A comprehensive pipe gallery entity classification method and system
By constructing a data set with hierarchical labels, using Transformer model and fuzzy logic for feature extraction, and building a model structure based on capsule network, the problems of insufficient fine-grained classification of comprehensive pipeline entities in the existing technology are solved, and efficient and accurate hierarchical classification is achieved.
Patent Information
- Application Number
- CN202411942803.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-27
- Publication Date
- 2025-03-14
- Estimated Expiration
- 2044-12-27
AI Technical Summary
The prior art is difficult to capture the deep correlation and hierarchical structure between entities in the comprehensive pipeline entity classification, and the classification is not fine-grained enough to effectively deal with fuzzy and uncertain information.
A comprehensive pipeline entity classification method is proposed. By obtaining standard text data and constructing a data set with hierarchical labels, feature extraction is used using the Transformer model, and fuzzy logic is incorporated into attention weights. Then, a model structure based on capsule network is built, and the feature vector set is trained to achieve efficient, accurate and detailed hierarchical classification of pipeline entities.
It realizes efficient, accurate and detailed hierarchical classification of comprehensive pipeline entities at different levels, improves the accuracy and fine-grainedness of classification, and enhances the model's ability to deal with fuzzy and uncertainty situations.
Smart Images
Figure CN119377408B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of pipe gallery entity classification, and in particular to a comprehensive pipe gallery entity classification method and system. Background Art
[0002] The utility corridor is a key part of urban infrastructure, integrating multiple municipal pipelines such as electricity, communications, gas, water supply and drainage, forming a large and complex network. In order to achieve effective management and maintenance of the utility corridor, it is necessary to accurately classify and identify the various entities inside it (such as pipelines, equipment, structures, etc.).
[0003] At present, the classification of utility corridor entities mainly relies on rule-based methods and traditional machine learning methods. The existing technology has the following defects when dealing with the classification of utility corridor entities:
[0004] Rule-based methods mainly rely on manually formulated rules, such as keyword matching and regular expressions. They are difficult to fully cover all complex entity types and hierarchical relationships, and lack adaptability to newly emerging entity types. Traditional machine learning methods are difficult to capture the deep associations and hierarchical structures between entities. The classification granularity is not fine enough to accurately distinguish the different levels of entities, and they cannot effectively handle fuzzy and uncertain information, making it difficult to achieve efficient, accurate and hierarchical classification of integrated pipeline corridor entities. Summary of the invention
[0005] Based on this, the purpose of the present invention is to propose a comprehensive pipe gallery entity classification method and system to solve the above-mentioned problems.
[0006] According to a method for classifying integrated pipe gallery entities proposed by the present invention, the method comprises:
[0007] Obtain standard text data related to the integrated pipeline corridor, annotate the pipeline corridor entities in the data and their categories at each level, and construct a dataset with hierarchical labels;
[0008] The Transformer model is used to extract features from a dataset with hierarchical labels, and fuzzy logic is integrated into the attention weights to obtain a fuzzified feature vector set.
[0009] Build a model structure based on the capsule network, and train the capsule network model based on the feature vector set to obtain a trained capsule network model;
[0010] The feature vector of the pipeline gallery entity to be classified is input into the trained capsule network model, and the comprehensive pipeline gallery entity is classified by the capsule network model to obtain the classification results of the pipeline gallery entity at different levels.
[0011] Furthermore, the feature extraction of the dataset with hierarchical labels by using the Transformer model includes:
[0012] Preprocessing of standard text data in hierarchically labeled datasets;
[0013] Converting the preprocessed standard text into a high-dimensional vector representation to obtain text embedding, wherein the preprocessed standard text data includes a series of words, and the text embedding is pre-trained to capture semantic information of the entity;
[0014] For each word in the text, generate its standard position code through a combination of sine and cosine functions;
[0015] Generate a code related to the location characteristics of the pipeline corridor for each word in the text to obtain the pipeline corridor location code;
[0016] The standard position code and the pipe gallery position code are combined to obtain a fused position code for each word;
[0017] Inputting the input vector obtained by adding the text embedding and the fused position encoding into the self-attention mechanism, generating a query vector, a key vector and a value vector for each word, and calculating an attention score based on the query vector and the key vector;
[0018] The attention scores are scaled based on the scaled dot product attention mechanism and normalized by the softmax function to obtain the attention weight of each word;
[0019] Split the input vector into multiple heads and calculate the attention weights and outputs of each attention head in parallel;
[0020] The outputs of each attention head are concatenated to obtain the feature vector output.
[0021] Furthermore, the fuzzy logic is integrated into the attention weight to obtain the fuzzified feature vector, including:
[0022] Define fuzzy sets and define membership functions for each fuzzy set. Fuzzy sets are represented by the value range or fuzziness degree of weights. The membership function is used to calculate the degree to which each weight belongs to a fuzzy set.
[0023] For each attention head, input its set of attention weights into the corresponding membership function, calculate its membership to each fuzzy set, and obtain the membership matrix of the corresponding attention head, where each row of the membership matrix corresponds to an attention weight and each column corresponds to a fuzzy set membership;
[0024] The membership information of each row in the membership matrix is fused through the fuzzy aggregation function to obtain a set of fuzzy weights of the corresponding attention head;
[0025] Calculate the output of each attention head based on the fuzzy weight;
[0026] The outputs of each attention head are concatenated to obtain the blurred feature vector.
[0027] Furthermore, the preprocessing of the standard text data in the dataset with hierarchical labels includes:
[0028] Perform data cleaning and word segmentation on standard texts. Data cleaning includes removing irrelevant characters, processing missing values, and data standardization.
[0029] Create a vocabulary to map each unique word in the standard text to an integer ID;
[0030] Replace each word in the text with its corresponding integer ID to form an integer sequence;
[0031] Encode the label corresponding to the text to obtain an encoded integer sequence;
[0032] The step of converting the preprocessed standard text into a high-dimensional vector representation to obtain text embedding includes:
[0033] Convert the encoded integer sequence into an embedding vector to get the text embedding.
[0034] Furthermore, the model structure based on the capsule network includes:
[0035] Designing an input layer and a plurality of capsule layers, wherein the capsule layers include a low-level capsule layer, a middle-level capsule layer and a high-level capsule layer, each of the capsule layers includes a plurality of capsule units, a dynamic routing algorithm is applied between the capsule layers to automatically adjust the connection strength between the capsules, and each capsule unit applies a nonlinear activation function;
[0036] For classification labels at different levels, corresponding output capsule layers are designed, where each output capsule layer contains capsule units corresponding to the number of categories at the corresponding level, and the output of each capsule unit represents the probability of the existence of the corresponding category.
[0037] Furthermore, the training of the capsule network model based on the feature vector set includes:
[0038] The feature vector of the corridor entity is input into the capsule network model, and through multiple levels of the capsule network model, from basic features to complex structures and then to global features, information is gradually extracted and combined to classify and predict labels of different levels of the corridor entity;
[0039] For each level of classification, a loss function suitable for hierarchical classification is used to calculate the loss between the actual label and the predicted label, and the losses of all levels are weighted summed to obtain the total loss value.
[0040] Calculate the gradient based on the loss value and propagate the gradient to each layer of the model through the back-propagation algorithm;
[0041] Continuously iterate and optimize, update the model parameters until the training is completed, and obtain the trained capsule network model.
[0042] Furthermore, through the multiple levels of the capsule network model, from basic features to complex structures and then to global features, information is gradually extracted and combined to classify and predict different levels of labels of the corridor entity, including:
[0043] The capsule units of the low-level capsule layer capture basic and local features, form a preliminary capsule representation, and output a low-level capsule vector;
[0044] The low-level capsule vectors are received through the middle-level capsule layer, and the connection strength with the low-level capsules is adjusted through the dynamic routing algorithm to further combine features to form a more complex structure to aggregate the low-level capsule vectors into the middle-level capsule vectors;
[0045] The high-level capsule layer receives the mid-level capsule vector, further combines the mid-level capsule vector into a global feature, and forms a capsule vector representing the entire entity or attribute through a higher-level routing mechanism, and outputs a global feature capsule vector;
[0046] Through the output capsule layer, the global feature capsule vector is mapped to the hierarchical label space, and classified through softmax to obtain the hierarchical label prediction result.
[0047] Furthermore, the integrated pipe gallery entity is classified by the capsule network model to obtain classification results of the pipe gallery entity at different levels, including:
[0048] Input the feature vector of the pipeline corridor entity to be classified into the trained capsule network model;
[0049] The output of each capsule layer is calculated through forward propagation, and the classification results of the corridor entities at each corresponding level are obtained from each output capsule layer. The classification result of each level is a probability distribution, which is used to represent the existence probability of each category at the corresponding level.
[0050] The hierarchical classification results are displayed in the form of a hierarchical structure diagram.
[0051] The present invention also proposes a comprehensive pipe gallery entity classification system for implementing the above-mentioned comprehensive pipe gallery entity classification method, the system comprising:
[0052] Dataset module: used to obtain standard text data related to the integrated pipeline corridor, annotate the pipeline corridor entities in the data and their categories at each level, and construct a dataset with hierarchical labels;
[0053] Feature extraction module: It is used to extract features from a dataset with hierarchical labels through the Transformer model, and integrate fuzzy logic into the attention weight to obtain a fuzzified feature vector set;
[0054] Model building module: used to build a model structure based on the capsule network, and train the capsule network model based on the feature vector set to obtain a trained capsule network model;
[0055] Classification module: used to input the feature vector of the pipeline corridor entity to be classified into the trained capsule network model, and classify the comprehensive pipeline corridor entity through the capsule network model to obtain the classification results of the pipeline corridor entity at different levels.
[0056] In summary, the integrated pipeline corridor entity classification method of the present invention, by constructing a data set with hierarchical labels, introducing fuzzy logic into the Transformer model, and building a model structure based on a capsule network, realizes efficient, accurate and detailed hierarchical classification of integrated pipeline corridor entities at different levels. This method not only improves the accuracy and granularity of classification, but also enhances the model's ability to handle fuzzy and uncertain situations. Furthermore, through the trained capsule network model, the classification results of pipeline corridor entities at different levels can be accurately and quickly output, providing strong support for the management and maintenance of the integrated pipeline corridor.
[0057] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description or will be learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:
[0059] Figure 1 This is a flow chart of a method for classifying integrated pipe gallery entities according to the first embodiment of the present invention;
[0060] Figure 2 This is a system block diagram of an integrated pipe gallery entity classification system according to the second embodiment of the present invention. DETAILED DESCRIPTION
[0061] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Several embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.
[0062] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.
[0063] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.
[0064] Example 1: Please refer to Figure 1 The present invention proposes a comprehensive pipe gallery entity classification method, which includes steps S101 to S104:
[0065] S101, obtain standard text data related to the integrated pipeline corridor, mark the pipeline corridor entities in the data and their categories at each level, and construct a data set with hierarchical labels.
[0066] Understandably, the collection of standard text data related to the integrated pipe corridor can cover various aspects such as its structure, function, pipeline type, location information, etc. In the standard text data, the entities related to the pipe corridor are identified and marked, which may include specific pipelines (such as drainage pipes, power cables), facilities (such as pumping stations, ventilation equipment), and more abstract elements (such as construction areas, construction progress).
[0067] While labeling entities, hierarchical labels need to be assigned to these entities. These labels should be able to reflect the different levels and categories of entities in the integrated pipeline corridor system, and gradually refine from coarse-grained to fine-grained. For example, coarse-grained labels: "construction achievements", "construction attributes", "construction resources", "construction process", etc. These labels represent the overall classification of different aspects of the integrated pipeline corridor project. Medium-grained labels: Under the coarse-grained labels, more specific categories are further subdivided. "Construction achievements" can be subdivided into "buildings", "structures", "roads", etc.; "construction attributes" can be subdivided into "size attributes", "material attributes", "functional attributes", etc. Fine-grained labels: Under the medium-grained labels, more specific entities or attributes are further subdivided. "Buildings" can be subdivided into "tunnels", "bridges", "underground passages", etc.; "size attributes" can be subdivided into "length", "width", "height", etc. This structure of hierarchical labels helps the model better understand the context and hierarchical relationship of entities when classifying.
[0068] The labeled data is organized into a dataset with hierarchical labels. Each entity is assigned one or more hierarchical labels, reflecting its position and category in the utility corridor system. The dataset should contain rich samples, covering different aspects of the utility corridor and different types of entities, to ensure the generalization ability and accuracy of the model.
[0069] A fine-grained example of hierarchical labeling: the entity is "drainage pipe", the coarse-grained label is "construction results", the medium-grained label is "structure", and the fine-grained label is "drainage facilities". In this example, the drainage pipe is first classified as "construction results", which is a coarse-grained classification, then it is further subdivided into "structures", which is a medium-grained classification, and finally, it is classified as "drainage facilities", which is a fine-grained classification. Even finer-grained classification is also performed. This hierarchical label structure provides detailed information about the location and category of drainage pipes in the integrated pipe corridor system.
[0070] By constructing a dataset with hierarchical labels and assigning labels from coarse-grained to fine-grained to each entity, the subsequent model can learn the hierarchical and dependency relationships between these labels during the training process, thereby more accurately classifying the integrated pipeline corridor entities.
[0071] S102, extract features from the dataset with hierarchical labels through the Transformer model, and integrate fuzzy logic into the attention weights to obtain a fuzzified feature vector set.
[0072] It is understandable that the Transformer model, as a powerful sequence processing model, has a self-attention mechanism that enables the model to capture the relationship between any two positions in the input sequence, thereby capturing the complex dependencies in the standard text data of the integrated pipeline corridor. The Transformer model of this embodiment receives data sets with hierarchical labels as input, which contain labeled pipeline corridor entities and their categories at each level, providing rich contextual information for the model. The model encodes the input data through components such as the self-attention mechanism and extracts useful feature vectors that can capture the semantic information and contextual dependencies of the pipeline corridor entities.
[0073] In practical applications, the Transformer model has achieved remarkable results in the field of natural language processing with its powerful sequence modeling capabilities. However, when processing text data, uncertainty is often encountered. The specific description, attributes or categories of the pipeline entity may not be clearly defined. For example, there is a description in the standard text data: "This pipeline is partly used for drainage and partly for power transmission." There is a certain ambiguity in the definition of the two attributes "partly for drainage" and "partly for power transmission" in this description. Traditional methods may find it difficult to accurately classify this pipeline entity into a specific category. In order to deal with this ambiguity, fuzzy logic is introduced into the Transformer model, and the model incorporates fuzzy logic into the attention weight.
[0074] Specifically, when the model calculates the attention weight of each input position, it not only considers the influence of other positions in the input sequence, but also the fuzziness of these influences. This is achieved by adjusting the attention weights so that the weights can reflect the fuzzy relationship between elements in the input sequence. In this way, the model can generate a set of fuzzified feature vectors. These feature vectors contain not only the semantic information of the original input data, but also additional information about the fuzziness and uncertainty of this information, which is used in subsequent steps to train the capsule network model to achieve the classification of integrated pipeline corridor entities.
[0075] Since these feature vectors contain information about the fuzziness and uncertainty of the input data, they can better reflect the complexity of the standard text data of the integrated pipeline corridor, enabling the model to learn more accurate and robust classification rules during the training process.
[0076] In summary, step S102 achieves effective feature extraction of data sets with hierarchical labels by introducing the Transformer model and fuzzy logic. These feature vectors not only contain the semantic information of the original input data, but also contain additional information about the fuzziness and uncertainty of this information. These feature vectors provide a solid foundation for the training and classification in the subsequent steps.
[0077] S103, building a model structure based on the capsule network, and training the capsule network model based on the feature vector set to obtain a trained capsule network model.
[0078] It is understandable that the capsule network is an advanced neural network architecture. By introducing the new basic unit of capsule, the capsule network can better capture the spatial hierarchical relationship and posture changes in the data. In the present invention, the capsule network is used to process the fuzzified feature vector set, which is generated by step S102 and contains rich information about the standard text data of the integrated pipe gallery.
[0079] The capsule network model structure usually includes multiple layers, each layer consists of multiple capsules. These capsules are able to receive input from the previous layer and pass information to the next layer through a dynamic routing algorithm, which allows the capsule network to learn more refined feature representations and hierarchical structures during training.
[0080] The capsule network model is trained based on the feature vector set, and the training process involves adjusting the model parameters to minimize the loss function, thereby improving the performance of the model in the integrated pipe gallery entity classification task. In the present invention, since hierarchical labels are involved, the loss function may need to consider the hierarchical relationship between the labels to ensure that the model can accurately capture these relationships.
[0081] After sufficient training, a trained capsule network model is obtained, which can receive the feature vectors of the pipeline corridor entities to be classified as input and output the classification results of these entities at different levels. Since the capsule network has powerful feature representation capabilities and dynamic routing mechanisms, the trained model can accurately capture the complex hierarchical relationships and posture changes in the standard text data of the integrated pipeline corridor, which makes the model have high accuracy and robustness in the task of classifying integrated pipeline corridor entities.
[0082] S104, inputting the feature vector of the pipeline corridor entity to be classified into the trained capsule network model, and classifying the comprehensive pipeline corridor entity through the capsule network model to obtain the classification results of the pipeline corridor entity at different levels.
[0083] Understandably, the feature vector of the tunnel entity to be classified is input into the trained capsule network model. The capsule network model has been trained with a large amount of data with hierarchical labels and has learned how to extract and interpret the information in the feature vector. When the feature vector is input into the model, the model uses its internal capsule and dynamic routing mechanism to pass and process the features layer by layer. Each capsule tries to capture specific aspects of the features, such as spatial relationships, posture changes, etc. Through the collaborative work of multiple layers of capsules, the model can gradually build a more refined and accurate feature representation of the tunnel entity.
[0084] The last layer of the capsule network model usually contains one or more classifiers (such as a linear classifier or a softmax layer) to map the extracted feature representations to predefined category labels.
[0085] In the present invention, since hierarchical labels are involved, the classifier may need to consider the hierarchical relationship between labels. For example, it may need to first determine whether an entity belongs to a coarse-grained category (such as "construction achievements"), and then further determine whether it belongs to a fine-grained category under the category (such as "building").
[0086] The classifier outputs a probability or score for each class label, which can then be used to determine the final classification result.
[0087] Based on steps S101 to S104, by constructing a data set with hierarchical labels, introducing fuzzy logic into the Transformer model, and building a model structure based on capsule networks, efficient, accurate and detailed hierarchical classification of integrated pipeline corridor entities at different levels is achieved. This method not only improves the accuracy and granularity of classification, but also enhances the model's ability to handle fuzzy and uncertain situations. In addition, through the trained capsule network model, the classification results of pipeline corridor entities at different levels can be accurately and quickly output, providing strong support for the management and maintenance of integrated pipeline corridors.
[0088] The following is a further detailed introduction to a method for classifying integrated pipe gallery entities according to an embodiment of the present invention:
[0089] Further optionally, in step S102, the feature extraction of the data set with hierarchical labels by using the Transformer model includes:
[0090] Preprocessing of standard text data in hierarchically labeled datasets;
[0091] Converting the preprocessed standard text into a high-dimensional vector representation to obtain text embedding, wherein the preprocessed standard text data includes a series of words, and the text embedding is pre-trained to capture semantic information of the entity;
[0092] For each word in the text, generate its standard position code through a combination of sine and cosine functions;
[0093] Generate a code related to the location characteristics of the pipeline corridor for each word in the text to obtain the pipeline corridor location code;
[0094] The standard position code and the pipe gallery position code are combined to obtain a fused position code for each word;
[0095] Inputting the input vector obtained by adding the text embedding and the fused position encoding into the self-attention mechanism, generating a query vector, a key vector and a value vector for each word, and calculating an attention score based on the query vector and the key vector;
[0096] The attention scores are scaled based on the scaled dot product attention mechanism and normalized by the softmax function to obtain the attention weight of each word;
[0097] Split the input vector into multiple heads and calculate the attention weights and outputs of each attention head in parallel;
[0098] The outputs of each attention head are concatenated to obtain the feature vector output.
[0099] It can be understood that by integrating pre-trained text embedding, position encoding, and self-attention mechanism technologies, this embodiment achieves effective feature extraction of data sets with hierarchical labels. This feature extraction method not only improves the model's ability to understand text data, but also significantly enhances the quality and effect of feature vectors, laying a solid foundation for subsequent natural language processing tasks.
[0100] Further optionally, in step S102, the fuzzy logic is integrated into the attention weight to obtain a fuzzified feature vector, including:
[0101] Define fuzzy sets and define membership functions for each fuzzy set. Fuzzy sets are represented by the value range or fuzziness degree of weights. The membership function is used to calculate the degree to which each weight belongs to a fuzzy set.
[0102] For each attention head, input its set of attention weights into the corresponding membership function, calculate its membership to each fuzzy set, and obtain the membership matrix of the corresponding attention head, where each row of the membership matrix corresponds to an attention weight and each column corresponds to a fuzzy set membership;
[0103] The membership information of each row in the membership matrix is fused through the fuzzy aggregation function to obtain a set of fuzzy weights of the corresponding attention head;
[0104] Calculate the output of each attention head based on the fuzzy weight;
[0105] The outputs of each attention head are concatenated to obtain the blurred feature vector.
[0106] Understandably, suppose we define three fuzzy sets: Low, Medium, and High. These fuzzy sets represent the possible value range or degree of fuzziness of the attention weight. Low: the weight is close to 0, Medium: the weight is in the medium range, and High: the weight is close to 1.
[0107] Define membership functions for each fuzzy set, for example:
[0108] , where μ L (x) is the membership function of the “low” fuzzy set, and x is the weight.
[0109] , where μ M (x) is the membership function of the “middle” fuzzy set, and x is the weight.
[0110] , where μ H (x) is the membership function of the “high” fuzzy set, and x is the weight.
[0111] For a set of attention weights W = [w 1 , w 2 ,…,w n ], input these weights into the corresponding membership function, and calculate the membership of each weight to each fuzzy set. For example, if the weight w 1 =0.4, then the membership degree of the weight to each fuzzy set is: μ L (0.4)=0,μ M (0.4)≈0.67, μ H (0.4) = 0. Perform the same calculation on all weights W to obtain a membership matrix.
[0112] Use a fuzzy aggregation function (such as weighted average) to fuse the membership information of each row in the membership matrix to obtain a set of fuzzy weights, and then use these fuzzy weights to calculate the output of each attention head. Finally, concatenate the outputs of each attention head to obtain the fuzzified feature vector: Feature vector = [output 1 , output 2 , ..., outputm ], where m is the number of attention heads.
[0113] In this way, the embodiment of the present invention can introduce fuzzy logic into the attention mechanism of the Transformer model, making the model more effective and robust in dealing with uncertainty and ambiguity. The fuzzified feature vector not only contains the precise information of the original attention weight, but also incorporates the fuzziness of the weight, thereby increasing the amount of information and expressiveness of the feature vector. This enhanced feature expression ability helps the model to more accurately capture the complex relationships and hierarchical structures between the entities of the integrated pipeline corridor, thereby improving the accuracy of entity classification.
[0114] In addition, fields such as the digitalization of integrated pipe corridor standards usually involve a large number of professional and complex terms and concepts, and traditional entity classification methods often have difficulty in effectively processing data in these fields. After introducing fuzzy logic into the Transformer model, the model can better adapt to the data characteristics of these complex fields, thereby improving the accuracy and efficiency of entity classification.
[0115] Further optionally, the preprocessing of the standard text data in the dataset with hierarchical labels includes:
[0116] Perform data cleaning and word segmentation on standard texts. Data cleaning includes removing irrelevant characters, processing missing values, and data standardization.
[0117] Create a vocabulary to map each unique word in the standard text to an integer ID;
[0118] Replace each word in the text with its corresponding integer ID to form an integer sequence;
[0119] Encode the label corresponding to the text to obtain an encoded integer sequence;
[0120] The step of converting the preprocessed standard text into a high-dimensional vector representation to obtain text embedding includes:
[0121] Convert the encoded integer sequence into an embedding vector to get the text embedding.
[0122] It can be understood that a series of preprocessing is performed on the standard text data in the dataset with hierarchical labels, including data cleaning, word segmentation, vocabulary creation, text integerization and label encoding to ensure the consistency and accuracy of the data; then, the pre-trained word embedding table is used to convert the encoded integer sequence into a high-dimensional vector representation, namely text embedding, so as to capture the semantic and grammatical relationships between words and provide high-quality feature input for subsequent model training and prediction.
[0123] Further optionally, in step S103, building a model structure based on a capsule network includes:
[0124] Design an input layer and multiple capsule layers, wherein the capsule layers include a low-level capsule layer, a middle-level capsule layer, and a high-level capsule layer, each of the capsule layers includes multiple capsule units, a dynamic routing algorithm is applied between the capsule layers to automatically adjust the connection strength between the capsules, and each capsule unit applies a nonlinear activation function, such as a squashing function, and the activation mechanism enables the capsule network to handle more complex classification tasks;
[0125] The capsule layer includes a low-level capsule layer, a middle-level capsule layer and a high-level capsule layer, wherein the low-level capsule layer is used to capture basic, local features to form a preliminary feature representation, the middle-level capsule layer is used to receive the output of the low-level capsule layer, further combine features to form a more complex structure, and the high-level capsule layer is used to receive the output of the middle-level capsule layer to capture the global features of the entire pipeline corridor entity;
[0126] For classification labels at different levels, corresponding output capsule layers are designed, where each output capsule layer contains capsule units corresponding to the number of categories at the corresponding level, and the output of each capsule unit represents the probability of the existence of the corresponding category.
[0127] Further optionally, in step S103, the training of the capsule network model based on the feature vector set includes:
[0128] Initialize the weights and biases of the input layer, capsule layers, and output layer, and set the initial routing coefficients in the dynamic routing algorithm;
[0129] The feature vector of the corridor entity is input into the capsule network model, and through multiple levels of the capsule network model, from basic features to complex structures and then to global features, information is gradually extracted and combined to classify and predict labels of different levels of the corridor entity;
[0130] For each level of classification, a loss function suitable for hierarchical classification is used to calculate the loss between the actual label and the predicted label, and the losses of all levels are weighted summed to obtain the total loss value.
[0131] Calculate the gradient based on the loss value and propagate the gradient to each layer of the model through the back-propagation algorithm;
[0132] Use an optimization algorithm to update the model’s weights and biases based on the gradients;
[0133] Update the routing coefficients in the dynamic routing algorithm;
[0134] The model is iteratively optimized. When the performance of the model on the validation set is satisfactory, the training is stopped and the model parameters are saved to obtain the trained capsule network model.
[0135] Further optionally, the routing coefficient update step in the dynamic routing algorithm can be as follows (taking the middle-level capsule layer as an example, the same applies to the high-level capsule layer):
[0136] For each parent capsule in the middle-level capsule layer and each child capsule in the lower-level capsule layer, initialize the routing coefficients. These coefficients can be initialized to the same value (e.g., all 1 or all 1 / N, where N is the number of child capsules) to ensure that the contribution of each child capsule is equal in the initial iteration;
[0137] For each pair of parent capsule and child capsule, calculate the prediction vector and perform a weighted sum of the prediction vectors from all child capsules according to the current routing coefficient to obtain the input vector of the parent capsule.
[0138] Apply an activation mechanism to the input vector of the parent capsule to produce the output vector of the parent capsule;
[0139] Update the routing coefficient based on the similarity between the output vector of the parent capsule and the prediction vector of the child capsule. This can be achieved by calculating the dot product or cosine similarity between the output vector of the parent capsule and the prediction vector of each child capsule. Then use these similarity scores to update the routing coefficient. The goal of the update is to increase the routing coefficients of those child capsules that are more similar to the output vector of the parent capsule and reduce the routing coefficients of those child capsules that are not similar to the output vector of the parent capsule.
[0140] Iterations are performed until the routing coefficients converge or a predetermined number of iterations is reached. In each iteration, the input vector and output vector of the parent capsule are calculated using the new routing coefficients.
[0141] Further optionally, the multiple levels of the capsule network model are used to gradually extract and combine information from basic features to complex structures and then to global features to classify and predict labels of different levels of the corridor entity, including:
[0142] The capsule units of the low-level capsule layer capture basic and local features, form a preliminary capsule representation, and output a low-level capsule vector;
[0143] The low-level capsule vectors are received through the middle-level capsule layer, and the connection strength with the low-level capsules is adjusted through the dynamic routing algorithm to further combine features to form a more complex structure to aggregate the low-level capsule vectors into the middle-level capsule vectors;
[0144] The high-level capsule layer receives the mid-level capsule vector, further combines the mid-level capsule vector into a global feature, and forms a capsule vector representing the entire entity or attribute through a higher-level routing mechanism, outputting a global feature capsule vector, where each capsule represents a global feature or a higher-level entity category;
[0145] Through the output capsule layer, the global feature capsule vector is mapped to the hierarchical label space, and classified through softmax to obtain the hierarchical label prediction result.
[0146] It is understandable that, taking the entity "drainage pipe" as an example, the above-mentioned Transformer model can extract feature vectors related to "drainage pipe" from standard text data. These vectors contain rich information about "drainage pipe", such as location (underground, ground, etc.), function (drainage, sewage, etc.), material (metal, plastic, etc.) and relationship with other facilities (relationship with cable trench, relationship with water supply pipe, etc.).
[0147] Then, the local features in these feature vectors can be captured by the low-level capsule layer to form capsules representing local features such as "underground facilities", "drainage function", and "metal material". Then, the low-level capsules can be combined by the mid-level capsule layer to form capsules representing more complex features such as "structures", "functional attributes", and "material attributes". For example, "underground facilities" and "drainage function" can be combined into a capsule representing "underground drainage facilities". Then, the mid-level capsules can be further combined by the high-level capsule layer to form a capsule representing the global feature of "construction achievements". This capsule covers all entities and attributes related to "construction achievements" in the text. Then, the classification layer performs fine-grained classification based on the global feature capsule vector, maps the "construction achievements" capsule to the fine-grained label space, and obtains specific fine-grained labels, such as "drainage facilities-underground drainage pipes-metal materials". The predicted fine-grained labels are decoded according to the paths in the label ontology structure to ensure that the predicted labels conform to the hierarchical relationship. For example, if the predicted label is "drainage facilities-underground drainage pipes-metal materials", the decoded label path is "construction results-structures-drainage facilities-underground drainage pipes-metal materials".
[0148] Further optionally, in step S104, the integrated pipe gallery entity is classified by the capsule network model to obtain classification results of the pipe gallery entity at different levels, including:
[0149] Input the feature vector of the pipeline corridor entity to be classified into the trained capsule network model;
[0150] The output of each capsule layer is calculated through forward propagation, and the classification results of the corridor entities at each corresponding level are obtained from each output capsule layer. The classification result of each level is a probability distribution, which is used to represent the existence probability of each category at the corresponding level.
[0151] The hierarchical classification results are displayed in the form of a tree diagram or a hierarchical structure diagram.
[0152] It is understandable that when the feature vector is input into the capsule network model, the model will perform forward propagation calculation. During the forward propagation process, the feature vector will pass through each capsule layer layer by layer, and each capsule layer will further process and transform its input features. The capsules in each capsule layer will try to capture specific aspects of the features, such as spatial relationships, and pass the feature information to the next layer of capsules through a dynamic routing mechanism. Through the collaborative work of multiple layers of capsules, the model can gradually build a more refined and accurate feature representation of the corridor entity. In the fine-grained classification layer of the capsule network model, the extracted feature representation will be mapped to the category label, and each output capsule layer of the model will correspond to a level of classification results, which are expressed in the form of probability distribution, that is, the existence probability of each category. By calculating the probability distribution of each output capsule layer, the classification results of the corridor entity at each corresponding level can be obtained, such as coarse-grained, medium-grained, and fine-grained labels. Finally, the hierarchical classification results can be displayed in the form of a tree diagram or a hierarchical structure diagram to clearly reflect the attribution relationship of the corridor entity at each level and the hierarchical relationship between different levels. In a tree diagram or hierarchical structure diagram, each node represents a category label, the position of the node indicates the level to which it belongs, and the connection relationship between the nodes indicates the hierarchical relationship between categories.
[0153] Example 2: Please refer to Figure 2 The present invention proposes a comprehensive pipe gallery entity classification system, which is used to implement the comprehensive pipe gallery entity classification method in the first embodiment. The system includes:
[0154] Dataset module: used to obtain standard text data related to the integrated pipeline corridor, annotate the pipeline corridor entities in the data and their categories at each level, and construct a dataset with hierarchical labels;
[0155] Feature extraction module: It is used to extract features from a dataset with hierarchical labels through the Transformer model, and integrate fuzzy logic into the attention weight to obtain a fuzzified feature vector set;
[0156] Model building module: used to build a model structure based on the capsule network, and train the capsule network model based on the feature vector set to obtain a trained capsule network model;
[0157] Classification module: used to input the feature vector of the pipeline corridor entity to be classified into the trained capsule network model, and classify the comprehensive pipeline corridor entity through the capsule network model to obtain the classification results of the pipeline corridor entity at different levels.
[0158] Further optionally, the feature extraction module is also used for:
[0159] Preprocessing of standard text data in hierarchically labeled datasets;
[0160] Converting the preprocessed standard text into a high-dimensional vector representation to obtain text embedding, wherein the preprocessed standard text data includes a series of words, and the text embedding is pre-trained to capture semantic information of the entity;
[0161] For each word in the text, generate its standard position code through a combination of sine and cosine functions;
[0162] Generate a code related to the location characteristics of the pipeline corridor for each word in the text to obtain the pipeline corridor location code;
[0163] The standard position code and the pipe gallery position code are combined to obtain a fused position code for each word;
[0164] Inputting the input vector obtained by adding the text embedding and the fused position encoding into the self-attention mechanism, generating a query vector, a key vector and a value vector for each word, and calculating an attention score based on the query vector and the key vector;
[0165] The attention scores are scaled based on the scaled dot product attention mechanism and normalized by the softmax function to obtain the attention weight of each word;
[0166] Split the input vector into multiple heads and calculate the attention weights and outputs of each attention head in parallel;
[0167] The outputs of each attention head are concatenated to obtain the feature vector output.
[0168] Further optionally, the feature extraction module is also used for:
[0169] Define fuzzy sets and define membership functions for each fuzzy set. Fuzzy sets are represented by the value range or fuzziness degree of weights. The membership function is used to calculate the degree to which each weight belongs to a fuzzy set.
[0170] For each attention head, input its set of attention weights into the corresponding membership function, calculate its membership to each fuzzy set, and obtain the membership matrix of the corresponding attention head, where each row of the membership matrix corresponds to an attention weight and each column corresponds to a fuzzy set membership;
[0171] The membership information of each row in the membership matrix is fused through the fuzzy aggregation function to obtain a set of fuzzy weights of the corresponding attention head;
[0172] Calculate the output of each attention head based on the fuzzy weight;
[0173] The outputs of each attention head are concatenated to obtain the blurred feature vector.
[0174] Further optionally, the feature extraction module is also used for:
[0175] Perform data cleaning and word segmentation on standard texts. Data cleaning includes removing irrelevant characters, processing missing values, and data standardization.
[0176] Create a vocabulary to map each unique word in the standard text to an integer ID;
[0177] Replace each word in the text with its corresponding integer ID to form an integer sequence;
[0178] Encode the label corresponding to the text to obtain an encoded integer sequence;
[0179] The step of converting the preprocessed standard text into a high-dimensional vector representation to obtain text embedding includes:
[0180] Convert the encoded integer sequence into an embedding vector to get the text embedding.
[0181] Optionally, the above model building module is also used to:
[0182] Designing an input layer and a plurality of capsule layers, wherein the capsule layers include a low-level capsule layer, a middle-level capsule layer and a high-level capsule layer, each of the capsule layers includes a plurality of capsule units, a dynamic routing algorithm is applied between the capsule layers to automatically adjust the connection strength between the capsules, and each capsule unit applies a nonlinear activation function;
[0183] For classification labels at different levels, corresponding output capsule layers are designed, where each output capsule layer contains capsule units corresponding to the number of categories at the corresponding level, and the output of each capsule unit represents the probability of the existence of the corresponding category.
[0184] Optionally, the above model building module is also used to:
[0185] The feature vector of the corridor entity is input into the capsule network model, and through multiple levels of the capsule network model, from basic features to complex structures and then to global features, information is gradually extracted and combined to classify and predict labels of different levels of the corridor entity;
[0186] For each level of classification, a loss function suitable for hierarchical classification is used to calculate the loss between the actual label and the predicted label, and the losses of all levels are weighted summed to obtain the total loss value.
[0187] Calculate the gradient based on the loss value and propagate the gradient to each layer of the model through the back-propagation algorithm;
[0188] Continuously iterate and optimize, update the model parameters until the training is completed, and obtain the trained capsule network model.
[0189] Optionally, the above model building module is also used to:
[0190] The capsule units of the low-level capsule layer capture basic and local features, form a preliminary capsule representation, and output a low-level capsule vector;
[0191] The low-level capsule vectors are received through the middle-level capsule layer, and the connection strength with the low-level capsules is adjusted through the dynamic routing algorithm to further combine features to form a more complex structure to aggregate the low-level capsule vectors into the middle-level capsule vectors;
[0192] The high-level capsule layer receives the mid-level capsule vector, further combines the mid-level capsule vector into a global feature, and forms a capsule vector representing the entire entity or attribute through a higher-level routing mechanism, and outputs a global feature capsule vector;
[0193] Through the output capsule layer, the global feature capsule vector is mapped to the hierarchical label space, and classified through softmax to obtain the hierarchical label prediction result.
[0194] Further optionally, the classification module is used to:
[0195] Input the feature vector of the pipeline corridor entity to be classified into the trained capsule network model;
[0196] The output of each capsule layer is calculated through forward propagation, and the classification results of the corridor entities at each corresponding level are obtained from each output capsule layer. The classification result of each level is a probability distribution, which is used to represent the existence probability of each category at the corresponding level.
[0197] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.
Claims
1. A comprehensive pipe gallery entity classification method, characterized in that: The method comprises: Obtain standard text data related to the integrated pipeline corridor, annotate the pipeline corridor entities in the data and their categories at each level, and construct a dataset with hierarchical labels; The Transformer model is used to extract features from a dataset with hierarchical labels, and fuzzy logic is integrated into the attention weights to obtain a fuzzified feature vector set. Build a model structure based on the capsule network, and train the capsule network model based on the feature vector set to obtain a trained capsule network model; The feature vector of the pipeline corridor entity to be classified is input into the trained capsule network model, and the comprehensive pipeline corridor entity is classified by the capsule network model to obtain the classification results of the pipeline corridor entity at different levels; The feature extraction of the data set with hierarchical labels by using the Transformer model includes: Preprocessing of standard text data in hierarchically labeled datasets; Converting the preprocessed standard text into a high-dimensional vector representation to obtain text embedding, wherein the preprocessed standard text data includes a series of words, and the text embedding is pre-trained to capture semantic information of the entity; For each word in the text, generate its standard position code through a combination of sine and cosine functions; Generate a code related to the location characteristics of the pipeline corridor for each word in the text to obtain the pipeline corridor location code; The standard position code and the pipe gallery position code are combined to obtain a fused position code for each word; Inputting the input vector obtained by adding the text embedding and the fused position encoding into the self-attention mechanism, generating a query vector, a key vector and a value vector for each word, and calculating an attention score based on the query vector and the key vector; The attention scores are scaled based on the scaled dot product attention mechanism and normalized by the softmax function to obtain the attention weight of each word; Split the input vector into multiple heads and calculate the attention weights and outputs of each attention head in parallel; Concatenate the outputs of each attention head to get the feature vector output; The fuzzy logic is integrated into the attention weight to obtain the fuzzified feature vector, including: Define fuzzy sets and define membership functions for each fuzzy set. Fuzzy sets are represented by the value range or fuzziness degree of weights. The membership function is used to calculate the degree to which each weight belongs to a fuzzy set. For each attention head, input its set of attention weights into the corresponding membership function, calculate its membership to each fuzzy set, and obtain the membership matrix of the corresponding attention head, where each row of the membership matrix corresponds to an attention weight and each column corresponds to a fuzzy set membership; The membership information of each row in the membership matrix is fused through the fuzzy aggregation function to obtain a set of fuzzy weights of the corresponding attention head; Calculate the output of each attention head based on the fuzzy weight; Concatenate the outputs of each attention head to obtain the fuzzy feature vector; The model structure based on the capsule network is constructed, including: Designing an input layer and a plurality of capsule layers, wherein the capsule layers include a low-level capsule layer, a middle-level capsule layer and a high-level capsule layer, each of the capsule layers includes a plurality of capsule units, a dynamic routing algorithm is applied between the capsule layers to automatically adjust the connection strength between the capsules, and each capsule unit applies a nonlinear activation function; For classification labels at different levels, corresponding output capsule layers are designed, where each output capsule layer contains capsule units corresponding to the number of categories at the corresponding level, and the output of each capsule unit represents the probability of the corresponding category existing.
2. The method for classifying integrated pipe gallery entities according to claim 1 is characterized in that: The preprocessing of the standard text data in the dataset with hierarchical labels includes: Perform data cleaning and word segmentation on standard texts. Data cleaning includes removing irrelevant characters, processing missing values, and data standardization. Create a vocabulary to map each unique word in the standard text to an integer ID; Replace each word in the text with its corresponding integer ID to form an integer sequence; Encode the label corresponding to the text to obtain an encoded integer sequence; The step of converting the preprocessed standard text into a high-dimensional vector representation to obtain text embedding includes: Convert the encoded integer sequence into an embedding vector to get the text embedding.
3. The method for classifying integrated pipe gallery entities according to claim 1 is characterized in that: The training of the capsule network model based on the feature vector set includes: The feature vector of the corridor entity is input into the capsule network model, and through multiple levels of the capsule network model, from basic features to complex structures and then to global features, information is gradually extracted and combined to classify and predict labels of different levels of the corridor entity. For each level of classification, a loss function suitable for hierarchical classification is used to calculate the loss between the actual label and the predicted label, and the losses of all levels are weighted summed to obtain the total loss value. Calculate the gradient based on the loss value and propagate the gradient to each layer of the model through the back-propagation algorithm; Continuously iterate and optimize, update the model parameters until the training is completed, and obtain the trained capsule network model.
4. The method for classifying integrated pipe gallery entities according to claim 3 is characterized in that: The capsule network model is used at multiple levels to gradually extract and combine information from basic features to complex structures and then to global features to classify and predict labels at different levels of the corridor entity, including: The capsule units of the low-level capsule layer capture basic and local features, form a preliminary capsule representation, and output a low-level capsule vector; The low-level capsule vectors are received through the middle-level capsule layer, and the connection strength with the low-level capsules is adjusted through the dynamic routing algorithm to further combine features to form a more complex structure to aggregate the low-level capsule vectors into the middle-level capsule vectors; The high-level capsule layer receives the mid-level capsule vector, further combines the mid-level capsule vector into a global feature, and forms a capsule vector representing the entire entity or attribute through a higher-level routing mechanism, and outputs a global feature capsule vector; Through the output capsule layer, the global feature capsule vector is mapped to the hierarchical label space, and classified through softmax to obtain the hierarchical label prediction result.
5. The method for classifying integrated pipe gallery entities according to claim 1 is characterized in that: The integrated pipe gallery entity is classified by the capsule network model to obtain classification results of the pipe gallery entity at different levels, including: Input the feature vector of the pipeline corridor entity to be classified into the trained capsule network model; The output of each capsule layer is calculated through forward propagation, and the classification results of the corridor entities at each corresponding level are obtained from each output capsule layer. The classification result of each level is a probability distribution, which is used to represent the existence probability of each category at the corresponding level.
6. A comprehensive utility corridor entity classification system, used to implement the comprehensive utility corridor entity classification method according to any one of claims 1 to 5, characterized in that: The system comprises: Dataset module: used to obtain standard text data related to the integrated pipeline corridor, annotate the pipeline corridor entities in the data and their categories at each level, and construct a dataset with hierarchical labels; Feature extraction module: It is used to extract features from a dataset with hierarchical labels through the Transformer model, and integrate fuzzy logic into the attention weight to obtain a fuzzified feature vector set; Model building module: used to build a model structure based on the capsule network, and train the capsule network model based on the feature vector set to obtain a trained capsule network model; Classification module: used to input the feature vector of the pipeline corridor entity to be classified into the trained capsule network model, and classify the comprehensive pipeline corridor entity through the capsule network model to obtain the classification results of the pipeline corridor entity at different levels; Wherein, the feature extraction module is also used for: Preprocessing of standard text data in hierarchically labeled datasets; Converting the preprocessed standard text into a high-dimensional vector representation to obtain text embedding, wherein the preprocessed standard text data includes a series of words, and the text embedding is pre-trained to capture semantic information of the entity; For each word in the text, generate its standard position code through a combination of sine and cosine functions; Generate a code related to the location characteristics of the pipeline corridor for each word in the text to obtain the pipeline corridor location code; The standard position code and the pipe gallery position code are combined to obtain a fused position code for each word; Inputting the input vector obtained by adding the text embedding and the fused position encoding into the self-attention mechanism, generating a query vector, a key vector and a value vector for each word, and calculating an attention score based on the query vector and the key vector; The attention scores are scaled based on the scaled dot product attention mechanism and normalized by the softmax function to obtain the attention weight of each word; Split the input vector into multiple heads and calculate the attention weights and outputs of each attention head in parallel; Concatenate the outputs of each attention head to get the feature vector output; The feature extraction module is also used for: Define fuzzy sets and define membership functions for each fuzzy set. Fuzzy sets are represented by the value range or fuzziness degree of weights. The membership function is used to calculate the degree to which each weight belongs to a fuzzy set. For each attention head, input its set of attention weights into the corresponding membership function, calculate its membership to each fuzzy set, and obtain the membership matrix of the corresponding attention head, where each row of the membership matrix corresponds to an attention weight and each column corresponds to a fuzzy set membership; The membership information of each row in the membership matrix is fused through the fuzzy aggregation function to obtain a set of fuzzy weights of the corresponding attention head; Calculate the output of each attention head based on the fuzzy weight; Concatenate the outputs of each attention head to obtain the fuzzy feature vector; The model building module is also used to: Designing an input layer and a plurality of capsule layers, wherein the capsule layers include a low-level capsule layer, a middle-level capsule layer and a high-level capsule layer, each of the capsule layers includes a plurality of capsule units, a dynamic routing algorithm is applied between the capsule layers to automatically adjust the connection strength between the capsules, and each capsule unit applies a nonlinear activation function; For classification labels at different levels, corresponding output capsule layers are designed, where each output capsule layer contains capsule units corresponding to the number of categories at the corresponding level, and the output of each capsule unit represents the probability of the corresponding category existing.
Citation Information
Patent Citations
Antithreatening detection method based on multi-level semantic capsule routing
CN111507101A
Customer service multilevel intention recognition method based on knowledge fusion
CN116303953A