A comprehensive pipe gallery hidden danger risk prediction method based on a knowledge graph

Abstract

The application provides a comprehensive pipe gallery hidden danger risk prediction method based on a knowledge graph, comprising the following steps: S1, acquiring an abnormal monitoring data type sequence, wherein the abnormal monitoring data type sequence comprises a plurality of type labels; S2, inputting the abnormal monitoring data type sequence into an entity recommendation network and a response generation network respectively to predict risk component probability ranking; and S3, performing weighted summation on the risk component probability obtained by the entity generation network and the response generation network to obtain final risk component probability. The application proposes that a comprehensive pipe gallery operation and maintenance risk knowledge graph is applied, a generative model is combined, historical dynamic monitoring data is trained, a risk analysis conclusion is inferred and generated according to input dynamic monitoring data, and a plurality of hidden dangers in a risk ranking are acquired. Through the method, the risk type of the pipe gallery operation and maintenance risk can be quickly and accurately predicted, thereby providing operation and maintenance decision support based on risk information.

Description

Technical Field

[0001] This invention relates to the field of utility tunnel hazard risk prediction technology, and in particular to a comprehensive utility tunnel hazard risk prediction method based on knowledge graphs. Background Technology

[0002] Underground utility tunnels are prone to problems such as pipeline aging, leaks, and tunnel settlement and deformation during operation, leading to serious operational and economic consequences. Identifying and predicting potential risks provides a basis for decision-making in operation and maintenance risk management, effectively reducing the probability of risk occurrence. Building operation and maintenance risk prediction mainly refers to the process of predicting safety hazards in utility tunnels based on dynamic monitoring data and relevant standards. Current methods still have the following problems:

[0003] 1. Traditional integrated utility tunnel risk prediction relies on textual standard documents. The comparison of collected monitoring data with the standards depends on manual work, and the resulting risk conclusions are highly subjective and cannot effectively utilize the dynamic changes in monitoring data.

[0004] 2. The dynamic risk assessment method based on knowledge graphs requires the construction of complex knowledge graphs. The conclusions of risk prediction are generated through fixed template query statements, which cannot make reasonable decisions by utilizing the fluctuations of historical monitoring data over time.

[0005] 3. The knowledge graph of the integrated utility tunnel has a large number of nodes. Using conditional statements to query the risk ranking database is very inefficient, resulting in high time costs and poor real-time prediction. Summary of the Invention

[0006] This invention aims to at least solve the technical problems existing in the prior art, and in particular, it innovatively proposes a method for predicting the hidden risks of integrated utility tunnels based on knowledge graphs.

[0007] To achieve the above-mentioned objectives of this invention, this invention provides a method for predicting potential hazards in integrated utility tunnels based on knowledge graphs, characterized by comprising the following steps:

[0008] S1, Obtain an anomaly monitoring data type sequence, wherein the anomaly monitoring data type sequence includes several type labels;

[0009] S2, input the anomaly monitoring data type sequence into the entity recommendation network and response generation network respectively to predict the probability ranking of risk components;

[0010] S3, the risk component probabilities obtained from the entity generation network and the response generation network are weighted and summed to obtain the final risk component probability. Finally, the risk is determined: the risk component with a risk component probability higher than the set threshold is selected or the risk component with the highest ranking is selected. Finally, the risk type of the integrated utility tunnel corresponding to the risk component is obtained.

[0011] Furthermore, the probability of risky components is obtained through an entity recommendation network, including the following steps:

[0012] S2-1, The graph convolutional network RGCN is applied to encode the entire knowledge graph to obtain the embedded representation of each knowledge graph entity node;

[0013] S2-2, exhaustively match each type label in the monitoring data category sequence with the name of each entity in the knowledge graph to identify K entities in the monitoring data category sequence;

[0014] S2-3, For the embedding of K entities in the monitoring data category sequence, an attention mechanism is used to obtain an attention vector. The attention vector is then multiplied by the corresponding entity embedding matrix in the monitoring data category sequence to obtain a condensed representation of the entities appearing in the monitoring data category sequence.

[0015] S-4, Match the condensed representation of entities appearing in the monitoring data category sequence of step S2-3 with the embedded representation of each knowledge graph entity node in step S2-1 to obtain the probability of the component entity.

[0016] Furthermore, an attention mechanism is employed to obtain the attention vector, which is acquired in the following way:

[0017]

[0018] Where SoftMax() is the SoftMax activation function;

[0019] H represents E The transpose of H E This represents the entity embedding matrix of K entities;

[0020] W K and W q These are learnable parameters;

[0021] Thus, the entity condensation representation C appearing in the monitoring data category sequence is obtained. E :

[0022] C E =αH E .

[0023] Furthermore, the probability of risky components is obtained through the response generation network, including the following steps:

[0024] S201, the Transformer encoder is used to encode the anomaly monitoring data type sequence word by word to obtain the encoder output;

[0025] S202, and then the Transformer decoder is used for decoding: the output of the encoder is concatenated with the condensed representation of the entity generated by the entity recommendation network, and the concatenation result is first passed through two fully connected layers, and then through the softmax function to obtain the probability distribution of the context-level components;

[0026] P res =SoftMax(W v W a [s j :C E ])

[0027] Among them, P res This represents the probability distribution of the entire vocabulary, including the probability distribution of all components.

[0028] W v The word embedding matrix is ​​shared with the encoder;

[0029] W a For a trainable linear projection;

[0030] [s j :C E ] indicates the decoder state, i.e., s j With C E Perform concatenation (splicing); train the network using cross-entropy at each decoder time step.

[0031] s j Indicates the encoder output;

[0032] C E It represents a condensed representation of entities appearing in a risk category sequence.

[0033] Furthermore, it also includes: obtaining the final risk component probability through entity package loss:

[0034] First, for each time step of the decoder in the response generation network, we compute the score r of all M entities in the knowledge graph. j :

[0035] r j =HW b [s j :C E ]+b ent

[0036] Where, r j This represents the score matrix for all M entities at decoding time step j;

[0037] H represents the embedding of all knowledge graph entities generated by RGCN;

[0038] Wb It is a trainable matrix used for dimension alignment and bias;

[0039] b ent Indicates bias;

[0040] Then, it is determined whether the predicted component entity m has a first-order neighbor. If it does, then label 0 exists; otherwise, label 1 exists. The cross-entropy loss function is then used to calculate the difference between the label value and the score value, thereby updating the score r. j ;

[0041] The probability of component entity m being predicted in the response generation network is obtained as follows:

[0042]

[0043] Where L is the response length;

[0044] r jm It is r j The m-th component;

[0045] m is a single entity in the M component entity, and M is the total number of all entities;

[0046] sigmoid() is the sigmoid function;

[0047] Finally, the probability distribution P of the risk components of the response generation network is obtained. all :

[0048] P all =P res +λ3P BOE

[0049] In the formula, λ3 is a hyperparameter.

[0050] Furthermore, after obtaining the encoder's output and before the decoder, word and entity embedding alignment is performed:

[0051] s = W v[E] W c C E +b align

[0052] Among them, W v[E] Is it only W? v The matrix obtained by matching the rows corresponding to the entity tags;

[0053] W c It is a trainable matrix;

[0054] b align It is a bias;

[0055] s represents C E With Wc The similarity score for each word;

[0056] Alignment loss L align Represented as:

[0057] L align =||sq|| 2

[0058] Where q represents the indicator vector;

[0059] ||.|| represents the norm.

[0060] Furthermore, the objective function of the response generator network is:

[0061] L total =L gen +λ1L BOE +λ2L align

[0062] Among them, L gen In response to network losses;

[0063] L BOE Loss due to first-order neighbors;

[0064] L align For alignment loss;

[0065] λ1 and λ2 are two hyperparameters.

[0066] In summary, by adopting the above technical solutions, this invention proposes to apply a comprehensive utility tunnel operation and maintenance risk knowledge graph, combined with a generative model, and train it using historical dynamic monitoring data. Based on the input dynamic monitoring data, it infers and generates risk analysis conclusions, identifying the top-ranked potential risks. This invention's method can quickly and accurately predict the risk types of utility tunnel operation and maintenance, thereby providing risk-information-based operation and maintenance decision support.

[0067] Compared with existing methods, the present invention has the following advantages and improvements:

[0068] 1. The risk prediction based on dynamic data is modeled as a sequence generation process. Generative language models are used for training, which effectively utilizes historical prediction data from previous periods and fully captures the relationship between the changing characteristics of dynamic data and potential risks.

[0069] 2. Using knowledge graph embedding instead of knowledge graph querying overcomes the shortcomings of rigid templates based on database operations and reliance on human experience. By using training data to fit the optimal risk prediction logic, the accuracy of risk prediction is improved.

[0070] 3. Combining knowledge graphs with lightweight model reasoning further improves the processing speed and real-time performance of risk prediction.

[0071] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0072] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0073] Figure 1 This is a schematic diagram of the overall construction process of the present invention.

[0074] Figure 2 This is a framework diagram of the integrated utility tunnel hidden danger risk prediction method based on knowledge graph of the present invention.

[0075] Figure 3 This is an example diagram of the risk component of the present invention.

[0076] Figure 4 This is an example diagram of the risk classification entity of the present invention.

[0077] Figure 5 This is a schematic diagram illustrating the application of the model of the present invention. Detailed Implementation

[0078] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0079] Integrated utility tunnel risk prediction is defined as: inputting the abnormal data categories of dynamic risk monitoring of specified components, and outputting the ranking of components involved in the risks.

[0080] In this invention, the overall process of constructing the integrated utility tunnel risk prediction method based on knowledge graphs is as follows: Figure 1 As shown: First, historical sensor data is collected and organized, evaluated by experts, and labeled to form a risk monitoring dataset. Regarding the integrated utility tunnel operation and maintenance knowledge graph, entity classification and extraction are performed based on relevant laws and regulations, expert knowledge, and historical sensor data, abstracting basic triples (including nodes, attribute data, and relationships). Using the risk monitoring dataset combined with the integrated utility tunnel operation and maintenance knowledge graph, a model is trained, ultimately yielding a knowledge graph-based method for predicting potential risks in integrated utility tunnels.

[0081] A framework for a knowledge graph-based integrated utility tunnel hazard risk prediction method is as follows: Figure 2 As shown, the risk prediction task is defined as: using an entity recommendation network f(X,G) and a response generation network g(X,G,f(X,G)) to predict the probability ranking of risky components. The overall goal of the integrated utility tunnel hazard risk prediction is: based on the risk entity category information of real-time monitoring data and combined with the integrated utility tunnel operation and maintenance knowledge graph, to predict the relevant component entities with potential risks. The specific steps are: representing the input abnormal monitoring data type sequence as a sequence of length n: X = <x1,x2,x3…x n The anomaly detection data type sequences are input into the entity recommendation network and the response generation network respectively to obtain a ranking of risk component probabilities. Then, the risk component probabilities obtained from the entity generation network and the response generation network are weighted and summed to obtain the final risk component probability. The risk component probability obtained from the entity recommendation network is achieved through knowledge graph lookup and matching; the knowledge graph G = <v h ,r,v t > where the entity set is E, and the head entity is the monitoring data category v h ∈E, specifically as follows Figure 3 As shown. The tail entity is the risk component v. t ∈E, specifically as follows Figure 4 As shown, r represents the corresponding set of relations. The probability of risk components obtained from the response generation network is obtained by encoding and decoding through a Transformer encoder and decoder architecture.

[0082] (1) Entity Recommendation Network: First, we apply a Graph Convolutional Network (RGCN) to encode the entire knowledge graph and obtain the embedding of each entity node in the knowledge graph. Next, we exhaustively match each word in the risk category sequence X with the name of each entity in the knowledge graph. In this way, we identify K entities from the risk category sequence and sort them according to their original positions, thus obtaining a D-dimensional entity embedding matrix H for the K entities. E ∈R K ×D .

[0083] Subsequently, we apply an attention mechanism to obtain the attention vector α, which is constructed through two fully connected (FC) layers:

[0084]

[0085] C E =αH E

[0086] Where SoftMax() is the SoftMax activation function;

[0087] H representsE The transpose of H E This represents the entity embedding matrix of K entities;

[0088] In the formula W K and W q These are learnable parameters.

[0089] The resulting C E ∈R D It is a condensed representation of entities appearing in the risk category sequence.

[0090] At this point, C is placed again. E Each word (risk component) in the knowledge graph is exhaustively matched with the name of each entity in the knowledge graph to achieve classification, thereby obtaining the probability P that risk component i is recommended. rec (i), and furthermore, softmax is used to map the probability of risk component i being recommended to 0 to 1:

[0091]

[0092] Among them, e i The entity embedding of risk component i;

[0093] P rec (i) represents the probability that risk component i is recommended;

[0094] C represents E The transpose of C E A condensed representation of entities appearing in a risk category sequence;

[0095] ∝ indicates that it is proportional to, i.e., P rec (i) with It is a multiple of K, where K is a positive integer.

[0096] In addition, the entity recommendation network is trained using cross-entropy loss.

[0097] (2) Response Generation Network: The response generation network predicts risk components based on the risk category sequence. The entire risk category sequence is encoded using a classic Transformer encoder and decoder architecture. The encoder encodes the entire sequence input word by word. At decoding time step j, the output s of the Transformer encoder... j Condensed representation of entities C E The layers are cascaded, and the cascaded result is then passed through two fully connected layers in sequence before being computed using the softmax function.

[0098] Therefore, the component prediction probability distribution corresponding to the response generation network is obtained as follows:

[0099] Pres =SoftMax(W v W a [s j :C E ])

[0100] In the formula, P res This represents the probability distribution of the entire vocabulary, i.e., the overall component vocabulary; where the m-th probability is a component of the word list probability.

[0101] W v The word embedding matrix is ​​shared with the encoder;

[0102] W a It is a trainable linear projection.

[0103] [s j :C E ] indicates the decoder state, i.e., s j With C E Perform concatenation (splicing); train the network using cross-entropy at each decoder time step.

[0104] s j Indicates the encoder output;

[0105] C E It represents a condensed representation of entities appearing in a risk category sequence.

[0106] (3) Bag-of-Entity Loss: While response generation networks trained using cross-entropy at each step can predict risky components, they rarely mention the risky entity concepts associated with the components. Therefore, a Bag-of-Entity (BOE) loss is introduced, which prompts the decoder state to include additional information about the first-order neighbors of the real connections in the associated knowledge graph. Since there are no explicit constraints on the range of component entities predicted in the response generation network, word-level scores are aggregated.

[0107] The loss of physical packets causes the decoder state [s] to be affected. j :C E This includes additional information about the first-order neighbors of the actual recommendations on the knowledge graph. The specific process is as follows:

[0108] First, for each time step of the decoder in the response generation network, we compute the score r of all M entities in the knowledge graph. j r j ∈R M :

[0109] r j =HW b [s j :C E ]+bent

[0110] Where, r j This represents the score matrix for all M entities at decoding time step j;

[0111] H represents the embedding of all knowledge graph entities generated by RGCN;

[0112] W b It is a trainable matrix used for dimension alignment and bias;

[0113] b ent Indicates bias;

[0114] Then, determine if entity m has a first-order neighbor. If it does, then it has label 0; otherwise, it has label 1. Update the score r based on labels 0 and 1. j For example, if the true label is 0 and the predicted value (score) is 0.731, the loss function calculates the difference between this predicted value and 0, thus obtaining a loss value. This loss value indicates the gap between the model's current prediction and the true situation. The model will adjust its parameters based on this loss value to make the next predicted value closer to 0, thereby reducing the loss value. The loss function mentioned is cross-entropy loss.

[0115] Since there are no explicit constraints on which word in the generation should contain information, we summarize the word-level scores into a component-based sigmoid function, representing the probability that entity m is predicted in the response generation network:

[0116]

[0117] Where L is the response length;

[0118] r jm It is r j The m-th component;

[0119] m is a single entity in the M component entity, and M is the total number of all entities;

[0120] Thus, the probability distribution P of the risky components of the response generation network at time step j is obtained. all as follows:

[0121] P all =P res +λ3P BOE

[0122] In the formula, λ3 is a hyperparameter.

[0123] (4) Word and Entity Embedding Alignment: The entity recommendation network and response generation network apply two types of token embeddings. The first type corresponds to words appearing in the risk category sequence, and the second type represents entities appearing in the knowledge graph. To link the RGCN encoding of the knowledge graph with the encoding of the Transformer, an alignment loss strategy is applied here:

[0124] s = W v[E] W c C E +b align

[0125] Among them, W v[E] Is it only W? v The matrix obtained by matching the rows corresponding to the entity tags;

[0126] W c It is a trainable matrix;

[0127] b align It is a bias;

[0128] s represents C E With W c The similarity score for each word;

[0129] Alignment loss L align It is s and an indicator vector q∈{0,1} |E| Mean squared error between: (E represents the entities appearing in the risk category sequence. If there are four E, then s will have four similarity scores, and the dimension of q will also be 4. If the risk category sequence can be matched with an entity in the knowledge graph, then the component is 1; otherwise, it is 0. This means that if the words in the risk category sequence cannot be matched with an entity in the knowledge graph, then the loss value needs to be increased.)

[0130] L align =||sq|| 2

[0131] Specifically, if an entity e exists in the sequence (entity e represents an entity in the knowledge graph that corresponds to a risk category in the sequence), then the component corresponding to q is set to 1. Otherwise, the component is 0. Finally, to learn the parameters of the response generation network, we minimize the following objective function:

[0132] L total =L gen +λ1L BOE +λ2L align

[0133] Among them, L gen In response to network losses;

[0134] LBOE Loss due to first-order neighbors;

[0135] L align For alignment loss;

[0136] λ1 and λ2 are two hyperparameters.

[0137] The schematic diagram of the application of the model of this invention is as follows: Figure 5 As shown. First, input the sample: <Poor ventilation, cracks, abnormal humidity, poor drainage, electromagnetic interference>, then use the knowledge graph-based comprehensive utility tunnel hidden danger risk prediction model; finally, obtain the predicted risk probabilities: drainage system risk 0.9, gas pipeline risk 0.7.

[0138] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for predicting hidden dangers in integrated utility tunnels based on knowledge graphs, characterized in that, Includes the following steps: S1, Obtain an anomaly monitoring data type sequence, wherein the anomaly monitoring data type sequence includes several type labels; S2, input the anomaly monitoring data type sequence into the entity recommendation network and response generation network respectively to predict the probability ranking of risk components; S3, the risk component probabilities obtained from the entity generation network and the response generation network are weighted and summed to obtain the final risk component probability. Risk components with a risk component probability higher than the set threshold or the highest-ranked risk component are selected. Finally, the risk component corresponding to the integrated utility tunnel hidden risk type is obtained. Obtaining the probability of risk components through an entity recommendation network includes the following steps: S2-1, The graph convolutional network RGCN is applied to encode the entire knowledge graph to obtain the embedded representation of each knowledge graph entity node; S2-2, perform an exhaustive match between each type label in the monitoring data category sequence and the name of each entity in the knowledge graph to identify the entities in the monitoring data category sequence. One entity; S2-3, for the monitoring data category sequence The embedding of each entity is achieved by using an attention mechanism to obtain an attention vector. The attention vector is then multiplied by the entity embedding matrix corresponding to the monitoring data category sequence to obtain a condensed representation of the entities appearing in the monitoring data category sequence. S-4, Match the condensed representation of entities appearing in the monitoring data category sequence of step S2-3 with the embedded representation of each knowledge graph entity node in step S2-1 to obtain the probability of the component entity; The probability of risky components is obtained through a response generation network, including the following steps: S201, the Transformer encoder is used to encode the anomaly monitoring data type sequence word by word to obtain the encoder output; S202, and then the Transformer decoder is used for decoding: the output of the encoder is concatenated with the condensed representation of the entity generated by the entity recommendation network, and the concatenation result is first passed through two fully connected layers, and then through the softmax function to obtain the probability distribution of the context-level components; ， in, This represents the probability distribution of the entire vocabulary, including the probability distribution of all components. The word embedding matrix is ​​shared with the encoder; For a trainable linear projection; Indicates the decoder status; Indicates the encoder output; This represents a condensed representation of entities appearing in a risk category sequence.

2. The method for predicting hidden dangers in integrated utility tunnels based on knowledge graphs according to claim 1, characterized in that, Attention vectors are obtained using an attention mechanism, and are acquired in the following way: ， in, for Activation function; express transpose, express Entity embedding matrix of entities; and These are learnable parameters; Thus, a condensed representation of the entities appearing in the monitoring data category sequence is obtained. : 。 3. The method for predicting hidden dangers in integrated utility tunnels based on knowledge graphs according to claim 2, characterized in that, Also includes: The final risk component probability is obtained through the physical package loss: First, for each time step of the decoder in the response generation network, we compute the scores of all M entities in the knowledge graph. : ， in, Indicates the decoding time step The score matrix of all M entities; This represents the embedding of all knowledge graph entities generated by RGCN; It is a trainable matrix used for dimension alignment and bias; Indicates bias; Then, it is determined whether the predicted component entity m has a first-order neighbor. If it does, then label 0 exists; otherwise, label 1 exists. The cross-entropy loss function is then used to calculate the difference between the label value and the score value, thereby updating the score. ; The probability of component entity m being predicted in the response generation network is obtained as follows: ， in, For response length; yes The m-th component; m is a single entity in the M component entity, and M is the total number of all entities; yes function; Finally, the probability distribution of risk components in the response generation network is obtained. : ， In the formula This is a hyperparameter.

4. The method for predicting hidden dangers in integrated utility tunnels based on knowledge graphs according to claim 3, characterized in that, After obtaining the encoder's output, word and entity embedding alignment is performed: ， in, Is it only select? The matrix obtained by matching the rows corresponding to the entity tags; It is a trainable matrix; It is a bias; express and The similarity score for each word; Alignment loss Represented as: ， in, Indicates the indicator vector; Represents the norm.

5. The method for predicting hidden dangers in integrated utility tunnels based on knowledge graphs according to claim 4, characterized in that, The objective function of the response generator network is: ， in, In response to network losses; Loss due to first-order neighbors; For alignment loss; and These are two hyperparameters.

Images