Bridge stress monitoring method and system driven by Internet of Things

Abstract

The invention provides a bridge stress monitoring method and system driven by the Internet of Things, and relates to the technical field of the Internet of Things, and the method comprises the steps: constructing a digital twinborn body for bridge stress monitoring based on the sensing data flow of the Internet of Things; monitoring a current working intention generated when a user interacts with the digital twin and the uncertainty of the intention; deducing a current global optimal reference monitoring intention of the bridge in parallel; solving a guide parameter through an optimization decision model by taking the deviation between the current working intention and the reference monitoring intention as a target and cooperatively considering the uncertainty; wherein the solving target of the optimization decision model takes account of the autonomy of the user and the reliability of the monitoring decision; and driving the digital twin to enter and present a guide mode adaptive to the current cognitive state of the user according to the guide parameters. According to the invention, the intelligence and accuracy of bridge stress monitoring are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) technology, and in particular to an IoT-driven method and system for monitoring bridge stress. Background Technology

[0002] With the rapid development of IoT technology, bridge stress monitoring is gradually shifting from traditional manual inspection to intelligent and real-time monitoring. IoT sensors can collect multi-dimensional data such as bridge stress, deformation, and vibration in real time, providing a data foundation for building digital twins of bridges.

[0003] However, existing monitoring systems are mostly based on static data analysis, making it difficult to dynamically adapt to the real-time interactive needs and complex decision-making scenarios of users (such as engineers or managers). When users interact with the monitoring system, their working intentions often deviate from the optimal monitoring requirements of the actual stress state of the bridge due to differences in subjective cognition or information asymmetry, resulting in reduced monitoring efficiency or decision-making errors.

[0004] Furthermore, existing technologies lack quantitative analysis of the uncertainty of user intent and fail to achieve an effective balance between user autonomy and system reliability.

[0005] Therefore, there is an urgent need for an IoT-based method for monitoring bridge stress. This method should construct a digital twin, combining dynamic perception of user intent with a globally optimal monitoring strategy to generate adaptive guidance parameters, thereby improving the accuracy and intelligence of monitoring. This approach must ensure the scientific validity and reliability of monitoring decisions while also considering the user experience, providing an innovative solution for bridge safety management. Summary of the Invention

[0006] The purpose of this invention is to provide an Internet of Things-driven method and system for monitoring bridge stress, in order to solve the problems mentioned in the background art.

[0007] In a first aspect, an embodiment of the present invention provides an Internet of Things-driven bridge stress monitoring method, comprising: Based on IoT sensor data streams, a digital twin for bridge stress monitoring is constructed. Monitor the current working intention generated by the user when interacting with the digital twin and the uncertainty of that intention; in parallel, derive the current globally optimal benchmark monitoring intention of the bridge; With the goal of bridging the discrepancy between the current working intention and the baseline monitoring intention, and taking into account the uncertainty, the guiding parameters are solved by optimizing the decision model; wherein, the solution objective of the optimized decision model takes into account both user autonomy and monitoring decision reliability. Based on the guidance parameters, the digital twin is driven to enter and present a guidance mode that adapts to the user's current cognitive state.

[0008] Optionally, the construction of a digital twin for bridge stress monitoring based on IoT sensor data streams includes: The IoT sensor data stream is preprocessed; The preprocessed IoT sensor data stream is input into the digital twin engine to construct a digital twin for bridge stress monitoring.

[0009] Optionally, the monitoring of the user's current working intent and the uncertainty of that intent during interaction with the digital twin includes: High-dimensional behavioral features are extracted from user interactions and mapped to behavioral semantic vectors. In the preset intent representation vector space, calculate the phylogenetic matching degree between the behavior semantic vector and each intent representation vector; The intent represented by the intent representation vector with the maximum genealogical matching degree is determined as the current working intent; The uncertainty of the current working intention is quantified based on the spectral residual between the maximum spectral matching degree and the ideal matching degree.

[0010] Optionally, the derivation of the bridge's current globally optimal benchmark monitoring intent includes: Based on users’ historical monitoring trajectories and the current distribution of monitoring tasks within the system, a heat map of monitoring resource coverage and a task gap distribution map are constructed. Based on the monitoring resource coverage heat map and task gap distribution map, identify the current monitoring blind spots and load weaknesses; The blind spots and weak points, along with the real-time mechanical state of the bridge and the preset monitoring procedures, are input into a multi-agent task allocation simulation environment for game theory simulation. The key task derived from the deduction that can maximize the overall monitoring effectiveness of the system is established as the benchmark monitoring objective.

[0011] Optionally, the step of bridging the discrepancy between the current working intention and the baseline monitoring intention, and taking into account the uncertainty, involves solving for the guiding parameters through an optimized decision model, including: The deviation between the current working intention and the benchmark monitoring intention, and the uncertainty are input into the optimization decision model to obtain the guiding parameters. The optimized decision-making model is a supervised learning model trained on historical cases. Each historical case includes intention bias, uncertainty index and corresponding expert calibration guidance parameter true value. The expert calibration guidance parameter true value embeds a balance between user autonomy and monitoring decision reliability. During training, the aforementioned intention bias and uncertainty indicators are used as inputs, and the expert-calibrated guidance parameter true value is used as the supervision signal. By minimizing the regularized mean square error loss function between the model output and the true value, the nonlinear mapping relationship from the input to the optimal guidance parameter is learned. This nonlinear mapping relationship implies the synergistic optimization of user autonomy space and system security level.

[0012] Optionally, the guidance parameters may include at least: cognitive gap bridging degree, information abstraction and scheduling strategy, and autonomous decision-making space threshold.

[0013] Secondly, an embodiment of the present invention provides an Internet of Things-driven bridge stress monitoring system, comprising: The building module is used to construct a digital twin for bridge stress monitoring based on IoT sensor data streams; The monitoring module is used to monitor the current working intention generated by the user when interacting with the digital twin and the uncertainty of that intention; in parallel, it derives the current globally optimal benchmark monitoring intention of the bridge. The solution module is used to solve for the guiding parameters by optimizing the decision model, with the goal of bridging the deviation between the current working intention and the benchmark monitoring intention, and taking into account the uncertainty. The solution objective of the optimized decision model takes into account both user autonomy and the reliability of monitoring decisions. The driving module, based on the guidance parameters, drives the digital twin to enter and present a guidance mode that adapts to the user's current cognitive state.

[0014] Optionally, the construction of a digital twin for bridge stress monitoring based on IoT sensor data streams includes: The IoT sensor data stream is preprocessed; The preprocessed IoT sensor data stream is input into the digital twin engine to construct a digital twin for bridge stress monitoring.

[0015] Optionally, the monitoring of the user's current working intent and the uncertainty of that intent during interaction with the digital twin includes: High-dimensional behavioral features are extracted from user interactions and mapped to behavioral semantic vectors. In the preset intent representation vector space, calculate the phylogenetic matching degree between the behavior semantic vector and each intent representation vector; The intent represented by the intent representation vector with the maximum genealogical matching degree is determined as the current working intent; The uncertainty of the current working intention is quantified based on the spectral residual between the maximum spectral matching degree and the ideal matching degree.

[0016] Optionally, the derivation of the bridge's current globally optimal benchmark monitoring intent includes: Based on users’ historical monitoring trajectories and the current distribution of monitoring tasks within the system, a heat map of monitoring resource coverage and a task gap distribution map are constructed. Based on the monitoring resource coverage heat map and task gap distribution map, identify the current monitoring blind spots and load weaknesses; The blind spots and weak points, along with the real-time mechanical state of the bridge and the preset monitoring procedures, are input into a multi-agent task allocation simulation environment for game theory simulation. The key task derived from the deduction that can maximize the overall monitoring effectiveness of the system is established as the benchmark monitoring objective.

[0017] The present invention has achieved the following beneficial effects: By constructing an IoT-driven digital twin of a bridge, the system can perceive user intentions and their uncertainties in real time. Combined with globally optimal baseline monitoring intentions, adaptive guidance parameters are generated, significantly improving the intelligence and accuracy of bridge stress monitoring. Compared to traditional methods, this technology, while ensuring user autonomy, effectively bridges the gap between user intentions and actual monitoring needs by optimizing the decision-making model, reducing the risk of decision-making errors and improving monitoring efficiency and reliability. Its adaptive guidance mode dynamically matches the user's cognitive state, optimizing the interactive experience and providing an efficient and intelligent solution for bridge safety management.

[0018] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an IoT-driven bridge stress monitoring method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of an IoT-driven bridge stress monitoring system according to an embodiment of the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] Figure 1 A flowchart of an IoT-driven bridge stress monitoring method is provided as an embodiment of this application, such as... Figure 1 As shown, the method includes: Step S1: Based on IoT sensor data streams, construct a digital twin for bridge stress monitoring.

[0023] Step S2: Monitor the current working intention generated by the user when interacting with the digital twin and the uncertainty of that intention; in parallel, derive the current globally optimal benchmark monitoring intention of the bridge.

[0024] Step S3: With the goal of bridging the discrepancy between the current working intention and the baseline monitoring intention, and taking into account the uncertainty, the guiding parameters are solved by optimizing the decision model; wherein, the solution objective of the optimized decision model takes into account both user autonomy and monitoring decision reliability.

[0025] Step S4: Based on the guidance parameters, drive the digital twin to enter and present a guidance mode that adapts to the user's current cognitive state.

[0026] This invention constructs an IoT-driven digital twin of a bridge to perceive user intentions and their uncertainties in real time. Combined with a globally optimal baseline monitoring intention, it generates adaptive guidance parameters, significantly improving the intelligence and accuracy of bridge stress monitoring. Compared to traditional methods, this technology, while ensuring user autonomy, effectively bridges the gap between user intentions and actual monitoring needs by optimizing the decision-making model, reducing the risk of decision-making errors and improving monitoring efficiency and reliability. Its adaptive guidance mode dynamically matches the user's cognitive state, optimizing the interactive experience and providing an efficient and intelligent solution for bridge safety management.

[0027] In some embodiments, step S1, constructing a digital twin for bridge stress monitoring based on IoT sensor data streams, includes the following sub-steps: Step S11: Preprocess the IoT sensor data stream.

[0028] Specifically, IoT sensor data streams refer to multi-dimensional time-series data streams collected in real time by various sensors installed on bridges (such as strain sensors, acceleration sensors, temperature sensors, etc.), containing information such as the bridge's stress state, vibration characteristics, and environmental parameters. Typical data include strain values, acceleration, and temperature. The preprocessing process includes four stages: data cleaning, format standardization, time synchronization, and feature enhancement.

[0029] Data cleaning identifies and removes outliers using statistical anomaly detection methods. For example, using an anomaly detection algorithm based on the 3σ criterion, the mean (e.g., 1000 microstrains) and standard deviation (e.g., 50 microstrains) of time-series data collected by strain sensors are calculated. Data points that exceed the mean ± 3 times the standard deviation are marked as anomalies and filled using linear interpolation.

[0030] Secondly, format standardization unifies the data from different sensors into a standardized format. For example, it unifies the 16-bit floating-point format of strain sensors and the 32-bit integer format of accelerometers into the 64-bit floating-point format of the IEEE 754 standard.

[0031] To address the timestamp discrepancies in multi-sensor data, a clock synchronization mechanism based on the Network Time Protocol (NTP) is employed. Specifically, sensor nodes synchronize their time with the central server via an NTP client (with a synchronization accuracy of 1ms), align the timestamps of each sensor's data, calculate the timestamp discrepancy (e.g., sensor A's timestamp is 1000ms, and sensor B's is 1002ms), and adjust the data points to a unified time grid (e.g., one sampling point per second) through linear interpolation.

[0032] Finally, feature enhancement extracts key features from time-series data through signal processing techniques. For example, Fast Fourier Transform (FFT) is applied to acceleration data to extract frequency domain features (such as a dominant frequency of 2Hz and an amplitude of 0.5m / s²).

[0033] Step S12: Input the preprocessed IoT sensor data stream into the digital twin engine to construct a digital twin for bridge stress monitoring.

[0034] Specifically, a digital twin engine refers to a software system that integrates physical simulation, data mapping, and state updates. It is implemented as a computing platform based on a multiphysics simulation framework (such as COMSOL Multiphysics) combined with a machine learning model. The digital twin includes the bridge's geometric model (defined as a three-dimensional finite element mesh, such as containing 100,000 tetrahedral elements), material properties (such as a concrete elastic modulus of 30 GPa and a Poisson's ratio of 0.2), boundary conditions (such as a support constraint force of 100 kN), and real-time stress state (such as a maximum stress of 10 MPa).

[0035] The specific construction process of a digital twin is as follows: First, the geometric and material parameters of the bridge are loaded. In practice, geometric data is extracted from the bridge design document, a finite element preprocessing tool (such as Gmsh) is used to generate a three-dimensional mesh (mesh density of 1000 elements per meter), and material properties are read from the material database (containing parameters such as the elastic modulus and density of concrete and steel reinforcement) and boundary conditions are set (such as the displacement of the fixed support is 0).

[0036] Next, the preprocessed sensor data stream (including strain, acceleration, temperature, etc.) is mapped to the finite element model nodes of the digital twin. In specific implementation, a spatial interpolation algorithm (such as inverse distance weighted interpolation, IDW) is used to calculate the strain value of each grid node based on the distance between the sensor position (represented by latitude and longitude coordinates) and the coordinates of the finite element grid nodes (e.g., the strain of node i is 0.8 times that of sensor A + 0.2 times that of sensor B), and a full-bridge stress distribution map is generated.

[0037] In addition, to achieve real-time updates, the digital twin engine integrates a time series prediction model, specifically a prediction model based on a Long Short-Term Memory (LSTM) network. The model is constructed by collecting historical sensor data, including features such as strain, acceleration, and temperature, to build a training dataset. The LSTM model structure consists of 3 layers (128 units per layer), using the ReLU activation function and the Adam optimizer (learning rate of 0.001). It is trained for 100 epochs by minimizing the mean squared error loss function (MSE) on a GPU cluster. After training, the model is deployed on the digital twin engine to predict the stress state of the bridge in real time.

[0038] To enhance model accuracy, physical constraints are introduced. For example, the predicted stress value must satisfy Hooke's Law. In practice, a constraint layer is added after the LSTM output to check whether the predicted stress satisfies the corresponding physical equation. If the deviation exceeds 5%, the output is adjusted through gradient correction.

[0039] This embodiment effectively improves data quality and the accuracy of digital twin construction by preprocessing IoT sensor data streams and inputting them into a digital twin engine to build a digital twin of bridge stress monitoring. The preprocessing step filters noise and corrects deviations, ensuring the reliability and consistency of the input data. The digital twin engine then generates a high-fidelity bridge digital twin through efficient data integration and modeling. This process significantly improves the real-time performance and accuracy of bridge stress monitoring, providing a solid foundation for subsequent user interaction and optimization decisions, thereby enhancing the overall efficiency of bridge safety monitoring.

[0040] In some embodiments, step S2, monitoring the current working intent generated by the user when interacting with the digital twin and the uncertainty of that intent, includes the following sub-steps: Step S21: Extract high-dimensional behavioral features from the user's interaction behavior and map them into behavioral semantic vectors.

[0041] Specifically, user interaction behaviors include a sequence of operations in which the user interacts with the digital twin through a graphical interface (such as a touch screen or mouse operation), such as clicking on a specific area of ​​the bridge model (such as the coordinates of the main beam node [50m, 5m]), adjusting the viewing angle (such as rotating the angle by 30°), and entering a query command (such as querying the stress of a certain node).

[0042] High-dimensional behavioral features refer to multi-dimensional quantitative indicators extracted from interactive behaviors, including operation type (such as click, drag), operation position (such as screen coordinates [500, 300]), operation frequency (such as 2 clicks per second), and operation duration (such as click lasting 0.5 seconds).

[0043] To capture semantic information, a pre-trained deep learning model is used to map behavioral sequences into behavioral semantic vectors. Specifically, this is a Transformer-based sequence encoding model. The model is constructed as follows: historical user interaction data is collected (at least 1000 user sessions, each containing 100 operations). Each operation sequence is labeled with a semantic tag (e.g., "query stress", "adjust perspective"). A training dataset is constructed (input is operation sequence features, output is semantic tags). The model structure includes a 6-layer Transformer encoder (512 dimensions per layer, 8 attention heads). Dropout (scale 0.1) is used to prevent overfitting, and an Adam optimizer (learning rate 0.0001) is used. The model is trained for 50 epochs using the cross-entropy loss function on a GPU cluster. After training, the model encodes the operation sequence (e.g., [click, (500, 300), 0.5 seconds]) into a 128-dimensional behavioral semantic vector (e.g., [0.12, -0.45, ..., 0.33]), representing the semantic features of user behavior. To ensure mapping accuracy, the model output vector is compared with a predefined semantic template (stored in a vector database, such as Faiss, which contains at least 1000 semantic template vectors) using cosine similarity calculation to verify whether the vector accurately represents the user's intent (e.g., similarity > 0.9 indicates successful mapping).

[0044] Step S22: In the preset intent representation vector space, calculate the phylogenetic matching degree between the behavior semantic vector and each intent representation vector.

[0045] Specifically, the intent representation vector space is a high-dimensional vector space (dimension 128) containing representation vectors of various predefined intents, such as "query main beam stress" (vectors such as [0.15, -0.32, ..., 0.41]) and "check support status" (vectors such as [0.22, 0.17, ..., -0.19]). These vectors are generated through expert annotation and clustering algorithms and stored in a vector database (such as Faiss).

[0046] Lineage matching degree refers to the similarity measure between the behavioral semantic vector and the intention representation vector. It is calculated using cosine similarity and is defined as the dot product of the two vectors divided by the product of their magnitudes (range [-1, 1], where 1 indicates a perfect match).

[0047] To improve matching accuracy, a genealogical analysis method is introduced, specifically an intent grouping algorithm based on K-means clustering. The algorithm is constructed as follows: historical interaction data (at least 10,000 behavioral semantic vectors) is collected, and the vectors are clustered into 50 intent clusters using the K-means algorithm (K=50, initialized with random centers). Each cluster center serves as an intent representation vector and is stored in the Faiss database. In the algorithm implementation, the KMeans module of the scikit-learn library is used, with a maximum iteration count of 300 and a convergence threshold of 0.0001. Training is performed on a CPU cluster, generating cluster center vectors (e.g., cluster 1 centers are [0.10, -0.50, ..., 0.30]). During the matching process, the behavioral semantic vectors are first compared with the cosine similarity of each cluster center, and the top 5 clusters with the highest similarity (e.g., similarity > 0.8) are selected. Then, within these clusters, further similarity is calculated with specific intent vectors to obtain the final genealogical matching degree (e.g., the highest similarity is 0.95, corresponding to the intent "query main beam stress").

[0048] Step S23: Determine the intent represented by the intent representation vector with the maximum spectrum matching degree as the current working intent.

[0049] Specifically, the current working intent refers to the core objective of the user's current interaction behavior, such as "querying the stress at a certain node of the main beam" or "adjusting the view to the support area". The determination process first extracts the maximum spectral matching degree and its corresponding intent representation vector from the calculation results of step S22. In specific implementation, all matching degree values ​​(such as [behavior ID, intent ID, similarity]) are read from the Redis cache (stored in the format of [behavior ID, intent ID, similarity]). Figure 1 [0.92], [Behavior 1, Intention] Figure 2[0.85]), using Python's max function to filter the maximum matching score (e.g., 0.92) and its intent ID (e.g., "query_stress"). Next, query the intent representation vector and its semantic description in the Faiss database by intent ID to confirm the intent category (e.g., "query the stress at a certain node of the main beam"), and record it as the current working intent.

[0050] Step S24: Based on the spectral residual between the maximum spectral matching degree and the ideal matching degree, quantify the uncertainty of the current working intention.

[0051] Specifically, the phylogenetic residual refers to the difference between the maximum phylogenetic match degree and the ideal match degree (defined as 1.0, representing a perfect match), reflecting the degree of ambiguity in the intention determination. For example, if the maximum match degree is 0.92, then the residual is 1.0 - 0.92 = 0.08. Uncertainty is a scalar value (ranging from [0, 1], where 0 represents complete certainty), calculated by normalizing the residual. The formula is: Uncertainty = Residual / Maximum Possible Residual (the maximum possible residual is 1.0). The quantization process first obtains the maximum phylogenetic match degree (e.g., 0.92) from the Redis cache in step S22, calculates the phylogenetic residual (1.0 - 0.92 = 0.08), and then normalizes it to obtain the uncertainty (e.g., 0.08 / 1.0 = 0.08).

[0052] Furthermore, to improve quantification accuracy, a context correction mechanism can be introduced. Specifically, the stability of intent is calculated based on the user's recent interaction sequence (the past 5 minutes, stored in MongoDB, including behavior ID, intent ID, and matching degree). The method is to count the number of consecutive occurrences of the same intent (e.g., "query main beam stress" appears 3 times consecutively). If the number exceeds the threshold (e.g., 2 times), the uncertainty is reduced (e.g., multiplied by a correction factor of 0.8, resulting in 0.08 × 0.8 = 0.064).

[0053] By extracting high-dimensional behavioral features, mapping behavioral semantic vectors, calculating genealogical matching degree, and quantifying uncertainty, this embodiment achieves accurate identification of user intent and scientific assessment of uncertainty. The introduction of behavioral semantic vectors allows user interaction intent to be captured and represented in a structured manner, while genealogical matching degree and residual analysis ensure high accuracy in intent determination and quantifiable uncertainty. This method not only enhances the system's ability to understand user intent but also provides a reliable basis for generating subsequent guidance parameters, significantly improving the intelligence and adaptability of the interaction process.

[0054] In some embodiments, step S2, deriving the current globally optimal benchmark monitoring intent for the bridge, includes the following sub-steps: Step S25: Based on the user's historical monitoring trajectory and the current monitoring division of labor distribution within the system, construct a monitoring resource coverage heat map and a task gap distribution map.

[0055] Specifically, historical monitoring trajectories refer to the sequence of monitoring tasks executed by the user in the past (stored in MongoDB, including task ID, execution time, and monitoring area, such as the main beam [50m, 60m]), reflecting the distribution of the areas of interest to the user.

[0056] Monitoring task allocation refers to the distribution of monitoring tasks undertaken by each user or device in the current system (e.g., user A is responsible for the main beam, user B is responsible for the support), which is stored in a distributed database (e.g., Cassandra).

[0057] The monitoring resource coverage heatmap is a two-dimensional grid (1m×1m resolution) representing the monitoring intensity of each area of ​​the bridge (e.g., the monitoring frequency for the main beam area is 5 times / hour). The task gap distribution map is a two-dimensional matrix (same resolution) representing the degree of absence of unmonitored areas (e.g., the abutment area was not monitored for 2 hours). The construction process first counts historical monitoring trajectories. Specifically, it queries the monitoring records of the past 24 hours from MongoDB (approximately 1000 records, with fields including user ID, timestamp, and area coordinates), calculates the monitoring frequency of each area (e.g., the main beam [50m, 60m] was queried 10 times, with a frequency of 10 / 24 = 0.42 times / hour), and generates a heatmap grid (stored using a NumPy array with a shape of [bridge length / 1m, bridge width / 1m]). Next, the current monitoring task distribution is analyzed. The task allocation table (fields include user ID, task area, and task time) is read from the Cassandra database. The monitoring coverage rate of each area (e.g., the main beam coverage rate is 80%) is calculated and overlaid with the heat map to generate a comprehensive coverage heat map (value range is [0, 1], 1 indicates complete coverage).

[0058] The task gap distribution map is generated by identifying low-coverage areas (e.g., coverage < 0.5) in the heatmap. Specifically, threshold segmentation is applied to the heatmap (using OpenCV's threshold function), and gap areas are marked (e.g., the support [10m, 20m] has a coverage of 0.3 and is marked as a gap).

[0059] Step S26: Based on the monitoring resource coverage heat map and task gap distribution map, identify the current monitoring blind spots and load weaknesses.

[0060] Specifically, monitoring blind spots refer to areas on the bridge that are not effectively monitored (e.g., the main beam [70m, 80m] was not queried, resulting in a coverage rate of 0). Weak load points refer to areas where insufficient monitoring resources may lead to monitoring failure (e.g., the support [20m, 30m] monitoring frequency is 0.1 times / hour, below the threshold of 0.5 times / hour).

[0061] During the identification process, heatmaps and gap maps are first loaded. Parquet files (containing two-dimensional grid data with shapes of [bridge length / 1m, bridge width / 1m]) are read from HDFS. NumPy array operations are used to extract areas with coverage below a threshold (0.3) as candidate blind zones (e.g., the main beam [70m, 80m] has a coverage of 0.2). Next, the unmonitored time is calculated based on the gap distribution map. Specifically, the unmonitored duration of gap areas is statistically analyzed (by querying the most recent monitoring time from MongoDB and calculating the time difference). Areas with unmonitored duration exceeding a threshold (e.g., 1 hour) are marked as monitoring blind zones. Load weaknesses are identified by analyzing the monitoring frequency gradient in the heatmap. Specifically, SciPy's gradient function is used to calculate the spatial gradient of coverage (e.g., the gradient for the main beam [50m, 60m] is 0.1 / m). If the gradient is higher than the threshold (0.05 / m) and the coverage is lower than 0.5, it is marked as a weak point.

[0062] Furthermore, to improve identification accuracy, a priority evaluation model can be introduced, specifically a multi-objective optimization algorithm based on weighted scoring. The algorithm is constructed as follows: a scoring function is defined (blind zone priority = unmonitored duration × 0.6 + regional importance × 0.4, where regional importance is obtained from the bridge design document, e.g., main beam importance is 0.9, and support importance is 0.7). Candidate blind zones and weak points are ranked, and the top 10 high-priority regions are selected. In the algorithm implementation, Python's heapq module is used to maintain the priority queue. The input is the score vector of the candidate regions, and the output is the ranked list of regions.

[0063] Step S27: Input the blind spots and weak points, along with the real-time mechanical state of the bridge and the preset monitoring procedures, into the multi-agent task allocation simulation environment for game theory simulation.

[0064] Specifically, monitoring blind spots and weak points are represented by regional coordinates and priorities (e.g., main beam [70m, 80m], priority 0.85). The real-time mechanical state of the bridge is obtained from a digital twin, including stress distribution (e.g., maximum stress 10MPa) and vibration frequency (e.g., 2Hz). The preset monitoring procedures are monitoring requirements extracted from bridge management specifications (e.g., main beam stress monitoring frequency is 1 time / hour). The multi-agent task allocation simulation environment is a game framework based on reinforcement learning. Each agent represents a monitoring task executor (e.g., user or device), and the goal is to maximize the overall monitoring utility of the system (defined as coverage × 0.7 + response time × 0.3). The simulation process first initializes the agent state, specifically by reading the current monitoring assignment (including user ID and task area) from the Cassandra database, and assigning an initial task to each agent (e.g., user A monitors the main beam [50m, 60m]). Next, a game environment is constructed, specifically a custom environment based on the OpenAIGym framework. The environment states include blind zone coordinates, weak point priority, real-time stress, and monitoring frequency requirements. The action space is the task allocation decision (e.g., allocating the main beam [70m, 80m] to user B). The reward function is the monitoring utility (coverage is calculated based on the blind zone elimination ratio, and response time is calculated based on the task completion delay; a positive reward is given if the delay is <10 seconds).

[0065] The game simulation employs a multi-agent reinforcement learning algorithm (MADDPG). The algorithm is constructed as follows: each agent defines a policy network (a 3-layer fully connected network with 128 units per layer, ReLU activation) and a critic network (inputting global state and actions, outputting Q-values). The Adam optimizer is used (learning rate 0.001). The training dataset consists of 10,000 historical monitoring task records (including state, action, and reward). Training is conducted for 1000 episodes (each episode simulating 1 hour of monitoring) on ​​a GPU cluster. During the simulation, the agents dynamically adjust task allocation based on blind zone priority and real-time stress (e.g., allocating high-stress areas [70m, 80m] to idle users), converging to the optimal allocation scheme through 100 iterations. Simulation results are output as a task list (stored in MongoDB, with fields including task ID, region, executor, and utility value).

[0066] Step S28: The key task derived from the deduction that can maximize the overall monitoring effectiveness of the system is established as the benchmark monitoring objective.

[0067] Specifically, the focus task refers to the high-priority monitoring task derived from the deduction (such as monitoring the stress of the main beam [70m, 80m], with a utility value of 0.95). The baseline monitoring intent is the monitoring target recommended by the system, in the format of a semantic description (such as "prioritize monitoring the stress of the main beam [70m, 80m]") and a corresponding intent representation vector (such as [0.18, -0.41, ..., 0.37]).

[0068] During the establishment process, the simulation results (fields including task ID, region, and utility value) are first read from MongoDB. The task with the highest utility value (e.g., a task with a utility value of 0.95) is selected, and the real-time status of the digital twin is queried through the task region coordinates (e.g., stress of 10.2 MPa). A semantic description is then generated (using template filling, such as "monitoring {indicators} of {region}"). Next, the semantic description is mapped to an intent representation vector. Specifically, the Transformer model from step S21 is used, with the semantic description text as input and a 128-dimensional vector as output.

[0069] By constructing a heatmap of monitoring resource coverage and a task gap distribution map, identifying monitoring blind spots and load weaknesses, and combining multi-agent game theory, this embodiment successfully derived the current globally optimal baseline monitoring intention for the bridge. This process comprehensively considers historical data, real-time mechanical state, and monitoring procedures, achieving efficient allocation of monitoring resources and maximizing system utility. The obtained baseline monitoring intention accurately reflects the priority and key points of bridge monitoring, providing users with a scientific reference direction, thereby significantly improving the comprehensiveness and reliability of monitoring decisions.

[0070] In some embodiments, step S3, with the goal of bridging the discrepancy between the current working intention and the baseline monitoring intention, and taking into account the uncertainty, involves solving for the guiding parameters by optimizing the decision model, and includes the following sub-steps: Step S31: Input the deviation between the current working intention and the benchmark monitoring intention, and the uncertainty into the optimization decision model, and solve to obtain the guiding parameters.

[0071] The optimized decision-making model is a supervised learning model trained on historical cases. Each historical case includes intention bias, uncertainty index and corresponding expert calibration guidance parameter true value. The expert calibration guidance parameter true value embeds a balance between user autonomy and monitoring decision reliability.

[0072] During training, the aforementioned intention bias and uncertainty indicators are used as inputs, and the expert-calibrated guidance parameter true value is used as the supervision signal. By minimizing the regularized mean square error loss function between the model output and the true value, the nonlinear mapping relationship from the input to the optimal guidance parameter is learned. This nonlinear mapping relationship implies the synergistic optimization of user autonomy space and system security level.

[0073] The guiding parameters include at least: The cognitive gap bridging degree is calculated by the optimization decision model based on the deviation between the current work intention and the benchmark monitoring intention, which directly drives the intensity and range of visual / auditory guidance signals in the interactive interface to focus the user's attention on the intention deviation area; The information abstraction and scheduling strategy is made by the optimization decision model in conjunction with the quantification results of the uncertainty, and dynamically adjusts the granularity of the presented monitoring information and the depth of the explanatory content to adapt to and stabilize the user's current fluctuating cognitive state. The autonomous decision-making space threshold, as the direct output of the optimized decision-making model to achieve a balance between user autonomy and monitoring reliability, is used to define the boundary between user-operable autonomy and system-suggestive intervention, thereby maximizing the user's autonomous exploration space within the framework of ensuring the reliability of monitoring decisions.

[0074] Specifically, the optimized decision-making model is a supervised learning-based neural network that aims to learn the mapping from intention bias and uncertainty to guidance parameters. Guidance parameters include: cognitive gap bridging degree (range [0, 1], e.g., 0.8, representing guidance signal strength), information abstraction and scheduling strategy (information granularity such as "detailed" or "summary", depth such as "provide formula explanation"), and autonomous decision-making space threshold (range [0, 1], e.g., 0.7, representing the proportion of user autonomous operation). The model is constructed by collecting historical case data (10,000 records, stored in PostgreSQL, fields including intention bias, uncertainty, and expert-calibrated guidance parameters). Expert calibration involves manual evaluation to balance autonomy and reliability (e.g., setting high-intensity guidance when bias > 0.2). The model structure is a 5-layer fully connected network (256 units per layer, ReLU activation). The input consists of [bias, uncertainty] (2D), and the output consists of [bridge, information granularity, information depth, autonomy threshold] (4D). It uses the Adam optimizer (learning rate 0.0001) and is trained for 200 epochs using a regularized mean squared error loss function (L2 regularization coefficient 0.01) on a GPU cluster. The solution process first calculates the intent bias, specifically by using NumPy's `linalg.norm` function to calculate the Euclidean distance between two vectors (e.g., 0.15), and then inputting the uncertainty (e.g., 0.064) into the model to obtain the guiding parameters (e.g., bridge 0.8, information granularity "summary", information depth "simple description", autonomy threshold 0.7). To ensure reliability, the model output is validated using rules (e.g., when bridge > 0.9, the information granularity is forced to be "detailed").

[0075] By optimizing the decision-making model to solve for the guidance parameters, this embodiment effectively bridges the gap between the user's current work intention and the baseline monitoring intention, while taking into account both user autonomy and the reliability of monitoring decisions. The model is trained based on historical cases, integrating intention bias and uncertainty indicators to generate guidance parameters that include cognitive gap bridging degree, information abstraction and scheduling strategies, and autonomous decision-making space thresholds. These parameters can dynamically adjust the guidance signals and information presentation methods of the interactive interface to adapt to the user's cognitive state, significantly improving the accuracy of the monitoring process, user experience, and system reliability.

[0076] In some embodiments, in step S4, based on the guidance parameters, the digital twin is driven to enter and present a guidance mode adapted to the user's current cognitive state. The guidance mode refers to a set of interactive behaviors of the digital twin, including visual guidance (e.g., highlighting the main beam [70m, 80m]), auditory prompts (e.g., voice prompt "Please check the stress of the main beam"), and information presentation (e.g., displaying the stress value of 10.2MPa). The driving process first parses the guidance parameters, specifically reading parameter values ​​from Redis (key "guide_params"). The intensity of the visual / auditory signals is adjusted according to the degree of convergence. For example, if the degree of convergence is 0.8, the transparency of the highlighted area is set to 0.8 (using the alpha channel of WebGL), and the volume of the voice prompt is set to 80% (set via WebAudioAPI). Next, the information presentation is adjusted according to the information abstraction and scheduling strategy. Specifically, if the granularity is "summary", key indicators (e.g., "main beam stress: 10.2MPa") are displayed; if the depth is "simple description", a brief explanation is added (e.g., "stress is normal"). In implementation, React components are used to dynamically render interface elements. The autonomous decision-making space threshold controls the boundaries of user operations. For example, if the threshold is 0.7, then 70% of interface operations (such as clicking on areas) are performed autonomously by the user, and 30% are guided by system suggestions (such as automatically focusing on high-stress areas).

[0077] By analyzing guidance parameters and driving the digital twin into an adaptive guidance mode, this embodiment achieves dynamic optimization of the interactive interface and precise adaptation to the user's cognitive state. Flexible adjustments to visual guidance, auditory cues, and information presentation, along with the reasonable division of thresholds in the autonomous decision-making space, enable the system to maximize user autonomy while ensuring monitoring reliability. This adaptive guidance mode significantly improves the efficiency and experience of user interaction with the digital twin, providing intelligent and personalized technical support for bridge stress monitoring.

[0078] Figure 2 This application provides a schematic diagram of an IoT-driven bridge stress monitoring system, as shown in the embodiment of the present application. Figure 2 As shown, the system includes: Module 100 is used to build a digital twin for bridge stress monitoring based on IoT sensor data streams; The monitoring module 200 is used to monitor the current working intention generated by the user when interacting with the digital twin and the uncertainty of the intention; in parallel, it derives the current globally optimal benchmark monitoring intention of the bridge. The solution module 300 is used to solve for the guiding parameters by optimizing the decision model, with the goal of bridging the deviation between the current working intention and the benchmark monitoring intention, and taking into account the uncertainty. The solution objective of the optimization decision model takes into account both user autonomy and monitoring decision reliability. The driving module 400, based on the guidance parameters, drives the digital twin to enter and present a guidance mode that adapts to the user's current cognitive state.

[0079] The construction of a digital twin for bridge stress monitoring based on IoT sensor data streams includes: The IoT sensor data stream is preprocessed; The preprocessed IoT sensor data stream is input into the digital twin engine to construct a digital twin for bridge stress monitoring.

[0080] The monitoring of the user's current working intent and the uncertainty of that intent during interaction with the digital twin includes: High-dimensional behavioral features are extracted from user interactions and mapped to behavioral semantic vectors. In the preset intent representation vector space, calculate the phylogenetic matching degree between the behavior semantic vector and each intent representation vector; The intent represented by the intent representation vector with the maximum genealogical matching degree is determined as the current working intent; The uncertainty of the current working intention is quantified based on the spectral residual between the maximum spectral matching degree and the ideal matching degree.

[0081] The derivation of the current globally optimal benchmark monitoring intention for the bridge includes: Based on users’ historical monitoring trajectories and the current distribution of monitoring tasks within the system, a heat map of monitoring resource coverage and a task gap distribution map are constructed. Based on the monitoring resource coverage heat map and task gap distribution map, identify the current monitoring blind spots and load weaknesses; The blind spots and weak points, along with the real-time mechanical state of the bridge and the preset monitoring procedures, are input into a multi-agent task allocation simulation environment for game theory simulation. The key task derived from the deduction that can maximize the overall monitoring effectiveness of the system is established as the benchmark monitoring objective.

[0082] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An Internet of Things-driven method for monitoring bridge stress, characterized in that, include: Based on IoT sensor data streams, a digital twin for bridge stress monitoring is constructed. Monitor the current working intention generated by the user when interacting with the digital twin and the uncertainty of that intention; in parallel, derive the current globally optimal benchmark monitoring intention of the bridge; With the goal of bridging the discrepancy between the current working intention and the baseline monitoring intention, and taking into account the uncertainty, the guiding parameters are solved by optimizing the decision model; wherein, the solution objective of the optimized decision model takes into account both user autonomy and monitoring decision reliability. Based on the guidance parameters, the digital twin is driven to enter and present a guidance mode that adapts to the user's current cognitive state.

2. The IoT-driven bridge stress monitoring method as described in claim 1, characterized in that, The construction of a digital twin for bridge stress monitoring based on IoT sensor data streams includes: The IoT sensor data stream is preprocessed; The preprocessed IoT sensor data stream is input into the digital twin engine to construct a digital twin for bridge stress monitoring.

3. The IoT-driven bridge stress monitoring method as described in claim 1, characterized in that, The monitoring of the user's current working intent and the uncertainty of that intent during interaction with the digital twin includes: High-dimensional behavioral features are extracted from user interactions and mapped to behavioral semantic vectors. In the preset intent representation vector space, calculate the phylogenetic matching degree between the behavior semantic vector and each intent representation vector; The intent represented by the intent representation vector with the maximum genealogical matching degree is determined as the current working intent; The uncertainty of the current working intention is quantified based on the spectral residual between the maximum spectral matching degree and the ideal matching degree.

4. The IoT-driven bridge stress monitoring method as described in claim 1, characterized in that, The derivation of the current globally optimal benchmark monitoring intention for the bridge includes: Based on users’ historical monitoring trajectories and the current distribution of monitoring tasks within the system, a heat map of monitoring resource coverage and a task gap distribution map are constructed. Based on the monitoring resource coverage heat map and task gap distribution map, identify the current monitoring blind spots and load weaknesses; The blind spots and weak points, along with the real-time mechanical state of the bridge and the preset monitoring procedures, are input into a multi-agent task allocation simulation environment for game theory simulation. The key task derived from the deduction that can maximize the overall monitoring effectiveness of the system is established as the benchmark monitoring objective.

5. The IoT-driven bridge stress monitoring method as described in claim 1, characterized in that, The goal is to bridge the discrepancy between the current working intention and the baseline monitoring intention, and to take into account the uncertainties, by optimizing the decision model to solve for the guiding parameters, including: The deviation between the current working intention and the benchmark monitoring intention, and the uncertainty are input into the optimization decision model to obtain the guiding parameters. The optimized decision-making model is a supervised learning model trained on historical cases. Each historical case includes intention bias, uncertainty index and corresponding expert calibration guidance parameter true value. The expert calibration guidance parameter true value embeds a balance between user autonomy and monitoring decision reliability. During training, the aforementioned intention bias and uncertainty indicators are used as inputs, and the expert-calibrated guidance parameter true value is used as the supervision signal. By minimizing the regularized mean square error loss function between the model output and the true value, the nonlinear mapping relationship from the input to the optimal guidance parameter is learned. This nonlinear mapping relationship implies the synergistic optimization of user autonomy space and system security level.

6. The IoT-driven bridge stress monitoring method as described in claim 1, characterized in that, The guidance parameters include at least: cognitive gap bridging degree, information abstraction and scheduling strategy, and autonomous decision-making space threshold.

7. An Internet of Things-driven bridge stress monitoring system, characterized in that, include: The building module is used to construct a digital twin for bridge stress monitoring based on IoT sensor data streams; The monitoring module is used to monitor the current working intention generated by the user when interacting with the digital twin and the uncertainty of that intention; in parallel, it derives the current globally optimal benchmark monitoring intention of the bridge. The solution module is used to solve for the guiding parameters by optimizing the decision model, with the goal of bridging the deviation between the current working intention and the benchmark monitoring intention, and taking into account the uncertainty. The solution objective of the optimized decision model takes into account both user autonomy and the reliability of monitoring decisions. The driving module, based on the guidance parameters, drives the digital twin to enter and present a guidance mode that adapts to the user's current cognitive state.

8. The IoT-driven bridge stress monitoring system as described in claim 7, characterized in that, The construction of a digital twin for bridge stress monitoring based on IoT sensor data streams includes: The IoT sensor data stream is preprocessed; The preprocessed IoT sensor data stream is input into the digital twin engine to construct a digital twin for bridge stress monitoring.

9. The IoT-driven bridge stress monitoring system as described in claim 7, characterized in that, The monitoring of the user's current working intent and the uncertainty of that intent during interaction with the digital twin includes: High-dimensional behavioral features are extracted from user interactions and mapped to behavioral semantic vectors. In the preset intent representation vector space, calculate the phylogenetic matching degree between the behavior semantic vector and each intent representation vector; The intent represented by the intent representation vector with the maximum genealogical matching degree is determined as the current working intent; The uncertainty of the current working intention is quantified based on the spectral residual between the maximum spectral matching degree and the ideal matching degree.

10. The IoT-driven bridge stress monitoring system as described in claim 7, characterized in that, The derivation of the current globally optimal benchmark monitoring intention for the bridge includes: Based on users’ historical monitoring trajectories and the current distribution of monitoring tasks within the system, a heat map of monitoring resource coverage and a task gap distribution map are constructed. Based on the monitoring resource coverage heat map and task gap distribution map, identify the current monitoring blind spots and load weaknesses; The blind spots and weak points, along with the real-time mechanical state of the bridge and the preset monitoring procedures, are input into a multi-agent task allocation simulation environment for game theory simulation. The key task derived from the deduction that can maximize the overall monitoring effectiveness of the system is established as the benchmark monitoring objective.

Images