An intelligent customer service system for renting engineering machinery based on a knowledge graph

Abstract

This invention discloses an intelligent customer service system for engineering machinery leasing based on a knowledge graph, involving the fields of knowledge graph, intelligent customer service, and natural language processing. It includes seven modules such as data collection and preprocessing, fusion, and knowledge graph construction. The system collects multi-source data including structured business data, unstructured text, and visual data. After preprocessing, it extracts and fuses triples, then constructs a knowledge graph through logical relationship modeling and GCN iterative optimization. Combined with CNF conjunctive paradigm atomic sentences, it achieves user inquiry intent recognition and Bayesian semantic reasoning, generating intelligent responses or triggering manual transfer. This system solves the pain points of traditional customer service, such as slow response, low accuracy, and high cost, and can efficiently meet diverse user inquiry needs, improving service quality and efficiency.

Description

Technical Field

[0001] This invention relates to the fields of knowledge graphs, intelligent customer service, and natural language processing, and more specifically to an intelligent customer service system for engineering machinery leasing based on knowledge graphs. Background Technology

[0002] Against the backdrop of the rapid development of the construction machinery rental industry, the rental business involves a wide variety of equipment types and complex parameters and specifications. The rental process covers multiple stages such as consultation, application, signing, operation and maintenance, and return. At the same time, relevant industry policies and compliance requirements are constantly changing, resulting in diversified, professional, and frequent user consultation needs.

[0003] Existing customer service models primarily rely on human agents or simple keyword-matching automated response systems, which have significant technical shortcomings. Firstly, human agents need to possess extensive knowledge of equipment, processes, and policies, resulting in long training periods, high labor costs, and limitations imposed by working hours, making 24 / 7 uninterrupted service impossible. Delayed response times outside of working hours severely impact the service experience. Secondly, simple automated response systems lack structured understanding of domain-specific knowledge, failing to accurately grasp the deeper intentions behind users' natural language. This is particularly problematic when dealing with complex issues involving multi-parameter comparisons, process node queries, and policy details, leading to low accuracy and a lack of support for multi-turn, in-depth dialogue, hindering efficient resolution of user needs. Furthermore, as the leasing business expands and user inquiries continue to grow, the response efficiency and service quality bottlenecks of traditional customer service models become increasingly prominent, not only increasing operating costs but also posing a risk of customer churn due to poor service experiences.

[0004] Therefore, there is an urgent need for an intelligent customer service technology solution that can integrate domain knowledge, accurately understand needs, and efficiently respond to inquiries, in order to solve the core problems of slow customer service response, inaccurate answers, and high labor costs in existing technologies. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent customer service system for engineering machinery rental based on knowledge graphs to solve the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A knowledge graph-based intelligent customer service system for construction machinery rental includes: a data acquisition and preprocessing module, a fusion module, a knowledge graph construction module, a natural language understanding module, an intelligent interaction module, and a human transfer module. The data acquisition module collects multi-source data on construction machinery rental, outputs a raw multi-source dataset, preprocesses the raw multi-source data, and outputs a standardized structured dataset, a preprocessed text set, and a visual feature vector set. The fusion module extracts knowledge from the preprocessed standardized structured dataset, preprocessed text set, and visual feature vector set, generates triples, performs fusion verification, and outputs a fused triple set. The knowledge graph construction module initializes the fusion triple set to construct a knowledge graph, and outputs an optimized knowledge graph and atomic sentences in CNF conjunction paradigm through logical relationship modeling and GCN iterative updates. The natural language understanding module combines the atomic sentences to perform intent recognition and semantic reasoning on the user's inquiry text, and outputs intent tags and reasoning results. The intelligent interaction module generates responses and follow-up questions based on intent tags, reasoning results, and the optimized knowledge graph. If the question is resolved, a final response is output; if not, a human transfer signal is triggered. The human transfer module triggers a human transfer based on the human transfer signal and synchronizes relevant information to the human customer service terminal.

[0007] Preferably, the multi-source data includes structured business data, unstructured text data, and visual data; the structured business data includes equipment parameters, leasing process records, policy terms, and order data; the unstructured text data includes user consultation history, fault analysis reports, and industry standard documents; and the visual data includes equipment appearance images and scene environment images.

[0008] Preferably, the preprocessing includes: cleaning and standardizing the structured data based on the Spark framework to output a standardized structured dataset S1; segmenting the unstructured text to obtain a set of atomic sentences T0, determining whether the atomic sentences are simple factual propositions, and refining non-simple factual propositions into simple factual propositions T1 by querying related texts through an inverted index; performing word segmentation and synonym replacement based on a dictionary of engineering machinery on T1 to output a preprocessed text set T2; and using a CNN model to extract features from the visual data to output a visual feature vector set V1.

[0009] Preferably, the fusion module specifically includes: based on the Spark-Neo4j cluster, setting table names, attributes, and foreign key aliases for the relation tables in the standardized structured dataset S1, automatically generating triples in the form of entity, attribute, and value, and triples in the form of entity, relation, and entity, with the two types of triples jointly forming the cluster triple S2; using the BiLSTM CRF model to identify named entities in the preprocessed text set T2, extracting entity relations through the Text CNN model, and generating entity attribute triples T3 based on the joint dependency parsing of the named entities and entity relations; establishing a mapping relationship between the visual feature vector set V1 and device entities, generating a visual association triple V2 in the form of device entity, visual feature, and applicable scenario; comparing the cluster triple S2, entity attribute triple T3, and visual association triple V2 with the temporary graph database, unifying entity representations through synonym replacement, and outputting the fusion triple set K1.

[0010] Preferably, the knowledge graph construction module specifically includes: The structured logical feature generation unit decomposes the fused triple set into logical relational sentences containing entity fields, attribute fields, and value fields; it creates an entity field table ST, an attribute field table SX, a value field table Z, and a conclusion table Y; it performs synonym replacement and deduplication on the field values ​​to generate mapping relationships, using the following formula: Mapping relationships serve as a structural logical feature; The graph optimization unit constructs a GCN input matrix, performs inter-layer propagation of GCN based on the GCN input matrix to update node features, calculates the loss value of fusion logic constraints based on the propagation output, optimizes GCN parameters based on the loss value, iterates until the loss converges, and outputs the optimized knowledge graph and atomic sentences; wherein, the atomic sentences of CNF conjunctive normal form are generated by transforming the logical relational sentence after deduplication of the field table.

[0011] Preferably, the natural language understanding module specifically includes: The intent reasoning unit converts user inquiry text into a sequence of word vectors. And perform feature extraction: , The gating weight matrix of the LSTM model. This is the cell state weight matrix of the LSTM model. For the gating bias term of the LSTM model, This is the cell state bias term in the LSTM model. For the sigmoid function, For element-wise multiplication, This is the hidden layer state vector of the LSTM model when it processes up to the t-th word, t=1,2,...,N. The preorder hidden layer state vector of the LSTM model when processing up to the (t-1)th word. To initialize the zero vector, The final hidden layer feature vector after processing all words in the LSTM model represents the global feature representation of the user's consultation text; a fully connected layer mapping is then performed. , Given a user's query text X, output the conditional probability distribution of each intent category as a probability vector, where the vector dimension is equal to the number of intent categories. For the weight matrix of the fully connected layer intended for classification, To classify the bias terms of the fully connected layer, take The category with the highest probability value is used as the intent label; The semantic reasoning unit uses CNF conjunctive paradigm atomic sentences output by the knowledge graph construction module as the basis for reasoning, constructs a Bayesian probabilistic network, and clarifies the conditional probability relationships of each entity, attribute, and value in the atomic sentences. The intent tags output by the intent recognition sub-step and the core features in the user's consultation text are used as evidence nodes in the Bayesian network, and the posterior probability of the reasoning conclusion is calculated by substituting them into the Bayesian reasoning formula. The posterior probability is used as the reasoning confidence, and the semantic reasoning result with confidence is output.

[0012] Preferably, the intelligent interaction module specifically includes generating an initial response based on intent tags, inference results, and knowledge graph node relationships, using a combination of template matching and semantic generation; generating targeted follow-up questions based on knowledge graph entity attribute relationships for unclear user needs; and setting a confidence threshold. The formula for triggering manual transfer is: Where P represents the confidence level of the inference result. =1 indicates that manual transfer is triggered.

[0013] As can be seen from the above technical solution, compared with the prior art, the present invention discloses an intelligent customer service system for construction machinery rental based on knowledge graph. By integrating technologies such as multi-source data preprocessing, multi-model fusion knowledge extraction, GCN-optimized knowledge graph construction, and Bayesian semantic reasoning, it solves the problems of high labor costs, untimely response, inaccurate answers, and lack of deep interaction in traditional construction machinery rental customer service. It achieves uninterrupted service, can accurately understand the diverse and professional consultation needs of users through structured knowledge graph, output high-confidence responses and targeted follow-up questions, can ensure the efficiency of solving complex problems by synchronizing key information through manual transfer, and can maintain the timeliness of knowledge through incremental updates of the graph, which greatly reduces the operating costs of enterprises and significantly improves the user service experience and consultation resolution rate. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the structure provided by the present invention; Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] This invention discloses an intelligent customer service system for engineering machinery leasing based on knowledge graphs, such as... Figure 1 As shown, the system includes: a data acquisition and preprocessing module, a fusion module, a knowledge graph construction module, a natural language understanding module, an intelligent interaction module, and a human referral module. The data acquisition module collects multi-source data on construction machinery rental, outputting a raw multi-source dataset. It then preprocesses the raw multi-source data, outputting a standardized structured dataset, a preprocessed text set, and a visual feature vector set. The fusion module extracts knowledge from the preprocessed standardized structured dataset, preprocessed text set, and visual feature vector set, generating triples and performing fusion verification, outputting a fused triple set. The knowledge graph construction module initializes the fused triple set to construct a knowledge graph, and through logical relationship modeling and GCN iterative updates, outputs an optimized knowledge graph and atomic sentences in CNF conjunction paradigm. The natural language understanding module combines atomic sentences to perform intent recognition and semantic reasoning on user inquiry text, outputting intent tags and reasoning results. The intelligent interaction module generates responses and follow-up questions based on intent tags, reasoning results, and the optimized knowledge graph. If the question is resolved, a final response is output; otherwise, a human referral signal is triggered. The human referral module triggers a human referral based on the human referral signal, synchronizing relevant information to the human customer service terminal.

[0018] In one specific embodiment, the multi-source data includes structured business data, unstructured text data, and visual data; the structured business data includes equipment parameters, leasing process records, policy terms, and order data; the unstructured text data includes user consultation history, fault analysis reports, and industry standard documents; and the visual data includes equipment appearance images and scene environment images.

[0019] In one specific embodiment, preprocessing includes: cleaning and standardizing structured data based on the Spark framework to output a standardized structured dataset S1; segmenting unstructured text to obtain a set of atomic sentences T0, determining whether the atomic sentences are simple factual propositions, and refining non-simple factual propositions into simple factual propositions T1 by querying related texts through an inverted index; performing word segmentation and synonym replacement based on a dictionary of engineering machinery on T1 to output a preprocessed text set T2; and using a CNN model to extract features from visual data to output a visual feature vector set V1.

[0020] The specific structure of the CNN model is as follows: The input layer receives visual data related to engineering machinery and converts it into a tensor format that the model can process. The input dimension is H×W×3, where H is the image height, W is the image width, and 3 represents the RGB three channels. The first convolutional layer uses a 5×5 kernel, with 64 output channels, a stride of 1, and SamePadding for padding. This initially extracts low-dimensional features from the image, and the feature representation capability is enhanced by expanding the number of channels from 3 to 64. After the convolutional layer, the first activation layer is connected, using the ReLU activation function, with the function expression f(x)=max(0,x). This introduces a non-linear mapping to strengthen effective features and suppress ineffective features. After the activation layer, the first pooling layer is connected, using average pooling with a 2×2 kernel size, a stride of 2, and ValidPadding for padding. This achieves the first dimensionality reduction, preserves the global feature distribution, reduces computational complexity, and removes redundant information. The second convolutional layer uses a 3×3 kernel, increasing the number of output channels to 128, with a stride of 1 and SamePadding padding, to further extract mid-dimensional features from the image. It is then connected to a second activation layer, also using the ReLU activation function. Following the activation layer is a second pooling layer using max pooling with a 2×2 kernel, a stride of 2, and ValidPadding padding, to preserve prominent features and enhance key structural information. The third convolutional layer reuses a 5×5 kernel, maintaining the same 128 output channels, with a stride of 1 and SamePadding padding, further refining feature extraction accuracy. It is connected to the third activation layer and then to a third pooling layer using max pooling with a 2×2 kernel, a stride of 2, and ValidPadding padding, to further compress feature dimensions and focus on core feature information. The fourth convolutional layer uses a 3×3 kernel, increasing the number of output channels to 256, with a stride of 1 and SamePadding, to extract high-dimensional semantic features. It is then connected to the fourth activation layer without pooling. The fifth convolutional layer uses a 1×1 kernel, with 1 output channel, a stride of 1, and SamePadding, to integrate and compress feature dimensions, outputting core visual features directly associated with device entities. Connected to the fifth activation layer, this completes the entire feature extraction process, ultimately outputting a visual feature vector set V1.

[0021] In one specific embodiment, the fusion module includes: based on a Spark-Neo4j cluster, setting table names, attributes, and foreign key aliases for the relation tables in the standardized structured dataset S1, automatically generating triples in the form of entity, attribute, and value, and triples in the form of entity, relation, and entity. These two types of triples together form the cluster triple S2. These two types of triples are complementary knowledge representations, fully covering the two core knowledge categories in the engineering machinery leasing field, adapting to the consultation scenario requirements of intelligent customer service, and representing a universally optimal solution for structured storage and reasoning of knowledge graphs. The entity, attribute, and value triple focuses on the characteristics of the entity itself, addressing user needs for attribute queries such as equipment parameters and business rules, and is the basic building block of knowledge; the entity, relation, and entity triple focuses on the relationships between different entities, addressing user needs for relationship-related consultations such as processes and relational logic, and is the core of knowledge association.

[0022] The BiLSTM CRF model is used to identify named entities in the preprocessed text set T2. Entity relations are extracted using the Text CNN model. Entity attribute triples T3 are generated based on joint dependency parsing of named entities and entity relations. A mapping relationship between the visual feature vector set V1 and device entities is established to generate a visual association triple V2 of device entity, visual feature, and applicable scenario form. The cluster triple S2, entity attribute triple T3, and visual association triple V2 are compared with a temporary graph database. The entity representation is unified by synonym replacement, and the fusion triple set K1 is output.

[0023] The BiLSTM CRF model employs a three-stage cascaded architecture consisting of an embedding layer, a BiLSTM hidden layer, and a CRF output layer. Embedded layer word vector transformation, assuming the input text sequence is... N is the text length. For the i-th word (named entity), the word vector matrix is: Let V be the size of the domain vocabulary and d be the dimension of the word vector. Then the word vector of the i-th word is: And if If domain synonyms exist, they are first replaced with standard terms before vector mapping.

[0024] BiLSTM Hidden Layer

[0025] Forward LSTM state update Reverse LSTM state update ;in , These are the weight parameter matrices for the forward and backward LSTMs, respectively. , These represent the hidden states at the previous positive position and the next negative position, respectively. Bidirectional features are then concatenated. ,in These are the hidden state features after splicing.

[0026] Hidden layer to label emission probability mapping ; in, Let be the weight parameter matrix from the BiLSTM hidden layer to the tag emission probability. For bias terms, Let be the emission probability score for each tag corresponding to the i-th word.

[0027] CRF output layer global optimal path calculation, assuming the label sequence is... Introducing start and end virtual labels, the label transition probability matrix is: , The label sequence represents the transition probability from label a to label b. The total score of the label sequence is: ; For the i-th word The label predicted by the BiLSTM model The probability score, It is a word Corresponding entity tags; Find the globally optimal label sequence using the Viterbi algorithm: ; in This is the final entity annotation sequence, enabling accurate extraction of domain entities.

[0028] The Text CNN model employs an architecture consisting of embedding layers, multi-scale convolutional layers, pooling layers, dropout layers, and fully connected layers; in the embedding layer, vocabulary is set... The word vectors are The position vector is Based on the distance calculation from words to entity pairs, the enhanced embedding vector is: ;in, The augmentation embedding matrix for the text sequence is given, where N is the text length. The multi-scale convolutional layer sets up filter windows of three sizes, and the convolution operation is as follows: ;in Let m be the weight matrix of the m-th convolutional filter. For independent bias, for The window submatrix, For convolutional output, each filter generates N-T+1 output values, forming a feature sequence. The pooling layer, dropout layer, and fully connected layer outputs the data. In the pooling layer, the optimal value is obtained for the feature sequence of each filter. The pooled features from all filters are concatenated into a total feature vector. ; The dropout layer introduces a dropout rate. , For discard rate; Fully connected layer mapping and relational probability output: ;in, The core weight matrix is ​​used for mapping relationships in the fully connected layer. As an independent bias term, the relation type with the highest probability is ultimately output as the entity relation.

[0029] In one specific embodiment, the knowledge graph construction module specifically includes: The structured logical feature generation unit decomposes the fused triple set into logical relational sentences containing entity fields, attribute fields, and value fields; it creates entity field table ST, attribute field table SX, value field table Z, and conclusion table Y, performs synonym replacement and deduplication on field values, and generates mapping relationships, as shown in the formula: Mapping relationships serve as a structural logical feature; The map optimization unit includes: Initialize the basic knowledge graph, defining G=(V,E), where V is the node set and E is the edge set. Construct the GCN input matrix based on the field table and mapping relationships, and then construct the attribute matrix. , Let D be the total number of entities, and D be the feature dimension. The formula is D = D1 + D2 + D3, where D1 is the basic entity feature dimension, D2 is the attribute-value mapping feature dimension, and D3 is the conclusion logic feature dimension. The attribute-value mapping feature is derived from the value field table, and the formula is... OneHotEncode(·) is the one-hot encoding function for the attribute, and Weight(·) is the value weight mapping; the logical features of the conclusion are derived from the conclusion table, and the formula is: Embedding(·) is the conclusion embedding vector. The conclusion weight matrix; the GCN input matrix. , For the basic features of an entity; Construct an adjacency matrix based on the relationships between entities in the entity field table and the logical consistency constraints of the conclusion table, using the following formula: , This represents the connection relationship between nodes i and j, and Consist(·) is the logical consistency function for the conclusion. GCN Interlayer Propagation Formula ;in, I is the identity matrix. for The degree matrix, Let l be the weight matrix of the l-th layer. It is the ReLU activation function. ; Calculate the cross-entropy loss function: ; in, For the number of node categories, For the true class probability, Output feature values ​​for GCN; minimize the loss function using stochastic gradient descent, iterate training until the loss converges, and output the optimized knowledge graph and atomic sentences. The atomic sentences of the CNF conjunctive paradigm are generated by transforming the logical relational sentences after deduplication of the field table.

[0030] In one specific embodiment, the natural language understanding module specifically includes: The intent reasoning unit converts user inquiry text into a sequence of word vectors. And perform feature extraction: , The gating weight matrix of the LSTM model. This is the cell state weight matrix of the LSTM model. For the gating bias term of the LSTM model, This is the cell state bias term in the LSTM model. For the sigmoid function, For element-wise multiplication, This is the hidden layer state vector of the LSTM model when it processes up to the t-th word, t=1,2,...,N. The preorder hidden layer state vector of the LSTM model when processing up to the (t-1)th word. To initialize the zero vector, The final hidden layer feature vector after processing all words in the LSTM model represents the global feature representation of the user's consultation text; a fully connected layer mapping is then performed. , Given a user's query text X, output the conditional probability distribution of each intent category as a probability vector, where the vector dimension is equal to the number of intent categories. For the weight matrix of the fully connected layer intended for classification, To classify the bias terms of the fully connected layer, take The category with the highest probability value is used as the intent label; The semantic reasoning unit uses CNF conjunctive paradigm atomic sentences output by the knowledge graph construction module as the basis for reasoning, constructs a Bayesian probabilistic network, and clarifies the conditional probability relationships of each entity, attribute, and value in the atomic sentences. The intent tags output by the intent recognition sub-step and the core features in the user's consultation text are used as evidence nodes in the Bayesian network, and the posterior probability of the reasoning conclusion is calculated by substituting them into the Bayesian inference formula. The posterior probability is used as the reasoning confidence, and the semantic reasoning result with confidence is output.

[0031] A Bayesian probabilistic network is constructed based on atomic sentences in the CNF conjunction paradigm, and the inference formula is as follows: ; in, For the conclusion of the inference, For association features, The prior probability of the conclusion; the posterior probability calculated using the Bayesian inference formula. As a measure of confidence in reasoning, it is related to the conclusion of the reasoning. f Together they form the semantic reasoning result output.

[0032] In one specific embodiment, the intelligent interaction module includes generating an initial response based on intent tags, inference results, and knowledge graph node relationships, using a combination of template matching and semantic generation; generating targeted follow-up questions based on knowledge graph entity attribute relationships for unclear user needs; and setting a confidence threshold. The formula for triggering manual transfer is: Where P represents the confidence level of the inference result. =1 indicates that manual transfer is triggered.

[0033] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The methods disclosed in the embodiments are described simply because they correspond to the methods disclosed in the embodiments; relevant parts can be found in the method section.

[0034] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A knowledge graph-based intelligent customer service system for engineering machinery leasing, characterized in that, include: The system includes a data acquisition and preprocessing module, a fusion module, a knowledge graph construction module, a natural language understanding module, an intelligent interaction module, and a manual transfer module. The data acquisition module collects multi-source data on engineering machinery rental, outputs raw multi-source datasets, preprocesses the raw multi-source data, and outputs standardized structured datasets, preprocessed text sets, and visual feature vector sets. The fusion module extracts knowledge from the preprocessed standardized structured dataset, preprocessed text set, and visual feature vector set, generates triples, and performs fusion verification to output a fused triple set. The knowledge graph construction module initializes the fused triple set to construct a knowledge graph, and outputs an optimized knowledge graph and atomic sentences of CNF conjunction paradigm through logical relationship modeling and GCN iterative updates. The natural language understanding module combines the atomic sentences to perform intent recognition and semantic reasoning on the user's inquiry text, and outputs intent tags and reasoning results; the intelligent interaction module generates replies and follow-up questions based on intent tags, reasoning results and optimized knowledge graphs. If the problem is solved, the final reply is output; if not, a manual transfer signal is triggered. The manual transfer module triggers a manual transfer based on a manual transfer signal and synchronizes relevant information to the human customer service terminal.

2. The intelligent customer service system for engineering machinery leasing based on knowledge graphs according to claim 1, characterized in that, The multi-source data includes structured business data, unstructured text data, and visual data; the structured business data includes equipment parameters, leasing process records, policy terms, and order data; the unstructured text data includes user consultation history, fault analysis reports, and industry standard documents; and the visual data includes equipment appearance images and scene environment images.

3. The intelligent customer service system for engineering machinery leasing based on knowledge graphs according to claim 1, characterized in that, The preprocessing includes: cleaning and standardizing structured data based on the Spark framework to output a standardized structured dataset S1; segmenting the unstructured text to obtain a set of atomic sentences T0, determining whether the atomic sentences are simple factual propositions, and refining non-simple factual propositions into simple factual propositions T1 by querying related texts through an inverted index; performing word segmentation and synonym replacement based on a dictionary of engineering machinery on T1 to output a preprocessed text set T2; and using a CNN model to extract features from the visual data to output a visual feature vector set V1.

4. The intelligent customer service system for engineering machinery leasing based on knowledge graphs according to claim 1, characterized in that, The fusion module specifically includes: based on the Spark-Neo4j cluster, setting table names, attributes, and foreign key aliases for relation tables in the standardized structured dataset S1, automatically generating triples in the form of entity, attribute, and value, and triples in the form of entity, relation, and entity, with the two types of triples forming the cluster triple S2; using the BiLSTM CRF model to identify named entities in the preprocessed text set T2, extracting entity relations through the Text CNN model, and generating entity attribute triples T3 based on the joint dependency parsing of the named entities and entity relations; establishing a mapping relationship between the visual feature vector set V1 and device entities, generating a visual association triple V2 in the form of device entity, visual feature, and applicable scenario; comparing the cluster triple S2, entity attribute triple T3, and visual association triple V2 with the temporary graph database, unifying entity representations through synonym replacement, and outputting the fusion triple set K1.

5. The intelligent customer service system for engineering machinery leasing based on knowledge graphs according to claim 1, characterized in that, The knowledge graph construction module specifically includes: The structured logical feature generation unit decomposes the fused triple set into logical relational sentences containing entity fields, attribute fields, and value fields; it creates an entity field table ST, an attribute field table SX, a value field table Z, and a conclusion table Y; it performs synonym replacement and deduplication on the field values ​​to generate mapping relationships, using the following formula: Mapping relationships serve as a structural logical feature; The graph optimization unit constructs a GCN input matrix, performs inter-layer propagation of GCN based on the GCN input matrix to update node features, calculates the loss value of fusion logic constraints based on the propagation output, optimizes GCN parameters based on the loss value, iterates until the loss converges, and outputs the optimized knowledge graph and atomic sentences; wherein, the atomic sentences of CNF conjunctive normal form are generated by transforming the logical relational sentence after deduplication of the field table.

6. The intelligent customer service system for engineering machinery leasing based on knowledge graphs according to claim 1, characterized in that, The natural language understanding module specifically includes: The intent reasoning unit converts user inquiry text into a sequence of word vectors. And perform feature extraction: , The gating weight matrix of the LSTM model. This is the cell state weight matrix of the LSTM model. For the gating bias term of the LSTM model, This is the cell state bias term in the LSTM model. For the sigmoid function, For element-wise multiplication, This is the hidden layer state vector of the LSTM model when it processes up to the t-th word, t=1,2,...,N. The preorder hidden layer state vector of the LSTM model when processing up to the (t-1)th word. To initialize the zero vector, The final hidden layer feature vector after processing all words in the LSTM model represents the global feature representation of the user's consultation text; a fully connected layer mapping is then performed. , Given a user's query text X, output the conditional probability distribution of each intent category as a probability vector, where the vector dimension is equal to the number of intent categories. For the weight matrix of the fully connected layer intended for classification, To classify the bias terms of the fully connected layer, take The category with the highest probability value is used as the intent label; The semantic reasoning unit uses CNF conjunctive paradigm atomic sentences output by the knowledge graph construction module as the basis for reasoning, constructs a Bayesian probabilistic network, and clarifies the conditional probability relationships of each entity, attribute, and value in the atomic sentences. The intent tags output by the intent recognition sub-step and the core features in the user's consultation text are used as evidence nodes in the Bayesian network, and the posterior probability of the reasoning conclusion is calculated by substituting them into the Bayesian reasoning formula. The posterior probability is used as the reasoning confidence, and the semantic reasoning result with confidence is output.

7. The intelligent customer service system for engineering machinery leasing based on knowledge graphs according to claim 1, characterized in that, The intelligent interaction module specifically includes generating an initial response based on intent tags, inference results, and knowledge graph node relationships, using a combination of template matching and semantic generation; generating targeted follow-up questions based on knowledge graph entity attribute relationships for unclear user needs; and setting confidence thresholds. The formula for triggering manual transfer is: Where P represents the confidence level of the inference result. =1 indicates that manual transfer is triggered.

Images