Convolutional neural network-based intensive single-column pier overload overturning risk assessment method

Abstract

The invention provides an intensive single-column pier overload overturning risk assessment method based on a convolutional neural network, and belongs to the technical field of bridge engineering safety assessment. Vehicle load parameters, a temperature effect, concrete creep and a vehicle bridge interaction dynamic amplification factor are integrated into multi-dimensional tensor input; by means of staged training and gradient descent optimization of the convolutional neural network, second-level counter-force prediction and dynamic amplification adjustment value output are achieved, counter-force data are expanded through a synthetic overturning axis method, errors are controlled within a reasonable range, after the bridge health monitoring system is integrated with an existing bridge health monitoring system, sensor data can be collected in real time, four-level risk levels can be output in real time, and the safety of the bridge health monitoring system is improved. According to the method, the overloading and overturning risks of the single-column pier bridge are respectively safe, critical, dangerous and immediate intervention, so that the bridge maintenance cost is reduced, resource waste is avoided, the structural stability in a high-traffic-volume environment is improved, the method is suitable for various bridge types such as straight bridges and curved bridges, intensive assessment of the overloading and overturning risks of the single-column pier bridge is realized, and the real-time performance and accuracy of assessment are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of bridge engineering safety assessment technology, and more specifically, to an intensive method for assessing the risk of overloaded overturning of single-column piers based on convolutional neural networks. Background Technology

[0002] Single-column pier bridges, due to their advantages such as small footprint, convenient design, low construction cost, and easy adaptation to the surrounding environment, have become one of the most common beam bridges for urban overpasses and ramps. However, the increasing traffic volume and the rise in heavy-duty vehicles have increased the risk of collapse for single-column pier bridges, making them more prone to overall collapse accidents with extremely negative social impacts. Bridge health monitoring can effectively issue early warning signals for bridges under special climatic or traffic conditions, or when the bridge's operational status is abnormally high, providing a basis and guidance for bridge maintenance, repair, and management decisions. Current bridge health monitoring systems can perform routine bridge diagnosis and post-accident cause analysis, but they cannot yet provide rapid early warnings of accidents to prevent them from occurring. Finite element analysis software such as Midas, Abaqus, or Ansys can perform relatively accurate stress and strain calculations for bridge structures, but the analysis time is long, making it impossible to quickly assess the situation before an accident occurs, resulting in poor timeliness. Currently, artificial intelligence has entered a new era of development. Neural networks, based on their advantages of spanning physical quantities, high accuracy, and fast calculation speed, can effectively solve the problems of poor timeliness and slow calculation speed in current bridge health monitoring systems.

[0003] The following technical problems exist in the existing technology: 1. Existing health monitoring of single-column pier bridges relies on finite element software, which takes several hours to analyze and cannot meet the real-time early warning requirements of overload scenarios. This can easily lead to overturning accidents. In addition, traditional methods only consider static loads and ignore the influence of other factors such as ring and vehicle interaction, resulting in large prediction errors and insufficient accuracy. Especially in complex environments, there is a lack of models that integrate multi-source data, which cannot achieve second-level response, thus limiting the safety management of high-traffic bridges and is not conducive to practical use. 2. Existing technologies are limited to static analysis and time-based diagnosis, ignoring the influence of other factors such as temperature and concrete creep. This leads to inaccurate risk classification and neglect of potential failure paths, resulting in a passive approach to maintenance. In particular, in complex structures such as curved bridges, the assessment of the overturning process is not detailed enough, the risk level classification is simplistic and there is a lack of effective reinforcement recommendations. This results in insufficient inspection of the risk of bearing detachment, serious waste of resources, and difficulty in meeting the high load requirements of urban bridges. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for assessing the overload overturning risk of intensive single-column piers based on convolutional neural networks, the method comprising: A full-bridge finite element model is established based on the bridge's geometric parameters to simulate vehicle load parameters, temperature effects, concrete creep, and the interaction between vehicles and the bridge. Support reaction force data are calculated and subjected to hierarchical normalization to extract key reaction force parameters. Based on the full-bridge finite element model, a convolutional neural network is designed. The multidimensional tensor is used as input data, the output is the key reaction force parameter, and the mean square error is used as the loss function. The network is optimized to convergence through staged training and gradient descent, and the output is a dynamically amplified and adjusted value. The key reaction parameters are input into the linear regression model. Based on the bridge symmetry and the effects of temperature, concrete creep, and vehicle-bridge interaction, the key reaction parameters are expanded using the synthetic overturning axis method. The core reaction parameters are predicted by the linear regression model to generate the support concurrent reaction force combination. By combining bridge geometric parameters, the risk of voiding under temperature and concrete creep is examined, the overturning stability coefficient is calculated, dynamic amplification and concrete creep effect are incorporated, the overturning process is evaluated in four stages and classified into four risk levels, and a convolutional neural network is integrated into the bridge health monitoring system to collect sensor data in real time and output dynamic early warning and reinforcement suggestions.

[0005] In another aspect, embodiments of the present invention also provide a method for assessing the risk of overloaded overturning of intensive single-column piers based on convolutional neural networks. The hardware of the method includes a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code. The processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0006] Based on the above, the beneficial effects of the intensive single-column pier overload overturning risk assessment method based on convolutional neural networks of the present invention are as follows: This invention integrates vehicle load parameters, temperature effects, concrete creep, and vehicle-bridge interactive dynamic amplification factors into a multidimensional tensor input. Utilizing staged training and gradient descent optimization of a convolutional neural network, it achieves second-level reaction force prediction and dynamically amplified adjustment value output. By synthesizing the overturning axis, the reaction force data is expanded, with errors controlled within a reasonable range. After integration with existing bridge health monitoring systems, it can collect sensor data in real time and output four risk levels: safe, critical, dangerous, and requiring immediate intervention. This effectively avoids the limitations of post-hoc diagnosis in existing systems, significantly reduces the probability of overturning accidents, thereby reducing bridge maintenance costs, preventing resource waste, and improving structural stability under high traffic conditions. It is applicable to various bridge types, including straight and curved bridges, enabling intensive assessment of overload overturning risks in single-column pier bridges, greatly improving the real-time performance and accuracy of the assessment.

[0007] This invention, based on highway bridge and culvert design specifications and combined with bridge geometric parameters, examines the risk of bearing delamination, calculates the overturning stability coefficient, incorporates dynamic amplification and concrete creep effects, assesses the overturning process in four stages and classifies it into four risk levels, automatically outputs reinforcement recommendations based on the risk level, and simulates the reinforcement effect through a digital twin platform. This provides a comprehensive prevention strategy for the overload overturning risk management of single-column pier bridges, achieving closed-loop management from assessment to action, improving the accuracy of bridge remaining life prediction, reducing human intervention errors, promoting the transformation of bridge maintenance from passive repair to proactive prevention, and enhancing the safety and sustainability of long-term bridge operation. Attached Figure Description

[0008] Figure 1 This is a flowchart of the intensive single-column pier overload overturning risk assessment method based on convolutional neural networks provided in this embodiment of the invention.

[0009] Figure 2 This is a line graph showing the real-time model iteration count and loss value of the intensive single-column pier overload overturning risk assessment method based on convolutional neural networks provided in this embodiment of the invention.

[0010] Figure 3 This is a scatter plot of the target value and predicted value of the intensive single-column pier overload overturning risk assessment method based on convolutional neural network provided in this embodiment of the invention. Figure 4 This is a data processing relationship diagram of the intensive single-column pier overload overturning risk assessment method based on convolutional neural networks provided in this embodiment of the invention. Detailed Implementation

[0011] 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 only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Specific Implementation Example 1: like Figures 1 to 4 As shown, the intensive method for assessing the overload overturning risk of single-column piers based on convolutional neural networks includes: Step S100: Establish a full-bridge finite element model based on the bridge's geometric parameters. Set vehicle load conditions according to the "General Specifications for Highway Bridge and Culvert Design" (JTG D60-2015). Simulate data on vehicle load parameters, temperature effects, concrete creep, and vehicle-bridge interaction. Calculate bearing reaction force data and perform hierarchical normalization processing. Extract key reaction force parameters to ensure the assessment considers the dynamic impact of actual operation, improving data representativeness and model generalization ability. Bearing reaction force data refers to the reaction force generated by bridge bearings when subjected to superstructure and external loads, a key indicator reflecting the bridge's stress state. The dynamic amplification factor is the load amplification effect caused by vibration and road surface roughness when vehicles pass over the bridge, typically ranging from 1.1 to 1.5. Data acquisition can be achieved using a drone equipped with a LiDAR to scan the bridge and generate a 3D point cloud model, ensuring geometric parameter accuracy down to the millimeter level. Simulating the bridge's dynamic response under real operating conditions using the finite element model allows for the identification of bearing reaction force variation patterns, providing high-quality input data for subsequent convolutional neural networks while reducing the cost and risk of on-site testing.

[0013] Vehicle-bridge interaction refers to the phenomenon where the stress state of a bridge changes due to the interaction between vehicle motion (speed, acceleration, etc.), vehicle load (axle load, wheelbase, off-center loading position, etc.), and the bridge's dynamic response (vibration, road surface roughness, etc.) when vehicles cross the bridge. This interaction is represented by a dynamic amplification factor, typically ranging from 1.1 to 1.5. This factor reflects the load amplification effect caused by vibration and road surface roughness when vehicles cross the bridge, as well as the dynamic impact of vehicles on the bridge during actual operation. It is a key input for bridge stress state analysis. The parameters for vehicle-bridge interaction are set based on the "General Specifications for Highway Bridge and Culvert Design" (JTG D60-2015), finite element model simulation, on-site data collection, stochastic traffic flow model, and environmental factors. Relevant data on vehicle-bridge interaction (such as bearing reactions and dynamic amplification factors) are generated through finite element model simulation, then processed through hierarchical normalization, stored in CSV format, and uploaded to a local server or cloud database as multidimensional tensor input. This input is provided to the convolutional neural network in step S200 for training, used to predict key reaction parameters. Subsequently, in step S300, the vehicle-bridge interaction data is expanded using the synthetic overturning axis method and linear regression model to generate a combined bearing reaction force. Finally, in step S400, this data is used to calculate the overturning stability coefficient and risk classification. The vehicle-bridge interaction data is collected in real-time by sensors in the bridge health monitoring system and uploaded to the cloud database, supporting real-time risk assessment and dynamic early warning. By simulating the dynamic load of vehicles and the vibration response of the bridge, it accurately captures the stress changes of the bridge during actual operation.

[0014] Vehicle load parameters mainly include axle load, wheelbase, lateral offset position, vehicle speed, road surface roughness grade, and the resulting dynamic amplification factor. These parameters reflect the impact of vehicles on the bridge during actual operation and are derived from simulations of bridge design specifications and actual operational data. Vehicle load conditions are set according to the "General Specifications for Highway Bridge and Culvert Design" (JTG D60-2015). The dynamic response of the bridge under real operational conditions is simulated using a full-bridge finite element model. For example, axle load ranges from 10 to 50 tons, wheelbase from 2 to 6 meters, lateral offset position ranges from -1.5 meters to 1.5 meters, vehicle speeds range from 20 to 80 km / h, and road surface roughness grades range from A to C. The dynamic amplification factor is derived from the load amplification effect caused by vibration and road surface roughness when vehicles pass over the bridge. It is typically 1.1 to 1.5 and is generated through at least 200 simulations (each simulation lasting 5 minutes).

[0015] When processing vehicle load parameters, a segmented modeling strategy was first adopted in the full-bridge finite element model. Key bridge components such as beams, piers, and supports were modeled separately, and then integrated into a full-bridge model through coupling analysis to improve computational efficiency. Multiphysics coupling analysis was introduced, incorporating wind and seismic loads to further simulate the bridge response under extreme environments. Then, the generated support reaction force data underwent hierarchical normalization, using Python's NumPy library to map the values ​​to the 0-1 range, and monitoring the data distribution to avoid overflow. Outliers, such as data exceeding three standard deviations, were removed using data cleaning tools to ensure dataset quality. Key reaction force parameters extracted included the maximum value of the double-support reaction force at the side piers (unit: kN, precision: 0.1 kN) and the standard deviation of the distribution of the single-support reaction force at the middle pier (precision: 0.01). The processed vehicle load parameters and their related support reaction force data are extracted as key reaction force parameters, generating a dataset covering straight bridges, curved bridges, and at least 20 bridge types, containing at least 1000 samples, stored in CSV format, and uploaded to a local server or cloud database. This data is then used for training the convolutional neural network in subsequent step S200, as well as for expansion and risk assessment in step S300. This improves the representativeness of the data and the generalization ability of the model. By simulating the impact of vehicle load parameters on the interaction between the vehicle and the bridge, the variation law of support reaction force is identified, providing high-quality input data for the convolutional neural network, reducing the cost and risk of field tests, supporting real-time risk assessment, avoiding bridge overturning accidents, and improving the efficiency and safety of bridge risk management under complex working conditions.

[0016] Concrete creep specifically refers to the continuous deformation phenomenon of concrete under long-term loads, including short-term and long-term concrete creep. It is incorporated into bridge simulations using staged loading to differentiate between the different concrete creep effects during construction and operation. The concrete creep coefficient, as a key parameter, is included in the multidimensional tensor input, combined with temperature effects and vehicle-bridge interaction. The samples are grouped according to the concrete creep stage: short-term (28 days) and long-term (1 to 10 years), with at least 300 samples in each group. Concrete creep affects the long-term stability of bridges, leading to increased beam deflection, redistribution of support reactions, and potentially causing support disengagement risks, especially amplifying the risk of overturning under temperature gradients and heavy vehicle loads. For example, in curved bridges, concrete creep may cause an increase in beam rotation angle and a decrease in the stability coefficient. The concrete creep data comes from finite element model simulation and is updated periodically according to the JTG 3362-2018 standard. The bridge's geometric parameters are measured and imported and simulated using finite element software such as Midas Civil or Ansys. Data acquisition combines a total station and a laser scanner to obtain the pier tilt angle and beam deflection data, with the accuracy controlled within 0.005 meters.

[0017] Step S101: When establishing the finite element model of the whole bridge, first measure and collect the geometric parameters of the bridge, including the support spacing accurate to the centimeter level, the beam weight distribution per meter, the pier height, and material properties, such as concrete strength grade C30 to C60 and steel yield strength HRB40. Use measuring instruments such as laser scanners to acquire data to ensure that the accuracy is within 0.01 meters. Then, the bridge geometry parameters were imported using finite element software, such as Midas Civil or Ansys, and the vehicle load conditions were set. The axle load range was set to 10 to 50 tons, the wheelbase to 2 to 6 meters, and the lateral load position variation range to -1.5 meters to 1.5 meters. Temperature effect simulation was performed, with uniform temperature variation set from -30℃ to 60℃. Data was collected every 5℃ and updated every 12 hours. The gradient temperature was set to 5℃ per meter in both positive and negative directions. Finally, concrete creep was calculated using a long-term self-weight deformation model updated periodically based on the JTG 3362-2018 standard. Through vehicle-bridge interactive analysis, vehicle speeds were set to 20 to 80 km / h, step sizes to 10 km / h, and pavement roughness grades to A to C. At least 200 simulations were run, each lasting 5 minutes. The dynamic scaling factor, ranging from 1.1 to 1.5, was calculated and exported, and support reaction force data was generated and stored on the local server.

[0018] The long-term self-weight deformation model is a concrete creep simulation model based on the JTG 3362-2018 standard. It is used to calculate the continuous deformation of concrete under long-term loads, including the concrete creep coefficient and staged loading, covering both short-term and long-term stages, and incorporating temperature effects and vehicle interaction. After importing the bridge's geometric parameters into the finite element method software, the long-term self-weight deformation simulation was set up. First, the concrete creep data was updated periodically (e.g., temperature data was updated every 12 hours, but for long-term concrete creep, updates were timed periodically). Second, staged loading was used to simulate different concrete creep effects during construction and operation phases. Then, through vehicle-bridge interaction analysis, at least 200 simulations were run, each lasting 5 minutes, to generate bearing reaction force data. The concrete creep time ranged from 28 days to 10 years, and the synthetic overturning axis method was used to handle the expansion of the curved bridge reaction force. During processing, the bearing reaction force data were first grouped according to the concrete creep stage (short-term 28 days and long-term 1 to 10 years), with at least 300 samples in each group. Then, the values ​​were normalized using Python's NumPy library, mapping them to the 0-1 range. Outlier detection, combined with machine learning algorithms (such as Isolation Forest), removes outliers (e.g., data exceeding 3 standard deviations). Data is stored on a local server, generating a CSV format dataset. Concrete creep is calculated to ensure the simulation of the bridge's dynamic response under real operational conditions, providing high-quality input data to the convolutional neural network, reducing the cost and risk of field testing, and serving as training data for step S200 and expansion data for step S300.

[0019] Step S102: When processing the support reaction force data to extract key reaction force parameters, the support reaction force data is first grouped by temperature range: -30℃ to 0℃, 0℃ to 30℃, and 30℃ to 60℃, and by concrete creep stage: short-term (28 days) and long-term (1 to 10 years), with at least 300 samples in each group. Then, the NumPy library in Python is used to normalize each group of data, mapping the values ​​to the 0 to 1 range. During the processing, the data distribution is monitored to avoid overflow. Key reaction force parameters are extracted, including the reaction force of the double supports of the side pier. The maximum value (in kN, with a precision of 0.1 kN), the standard deviation of the distribution of the reaction force at a single support of the pier (with a precision of 0.01), the beam rotation angle (with a precision of 0.001 radians), and the long-term coefficient of variation (with a precision of 0.001) are used to generate a dataset covering straight bridges, curved bridges, and at least 20 other bridge types. The dataset contains at least 1000 samples, is stored in CSV format, and outliers, such as data exceeding 3 times the standard deviation, are removed using data cleaning tools to ensure the quality of the dataset. Through the complete chain from simulation to extraction, standardized inputs are provided for step S200 to realize the transformation of physical data into machine learning data.

[0020] In this embodiment, a laser scanner, using a laser beam to measure distances on the surface of an object, can generate high-precision three-dimensional point cloud data for bridge geometric modeling. The continuous deformation of concrete under long-term load is called concrete creep, which affects the long-term stability of the bridge. Road surface smoothness is categorized into A (smooth), B (moderate), and C (poor) grades, affecting the dynamic effects of vehicle-bridge interaction. During measurement, a total station and laser scanner can be combined to obtain pier tilt angle and beam deflection data, with accuracy controlled within 0.005 meters. When simulating vehicle loads, a random traffic flow model is introduced to simulate scenarios with multiple vehicles passing simultaneously during peak hours. Temperature effect simulation incorporates a solar radiation model to consider the non-uniform temperature distribution on the bridge surface caused by solar radiation. Concrete creep simulation employs staged loading to simulate different concrete creep effects during construction and operation. Precise geometric parameters and multi-condition simulations ensure that the model reflects the true stress state of the bridge under complex environments, providing diverse data samples for subsequent machine learning training.

[0021] Key reaction force parameters are core indicators reflecting the stress state of bridges, including maximum reaction force and standard deviation of distribution, used to assess overturning risk. Time series analysis can be introduced when grouping data, and the reaction force data can be dynamically grouped in conjunction with daily traffic flow changes. Normalization processing adopts Min-Max standardization or Z-score standardization, and the optimal method is selected according to the data distribution. Outlier detection can be combined with machine learning algorithms (such as Isolation Forest) to further improve data cleaning efficiency. Visualization tools (such as Matplotlib to generate reaction force distribution maps) can be developed to help engineers verify data quality. Through grouping and normalization processing, data noise is reduced, the representativeness of key reaction force parameters is improved, and standardized feature inputs are provided for subsequent neural network training.

[0022] Step S200: Design a convolutional neural network based on the full-bridge finite element model, take multidimensional tensors as input data, output key reaction force parameters and use mean square error as the loss function, optimize to convergence through staged training and gradient descent method, output dynamic amplified adjustment value, learn the nonlinear relationship between vehicle load parameters and key reaction force parameters from multi-source data, realize efficient prediction, replace traditional finite element iterative calculation, and improve the calculation speed and adaptability of evaluation; The specific method of convolution operation in neural networks is as follows: Suppose the input data is a one-dimensional sequence Using a convolution kernel of length k, The convolution operation at position j is then: ; in, Let x be the nth element of the input sequence, representing the last value of the vehicle load parameter sequence, and the input sequence x contains a multidimensional tensor; is the k-th weight of the convolution kernel, representing the last parameter of the convolution kernel, which is learned through training of the convolutional neural network and is used to perform weighted summation on a local window of the input sequence; It is the j-th output value after the convolution calculation, representing an element in the result sequence after the convolution calculation. It is the output after local feature extraction in the convolutional neural network and is used for subsequent layer processing. It refers to the local window data in the input sequence, specifically the i-th input value in the window starting from position j; It is the i-th weight value of the convolution kernel parameters, which is a parameter learned during the training of the convolutional neural network and used to extract features from the input data; It is a bias term used to adjust the offset of the convolution output, thereby improving the model's fitting ability.

[0023] By taking a one-dimensional sequence as input, such as a multi-dimensional tensor, and selecting a convolution kernel of length k, for each position j in the output sequence, a local window of length k is extracted from the input sequence (from...). arrive However, the index is ), each element in the window With the corresponding convolutional kernel weights Multiply and sum the results, then add the bias b to obtain the output. This process is repeated, with the sliding window traversing the entire input sequence to generate the output sequence. The step size s determines the sliding interval (if s > 1, the calculation is skipped), and the padding p determines whether to pad the input with zeros (SAME mode keeps the output length the same, while VALID mode does not pad with zeros, resulting in a shortened output). This is used to extract key features from the multidimensional tensor, quickly predict key reaction force parameters, and construct a predictive model of vehicle load parameters and key reaction force parameters, improving computational efficiency and achieving second-level evaluation. This replaces time-consuming finite element simulations, and overall, supports real-time risk assessment to avoid bridge overturning accidents.

[0024] The output length is calculated as follows: ; in, This is the length of the output sequence; is the length of the input sequence; k is the kernel length; p is the padding length, which is padded with zeros to make the output length the same as the input length, and without padding, the output is shorter than the input; s is the stride, which represents the distance the kernel slides each time, and the default value is 1. If the stride s > 1, the kernel will jump in units of s; if s = 1 and p = 0, it is VALID mode, and the output is shortened; if p = (k-1) / 2, it is SAME mode, and the output length is close to the input. By calculating the output length, the hyperparameters of the convolutional neural network can be optimized, such as the kernel size and pooling strategy, to improve the model training efficiency and prediction accuracy, and ultimately serve the rapid assessment of the risk of bridge overturning due to overload.

[0025] The network was trained using data samples generated by finite element analysis. The hyperparameters of the network, including convolution kernel size, pooling strategy, and activation function, were optimized. The accuracy of the model was evaluated by the mean square error of the loss function until convergence. Figure 2 The diagram shows the training iterations versus the loss function of the convolutional network in this example.

[0026] Step S201: Construct a convolutional neural network. Using the TensorFlow framework, set the input layer to a multidimensional tensor containing the axial weight distribution vector, load location vector, temperature gradient distribution, concrete creep coefficient, and vehicle-bridge interaction dynamic amplification factor, with dimensions such as [number of samples, number of features, time step]. Set up 4 convolutional layers with kernel lengths of 3 to 7 and stride of 1. Use zero padding to ensure consistent output and avoid information loss. Configure 3 fully connected layers using the ReLU activation function to output key reaction force parameters, including the maximum value and distribution characteristics of the reaction forces of the side piers and the middle pier. Use mean squared error as the loss function and gradient descent for optimization. Monitor gradient explosion by controlling gradient clipping (threshold 1.0).

[0027] Step S202: Based on the basic load data, including only axle load, axle spacing, and off-center load location, at least 600 sets of samples are pre-trained using a Python script, splitting them into an 80% training set and a 20% validation set. The model is initially trained on the basic dataset to obtain general weights, providing a foundation for subsequent fine-tuning. The batch size is set to 32, and the training epochs are 100 to 200. The validation loss is evaluated every 10 epochs, and the intermediate model is saved for easy resume training. Fine-tuning of the intermediate model is then performed using data including temperature effects, concrete creep, and vehicle-bridge interaction effects. The pre-trained model is then refined based on a specific dataset. To optimize and improve the prediction accuracy for specific scenarios, the batch size is set to 16, and the training epochs are 50 to 100. The hyperparameters, including the convolution kernel size and pooling strategy, are validated using actual data from at least 20 bridge cases (including reaction force measurements of straight and curved bridges). Training stops when the loss function converges to below 0.005. The output includes reaction force parameters with dynamically amplified adjustment values, which are then uploaded to the cloud server. The data from step S100 is transformed into prediction output, ensuring a complete iteration of the process from model design to optimization. This provides an extended foundation for step S300 and enables engineering applications of real-time prediction.

[0028] In this embodiment, a learning rate decay strategy (e.g., decreasing from 0.01 to 0.001) is adopted during training to improve the convergence accuracy in the later stages. An early stopping mechanism is introduced, which stops training when the validation loss does not decrease for 10 consecutive rounds. During fine-tuning, customized optimization is performed for specific bridge types (e.g., curved bridges with a curvature radius of <50 meters). An automated pipeline is developed, and training data is updated regularly from the cloud database. Through phased training and fine-tuning, the model can adapt to different bridge types and working conditions, improving the generalization ability and engineering applicability of predictions.

[0029] Step S300: Input the key reaction force parameters into the linear regression model. Based on the bridge symmetry and the effects of temperature, concrete creep, and vehicle-bridge interaction, expand the key reaction force parameters using the synthetic overturning axis method. Generate the support concurrent reaction force combination by predicting the core reaction force parameters through the linear regression model. Expand the key reaction force parameters using symmetry and the regression model to generate multiple scenario concurrent combinations, so that the evaluation meets the requirements of the specifications and improves the intensification and robustness of the method. The expression for the linear regression model is as follows: ; in, y is the predicted value, representing the core reaction force parameter output by the model, used for subsequent expansion of key reaction force parameters; x is the input value, identifying the key reaction force parameter; w is the weight, representing the degree of influence of input x on y; b is the bias, and both are scalars; in linear regression prediction, the smaller the difference between the predicted value and the true value, the higher the accuracy of the model, and the error follows a normal distribution with a mean of 0. Figure 3This is a scatter plot of the predictions from the linear regression model in the example.

[0030] Among them, the synthetic overturning axis method is a technique based on bridge symmetry and multi-source influence to extend key reaction parameters. It is used to generate concurrent reaction force combinations at supports, making the evaluation meet the specifications and improving the method's efficiency and robustness. The synthetic overturning axis method is an algorithm for handling the extension of reaction forces in curved bridges. Based on the symmetrical distribution of double supports at the side piers and the axial symmetry of a single support at the central pier, it considers the effects of temperature gradient, concrete creep, and vehicle-bridge interaction. It simulates the reaction force distribution under different scenarios by synthesizing a virtual overturning axis, setting the error control range to -38.8% to 33.9%, to extend a single key reaction parameter into concurrent combinations for multiple scenarios. When using this method, input the key reaction force parameters output from step S200 into the linear regression model. Use a Python script to divide the model into symmetrical units (e.g., divide the abutment into left and right symmetrical parts), analyze the effects of temperature gradient and concrete creep, and use a temperature of 5°C per meter in both positive and negative directions. Collect data every 4 hours. The concrete creep time is 28 days to 10 years. Filter out abnormal expansions through pre-support, run at least 300 iterations, optimize a limited number of locations in each iteration, and generate expanded key reaction force data, covering at least 30 combinations of concurrent support reactions.

[0031] When expanding key reaction parameters, symmetry and regression models are used to generate multiple scenario combinations from a single input, ensuring comprehensive coverage of specification requirements (such as JTG 3362-2018). Using the key reaction parameters output by the convolutional neural network in step S200 (including the maximum values ​​and distribution characteristics of the reactions at the side piers and central piers) as input, and combining bridge geometric parameters, a Python script is used to divide the bridge into symmetrical units based on its symmetry. For example, side piers are divided into left-right symmetrical parts, and central piers are based on axial symmetry. The effects of temperature gradient and concrete creep are analyzed. For curved bridges, a virtual overturning axis is synthesized to simulate beam rotation and reaction redistribution. An error control range of -38.8% to 33.9% is set. Abnormal expanded values ​​are filtered by pre-support, and at least 300 iterations are run. Each iteration optimizes a limited number of locations, generating expanded key reaction data. For example, the maximum value of the reaction force at the dual supports of the side piers is expanded into multiple combinations, covering at least 30 concurrent support reaction force combinations, and stored as an index table in a MySQL database, including timestamp and expansion type fields for easy retrieval. The expanded data is provided to the linear regression model to further predict core reaction parameters.

[0032] The core reaction parameters predicted by the linear regression model mainly include the support reaction force values ​​under failure conditions, such as the reaction force value of the support on the i-th pier under dead load when it is in a failure state. The reaction force value when the support on the i-th pier is in a failed state under live load. and the support reaction force under normal working conditions This system is used to calculate the overturning stability coefficient. It utilizes key reaction force parameters derived from the full-bridge finite element model simulation and data output from the convolutional neural network in step S200. The input is a feature vector containing a dynamic amplification factor of the concrete creep coefficient and the vehicle-bridge interaction effect, based on the expanded key reaction force data. At least 900 training samples and 300 test samples are used, conforming to the JTG 3362-2018 standard. The predicted core reaction force parameters generate concurrent support reaction force combinations, stored in CSV format, supporting subsequent steps in S400 for calculating the overturning stability coefficient and risk classification. These combinations are uploaded to a cloud database for real-time access by the health monitoring system. The core reaction force parameters expand the key reaction force parameters, generating multi-scenario concurrent combinations to predict overturning risk, supporting four-stage assessment and generating four risk levels.

[0033] Step S301: Input the key reaction parameters output in step S200 into the linear regression model. Based on the symmetrical distribution of double supports on the side piers and the axial symmetry of the single support on the middle pier, according to the "Design Specification for Highway Reinforced Concrete and Prestressed Concrete Bridges and Culverts" (JTG3362-2018), use Python scripts to divide symmetrical units. For example, divide the side piers into left and right symmetrical parts. Analyze the effects of temperature gradient and concrete creep. The temperature is set at 5℃ per meter in both positive and negative directions. Data is collected every 4 hours. The concrete creep time is 28 days to 10 years. Use the synthetic overturning axis method to handle the reaction force expansion of the curved bridge. Set the error control range to -38.8% to 33.9%. Filter abnormal expansion through pre-support. Run at least 300 iterations, optimizing a limited number of locations in each iteration.

[0034] The system generates expanded key reaction force data, covering at least 30 concurrent support reaction force combinations. It also expands the storage of key reaction force data to generate an index table using a MySQL database, which includes timestamp and extended type fields for easy retrieval and updating.

[0035] Step S302: Extract the key reaction force parameters simulated in the full-bridge finite element model and the core reaction force parameters predicted by the linear regression model. Set the input as a feature vector containing the concrete creep coefficient and the vehicle-bridge interactive dynamic amplification factor. Train the linear regression model, allocate at least 900 training samples and 300 test samples to adjust the weights and biases, so as to minimize the difference between the predicted value and the true value. Generate the support concurrent reaction force combination, which meets the requirements of JTG 3362-2018 specification, and store it in CSV format to support subsequent step S400. Extend the prediction of step S200 to standard data, ensure the efficient conversion of the process from a single input to multiple combination outputs, and achieve comprehensive evaluation coverage and data optimization.

[0036] Step S400: Combining bridge geometric parameters, which are loaded from the local server, such as bearing spacing and beam self-weight distribution, according to Article 3.6.8 of the "General Specifications for Design of Highway Bridges and Culverts" (JTG D60-2015), the setting of transverse and vertical bearings of bridges should consider the issue of bearing detachment. The bearings should maintain horizontal upper and lower force transmission surfaces. Check the risk of detachment under temperature and concrete creep, calculate the overturning stability coefficient, incorporate dynamic amplification and concrete creep effects, evaluate the overturning process in four stages and classify it into four risk levels, integrate the convolutional neural network into the bridge health monitoring system, collect sensor data in real time and output dynamic early warning and practical reinforcement suggestions, integrate the aforementioned data for final calculation and output, provide risk classification and prevention guidance, realize the closed loop of the assessment process and practical application, and ensure the initiative and sustainability of bridge management.

[0037] According to Article 4.1.8 and its explanation in the "Design Specification for Highway Reinforced Concrete and Prestressed Concrete Bridges and Culverts" (JTG 3362-2018), under sustained conditions, the structural system of a beam bridge should not change, and the following requirements should be met simultaneously: ① Under the basic combination of actions, the individual compression supports should always remain under compression; ② When combined according to the standard values ​​of actions, the action effects of simply supported beams and continuous beams with integral sections should meet the following requirements: ; ; ; In the formula: For the lateral overturning stability coefficient, take... ; This is the design value for the effect of superstructure stability; The design value for the effect of superstructure instability; This refers to the reaction force value when the support on the i-th pier is in a failure state under constant load; This refers to the reaction force value when the support on the i-th pier is in a failed state under live load; This refers to the center-to-center distance between the supports on the i-th pier when they are in normal working condition and when they are in failure condition. Based on the overturning coefficient threshold, such as k≥2.5 for safety, bridge stability is classified into four levels: safe, critical, dangerous, and requiring immediate intervention.

[0038] Step S401: Input the combined reaction forces of the supports, and evaluate the overturning process by combining the dynamic amplification factor and concrete creep effect. The four-stage evaluation of the overturning process includes: a stable stage with all supports under compression (all supports are working normally and the reaction force is uniform); a transitional stage with the first support failing (the first support detaches and the reaction force is redistributed); a risk stage with three supports (the remaining three supports bear the load and the stability decreases); and an overturning stage with two supports (only two supports remain and the structure loses stability). Calculate the overturning stability coefficient according to Article 3.6.8 of JTG D60-2015 and JTG... Section 4.1.8 of 3362-2018 examines the risk of voiding under temperature and concrete creep (simulating beam rotation angle threshold of 0.02 radians). Based on the results of the four-stage assessment of the overturning process, four risk levels are classified: safe (stability coefficient ≥ 2.5, normal operation), critical (coefficient 1.5 to 2.5, monitoring required), dangerous (coefficient 1.0 to 1.5, immediate repair), and immediate intervention (coefficient < 1.0, cessation of passage). An assessment report in PDF format is generated and updated daily, and the manager is notified via email.

[0039] Step S402: Integrate the trained convolutional neural network model into the bridge health monitoring system. Configure sensors to collect vehicle load, temperature, and vibration data in real time, such as 15 sensors, including 8 temperature sensors and 7 vibration sensors, to collect vehicle load, temperature (range -30℃ to 60℃), and vibration data in real time, with a collection frequency of once every 3 seconds, supporting wireless transmission; set a beam rotation angle alarm with a threshold of 0.02 radians, which is displayed in real time through the monitoring system's web interface. Generate reinforcement suggestions based on the risk level, including adding steel supports to form steel cap beams and connecting them with ultra-high performance concrete post-cast strips, pier-beam consolidation, widening pier columns, or setting auxiliary pull-out constraints, prioritizing the risk of concrete creep in curved bridges, considering cases with a curvature radius <50 meters, simulating the reinforcement effect through a digital twin platform, running at least 100 simulations (each lasting 10 minutes) to verify the force transmission path and bending moment resistance performance, generating a simulation report including 3D visualization, and storing it in a cloud database. Transform the data from the aforementioned steps into actionable outputs, ensuring a complete closed loop from input to prevention, and achieving dynamic monitoring and optimized management of bridge risks.

[0040] According to the above technical solution, when applied to single-pier highway bridges, it faces challenges such as frequent heavy vehicle traffic (an average of 6,000 vehicles per day, with a maximum axle load of 50 tons) and high temperature and humidity (-10℃ to 50℃). To address this challenge, the team first used LiDAR and a total station to measure the bridge's geometric parameters (pier height 9 meters, support spacing 2.8 meters). A full-bridge finite element model was then established using Midas Civil to simulate the effects of heavy vehicles and temperature gradients, generating support reaction force data. The peak reaction force at the side piers was 1600 kN, with a dynamic amplification factor of 1.35. After processing the data using Python, a convolutional neural network was trained to predict the reaction force, with an error controlled within 1%, replacing traditional finite element iteration and significantly saving computation time. Based on the bridge's symmetry, 40 reaction force combinations were extended, and a linear regression model predicted concurrent reaction forces, conforming to the JTG 3362-2018 standard. Fifteen sensors integrated into the health monitoring system collected data in real time, discovering that the beam rotation angle reached 0.019 radians during a typhoon, with a stability coefficient of 2.6 (safety level). The system recommended pier-beam consolidation reinforcement. Digital twin simulation verified that the coefficient improved to 3.0 after reinforcement. Based on this, the management implemented reinforcement and optimized traffic control, successfully avoiding the risk of derailment and extending the bridge's lifespan by more than 5 years. This demonstrates that the technical solution significantly improves the efficiency and safety of bridge risk management under complex conditions through accurate modeling, rapid prediction, and real-time monitoring.

[0041] When applied to a single-column pier bridge with a curvature radius of 40 meters in mountainous areas and a span of 45 meters, this technology addresses the impact of heavy-duty mining trucks (axle load 45 tons) and long-term concrete creep (10 years). The team used UAV laser scanning to acquire geometric parameters: pier height 7.5 meters, support spacing 2.5 meters. A finite element model was built in Ansys to simulate vehicle eccentric loading and temperature variations from -20℃ to 45℃, generating a reaction force dataset. The reaction force at the middle pier was 850 kN, with a dynamic amplification factor of 1.28. After normalization using Python, key reaction force parameters were extracted, and a convolutional neural network was trained, achieving a reaction force prediction error of 0.8% in just 2 seconds. Based on the symmetry of the curved bridge, 35 reaction force combinations were generated. A linear regression model was used to optimize the concurrent reaction forces, achieving an error of 3.5%, meeting the JTG (Junior High-Temperature Tolerance) standard. The D60-2015 standard; the health monitoring system is equipped with 12 sensors (6 temperature, 6 vibration). Real-time monitoring detected a beam rotation angle of 0.021 radians when a heavy-load convoy passed by, with a stability coefficient of 1.8 (critical level), triggering an alarm. The system recommended widening the piers and adding pull-out restraints. Digital twin simulation showed that the coefficient increased to 2.7 after reinforcement. The management implemented reinforcement and restricted heavy-load traffic at night. A retest three months later confirmed that the risk was eliminated. The system demonstrates adaptability to complex stress scenarios of curved bridges, effectively preventing overturning risks through efficient prediction and dynamic management, ensuring the safe operation of bridges in mountainous areas.

[0042] In summary, by using finite element modeling and convolutional neural networks, bridge reaction forces can be predicted efficiently, saving computation time. The expansion of reaction force parameters and linear regression ensure the comprehensiveness of the assessment and compliance with standards. Real-time monitoring and digital twin technology enable dynamic risk warning and optimization of reinforcement schemes. The solution is applicable to different bridge types and complex working conditions, demonstrating its robustness and engineering practicality.

[0043] This embodiment proposes an intensive method for assessing the overload overturning risk of single-column piers based on convolutional neural networks. By establishing a full-bridge finite element model, designing a convolutional neural network, expanding reaction force parameters, and integrating a health monitoring system, this method provides a complete closed-loop application from data acquisition to risk assessment for single-column pier bridges with a span of 50 meters on highways and elevated bridges on urban roads. It generates an anti-overturning stability coefficient, outputs a beam rotation angle of 0.015 radians in real time, and provides reinforcement suggestions. These suggestions are updated daily for managers through PDF reports and a web interface, ensuring the safe operation of bridges under heavy vehicle loads, extreme temperatures, and long-term concrete creep. This significantly improves management efficiency and structural safety. The entire process can be deployed on a cloud platform (such as AWS or Alibaba Cloud) to integrate data storage, model training, and real-time monitoring. An AI-assisted decision-making system is introduced to automatically generate maintenance plans. Combined with IoT technology, the method extends to the collaborative management of regional bridge groups. Regular expert reviews are organized to optimize the model and reinforcement schemes. This technical solution, through the combination of physical modeling, machine learning, and real-time monitoring, achieves accurate prediction and dynamic management of bridge overturning risk. It is applicable to various bridge types and reduces operational risks.

[0044] To address the overturning risk of single-column pier bridges under overload, environmental changes, and dynamic interactions, this project provides a complete chain from data simulation to real-time early warning. First, data is generated through finite element simulation. Then, a convolutional neural network is built and trained for prediction. Next, the reaction force data is expanded and optimized, ultimately achieving risk assessment and prevention output. Based on physical simulation, machine learning and standardized calculations are integrated to combine offline data preparation with online real-time application. An iterative optimization mechanism is employed, processing the dataset and training the model on a local server before deployment to a bridge health monitoring system, supporting cloud data synchronization to ensure the accuracy, real-time performance, and scalability of the assessment. Vehicle load parameters, temperature effects, concrete creep, and vehicle-bridge interaction dynamic amplification factors are integrated into a multi-dimensional tensor input. Phased training and gradient descent optimization of the convolutional neural network achieve second-level reaction force prediction and dynamic amplification adjustment value output. The reaction force data is expanded using a synthetic overturning axis method, with errors controlled within a reasonable range. After integration with existing bridge health monitoring systems, sensor data can be collected in real time. This system outputs four risk levels: safe, critical, dangerous, and immediate intervention. It effectively avoids the limitations of post-event diagnosis in existing systems, significantly reduces the probability of overturning accidents, thereby reducing bridge maintenance costs and avoiding resource waste. It also improves structural stability in high-traffic environments and is applicable to various bridge types, including straight and curved bridges. It achieves a centralized assessment of overload overturning risks for single-column pier bridges, greatly improving the real-time nature and accuracy of the assessment. Based on highway bridge and culvert design specifications and bridge geometric parameters, it checks the risk of bearing dislodgement, calculates the overturning stability coefficient, incorporates dynamic amplification and concrete creep effects, and assesses the overturning process in four stages, classifying it into four risk levels. Based on the risk level, it automatically outputs reinforcement recommendations and simulates the reinforcement effect through a digital twin platform. This provides a comprehensive prevention strategy for overload overturning risk management of single-column pier bridges, achieving closed-loop management from assessment to action. It improves the accuracy of bridge remaining life prediction, reduces human intervention errors, promotes a shift in bridge maintenance from passive repair to proactive prevention, and enhances the long-term safety and sustainability of bridge operations. Specific Implementation Example 2: like Figures 1 to 4 As shown, based on the content of the above specific embodiments, the following content is further disclosed: The hardware of the intensive single-column pier overload overturning risk assessment method includes a processor and a machine-readable storage medium. The machine-readable storage medium and the processor are connected. The machine-readable storage medium is used to store programs, instructions or code. The processor is used to execute the programs, instructions or code in the machine-readable storage medium to implement the above method.

[0046] To achieve the above technical solution, at least the following hardware structure is required in practical use: The hardware architecture adopts a layered structure of "sensor-edge device-cloud server-user terminal" to ensure real-time data flow from on-site bridge data acquisition to cloud analysis. It supports offline training and online deployment, features high reliability, low latency, and high computing power, and is suitable for urban bridge environments, including: Bridge-end sensor system (data acquisition layer): Responsible for real-time acquisition of bridge load and environmental data, supporting analog input and real-time monitoring in steps S100 and S400. Hardware includes: Sensor array: Deploy at least 15 sensors, including 8 temperature sensors, such as PT100 platinum resistance thermometers, with a measurement range of -50℃ to 100℃ and an accuracy of ±0.1℃; 7 vibration sensors, such as IEPE accelerometers, with a frequency response of 10Hz to 10kHz and a sensitivity of 100mV / g; and vehicle load sensors and strain gauge load cells, with a range of 10 tons to 50 tons and an accuracy of ±0.5%. These sensors are installed at key locations on piers, supports, and beams, and support wireless transmission based on LoRa or NB-IoT protocols. The data acquisition frequency is once every 3 seconds to ensure the capture of dynamic changes such as temperature gradients of 5℃ per meter and concrete creep effects. Data Acquisition Unit: Uses an embedded microcontroller, such as the STM32 series, with an ARM Cortex-M core and 512KB of memory. It integrates an ADC conversion module to perform preliminary filtering of noisy data (threshold filtering of outliers) and summarizes the data through an RS485 interface. This layer outputs a raw data stream in CSV format and transmits it to the edge device. It supports load simulation input in step S100 and real-time acquisition in step S400, ensuring data accuracy and anti-interference capability, and has an IP67 protection rating.

[0047] Edge computing devices (local processing layer): Located at or near the bridge site, responsible for initial data processing and model inference, supporting rapid prediction in steps S200 and S300, reducing cloud load, and achieving low-latency response. Hardware includes: Edge server: Employs an industrial-grade embedded computer, such as the NVIDIA Jetson Xavier NX, equipped with a 6-core ARM CPU and a 384-core Volta GPU, 16GB LPDDR4 memory, and 256GB NVMe SSD storage. It runs a simplified version of the convolutional neural network, processes multi-dimensional tensor inputs, performs the inference part of staged training and linear regression prediction, supports GPIO interface connection to sensors, and integrates Wi-Fi / 5G module to transmit data to the cloud, ensuring local data storage in the event of network interruption. Power supply and protection system: Equipped with a UPS uninterruptible power supply (500Wh capacity, supports solar supplementation) and a protective enclosure (IP65 rating, dustproof and waterproof), operating temperature -20℃ to 60℃. The system processes hierarchical normalization in step S100 (using the NumPy library) and reaction force expansion in step S300 (synthetic overturning axis method, 300 iterations), outputting preprocessed key reaction force parameters (JSON format) to provide efficient input to the cloud and enable edge intelligent computing.

[0048] Cloud servers (computing and storage layer): responsible for high-intensity computing tasks, such as finite element simulation, full model training, and data expansion, supporting offline processing and global optimization of steps S100, S200, and S300. Hardware includes: Cloud computing nodes: Use high-performance server clusters, such as AWS EC2 or Alibaba Cloud ECS instances, configured with 64 Intel Xeon Platinum CPUs, 4 NVIDIA A100 GPUs, 512GB of memory, and 10TB of SSD storage. The servers run finite element software (Midas Civil) to perform 200 vehicle-bridge interaction simulations (axle load 10-50 tons, temperature -30℃ to 60℃), and perform CNN training, with 150 pre-training rounds and 80 fine-tuning rounds, using the Adam optimizer, and integrating a database management system (such as MySQL) to store support reaction force data and index tables, supporting big data queries, with at least 1000 sets of samples; Cloud storage and backup: Utilizes distributed storage, such as Amazon S3 compatible object storage, with unlimited capacity expansion. Backs up CSV and JSON data with daily automatic synchronization. The server supports containerized deployment, running Python scripts to handle concrete creep model updates and linear regression training (900 training samples), ensuring error control in step S300. It receives edge data via API and outputs extended reaction force combinations to achieve computationally intensive tasks.

[0049] User terminal (output and interaction layer): Responsible for risk visualization and early warning output, supporting the display of assessment reports and hardening recommendations from step S400, and enabling human-computer interaction. Hardware includes: Monitoring terminal: Use a laptop or tablet (such as Dell XPS series, Intel i7 CPU, 16GB memory, 512GB SSD) or mobile APP device (Android / iOS smartphone with high-resolution screen). The terminal runs a web interface (based on HTML5 and JavaScript) to display beam rotation angle alarms and four-level risk levels in real time.

[0050] Digital Twin Platform: Integrates a dedicated workstation, such as a desktop PC equipped with an RTX 3080 GPU, to run Bentley Systems software, conduct 100 reinforcement simulations, verify steel cap beam connections, pier-beam consolidation, etc., generate PDF reports and 3D visualizations, receive cloud data via 5G or wired network, support daily updates of assessment reports, and ensure users receive reinforcement recommendations in a timely manner.

[0051] The hardware architecture achieves end-to-end operation of the solution through a process of sensor acquisition, edge preprocessing, cloud computing, and terminal output. It supports real-time monitoring and preventive maintenance of bridges, with total power consumption controlled within 500W and high system availability.

[0052] The aforementioned hardware structure and its corresponding hierarchical structure are merely one way to implement the intensive single-column pier overload overturning risk assessment method of this application, including but not limited to the aforementioned hardware structure and its corresponding parameters. In actual use, the hardware structure can be further configured according to specific scenarios and usage requirements.

[0053] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing the risk of overloaded overturning of a single-column pier based on a convolutional neural network, characterized in that, The method for assessing the risk of overloaded overturning of the intensive single-column pier includes: A full-bridge finite element model is established based on the bridge's geometric parameters. Data on vehicle load parameters, temperature effects, concrete creep, and vehicle-bridge interaction are simulated. Support reaction force data are calculated and subjected to hierarchical normalization to extract key reaction force parameters. Based on the full-bridge finite element model, a convolutional neural network is designed. The multidimensional tensor is used as input data, the output is the key reaction force parameter, and the mean square error is used as the loss function. The network is optimized to convergence through staged training and gradient descent, and the output is a dynamically amplified and adjusted value. The key reaction parameters are input into the linear regression model. Based on the bridge symmetry and the effects of temperature, concrete creep, and vehicle-bridge interaction, the key reaction parameters are expanded using the synthetic overturning axis method. The core reaction parameters are predicted through the linear regression model, and the combination of concurrent reaction forces at the supports is generated. By combining bridge geometric parameters, the risk of voiding under temperature and concrete creep is examined, the overturning stability coefficient is calculated, dynamic amplification and concrete creep effect are incorporated, the overturning process is evaluated in four stages and classified into four risk levels, and a convolutional neural network is integrated into the bridge health monitoring system to collect sensor data in real time and output dynamic early warning and reinforcement suggestions.

2. The method for assessing the risk of overloaded overturning of intensive single-column piers based on convolutional neural networks according to claim 1, characterized in that, The establishment of the full-bridge finite element model based on the bridge's geometric parameters includes: Measure and collect bridge geometry parameters, including bearing spacing accurate to centimeters, beam weight distribution per meter, pier height, and material properties; The bridge's geometric parameters were imported using finite element software, and vehicle load conditions were set. The axle load range was configured to be 10 to 50 tons, the wheelbase to be 2 to 6 meters, and the lateral eccentric load position variation range to be -1.5 meters to 1.5 meters. Temperature effect simulation was also performed, with uniform temperature variation set from -30℃ to 60℃. Concrete creep is calculated by updating the long-term self-weight deformation model periodically. Through vehicle-bridge interactive analysis, vehicle speeds are set from 20 to 80 km / h, and road surface roughness levels are set from A to C. At least 200 simulations are run to calculate and export the dynamic amplification factor, generate support reaction force data, and store it on the local server.

3. The method for assessing the overload overturning risk of intensive single-column piers based on convolutional neural networks according to claim 2, characterized in that, The calculated support reaction force data is subjected to hierarchical normalization processing to extract key reaction force parameters, including: The support reaction force data were grouped by temperature range, namely -30℃ to 0℃, 0℃ to 30℃, and 30℃ to 60℃, and by concrete creep stage, namely short-term 28 days and long-term 1 to 10 years, with at least 300 samples in each group. Each set of data is normalized to map the values ​​to the interval between 0 and 1, and key reaction parameters are extracted. These key reaction parameters include the maximum value of the double support reaction force of the side pier, the standard deviation of the distribution of the single support reaction force of the middle pier, the beam rotation angle, and the long-term coefficient of variation. A dataset covering straight bridges, curved bridges, and at least 20 bridge types is generated, containing at least 1,000 sets of samples.

4. The method for assessing the overload overturning risk of intensive single-column piers based on convolutional neural networks according to claim 1, characterized in that, The convolutional neural network designed based on the full-bridge finite element model takes multidimensional tensors as input data, outputs key reaction force parameters, and uses mean square error as the loss function, including: Construct a convolutional neural network, setting the input layer to a multidimensional tensor containing the axial weight distribution vector, load position vector, temperature gradient distribution, concrete creep coefficient, and vehicle-bridge interaction dynamic amplification factor, configuring the convolutional kernel length to be 3 to 7, the stride to be 1, and the padding mode to be zero padding. Three fully connected layers are configured to output key reaction force parameters, including the maximum values ​​and distribution characteristics of the reaction forces of the side piers and the middle piers. The mean square error is used as the loss function, and the gradient descent method is used for optimization.

5. The method for assessing the overload overturning risk of intensive single-column piers based on convolutional neural networks according to claim 4, characterized in that, The process of optimizing to convergence through phased training and gradient descent, and outputting dynamically amplified adjustment values, includes: Based on basic load data, including only axle load, axle spacing, and off-center load location, at least 600 sets of samples were divided into 80% training set and 20% validation set for pre-training, with 100 to 200 training rounds, and intermediate models were saved. The intermediate model is fine-tuned using data that includes temperature effects, concrete creep, and vehicle-bridge interaction effects. The training epochs are 50 to 100 times. The hyperparameters, including convolution kernel size and pooling strategy, are validated using real data from at least 20 bridge cases. The output includes reaction force parameters with dynamically amplified adjustment values ​​and is uploaded to a cloud server.

6. The method for assessing the overload overturning risk of intensive single-column piers based on convolutional neural networks according to claim 1, characterized in that, In the linear regression model inputting the key reaction parameters, based on bridge symmetry and the effects of temperature, concrete creep, and vehicle-bridge interaction, the key reaction parameters are expanded using a synthetic overturning axis method, including: The extracted key reaction force parameters are input into the linear regression model. Based on the symmetrical distribution of the double supports of the side piers and the axial symmetry of the single support of the middle pier, symmetrical elements are divided. The effects of temperature gradient and concrete creep are analyzed. The synthetic overturning axis method is used to process the reaction force expansion of the curved bridge. The error control range is set to -38.8% to 33.9%, and at least 300 iterations are run. Generate expanded key reaction force data, covering at least 30 combinations of concurrent support reaction forces, and expand the storage of key reaction force data to generate an index table.

7. The method for assessing the overload overturning risk of intensive single-column piers based on convolutional neural networks according to claim 1, characterized in that, The method of generating support concurrent reaction force combinations by predicting core reaction force parameters through a linear regression model includes: Extract the key reaction force parameters simulated in the full bridge finite element model and the core reaction force parameters predicted by the linear regression model, and set the input as a feature vector containing the concrete creep coefficient and the vehicle-bridge interactive dynamic amplification factor. Train a linear regression model, allocate at least 900 training samples and 300 test samples to adjust weights and biases, minimize the difference between predicted and true values, and generate a combination of support and reaction forces.

8. The method for assessing the risk of overloaded overturning of intensive single-column piers based on convolutional neural networks according to claim 1, characterized in that, The process involves combining bridge geometric parameters, examining the risk of voiding under temperature and concrete creep, calculating the overturning stability coefficient, and incorporating dynamic amplification and concrete creep effects, including: The input support reaction force combination is combined with the dynamic amplification factor and concrete creep effect to evaluate the overturning process. The four-stage evaluation of the overturning process includes a stable stage with all supports under pressure, a transition stage with the first support failure, a risk stage with three supports, and an overturning stage with two supports. Based on the results of the four-stage evaluation of the overturning process, four risk levels are divided, namely safe, critical, dangerous, and immediate intervention. An evaluation report is generated and updated daily.

9. The method for assessing the overload overturning risk of intensive single-column piers based on convolutional neural networks according to claim 8, characterized in that, The integration of convolutional neural networks into the bridge health monitoring system, which collects sensor data in real time and outputs dynamic early warnings and reinforcement suggestions, includes: The trained convolutional neural network model is integrated into the bridge health monitoring system. Sensors are configured to collect vehicle load, temperature and vibration data in real time. Beam rotation angle alarms are set and displayed in real time through the monitoring system's web interface. Based on the risk level, reinforcement recommendations are generated, including adding steel supports to form steel cap beams and connecting them with post-cast concrete strips, pier-beam consolidation, widening pier columns, or setting auxiliary pull-out restraints. Priority is given to the risk of concrete creep in curved bridges. The reinforcement effect is simulated through a digital twin platform to verify the force transmission path and bending moment resistance.

10. The method for assessing the overload overturning risk of intensive single-column piers based on convolutional neural networks according to claim 1, characterized in that, The hardware of the intensive single-column pier overload overturning risk assessment method includes a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code. The processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above method.

Images