Concrete pole damage monitoring method and system based on modal recognition

Abstract

The invention relates to the technical field of concrete pole damage monitoring, and discloses a concrete pole damage monitoring method based on modal recognition, comprising the following steps: S1, multi-modal data acquisition; s2, edge side real-time processing and micro-damage identification; s3, performing model diagnosis based on cloud depth; s4, performing digital twin driven damage evolution prediction; and S5, performing dynamic risk assessment and decision support. According to the method, through synchronous acquisition and fusion of multi-source data, the limitation of a single sensing mode is overcome, multi-dimensional and cross-modal collaborative recognition of the microcrack germination stage is realized, and the sensitivity and reliability of early damage detection are remarkably improved; a lightweight damage identification model is deployed on the edge side, so that the instant early warning requirement of micro-damage germination is met; damage positioning and degree evaluation are carried out through the cloud based on a multi-modal fusion deep learning model, an efficient collaborative architecture of edge real-time early warning and cloud accurate diagnosis is formed, and the system response speed and decision accuracy are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of cement pole damage monitoring technology, specifically a cement pole damage monitoring method and system based on modal recognition. Background Technology

[0002] Concrete poles, as crucial supporting structures for infrastructure such as power and communications, are exposed to the natural environment for extended periods, enduring wind loads, temperature variations, and material aging. This makes them prone to developing microcracks that gradually propagate, ultimately leading to structural failure and safety accidents. Traditional concrete pole damage detection relies primarily on manual inspections and periodic maintenance, which suffers from limitations such as low efficiency, high subjectivity, and the inability to detect early damage in real time.

[0003] In recent years, structural health monitoring technology based on vibration modal analysis has made some progress, identifying damage by analyzing changes in overall modal parameters such as the structure's natural frequency and damping ratio. However, such methods are not sensitive to early local damage such as microcrack initiation, making accurate early warning difficult. Furthermore, single-sensor modes are susceptible to environmental interference, resulting in a high false alarm rate. In addition, existing monitoring systems mostly rely on centralized cloud data processing, leading to large transmission delays that fail to meet real-time requirements. Moreover, complex deep learning models have high computational resource requirements, making direct deployment on resource-constrained edge devices impossible. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for monitoring damage to cement poles based on modal recognition, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for monitoring damage to cement poles based on modal recognition includes the following steps: S1. Multimodal data acquisition: Multimodal data of the cement pole structure are acquired synchronously through a sensor array deployed on the cement pole; the multimodal data includes: vibration signals, acoustic emission signals, infrared thermal imaging data, environmental temperature and humidity data, and environmental and load data; S2. Edge-side real-time processing and micro-damage identification: In the edge computing device deployed locally on the cement pole, the collected multimodal data is preprocessed by time synchronization and spatial registration; and input into the lightweight damage identification model to identify and provide real-time early warning of micro-crack initiation. S3. Cloud-based deep learning model diagnosis: Upload the pre-processed data and early warning information from the edge to the cloud platform, and use the computing power of the cloud to accurately identify the location and extent of damage based on a deep learning diagnostic model that integrates multimodal data fusion. S4. Digital twin-driven damage evolution prediction: Based on a deep learning diagnostic model, combined with historical and real-time data, the damage evolution trend of cement poles under future load spectrum is simulated to predict their remaining service life. S5. Dynamic Risk Assessment and Decision Support: Integrates damage diagnosis results, real-time environmental information, and load information to calculate a quantitative dynamic risk assessment index, and outputs corresponding risk levels and maintenance strategies accordingly.

[0006] As a further aspect of the present invention: In step S2, the lightweight damage identification model employs a lightweight neural network that has undergone knowledge distillation and model quantization. This lightweight damage identification model is transferred from a complex teacher model trained in the cloud, and its task is to output an anomaly score.

[0007] As a further aspect of the present invention: in step S2, the specific steps for identifying the microcrack initiation stage are as follows: S21. Acoustic emission energy surge monitoring: The original acoustic emission signal is bandpass filtered in a set frequency band, and the energy of the signal in that frequency band is calculated within a continuous time window; the energy value is then converted into a decibel value; and the decibel value is monitored in real time. S22. Vibration mode change monitoring: Rapidly identify modal parameters of the vibration signal of the local area corresponding to acoustic emission; S23. Infrared thermal imaging temperature difference anomaly monitoring: In the infrared thermal image, locate the local area corresponding to vibration and acoustic emission; through differential processing, monitor whether there are local minor temperature rise anomalies; S24. Joint Logic Decision Mechanism: The judgment results of the three channels, acoustic emission energy, vibration mode parameters and infrared temperature difference anomaly, are converted into three Boolean quantities; at the end of an analysis cycle, if at least two conditions are met at the same time, it is determined to be a microcrack initiation, and an early warning is triggered.

[0008] As a further aspect of the present invention: in step S3, the specific steps for constructing the deep learning diagnostic model are as follows: S31. Multimodal data construction: Preprocess the vibration signal, acoustic emission signal, and infrared thermal imaging data respectively; use the time-frequency diagram of the vibration signal, the characteristic parameter sequence of the acoustic emission signal, and the registered thermal image of the infrared thermal imaging as multi-channel input; S32. Cross-modal feature extraction and fusion: A convolutional neural network is used to process the time-frequency plot, and frequency domain and time domain features related to damage are learned from the time-frequency plot, and feature vectors are output. Long Short-Term Memory (LSTM) networks are used to process feature parameter sequences, capturing the temporal evolution patterns of damage events from acoustic emission feature sequences and outputting feature vectors. The registered thermal images are processed using a convolutional neural network to learn spatial thermal anomaly patterns related to damage, and output feature vectors. F ir ; The three feature vectors are then concatenated to form a joint feature vector; this is then processed through one or more fully connected layers to achieve cross-modal information fusion. S33. Multi-task learning and output: A multi-task learning framework is adopted to realize damage localization and damage degree assessment.

[0009] As a further aspect of the present invention: in step S4, the specific steps for predicting the remaining useful life are as follows: S41. Digital Twin Model Construction and Dynamic Update: Based on the design drawings of the cement pole, a parametric three-dimensional finite element model is established; the damage location and degree identified in step S3 are injected into the finite element model as initial defects; and the parameters are corrected to obtain a corrected digital twin model that is highly consistent with the current physical entity state. S42. Damage Evolution Model Integration and Future Load Spectrum Input: In the digital twin model, a physical evolution model is introduced for the identified damage; then, a predicted load spectrum for a future period is input; this load spectrum is used as the dynamic boundary condition of the finite element model. S43. Real-time simulation and remaining service life prediction: Run a modified digital twin with integrated damage evolution model for accelerated cyclic simulation in the cloud; continuously monitor damage indicators in the simulation; stop the simulation when the damage indicators reach the preset critical failure size; output damage evolution curve and remaining service life.

[0010] As a further aspect of the present invention: in step S5, the specific steps for outputting the risk level and maintenance strategy are as follows: S51. Quantification of multi-source risk factors: unifying heterogeneous data from different sources into dimensionless indices; among which, the indices include: damage index, environmental index and load index; S52. Calculation of dynamic risk assessment index: Multiply the damage index, environmental index and load index to obtain the dynamic risk assessment index. S53. Risk Level Classification and Decision Support: Based on the calculated dynamic risk assessment index, risk levels are classified and corresponding preset operation and maintenance strategies are automatically triggered.

[0011] A cement pole damage monitoring system based on modal recognition includes: The sensing module includes a vibration acceleration sensor, an acoustic emission sensor, an infrared thermal imager, a temperature and humidity sensor, and strain gauges, used to acquire multimodal data; The edge computing module, deployed on-site at the cement pole, has a built-in embedded AI chip for running the lightweight damage identification model, enabling real-time data processing and micro-damage early warning; A cloud-based data analysis platform is used to run the deep learning diagnostic model, digital twin model, perform model correction, damage evolution prediction, and dynamic risk assessment. The communication module is used to enable data interaction between the edge computing module and the cloud data analysis platform.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention overcomes the limitations of a single sensing mode by synchronously acquiring and fusing multi-source data, enabling multi-dimensional, cross-modal collaborative identification of microcrack initiation, and significantly improving the sensitivity and reliability of early damage detection. By deploying a lightweight damage recognition model at the edge, the system meets the need for immediate early warning of micro-damage budding; and by using a cloud-based deep learning model based on multimodal fusion for damage localization and severity assessment, a highly efficient collaborative architecture of real-time edge early warning + precise cloud diagnosis is formed, which significantly improves system response speed and decision accuracy. By injecting identified damage as an initial defect into a digital twin model, and combining it with the laws of fracture mechanics evolution and future load spectrum, dynamic simulation of crack propagation and prediction of remaining service life can be achieved, providing a scientific basis for preventive maintenance. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating a method for monitoring damage to cement poles based on modal recognition. Detailed Implementation

[0014] Please see Figure 1 In this embodiment of the invention, the cement pole damage monitoring method based on modal recognition includes the following steps: S1. Multimodal Data Acquisition: Multimodal data of the cement pole structure is simultaneously acquired via a sensor array deployed on the pole. This multimodal data includes vibration signals, acoustic emission signals, infrared thermal imaging data, environmental temperature and humidity data, and environmental and load data. Vibration signals are acquired through an accelerometer and used to analyze the overall modal characteristics of the structure. Acoustic emission signals are acquired through acoustic emission sensors and used to capture transient stress waves during the initiation and propagation of microcracks. Infrared thermal imaging, acquired through an infrared camera, is used to identify abnormal local energy dissipation caused by damage. Environmental and load information is obtained through sensors such as temperature and humidity sensors and anemometers. S2. Edge-side Real-time Processing and Micro-damage Identification: In the edge computing device deployed locally on the cement pole, the collected multimodal data undergoes time synchronization and spatial registration preprocessing; this data is then input into the lightweight damage identification model for the identification and real-time early warning of micro-crack initiation. The methods for achieving time synchronization are as follows: All sensors are sampled using the same hardware clock source (such as a GPS module or a high-precision crystal oscillator) to ensure that all data timestamps are consistent; assuming that the sampling start point is set to T0, the vibration signal is sampled at 1kHz and the acoustic emission signal is sampled at 1MHz, but the time axes of the two are strictly aligned. The spatial registration method is as follows: In the data processing logic, a coordinate point is established for each physical location (such as a pole height of 5 meters); the vibration sensor data, acoustic emission sensor data, and temperature data of the area in the infrared image at this point are bound together and regarded as a monitoring unit.

[0015] The lightweight damage recognition model employs a lightweight neural network that has undergone knowledge distillation and model quantization, such as a 1-D version of MobileNetV2 or a simple multilayer perceptron (MLP). This lightweight damage recognition model is transferred from a complex teacher model trained in the cloud. Its task is to output an anomaly score, such as a scalar between 0 and 1, where a higher score indicates a greater likelihood of the presence of microcracks. An alarm threshold for the anomaly score can be set, such as 0.75. If the anomaly score exceeds this value, an alarm is triggered. The model size is compressed to less than 10MB and achieves an inference speed of no less than 20 frames per second on an embedded AI chip.

[0016] The specific steps for identifying the initiation stage of microcracks are as follows: S21. Acoustic Emission Energy Surge Monitoring: The original acoustic emission signal is bandpass filtered within a set frequency band, such as a 1kHz to 10kHz bandpass filter for a 1MHz original acoustic emission signal; the energy of the signal in this frequency band is calculated within a continuous time window (e.g., every 10 milliseconds); the formula for calculating energy E is: E=Σ(x i ） 2 ,in x i The signal data points are filtered within the window; then the energy value is converted into a decibel value; the formula for calculating the decibel value (EdB) is: ：E dB =10×log10 (E) It also monitors the decibel value in real time; if the threshold for the decibel value is set to 40dB, it calculates the energy and checks for any sudden increase of more than 40dB within 0.1 seconds. S22. Vibration Modal Change Monitoring: Rapidly identify modal parameters of the vibration signal in the local area corresponding to acoustic emission; such as root mean square (RMS) value, curvature mode, damping ratio, and other modal parameters; taking vibration RMS as an example, calculate the root mean square of the vibration signal within a 0.1-second time window. Determine whether the RMS value exceeds the set threshold, such as a threshold of 1.5 times the baseline level. S23. Infrared thermal imaging temperature difference anomaly monitoring: In the infrared thermal image, lock the local area corresponding to vibration and acoustic emission; through differential processing (i.e., infrared temperature difference anomaly = current temperature - previous moment or reference background temperature), monitor whether there are local small temperature rise anomalies; if the threshold for infrared temperature difference anomaly is set to 2℃, then monitor whether the infrared temperature difference anomaly exceeds 2℃. S24. Joint Logic Decision Mechanism: The judgment results of the three channels, acoustic emission energy, vibration mode parameters and infrared temperature difference anomaly, are converted into three Boolean quantities, namely True / False. At the end of an analysis cycle, when at least two conditions are met simultaneously, that is, at least two True values, it is determined that a microcrack has started, and an early warning is triggered.

[0017] I. The effectiveness of the lightweight damage recognition model will be verified by simulating the following scenario: Edge device: NVIDIA Jetson Nano; Monitoring point: 5 meters above the pole; Preset damage: A 0.08 mm microcrack is artificially created at the 10.0 second mark of the experiment; Acoustic emission energy signal (dB), vibration RMS signal (g), and infrared thermal imaging data (infrared temperature difference anomaly (Δ℃)) were recorded between 9.98 and 10.02 seconds, and the anomaly score and edge device action were output; the results are shown in Table 1 below. Table 1 Multimodal Data Analysis Table Serial Number Timestamp Acoustic emission energy Vibration RMS Infrared temperature difference abnormality Abnormal scores Edge device actions 1 9.98 38 0.0052 0.1 0.15 Continuous monitoring 2 9.99 42 0.0051 0.2 0.31 Continuous monitoring 3 10.00 25 0.0050 0.1 0.11 Continuous monitoring 4 10.01 55 0.0053 0.8 0.82 Triggering an alert 5 10.02 48 0.0055 0.5 0.79 Continuous warning 6 10.03 41 0.0052 0.3 0.45 Warning ended, data packet uploaded. From Table 1 above, we can conclude that: (1) When the microcrack is generated at 10.00 seconds, the single vibration signal (RMS) hardly changes (0.0050g~0.0053g), which is insufficient to trigger an alarm; however, the acoustic emission energy increases dramatically (25dB~55dB), and the infrared also captures a small temperature rise; the lightweight AI model that integrates all this information has its anomaly score soar from 0.11 to 0.82 in an instant, successfully crossing the warning threshold of 0.75, and realizing real-time warning within 0.01 seconds after the damage occurs.

[0018] (2) The average processing latency of the entire process on the edge device is less than 10 milliseconds, which meets the real-time requirements. After testing, the size of the lightweight model used is 4.2MB, which is much smaller than the original teacher model of 450MB, proving the effectiveness of knowledge distillation and quantization technology, enabling complex AI algorithms to run stably on resource-constrained edge devices.

[0019] S3. Cloud-based Deep Learning Model Diagnosis: Preprocessed data and early warning information from the edge are uploaded to the cloud platform. Utilizing cloud computing power, a deep learning diagnostic model based on multimodal data fusion is used to accurately identify the location and extent of damage. The specific steps for building a deep learning diagnostic model are as follows: S31. Multimodal data construction: Preprocess the vibration signal, acoustic emission signal, and infrared thermal imaging data respectively; use the time-frequency diagram of the vibration signal, the characteristic parameter sequence of the acoustic emission signal, and the registered thermal image of the infrared thermal imaging as multi-channel input; S32. Cross-modal feature extraction and fusion: A convolutional neural network is used to process the time-frequency plot, learning the frequency and time domain features related to the damage from the plot, and outputting a feature vector. F v ; Long Short-Term Memory (LSTM) networks are used to process feature parameter sequences to capture the temporal evolution patterns of damage events from acoustic emission feature sequences, and feature vectors are output. F ae ; The registered thermal images are processed using a convolutional neural network to learn spatial thermal anomaly patterns related to damage, and output feature vectors. F ir ; Then take the three feature vectors F v ,F ae , F ir The features are concatenated to form a joint feature vector; then processed through one or more fully connected layers to achieve cross-modal information fusion. S33. Multi-task learning and output: A multi-task learning framework is adopted to achieve damage localization (regression task) and damage severity assessment (classification task), wherein... The output layer for lesion localization has a fully connected layer with two neurons, using a linear activation function; it outputs the coordinates of the lesion center. (X,Y) ;in, X Represents the position along the height of the pole. Y Represents angle or clock direction; The output layer for assessing the severity of injury has a fully connected layer with four neurons, using the Softmax activation function; it outputs a probability distribution for four severity levels, as shown below. P (No damage) P (slight), P (medium), P (serious)]; II. The effectiveness of the model is verified by simulating the following scenarios: A labeled dataset containing 3000 samples was used for training and testing in the cloud; the average error (meters) of damage localization and the accuracy of damage classification were obtained under the Benostat modality data; see Table 2 below for details; Table 2 Model Accuracy Analysis Table Serial Number Modal data Average error in damage localization (meters) Damage severity classification accuracy 1 Vibration time-frequency diagram only 0.85 78.5% 2 Acoustic emission characteristics only Unable to locate 82.1% 3 Infrared thermal images only 0.45 80.3% 4 Multimodal fusion 0.15 95.8% Table 2 above shows that the multimodal fusion model is significantly better than any single-modal model in terms of both localization accuracy and classification accuracy, thus proving the necessity and effectiveness of cross-modal feature fusion.

[0020] Specific diagnostic cases: A simulated diagnosis was performed on a 10-meter-high concrete pole. The actual damage location of the concrete pole was (7.5 meters, 2 o'clock position), with a crack about 15mm long, which is classified as moderate damage. The model input is as follows: Vibration time-frequency diagram: shows energy decay at 120Hz and 250Hz; Acoustic emission sequence: Shows continuous acoustic emission events with high energy and high count; Infrared image: A localized high-temperature area was found at a height of about 7.5 meters on the pole, which was +8°C higher than the background temperature; The model output is as follows: Damage location: (7.3, 2 o'clock direction); very close to the actual damage location, with an error of only 0.2 meters; Damage level: [0.01, 0.04, 0.90, 0.05], judged as "moderate damage" (probability 90%); same as the actual damage level.

[0021] S4. Damage Evolution Prediction Driven by Digital Twins: Based on a deep learning diagnostic model, combining historical and real-time data, the damage evolution trend of cement poles under future load spectra is simulated to predict their remaining service life; the specific steps for predicting the remaining service life are as follows: S41. Digital Twin Model Construction and Dynamic Update: Based on the design drawings of the cement pole, a parametric three-dimensional finite element model is established; the damage location and degree identified in step S3 are injected into the finite element model as initial defects; and the parameters are corrected to obtain a corrected digital twin model that is highly consistent with the current physical entity state; if Bayesian update or Kalman filter algorithms are used to adjust the parameters in the model, the simulated modal frequencies and mode shapes are most closely matched with the current measured data. S42. Damage Evolution Model Integration and Future Load Spectrum Input: In the digital twin model, a physical evolution model is introduced for the identified damage; for example, for cracks, Paris's law from fracture mechanics is used to describe the propagation rate, and the formula for calculating the propagation rate is as follows: ;in, This represents the amount of crack propagation per stress cycle. ΔK The stress intensity factor amplitude (calculated from the load by the finite element model); C,m The material constant is then used; the predicted load spectrum for a future period (such as the next six months) is then input (e.g., wind speed spectrum predicted based on historical meteorological data, to calculate wind pressure load); this load spectrum is used as the dynamic boundary condition of the finite element model. S43. Real-time simulation and remaining service life prediction: Run the modified digital twin with integrated damage evolution model for accelerated cyclic simulation in the cloud; the simulation process is as follows: Apply load → Calculate stress intensity factor ΔK → Update crack size using Paris's law → Update crack geometry in the model → Loop; The simulation continuously monitors damage indicators (such as crack length); when the damage indicators reach the preset critical failure size (i.e., it is determined to be failure), the simulation stops; and the damage evolution curve and remaining service life are output. III. Damage evolution is predicted by simulating the following scenario: Diagnostic results: A 10mm crack was found at 5m on the cement pole; the material constants in Paris's law are C=2e-10, m=3.2, and the critical crack length is taken as 50mm; the prediction results are shown in Table 3 below. Table 3 Analysis of Damage Evolution Prediction Results Serial Number Time (month) Predicted crack length (mm) Average wind speed for the month (m / s) Remark 1 0 10.0 8.3 Current status 2 3 12.5 8.5 slow growth 3 6 16.2 10.2 Accelerated growth 4 9 22.1 12.5 Expansion Acceleration 5 12 31.5 9.0 Extended persistence 6 15 45.8 10.3 Approaching the critical value, triggering an advanced alert. 7 15.5 50.0 9.6 Predicted failure Table 3 above shows that the predicted remaining service life is approximately 15.5 months; that is, the critical damage state will be reached after 15.5 months; it is recommended to carry out preventive maintenance or replacement within 12 months, that is, before the crack expands to about 31.5 mm.

[0022] S5. Dynamic Risk Assessment and Decision Support: Integrates damage diagnosis results, real-time environmental information, and load information to calculate a quantified dynamic risk assessment index, and outputs corresponding risk levels and maintenance strategies accordingly; among which, The specific steps for outputting risk levels and maintenance strategies are as follows: S51. Quantification of Multi-Source Risk Factors: This involves uniformly quantifying heterogeneous data from different sources into dimensionless indices; these indices include: damage index, environmental index, and load index; among which, (1) Damage index: The severity level of damage obtained from step S3 (e.g., 0-none, 1-slight, 2-moderate, 3-severe); and directly mapped to an index. DI ; No damage → 0; Minor → 1; Moderate → 2; Severe → 3; (2) Environmental Index: Input real-time environmental data, such as temperature, humidity, and freezing conditions; and map environmental conditions to an index by looking up a predefined environmental severity table. EC ; For example, room temperature drying (20℃, 50%RH) → 1.0; high temperature and high humidity (35℃, 90%RH) → 1.3 (accelerated corrosion); frozen environment (<0℃) → 1.8 (material embrittlement, freeze-thaw cycle); (3) Load index: Input real-time load data; such as stress σ converted from strain gauges. actual The wind pressure can be calculated from the anemometer reading or by converting it from the wind speed meter; and the real-time load and design load σ can be calculated. design The ratio of , and map it to the exponent LC; For example, LC=max(1.0,σ actual / σ design ×1.5); This means that when the actual load reaches 2 / 3 of the design load, LC≈1.0; when it approaches or exceeds the design load, LC increases rapidly; such as above 1.5; S52. Calculation of Dynamic Risk Assessment Index: Multiply the damage index, environmental index, and load index to obtain the dynamic risk assessment index; Dynamic Risk Assessment Index DRI = DI × EC × LC ; S53. Risk Level Classification and Decision Support: Based on the calculated dynamic risk assessment index, risk levels are classified, and corresponding preset operation and maintenance strategies are automatically triggered; for example, DRI < 1.5, the risk level is defined as low risk, and the color code is green; Automated decision and recommendation: Continue monitoring and record data, no special operation required; 1.5≤DRI<3.0, the risk level is defined as medium risk, and the color code is yellow; Automatic decision and recommendation: planned maintenance, generate work orders, and recommend on-site inspection and maintenance in the next planned maintenance cycle (e.g., within 1 month); 3.0≤DRI<6.0, the risk level is defined as high risk, and the color code is orange; Automatic decision and suggestion: strengthen monitoring and maintenance preparation, increase data collection frequency, send alarms to maintenance personnel, and suggest arranging maintenance within 1 week; DRI≥6.0, the risk level is defined as extremely high risk, and the color code is red; Automatic decision and suggestion: Immediately shut down for maintenance, issue the highest level alarm, notify the person in charge via SMS, App push, etc., suggest that they go to the site immediately for handling, and consider temporary isolation measures; IV. By simulating the following scenarios, calculate the Dynamic Risk Assessment Index (DRI) under different scenarios; and classify risk levels and provide decision support; the results are shown in Table 4 below. Table 4 Risk Level Classification and Decision Support Analysis Table Scene Description Index DI EC Index Index LC DRI Index Risk level System Decision 1. Normal weather conditions, no damage. 0 1.0 1.0 0 Low Continue monitoring 2. Under normal weather conditions, minor cracks 1 1.0 1.0 1 Low Continue monitoring 3. Minor cracks may occur during strong winds. 1 1.0 1.8 1.8 middle Planned maintenance 4. Minor cracks in freezing weather 1 1.8 1.0 1.8 middle Planned maintenance 5. Strong winds and freezing temperatures caused minor cracks. 1 1.8 1.8 3.24 high Repair within one week 6. Normal weather conditions, severe damage 3 1.0 1.0 3 high Repair within one week 7. Strong winds + freezing temperatures, causing severe damage. 3 1.8 1.8 9.72 Extremely high Immediately stop the machine for inspection and repair. From Table 4 above, we can conclude that: (1) Comparison of scenarios 2, 3, 4, and 5: all are minor damages. The risk is low under normal conditions, but rises to medium under strong winds or freezing conditions, and rises sharply to high under the combined effects of strong winds and freezing conditions. This perfectly reflects the dynamic change characteristics of risk, which static assessment methods cannot achieve. (2) Scenario 6 shows that even if the environment is good, the risk is still high if the damage itself is severe enough; Scenario 7 is the most dangerous working condition, where damage, environment, and load are all in an extremely unfavorable state, the DRI index is off the charts, triggering the most urgent response; this is completely in line with common sense in engineering and risk management principles. (3) The system outputs no longer raw data, but direct, tiered action guidelines; this greatly reduces the technical threshold and decision-making pressure for maintenance personnel. For example, for scenario 5, the system will not simply display: crack length 10mm, temperature -5℃, wind speed 20m / s; but will directly tell the maintenance team: high risk, it is recommended to inspect within a week, and provide detailed data support.

[0023] A cement pole damage monitoring system based on modal recognition includes: The sensing module includes a vibration acceleration sensor, an acoustic emission sensor, an infrared thermal imager, a temperature and humidity sensor, and strain gauges, used to acquire multimodal data; The edge computing module, deployed on-site at the cement pole, has a built-in embedded AI chip for running the lightweight damage identification model, enabling real-time data processing and micro-damage early warning; A cloud-based data analysis platform is used to run the deep learning diagnostic model, digital twin model, perform model correction, damage evolution prediction, and dynamic risk assessment. The communication module is used to enable data interaction between the edge computing module and the cloud data analysis platform.

[0024] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring damage to cement poles based on modal recognition, characterized in that, Includes the following steps: S1. Multimodal data acquisition: Multimodal data of the cement pole structure are acquired synchronously through a sensor array deployed on the cement pole; the multimodal data includes: vibration signals, acoustic emission signals, infrared thermal imaging data, environmental temperature and humidity data, and environmental and load data; S2. Edge-side real-time processing and micro-damage identification: In the edge computing device deployed locally on the cement pole, the collected multimodal data is preprocessed by time synchronization and spatial registration; and input into the lightweight damage identification model to identify and provide real-time early warning of micro-crack initiation. S3. Cloud-based deep learning model diagnosis: Upload the pre-processed data and early warning information from the edge to the cloud platform, and use the computing power of the cloud to accurately identify the location and extent of damage based on a deep learning diagnostic model that integrates multimodal data fusion. S4. Digital twin-driven damage evolution prediction: Based on a deep learning diagnostic model, combined with historical and real-time data, the damage evolution trend of cement poles under future load spectrum is simulated to predict their remaining service life. S5. Dynamic Risk Assessment and Decision Support: Integrates damage diagnosis results, real-time environmental information, and load information to calculate a quantitative dynamic risk assessment index, and outputs corresponding risk levels and maintenance strategies accordingly.

2. The method for monitoring damage to cement poles based on modal recognition according to claim 1, characterized in that, In step S2, the lightweight damage identification model employs a lightweight neural network that has undergone knowledge distillation and model quantization. This lightweight damage identification model is transferred from a complex teacher model trained in the cloud, and its task is to output an anomaly score.

3. The method for monitoring damage to cement poles based on modal recognition according to claim 1, characterized in that, In step S2, the specific steps for identifying the microcrack initiation stage are as follows: S21. Acoustic emission energy surge monitoring: The original acoustic emission signal is bandpass filtered in a set frequency band, and the energy of the signal in that frequency band is calculated within a continuous time window; the energy value is then converted into a decibel value; and the decibel value is monitored in real time. S22. Vibration mode change monitoring: Rapidly identify modal parameters of the vibration signal of the local area corresponding to acoustic emission; S23. Infrared thermal imaging temperature difference anomaly monitoring: In the infrared thermal image, locate the local area corresponding to vibration and acoustic emission; through differential processing, monitor whether there are local minor temperature rise anomalies; S24. Joint Logic Decision Mechanism: The judgment results of the three channels, acoustic emission energy, vibration mode parameters and infrared temperature difference anomaly, are converted into three Boolean quantities; at the end of an analysis cycle, if at least two conditions are met at the same time, it is determined to be a microcrack initiation, and an early warning is triggered.

4. The method for monitoring damage to cement poles based on modal recognition according to claim 1, characterized in that, In step S3, the specific steps for constructing the deep learning diagnostic model are as follows: S31. Multimodal data construction: Preprocess the vibration signal, acoustic emission signal, and infrared thermal imaging data respectively; use the time-frequency diagram of the vibration signal, the characteristic parameter sequence of the acoustic emission signal, and the registered thermal image of the infrared thermal imaging as multi-channel input; S32. Cross-modal feature extraction and fusion: A convolutional neural network is used to process the time-frequency plot, and frequency domain and time domain features related to damage are learned from the time-frequency plot, and feature vectors are output. Long Short-Term Memory (LSTM) networks are used to process feature parameter sequences, capturing the temporal evolution patterns of damage events from acoustic emission feature sequences and outputting feature vectors. The registered thermal images are processed using a convolutional neural network to learn spatial thermal anomaly patterns related to damage, and output feature vectors. F ir ; The three feature vectors are then concatenated to form a joint feature vector; this is then processed through one or more fully connected layers to achieve cross-modal information fusion. S33. Multi-task learning and output: A multi-task learning framework is adopted to realize damage localization and damage degree assessment.

5. The method for monitoring damage to cement poles based on modal recognition according to claim 1, characterized in that, In step S4, the specific steps for predicting the remaining useful life are as follows: S41. Digital Twin Model Construction and Dynamic Update: Based on the design drawings of the cement pole, a parametric three-dimensional finite element model is established; the damage location and degree identified in step S3 are injected into the finite element model as initial defects; Then, the parameters are corrected to obtain a corrected digital twin model that is highly consistent with the current physical entity state; S42. Damage Evolution Model Integration and Future Load Spectrum Input: In the digital twin model, a physical evolution model is introduced for the identified damage; then, a predicted load spectrum for a future period is input; this load spectrum is used as the dynamic boundary condition of the finite element model. S43. Real-time simulation and remaining service life prediction: Run a modified digital twin with integrated damage evolution model for accelerated cyclic simulation in the cloud; continuously monitor damage indicators in the simulation; stop the simulation when the damage indicators reach the preset critical failure size; output damage evolution curve and remaining service life.

6. The method for monitoring damage to cement poles based on modal recognition according to claim 1, characterized in that, In step S5, the specific steps for outputting the risk level and maintenance strategy are as follows: S51. Quantification of multi-source risk factors: unifying heterogeneous data from different sources into dimensionless indices; among which, the indices include: damage index, environmental index and load index; S52. Calculation of dynamic risk assessment index: Multiply the damage index, environmental index and load index to obtain the dynamic risk assessment index. S53. Risk Level Classification and Decision Support: Based on the calculated dynamic risk assessment index, risk levels are classified and corresponding preset operation and maintenance strategies are automatically triggered.

7. A system for implementing the modal recognition-based cement pole damage monitoring method according to any one of claims 1-6, characterized in that, include: The sensing module includes a vibration acceleration sensor, an acoustic emission sensor, an infrared thermal imager, a temperature and humidity sensor, and strain gauges, used to acquire multimodal data; The edge computing module, deployed on-site at the cement pole, has a built-in embedded AI chip for running the lightweight damage identification model, enabling real-time data processing and micro-damage early warning; A cloud-based data analysis platform is used to run the deep learning diagnostic model, digital twin model, perform model correction, damage evolution prediction, and dynamic risk assessment. The communication module is used to enable data interaction between the edge computing module and the cloud data analysis platform.

Images