A multi-source data driven bridge electromechanical installation safety monitoring method and system

Abstract

The application discloses a kind of multi-source data driven bridge electromechanical installation safety monitoring method and system, belong to safety monitoring technical field.The method first constructs cooperative collection architecture, and obtains structural stress, vibration, displacement and environmental monitoring data by the synchronous acquisition of unified data acquisition protocol and time reference synchronization;Second, the data are filtered and processed with multiple levels and outliers are removed, and the time series interpolation method based on adjacent valid data points is used to repair data missing area;Then, the pretreated data are input into the multi-source data correlation model based on multi-task deep learning network, high-order features are extracted and multiple prediction and evaluation results are output;Finally, based on the output results, the analytic hierarchy process and fuzzy comprehensive evaluation method are used for comprehensive scoring and graded early warning.The application solves the problems of data island and poor data quality, deeply mines the coupling relationship between parameters, breaks through the limitations of single index, and improves the risk comprehensive judgment ability.

Description

Technical Field

[0001] This invention relates to the field of bridge engineering and safety monitoring technology, specifically to a multi-source data-driven method and system for safety monitoring of bridge electromechanical installations. Background Technology

[0002] With the continuous expansion of modern bridge construction and the increasing service life, the safety of bridge structures has become a growing focus in the engineering field. As a crucial link in bridge construction and operation, the safety of bridge electromechanical installation directly affects the stable operation and service life of the overall bridge structure. Safety monitoring of bridge electromechanical installation aims to comprehensively utilize various sensing technologies and data processing methods to perceive and assess the structural state, equipment operating parameters, and environmental conditions during the installation process in real time, promptly identifying abnormalities and providing risk warnings.

[0003] In recent years, multi-source data-driven monitoring methods have gradually become an important branch of technology in this field. However, existing technologies have many shortcomings in practical applications:

[0004] (1) Current monitoring systems mainly rely on single sensing technologies. There is a lack of effective coordination mechanisms between subsystems. Different types of monitoring equipment have obvious protocol differences and inconsistent time bases, making it difficult to achieve synchronous acquisition and spatiotemporal alignment of multi-dimensional parameters, resulting in serious data silos.

[0005] (2) The installation of bridge electromechanical equipment is often in harsh working conditions and strong electromagnetic interference environment in the field. The original monitoring data is easily affected by environmental noise, resulting in quality problems such as missing data, distortion and abnormal jump points, which affects the authenticity and effectiveness of subsequent analysis.

[0006] (3) Existing technologies generally lack in-depth exploration of the intrinsic correlation between multi-source monitoring data. It is difficult to establish the potential coupling relationship between monitoring data and structural safety risks, resulting in a large amount of monitoring data that cannot be directly applied to safety assessment. It is difficult to break through the limitations of traditional single-indicator early warning, which restricts the improvement of comprehensive early warning and risk assessment capabilities. Summary of the Invention

[0007] This invention provides a multi-source data-driven method and system for monitoring the safety of bridge electromechanical installations, addressing the technical problems in existing bridge monitoring technologies, such as data silos caused by a lack of collaboration among monitoring equipment, poor data quality under complex working conditions, and a lack of in-depth mining of potential coupling relationships in multi-dimensional data, which restricts the ability to provide early warnings.

[0008] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0009] In a first aspect, the present invention provides a multi-source data-driven method for safety monitoring of bridge electromechanical installation, comprising the following steps:

[0010] Step 1: Construct a collaborative acquisition architecture. Through a unified data acquisition protocol and time reference, synchronously acquire multi-source monitoring data during the bridge electromechanical installation process. The multi-source monitoring data includes at least structural stress data, vibration data, displacement data, and environmental data.

[0011] Step 2: Perform multi-level filtering and outlier removal on the acquired multi-source monitoring data, and repair the missing data areas based on the time series interpolation method of adjacent valid data points to obtain preprocessed multi-source monitoring data;

[0012] Step 3: Input the preprocessed multi-source monitoring data into a pre-trained multi-source data association model for association modeling and analysis, extract high-order feature representations between multi-dimensional parameters, and output stress prediction results, vibration prediction results, displacement prediction results and environmental impact assessment results. The multi-source data association model is built based on a multi-task deep learning network architecture.

[0013] Step 4: Based on the output results of the multi-source data association model, the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation method are used to comprehensively score the safety status of electromechanical installation, and different warning levels are assigned according to the scoring results for graded warning.

[0014] Furthermore, in step 1, the data acquisition protocol follows a unified data encapsulation format, and a single frame data packet sequentially includes four fields: acquisition timestamp, sensor type identifier, original measurement value, and cyclic redundancy check code; various sensing devices used to acquire the multi-source monitoring data perform clock calibration through a unified time synchronization protocol to unify the time reference.

[0015] In step 2, the multi-stage filtering process specifically includes a first-stage filtering, a second-stage filtering, and a third-stage filtering. The first-stage filtering uses a second-order Butterworth low-pass filter to suppress high-frequency interference signals, and the cutoff frequency of the second-order Butterworth low-pass filter is set to half of the data sampling frequency. The second-stage filtering uses time-domain statistical filtering based on the Grubbs criterion to detect and remove outliers. The third-stage filtering uses a spatial smoothing bilateral filtering algorithm to smooth noise interference while maintaining the characteristics of data mutation.

[0016] In step 2, the interpolation method is cubic spline interpolation. When repairing missing data areas, if the length of a single continuous missing data repair exceeds the preset sampling period threshold, the corresponding missing data area is marked as an invalid area and a manual verification report is triggered.

[0017] In step 3, the multi-task deep learning network architecture includes a shared feature extraction layer and multiple task-specific output layers. The shared feature extraction layer includes three sequentially connected one-dimensional convolutional neural networks, with a max pooling layer for feature dimensionality reduction after each one-dimensional convolutional neural network layer. The multiple task-specific output layers correspond to four sub-tasks: stress prediction, vibration prediction, displacement prediction, and environmental impact assessment. Each task-specific output layer consists of a fully connected layer and an output layer.

[0018] In step 3, the multi-source data association model adopts a hybrid network architecture that combines a multi-task deep learning network with a graph neural network. The multi-task deep learning network is used to extract temporal feature representations, and the graph neural network is used to model the spatial topological relationships of the sensors. The graph neural network is constructed based on the spatial coordinates of the deployment locations of each sensor and establishes edge relationships in the graph structure between adjacent sensors. The edge weight values ​​of the edge relationships are the reciprocal of the Euclidean distance between the two corresponding sensors in the actual deployment space.

[0019] In step 4, the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation method are executed based on a constructed multi-level evaluation index system. This multi-level evaluation index system includes four primary indicators: structural safety indicators, equipment status indicators, environmental impact indicators, and historical trend indicators. Each primary indicator has several secondary indicators. The specific process for comprehensively scoring the safety status of electromechanical installation is as follows: a judgment matrix is ​​constructed according to the scaling method of the AHP, and the weights corresponding to each primary and secondary indicator are calculated; a fuzzy evaluation matrix is ​​established, and the qualitative evaluation is transformed into a quantitative score based on the weights of each indicator. Finally, the comprehensive score result is calculated and accumulated, and the warning level is divided into four warning levels from low to high: normal, attention, alert, and danger.

[0020] Furthermore, the bridge electromechanical installation safety monitoring method is applied to a three-level network architecture consisting of a data aggregation unit, an edge computing unit, and a cloud analysis unit. Step 1 is executed by the data aggregation unit, and steps 2 to 4 are executed by the edge computing unit deployed on-site to achieve local risk warning decision-making. The edge computing unit synchronizes the processed data to the cloud analysis unit, which is used for historical data storage, periodic model updates, and comprehensive evaluation report generation. The sensing devices used to acquire the multi-source monitoring data are all equipped with protective housings made of metal. The protective housings have a double sealing structure inside, including a first rubber sealing ring and a second skeleton oil seal. The internal circuits are electromagnetically isolated using a multi-layer metal shielding structure with alternating copper foil and aluminum foil.

[0021] Secondly, the present invention provides a multi-source data-driven bridge electromechanical installation safety monitoring system, comprising:

[0022] The multi-source collaborative acquisition module is used to build a collaborative acquisition architecture. Through a unified data acquisition protocol and time reference, it synchronously acquires multi-source monitoring data during the bridge electromechanical installation process. The multi-source monitoring data includes at least structural stress data, vibration data, displacement data, and environmental data.

[0023] The data quality control and preprocessing module is used to perform multi-level filtering and outlier removal on the acquired multi-source monitoring data, and to repair the missing data areas based on the time series interpolation method of adjacent valid data points, so as to obtain the preprocessed multi-source monitoring data.

[0024] The multi-source data association modeling module is used to input the preprocessed multi-source monitoring data into a pre-trained multi-source data association model for association modeling analysis, extract high-order feature representations between multi-dimensional parameters, and output stress prediction results, vibration prediction results, displacement prediction results, and environmental impact assessment results. The multi-source data association model is constructed based on a multi-task deep learning network architecture.

[0025] The risk comprehensive assessment module is used to comprehensively score the safety status of electromechanical installation based on the output results of the multi-source data association model, using the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation method, and to classify the results into different warning levels for graded warning.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] (1) By constructing a collaborative acquisition architecture covering multiple dimensions of parameters, this invention unifies the data acquisition protocol and time reference, and realizes the synchronous acquisition and accurate spatiotemporal alignment of multi-source heterogeneous data such as structural stress, vibration, displacement and environment, providing a complete and consistent data base for subsequent correlation analysis and risk assessment.

[0028] (2) By introducing multi-level filtering and outlier removal, combined with time series interpolation repair method based on adjacent valid data points, the influence of severe environmental noise and electromagnetic interference was eliminated, missing data was filled, and the authenticity and integrity of the underlying data entering the evaluation model were ensured.

[0029] (3) By using a multi-task deep learning network to build an association model, the high-order features and potential coupling relationships between multi-dimensional parameters were explored. The analytic hierarchy process and fuzzy comprehensive evaluation method were integrated to conduct scientific scoring. This mechanism changed the traditional fragmented and single-factor early warning mode and realized accurate and three-dimensional risk comprehensive judgment and multi-level early warning.

[0030] (4) Introducing graph neural networks into the hybrid network architecture to model the spatial topological relationship between sensors and mapping spatial distance to edge weights, which complements the temporal feature extraction of multi-task networks, this design enhances the model’s expressive power and generalization performance in complex scenarios.

[0031] (5) In the multi-level filtering strategy, Butterworth low-pass suppression, Grubbs time-domain elimination and bilateral filtering spatial smoothing are connected in series. This combination method not only thoroughly purifies multi-source noise, but also protects the abrupt edge characteristics of the data and avoids the loss of key risk signals in the filtering process.

[0032] (6) The edge computing unit performs real-time data preprocessing and local early warning decision-making, and the cloud analysis unit performs long-term data storage and model updates, which reduces response latency and improves the system's autonomous operation capability; at the same time, the dual-sealed and multi-layer metal shielding shell design ensures the survival capability of the field acquisition nodes under extreme working conditions.

[0033] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, embodiments of the present invention are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a schematic diagram of the bridge electromechanical installation safety monitoring method in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of multi-dimensional feature fusion in an embodiment of the present invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0038] like Figure 1As shown in the figure, the bridge electromechanical installation safety monitoring method driven by multi-source data proposed in this embodiment of the invention relies on a three-level distributed system architecture of end-edge-cloud. This system architecture consists of a data aggregation unit (end-side), an edge computing unit (edge-side), and a cloud analysis unit (cloud-side), forming a complete technical closed loop from data acquisition, transmission, local real-time early warning to long-term cloud storage and model evolution. Under harsh working conditions, the system uses highly reliable hardware to collaboratively acquire multimodal data throughout the entire bridge electromechanical installation cycle. Through real-time data cleaning, high-order feature fusion, and multi-dimensional risk assessment at the edge side, it completely breaks through the technical bottlenecks of data silos, singular analysis, and delayed early warning in traditional monitoring systems. The specific bridge electromechanical installation safety monitoring method described in this embodiment of the invention is as follows:

[0039] Step 1: Multi-source collaborative acquisition and spatiotemporal benchmark alignment

[0040] In this step, the core task is to build a collaborative acquisition architecture that covers key structural and environmental parameters of the bridge. By enforcing a unified data acquisition protocol and a high-precision time synchronization benchmark, the synchronous acquisition and precise spatiotemporal alignment of multi-source monitoring data, such as structural stress data, vibration data, displacement data, environmental data, structural surface crack data, and overall tilt data, can be achieved.

[0041] To adapt to the electromechanical installation scenarios of ultra-large-scale and complex bridges (such as long-span cable-stayed bridges and suspension bridges), the collaborative data acquisition architecture of this invention adopts a distributed multi-node network design. Sensor selection and deployment cover the following six sensing units:

[0042] (1) Stress measurement unit: Industrial-grade high-sensitivity resistance strain gauges or fiber optic strain gauges are used as sensitive elements and are deployed in key stress and stress concentration areas such as the root of the main tower, the mid-span of the main beam, and the anchorage area of ​​the stay cables. The sampling frequency is set to 100 Hz to continuously capture transient stress changes during electromechanical installation or load application.

[0043] (2) Vibration measurement unit: Microelectromechanical systems (MEMS) triaxial accelerometers or piezoelectric accelerometers are used and deployed at the mid-span, quarter-span, and end-span points of the main beam of the bridge, which are sensitive points of structural dynamic response. They are fixed with strong magnetic bases to ensure high-frequency response fidelity. The sampling frequency is set to 500 Hz to capture high-frequency dynamic characteristics such as construction machinery vibration and wind-induced vibration.

[0044] (3) Displacement measurement unit: High-precision wire-type displacement sensors or laser displacement sensors are used and deployed at the locations of maximum structural deformation or most sensitive displacement, such as the mid-span deflection point and the end of the cantilever. They are fixed to a stable reference surface using rigid brackets. The sampling frequency is set to 50 Hz to monitor spatial deformation and cumulative deflection during the installation process.

[0045] (4) Environmental Measurement Unit: Integrates high-precision temperature sensors (thermocouples or PT100), capacitive humidity sensors, and ultrasonic anemometers, which are deployed at different elevations on the bridge deck and inside the installation compartments of key equipment. The sampling frequency is set to 1 Hz to record in real time the environmental boundary conditions that may cause structural thermal stress or wind loads.

[0046] (5) Crack Measurement Unit: A high-precision surface crack gauge or a machine vision-based micro-crack detection camera system is used, deployed in historically vulnerable areas or at concrete pouring joints, to monitor the initiation and width expansion of micro-cracks on the structural surface. The sampling frequency is set to 0.1 Hz, and the detection resolution can reach 0.01 mm.

[0047] (6) Tilt Measurement Unit: A dual-axis tilt sensor or a high-precision attitude reference system (AHRS) is used to monitor the overall tilt angle and tilt change rate of the main tower or main beam. The sampling frequency is set to 10 Hz, and the measurement accuracy reaches 0.001 degrees.

[0048] Due to the wide variety of sensor types and communication interfaces, using a traditional direct connection mode would inevitably lead to severe data silos and timing discrepancies. Therefore, this invention implements a strict "protocol unification" and "time base unification" mechanism within the data aggregation unit.

[0049] Regarding the data acquisition protocol, all data from the front-end sensor nodes is initially encapsulated by the local data acquisition station, following a unified data encapsulation format across the entire network. A single data packet sequentially contains four core fields along the logical link:

[0050] The first field is the collection timestamp, represented by a 64-bit unsigned integer, recording the absolute number of milliseconds or microseconds from 00:00:00 UTC on January 1, 1970 to the moment of sampling.

[0051] The second field is the sensor type identifier, which uses an 8-bit unsigned integer encoding. For example, 00H represents stress data, 01H represents vibration data, 02H represents displacement data, 03H represents environmental data, 04H represents crack data, and 05H represents tilt data, ensuring that the decoding end can quickly route the data.

[0052] The third field is the original measurement value, which is uniformly converted into a 32-bit or 64-bit floating-point number format according to the accuracy requirements, eliminating the differences between different units and the number of bits of the ADC (Analog-to-Digital Converter).

[0053] The fourth field is the Cyclic Redundancy Check (CRC-16), which uses a 16-bit polynomial to hash the first three fields. It is used for error checking of the transmission link to ensure the integrity of data in complex electromagnetic environments.

[0054] Regarding the time reference, to meet the stringent requirements of microsecond-level phase consistency in bridge dynamics analysis (such as modal decomposition and modal analysis), this invention abandons the traditional Network Time Protocol (NTP) and fully upgrades to the Precision Time Protocol (PTP, conforming to the IEEE 1588v2 standard). The data aggregation unit incorporates a rubidium atomic clock or a high-precision temperature-controlled crystal oscillator as the master clock, sending Sync messages to each local data acquisition station (slave clock) via a fiber optic ring network. By calculating the four-step timestamp difference between the Sync, Follow_Up, Delay_Req, and Delay_Resp messages, link transmission delays and dwell times are automatically compensated, strictly controlling the sampling clock synchronization error of all sensors in the network to within 0.1 milliseconds. This extreme spatiotemporal alignment precision provides an impeccable data foundation for subsequent multi-dimensional joint analysis of multi-source data.

[0055] In terms of hardware physical protection, considering the extremely harsh environment of bridge construction in the field, including extreme temperatures, strong winds, rain, snow, dust, and strong electromagnetic interference from large construction machinery, this invention incorporates in-depth environmental adaptability design for all data aggregation units and edge computing units. The equipment casing is made of customized aluminum alloy or 316L stainless steel with an IP67 protection rating, and the surface is treated with micro-arc oxidation or heavy-duty anti-corrosion powder coating, capable of withstanding 50g of impact acceleration and 20 joules of mechanical collision energy. The casing sealing structure adopts a dual-redundancy design: the first seal uses a high- and low-temperature resistant fluororubber O-ring, and the second seal uses a spring-loaded skeleton oil seal, completely preventing the intrusion of moisture and fine dust. The internal circuit board of the equipment adopts a multi-layer metal shielding structure, with the shielding layer composed of alternating layers of copper foil and aluminum foil, separated by polytetrafluoroethylene insulation material. Based on Faraday's electromagnetic shielding principle and the skin effect, this composite shielding layer attenuates external electromagnetic interference in the 100kHz to 5GHz frequency band by more than 60 dB, ensuring that weak analog signals are not contaminated by high-power welding machines or radio transmitters at the construction site before analog-to-digital conversion. The overall system operating temperature range is extended to -40 degrees Celsius to 85 degrees Celsius.

[0056] Step 2: Multi-level data quality control and preprocessing

[0057] After the raw monitoring data is transmitted to the edge computing unit, due to unavoidable sudden changes in the physical environment, sensor zero-point drift, or transient electromagnetic pulses, the data sequence often contains high-frequency noise, abnormal jumps, and even lost data segments. If such dirty data is directly input into the correlation model, it will cause a serious garbage-in, garbage-out effect, leading to frequent false alarms in the early warning system. To address this, this invention designs a three-level filtering and outlier removal mechanism, supplemented by time series interpolation repair, to deeply clean the data at the edge.

[0058] The first-stage filtering is a hardware-based low-pass filter. For continuous signals such as vibration and strain, a second-order Butterworth low-pass filter is configured at the front end of the signal chain. The Butterworth filter is characterized by its maximally flat amplitude-frequency response within the passband, and its transfer function is expressed as:

[0059]

[0060] in, Let be the order (set to 2 here). The signal angular frequency, This is the cutoff angular frequency. To prevent aliasing, the cutoff frequency is strictly set to half the data sampling frequency (i.e., satisfying the Nyquist sampling theorem). This stage of filtering effectively removes high-frequency thermal noise and radio frequency interference that are beyond the physical essence of the phenomenon.

[0061] The second-stage filtering involves time-domain statistical filtering and outlier removal. For isolated outliers caused by transient disturbances, this invention employs a sliding window algorithm based on the Grubbs criterion. A neighborhood window size is set as... (For example For time-series data sequences within a window Calculate its sample mean and sample standard deviation .

[0062] Calculate each data point in the sequence Grubbs statistic :

[0063]

[0064] Given a confidence level (e.g., 95%, i.e., significance level) The Grubbs critical value can be obtained by looking up a table or calculating using the t-distribution. When it exists At that time, determine the data point Outliers are identified and removed from the valid data sequence, marked as null values ​​(NaN). Grubbs' criterion, compared to the traditional... The criteria demonstrate higher detection rigor and robustness within a small sample sliding window.

[0065] The third-level filtering is a spatial smoothing bilateral filtering algorithm. Traditional moving averages or Gaussian smoothing, while eliminating noise, often blur the step signals caused by sudden changes in structural stress or micro-damage, which is fatal for safety monitoring. This invention introduces a bilateral filtering algorithm, which considers not only the proximity of data points in the spatiotemporal domain but also the similarity of data value ranges. For data points... Its filtered output value The calculation formula is:

[0066]

[0067] in, For neighborhood windows, combine weights Spatial weights Sum range weight Multiplication decision:

[0068]

[0069] In the formula, Controlling the decay rate of the spatial neighborhood, Control the decay rate of amplitude differences. When encountering real physical abrupt changes (such as stress steps), due to Maximum, range weight Approaching 0 makes the weight of smoothly crossing the mutation boundary extremely small, thus perfectly eliminating random white noise while protecting the mutation edge characteristics of the data to the extreme.

[0070] To address data gaps caused by outliers or network interruptions, this invention employs a time-series interpolation method based on adjacent valid data points for data repair, specifically cubic spline interpolation. Compared to linear interpolation or polynomial fitting, cubic spline interpolation ensures that the interpolation curve is not only continuous at data nodes, but also that its first derivative (rate of change) and second derivative (acceleration) are continuous, closely matching the constitutive relationship of continuous changes in bridge physical quantities.

[0071] Suppose there are on both sides of the missing interval 1 valid data point Cubic spline interpolation in each subinterval Construct a cubic polynomial on :

[0072]

[0073] By jointly solving the boundary conditions (such as natural boundary conditions, i.e., the second derivative at the endpoints is 0) and the continuity equations, all polynomial coefficients can be uniquely determined. After the calculation is complete, by substituting the polynomials of the corresponding sub-intervals according to the missing timestamps, smooth and dynamically consistent repair data can be generated.

[0074] To ensure that security assessments are not rendered ineffective by excessively fabricated data, the system employs a strict repair tolerance mechanism. If the length of repaired consecutive missing data exceeds a preset sampling period threshold (e.g., 10 consecutive sampling periods, depending on the parameter type), the system determines that the data segment has lost too many high-frequency features, and forced repair would cause information distortion. In this case, the system refuses automatic repair, marks the corresponding missing data area as invalid, and triggers and reports a manual verification command in the management backend, allowing engineers to intervene and check the health status of sensors and lines. After a complete three-stage filtering and interpolation repair process, the data integrity rate output from the edge can stably exceed 98%, and the data availability rate is greater than 95%, providing a high-standard, high-quality dataset for downstream correlation analysis.

[0075] Step 3: Multi-source data correlation modeling and analysis

[0076] After acquiring high-quality multi-source synchronous data, simply comparing isolated thresholds for each parameter would completely mask the complex risks arising from multi-physics coupling during bridge electromechanical installation. For example, the combined effect of cable stress redistribution caused by rapid temperature changes and light winds could lead to destructive fatigue induced by stress levels that are normally considered safe. To address this, this invention deploys a highly complex multi-source data association model in the edge computing unit to deeply mine high-order features and potential coupling relationships between multi-dimensional parameters.

[0077] This multi-source data association model abandons the traditional empirical formula fitting method and is instead built based on a cutting-edge hybrid deep learning network architecture. In order to fully express the interactive influence of the bridge sensor network in spatial topology and extract the dynamic evolution features in time series, this model deeply integrates a multi-task one-dimensional convolutional neural network (1D-CNN) with a graph neural network (GNN).

[0078] Combination Figure 2 As shown, the data input layer of the hybrid network architecture receives preprocessed tensor blocks, with tensor dimensions of... ,in This represents the total number of sensor nodes. The length of the observation time window (e.g., a sequence of the past 60 seconds). This represents the number of multidimensional feature channels output by the sensor.

[0079] One of the core components of the network is the shared feature extraction layer, which includes a three-layer one-dimensional convolutional neural network module for temporal feature extraction and a graph neural network module for spatial topological interaction.

[0080] First, the time-series data for each sensor node is input into a 1D-CNN module containing three cascaded layers:

[0081] The first layer of the one-dimensional convolutional neural network: It uses 64 convolutional kernels of size 8, performing convolution operations by sliding them along a one-dimensional time axis. For the input signal... Convolution operation outputs feature maps :

[0082]

[0083] in As weight, For bias, the activation function is uniformly a linear rectified function.

[0084]

[0085] To alleviate the gradient vanishing problem, a max pooling layer with a window size of 2 and a stride of 2 is then applied, which halves the feature dimension while preserving the most significant temporal features.

[0086] The second layer is a one-dimensional convolutional neural network: configured with 128 convolutional kernels of size 5, which receive the output of the first layer.

[0087] The third layer is a one-dimensional convolutional neural network with 256 convolutional kernels of size 3, which receives the output of the second layer. Through the nesting of three layers of convolution and pooling, the network can automatically abstract high-order deep temporal feature vectors reflecting structural damage and abnormal working conditions from the original high-frequency vibration or strain sequences.

[0088] After extracting the independent temporal features of each node, a graph neural network module was introduced to establish the force transmission and coupling effects between sensors at different spatial locations. The graph neural network is based on the spatial topology of the physical deployment locations. In the graph, each sensor is considered a node, and edge relationships are established between adjacent sensors or sensors with related force transmission paths.

[0089] edge weight The calculation basis node and nodes Euclidean distance in actual three-dimensional layout space Use the reciprocal to quantify precisely:

[0090]

[0091] If the distance exceeds the preset physical influence cutoff radius, the weight is set to 0. Based on this, an adjacency matrix representing the entire bridge sensor topology is constructed. To prevent gradient explosion, the adjacency matrix is ​​processed using normalized Laplacian:

[0092]

[0093] in It is the identity matrix. This is the angle matrix.

[0094] Building upon this, the graph neural network module employs a graph attention mechanism (GAT). This attention mechanism allows nodes to adaptively learn the importance of features from different neighboring nodes when aggregating neighbor information.

[0095] and Attention coefficient between The calculation is as follows:

[0096]

[0097] in The high-order feature vector output by the CNN module. The feature transformation matrix is ​​a learnable feature. This indicates a splicing operation. This is the attention weight vector. It is obtained by normalizing the attention coefficients within the neighborhood using the Softmax function. Subsequently, the node Update its own spatial coupling feature representation by aggregating neighbor features:

[0098]

[0099] Through this ingenious GNN architecture, the displacement sensor data not only includes local deformation, but also incorporates the dynamic response influence transmitted from the remote vibration sensor, achieving true deep coupling of multi-dimensional physical fields.

[0100] After the shared feature extraction layer, the hybrid network architecture differentiates into multiple task-specific output layers, corresponding to four sub-tasks: stress prediction (predicting future extreme values ​​based on the current state), vibration prediction (predicting modal evolution), displacement prediction (predicting cumulative deflection), and environmental impact assessment (quantifying the additional response coefficient of the environment to the structural state).

[0101] Each task-specific output layer consists of a fully connected layer with 128 neurons and a final regression output layer. Through this "shared bottom extraction + top multi-task branching" architecture, the model achieves knowledge transfer and feature sharing across multiple tasks, which not only reduces the risk of overfitting but also significantly improves the consistency and overall accuracy of multi-dimensional parameter predictions.

[0102] Model training is conducted in the cloud-based analytics unit, relying on a large-scale GPU cluster. The training dataset contains tens of thousands of historical monitoring data sets, comprehensively covering the structure's resting state, routine construction conditions, extreme high / low temperature conditions, typhoon and thunderstorm attacks, and past occasional dangerous stress exceeding events.

[0103] To simultaneously optimize the four prediction tasks and ensure the stability of the graph topology, the network employs a multi-task joint optimization strategy during training. The total loss function of the hybrid network architecture is as follows. Joint regression loss of multi-task deep learning networks Topology Reconstruction Loss of Graph Neural Networks The weighted summation consists of:

[0104]

[0105] The weighting ratio is set based on empirical verification. .

[0106] For regression loss It is itself a weighted sum of the mean squared error (MSE) loss functions of the four subtasks, that is:

[0107]

[0108] Weight parameters During training, the network can be dynamically and adaptively adjusted through the homoscedasticity uncertainty mechanism, so that it can automatically focus on the more difficult tasks to learn at different convergence stages.

[0109] The training optimizer employs an adaptive moment estimation algorithm, dynamically adjusting the learning rate of each parameter using first-order and second-order moment estimations of the gradient. The initial learning rate is set to 0.001, and a cosine annealing learning rate decay strategy is used to ensure the model converges quickly and stably to the global optimum in complex high-dimensional error surfaces. The optimal model parameter file, after incremental training in the cloud, is periodically distributed and updated to each edge computing unit. With hardware acceleration from the NPU (Neural Processing Unit) in the edge computing unit, the time for a single model forward inference is strictly controlled to within 100 milliseconds, perfectly meeting the stringent computational performance requirements for real-time risk assessment in bridge construction safety.

[0110] Step 4: Comprehensive assessment and tiered early warning of safety risks

[0111] While multi-source data association models output quantitative indicators of stress, vibration, displacement, and environmental impact for future periods, these are merely underlying engineering physical quantities. Engineering managers require intuitive, comprehensive, and instructive safety status assessments. Therefore, this invention integrates a risk comprehensive assessment module at the edge. Based on the aforementioned prediction outputs, it employs a deep combination of the Analytic Hierarchy Process (AHP) and Fuzzy Comprehensive Evaluation (FCE) to perform multi-level quantitative scoring of the safety status of electromechanical installations.

[0112] First, the system constructs a logically rigorous multi-level evaluation index system. This system comprises four primary indicators from top to bottom: structural safety indicators (weighted) Equipment status indicators (weight) Environmental impact indicators (weights) ) and historical trend indicators (weight) ).

[0113] Each primary indicator has several detailed secondary indicators, totaling 20 evaluation factors. For example:

[0114] Structural safety indicators include: stress level indicators, vibration amplitude indicators, displacement deformation indicators, fatigue damage indicators, and load-bearing capacity indicators.

[0115] The environmental impact indicators include: temperature impact indicators, humidity corrosion indicators, wind speed and aerodynamic indicators, corrosive medium concentration indicators, and micro-vibration effects indicators.

[0116] The determination of weights relies on the Analytic Hierarchy Process (AHP). Combining bridge mechanics experts with a large database of historical failure cases, the Saaty 1-9 scaling method is used to pairwise compare and score the importance of indicators within the same level relative to the target at the next higher level, constructing a judgment matrix. :

[0117]

[0118] in Indicators relative to indicators The degree of importance. This is determined by finding the largest eigenvalue of the judgment matrix. and its corresponding normalized eigenvectors This yields the weight vector for each indicator. To ensure the logical rationality of the scoring, a consistency check is required. The consistency index is then calculated. And find the random consistency index based on the matrix order. Finally, the consistency ratio is calculated. .when At that time, it was assumed that the judgment matrix had satisfactory consistency and the weight vector was effective. In the implementation of this system, the weights of the four primary indicators were established as follows: structural safety 0.45, equipment status 0.20, environmental impact 0.15, and historical trend 0.20, highlighting the core position of structural mechanical response.

[0119] Determining the weights solves the problem of "what proportion" each indicator should account for. The subsequent fuzzy comprehensive evaluation rule is used to solve the problem of "how to quantify the evaluation scale".

[0120] The system predefines a set of comments, corresponding to different security levels:

[0121]

[0122] For each underlying secondary indicator, the system constructs a fuzzy membership function based on real-time monitoring values, model prediction values, and current bridge design specifications (such as the "Specifications for Testing and Evaluation of Bearing Capacity of Highway Bridges"). For example, for the stress level indicator, a semi-trapezoidal membership function is used. When the stress value is between 0% and 60% of the allowable stress, the membership degree for "normal" is 1; when it exceeds 85% of the allowable stress, the membership degree for "dangerous" rapidly rises to 1.

[0123] By calculating the membership vector of each secondary indicator at each rating level, a single-factor evaluation fuzzy matrix is ​​constructed. Combine the weight vectors of each secondary indicator under its parent node (primary indicator). Calculate the fuzzy evaluation vector of the primary indicator. Finally, the weight vector of the primary indicators is used. For all High-level fuzzy synthesis is performed to obtain the final total fuzzy evaluation vector. .

[0124] The system evaluates the total fuzzy evaluation vector. The weighted average method is applied to defuzzify the qualitative membership degree, transforming it into a quantitative percentage-based comprehensive score. :

[0125]

[0126] Based on the calculated percentage-based comprehensive score, the system implements strict tiered early warning systems and issues warning commands in real time. The specific early warning level classifications and response strategies are as follows:

[0127] Level 1: Normal Level, corresponding to a comprehensive score of 85 to 100. This indicates that the internal forces of the structure are evenly distributed, environmental disturbances are minimal, and electromechanical installation work can proceed normally according to the established construction schedule without intervention.

[0128] Level 2: Attention Level, corresponding to a comprehensive score of 70 to 84. This indicates that the system has detected slight nonlinear deformation or vibration energy accumulation at a specific frequency, which may indicate potential stress redistribution. The system backend will automatically increase the frequency of high-frequency sensor wake-up and data upload, generate automated monitoring reports, and push "inspection suggestions" to on-site construction safety management personnel, requiring manual visual verification of suspicious cable towers or anchorages.

[0129] Level 3: Alert Level, corresponding to a comprehensive score of 55 to 69. This indicates that the model predicts some components are approaching the allowable stress boundary, or that environmental wind loads are undergoing initial fluid-structure interaction resonance with the structure, posing a significant risk. At this point, the edge computing unit will immediately cut off the power to secondary, non-critical equipment to reduce on-site hazards and send the highest-priority audible and visual alarm SMS and email to the mobile applications of the project's chief engineer and supervision unit. A multi-departmental joint expert consultation mechanism will be forcibly activated, and the application of new loads at the construction site must be suspended. Preventive measures such as adding temporary wind cables and adjusting counterweights must be taken.

[0130] Level 4: Danger Level, corresponding to a comprehensive score of 0 to 54. This indicates that the bridge's electromechanical installation is on the verge of catastrophic failure (such as a sudden loss of stress in the main beam, or unexpected and significant deflection and tearing). The system directly triggers a mandatory on-site construction alarm broadcast, initiating an automated emergency response plan. All on-site construction personnel are required to immediately cease all work and evacuate to a safe zone according to the planned route. Core data is forcibly and directly written to the cloud black box area at high frequency until structural intervention takes effect and the assessment indicators move out of the danger zone.

[0131] It is worth mentioning that the scoring thresholds for each warning level in this system are not static, but rather incorporate a dynamic adaptive mechanism. The edge computing node extracts prior structural knowledge about the bridge's current season (e.g., steel expansion in summer, contraction in winter) and specific construction stage (e.g., cantilever construction stage, large-tonnage electromechanical hoisting and closure stage). Combined with statistical patterns of historical data distribution, this information is used to fine-tune the warning thresholds in real time, fundamentally eliminating false alarms and missed alarms caused by rigid thresholds. Since most of the computational load is pre-processed at the edge computing nodes, the entire process—from sensor capture of abnormal data mutations, through three levels of filtering and purification, deep network prediction, to fuzzy comprehensive evaluation and finally triggering the audible and visual warning—is remarkably compressed to less than 30 seconds (compared to just over ten seconds in normal systems). This provides an extremely valuable golden window for intervention to avoid major safety accidents.

[0132] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-source data-driven method for safety monitoring of bridge electromechanical installation, characterized in that, Includes the following steps: Step 1: Construct a collaborative acquisition architecture. Through a unified data acquisition protocol and time reference, synchronously acquire multi-source monitoring data during the bridge electromechanical installation process. The multi-source monitoring data includes at least structural stress data, vibration data, displacement data, and environmental data. Step 2: Perform multi-level filtering and outlier removal on the acquired multi-source monitoring data, and repair the missing data areas based on the time series interpolation method of adjacent valid data points to obtain preprocessed multi-source monitoring data; Step 3: Input the preprocessed multi-source monitoring data into a pre-trained multi-source data association model for association modeling and analysis, extract high-order feature representations between multi-dimensional parameters, and output stress prediction results, vibration prediction results, displacement prediction results and environmental impact assessment results. The multi-source data association model is built based on a multi-task deep learning network architecture. Step 4: Based on the output results of the multi-source data association model, the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation method are used to comprehensively score the safety status of electromechanical installation, and different warning levels are assigned according to the scoring results for graded warning.

2. The method for multi-source data-driven safety monitoring of bridge electromechanical installation according to claim 1, characterized in that, In step 1, the data acquisition protocol follows a unified data encapsulation format. A single frame data packet includes four fields in sequence: acquisition timestamp, sensor type identifier, original measurement value, and cyclic redundancy check code. Various sensing devices used to acquire the multi-source monitoring data undergo clock calibration through a unified time synchronization protocol to unify the time reference.

3. The method for multi-source data-driven safety monitoring of bridge electromechanical installation according to claim 1, characterized in that, In step 2, the multi-stage filtering process specifically includes a first-stage filtering, a second-stage filtering, and a third-stage filtering. The first-stage filtering uses a second-order Butterworth low-pass filter to suppress high-frequency interference signals, and the cutoff frequency of the second-order Butterworth low-pass filter is set to half of the data sampling frequency. The second-stage filtering uses time-domain statistical filtering based on the Grubbs criterion to detect and remove outliers. The third-stage filtering uses a spatial smoothing bilateral filtering algorithm to smooth noise interference while maintaining the characteristics of data mutation.

4. The method for multi-source data-driven safety monitoring of bridge electromechanical installation according to claim 1, characterized in that, In step 2, the interpolation method is cubic spline interpolation. When repairing missing data areas, if the length of a single continuous missing data repair exceeds the preset sampling period threshold, the corresponding missing data area is marked as an invalid area and a manual verification report is triggered.

5. The method for multi-source data-driven safety monitoring of bridge electromechanical installation according to claim 1, characterized in that, In step 3, the multi-task deep learning network architecture includes a shared feature extraction layer and multiple task-specific output layers. The shared feature extraction layer includes three sequentially connected one-dimensional convolutional neural networks, with a max pooling layer for feature dimensionality reduction after each one-dimensional convolutional neural network layer. The multiple task-specific output layers correspond to four sub-tasks: stress prediction, vibration prediction, displacement prediction, and environmental impact assessment. Each task-specific output layer consists of a fully connected layer and an output layer.

6. The method for multi-source data-driven safety monitoring of bridge electromechanical installation according to claim 1, characterized in that, In step 3, the multi-source data association model adopts a hybrid network architecture that combines a multi-task deep learning network with a graph neural network. The multi-task deep learning network is used to extract temporal feature representations, and the graph neural network is used to model the spatial topological relationships of the sensors. The graph neural network is constructed based on the spatial coordinates of the deployment locations of each sensor and establishes edge relationships in the graph structure between adjacent sensors. The edge weight values ​​of the edge relationships are the reciprocal of the Euclidean distance between the two corresponding sensors in the actual deployment space.

7. The method for safety monitoring of bridge electromechanical installation driven by multi-source data according to claim 1, characterized in that, In step 4, the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation method are performed based on a constructed multi-level evaluation index system. The multi-level evaluation index system includes four primary indicators: structural safety index, equipment status index, environmental impact index, and historical trend index. Each primary indicator has several secondary indicators. The specific process for comprehensively scoring the safety status of electromechanical installation is as follows: construct a judgment matrix according to the scaling method of the AHP and calculate the weights corresponding to each primary indicator and each secondary indicator. A fuzzy evaluation matrix is ​​established, and the qualitative evaluation is transformed into a quantitative score based on the weight of each indicator. Finally, the comprehensive score is calculated and accumulated, and the warning level is divided into four warning levels from low to high: normal, attention, alert, and danger.

8. The method for safety monitoring of bridge electromechanical installation driven by multi-source data according to claim 1, characterized in that, The bridge electromechanical installation safety monitoring method is applied to a three-level network architecture consisting of a data aggregation unit, an edge computing unit, and a cloud analysis unit. Step 1 is executed by the data aggregation unit, and steps 2 to 4 are executed by the edge computing unit deployed on-site to achieve local risk warning decision-making. The edge computing unit synchronizes the processed data to the cloud analysis unit, which is used for historical data storage, regular model updates, and comprehensive evaluation report generation. The sensing devices used to acquire the multi-source monitoring data are all equipped with protective housings made of metal. The protective housings have a double sealing structure inside, including a first rubber sealing ring and a second skeleton oil seal. The internal circuits use a multi-layer metal shielding structure with alternating copper foil and aluminum foil for electromagnetic isolation.

9. A multi-source data-driven bridge electromechanical installation safety monitoring system, characterized in that, include: The multi-source collaborative acquisition module is used to build a collaborative acquisition architecture. Through a unified data acquisition protocol and time reference, it synchronously acquires multi-source monitoring data during the bridge electromechanical installation process. The multi-source monitoring data includes at least structural stress data, vibration data, displacement data, and environmental data. The data quality control and preprocessing module is used to perform multi-level filtering and outlier removal on the acquired multi-source monitoring data, and to repair the missing data areas based on the time series interpolation method of adjacent valid data points, so as to obtain the preprocessed multi-source monitoring data. The multi-source data association modeling module is used to input the preprocessed multi-source monitoring data into a pre-trained multi-source data association model for association modeling analysis, extract high-order feature representations between multi-dimensional parameters, and output stress prediction results, vibration prediction results, displacement prediction results, and environmental impact assessment results. The multi-source data association model is constructed based on a multi-task deep learning network architecture. The risk comprehensive assessment module is used to comprehensively score the safety status of electromechanical installation based on the output results of the multi-source data association model, using the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation method, and to classify the results into different warning levels for graded warning.

Images