A tunnel unmanned intelligent monitoring method and control system
Through neural network technology, the template exception information and implicit representation of tunnel sections in the tunnel monitoring network are processed, and the target tunnel monitoring network is generated, which solves the problems of tunnel monitoring data integration and feature extraction, improves monitoring efficiency and accuracy, and reduces the false alarm rate .
Patent Information
- Application Number
- CN202411266906.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-11
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-09-11
AI Technical Summary
The existing tunnel monitoring technology has problems such as low monitoring efficiency, slow response speed and high false alarm rate. It is difficult to effectively integrate tunnel monitoring data and extract valuable features, resulting in insufficient intelligent monitoring level.
Using neural network technology, the initial tunnel monitoring network is debugged and the target tunnel monitoring network is generated to identify abnormal warning information by obtaining and processing the implicit representation of template exception information in the tunnel monitoring network.
It realizes efficient integration of tunnel monitoring data, accurately identify abnormal situations, improves the intelligence level and response speed of the monitoring system, and reduces the false alarm rate.
Smart Images

Figure CN118821024B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of data processing technology, and in particular to an unmanned intelligent monitoring method and control system for tunnels. Background Art
[0002] With the acceleration of urbanization, tunnels, as important transportation infrastructure, are receiving more and more attention for their safety and stability. Traditional tunnel monitoring systems mainly rely on manual inspections and video monitoring, which have problems such as low monitoring efficiency, slow response speed, and high false alarm rate. In order to overcome these defects and realize unmanned and intelligent monitoring of tunnels, it is particularly important to develop an efficient and accurate tunnel monitoring technology. At present, the field of tunnel monitoring has begun to explore the use of advanced technical means such as machine learning and deep learning to improve the intelligence level of monitoring systems. However, existing technologies still face many challenges in practical applications. Tunnel monitoring data is diverse and complex, and how to effectively integrate this information and extract valuable features is a difficult problem. In order to meet these challenges, researchers have proposed a variety of optimization schemes for tunnel monitoring systems. Some schemes improve the accuracy of feature extraction and classification by introducing complex machine learning models, but these models often have high computational complexity and are difficult to maintain real-time and efficiency in actual deployment. Other schemes focus on building a multi-source data fusion system, trying to improve the comprehensiveness and accuracy of monitoring by integrating different types of data, but there are still technical bottlenecks in data fusion and feature matching. Therefore, there is an urgent need for an intelligent monitoring technology that can efficiently integrate tunnel monitoring data and accurately identify abnormal situations. Summary of the invention
[0003] In view of this, the present application provides an unmanned intelligent monitoring method and control system for tunnels. The technical solution of the present application is implemented as follows:
[0004] On the one hand, an embodiment of the present application provides an unmanned intelligent monitoring method for a tunnel, comprising: obtaining p first target relationship chains and q second target relationship chains, where p and q are both natural numbers not equal to 0; extracting the implicit link representation of each of the first target relationship chains and the implicit link representation of each of the second target relationship chains; using the implicit link representations of the p first target relationship chains and the implicit link representations of the q second target relationship chains as execution data of a target tunnel monitoring network, obtaining an implicit representation of target abnormal information and an implicit representation of a target tunnel section, wherein the target tunnel monitoring network is an initial tunnel monitoring network that uses the template implicit representation of abnormal information and the template implicit representation of tunnel section as a debugging template. The template abnormality information is implicitly represented as a neural network obtained by debugging a network, the template abnormality information is implicitly represented as a link implicit representation of q second template relationship chains based on the first branch network in the initial tunnel monitoring network, the template tunnel section is implicitly represented as a link implicit representation of the template abnormality information or a first enhanced implicit representation based on the second branch network in the initial tunnel monitoring network, and p first template relationship chains, the first enhanced implicit representation is used to guide the second branch network to perform adaptive adjustment according to the result of the first branch network; the target tunnel abnormality warning information is determined according to the target abnormality information implicit representation and the target tunnel section implicit representation.
[0005] As an implementation manner, determining target tunnel abnormality warning information according to the target abnormality information implicit representation and the target tunnel section implicit representation includes:
[0006] Obtaining a dot product between the implicit representation of the target abnormal information and the implicit representation of the target tunnel section, and obtaining a commonality metric value between the implicit representation of the target abnormal information and the implicit representation of the target tunnel section;
[0007] The target tunnel abnormality warning information is determined from the reference abnormality warning information set according to the commonality metric value.
[0008] As an implementation mode, the debugging process of the target tunnel monitoring network includes:
[0009] Obtain p first template relationship chains and q second template relationship chains, each of the first template relationship chains is used to represent a historical inclusion relationship between the tunnel section template and the abnormal information template, and each of the second template relationship chains is used to represent a historical inclusion relationship between the abnormal information template and the abnormal information template in the tunnel section template, and both p and q are natural numbers that are not 0;
[0010] Extracting an implicit link representation of each first template relationship chain and an implicit link representation of each second template relationship chain;
[0011] Using the implicit representations of the links of the q second template relationship chains as the execution data of the first branch network in the initial tunnel monitoring network, obtaining the implicit representation of the template abnormality information, and using the implicit representation of the template abnormality information or the first enhanced implicit representation and the implicit representations of the links of the p first template relationship chains as the execution data of the second branch network in the initial tunnel monitoring network, obtaining the implicit representation of the template tunnel section, wherein the first enhanced implicit representation is used to guide the second branch network to perform adaptive adjustment according to the result of the first branch network;
[0012] The initial tunnel monitoring network is debugged according to the implicit representation of the template abnormality information and the implicit representation of the template tunnel section to obtain the target tunnel monitoring network, which is used to process the implicit representation of the links of p first target relationship chains and the implicit representation of the links of q second target relationship chains to obtain the implicit representation of the target abnormality information and the implicit representation of the target tunnel section, which are used to determine the abnormality warning information of the target tunnel.
[0013] As an implementation mode, the first branch network includes a first fusion component and a first weight distribution component; the implicit representation of the links of the q second template relationship chains is used as the execution data of the first branch network in the initial tunnel monitoring network to obtain the implicit representation of the template abnormality information, including:
[0014] Loading the implicit link representation of each second template relationship chain to the first fusion component respectively, obtaining the semantic implicit representation of each second template relationship chain, wherein the semantic implicit representation of each second template relationship chain is used to represent the semantics of the abnormal information inclusion relationship in the corresponding second template relationship chain;
[0015] Using the semantic implicit representations of the q second template relationship chains as execution data of the first weight allocation component, to obtain a relationship chain weight of each second template relationship chain;
[0016] The semantic implicit representations of the q second template relationship chains and the relationship chain weights are combined according to the weights by the first weight allocation component to obtain an implicit representation of template exception information.
[0017] As an implementation manner, before loading the implicit link representation of each second template relationship chain into the first fusion component to obtain the semantic implicit representation of each second template relationship chain, the method further includes:
[0018] Acquire a second enhanced implicit representation, where the second enhanced implicit representation is used to guide the first branch network to perform adaptive adjustment according to a result of the second branch network;
[0019] Loading the implicit link representation of each second template relationship chain into the first fusion component to obtain the semantic implicit representation of each second template relationship chain includes:
[0020] The second enhanced implicit representation is stacked with the link implicit representation of each second template relationship chain to obtain an integrated implicit representation of each second template relationship chain;
[0021] Loading the integrated implicit representation of each second template relationship chain into the first fusion component respectively to obtain the semantic implicit representation of each second template relationship chain;
[0022] The step of combining the semantic implicit representations of the q second template relationship chains and the relationship chain weights according to the first weight distribution component to obtain the template exception information implicit representation includes:
[0023] According to the first weight allocation component, the semantic implicit representations of the q second template relationship chains and the relationship chain weights are combined according to the weights to obtain a target integrated implicit representation;
[0024] stacking the target integrated implicit representation and the second enhanced implicit representation to obtain an implicit representation of template abnormal information;
[0025] The second branch network includes a second fusion component and a second weight allocation component; the implicit representation of the template abnormal information or the first enhanced implicit representation, and the implicit representation of the links of p first template relationship chains are used as the execution data of the second branch network in the initial tunnel monitoring network to obtain the implicit representation of the template tunnel section, including:
[0026] stacking the first enhanced implicit representation with the link implicit representation of each of the first template relationship chains respectively to obtain a first integrated implicit representation of each of the first template relationship chains;
[0027] Loading the first integrated implicit representation of each of the first template relationship chains into the second fusion component respectively, to obtain the semantic implicit representation of each of the first template relationship chains, wherein the semantic implicit representation of each of the first template relationship chains is used to represent the semantics of the tunnel section inclusion relationship in the corresponding first template relationship chain;
[0028] Using the semantic implicit representation of each first template relationship chain as execution data of the second weight allocation component to obtain the relationship chain weight of each first template relationship chain;
[0029] According to the second weight allocation component, the semantic implicit representations of the p first template relationship chains and the relationship chain weights are combined according to the weights to obtain a first focused implicit representation;
[0030] The first focused implicit representation and the first enhanced implicit representation are stacked to obtain an implicit representation of a template tunnel section.
[0031] As an implementation mode, loading the first integrated implicit representation of each first template relationship chain into the second fusion component respectively to obtain the semantic implicit representation of each first template relationship chain includes:
[0032] In a first integrated implicit representation of a first branch chain, an implicit representation of a target tunnel section template and an implicit representation of each adjacent template connected to the target tunnel section template are extracted, wherein the first branch chain is a relationship chain in the p first template relationship chains, and the target tunnel section template is any tunnel section template in the first branch chain;
[0033] Performing inner product calculations on the implicit representation of the target tunnel section template and the implicit representation of each of the adjacent templates to obtain a template weight of each of the adjacent templates in the first branch chain;
[0034] According to the template weight of each of the adjacent templates, the implicit representations corresponding to the adjacent templates are merged according to the weights to obtain the semantic implicit representation of the first branch.
[0035] As an embodiment, the method further includes:
[0036] determining a first error between the template tunnel section implicit representation and the second enhanced implicit representation;
[0037] optimizing the first enhanced implicit representation and the second enhanced implicit representation according to the first error;
[0038] determining a second error between the template anomaly information implicit representation and the first enhanced implicit representation;
[0039] The second enhanced implicit representation and the first enhanced implicit representation are optimized according to the second error.
[0040] As an implementation mode, the second branch network includes s task adaptation components, a dynamic routing component, and an average dimension reduction component, s≥1; the template abnormal information implicit representation or the first enhanced implicit representation, and the implicit representation of the link of p first template relationship chains are used as the execution data of the second branch network in the initial tunnel monitoring network to obtain the implicit representation of the template tunnel section, including:
[0041] Loading the implicit link representation of each of the first template relationship chains and s task preference arrays into the s task adaptation components, obtaining s abnormal concern arrays of each of the first template relationship chains, each of the task adaptation components outputting one abnormal concern array, each of the task preference arrays being used to characterize the preference of the corresponding task adaptation component for the abnormal information template;
[0042] Loading the s exception concern arrays, the s task preference arrays and the template exception information implicit representation of each of the first template relationship chains into the dynamic routing component to obtain a second integrated implicit representation of each of the first template relationship chains;
[0043] The second integrated implicit representations of the p first template relationship chains are loaded into the average dimension reduction component to obtain an implicit representation of the template tunnel section.
[0044] As an implementation mode, the implicit link representations of p first template relationship chains and s task preference arrays are loaded into the s task adaptation components to obtain s abnormal attention arrays of each first template relationship chain, including:
[0045] In the implicit representation of the link of the second branch, the implicit representation of the target tunnel section template and the implicit representation of each adjacent template connected to the target abnormal information template are extracted, the second branch is a relationship chain in the p first template relationship chains, and the target tunnel section template is any abnormal information template in the second branch;
[0046] stacking the implicit representation of the target tunnel section template with the s task preference arrays respectively to obtain s first stacked arrays;
[0047] Performing inner product calculations on the s first stacked arrays and the implicit representation of each of the adjacent templates, respectively, to obtain a template weight of each of the adjacent templates in the second branch chain;
[0048] According to the template weight of each of the adjacent templates, implicit representations corresponding to the adjacent templates are merged according to the weights to obtain s abnormal attention arrays of the second branch chain;
[0049] The s exception concern arrays, the s task preference arrays and the template exception information implicit representation of each of the first template relationship chains are loaded into the dynamic routing component to obtain a second integrated implicit representation of each of the first template relationship chains, including:
[0050] For each of the first template relationship chains, inner product calculations are performed on the s task preference arrays and the implicit representation of the template exception information to obtain the weight of each of the exception attention arrays in each of the first template relationship chains;
[0051] For each of the first template relationship chains, the corresponding s exception attention arrays are merged according to the weights of the s exception attention arrays to obtain a second integrated implicit representation of each of the first template relationship chains.
[0052] In a second aspect, the present application provides a management and control system, including a memory and a processor, wherein the memory stores a computer program that can be executed on the processor, and the processor implements the steps in the above-mentioned method when executing the program.
[0053] The beneficial effects of the present application include: after obtaining p first template relationship chains and q second template relationship chains, the present application extracts the implicit link representation of each first template relationship chain and the implicit link representation of each second template relationship chain. In addition, because the first enhanced implicit representation can guide the second branch network to adjust according to the structure of the first branch network, after extracting and obtaining the implicit link representation of each first template relationship chain and the implicit link representation of each second template relationship chain, the implicit link representation of the q second template relationship chains is used as the execution data of the first branch network in the initial tunnel monitoring network to obtain the implicit representation of template abnormality information; and the implicit representation of template abnormality information or the first enhanced implicit representation and the implicit link representation of p first template relationship chains are used as the execution data of the second branch network in the initial tunnel monitoring network to obtain the implicit representation of template tunnel section. Based on this, after obtaining the implicit representation of template abnormality information and the implicit representation of template tunnel section, the initial tunnel monitoring network is debugged based on the implicit representation of template abnormality information and the implicit representation of template tunnel section to obtain the target tunnel monitoring network. Based on the target tunnel monitoring network, it is possible to process implicit representations of links of p first target relationship chains and implicit representations of links of q second target relationship chains, and obtain implicit representations of target abnormal information and implicit representations of target tunnel sections for determining abnormal warning information of the target tunnel. Based on the above process, the present application can use implicit representations of template abnormal information output by the first branch network, or the first enhanced implicit representation, and combine implicit representations of links of p first template relationship chains to perform feature mixing in the network. Based on this, in the second branch network of the initial tunnel monitoring network, the influence of implicit representations of template abnormal information output by another first branch network, or the first enhanced implicit representation on the implicit representation of the template tunnel section to be determined by the second branch network can be fully weighed, so that the obtained implicit representation of the template tunnel section represents more obvious abnormal features. In addition, the implicit representation of the template tunnel section and the implicit representation of the template abnormal information obtained by the present application iteratively optimize the initial tunnel monitoring network, so that the abnormal monitoring capability of the obtained target tunnel monitoring network is enhanced, and the abnormal warning information of the target tunnel is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present application and are used together with the specification to illustrate the technical solution of the present application.
[0055] Figure 1 A schematic diagram of the implementation process of an unmanned intelligent monitoring method for a tunnel provided in an embodiment of the present application.
[0056] Figure 2 A schematic diagram of the hardware entity of a management and control system provided in an embodiment of the present application. DETAILED DESCRIPTION
[0057] In order to make the purpose, technical solutions and advantages of the present application clearer, the technical solutions of the present application are further elaborated in detail below in conjunction with the drawings and embodiments. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.
[0058] Figure 1 A schematic diagram of the implementation process of an unmanned intelligent monitoring method for a tunnel provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method includes the following operations 110 to 150:
[0059] The present application provides an unmanned intelligent monitoring method for a tunnel. The method is performed based on a neural network. The neural network is a target tunnel monitoring network. The training process of the target tunnel monitoring network is first introduced below, which specifically includes the following steps:
[0060] Step S10: Obtain p first template relationship chains and q second template relationship chains, each first template relationship chain is used to represent the historical inclusion relationship between the tunnel section template and the abnormal information template, and each second template relationship chain is used to represent the historical inclusion relationship between the abnormal information template and the abnormal information template appearing in the tunnel section template, and both p and q are natural numbers not equal to 0.
[0061] In step S10, the control system can access the historical database of the tunnel monitoring system, which stores various types of data in the past tunnel monitoring process, including but not limited to monitoring videos of tunnel sections, sensor data, maintenance records, and abnormal event reports. The control system identifies and extracts records associated with tunnel sections and abnormal information from these historical data. For the acquisition of the first template relationship chain, the control system analyzes each record and identifies those historical records that clearly indicate that the tunnel section (such as "A tunnel section") has specific abnormal information (such as "water seepage"). These records constitute a relationship chain from the tunnel section template to the abnormal information template, for example, "A tunnel section → water seepage". The system organizes these relationship chains into a first template relationship chain set. Assuming that a total of p (for example, p=100) are extracted, each relationship chain represents a specific inclusion relationship between the tunnel section template and the abnormal information template that has appeared in its history. Next, in order to obtain the second template relationship chain, the control system adopts the opposite logic, that is, to find the historical records of which tunnel section templates those abnormal information templates appear in. For example, if multiple records indicate that the abnormal information "water seepage" has appeared in "Tunnel Section A" and "Tunnel Section B", two second template relationship chains can be formed: "Water seepage → Tunnel Section A" and "Water seepage → Tunnel Section B". The system also organizes these relationship chains into a set of second template relationship chains. Assuming that a total of q (for example, q=50) are extracted, each relationship chain represents the historical inclusion relationship of the abnormal information template and the tunnel section templates in which the abnormal information appears.
[0062] In the process of extracting these relationship chains, the control system can use natural language processing (NLP) technology to parse text records, or use machine learning models (such as classifiers or sequence labeling models) to automatically identify and classify relationship chains. For each relationship chain, the system converts it into a machine-understandable format, such as a vector or array form, to facilitate neural network processing in subsequent steps. For example, tunnel section templates and abnormal information templates can be encoded as one-hot encoding (One-Hot Encoding) or embedding vectors (Embedding Vectors), where each element or dimension represents a specific tunnel section or abnormal information. When constructing these relationship chains, the control system can consider the diversity and representativeness of the relationship chains to ensure that the trained target tunnel monitoring network can be generalized to a wider range of monitoring scenarios. For example, the system can filter and sort these relationship chains according to their frequency, importance, or timestamp to ensure the balance and effectiveness of the training data.
[0063] Step S20: extracting the implicit representation of the links of each first template relationship chain and the implicit representation of the links of each second template relationship chain.
[0064] First, the management and control system traverses all the first template relationship chains and second template relationship chains obtained in step S10. For each relationship chain, the system needs to convert it from text or symbolic form to an implicit link representation in the form of a vector or array. This conversion process, for example, involves encoding and embedding elements in the relationship chain so that the neural network can understand and process it. Taking the first template relationship chain as an example, suppose there is a relationship chain "A tunnel section → water seepage". In this relationship chain, "A tunnel section" and "water seepage" are two key elements, representing the tunnel section template and the abnormal information template, respectively. The management and control system first needs to encode these two elements. The encoding method can be a simple one-hot encoding or an embedded vector.
[0065] For "A tunnel section", assume that there is a pre-trained embedding matrix E, in which each row represents the embedding vector of a tunnel section. The control system obtains its corresponding embedding vector v_tunnel by finding the index of "A tunnel section" in the embedding matrix. Similarly, for the abnormal information template "water seepage", the control system also finds its corresponding embedding vector v_anomaly from another embedding matrix.
[0066] Then, the control system needs to combine these two embedded vectors into an implicit link representation. This can be achieved in many ways, such as simple concatenation, summation, or more complex function mapping. Assuming that the concatenation method is chosen, the implicit link representation of "A tunnel segment → water seepage" can be expressed as [v_tunnel;v_anomaly], where ";" represents the vector concatenation operation.
[0067] For the second template relationship chain, such as "water seepage → tunnel section A", the management and control system also performs the above encoding and combination process to generate the corresponding implicit representation of the link. It should be noted that although the elements are the same, due to the different directions of the relationship chains (from abnormal information to tunnel sections), the generated implicit representations of the links may also be different, depending on whether the combination method takes directionality into account. In practical applications, the generation of the embedding matrix, for example, relies on a large amount of historical data and is trained through machine learning algorithms such as Word2Vec, GloVe, or FastText. These algorithms are able to learn distributed representations of words or entities from a text corpus and capture the semantic and contextual relationships between them. In the tunnel monitoring scenario, the embedding matrix can contain the embedding vectors of all known tunnel sections and abnormal information, providing a rich feature representation for the neural network.
[0068] Finally, through step S20, the management and control system converts each first template relationship chain and second template relationship chain into a link implicit representation, providing basic data support for subsequent neural network training and processing. These link implicit representations not only contain the key element information in the relationship chain, but also imply the semantic and contextual relationship between them, which helps the neural network better understand the tunnel monitoring data and make accurate predictions and judgments.
[0069] Step S30: Using the implicit representation of the links of q second template relationship chains as the execution data of the first branch network in the initial tunnel monitoring network to obtain the implicit representation of the template exception information, and using the implicit representation of the template exception information or the first enhanced implicit representation and the implicit representation of the links of p first template relationship chains as the execution data of the second branch network in the initial tunnel monitoring network to obtain the implicit representation of the template tunnel section, and the first enhanced implicit representation is used to guide the second branch network to perform adaptive adjustments according to the results of the first branch network.
[0070] Step S30 is a crucial step in the unmanned intelligent monitoring method for tunnels. It is executed by the management and control system and involves data processing and interaction of two branch networks in the initial tunnel monitoring network, the first branch network and the second branch network. The management and control system first processes the implicit link representations of q second template relationship chains, which have been generated in step S20 and serve as input data for the first branch network. The implicit link representation of each second template relationship chain, such as "[v_anomaly;v_tunnel]" (where v_anomaly is the embedding vector of the abnormal information, v_tunnel is the embedding vector of the tunnel section, and it is assumed here that the implicit link representation takes into account the directionality of the relationship chain in some way), reflects the historical pattern of abnormal information appearing in a specific tunnel section. The first branch network is a specially designed neural network structure used to capture the correlation characteristics between the abnormal information and the tunnel section starting from the abnormal information. The network may contain multiple hidden layers, each of which processes the input data through a nonlinear activation function (such as ReLU) and transforms it through a weight matrix. In the forward propagation process of the first branch network, the input data (i.e., the implicit representation of the links of the q second template relationship chains) passes through each layer in turn, and finally outputs the implicit representation of the template abnormal information. These implicit representations are abstractions and generalizations of the abnormal information features in the input data, and may contain higher-level semantic information or patterns.
[0071] At the same time, in order to enhance the adaptability and accuracy of the second branch network, the control system also introduces the first enhanced implicit representation. This enhanced representation is not directly derived from the data, but is dynamically generated based on the output of the first branch network or other prior knowledge, and is intended to guide how the second branch network adjusts according to the results of the first branch network. For example, if the first branch network identifies that a certain type of abnormal information frequently appears in a specific type of tunnel section, the first enhanced implicit representation can emphasize this association so that the second branch network pays more attention when processing the relevant tunnel section. Next, the control system uses the template abnormal information implicit representation (or combined with the first enhanced implicit representation) and the link implicit representation of p first template relationship chains as input data for the second branch network. The link implicit representation of the first template relationship chain (such as "tunnel section A→water seepage") starts from the tunnel section and describes the historical situation of what abnormal information the tunnel section has. The task of the second branch network is to start from the tunnel section and combine the abnormal information characteristics to further analyze and infer the status or possible risks of the tunnel section. The second branch network also adopts a multi-layer neural network structure, which may include convolutional layers, pooling layers, fully connected layers, etc., depending on the complexity of the problem and the characteristics of the data. When processing input data, the second branch network comprehensively considers the implicit representation of template abnormal information (which may have been combined with the first enhanced implicit representation) and the implicit representation of the link of the first template relationship chain, and finally outputs the implicit representation of the template tunnel section through multiple layers of nonlinear transformation. These implicit representations reflect the comprehensive status or risk assessment of the tunnel section in the context of specific abnormal information. Throughout the process, the first branch network and the second branch network are closely connected through the implicit representation of template abnormal information (which may be affected by the first enhanced implicit representation), forming an end-to-end processing flow. This structure enables the network to simultaneously capture the bidirectional association from abnormal information to tunnel section and from tunnel section to abnormal information, improving the accuracy and robustness of the monitoring system.
[0072] Step S40: Debug the initial tunnel monitoring network according to the implicit representation of template abnormality information and the implicit representation of template tunnel section to obtain the target tunnel monitoring network, which is used to process the implicit representation of links of p first target relationship chains and the implicit representation of links of q second target relationship chains to obtain the implicit representation of target abnormality information and the implicit representation of target tunnel section, which are used to determine the abnormality warning information of the target tunnel.
[0073] In step S30, the management and control system has obtained the implicit representation of the template abnormal information through the first branch network, and the implicit representation of the template tunnel section through the second branch network. These implicit representations capture the associated features from abnormal information to tunnel sections and from tunnel sections to abnormal information, respectively. The goal of step S40 is to use these implicit representations to debug and optimize the initial tunnel monitoring network. The debugging process, for example, involves one or more training cycles, each cycle including steps such as forward propagation, loss calculation, back propagation, and weight update. In the forward propagation stage, the management and control system inputs the implicit representation of the template abnormal information and the implicit representation of the template tunnel section into the initial tunnel monitoring network, and observes the difference between the network output and the target value. The "target value" here can be a predefined label or expected output feature based on historical data.
[0074] In order to quantify the difference between the network output and the target value, the control system uses a loss function (such as cross entropy loss, mean square error loss, etc.) to calculate the loss value. Taking cross entropy loss as an example, if the network output is the probability distribution of each category, and the target value is a one-hot encoded vector, the cross entropy loss calculation formula can be expressed as:
[0075] ;
[0076] Where M is the total number of categories, is the true probability of the cth category in the target value (in one-hot encoding, only one element is 1 and the rest are 0), is the predicted probability of the cth category output by the network. After obtaining the loss value, the control system enters the back propagation stage. In this stage, the system calculates the gradient of the loss function with respect to the network weights according to the chain rule and back propagates these gradients from the output layer to the input layer. After the gradient calculation is completed, the control system updates the weights of the network according to the gradient descent method (or other optimization algorithms) to reduce the loss value. This process is repeated many times until the loss value converges to an acceptable level or reaches the preset upper limit of the training round. Through continuous iterative debugging, the weights of the initial tunnel monitoring network are gradually optimized, and finally the target tunnel monitoring network is generated. This network can more accurately process new tunnel monitoring data, that is, the implicit link representations of the p first target relationship chains and the implicit link representations of the q second target relationship chains. For each pair of input relationship chain implicit representations, the target tunnel monitoring network will output the corresponding target anomaly information implicit representation and the target tunnel section implicit representation. Finally, the control system uses these target implicit representations to determine the abnormal warning information of the target tunnel. This involves, for example, further decoding or classification of the implicit representation to extract specific anomaly types and location information. For example, the system can input the implicit representation of target abnormal information into a classifier to determine whether there are abnormal phenomena such as water seepage and cracks in the tunnel; at the same time, the system will also analyze the implicit representation of the target tunnel section to determine the specific location where the abnormal phenomenon occurs. With this information, the management and control system can generate a detailed abnormal warning report for reference and decision-making by tunnel managers.
[0077] As an implementation mode, step S10, obtaining p first template relationship chains and q second template relationship chains, includes: obtaining a relationship map, the relationship map includes p first template relationship chains and q second template relationship chains, and the relationship map is used to characterize the inclusion relationship between the tunnel section template and the abnormal information template.
[0078] In step S10, the management and control system collects p first template relationship chains and q second template relationship chains by acquiring a relationship map. A relationship map is a structured graphical representation method for representing the relationship between entities, which can present complex data relationships in an intuitive and clear manner. In the scenario of unmanned intelligent monitoring of tunnels, the relationship map is used to capture and display the historical association and inclusion relationship between the tunnel section (as an entity) and the abnormal information (also as an entity).
[0079] The control system first accesses a data warehouse or database that stores a large amount of historical monitoring data. These data contain various records of the tunnel monitoring system over a period of time, such as sensor readings, video monitoring logs, maintenance reports, and abnormal event records. From these raw data, the system uses natural language processing (NLP) technology, data mining algorithms, or knowledge graph construction tools to extract the association information between tunnel sections and abnormal information. Specifically, the system can identify text descriptions such as "Water seepage has frequently occurred in tunnel section A recently", and then parse out the two entities "Tunnel section A" and "Water seepage", as well as the "inclusion relationship" between them (that is, water seepage is an abnormal manifestation of tunnel section A). Similarly, the system can also identify situations where abnormal information appears in multiple tunnel sections, such as "Water seepage not only appears in tunnel section A, but also has records in tunnel sections B and C", thereby constructing a second template relationship chain.
[0080] In the process of constructing the relationship graph, each tunnel section and abnormal information is regarded as a node in the graph, and the relationship between them is represented as an edge. For example, in the relationship graph, the two nodes "A tunnel section" and "water seepage" are connected by a directed edge, with the direction from "A tunnel section" to "water seepage", indicating that "A tunnel section" has a historical inclusion relationship with the abnormal information "water seepage". Similarly, for the second template relationship chain, the "water seepage" node points to multiple tunnel section nodes, indicating that the abnormal information appears in multiple tunnel sections. By traversing the entire relationship graph, the management and control system can easily extract the required p first template relationship chains and q second template relationship chains. These relationship chains not only contain specific tunnel sections and abnormal information entities, but also imply the historical associations and patterns between them, providing rich contextual information for the subsequent training of neural network models.
[0081] For example, suppose the relationship graph contains the following relationship chains:
[0082] The first template relationship chain: "A tunnel section → water seepage", "B tunnel section → cracks" ... (a total of p);
[0083] The second template relationship chain: "water seepage → tunnel section A", "cracks → tunnel section B", "water seepage → tunnel section C" ... (q links in total);
[0084] By parsing these relationship chains, the management and control system can provide necessary input data for subsequent steps (such as the extraction of implicit representation of links in step S20), thereby promoting the implementation of the unmanned intelligent monitoring method for the entire tunnel.
[0085] As an implementation method, the first branch network includes a first fusion component and a first weight distribution component. Based on this, in step S30, the implicit representation of the links of the q second template relationship chains is used as the execution data of the first branch network in the initial tunnel monitoring network to obtain the implicit representation of the template abnormal information, including:
[0086] Step S31: loading the implicit link representation of each second template relationship chain into the first fusion component respectively, obtaining the semantic implicit representation of each second template relationship chain, wherein the semantic implicit representation of each second template relationship chain is used to characterize the semantics of the abnormal information inclusion relationship in the corresponding second template relationship chain;
[0087] Step S32: using the semantic implicit representations of the q second template relationship chains as execution data of the first weight allocation component to obtain the relationship chain weight of each second template relationship chain;
[0088] Step S33: The semantic implicit representations of the q second template relationship chains and the relationship chain weights are combined according to the first weight allocation component to obtain an implicit representation of the template exception information.
[0089] The first branch network consists of a first fusion component and a first weight assignment component, and is intended to extract implicit representation of template anomaly information from the second template relationship chain.
[0090] In step S31, the management and control system has obtained the implicit link representations of q second template relationship chains through step S20. These implicit link representations exist in the form of vectors or matrices, for example, which encode the historical association between abnormal information and tunnel sections. For example, an implicit link representation may be a vector containing multiple features, which describe the type, frequency, severity, and association with a specific tunnel section of the abnormal information. The task of the first fusion component is to further process these implicit link representations to extract higher-level semantic information. The "semantics" here refers to the abstract representation of the relationship between abnormal information and tunnel sections, which goes beyond simple feature combinations and can capture more complex patterns and meanings.
[0091] Assume that the link of the second template relationship chain "water seepage → tunnel segment A" is implicitly represented as a vector v_leak_A=[f1,f2,...,fn], where each element fi represents a feature value (such as the frequency, duration, impact range, etc. of water seepage). The first fusion component may be a multi-layer perceptron (MLP) or a variant of a recurrent neural network (RNN), such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU).
[0092] The control system passes v_leak_A as input to the first fusion component. Inside the component, through a series of hidden layers, each hidden layer applies a nonlinear activation function (such as ReLU) to transform the input and linearly combines it through the weight matrix. This process can be expressed as:
[0093] ;
[0094] ;
[0095] ;
[0096] in, and are the weight matrix and bias vector of the lth layer respectively, σ is the activation function (σ is used here to represent any activation function, which may be ReLU, tanh, etc. in specific implementation), is the output of the lth layer. Finally, the output of the first fusion component sem_leak_A (i.e., the semantic implicit representation of "water seepage → A tunnel segment") is the output of the last hidden layer. For example, assuming v_leak_A=[0.8,0.3,0.5] (just an example, the actual vector dimension may be higher), after being processed by the first fusion component, the obtained semantic implicit representation sem_leak_A may be a higher-dimensional vector, such as [0.4,0.6,-0.2,0.1]. Each element in this vector no longer directly corresponds to the original feature, but is an abstract representation of these features after nonlinear combination and transformation, which can better capture the semantics of the relationship between "water seepage" and "A tunnel segment".
[0097] In step S32, after obtaining the semantic implicit representation of each second template relationship chain, the next step is to assign weights according to the importance of these representations. This is because different second template relationship chains may have different amounts of information and predictive values, and therefore need to be given different levels of importance in subsequent processing. The task of the first weight assignment component is to dynamically calculate the weight of each relationship chain based on the content of the semantic implicit representation. This involves, for example, further analysis and comparison of the semantic implicit representations to determine which relationship chains contribute more to the final result. Specific implementation method: A common practice is to use an attention mechanism to implement weight assignment. The attention mechanism allows the model to dynamically adjust the degree of attention to different parts when processing input data. In the first weight assignment component, a scoring function based on attention weights can be designed, which accepts a semantic implicit representation as input and outputs a scalar weight. For example, for each semantic implicit representation (i represents different second template relationship chains), a scoring function score( )) to calculate its weight wi:
[0098] ;
[0099] Among them, the specific form of the score function can be designed according to the actual situation, for example, it can be a simple linear transformation plus nonlinear activation:
[0100] ;
[0101] here, and are the parameters of the scoring function, and σ is the activation function (such as a variant of the softmax function, but note that softmax is actually used in the denominator to normalize the weights, while exp(score(sem_i)) in the numerator is just a nonlinear transformation of the scoring function).
[0102] In step S33, after obtaining the weight of each second template relationship chain, the next step is to apply these weights to the corresponding semantic implicit representations and merge them to obtain the final implicit representation of template abnormal information. This process is similar to weighted averaging, but the data in vector or matrix form is processed here. Specific implementation method: The management and control system first weights the corresponding semantic implicit representation according to the weight calculated in step S32. Then, these weighted representations are merged (such as summation or weighted average) to obtain the implicit representation of template abnormal information. Assume that there are two second template relationship chains "seepage → tunnel section A" and "crack → tunnel section B", their semantic implicit representations are sem_leak_A=[0.4,0.6,-0.2,0.1] and sem_crack_B=[-0.3,0.5,0.1,0.7], and the weights calculated by the first weight allocation component are w_leak_A=0.7 and w_crack_B=0.3, respectively. Then, the final implicit representation of the template abnormal information can be obtained by weighted summation [0.19, 0.57, -0.07, 0.34]. This representation integrates the association between the two abnormal information of "water seepage" and "cracks" and the tunnel section, and is appropriately weighted according to their historical importance and characteristics. In summary, the various sub-steps in step S30 extract the implicit representation of the template abnormal information from the second template relationship chain through the collaborative work of the first fusion component and the first weight allocation component. This process not only takes into account the complex relationship between the abnormal information and the tunnel section, but also dynamically weights and merges these relationships through advanced technologies such as the attention mechanism, providing strong support for subsequent tunnel monitoring and abnormal warning.
[0103] As an implementation mode, in step S31, that is, before loading the implicit representation of the link of each second template relationship chain into the first fusion component respectively and obtaining the semantic implicit representation of each second template relationship chain, the method may also include: step S31a: obtaining the second enhanced implicit representation, and the second enhanced implicit representation is used to guide the first branch network to perform adaptive adjustment according to the result of the second branch network.
[0104] Then, step S31, loading the implicit link representation of each second template relationship chain into the first fusion component to obtain the semantic implicit representation of each second template relationship chain, may include:
[0105] Step S311: stacking the second enhanced implicit representation with the link implicit representation of each second template relationship chain respectively to obtain the integrated implicit representation of each second template relationship chain;
[0106] Step S312: Load the integrated implicit representation of each second template relationship chain into the first fusion component to obtain the semantic implicit representation of each second template relationship chain.
[0107] Before formally processing the implicit representation of the link of the second template relationship chain, the management and control system first executes step S31a to obtain a second enhanced implicit representation. This enhanced representation is not directly extracted from the data, but is generated based on other information or knowledge within the system, and is intended to provide an additional guidance signal to help the first branch network better adapt to subsequent processing tasks.
[0108] The second enhanced implicit representation can be generated in many ways. Here, it is assumed that it comes from some preprocessing result of the second branch network or the output of an independent machine learning model. For example, the second branch network may have generated some intermediate results or feature representations when processing the first template relationship chain. These results reflect the deep connection between the tunnel section and the abnormal information. The management and control system can extract the most critical or representative features from these results and generate the second enhanced implicit representation through some aggregation or encoding method.
[0109] In addition, if there is an independent reinforcement learning model or meta-learning model in the system, it can also generate a second reinforcement implicit representation based on the system's historical performance, current state or external environment. This model can learn how to dynamically adjust internal parameters or behavior strategies according to the actual needs of the system, thereby providing targeted guidance signals.
[0110] Regardless of the approach adopted, the second enhanced implicit representation should be a high-dimensional vector or matrix that contains rich contextual information or prior knowledge, which can guide the first branch network to focus more on important features or patterns when processing the second template relationship chain.
[0111] For example, suppose that when the second branch network processes the first template relationship chain "Tunnel Segment A → Water Seepage", it generates an intermediate feature representation vector feat_A=[0.6,-0.2,0.3,0.1], which reflects the specific correlation characteristics between "Tunnel Segment A" and the "Water Seepage" anomaly. The management and control system can use some feature selection or dimensionality reduction technology (such as PCA, LDA, etc.) to process this feature representation and obtain a more compact and representative second enhanced implicit representation boost_rep=[0.8,-0.5]. This enhanced representation emphasizes the significance of the "Water Seepage" anomaly in "Tunnel Segment A" and provides important guidance information for subsequent processing.
[0112] After obtaining the second enhanced implicit representation, the management and control system enters step S31 to combine the link implicit representation of each second template relationship chain with the second enhanced implicit representation to obtain a higher-level semantic implicit representation. In order to integrate the information of the second enhanced implicit representation into the processing flow of the second template relationship chain, the management and control system first stacks the second enhanced implicit representation with the link implicit representation of each second template relationship chain. Stacking is a simple and effective feature combination method that forms a new and richer representation by splicing feature vectors or matrices from different sources in dimensions.
[0113] Assume that there is a second template relationship chain “water seepage → tunnel segment B”, whose implicit link representation is link_rep_B=[0.4,0.2,-0.1,0.3]. The management and control system stacks this implicit link representation with the second enhanced implicit representation boost_rep=[0.8,-0.5] obtained earlier to form a new integrated implicit representation [0.8,-0.5,0.4,0.2,-0.1,0.3]. This integrated implicit representation not only contains the information in the original implicit link representation, but also introduces the guidance signal in the second enhanced implicit representation, providing a more comprehensive context for subsequent processing.
[0114] After the integrated implicit representation is loaded into the first fusion component and the stacking operation is completed, the management and control system loads the integrated implicit representation of each second template relationship chain into the first fusion component respectively. The first fusion component is a complex machine learning model (such as a deep neural network), which can perform nonlinear transformation and feature extraction on the input high-dimensional data to generate a higher-level semantic implicit representation. Specific implementation method: The first fusion component may contain multiple hidden layers, each of which applies a series of nonlinear activation functions (such as ReLU, tanh, etc.) and weight matrices to transform the input data. When processing the integrated implicit representation, the component extracts and combines features layer by layer, and finally outputs an implicit representation containing rich semantic information. For example, in the first hidden layer, the component can apply a linear transformation and a ReLU activation function to the integrated implicit representation. This process is repeated for multiple hidden layers until the final semantic implicit representation is generated. For example (continuing the previous example): Assume that the first fusion component is an MLP with two hidden layers. In the first hidden layer, after linear transformation and ReLU activation, the output may become: (The values here are hypothetical, and the actual output will depend on the specific weights and bias values as well as the input data.) This output is then passed to the second hidden layer for further processing until the final semantic implicit representation sem_rep_B is generated. This sem_rep_B not only contains the original information of the relationship chain "water seepage → tunnel segment B", but also incorporates the guidance signal in the second enhanced implicit representation, so that it can more accurately reflect the deep-level association and semantic meaning between the "water seepage" anomaly and the "tunnel segment B". In summary, through the processing flow of steps S31a and S31, the management and control system combines the second enhanced implicit representation with the link implicit representation of the second template relationship chain, and generates a higher-level semantic implicit representation through the first fusion component. This process not only improves the processing capability and adaptability of the first branch network, but also provides a more accurate and comprehensive information basis for subsequent anomaly detection and early warning.
[0115] As an implementation mode, step S33, according to the first weight distribution component, combines the semantic implicit representations of the q second template relationship chains and the relationship chain weights according to the weights to obtain the implicit representation of the template abnormal information, including:
[0116] Step S331: merging the semantic implicit representations of the q second template relationship chains and the relationship chain weights according to the first weight allocation component to obtain a target integrated implicit representation;
[0117] Step S332: stacking the target integrated implicit representation and the second enhanced implicit representation to obtain the template abnormality information implicit representation.
[0118] In the embodiment of the present application, step S33 involves merging the semantic implicit representations of multiple second template relationship chains according to their respective importance (i.e., weight), and further combining the second enhanced implicit representation to generate the final template abnormal information implicit representation. This process not only makes full use of the information provided by different relationship chains, but also enhances the adaptability and accuracy of the model by introducing an enhanced mechanism.
[0119] In step S331, the management and control system uses the weight of each second template relationship chain calculated by the first weight distribution component to merge the semantic implicit representations of these relationship chains according to the weights to generate a target integrated implicit representation that integrates all relationship chain information. This process is similar to weighted averaging, but here it processes data in the form of high-dimensional vectors or matrices.
[0120] Assume that there are q second template relationship chains, and the semantic implicit representation of each relationship chain is sem_rep_1, sem_rep_2, ..., sem_rep_q, and the first weight allocation component allocates corresponding weights w_1, w_2, ..., w_q to these semantic implicit representations. The management and control system first ensures that the sum of these weights is 1 (through normalization), and then performs weighted averaging of the semantic implicit representations according to these weights. The weighted averaging operation here is performed between corresponding elements, that is, each element in each semantic implicit representation vector is weighted and summed according to its corresponding weight. Since the semantic implicit representation is, for example, a high-dimensional vector, this weighted averaging operation will traverse all dimensions of the vector. For example, suppose there are two second template relationship chains "water seepage → tunnel section A" and "cracks → tunnel section B", their semantic implicit representations are: sem_rep_A=[0.4,0.6,-0.2,0.1] (representing the semantics of "water seepage → tunnel section A") and sem_rep_B=[-0.3,0.5,0.1,0.7] (representing the semantics of "cracks → tunnel section B"), and the weights assigned by the first weight allocation component to these two relationship chains are w_A=0.6 and w_B=0.4 respectively. The control system performs weighted averaging of the semantic implicit representations according to these weights to obtain the target integrated implicit representation [0.12,0.56,-0.06,0.34]. This integrated implicit representation integrates the association between the two abnormal information of "water seepage" and "cracks" and the tunnel section, and performs appropriate weighting according to their importance (i.e., weight). Step S332: Stack the target integrated implicit representation with the second enhanced implicit representation to obtain the template abnormality information implicit representation. After obtaining the target integrated implicit representation, the control system enters step S332 to stack this representation with the second enhanced implicit representation to generate the final template abnormality information implicit representation. Stacking is a simple feature combination method that forms a new and richer representation by splicing feature vectors or matrices from different sources in dimensions. Specific implementation method: After the second enhanced implicit representation boost_rep has been obtained in some way (such as step S31a above), the control system splices it with the target integrated implicit representation in dimensions. Assuming that the target integrated implicit representation is an n-dimensional vector and the second enhanced implicit representation is an m-dimensional vector, the spliced template abnormality information implicit representation will be an n+m-dimensional vector. For example (continuing the previous example): Assuming that the second enhanced implicit representation is boost_rep=[0.8,-0.5], this is a two-dimensional vector containing additional guidance information. The control system stacks this vector with the target integrated implicit representation [0.12, 0.56, -0.06, 0.34] to obtain the template abnormal information implicit representation [0.12, 0.56, -0.06, 0.34, 0.8, -0.5].This representation not only includes the comprehensive association information between the two abnormal information of "water seepage" and "cracks" and the tunnel section (embodied by the target integrated implicit representation), but also introduces the guidance signal in the second enhanced implicit representation. These guidance signals may come from multiple aspects such as the system's historical performance, current state or external environment. They can help the model better understand and process the input data, thereby improving the accuracy of anomaly detection and early warning. Through the processing flow of the above step S33 and its sub-steps S331 and S332, the control system merges the semantic implicit representations of multiple second template relationship chains according to the weights, and further combines the second enhanced implicit representation to generate the template abnormal information implicit representation. This process makes full use of the information provided by different relationship chains and the guidance signals provided by the reinforcement mechanism, providing a more comprehensive and accurate basis for subsequent anomaly detection and early warning. At the same time, since this process is based on the operation of high-dimensional vectors or matrices and involves complex nonlinear transformations and feature combination techniques (such as hidden layer processing in neural networks), it can capture the deep-level associations and pattern features in the data, thereby improving the intelligence level and adaptability of the entire system.
[0121] As an implementation method, the second branch network includes a second fusion component and a second weight distribution component. Based on this, in step S30, the template abnormality information implicit representation or the first enhanced implicit representation, and the implicit representation of the links of the p first template relationship chains are used as the execution data of the second branch network in the initial tunnel monitoring network to obtain the implicit representation of the template tunnel section, including:
[0122] Step S34: stacking the first enhanced implicit representation with the link implicit representation of each first template relationship chain respectively to obtain a first integrated implicit representation of each first template relationship chain;
[0123] Step S35: loading the first integrated implicit representation of each first template relationship chain into the second fusion component respectively, obtaining the semantic implicit representation of each first template relationship chain, wherein the semantic implicit representation of each first template relationship chain is used to represent the semantics of the tunnel section inclusion relationship in the corresponding first template relationship chain;
[0124] Step S36: using the semantic implicit representation of each first template relationship chain as execution data of the second weight allocation component to obtain the relationship chain weight of each first template relationship chain;
[0125] Step S37: merging the semantic implicit representations of the p first template relationship chains and the relationship chain weights according to the second weight allocation component to obtain a first focused implicit representation;
[0126] Step S38: stacking the first focused implicit representation and the first enhanced implicit representation to obtain an implicit representation of the template tunnel section.
[0127] In step S34, the control system first stacks the first enhanced implicit representation with the implicit representation of the link of each first template relationship chain. The first enhanced implicit representation is an additional guidance signal, which may come from the system's historical experience, an external knowledge base, or the output of other related models, and is intended to provide the second branch network with global guidance on the relationship between the tunnel section and the abnormal information. Assume that the first enhanced implicit representation is a vector boost_vec=[b1,b2,...,bn], where each element bi represents a kind of enhanced information. At the same time, assuming that there are p first template relationship chains, the implicit representation of the link of each relationship chain is also a vector, for example, the implicit representation of the link of the first relationship chain is link_rep_1=[l11,l12,...,l1m]. The control system stacks these implicit representations of the link with the first enhanced implicit representation to form a new integrated implicit representation. The stacking operation is a splicing operation in the vector dimension, that is, simply connecting two vectors together to form a longer vector.
[0128] For the first relationship chain, the first integrated implicit representation after stacking is [b1, b2, ..., bn, l11, l12, ..., l1m]. Similarly, a corresponding first integrated implicit representation can be generated for each first template relationship chain. For example, assuming boost_vec = [0.5, -0.2], and the link implicit representation of the first first template relationship chain "A tunnel segment → water seepage" is link_rep_A = [0.8, 0.3, -0.1]. The first integrated implicit representation after stacking is [0.5, -0.2, 0.8, 0.3, -0.1].
[0129] In step S35, the control system takes the first integrated implicit representation of each first template relationship chain as input and loads it into the second fusion component for processing. The second fusion component is a complex machine learning model that can perform nonlinear transformation and feature extraction on the input high-dimensional data, thereby generating a higher-level semantic implicit representation. Specific implementation method: The second fusion component may be a combination of models such as a multi-layer perceptron (MLP), a convolutional neural network (CNN) or a recurrent neural network (RNN). Taking MLP as an example, it contains multiple hidden layers, each hidden layer is composed of multiple neurons, each neuron receives the output of the previous layer of neurons as input, and performs nonlinear transformation through an activation function. In the second fusion component, each first integrated implicit representation is processed by each hidden layer in turn, and finally generates the corresponding semantic implicit representation. For example (continued from the previous example): Assume that the second fusion component is an MLP containing two hidden layers. The first hidden layer may apply a ReLU activation function to perform a nonlinear transformation on the input and output a new feature vector. The second hidden layer further processes this feature vector and generates the final semantic implicit representation. For the first first template relationship chain, the semantic implicit representation obtained after processing by the second fusion component may be sem_rep_A=[0.2,-0.3,0.4,0.1].
[0130] In step S36, the control system passes the semantic implicit representation of each first template relationship chain as input to the second weight allocation component. The task of this component is to assign a weight to each semantic implicit representation, which reflects the importance of the semantic implicit representation in subsequent processing. The second weight allocation component may be a simple linear layer that calculates the weight by linearly transforming the semantic implicit representation. Alternatively, the attention mechanism is used to implement weight allocation, because the attention mechanism can dynamically adjust the weight according to the input content, so as to handle complex data relationships more flexibly. In the attention mechanism, for example, a scoring function is defined to calculate the score of each semantic implicit representation, and then the score is converted into a normalized weight through the softmax function. The specific form of the scoring function can be designed according to the actual situation, for example, it can be a dot product, cosine similarity or a complex neural network. For example, assume that the dot product is used as the scoring function, and the second weight allocation component has learned the corresponding parameters. For the semantic implicit representation sem_rep_A of the first first template relationship chain, its weight w_A can be calculated as follows:
[0131] ;
[0132] ;
[0133] Among them, context_vec is a context vector, which may represent the global context information of the current tunnel monitoring task. The weight w_A is obtained by performing softmax normalization on the score of sem_rep_A and the scores of all first template relationship chains.
[0134] In step S37, the management and control system performs weighted averaging on the semantic implicit representations of p first template relationship chains according to the weights calculated by the second weight allocation component to generate a first focused implicit representation. This representation integrates the information of all first template relationship chains and performs appropriate weighting according to their importance. Specific implementation method: The weighted average operation is performed between corresponding elements. For each element in each semantic implicit representation, a weighted sum is performed according to its corresponding weight. Assume that the semantic implicit representation of the first first template relationship chain is sem_rep_1 with a weight of w_1; the second is sem_rep_2 with a weight of w_2; and so on, until the pth. Then the first focused implicit representation focused_rep can be calculated in the following way:
[0135] ;
[0136] The weighted average operation here is performed independently on each dimension of the semantic implicit representation. For example (continuing the previous example): Assume that in addition to sem_rep_A, there is another semantic implicit representation of the first template relationship chain "B tunnel segment → crack" sem_rep_B=[-0.1,0.2,0.5,-0.3], and its weight is w_B. Then the first focused implicit representation can be calculated as follows:
[0137] ;
[0138] In step S38, the management and control system stacks the first focused implicit representation with the first enhanced implicit representation to generate the final implicit representation of the template tunnel section. The purpose of this step is to combine the global enhanced information with the local focused information to obtain a comprehensive and accurate representation of the tunnel section. Specific implementation method: The stacking operation is similar to the stacking in step S34, both of which splice two vectors in dimension. This representation integrates the information of multiple first template relationship chains, global enhanced information, and focused information obtained by weighted averaging, providing strong support for subsequent tunnel monitoring and abnormal warning. Through the processing flow of the above-mentioned step S30 and its sub-steps, the management and control system combines the implicit representation of the template abnormal information, the first enhanced implicit representation, and the implicit representation of the link of the first template relationship chain, and generates the implicit representation of the template tunnel section through the processing of the second branch network. This process not only makes full use of information resources from multiple sources, but also deeply processes and abstractly expresses the information through complex machine learning models.
[0139] As an implementation mode, step S35, loading the first integrated implicit representation of each first template relationship chain into the second fusion component to obtain the semantic implicit representation of each first template relationship chain, includes:
[0140] Step S351: extracting the implicit representation of the target tunnel section template and the implicit representation of each adjacent template connected to the target tunnel section template from the first integrated implicit representation of the first branch chain, the first branch chain being a relationship chain in the p first template relationship chains, and the target tunnel section template being any tunnel section template in the first branch chain;
[0141] Step S352: performing inner product calculations on the implicit representation of the target tunnel section template and the implicit representation of each adjacent template, respectively, to obtain a template weight of each adjacent template in the first branch chain;
[0142] Step S353: Based on the template weight of each adjacent template, the implicit representations of the corresponding adjacent templates are merged according to the weights to obtain the semantic implicit representation of the first branch.
[0143] In step S35, the control system is responsible for inputting the first integrated implicit representation of each first template relationship chain (the representation is obtained by stacking the first enhanced implicit representation and the link implicit representation in step S34) into the second fusion component. The second fusion component is a complex machine learning model that can deeply process the input high-dimensional data, extract key features and generate semantic implicit representations. This representation not only contains the information of the original data, but also incorporates the model's understanding and abstraction of the data. The second fusion component may adopt the structure of a variety of machine learning models, such as a deep neural network (DNN), a convolutional neural network (CNN) or a recurrent neural network (RNN). However, in this embodiment, a model that can process graph structure data (such as the relationship chain between tunnel sections and abnormal information) is particularly focused on, such as a graph neural network (GNN) or its variants. Graph neural networks are particularly suitable for processing network structure data composed of nodes (such as tunnel sections) and edges (such as relationship chains) because they can capture complex interactions and dependencies between nodes. In order to explain step S35 and its sub-steps more directly, it can be assumed that the second fusion component is a customized neural network layer or module that contains specific feature extraction and combination mechanisms.
[0144] In step S351, the control system first extracts key information from the first integrated implicit representation of the first branch (i.e., any one of the p first template relationship chains). This information includes the implicit representation of the target tunnel section template and the implicit representation of each adjacent template connected to the tunnel section template. Assuming that the first branch is "A tunnel section → water seepage", its first integrated implicit representation is [b1, b2, ..., bn, l1, l2, ..., lm], where [b1, b2, ..., bn] is the first enhanced implicit representation and [l1, l2, ..., lm] is the link implicit representation. The control system needs to separate the information directly related to the target tunnel section template "A tunnel section" from this integrated implicit representation. Since the integrated implicit representation may not directly distinguish the sources of different parts (i.e., which elements belong to the target tunnel section template and which belong to the adjacent templates), the control system can rely on additional structural information or labels to determine the ownership of these elements. However, to simplify the description, it is assumed that the integrated implicit representation has been preprocessed in some way to a form that can be directly extracted.
[0145] In practical applications, the implicit representations of the target tunnel section template and the neighboring templates may not be directly stored as continuous subvectors in the integrated implicit representation, but are associated with the elements in the integrated implicit representation through some index or mapping relationship. The control system extracts the implicit representation of the target tunnel section template (such as tunnel_rep=[t1,t2,...,tk]) and the implicit representation of each neighboring template connected to it (such as neighbor_rep_1=[n11,n12,...,n1p], neighbor_rep_2=[n21,n22,...,n2q], etc.) based on these relationships.
[0146] In step S352, the control system determines the importance or weight of each neighboring template connected to the target tunnel section template by inner product calculation. This step is based on an assumption that the similarity or correlation between the neighboring template and the target tunnel section template can be measured by the inner product between their implicit representations. For the implicit representation tunnel_rep of the target tunnel section template and the implicit representation of each neighboring template (such as neighbor_rep_1, neighbor_rep_2, etc.), the control system calculates their inner products as weights respectively. Inner product calculation is a common method to measure the similarity of two vectors. It calculates the sum of the products of the corresponding elements of the two vectors.
[0147] In other embodiments, a learnable weight matrix or neural network layer may be used to dynamically calculate weights, so that the weight values may be adjusted according to different inputs and context information.
[0148] In step S353, the control system performs weighted averaging of the implicit representations of the adjacent templates according to the weights calculated in step S352 to generate a semantic implicit representation of the first branch. This representation integrates the information of all adjacent templates and performs appropriate weighting processing according to their similarity or correlation with the target tunnel section template. Specific implementation method The weighted averaging operation is performed between corresponding elements implicitly represented by adjacent templates. For each element position j (assuming that all elements implicitly represented by adjacent templates at this position exist and are comparable), the control system calculates the weighted sum of all adjacent templates at this position as the element value of the semantic implicit representation at this position.
[0149] ;
[0150] Where r is the number of neighboring templates, semantic_rep_j is the value of the element in the semantic implicit representation at position j, weight_i is the weight of neighboring template i, and neighbor_rep_ij is the value of the element in the implicit representation of neighboring template i at position j. It should be noted that the above weighted average operation does not actually directly consider the implicit representation tunnel_rep of the target tunnel section template. In practical applications, tunnel_rep can be added to the semantic implicit representation as an additional feature or context information, or combined with the weighted averaged neighbor template implicit representation in some way (such as concatenation, dot product, etc.). However, here, in order to simplify the description and maintain consistency with the original step description, it is assumed that the semantic implicit representation is only composed of the weighted averaged neighbor template implicit representation. The final semantic implicit representation semantic_rep is a high-dimensional vector that captures the complex relationship and pattern characteristics between the target tunnel section template and its neighboring templates. This representation will be used as one of the input data for subsequent processing steps (such as weight allocation in step S36 and merging operation in step S37). Through the processing flow of step S35 and its sub-steps S351, S352, and S353, the management and control system converts the first integrated implicit representation of each first template relationship chain into a semantic implicit representation. This process not only extracts key feature information but also effectively combines and utilizes this information through inner product calculation and weighted average operation.
[0151] As an embodiment, the method further includes:
[0152] Step S3A: determining a first error between the template tunnel section implicit representation and the second enhanced implicit representation;
[0153] Step S3B: Optimize the first enhanced implicit representation and the second enhanced implicit representation according to the first error.
[0154] In step S3A, the control system calculates the difference between the implicit representation of the template tunnel section and the second enhanced implicit representation. This difference is measured by a certain error metric, and the mean square error (MSE) is used as an example here. MSE is a commonly used error metric that calculates the average of the squares of the differences between the predicted value and the true value, which can intuitively reflect the accuracy of the prediction result.
[0155] Assume that the template tunnel segment is implicitly represented as tunnel_seg_rep=[t1,t2,...,tn], and the second boost is implicitly represented as boost_rep_2=[b1,b2,...,bn] (note that the dimensions of the two vectors are assumed to be the same, and appropriate dimension matching or conversion can be performed in actual applications). The control system first calculates the difference between the corresponding elements of the two vectors, then squares these differences and calculates the average to obtain the MSE value.
[0156] In step S3B, the control system uses the first error (i.e., MSE value) calculated in step S3A to guide the optimization process of the first enhanced implicit representation and the second enhanced implicit representation. This process involves, for example, the application of optimization techniques such as backpropagation and gradient descent.
[0157] The core idea of the optimization process is to minimize the MSE value by adjusting the relevant parameters of the first reinforced implicit representation and the second reinforced implicit representation (such as the weights and biases in the neural network that generates these representations). This requires multiple iterations during the training phase of the system, and each iteration includes two steps: forward propagation (calculating the MSE value) and back propagation (updating the parameters according to the gradient of the MSE value).
[0158] Forward propagation: In each iteration, the control system first calculates the MSE value through forward propagation based on the current first reinforced implicit representation and second reinforced implicit representation (and their corresponding generation network parameters).
[0159] Backpropagation: The system then calculates the gradients of the MSE value with respect to the first reinforced implicit representation and the second reinforced implicit representation (or more specifically, the neural network parameters that generate these representations). These gradients indicate how to adjust the parameters to reduce the MSE value.
[0160] Parameter update: Finally, the system updates the parameters of the generated network according to the direction and size of the gradient, for example, making small adjustments in the opposite direction of the gradient (i.e., the direction in which the MSE value decreases).
[0161] This process is repeated multiple times until the MSE value converges to an acceptable level or reaches the preset upper limit of the number of iterations.
[0162] After multiple iterations, as the MSE value gradually decreases, the first enhanced implicit representation and the second enhanced implicit representation will become closer and closer to the actual tunnel section state or characteristics described by the template tunnel section implicit representation. This can not only improve the system's monitoring accuracy of the tunnel section, but also enhance the system's adaptability and robustness to different monitoring scenarios.
[0163] Steps S3A and S3B provide an important feedback mechanism for the training process of the unmanned intelligent monitoring system of the tunnel by quantifying the difference (i.e., the first error) between the implicit representation of the template tunnel section and the second enhanced implicit representation and optimizing the first enhanced implicit representation and the second enhanced implicit representation accordingly. This process not only helps to improve the monitoring accuracy of the system but also enhances the adaptability and robustness of the system. At the same time, it also demonstrates the powerful ability and flexibility of machine learning technology in handling complex monitoring tasks.
[0164] As an embodiment, the method further includes:
[0165] Step S3C: determining a second error between the implicit representation of the template abnormal information and the first enhanced implicit representation;
[0166] Step S3D: Optimize the second enhanced implicit representation and the first enhanced implicit representation according to the second error.
[0167] In step S3C, the control system is responsible for calculating the difference between the implicit representation of the template abnormal information and the first enhanced implicit representation, which is measured by a certain error metric. Similar to step S3A, the mean square error (MSE) can also be used as an error metric here because it can intuitively reflect the size of the difference between two vectors or matrices. However, it is worth noting that the implicit representation of template abnormal information and the first enhanced implicit representation play different roles in the system. The former focuses on the abstract expression of abnormal information, while the latter serves as a global guidance signal for monitoring tunnel sections. Therefore, the difference between them not only reflects the accuracy of information expression, but also reflects the synergy between different components within the system.
[0168] Assume that the template anomaly information is implicitly represented as anomaly_rep=[a1,a2,...,am], and the first enhancement is implicitly represented as boost_rep_1=[b1,b2,...,bn] (the dimensions of the two may be different here, and the dimensions can be matched or converted in actual applications). The control system first calculates the difference between the corresponding elements of the two vectors (or the elements after some transformation), then squares these differences, and calculates the average value to obtain the MSE value. However, since the dimensions may not match, it may not be feasible to directly calculate the MSE, so a feasible solution is to perform some form of dimensionality reduction or mapping on the two so that they can be compared in the same dimensional space. In order to simplify the description and maintain the relevance to the original description, it can be assumed that the two have been converted to the same dimensional space through some preprocessing steps (such as principal component analysis PCA, autoencoder dimensionality reduction, etc.), or an error measurement method that can handle inputs of different dimensions (such as weighted Kullback-Leibler divergence, etc.) is adopted. However, here, MSE is still used as an example for explanation, and it is assumed that the dimensions have been matched.
[0169] In step S3D, the control system uses the second error (i.e., MSE value) calculated in step S3C to guide the optimization process of the second enhanced implicit representation and the first enhanced implicit representation. Similar to step S3B, this process also involves the application of optimization techniques such as back propagation algorithm and gradient descent method. However, unlike step S3B, the optimization goal here is to adjust the neural network parameters that generate the second enhanced implicit representation and the first enhanced implicit representation (if they are generated by a neural network) to minimize the second error.
[0170] The core idea of the optimization process is to reduce the second error by adjusting the relevant parameters. This requires multiple iterations during the training phase of the system, each of which includes two steps: forward propagation (calculating the second error) and back propagation (updating the parameters according to the gradient of the second error).
[0171] Since the first reinforced implicit representation and the second reinforced implicit representation may be generated by different neural networks, and their positions and roles in the network may also be different, these differences need to be considered when performing backpropagation and parameter updates. For example, the first reinforced implicit representation may focus more on the immediate feedback of tunnel section monitoring, while the second reinforced implicit representation may focus more on the long-term guidance and optimization of the entire monitoring system.
[0172] After multiple iterations, as the second error gradually decreases, the first enhanced implicit representation and the second enhanced implicit representation will work more and more harmoniously in the unmanned intelligent monitoring system of the tunnel. This coordination is not only reflected in their respective contributions to the monitoring task, but also in their mutual dependence and mutual promotion. Ultimately, this optimization process will help improve the overall monitoring performance and accuracy of the system, allowing the system to more accurately identify abnormal information in the tunnel and issue early warnings in a timely manner.
[0173] As an implementation method, the second branch network includes s task adaptation components, a dynamic routing component, and an average dimension reduction component, where s≥1. Based on this, in step S30, the template abnormality information implicit representation or the first enhanced implicit representation, and the implicit representation of the links of the p first template relationship chains are used as the execution data of the second branch network in the initial tunnel monitoring network to obtain the implicit representation of the template tunnel section, including:
[0174] Step S30A: Load the implicit link representation of each first template relationship chain and s task preference arrays into s task adaptation components, obtain s abnormal concern arrays of each first template relationship chain, each task adaptation component outputs an abnormal concern array, and each task preference array is used to represent the preference of the corresponding task adaptation component for the abnormal information template;
[0175] Step S30B: loading the s exception concern arrays, s task preference arrays and template exception information implicit representation of each first template relationship chain into the dynamic routing component to obtain a second integrated implicit representation of each first template relationship chain;
[0176] Step S30C: Load the second integrated implicit representation of the p first template relationship chains into the average dimension reduction component to obtain the implicit representation of the template tunnel section.
[0177] In the embodiment of the present application, step S30 is a key link in the neural network processing flow, especially when the second branch network adopts a more complex structure, including multiple task adaptation components, dynamic routing components and average dimensionality reduction components, the implementation of this step becomes richer and more flexible.
[0178] In step S30A, the management and control system loads the implicit link representation of each first template relationship chain together with the pre-defined s task preference arrays into s task adaptation components. Each task adaptation component is an expert network, which is designed to process a specific type of task or abnormal information and adjust its processing strategy according to the task preference array.
[0179] Each task adaptation component is an independent neural network that may contain multiple hidden layers, activation functions, and output layers. Their role is to extract and transform the input link implicit representation to generate the corresponding anomaly attention array. The anomaly attention array reflects the component's attention to the abnormal information in the current input link implicit representation and the recognition results. Each task adaptation component is associated with a task preference array, which is defined before training and is used to guide the component's behavior when processing specific types of abnormal information. The task preference array may contain a series of weights or probability values, indicating the component's preference for different features or types of anomalies.
[0180] For each first template relationship chain's implicit link representation, the control system passes it along with s task preference arrays as input to the corresponding task adaptation component. Each component processes the input independently and outputs an anomaly concern array. Assuming there are p first template relationship chains, each component will generate an anomaly concern array for each relationship chain, so a total of ps anomaly concern arrays will be generated.
[0181] For example, suppose there are two task adaptation components (s=2), and the task preference array of each component is biased towards identifying two types of abnormal information, "water seepage" and "cracks". For the first template relationship chain "A tunnel segment → water seepage", its link implicit representation is link_rep_A=[0.6,0.2,-0.1,0.3]. The management and control system loads this representation together with the two task preference arrays into the two task adaptation components.
[0182] Component 1 (biased towards water seepage identification): Based on link_rep_A and its own task preference array, it generates an abnormal focus array focus_arr_A1=[0.8,0.1,0.05,0.05], indicating a strong focus on the "water seepage" feature.
[0183] Component 2 (biased towards crack identification): link_rep_A is processed in the same way, but due to different task preferences, the generated abnormal focus array focus_arr_A2=[0.2, 0.3, 0.4, 0.1] may focus more on features that are not directly related to cracks.
[0184] In step S30B, the management and control system loads the s abnormal concern arrays, s task preference arrays and template abnormal information implicit representation of each first template relationship chain into the dynamic routing component. The role of the dynamic routing component is to dynamically determine the weight of each abnormal concern array based on these inputs and integrate them into a second integrated implicit representation.
[0185] Dynamic routing is a mechanism that can dynamically adjust the information flow path based on the input content. Here, it calculates the weight of each abnormal attention array based on the task preference array and the implicit representation of template abnormal information, thereby realizing the intelligent integration of the output results of different components. The dynamic routing component may use an attention mechanism or a gating mechanism to calculate the weight. For example, it can interact the implicit representation of template abnormal information with each task preference array to evaluate the relevance of each abnormal attention array to the current monitoring task and assign weights accordingly. Based on the calculated weights, the dynamic routing component weighted sums the s abnormal attention arrays of each first template relationship chain (or adopts other merging strategies) to generate the corresponding second integrated implicit representation.
[0186] In step S30C, the management and control system loads the second integrated implicit representations of the p first template relationship chains into the average dimension reduction component. The function of this component is to further process and reduce the dimensions of these integrated representations to generate the final template tunnel section implicit representation. The average dimension reduction component may first average the p second integrated implicit representations to reduce individual differences and extract common features. This averaging operation helps to generate a more representative tunnel section representation. After averaging, the component may use principal component analysis (PCA), linear discriminant analysis (LDA) or other dimension reduction techniques to further reduce the dimension of the representation. The purpose of dimensionality reduction is not only to reduce the amount of calculation, but also to remove redundant information and highlight key features. After averaging and dimension reduction processing, the average dimension reduction component outputs the final template tunnel section implicit representation. This representation integrates the information of all first template relationship chains and reflects the overall state or characteristics of the tunnel section under the current monitoring task. Through the processing flow of steps S30A, S30B and S30C, the management and control system integrates the implicit representation of template abnormal information, the implicit representation of the links of the first template relationship chain and the task preference array into the second branch network, and generates the implicit representation of the template tunnel section through the collaborative work of multiple task adaptation components, dynamic routing components and average dimension reduction components. This process not only makes full use of the expertise of multiple expert networks, but also realizes the effective integration and compression of information through dynamic routing and dimension reduction processing.
[0187] As an implementation mode, in step S30A, implicit link representations of p first template relationship chains and s task preference arrays are loaded into s task adaptation components to obtain s abnormal attention arrays of each first template relationship chain, including:
[0188] Step S30A1: extracting the implicit representation of the target tunnel section template and the implicit representation of each adjacent template connected to the target abnormal information template from the implicit representation of the link of the second branch, the second branch being a relationship chain in the p first template relationship chains, and the target tunnel section template being any abnormal information template in the second branch;
[0189] Step S30A2: stacking the implicit representation of the target tunnel section template with the s task preference arrays respectively to obtain s first stacked arrays;
[0190] Step S30A3: performing inner product calculations on the s first stacked arrays and the implicit representation of each adjacent template, respectively, to obtain the template weight of each adjacent template in the second branch chain;
[0191] Step S30A4: Based on the template weight of each adjacent template, the implicit representations of the corresponding adjacent templates are merged according to the weights to obtain s abnormal attention arrays of the second branch.
[0192] In the embodiment of the present application, step S30A involves combining the implicit representation of the links of the multiple first template relationship chains with the task preference array, and generating an abnormal focus array through the task adaptation component. This process makes full use of the professionalism of the expert network and the guidance of the task preference array, so that the system can generate more targeted focus for different types of abnormal information.
[0193] In step S30A1, the management and control system extracts key information from the implicit representation of the link of the first template relationship chain currently being processed (herein referred to as the "second branch chain" to distinguish different contexts in the step description). The key information includes the implicit representation of the target tunnel section template and the implicit representation of each adjacent template connected to the target abnormal information template.
[0194] Implicit representation of target tunnel section template: In the tunnel monitoring system, each tunnel section can be regarded as a template, and its implicit representation is an abstraction and encoding of the characteristics of the section. In step S30A1, the management and control system needs to identify the tunnel section template (i.e., the target tunnel section template) that is directly related to the target abnormal information template from the implicit representation of the link, and extract its corresponding implicit representation. This implicit representation may be a high-dimensional vector that contains a variety of feature information of the tunnel section. In addition to the target tunnel section template, the management and control system also needs to extract the implicit representation of each adjacent template connected to the target abnormal information template. These adjacent templates may represent other sections adjacent to the target tunnel section, or other entities associated with the target abnormal information (such as sensors, equipment, etc.). Their implicit representations also exist in the form of high-dimensional vectors for subsequent calculations and analysis.
[0195] For example, assuming that the second branch currently being processed is "Tunnel Section A → Water Seepage", its implicit link representation is link_rep_A=[l1,l2,...,ln]. The control system identifies the target tunnel section template as "Tunnel Section A", and its corresponding implicit representation is tunnel_rep_A=[t1,t2,...,tm] (here it is assumed that it has been extracted or mapped from link_rep_A in some way). At the same time, the system also extracts the implicit representations of the two adjacent templates connected to the target abnormal information template "water seepage", namely neighbor_rep_1=[n11,n12,...,n1p] and neighbor_rep_2=[n21,n22,...,n2q].
[0196] In step S30A2, the management and control system stacks the implicit representation of the target tunnel section template with s task preference arrays to generate s first stack arrays. These task preference arrays are pre-defined and used to guide the behavior of the task adaptation component when processing different types of abnormal information.
[0197] Each task adaptation component is associated with a task preference array, which contains a series of weights or probability values that represent the component's preference for different types of abnormal information. These preference values may be set based on domain knowledge, historical data, or expert experience. The control system concatenates the implicit representation of the target tunnel section template with each task preference array in terms of dimension (if the dimensions do not match, appropriate scaling, padding, or truncation can be performed). In this way, each task preference array generates a corresponding first stacked array.
[0198] For example, suppose there are s=3 task adaptation components, and the task preference arrays of each component are pref_arr_1, pref_arr_2, and pref_arr_3. The control system stacks the implicit representation tunnel_rep_A of the target tunnel section template with these three task preference arrays to generate three first stacking arrays:
[0199] stacked_arr_1=[tunnel_rep_A;pref_arr_1]=[t1,t2,...,tm;p11,p12,...,p1r]
[0200] stacked_arr_2=[tunnel_rep_A;pref_arr_2]=[t1,t2,...,tm;p21,p22,...,p2s]
[0201] stacked_arr_3=[tunnel_rep_A;pref_arr_3]=[t1,t2,...,tm;p31,p32,...,p3t]
[0202] Among them, pij represents the jth element in the i-th task preference array, and r, s, and t represent the lengths of the three task preference arrays respectively (note that they may not be equal).
[0203] In step S30A3, the management and control system performs inner product calculations on each first stacked array and the implicit representation of each adjacent template to obtain template weights of the adjacent templates. These weights reflect the importance or relevance of the adjacent templates in the current processing task.
[0204] The inner product is a method to measure the similarity of two vectors. It calculates the sum of the products of the corresponding elements of the two vectors. Here, the control system calculates the inner product of each first stacked array with the implicit representation of each adjacent template, and obtains a series of template weight values. In order to improve the comparability and effectiveness of the weights, the control system can normalize the calculated template weights (such as using a softmax function) so that the sum of the weights of all adjacent templates is 1. However, this step may be optional in some implementations, depending on the design of the task adaptation component and system requirements.
[0205] For example, the management and control system respectively calculates the inner product of the three first stacked arrays stacked_arr_1, stacked_arr_2 and stacked_arr_3 with the implicit representations of the two adjacent templates neighbor_rep_1 and neighbor_rep_2. Similarly, the template weights between all stacked arrays and all adjacent templates can be calculated. Step S30A4: Merge the implicit representations of adjacent templates by weight In step S30A4, the management and control system performs weighted averaging (or other forms of merging) on the implicit representations of adjacent templates according to the template weights calculated in step S30A3 to generate s abnormal attention arrays of the second branch.
[0206] The control system performs a weighted average operation on the implicit representation of each adjacent template according to the template weight of each adjacent template. This operation is performed, for example, between corresponding elements of the implicit representation, that is, the value at each element position is weighted and summed according to its corresponding weight. The vector or matrix obtained after weighted averaging is the abnormal attention array. Each task adaptation component generates an abnormal attention array, which reflects the component's attention level and recognition results for different types of abnormal information in the current processing task. For example (continuing the previous example), assuming that the template weights obtained after inner product calculation and normalization are weights_1=[w11,w12] (corresponding to the inner product result of stacked_arr_1 and two adjacent templates), the control system can perform weighted averaging on neighbor_rep_1 and neighbor_rep_2 according to these weights. Similarly, the weighted average results based on stacked_arr_2 and stacked_arr_3 can be calculated to obtain the other two abnormal attention arrays focus_arr_A2 and focus_arr_A3. Through the processing flow of steps S30A1 to S30A4, the management and control system combines the implicit representation of the link of the first template relationship chain with the task preference array, and generates multiple abnormal attention arrays through the task adaptation component. This process not only makes full use of the guidance of the task preference array, but also realizes the effective integration and refinement of the implicit representation of adjacent templates through operations such as inner product calculation and weighted average. The finally generated abnormal attention array provides an important basis for subsequent dynamic routing and dimensionality reduction processing, and also lays a solid foundation for the abnormal detection and early warning functions of the unmanned intelligent monitoring system of the tunnel.
[0207] As an implementation mode, step S30B, loading the s exception concern arrays, s task preference arrays and template exception information implicit representation of each first template relationship chain into the dynamic routing component, and obtaining the second integrated implicit representation of each first template relationship chain, includes:
[0208] Step S30B1: for each first template relationship chain, perform inner product calculations on the s task preference arrays and the implicit representation of the template abnormal information to obtain the weight of each abnormal concern array in each first template relationship chain;
[0209] Step S30B2: For each first template relationship chain, based on the weights of the s abnormal concern arrays, the corresponding s abnormal concern arrays are merged according to the weights to obtain a second integrated implicit representation of each first template relationship chain.
[0210] In the embodiment of the present application, step S30B is a key link in the dynamic routing component processing flow, which is responsible for intelligently integrating multiple abnormal concern arrays (generated by the task adaptation component) of each first template relationship chain according to the task preference array and the template abnormal information implicit representation to generate a more representative second integrated implicit representation. This process not only reflects the system's comprehensive processing capabilities for multi-source information, but also achieves effective focus on key information through the dynamic routing mechanism.
[0211] In step S30B1, the management and control system calculates the inner product of the corresponding s task preference arrays and the implicit representation of the template abnormal information for each first template relationship chain to obtain the weight of each abnormal attention array. This calculation process is intended to determine the relative importance of different abnormal attention arrays based on the needs of the current monitoring task (reflected by the implicit representation of the template abnormal information) and the professionalism of the component (represented by the task preference array).
[0212] In step S30B2, the management and control system performs weighted averaging (or other forms of merging) on the s abnormal attention arrays of each first template relationship chain according to the weights calculated in step S30B1 to generate the corresponding second integrated implicit representation. This process realizes the intelligent integration of multiple abnormal attention arrays, so that the generated representation is more in line with the needs of the current monitoring task. Weighted averaging is a commonly used data merging method, which calculates the contribution of each data point according to its weight, and accumulates the contribution of all data points to obtain the final result. Here, the management and control system regards each abnormal attention array as a data point, and uses its corresponding weight as a contribution factor for weighted averaging. The result after weighted averaging is the second integrated implicit representation, which integrates the information of multiple abnormal attention arrays and performs intelligent focusing based on the guidance of task preferences and implicit representation of template abnormal information. This representation not only contains rich abnormal information but also reflects the system's understanding and response to the current monitoring task. For example (continuing the previous example), assuming that the weights calculated in step S30B1 are weight_{A1}=w1, weight_{A2}=w2, and weight_{A3}=w3, the management and control system can calculate the second integrated implicit representation by weighted average of focus_arr_A1, focus_arr_A2, and focus_arr_A3 according to these weights. Through the processing flow of the above steps S30B1 and S30B2, the management and control system intelligently integrates the multiple abnormal focus arrays of each first template relationship chain according to the task preference array and the template abnormal information implicit representation. This process not only makes full use of the expertise of multiple task adaptation components, but also realizes the effective focus on key information through the dynamic routing mechanism. The final generated second integrated implicit representation not only contains rich abnormal information, but also reflects the system's understanding and response to the current monitoring task, which provides strong support for the subsequent average dimensionality reduction processing and the generation of the final template tunnel section implicit representation.
[0213] As another implementation, step S31, loading the implicit link representation of each second template relationship chain into the first fusion component to obtain the semantic implicit representation of each second template relationship chain, includes:
[0214] Step S31a: extracting the implicit representation of the target abnormal information template and the implicit representation of each adjacent template connected to the target abnormal information template from the implicit representation of the link of the third branch, the third branch being a relationship chain in the q second template relationship chains, and the target abnormal information template being any abnormal information template in the third branch;
[0215] Step S31b: performing inner product calculations on the implicit representation of the target abnormal information template and the implicit representation of each adjacent template, respectively, to obtain the template weight of each adjacent template in the third branch chain;
[0216] Step S31c: Based on the template weight of each adjacent template, the implicit representations of the corresponding adjacent templates are merged according to the weights to obtain the semantic implicit representation of the third branch.
[0217] In the embodiment of the present application, step S31 is a key link in the first branch network processing flow, and its implementation methods are various. The above another implementation method focuses on generating the semantic implicit representation of each second template relationship chain through refined feature extraction and weight allocation.
[0218] In step S31a, the management and control system focuses on the second template relationship chain currently being processed (herein referred to as the "third branch" to distinguish different contexts) and extracts key information from it. The key information includes the implicit representation of the target abnormal information template and the implicit representation of each adjacent template connected to the target abnormal information template.
[0219] Implicit representation of target anomaly information template: The target anomaly information template refers to the template that represents the anomaly information in the current third branch, such as "water seepage", "cracks", etc. Its implicit representation is a high-dimensional vector or matrix that contains multiple feature descriptions of the anomaly information. These features may include the type, severity, frequency of occurrence, scope of impact, etc. of the anomaly, which are pre-defined and encoded through historical monitoring data and domain knowledge.
[0220] Adjacent templates refer to templates that are directly connected to the target anomaly information template in the relationship chain. These templates may represent entities such as tunnel sections, equipment, timestamps, etc. that are associated with the anomaly information. Each adjacent template also has a corresponding implicit representation that describes its unique attributes and status.
[0221] For example, assuming that the third branch currently being processed is "water seepage → tunnel segment A", where "water seepage" is the target anomaly information template, its implicit representation can be assumed to be anomaly_rep_leak=[l1,l2,...,ln]. The adjacent templates connected to the anomaly information template may include "tunnel segment A" (representing the specific tunnel section where water seepage occurs) and other possible adjacent entities (such as "May 1, 2023" represents the timestamp of water seepage). The implicit representations of these adjacent templates are tunnel_seg_rep_A=[t1,t2,...,tm] and timestamp_rep=[ts1,ts2,...,tsp], etc.
[0222] In step S31b, the control system uses inner product calculation to evaluate the correlation between the target abnormal information template and each adjacent template, and assigns a weight to each adjacent template accordingly. This weight reflects the importance or contribution of the adjacent template in describing the target abnormal information.
[0223] In step S31c, the control system performs weighted averaging or other forms of merging operations on the implicit representations of adjacent templates according to the template weights calculated in step S31b to generate the semantic implicit representation of the current third branch. This representation not only contains the feature description of the target abnormal information template, but also integrates the information of the adjacent templates associated with it, thereby forming a more comprehensive and in-depth semantic understanding of the abnormal information.
[0224] Weighted average is a commonly used data merging method, which calculates the contribution of each data point according to its weight, and accumulates the contribution of all data points to obtain the final result. Here, the management and control system uses the weight calculated in step S31b as the contribution factor to perform weighted average on the implicit representations of adjacent templates. The result after weighted average is the semantic implicit representation of the current third branch. This representation integrates the information of the target anomaly information template and its adjacent templates, reflecting the complex relationship and interaction between them. In the tunnel monitoring system, such semantic implicit representation helps to more accurately identify the type of anomaly, assess the scope of anomaly impact, and predict the development trend of anomalies.
[0225] Through the processing flow of steps S31a to S31c, the management and control system extracts rich semantic information from each second template relationship chain and generates corresponding semantic implicit representation. This process not only relies on refined feature extraction technology but also makes full use of mathematical tools such as inner product calculation and weighted average to achieve effective integration of multi-source information.
[0226] As an implementation mode, step S40, debugging the initial tunnel monitoring network according to the template abnormal information implicit representation and the template tunnel section implicit representation to obtain the target tunnel monitoring network, includes:
[0227] Step S41: determining the difference between the implicit representation of the template abnormal information and the implicit representation of the template tunnel section to obtain a target error;
[0228] Step S42: Optimize the network parameters of the initial tunnel monitoring network according to the target error to obtain the target tunnel monitoring network.
[0229] In the embodiment of the present application, step S40 is a core link in the neural network training process, which involves debugging and optimizing the initial tunnel monitoring network by using the difference between the implicit representation of the template abnormal information and the implicit representation of the template tunnel section, thereby generating a target tunnel monitoring network with better performance. This process is not only an important improvement in the adaptability and accuracy of the network model, but also a key step in achieving the goal of intelligent tunnel monitoring.
[0230] In step S41, the primary task of the control system is to quantify the difference between the implicit representation of the template abnormal information and the implicit representation of the template tunnel section. This difference reflects the accuracy of the network in understanding the relationship between the abnormal information and the tunnel section in the tunnel monitoring data, and is an important indicator for evaluating network performance. In order to quantify this difference, the control system uses a loss function (or cost function) to calculate the target error.
[0231] Here, cross entropy is used as an example of a loss function. Cross entropy is a way to measure the difference between two probability distributions, and it is particularly suitable for multi-category predictions in classification problems. Although in this example, implicit representations (rather than direct probability distributions) are processed, the implicit representations can be converted to probability form in some way (such as softmax function), or the concept of cross entropy can be extended to more general difference measures. In order to calculate cross entropy, the control system first needs to determine a "real" or "expected" implicit representation of the template tunnel section (although in actual training, this "real" representation may not exist, but is replaced by some standard or target representation generated internally by the system). The system then compares the implicit representation of the template abnormal information with the implicit representation of the template tunnel section (whether it is the actual output or the target output), and calculates the difference between them through the cross entropy formula.
[0232] The calculated cross entropy value is the target error, which reflects the difference between the implicit representation of the template anomaly information and the implicit representation of the template tunnel section. The smaller the error, the more accurate the network's understanding of the relationship between the anomaly information and the tunnel section; the larger the error, the more the network needs to be optimized to improve its performance.
[0233] In step S42, the control system uses the target error calculated in step S41 to guide the optimization process of the initial tunnel monitoring network. This process involves, for example, the application of optimization techniques such as back propagation and gradient descent. Back propagation is an effective method for training neural networks. It guides the update direction of parameters by calculating the gradient of the loss function with respect to the network parameters. In step S42, the control system first calculates the gradient according to the target error, which indicates how to adjust the network parameters to reduce the error. After obtaining the gradient, the control system uses the gradient descent method to update the network parameters. Specifically, it moves each parameter in the opposite direction of its gradient by a small step (the step size is controlled by the learning rate) to reduce the error. This process is repeated for multiple iterations until the error converges to an acceptable range or reaches a preset upper limit of the number of iterations. In each iteration, the control system updates the network parameters such as weights and biases based on the latest gradient information. The optimization of these parameters enables the network to better capture the relationship characteristics between abnormal information and tunnel sections, thereby improving its monitoring performance. For example (continuing the previous example), suppose that the target error calculated after the first iteration is large, and the control system calculates the gradients of each network parameter through the back-propagation algorithm. Then, it uses the gradient descent method to update the weights and biases of the network based on these gradients. After multiple iterations, as the target error gradually decreases, the performance of the initial tunnel monitoring network will be improved, and the target tunnel monitoring network will eventually be generated. This network can more accurately process abnormal information in tunnel monitoring data and generate effective early warning information, providing strong protection for the safe operation of the tunnel. Through the processing flow of the above steps S41 and S42, the control system uses the difference between the implicit representation of the template abnormal information and the implicit representation of the template tunnel section to debug and optimize the initial tunnel monitoring network.
[0234] Based on the above steps, the target tunnel monitoring network is obtained through training. The following describes an unmanned intelligent monitoring method for tunnels based on the target tunnel monitoring network, which may specifically include the following steps:
[0235] Step S100: obtaining p first target relationship chains and q second target relationship chains, where p and q are both natural numbers not equal to 0;
[0236] Step S200: extracting the implicit link representation of each first target relationship chain and the implicit link representation of each second target relationship chain;
[0237] Step S300: taking the implicit representations of the links of the p first target relationship chains and the implicit representations of the links of the q second target relationship chains as the execution data of the target tunnel monitoring network, and obtaining the implicit representation of the target abnormal information and the implicit representation of the target tunnel section;
[0238] Step S400: determining target tunnel abnormality warning information according to the target abnormality information implicit representation and the target tunnel section implicit representation.
[0239] As before, the target tunnel monitoring network is a neural network obtained by debugging the initial tunnel monitoring network with the template exception information implicit representation and the template tunnel section implicit representation as debugging templates. The template exception information implicit representation is obtained by processing the implicit representation of the links of q second template relationship chains based on the first branch network in the initial tunnel monitoring network. The template tunnel section implicit representation is obtained by processing the implicit representation of the template exception information or the first enhanced implicit representation and the implicit representation of the links of p first template relationship chains based on the second branch network in the initial tunnel monitoring network. The first enhanced implicit representation is used to guide the second branch network to make adaptive adjustments according to the results of the first branch network.
[0240] In the embodiment of the present application, after the complex training process described above, the management and control system successfully built an efficient and accurate target tunnel monitoring network. This network can automatically analyze and identify potential abnormal information based on the input tunnel monitoring data, and then generate corresponding abnormal warning information.
[0241] In step S100, p first target relationship chains and q second target relationship chains are obtained. In the actual operation process of unmanned intelligent monitoring of tunnels, the management and control system first needs to obtain tunnel monitoring data in real time or regularly, and extract key relationship chain information from it. These relationship chain information includes two types: first target relationship chain and second target relationship chain. First target relationship chain: This type of relationship chain represents the historical or real-time inclusion relationship of tunnel sections with specific abnormal information starting from the tunnel section. For example, "A tunnel section → current water seepage" constitutes a first target relationship chain, indicating that there is currently water seepage abnormality in A tunnel section. In practical applications, the management and control system may discover this type of relationship chain through real-time sensor data, video monitoring or manual inspection. Second target relationship chain: In contrast to the first target relationship chain, the second target relationship chain represents the historical or real-time inclusion relationship of the abnormal information in which tunnel sections the abnormal information appears. For example, "water seepage → A tunnel section, B tunnel section" constitutes two second target relationship chains, indicating that the water seepage abnormality appears simultaneously in A tunnel section and B tunnel section. The acquisition of this type of relationship chain also relies on the comprehensive analysis of real-time monitoring data and historical records.
[0242] In step S100, the management and control system obtains p first target relationship chains and q second target relationship chains from the tunnel monitoring system through its built-in data acquisition module or interface. The number of these relationship chains (p and q) will vary according to the actual monitoring requirements and the size of the tunnel, but they are all natural numbers that are not 0. For example, suppose that in a real-time monitoring, the management and control system finds the following relationship chain information:
[0243] The first target relationship chain (p=3): "Tunnel section A → current water seepage"; "Tunnel section B → crack expansion"; "Tunnel section C → temperature anomaly";
[0244] The second target relationship chain (q=2): "water seepage → tunnel section A, tunnel section D"; "cracks → tunnel section B"; these relationship chains constitute the basic data for subsequent processing.
[0245] After obtaining the relationship chain information, the management and control system converts these relationship chains into a machine-understandable format, namely, implicit link representation. Implicit link representation is an abstract encoding of entities and their relationships in the relationship chain, for example, in the form of a high-dimensional vector or matrix. In step S200, the management and control system converts each first target relationship chain and second target relationship chain into a corresponding implicit link representation using predefined encoding rules or machine learning models (such as embedding layers, autoencoders, etc.). These representations not only contain key information in the relationship chain (such as tunnel section names, abnormal information types, etc.), but also reflect the similarities and differences between them through positional relationships in the vector space.
[0246] For simple systems, the control system may use a rule-based encoding method to map the tunnel section name and the abnormal information type to a fixed vector index or binary code. However, this method often has difficulty capturing complex relationships and contextual information. A more advanced method is to use machine learning models to automatically generate implicit link representations. For example, an embedding layer can be used to learn distributed representations (embeddings) of tunnel sections and abnormal information, which are optimized during the training process to capture the semantic relationship between entities. For example (continuing the previous example), assume that the control system uses an embedding layer to generate implicit link representations. For the first target relationship chain "A tunnel section → current water seepage", its implicit link representation may be a long vector that concatenates two embedded vectors: the embedding vector of "A tunnel section": [a1, a2, ..., an]; the embedding vector of "current water seepage": [w1, w2, ..., wm]; the concatenated implicit link representation is: [a1, a2, ..., an, w1, w2, ..., wm].
[0247] Similarly, the implicit representation of the link of the second target relation chain can also be generated by concatenating the embedding vectors of the corresponding entities.
[0248] In step S300, the management and control system uses the implicit representation of the link generated in step S200 as input data of the target tunnel monitoring network. The target tunnel monitoring network is a neural network model that has undergone a complex training process, which can automatically process input data and generate corresponding output representations. The target tunnel monitoring network may contain multiple levels of structures, such as convolutional layers, pooling layers, fully connected layers, and specially designed fusion components and routing components. These components work together to extract key features from the input data and generate the final output representation. In the target tunnel monitoring network, the output includes, for example, two types of implicit representations: implicit representation of target abnormal information and implicit representation of target tunnel sections. The former reflects the comprehensive characteristics of the currently monitored abnormal information; the latter describes the comprehensive status of the tunnel sections associated with these abnormal information.
[0249] The management and control system passes the implicit link representations of the p first target relationship chains and the implicit link representations of the q second target relationship chains to different input channels of the target tunnel monitoring network (if the network is designed to support multi-channel input). The network gradually extracts and transforms input features through multi-layer processing (such as convolution, pooling, nonlinear activation, etc.), and finally generates implicit representations of target abnormality information and target tunnel sections. For example (continuing the previous example), assume that the target tunnel monitoring network has been optimized according to the training process in the previous article. The management and control system passes the implicit link representation generated in step S200 as input to the network, and waits for the network to output implicit representations of target abnormality information and target tunnel sections. These representations may be in the form of high-dimensional vectors or matrices, containing rich feature information and contextual relationships.
[0250] In step S400, the management and control system determines the final abnormal warning information according to the target abnormal information implicit representation and the target tunnel section implicit representation generated in step S300. This process may involve further feature analysis, pattern matching or classification decision-making operations.
[0251] The control system first performs an in-depth feature analysis of the implicit representation of the target abnormal information and the implicit representation of the target tunnel section. This may include calculating statistics of specific features (such as mean, variance, maximum value, etc.), performing principal component analysis (PCA) to reduce the dimension and extract the main feature direction, or using clustering algorithms to discover potential abnormal patterns. Based on the feature analysis, the control system attempts to match the current feature representation with a preset abnormal pattern library. The abnormal pattern library may contain feature templates or prototypes of various abnormal situations that have been identified and classified in history. The degree of matching is evaluated by calculating indicators such as similarity scores or distance metrics, and based on this, it is determined whether the current monitoring data belongs to a known abnormal category. If the system adopts a classification-based approach to handle the anomaly detection problem, the implicit representation of the target abnormal information and the implicit representation of the target tunnel section can be directly used as the input of the classifier, and the abnormal warning information can be determined based on the output of the classifier. The classifier may be a support vector machine (SVM), random forest, neural network or other type of machine learning model. Once the abnormal category is determined or the pattern match is successful, the control system can generate the corresponding abnormal warning information. This information may include the type of anomaly, location, scope of impact, severity, and recommended countermeasures, etc. The early warning information will be passed to the tunnel management department or relevant responsible persons so that necessary actions can be taken in time to eliminate safety hazards.
[0252] For example, suppose that after feature analysis and pattern matching, the control system determines that there is water seepage anomaly in the current monitoring data and it mainly affects tunnel sections A and D. The system generates the following abnormal warning information: abnormal type: water seepage; location: tunnel section A, tunnel section D; affected range: section XX to YY meters in tunnel section A, section ZZ to WW meters in tunnel section D; severity: medium (based on a comprehensive evaluation of historical data and current monitoring indicators); recommended measures: immediately launch the emergency plan, organize personnel to go to the site for investigation and take measures to prevent the spread of water seepage; at the same time, strengthen the monitoring and inspection of relevant areas to ensure safety. This warning information will be promptly transmitted to the tunnel management department so that they can quickly respond and handle abnormal situations to ensure the safe operation of the tunnel.
[0253] As an implementation mode, step S400, determining target tunnel abnormality warning information according to the target abnormality information implicit representation and the target tunnel section implicit representation, includes:
[0254] Step S410: obtaining a dot product between the implicit representation of target abnormal information and the implicit representation of target tunnel section, and obtaining a commonality measurement value between the implicit representation of target abnormal information and the implicit representation of target tunnel section;
[0255] Step S420: determining target tunnel abnormality warning information from the reference abnormality warning information set according to the commonality metric value.
[0256] In the embodiment of the present application, step S400 is the key link in generating abnormal warning information. It determines the current tunnel monitoring status based on the implicit representation of the target abnormal information output by the target tunnel monitoring network and the implicit representation of the target tunnel section, and generates corresponding abnormal warning information accordingly.
[0257] In step S410, the management and control system first calculates the dot product between the implicit representation of the target abnormal information and the implicit representation of the target tunnel section. As a mathematical operation, the dot product can measure the similarity or "alignment" degree of two vectors in direction. Here, the calculation result of the dot product reflects the correlation strength or common characteristics between the target abnormal information and the target tunnel section, and provides a basis for the subsequent selection of appropriate warning information from the reference abnormal warning information set. The calculated dot product value is the commonality metric value, which reflects the similarity or correlation strength between the implicit representation of the target abnormal information and the implicit representation of the target tunnel section. The larger the dot product value, the more obvious the common characteristics between the two, that is, the closer the correlation between the currently monitored abnormal information and the specific tunnel section.
[0258] In step S420, the management and control system uses the commonality metric calculated in step S410 to search for the warning information that best matches the current monitoring status in the predefined reference abnormality warning information set. The reference abnormality warning information set is a database or knowledge base that contains information such as various abnormality types, occurrence locations, impact ranges, and recommended countermeasures, and is used to provide support for the system to generate specific warning information. The reference abnormality warning information set may exist in various forms such as database tables, files, and memory data structures, each of which contains a detailed description of one or more abnormality warning information. These warning information may be pre-defined and encoded based on historical monitoring data, domain knowledge, or expert experience. The management and control system searches for matching warning information in the reference abnormality warning information set based on the commonality metric.
[0259] Matching strategies may include the following:
[0260] One or more thresholds are set and the commonality metric values are compared with these thresholds to determine the category or level of the warning information. For example, if the commonality metric value is greater than a certain threshold, it is considered that the current monitoring state is highly correlated with a certain type of anomaly in the set, and the corresponding warning information is selected.
[0261] Calculate the similarity or distance between the commonality metric and the metric corresponding to each warning message in the set (which may be pre-calculated and stored in some way), and sort the warning messages from high to low according to the similarity. Select the warning message with the highest similarity as the output. Define a series of rules to map the relationship between the commonality metric and the warning message. When the commonality metric meets the conditions of a rule, the warning message generation logic corresponding to the rule is triggered.
[0262] Once the appropriate warning information is found, the control system formats it into a text or instruction form that is easy to understand and operate, and prepares to pass it to the tunnel management department or relevant responsible persons. The warning information may include the type of anomaly, the location of occurrence, the scope of impact, the severity, the recommended response measures, etc. For example (continuing the previous example), assume that the reference abnormal warning information set contains warning information of multiple types of anomalies, and each piece of information is associated with a specific commonality metric value range or threshold. The control system compares the commonality metric value (0.44) calculated in step S410 with the threshold in the set and finds that the value falls within the threshold range of "water seepage anomaly". Therefore, the system selects the warning information corresponding to the anomaly type as output: the specific tunnel section inferred from the implicit representation of the target tunnel section (which can be further processed to determine the specific location), the scope of impact-severity presumed based on the preliminary estimate or historical data, the judgment based on the size of the commonality metric value and the definition in the set (for example, the larger the commonality metric value, the more serious the anomaly), the standard response measures or emergency plans formulated for the water seepage anomaly, and the warning information finally generated will describe in detail the currently monitored water seepage anomaly and its related information, providing timely and accurate warning support for the tunnel management department. Through this process, the tunnel's unmanned intelligent monitoring system can achieve continuous monitoring and rapid response to the tunnel status, effectively ensuring the safe operation of the tunnel.
[0263] Figure 2 A schematic diagram of a hardware entity of a management and control system provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the hardware entity of the management and control system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.
[0264] The above is only an implementation method of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.
Claims
1. A method for unmanned intelligent monitoring of tunnels, characterized in that: include: Obtain p first target relationship chains and q second target relationship chains, where p and q are both natural numbers not equal to 0; Extracting an implicit link representation of each of the first target relationship chains and an implicit link representation of each of the second target relationship chains; The implicit representations of the links of the p first target relationship chains and the implicit representations of the links of the q second target relationship chains are used as the execution data of the target tunnel monitoring network to obtain the implicit representation of the target abnormality information and the implicit representation of the target tunnel section. The target tunnel monitoring network is a neural network obtained by debugging the initial tunnel monitoring network with the implicit representation of the template abnormality information and the implicit representation of the template tunnel section as debugging templates. The implicit representation of the template abnormality information is obtained by processing the implicit representations of the links of the q second template relationship chains based on the first branch network in the initial tunnel monitoring network. The implicit representation of the template tunnel section is obtained by processing the implicit representation of the template abnormality information or the first enhanced implicit representation and the implicit representations of the links of the p first template relationship chains based on the second branch network in the initial tunnel monitoring network. The first enhanced implicit representation is used to guide the second branch network to perform adaptive adjustment according to the result of the first branch network. Determining target tunnel abnormality warning information according to the target abnormality information implicit representation and the target tunnel section implicit representation; The debugging process of the target tunnel monitoring network includes: Obtain p first template relationship chains and q second template relationship chains, each of the first template relationship chains is used to represent a historical inclusion relationship between the tunnel section template and the abnormal information template, and each of the second template relationship chains is used to represent a historical inclusion relationship between the abnormal information template and the abnormal information template in the tunnel section template, and both p and q are natural numbers that are not 0; Extracting an implicit link representation of each first template relationship chain and an implicit link representation of each second template relationship chain; Using the implicit representations of the links of the q second template relationship chains as the execution data of the first branch network in the initial tunnel monitoring network, obtaining the implicit representation of the template abnormality information, and using the implicit representation of the template abnormality information or the first enhanced implicit representation and the implicit representations of the links of the p first template relationship chains as the execution data of the second branch network in the initial tunnel monitoring network, obtaining the implicit representation of the template tunnel section, wherein the first enhanced implicit representation is used to guide the second branch network to perform adaptive adjustment according to the result of the first branch network; The initial tunnel monitoring network is debugged according to the implicit representation of the template abnormality information and the implicit representation of the template tunnel section to obtain the target tunnel monitoring network, which is used to process the implicit representation of the links of p first target relationship chains and the implicit representation of the links of q second target relationship chains to obtain the implicit representation of the target abnormality information and the implicit representation of the target tunnel section, which are used to determine the abnormality warning information of the target tunnel.
2. The method according to claim 1, characterized in that Determining target tunnel abnormality warning information according to the target abnormality information implicit representation and the target tunnel section implicit representation includes: Obtaining a dot product between the implicit representation of the target abnormal information and the implicit representation of the target tunnel section, and obtaining a commonality metric value between the implicit representation of the target abnormal information and the implicit representation of the target tunnel section; The target tunnel abnormality warning information is determined from the reference abnormality warning information set according to the commonality metric value.
3. The method according to claim 1, characterized in that The first branch network includes a first fusion component and a first weight distribution component; the implicit representation of the links of the q second template relationship chains is used as the execution data of the first branch network in the initial tunnel monitoring network to obtain the implicit representation of the template abnormal information, including: Loading the implicit link representation of each second template relationship chain to the first fusion component respectively, obtaining the semantic implicit representation of each second template relationship chain, wherein the semantic implicit representation of each second template relationship chain is used to represent the semantics of the abnormal information inclusion relationship in the corresponding second template relationship chain; Using the semantic implicit representations of the q second template relationship chains as execution data of the first weight allocation component, to obtain a relationship chain weight of each second template relationship chain; The semantic implicit representations of the q second template relationship chains and the relationship chain weights are combined according to the weights by the first weight allocation component to obtain an implicit representation of the template exception information.
4. The method according to claim 3, characterized in that Before loading the implicit link representation of each second template relationship chain into the first fusion component to obtain the semantic implicit representation of each second template relationship chain, the method further includes: Acquire a second enhanced implicit representation, where the second enhanced implicit representation is used to guide the first branch network to perform adaptive adjustment according to a result of the second branch network; Loading the implicit link representation of each second template relationship chain into the first fusion component to obtain the semantic implicit representation of each second template relationship chain includes: The second enhanced implicit representation is stacked with the link implicit representation of each second template relationship chain to obtain an integrated implicit representation of each second template relationship chain; Loading the integrated implicit representation of each second template relationship chain into the first fusion component respectively to obtain the semantic implicit representation of each second template relationship chain; The step of combining the semantic implicit representations of the q second template relationship chains and the relationship chain weights according to the first weight distribution component to obtain the template exception information implicit representation includes: According to the first weight allocation component, the semantic implicit representations of the q second template relationship chains and the relationship chain weights are combined according to the weights to obtain a target integrated implicit representation; stacking the target integrated implicit representation and the second enhanced implicit representation to obtain an implicit representation of template abnormal information; The second branch network includes a second fusion component and a second weight allocation component; the implicit representation of the template abnormal information or the first enhanced implicit representation, and the implicit representation of the links of p first template relationship chains are used as the execution data of the second branch network in the initial tunnel monitoring network to obtain the implicit representation of the template tunnel section, including: stacking the first enhanced implicit representation with the link implicit representation of each of the first template relationship chains respectively to obtain a first integrated implicit representation of each of the first template relationship chains; Loading the first integrated implicit representation of each of the first template relationship chains into the second fusion component respectively, to obtain the semantic implicit representation of each of the first template relationship chains, wherein the semantic implicit representation of each of the first template relationship chains is used to represent the semantics of the tunnel section inclusion relationship in the corresponding first template relationship chain; Using the semantic implicit representation of each first template relationship chain as execution data of the second weight allocation component to obtain the relationship chain weight of each first template relationship chain; According to the second weight allocation component, the semantic implicit representations of the p first template relationship chains and the relationship chain weights are combined according to the weights to obtain a first focused implicit representation; The first focused implicit representation and the first enhanced implicit representation are stacked to obtain an implicit representation of a template tunnel section.
5. The method according to claim 4, characterized in that The first integrated implicit representation of each of the first template relationship chains is loaded into the second fusion component to obtain the semantic implicit representation of each of the first template relationship chains, including: In a first integrated implicit representation of a first branch chain, an implicit representation of a target tunnel section template and an implicit representation of each adjacent template connected to the target tunnel section template are extracted, wherein the first branch chain is a relationship chain in the p first template relationship chains, and the target tunnel section template is any tunnel section template in the first branch chain; Performing inner product calculations on the implicit representation of the target tunnel section template and the implicit representation of each of the adjacent templates to obtain a template weight of each of the adjacent templates in the first branch chain; According to the template weight of each of the adjacent templates, the implicit representations corresponding to the adjacent templates are merged according to the weights to obtain the semantic implicit representation of the first branch.
6. The method according to claim 4, characterized in that The method further comprises: determining a first error between the template tunnel section implicit representation and the second enhanced implicit representation; optimizing the first enhanced implicit representation and the second enhanced implicit representation according to the first error; determining a second error between the template anomaly information implicit representation and the first enhanced implicit representation; The second enhanced implicit representation and the first enhanced implicit representation are optimized according to the second error.
7. The method according to claim 3, characterized in that The second branch network includes s task adaptation components, a dynamic routing component, and an average dimension reduction component, s≥1; the template abnormal information implicit representation or the first enhanced implicit representation, and the p implicit representations of the links of the first template relationship chain are used as the execution data of the second branch network in the initial tunnel monitoring network to obtain the implicit representation of the template tunnel section, including: Loading the implicit link representation of each of the first template relationship chains and s task preference arrays into the s task adaptation components, obtaining s abnormal concern arrays of each of the first template relationship chains, each of the task adaptation components outputting one abnormal concern array, each of the task preference arrays being used to characterize the preference of the corresponding task adaptation component for the abnormal information template; Loading the s exception concern arrays, the s task preference arrays and the template exception information implicit representation of each of the first template relationship chains into the dynamic routing component to obtain a second integrated implicit representation of each of the first template relationship chains; The second integrated implicit representations of the p first template relationship chains are loaded into the average dimension reduction component to obtain an implicit representation of the template tunnel section.
8. The method according to claim 7, characterized in that Loading the implicit link representations of the p first template relationship chains and the s task preference arrays into the s task adaptation components, and obtaining s abnormal attention arrays of each of the first template relationship chains, including: In the implicit representation of the link of the second branch, the implicit representation of the target tunnel section template and the implicit representation of each adjacent template connected to the target abnormal information template are extracted, the second branch is a relationship chain in the p first template relationship chains, and the target tunnel section template is any abnormal information template in the second branch; stacking the implicit representation of the target tunnel section template with the s task preference arrays respectively to obtain s first stacked arrays; Performing inner product calculations on the s first stacked arrays and the implicit representation of each of the adjacent templates, respectively, to obtain a template weight of each of the adjacent templates in the second branch chain; According to the template weight of each of the adjacent templates, implicit representations corresponding to the adjacent templates are merged according to the weights to obtain s abnormal attention arrays of the second branch chain; The s exception concern arrays, the s task preference arrays and the template exception information implicit representation of each of the first template relationship chains are loaded into the dynamic routing component to obtain a second integrated implicit representation of each of the first template relationship chains, including: For each of the first template relationship chains, inner product calculations are performed on the s task preference arrays and the implicit representation of the template exception information to obtain the weight of each of the exception attention arrays in each of the first template relationship chains; For each of the first template relationship chains, the corresponding s exception attention arrays are merged according to the weights of the s exception attention arrays to obtain a second integrated implicit representation of each of the first template relationship chains.
9. A management and control system, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the program, the steps in the method according to any one of claims 1 to 8 are implemented.
