Iron tower structure hidden danger detection method, system and device and medium

Abstract

The invention relates to the technical field of iron tower monitoring, in particular to an iron tower structure hidden danger detection method, system and device and a medium. Predicting a dynamic baseline of a health state according to the environmental data by using a generative adversarial network so as to calculate and monitor a residual signal without environmental influence, and when a residual continuously exceeds a threshold value, starting a physical information neural network to invert, solve and output a variable quantity of a structure physical parameter according to the residual signal; therefore, the hidden danger of the iron tower structure can be accurately detected. According to the method, the environmental effect is accurately modeled through the GAN model, the signal-to-noise ratio of the residual signal is improved, the system can distinguish real damage and environmental fluctuation with extremely high confidence, the structural safety of the iron tower is improved, and the maintenance cost of the whole life cycle is optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of iron tower monitoring. BACKGROUND

[0002] Vibration-based structural health monitoring aims to assess the safety and integrity of the structure by analyzing the changes in its dynamic response. The core theoretical basis is that the physical properties of the structure directly determine the modal parameters. When the structure has damage such as loose bolts, component micro-cracks, and material fatigue, it will inevitably lead to a decrease in local or overall stiffness. This change will directly reflect the change in modal parameters, especially the decrease in natural frequency. Therefore, by continuously monitoring the modal parameters of the structure over a long period of time and comparing them with the baseline values in the healthy state, early warning of structural damage can be achieved in theory.

[0003] The existing technology mainly uses data normalization or compensation models, such as multiple linear regression-based models and principal component analysis statistical methods, to try to establish a linear relationship between environmental factors and modal frequencies, and then "filter out" the environmental impact from the measured data. However, these methods have the following limitations:

[0004] The assumption premise is unrealistic: existing methods generally assume a linear, non-time-varying mapping relationship between environmental factors and modal frequencies. However, the response of the structure to environmental factors is actually a highly nonlinear, hysteresis effect, and multi-factor coupled complex process. This assumption cannot match the real working conditions.

[0005] Conceptual cognition has bias: existing methods treat environmental impact as statistical noise, but statistical noise is usually random and unrelated to the target signal. However, environmental effects are deterministic physical processes, and the signals they produce are highly similar to damage signals and difficult to distinguish. Environment factors are essentially strong confounding variables, not noise.

[0006] Separation effect is difficult to guarantee: traditional signal filtering or statistical noise reduction techniques cannot effectively separate the confounding effects produced by deterministic physical processes at the mathematical level.

[0007] In summary, existing technologies have not provided a robust and accurate technical solution to decouple environmental effects from real structural degradation effects. Without this core capability, reliable, efficient, and low false alarm rate automated structural health monitoring systems remain unsolved technical problems, and the field urgently needs a new technical paradigm to fundamentally solve the environmental interference problem and achieve accurate identification of structural hazards. SUMMARY

[0008] In order to overcome the problems of existing technologies that the assumptions do not match the real working conditions, the environmental factors are not correctly recognized, and the separation effect is difficult to guarantee, the present application provides an iron tower structure hazard detection method, system, device, and medium.

[0009] The technical scheme adopted by the present application to achieve the above-mentioned purpose is: a tower structure hidden danger detection method, comprising the following steps:

[0010] S1, deploying sensors on the tower structure, collecting multi-dimensional original data in real time, including structure dynamic response data, structure temperature field data and three-dimensional wind load data;

[0011] S2, based on the structure dynamic response data, obtaining real-time modal parameters through online modal identification;

[0012] S3, establishing a first deep learning model, the first deep learning model being a time series generative adversarial network (GAN), inputting the structure temperature field data and the three-dimensional wind load data into the trained first deep learning model to generate dynamic baseline modal parameters representing the health state of the tower structure under the current environmental conditions;

[0013] S4, performing point-by-point subtraction operation on the real-time modal parameters and the dynamic baseline modal parameters according to time points to obtain residual signals;

[0014] S5, establishing a second deep learning model, the second deep learning model being a physical information neural network (PINN), real-time monitoring the residual signals, when the residual signals continuously exceed the preset threshold, inputting the abnormal residual signals and the real-time modal parameters in the corresponding time period into the second deep learning model, and outputting the change amount of multiple physical parameters of the tower structure to represent the tower structure hidden danger.

[0015] Preferably, in step S1, the sensors include a three-axis acceleration sensor, a temperature sensor and a three-dimensional wind speed and direction instrument; the three-axis acceleration sensor is deployed at a key structure node of the tower to collect the structure dynamic response data; the temperature sensor is deployed at different heights and different directions of the tower to collect the structure temperature field data; and the three-dimensional wind speed and direction instrument is deployed at the top of the tower to collect the three-dimensional wind load data.

[0016] Preferably, in step 2, a fixed-length sliding time window mechanism is used to continuously segment the structure dynamic response data, and a stochastic subspace identification (SSI) algorithm is used for online modal identification in each time window to generate continuous real-time modal parameters.

[0017] Preferably, in step S3, the first deep learning model includes a generator and a discriminator, the generator adopts a structure containing a recurrent neural network unit, and the discriminator is used to distinguish the modal parameters generated by the generator from the real modal parameters under the historical health state of the tower structure; and the model is adversarially trained using a historical data set collected under the historical health state of the tower structure.

[0018] Preferably, in step S4, the real-time modal parameters and the dynamic baseline modal parameters are point-by-point subtracted at time points:

[0019] ;

[0020] wherein, is a residual signal, is a real-time modal parameter, is a dynamic baseline modal parameter.

[0021] Preferably, in step S5, the training optimization of the physics-informed neural network (PINN) is driven by a compound loss function, and the formula of the compound loss function is:

[0022] ;

[0023] wherein, is a compound loss function, is a data loss term, is a physics loss term, and are weight coefficients.

[0024] Preferably, the data loss term is used to quantify the difference between the predicted modal parameters of the PINN and the abnormal real-time modal parameters, and the physics loss term is used to quantify the satisfaction of the predicted displacement field of the PINN to the structure dynamics control partial differential equation containing the trainable stiffness variable, and the stiffness matrix in the structure dynamics control partial differential equation is parameterized.

[0025] A tower structure hidden danger detection system, comprising:

[0026] a sensor for real-time acquisition of multi-dimensional original data of the tower structure, the sensor comprising a three-axis acceleration sensor, a temperature sensor, and a three-dimensional wind speed and direction instrument;

[0027] a modal identification module for obtaining real-time modal parameters through online modal identification based on structure dynamic response data;

[0028] a dynamic baseline generation module, which is internally provided with a trained first deep learning model, for generating dynamic baseline modal parameters;

[0029] a residual calculation module for calculating a residual signal between the real-time modal parameters and the dynamic baseline modal parameters;

[0030] a hidden danger determination module, which is internally provided with a trained second deep learning model, for monitoring the residual signal and outputting a change amount of the tower structure physical parameters.

[0031] A tower structure hidden danger detection device, comprising a memory and a processor, the memory is used to store a computer program, the computer program is used to execute the above-mentioned tower structure hidden danger detection method when loaded by the processor.

[0032] A readable storage medium, the storage medium stores a computer program, the computer program is used to execute the above-mentioned tower structure hidden danger detection method when loaded by the processor.

[0033] The beneficial effects of the present application are:

[0034] The present application accurately models the environmental effect through the GAN model, eliminates the main factors leading to false positives, greatly improves the signal-to-noise ratio of the residual signal, enables the system to distinguish between real damage and environmental fluctuations with very high confidence, and reduces the false positive rate to a very low level; The present application identifies deviations from the physical laws of healthy structures by learning these laws, without needing to access any "damage" samples during the training phase, solving the bottleneck problem of being difficult to obtain labeled data covering all damage modes and degrees in actual engineering, greatly enhancing the practicality and generalization ability of the present application; The high-confidence, quantitative damage assessment results provided by the present application enable asset managers to abandon expensive, fixed-cycle-based inspection modes or passive maintenance after accidents, and instead adopt a proactive maintenance strategy based on the true physical state of the structure, with a clear goal, which can accurately intervene in the early stages of hidden dangers, significantly improve structural safety and optimize life cycle maintenance costs. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is the overall flowchart of the embodiment of the present application. DETAILED DESCRIPTION

[0036] The embodiment of the present application provides a tower structure hidden danger detection method, as shown in Figure 1 The embodiment of the present application provides a tower structure hidden danger detection method, as shown in

[0037] S1, deploying a sensor suite on the tower structure, including a three-axis acceleration sensor, a temperature sensor and a three-dimensional wind speed and direction instrument, real-time collecting multi-dimensional original data and preprocessing to obtain synchronous heterogeneous data streams; Specifically:

[0038] Deploying a sensor suite at a key position of the tower structure, including a three-axis acceleration sensor, a temperature sensor and a three-dimensional wind speed and direction instrument, presetting a sampling frequency, and real-time collecting multi-dimensional original data:

[0039] Three-axis acceleration sensor: deployed at key structural nodes of the tower, such as the middle of the tower leg, the intersection of the main diagonal material, and the tower top, used to capture the micro-vibration of the structure under environmental excitation, and output multiple three-dimensional acceleration time series at a preset sampling frequency to form structural dynamic response data;

[0040] Temperature sensors: deployed at different heights of the tower and different orientations of the sunny and shady sides, used to accurately capture the structural temperature field and its gradient changes, and output the temperature time series of the corresponding measurement points, forming the structural temperature field data;

[0041] Three-dimensional wind speed and direction instrument: deployed at the top of the tower, used to measure horizontal wind speed, direction and vertical wind speed in real time, output time series data set containing horizontal wind speed, direction and vertical wind speed, forming three-dimensional wind load data;

[0042] All sensors are connected to the field data acquisition unit for synchronous sampling, A / D conversion and data packaging, and high-precision time stamps are attached. The above structural dynamic response data, structural temperature field data and three-dimensional wind load data are transmitted in real time to the central processing server through wireless or wired network, and preprocessed such as time synchronization alignment and outlier processing to obtain synchronous heterogeneous data stream.

[0043] S2, based on the structural dynamic response data in the synchronous heterogeneous data stream, real-time modal parameters are obtained through online modal identification; specifically:

[0044] Extract the structural dynamic response data in the synchronous heterogeneous data stream, use a fixed-length sliding time window mechanism to continuously segment the data, and form an independent analysis data block in each time window; process each data block using the random subspace identification algorithm in the time domain to identify a set of real-time modal parameters in the time window; by repeatedly executing the above identification process on the continuous sliding time window, continuous real-time modal parameters are generated, and the parameters are completely synchronized and corresponding with the structural temperature field data and three-dimensional wind load data in the synchronous heterogeneous data stream in the time dimension. Real-time modal parameters are dynamic characteristics indicators of the tower structure.

[0045] In this embodiment, the random subspace identification (SSI) algorithm in the time domain is preferably used. SSI algorithm is widely used in the field of structural health monitoring due to its high computational efficiency, strong robustness to noise and advantages of suitable for automation processing. This method processes acceleration data with a sliding data window, for example, 10 minutes in length and 5 minutes in sliding step. In each window, the SSI algorithm automatically identifies the main modal parameters of the structure in that time period, especially the first few natural frequencies, by constructing a data-driven state space model and solving it. By continuously processing the data of the sliding window, a real-time modal parameter that is time-synchronized with the environmental parameter data (temperature field, wind load) is generated.

[0046] S3, a first deep learning model is established, and the structural temperature field data, three-dimensional wind load data and real-time modal parameters in the synchronous heterogeneous data stream are used to generate dynamic baseline modal parameters, which are used to represent the health status of the tower structure under the current environmental conditions; specifically:

[0047] establishing a first deep learning model, which is a time series generative adversarial network (GAN) including a generator and a discriminator;

[0048] The generator adopts a structure including a recurrent neural network unit, and the input thereof includes historical modal parameters, environmental parameters and a random noise vector;

[0049] The discriminator is used to distinguish the modal parameters generated by the generator from the real modal parameters under the historical health state of the tower structure;

[0050] The generator and the discriminator are adversarially trained using a historical data set collected under the historical health state of the tower structure as training data, and a trained first deep learning model is obtained;

[0051] The temperature field data of the combined structure, the three-dimensional wind load data and the historical part of the real-time modal parameters are combined to form a multi-channel input time series;

[0052] The multi-channel input time series is input into the trained first deep learning model, and the output of the forward propagation calculation of the first deep learning model is a dynamic baseline modal parameter, which is used to represent the health state of the tower under the current environmental condition.

[0053] S4, based on the real-time modal parameters and the dynamic baseline modal parameters, a residual signal between the two is calculated; specifically:

[0054] The real-time modal parameters and the dynamic baseline modal parameters are subtracted point by point according to the time point:

[0055] ;

[0056] wherein, is the residual signal, is the real-time modal parameter, is the dynamic baseline modal parameter;

[0057] The residual signal with the environmental interference removed is obtained and used as an indication feature of the tower structure damage.

[0058] S5, establishing a second deep learning model to statistically monitor the residual signal, and outputting a tower structure physical parameter change amount through the second deep learning model when the residual signal is abnormal, which is used to represent the tower structure hidden danger; specifically:

[0059] establishing a second deep learning model, which is a physical information neural network (PINN);

[0060] statistical process control method is used to monitor the residual signal in real time: the statistical characteristics of the residual signal are evaluated in a preset length of sliding time window, and compared with the preset statistical threshold determined according to the historical health data; when the statistical characteristics of the residual signal continuously exceed the preset statistical threshold, it is determined to be abnormal, and the abnormal residual signal and the real-time modal parameters in the corresponding time period are input into the second deep learning model as target data.

[0061] A composite loss function is constructed to drive the training optimization of the PINN, which includes a data loss term quantifying the difference between the predicted modal parameters of the PINN and the abnormal real-time modal parameters, and a physical loss term quantifying the satisfaction of the structure dynamics control partial differential equation containing trainable stiffness variables by the predicted displacement field of the PINN; the stiffness matrix in the structure dynamics control partial differential equation is parameterized, for example, the structure is divided into multiple element groups, and the stiffness of each group is taken as a trainable variable, and the physical loss term is calculated by automatic differentiation technology.

[0062] The formula for calculating the composite loss function is:

[0063] ;

[0064] wherein, is the composite loss function, representing the total target of PINN training optimization; is the data loss term, is the physical loss term, and are weight coefficients.

[0065] By minimizing the composite loss function, the weight coefficients of the PINN and the trainable stiffness variables are optimized, and the goal of the optimization process is to find a set of parameters that can not only fit the observed abnormal modal parameters, but also satisfy the physical law. Compare the values of the optimized trainable stiffness variables with the health state stiffness values to obtain and output the change amount of multiple physical parameters of the tower structure, which is the characterization result of the tower structure hidden danger, indicating the location and severity of the damage.

[0066] The real-time modal parameters, dynamic baseline modal parameters and residual signals mentioned above are all time series.

[0067] Working principle: real-time modal parameters are obtained by real-time acquisition of tower vibration and environmental data and online identification, and the dynamic baseline of the health state is predicted according to the environmental data by using the generative adversarial network to calculate and monitor the residual signal excluding environmental impact. When the residual signal continuously exceeds the threshold, the physical information neural network is started to inverse solve and output the change amount of the structure physical parameters according to the residual signal, so as to realize the accurate detection of the tower structure hidden danger.

[0068] A tower structure hidden danger detection system, comprising:

[0069] a sensor for collecting multi-dimensional raw data of the tower structure in real time, the sensor comprising a three-axis acceleration sensor, a temperature sensor and a three-dimensional wind speed and direction meter;

[0070] a modal identification module for obtaining real-time modal parameters through online modal identification based on the structural dynamic response data;

[0071] a dynamic baseline generation module, which is internally provided with a first deep learning model trained to generate dynamic baseline modal parameters;

[0072] a residual calculation module for calculating residual signals between the real-time modal parameters and the dynamic baseline modal parameters;

[0073] a hidden danger determination module, which is internally provided with a second deep learning model trained to monitor the residual signals and output the change amount of the physical parameters of the tower structure.

[0074] A tower structure hidden danger detection device, comprising a memory and a processor, the memory being used to store a computer program, the computer program being used to execute the tower structure hidden danger detection method described above when loaded by the processor.

[0075] A readable storage medium, the storage medium storing a computer program, the computer program being used to execute the tower structure hidden danger detection method described above when loaded by the processor.

[0076] The present application is described by way of examples, and those skilled in the art know that various changes or equivalent replacements can be made to these features and examples without departing from the spirit and scope of the present application. In addition, under the guidance of the present application, these features and examples can be modified to adapt to specific conditions and materials without departing from the spirit and scope of the present application. Therefore, the present application is not limited by the specific examples disclosed herein, and all examples falling within the scope of the claims of the present application are within the protection scope of the present application.

Claims

1. A method for detecting potential structural hazards in iron towers, characterized in that, Includes the following steps: S1. Deploy sensors on the tower structure to collect multi-dimensional raw data in real time, including structural dynamic response data, structural temperature field data, and three-dimensional wind load data. S2. Based on structural dynamic response data, real-time modal parameters are obtained through online modal identification; S3. Establish the first deep learning model, which is a time series generative adversarial network (GAN). Input the structural temperature field data and three-dimensional wind load data into the trained first deep learning model to generate dynamic baseline modal parameters that characterize the health status of the tower structure under the current environmental conditions. S4. Subtract the real-time modal parameters and the dynamic baseline modal parameters point by point according to the time point to obtain the residual signal; S5. Establish a second deep learning model, which is a physical information neural network (PINN), to monitor the residual signal in real time. When the residual signal continuously exceeds the preset threshold, the abnormal residual signal and its corresponding real-time modal parameters within the time period are input into the second deep learning model, and the changes in multiple physical parameters of the tower structure are output to represent potential hazards in the tower structure.

2. The method for detecting hidden dangers in iron tower structures according to claim 1, characterized in that, In step S1, the sensors include a triaxial accelerometer, a temperature sensor, and a three-dimensional anemometer; the triaxial accelerometer is deployed at key structural nodes of the tower to collect dynamic response data of the structure; the temperature sensor is deployed at different heights and orientations of the tower to collect temperature field data of the structure; and the three-dimensional anemometer is deployed at the top of the tower to collect three-dimensional wind load data.

3. The method for detecting hidden dangers in iron tower structures according to claim 1, characterized in that, In step 2, a fixed-length sliding time window mechanism is used to continuously segment the structural dynamic response data. Within each time window, the random subspace identification (SSI) algorithm is used for online modal identification to generate continuous real-time modal parameters.

4. The method for detecting hidden dangers in iron tower structures according to claim 1, characterized in that, In step S3, the first deep learning model includes a generator and a discriminator. The generator adopts a structure containing recurrent neural network units, and the discriminator is used to distinguish the modal parameters generated by the generator from the real modal parameters under the historical health state of the tower structure. The model was trained adversarially using historical datasets collected from the historical health status of the tower structure.

5. The method for detecting hidden dangers in iron tower structures according to claim 1, characterized in that, In step S4, the real-time modal parameters and the dynamic baseline modal parameters are subtracted point by point according to time points: ； in, For residual signals, For real-time modal parameters, These are dynamic baseline modal parameters.

6. The method for detecting hidden dangers in iron tower structures according to claim 1, characterized in that, In step S5, the training optimization of the Physical Information Neural Network (PINN) is driven by a composite loss function, the formula of which is: ； in, For composite loss function, For data loss items, For physical loss items, and These are the weighting coefficients.

7. The method for detecting hidden dangers in iron tower structures according to claim 6, characterized in that, The data loss term is used to quantify the difference between the PINN predicted modal parameters and the abnormal real-time modal parameters, and the physical loss term is used to quantify the degree to which the PINN predicted displacement field satisfies the structural dynamics control partial differential equation containing trainable stiffness variables. The stiffness matrix in the structural dynamics control partial differential equation is parameterized.

8. A system for detecting potential structural hazards in iron towers, characterized in that, include: Sensors are used to collect multi-dimensional raw data of the tower structure in real time. The sensors include a triaxial accelerometer, a temperature sensor, and a three-dimensional anemometer. The modal identification module is used to obtain real-time modal parameters based on structural dynamic response data through online modal identification; The dynamic baseline generation module has a built-in first deep learning model that has been trained, which is used to generate dynamic baseline modal parameters. The residual calculation module is used to calculate the residual signal between the real-time modal parameters and the dynamic baseline modal parameters; The hazard assessment module has a built-in, trained second deep learning model, which is used to monitor the residual signal and output the changes in the physical parameters of the tower structure.

9. A device for detecting potential structural hazards in iron towers, characterized in that, include: The system includes a memory and a processor. The memory is used to store a computer program, which is used to execute the tower structure hazard detection method described in claims 1-7 when loaded by the processor.

10. A readable storage medium, characterized in that, The storage medium contains a computer program that is adapted to execute the tower structure hazard detection method described in claims 1-7 when loaded by a processor.

Images