance candidate set of each module are input into the trained ranking model of the corresponding module, and the recommended score of all module instances in the module instance candidate set of each module is output, and the module instance with the highest recommended score in the module instance candidate set of each module is selected as the configuration design scheme of the current customer order data, and the corresponding customized product is manufactured according to the current configuration design scheme.

[0115] The embodiment of the application takes the traction machine module of the customized elevator product as an example, the Top-10 module instance candidate result of the elevator traction machine module based on the retrieval model of the elevator traction machine module is shown in Table 3, and the recommended score calculated by the ranking model of the elevator traction machine module based on the Top-10 module instance candidate screened by the retrieval model is shown in Table 4.

[0116] Table 3: Top-10 module instance candidate result of the traction machine module of the customized elevator product

[0117]

[0118]

[0119] Table 4: Recommendation scores for Top-10 module instance candidates for the hoist module of a custom elevator product

[0120]

[0121]

Claims

1. A method for recommending a customized product configuration design scheme based on a search and ranking model, characterized by, The method comprises the following steps: 1) According to historical customer order data of customized products and a corresponding configuration design bill of materials, all demand features of each customer order data are identified and extracted to form a demand feature vector of the corresponding customer order data; according to the configuration design bill of materials, all module instances corresponding to each customer order data are screened and obtained, and each instance feature of all module instances is extracted to form an instance feature vector of the corresponding module instance; 2) The module classification is performed on all module instances to obtain a module instance set of different modules, each instance feature vector of each module instance of each module is combined with the corresponding demand feature vector to form each data sample feature of the current module, each data sample feature of the current module forms a search data set of the current module, each data sample feature and the corresponding true recommendation score form a sorting data set of the current module, the module instance set of each module is traversed and the data set is constructed to obtain the search data set and the sorting data set of each module; 3) The search data set and the sorting data set of each module are respectively preprocessed to obtain the preprocessed search data set and the preprocessed sorting data set corresponding to each module; 4) The preprocessed search data set of each module is input into the search model corresponding to the module for training, each search model outputs the affinity score between each instance feature vector and the corresponding demand feature vector, and the trained search model corresponding to each module is obtained through training; 5) The preprocessed sorting data set of each module is input into the sorting model corresponding to the module for training, each sorting model outputs the predicted recommendation score between each instance feature vector and the corresponding demand feature vector, and the trained sorting model corresponding to each module is obtained through training; 6) According to new customer order data, the demand feature vector of the current customer order data and the modules required by the current customer order data are obtained, then the instance feature vectors of all module instances in the module instance set corresponding to each module are combined with the demand feature vector of the current customer order data to form all new data sample features, and the new data sample features of each module are preprocessed and input into the trained search model corresponding to the module, the affinity scores between all module instances of each module and the current customer order data are output; a plurality of module instances of each module are selected in the order from high to low according to the affinity scores and recorded as a module instance candidate set of the corresponding module; then the new data sample features of all module instances in the module instance candidate set of each module are input into the trained sorting model corresponding to the module, the recommendation scores of all module instances in the module instance candidate set of each module are output, the module instance with the highest recommendation score in the module instance candidate set of each module is selected as the configuration design scheme of the current customer order data, and the corresponding customized product is manufactured according to the current configuration design scheme. 2.The customized product configuration design scheme recommendation method based on a retrieval and ranking model according to claim 1, wherein, The step 3) includes data supplementing processing of missing data sample features in the search data set and the sorting data set, and one-hot encoding processing of the data sample features of the discrete type. 3.The customized product configuration design scheme recommendation method based on a retrieval and ranking model according to claim 1, characterized in that, The search model structures of the modules are the same, and each search model structure includes a first embedding layer, a query network, a candidate network, and a first output layer. The input of the search model is input to the first embedding layer as the input of the first embedding layer. The demand feature vector of each data sample feature in the search data set is input to the query network after passing through the corresponding first embedding layer. The instance feature vector of each data sample feature in the search data set is input to the candidate network after passing through the corresponding first embedding layer. The output of the query network and the output of the candidate network are input to the first output layer. The first output layer outputs the affinity score between each instance feature vector and the corresponding demand feature vector. 4.The customized product configuration design scheme recommendation method based on a retrieval and ranking model according to claim 3, characterized in that, The query network and the candidate network have the same structure, which is a forward neural network including two hidden layers. 5.The customized product configuration design scheme recommendation method based on a retrieval and ranking model according to claim 1, wherein, The sorting model structures of the modules are the same, and each sorting model structure includes a second embedding layer, an interaction network, and a second output layer. The input of the sorting model is input to the second embedding layer as the input of the second embedding layer. The customer demand feature and the module instance feature of each data sample feature in the sorting data set are input to the corresponding second embedding layer. The interaction network includes an adaptive weight interaction block and an adaptive power order interaction block. The output of each second embedding layer is input to the adaptive weight interaction block and the adaptive power order interaction block, respectively. The output of the adaptive weight interaction block and the output of the adaptive power order interaction block are input to the second output layer. The second output layer outputs the predicted recommendation score between each instance feature vector and the corresponding demand feature vector.

6. The method of claim 5, wherein the method further comprises: The adaptive weight interaction block includes two multi-head self-attention interaction layers. The input of the adaptive weight interaction block is input to the first multi-head self-attention interaction layer. The output of the first multi-head self-attention interaction layer is added to the input of the adaptive weight interaction block and then input to the second multi-head self-attention interaction layer. The output of the second multi-head self-attention interaction layer is the output of the adaptive weight interaction block.

7. The method of claim 5, wherein the method further comprises: The adaptive power order interaction block includes a logarithmic transformation layer and two fully connected hidden layers. The input of the adaptive power order interaction block is input to the logarithmic transformation layer. The output of the logarithmic transformation layer is connected to the first fully connected hidden layer and the second fully connected hidden layer. The output of the second fully connected hidden layer is the output of the adaptive power order interaction block. 8.The customized product configuration design scheme recommendation method based on a retrieval and ranking model of claim 1, wherein, The parameters of the embedding layer of the sorting model of each module are initialized according to the parameters of the embedding layer of the trained search model of the corresponding module.

9. A storage medium storing a computer program, characterized by The computer program is executed by the processor to implement the method of any one of claims 1-8.

10. The storage medium of claim 9, wherein, The computer program is an instruction for implementing the method of any one of claims 1-8.

Images