Real-time diagnosis method for damage of metal structure node based on acoustic emission and deep learning

Abstract

The present application relates to the technical field of metal structure health monitoring, and discloses a kind of acoustic emission deep learning real-time diagnosis method for metal structure node damage, comprising: the acoustic transfer function fingerprint under the calculation environment noise is generated, and its perturbation time series relative to health benchmark is generated, to continuously perceive the evolution of structure physical state;When the sequence presents continuous deterioration trend, trigger active acoustic wave challenge to identify the physical properties of damage, and periodically carry out sensor network self-checking, finally, the perturbation time series is input into deep learning model together with damage property and sensor state, and output diagnostic information.The present application changes the diagnostic focus from rare events to continuous state monitoring, and solves the monitoring blank problem of traditional acoustic emission technology during the event interval.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of acoustic emission deep learning real-time diagnosis method of metal structure node damage, belong to metal structure health monitoring technical field. BACKGROUND

[0002] At present, the transient elastic wave generated by crack initiation or expansion in the material is captured and analyzed by using acoustic emission technology, which is a commonly used technical method, which can effectively respond to the local damage event that has released enough energy, and provides direct basis for structure safety evaluation.

[0003] However, for long-term service of major infrastructure, its structure safety not only needs to deal with occasional rapid damage, but also needs to effectively manage the long-term slow deterioration process of material performance caused by fatigue corrosion or creep. In this application background, the event-driven working mode of acoustic emission technology may have a long monitoring information blank period between two high-energy acoustic emission events that can be identified, and during this period, the health status of the structure may be continuously deteriorating in a silent acoustic way. The prior art is difficult to perceive the accumulation process of this risk.

[0004] To improve the detection ability of early weak damage, a direct technical path is to improve the receiving sensitivity of the sensor or increase the deployment density of the sensor. However, the energy of the environment and the running noise in the real structure is usually much higher than the weak elastic wave released by the early micro-crack. Simply improving the sensitivity of the system will make the weak damage signal more thoroughly submerged in the background noise amplified synchronously, thereby increasing the false alarm rate and affecting the reliability and practicability of the monitoring system. Specifically, the prior art mainly has the following two technical problems: 1. Weak damage signal is easily submerged in high-energy background noise, affecting the sensitivity and accuracy of monitoring; 2. The event monitoring method has a monitoring blank period between events, and cannot continuously track the slow deterioration process of material physical properties. Therefore, how to establish a method that can continuously measure the gradual change of the physical properties of the structure node material itself without relying on the occurrence of high-energy discrete damage events has become a technical problem to be solved by the present application. SUMMARY

[0005] The present application provides an acoustic emission deep learning real-time diagnosis method for metal structure node damage, which mainly aims to solve the problem that the prior art cannot continuously measure the gradual change of the physical properties of the structure material in an environment full of background noise.

[0006] To achieve the above purpose, the present application provides an acoustic emission deep learning real-time diagnosis method for metal structure node damage, which establishes and executes the following monitoring and diagnosis procedures:

[0007] Based on the environmental background noise collected by the reference sensor and the monitoring sensors, a baseline acoustic transfer function fingerprint representing the health state of the metal structure node is calculated and stored;

[0008] Periodically, a real-time acoustic transfer function fingerprint is obtained, and a perturbation degree between the real-time acoustic transfer function fingerprint and the baseline acoustic transfer function fingerprint is calculated, thereby generating a perturbation degree time series;

[0009] It is judged whether the value of the perturbation degree time series continuously increases within a predetermined number of calculation periods and the growth slope exceeds a degradation rate threshold. If the judgment is yes, an active challenge is triggered. The active challenge controls one of the monitoring sensors to switch to a sound wave emission mode to emit a challenge sound wave, and the remaining monitoring sensors receive a response sound wave. By analyzing the harmonic nonlinear acoustic characteristics generated in the response sound wave, the physical properties of the damage are determined;

[0010] Periodically, at least one pair of monitoring sensors is selected, which are sound wave emitters and receivers, respectively. Based on the acoustic reciprocity principle, the forward and reverse acoustic transfer functions are compared to diagnose the coupling state of the sensor itself and the metal structure node;

[0011] Finally, the perturbation degree time series is input into a deep learning model, and according to the physical properties of the damage and the state of the sensor and the coupling, diagnostic information about the damage state of the metal structure node is output.

[0012] Preferably, the harmonic nonlinear acoustic characteristics generated in the response sound wave are analyzed by calculating a nonlinear response index R_nl (t), where: R_nl (t)=E_harmonic / E_fundamental, where E_fundamental is the energy of the response sound wave at the reference frequency of the challenge sound wave, and E_harmonic is the energy of the response sound wave at the newly generated harmonic frequency; and the nonlinear response index R_nl (t) is compared with a nonlinear damage threshold preset according to the fracture mechanics properties of the metal structure material to distinguish the physical properties of the damage.

[0013] Preferably, the method further establishes and implements the following rules: continuously monitor the energy level of the environmental background noise signal, and when the energy level is lower than a silence threshold determined based on the average energy level of the baseline acoustic transfer function fingerprint calculation period, automatically activate the autonomous heartbeat mode; the autonomous heartbeat mode includes: controlling one of the monitoring sensors to emit an internally generated energy-stable heartbeat sound wave, and using the heartbeat sound wave as an alternative excitation source to maintain the continuous generation of the perturbation degree time series.

[0014] Preferably, the method further establishes and implements the following rules: in the process of calculating the acoustic transfer function fingerprint, the components coherent with the signals collected by the reference sensor are removed from the signals collected by the monitoring sensor to obtain a residual signal; the energy and spectral kurtosis of the residual signal are analyzed to generate an index representing the coupling quality between the monitoring sensor and the node of the metal structure; and when outputting the diagnostic information, the monitoring data is weighted according to the coupling quality index.

[0015] Preferably, the method further establishes and implements the following rules: when generating the perturbation time series, the power spectrum of the environmental background noise signal collected by the reference sensor is calculated, and a spectral entropy index representing the signal as the excitation source quality is generated based on the power spectrum; and the weight of the perturbation time series in outputting the final diagnostic information is dynamically adjusted according to the spectral entropy index.

[0016] Preferably, the rules for diagnosing the sensor itself and the coupling state between the sensor and the node of the metal structure further include: when the difference between the forward and reverse acoustic transfer functions exceeds an reciprocity error threshold set according to the sensor factory calibration data, it is determined that the corresponding sensor or its coupling state has failed, and the data weight from the corresponding sensor is automatically reduced or logically isolated when outputting the diagnostic information.

[0017] Preferably, the perturbation is determined by calculating the integral of the difference between the logarithmic amplitude of the real-time acoustic transfer function fingerprint and the logarithmic amplitude of the reference acoustic transfer function fingerprint over the entire frequency band.

[0018] Preferably, the deterioration rate threshold is set in advance according to the design safety level and material fatigue properties of the node of the metal structure.

[0019] Preferably, the method further establishes and implements the following rules: the continuous state sequence composed of the perturbation time series and the discrete event sequence composed of the monitored transient high-energy acoustic emission events are obtained in parallel; the transfer entropy algorithm is used to calculate the information flow from the discrete event sequence to the continuous state sequence to generate an index representing the time sequence causal correlation degree of the two; and according to the index, the external pseudo-source storm interference working condition is identified and false positives are suppressed.

[0020] Preferably, the deep learning model is an autoencoder model, which determines whether the perturbation time series has an abnormal evolution pattern by identifying whether the reconstruction error of the perturbation time series exceeds a dynamic baseline determined by the reconstruction error statistical distribution generated by the model when processing the perturbation time series in a healthy state, thereby outputting the diagnostic information.

[0021] Compared with the prior art, the present application has the following advantages:

[0022] 1. The method establishes a working mode for continuously monitoring the physical properties of a structure, which uses the always-existing environmental vibration as a detection medium, and obtains a stable basis representing the current physical state of the node by calculating the acoustic transfer function generated when the vibration propagates in the structure node. When progressive physical changes occur inside the material, the conduction characteristics of the material to the acoustic wave will also change, which is reflected in the continuous drift of the acoustic transfer function. The method shifts the focus of diagnosis from capturing rare high-energy burst damage events to observing the continuous evolution of the physical state of the structure by analyzing the time series of the differences in the function, so that whether the structure health state has a continuous small deterioration process between two burst events becomes a directly observable fact, avoiding the blank area of monitoring information due to no event occurrence in the traditional event monitoring mode.

[0023] 2. The method constructs a diagnosis process triggered by state monitoring and property identification. After confirming that the acoustic transfer function of the structure node shows an irreversible continuous evolution trend, the system will use this trend as a trigger condition to temporarily change the working mode of the monitoring network. It controls one of the monitoring sensors to emit a preset challenge acoustic wave, and the response acoustic wave after passing through the node is received by the remaining sensors. Due to the physical properties of nonlinear damage such as micro-cracks, the passing acoustic wave will be modulated and generate new frequency components, while linear physical changes such as bolt relaxation do not have this property. By analyzing the differences in the frequency spectrum composition of the response acoustic wave relative to the challenge acoustic wave, the system can qualitatively distinguish the physical reasons that cause the change in the transfer function. This combination of continuous state trend monitoring and discrete physical property identification makes the final output of the diagnosis information not only contain the conclusion that the state is changing, but also contain the physical evidence of why it is changing, improving the completeness and credibility of the diagnosis conclusion.

[0024] 3. The method integrates a self-checking mechanism for verifying the diagnosis premise. The system for structural damage diagnosis uses the information obtained during the calculation of the acoustic transfer function to evaluate the coupling state. The residual signal in the monitoring signal that cannot be explained by the reference signal can directly reflect the physical coupling quality between the monitoring sensor and the structure. Poor coupling will cause the energy of the residual signal to rise and the frequency spectrum to be abnormal. By periodically controlling pairs of sensors to transmit and receive each other and comparing the acoustic transfer functions in the forward and reverse directions, the sensor performance can be judged according to the acoustic reciprocity principle. These two parallel running self-checking processes ensure that all data sources used for structural diagnosis are reliable, and the final diagnosis conclusion is based on a stable and reliable measurement basis that has been internally verified. BRIEF DESCRIPTION OF DRAWINGS

[0025] Fig. 1The state monitoring driven closed-loop diagnostic flowchart of the present application;

[0026] Fig. 2 The excitation source spectrum entropy and diagnostic weight mapping relationship diagram of the present application;

[0027] Fig. 3 The state machine diagram of the multi-mode operation logic of the monitoring system of the present application. DETAILED DESCRIPTION

[0028] In order to make the technical solutions and advantages of the present application clearer and more apparent, the present application will be further described in detail below with specific embodiments, and it should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the protection scope of the present application.

[0029] The application discloses a kind of metal structure node damage acoustic emission deep learning real-time diagnosis method, its overall architecture is established on a monitoring procedure by continuous state perception discrete event discrimination and system self-checking;The procedure first passes through a calibration stage, using the environmental background noise that always exists, establishes a reference acoustic transfer function fingerprint that can represent the physical characteristics of elastic wave conduction of metal structure node in healthy state, and then in a continuous monitoring stage, by comparing the acoustic transfer function fingerprint obtained in real time with the reference, a perturbation time series that can quantify the progressive change of the physical characteristics is generated, finally, the procedure uses this time series as the core diagnostic basis, supplemented by an active sound wave interrogation mechanism for identifying damage physical properties triggered by the abnormal evolution trend of the sequence, and a sensor network self-checking mechanism running in parallel to ensure the reliability of measurement premise, jointly output the diagnosis information about node damage state;In a typical application scenario, for example, long-term health monitoring is carried out on the key node weld area of a large-span metal truss bridge (such as a steel truss or aluminum alloy truss bridge), which continuously bears complex stress caused by traffic wind load and environmental temperature change, and there is a risk of fatigue microcrack initiation;In order to realize continuous measurement of the progressive change of material physical properties in such high-energy background noise engineering environment, the monitoring system using the method of the application is configured to perform the following operation procedure;Monitoring system is laid out and initial state is defined, a monitoring network composed of multiple piezoelectric acoustic emission sensors is arranged in the key node area, wherein at least one sensor is defined as a reference sensor, which is installed on the base material away from the weld heat-affected zone, and the structure response is relatively stable, used to collect the original environmental background noise signal S_ref (t) as excitation source, the remaining sensors are used as monitoring sensors, and are directly arranged in the weld and surrounding area to collect the response signal S_(mon_i ) (t) after propagation via node path, all sensors are connected with a multi-channel synchronous data acquisition card through low-noise cable, the sampling frequency of the acquisition card is set to not less than 2MHz to cover the main energy band of acoustic emission signal.

[0030] After the monitoring system is deployed and enters the calibration phase, an objective basis that can stably represent the physical characteristics of the structure needs to be extracted from the random environmental noise; for this purpose, the method performs the calculation and storage steps of the reference acoustic transfer function fingerprint, which is performed when the structure is confirmed to be in a healthy state, the system continuously collects the signals of the reference sensor and each monitoring sensor for 24 hours without interruption, and the collected long-term signals are divided into several data frames, for example, each frame has a length of 1024 sampling points, for each data frame, the data processing unit calculates the cross-power spectral density between the reference signal and the monitoring signal and the self-power spectral density of the reference signal by performing Fourier transform on the reference signal and the monitoring signal, and finally obtains the cross-power spectral density from the reference point to the first monitoring point, the second monitoring point, and the nth monitoring point. The acoustic transfer function H_i (f) between each monitoring point, the system will average all H_i (f) calculated within 24 hours to eliminate the influence of short-term environmental noise fluctuations, so as to obtain a statistically stable baseline acoustic transfer function fingerprint set {H_1 (f), H_2(f),...} describing the health node's wideband sound conduction characteristics, which is stored in the local memory as the baseline for all subsequent diagnostic analysis; After completing the calibration and entering the long-term continuous monitoring phase, it is necessary to continuously quantify the possible small changes in the physical properties of the node; To achieve this purpose, the system is configured to periodically generate a perturbation time series, the system repeatedly performs the operation of acquiring the acoustic transfer function at a fixed time interval, for example every 60 seconds, so as to obtain the real-time acoustic transfer function fingerprint {H_1_' (f), H_2_' (f),...} at the current time, then the processor compares the real-time acoustic transfer function fingerprint with the stored baseline acoustic transfer function fingerprint, and generates a single scalar value, i.e. perturbation D(t), through a deterministic calculation procedure, which is specifically, calculating the difference between the logarithmic amplitude of the real-time acoustic transfer function fingerprint and the logarithmic amplitude of the baseline acoustic transfer function fingerprint, and integrating it within the entire effective frequency band, for example 50kHz to 400kHz, so as to obtain a quantitative index reflecting the small difference; For example, at a certain calculation period t, if the logarithmic amplitude of the baseline acoustic transfer function fingerprint at three discrete frequency points f_1, f_2, f_3 is [10.5, 12.1, 9.8], and the logarithmic amplitude of the real-time acoustic transfer function fingerprint at the corresponding frequency points is [10.6, 12.0, 9.9], then the perturbation D(t) at this time is proportional to the cumulative value of the difference between the two sets of values, which is 0.3 in this example, by continuously performing this operation, the system converts complex multi-dimensional function changes into a single variable perturbation time series D(t), any irreversible increasing trend of the sequence directly reflects that the physical properties of the node's internal materials are undergoing continuous irreversible changes.

[0031] When the perturbation time series D(t) is observed to exhibit a sustained growth trend, it is necessary to further distinguish whether the change is caused by crack initiation or other physical changes such as bolt pre-tightening force relaxation; In view of this, the system establishes and executes a damage physical property active interrogation procedure triggered by state monitoring, and the triggering condition of the procedure is set as: when the value of the perturbation time series D(t) continuously and monotonously increases for a predetermined number of calculation periods, for example, for 10 consecutive calculation periods, and the growth slope of the sequence calculated by the least square method exceeds a preset degradation rate threshold, an active interrogation is automatically triggered; wherein the degradation rate threshold is set according to the design safety level of the monitored metal structure node and the fatigue crack propagation rate curve of the metal material used, a higher safety level corresponds to a lower degradation rate threshold; Once the interrogation is triggered, the system control logic temporarily changes the working mode of the sensor network, selects one of the monitoring sensors, for example, sensor A, changes its working mode from signal reception to sound wave emission, and applies an energy-controlled standardized electric pulse, for example, a 5-cycle sine wave packet with a center frequency of 150 kHz, to it through a digital analog converter, so that it emits a beam of interrogation sound waves, and the remaining monitoring sensors in the region, for example, sensors B and C, synchronously capture the response sound waves after propagation through the node; Since micro-cracks will produce crack interface opening and closing under the action of sound wave stress, this nonlinear physical behavior will modulate the passing sound wave and generate new high-frequency harmonic components, while other linear physical changes do not have this feature, therefore, by performing frequency spectrum analysis on the response sound wave, and calculating a nonlinear response index , that is, R_nl(t)=E_harmonic / E_fundamental, wherein E_fundamental is the energy of the response sound wave at the reference frequency of the interrogation sound wave, and E_harmonic is the energy of the response sound wave at the newly generated second or third harmonic frequency; If the calculated R_nl(t) is higher than a nonlinear damage threshold preset according to the fracture mechanics properties of the metal structure material, the diagnosis conclusion is updated to exist nonlinear damage, thereby providing physical evidence for the sustained growth of D(t).

[0032] It should be noted that the effectiveness of all the above diagnostic logics is based on the premise that the sensor itself and its acoustic coupling with the structure are stable; in order to exclude the interference of sensor performance drift or coupling state degradation on the diagnostic results, the method also establishes a set of periodic sensor network state self-checking procedures, which utilizes the principle of acoustic reciprocity; the specific operation is that the system selects at least one pair of monitoring sensors, such as sensor A and sensor B, switches sensor A to the transmitting mode, transmits a standardized probe pulse, receives and calculates the forward transfer function H ab by sensor B, immediately switches sensor B to the transmitting mode, transmits the same probe pulse, receives and calculates the reverse transfer function H ba by sensor A, and the processor obtains a reciprocity error E ab = | H ab - H ba | by calculating the difference between the two, if the error exceeds a reciprocity error threshold set according to the sensor factory calibration data and the acoustic characteristics of the coupling agent, it directly indicates that the performance of sensor A or B or its coupling state has changed asymmetrically, the system can locate the specific sensor that has failed or drifted by polling different sensor pairs in the test network, and automatically reduces the data weight from the faulty sensor or logically isolates it when outputting the final diagnostic information.

[0033] Further, in order to cope with different boundary conditions in different application scenarios, the method also integrates several adaptive adjustment mechanisms; for example, when the monitoring object enters the working condition with low environmental noise energy at night and other time periods, it may lead to the failure to calculate the effective acoustic transfer function, for this purpose, the system is configured to continuously monitor the real-time energy level of the signal collected by the reference sensor, when the energy level is continuously lower than a silence threshold determined based on the average energy level in the calibration stage, the system automatically activates an autonomous heartbeat mode, which controls one of the monitoring sensors to periodically emit an internally generated and energy stable heartbeat sound wave, and uses the internal sound source as an alternative excitation source to maintain the continuous generation of the perturbation time series, once the environmental noise energy recovers, the heartbeat mode automatically stops and switches back to the passive monitoring mode; in addition, the system also evaluates the quality of the environmental noise signal itself as the excitation source, the system calculates a normalized spectral entropy index representing the spectral complexity of the reference sensor signal in parallel with the system for calculating the perturbation time series, an excitation source with a flat and broad spectrum has a higher spectral entropy, otherwise, an excitation source with a spectrum concentrated in a few narrow peaks has a lower quality, when making a diagnostic decision, the system will dynamically adjust the weight of the currently calculated perturbation in the time series analysis according to the spectral entropy index, set a higher weight for the perturbation generated by a higher quality excitation source, and vice versa; finally, a perturbation time series D(t) corrected by the above self-checking procedures and adaptive adjustment mechanisms is input into a deep learning model, the model is an autoencoder model, specifically, the autoencoder model is a fully connected neural network containing an input layer, two encoding layers, a bottleneck layer, two decoding layers and an output layer, using ReLU activation function and Adam optimizer, trained to minimize mean square error (MSE); the model has established a baseline model capable of reconstructing such normal sequences and a statistical distribution of reconstruction errors generated by the model when processing healthy sequences by learning the perturbation time series in the healthy state during the calibration stage; in the monitoring stage, the autoencoder model continuously reconstructs the input real-time D(t) sequence and calculates its reconstruction error, when the D(t) sequence presents an irreversible, monotonically increasing abnormal evolution pattern due to structural damage accumulation, its morphology will deviate from the healthy pattern learned by the model, leading to an increase in reconstruction error, when the reconstruction error exceeds the dynamic baseline determined by the statistical distribution of the healthy state, such as the average value plus three times the standard deviation, the model judges that the time series has an abnormal evolution pattern, and combines the qualitative conclusions about the physical properties of the damage output by the active interrogation procedure and the state information about the sensors and coupling output by the self-checking procedure to jointly generate and output the final diagnostic information about the damage state of the metal structure node.

[0034] Embodiment 1: This embodiment aims to illustrate in detail how the method claimed in the present invention works in a systematic way to solve the limitations inherent in the prior art methods in the application level, in combination with a specific challenging engineering application scenario; the scenario is set as long-term health monitoring of a critical steel box girder web butt weld joint of a sea-crossing bridge, which not only bears huge temperature stress caused by day and night temperature difference, but also bears random impact load caused by tens of thousands of heavy trucks passing through during daily peak hours, while in the midnight to early morning period, it enters a quasi-quiet working condition with extremely weak environmental excitation, the complexity and extremity of this working condition put strict requirements on the sensitivity and all-weather adaptability of the monitoring method; in the initial stage of this monitoring task, a high-sensitivity acoustic emission monitoring system based on traditional event monitoring method was deployed, but during the three-month trial operation, the system exposed two insurmountable application obstacles; first, during the daytime traffic peak period, the energy of the structural vibration caused by vehicles passing through and various non-structural friction noises inside the steel box girder is much higher than the weak acoustic emission signals that early fatigue micro-crack propagation may release, in order to avoid mass noise signals triggering false alarms, the monitoring threshold of the system must be set at a higher level, but this directly leads to the loss of the system's detection ability for real weak damage signals; second, in the quasi-quiet working condition from two o'clock to four o'clock in the morning, the entire monitoring system is completely ineffective due to the lack of sufficient external excitation, and it cannot effectively monitor any crack propagation that may occur during this period due to residual stress or temperature stress changes, thus forming a blank area of monitoring information for several hours; to cope with the above challenges, the method claimed in the present invention was adopted for monitoring deployment in the subsequent stage; after completing the sensor layout and connection, the system performed a 48-hour calibration process, using the complete traffic and environmental noise samples in this period, it calculated and stored a set of baseline acoustic transfer function fingerprints {H_1 (f), H_2 (f),...} representing the weld joint in a healthy state; after switching to long-term monitoring, the system calculates the real-time acoustic transfer function fingerprints at a frequency of once per minute, and compares them with the baseline, thus generating a continuous perturbation time series D(t); on the 45th day of monitoring, the data showed that the D(t) sequence began to deviate from its initial baseline level of 0.05 fluctuating randomly above and below, showing a slow but continuous and irreversible growth trend, by the 90th day, its value had stabilized at a level of 0.82, while the traditional event monitoring system running in parallel did not record any high-energy acoustic emission events exceeding the alarm threshold.

[0035] The explicit irreversible evolution trend presented by the D(t) sequence indicates that the material physical properties inside the node are undergoing continuous changes. When the growth slope of the sequence exceeds the degradation rate threshold set according to the bridge design safety level for 20 consecutive calculation periods, the system automatically triggers an active interrogation procedure. One sensor in the monitoring network is temporarily switched to the sound wave emission mode to emit a standardized interrogation sound wave to the node area. After the spectrum analysis of the response sound wave received by the remaining sensors, the nonlinear response index R_nl (t) calculated is 0.35, which exceeds the nonlinear damage threshold of 0.1 set according to the material properties of the Q345qDNH weather-resistant bridge steel used in the bridge. This result provides direct physical evidence for the judgment that the growth trend of the D(t) sequence is caused by a damage source with nonlinear physical characteristics, i.e., a suspected micro-crack. It should be noted that during the multiple night quasi-quiet periods in the entire monitoring period, the system automatically activates the autonomous heartbeat mode to maintain the continuous generation of the D(t) sequence by emitting internally generated weak heartbeat sound waves as an alternative excitation source, ensuring that the monitoring process of the growth trend is never interrupted. Based on the above diagnostic information containing the state continuous deterioration trend and the damage nonlinear physical property confirmation, the operation and maintenance team conducts an ultrasonic flaw detection review on the node, and finally discovers a surface micro-crack with a length of about 3mm at the root of the weld. The results of this embodiment show that the core value of the method claimed in the present application is not just the simple combination of various technologies, but the construction of a state monitoring driven diagnostic work mode. It uses perturbation analysis of acoustic transfer function fingerprints to shift the focus of monitoring from finding rare signals in the sea to observing the continuous changes of the entire sea, i.e., the structure medium itself, thereby avoiding the limitation of insufficient signal-to-noise ratio in the traditional way. By combining this continuous state perception ability with a discrete diagnostic tool, i.e., the active interrogation procedure, which is intelligently triggered by it to identify physical properties, a logically self-consistent diagnostic closed loop from discovering abnormalities to confirming properties is formed. Finally, this change in work mode makes it possible to continuously quantify the tracking of the structure health state during the long period between two high-energy damage events.

[0036] Example 2: This example is to verify the data of the ability of the monitoring system using the method of the present application to continuously monitor the early fatigue crack propagation process of the node of the metal structure under the condition of controlled noise, through an accelerated fatigue test carried out in a laboratory environment, and to compare it with the traditional acoustic emission event monitoring method; the test is carried out on an MTS electro-hydraulic servo fatigue testing machine platform, which can apply controlled cyclic load; the test object is a metal test piece, for example a Q345B steel plate test piece with a size of 300mm x 100mm x 10mm, the middle part of which is welded by V groove butt welding process, and at the weld toe on one side of the weld, a 0.5mm deep initial notch is pre-fabricated by electric spark machining as the crack initiation point of the fatigue crack; during the test, the test piece is subjected to axial tensile sinusoidal load, the load frequency is set to 10Hz, which is designed to balance the test efficiency and avoid the heat effect that may be caused by high frequency loading, the maximum nominal stress is set to 180MPa, and the stress ratio R is 0.1; in order to simulate the background noise interference in the engineering environment, an industrial sound source is arranged 1 meter away from the test piece, which plays broadband white noise, and the noise level measured on the surface of the test piece by the sound level meter is maintained at 75dB; the initial state of the test piece is defined as no load applied, and its physical characteristics are no macroscopic cracks; the temperature of the whole test environment is maintained at 25°C±2°C; the monitoring system using the method of the present application (hereinafter referred to as the present application group) and a set of standard acoustic emission event monitoring system (hereinafter referred to as the control group) are synchronously deployed in the weld node area of the test piece, wherein the event discrimination threshold of the control group is set to 45dB according to the background noise level, while the present application group first completes the establishment of the reference acoustic transfer function fingerprint of the test piece in the silent state without load.

[0037] After the start of the test, the fatigue testing machine is started and cyclic loading is applied, and the noise source is turned on; the data acquisition system of the sample group and the control group starts to work synchronously, wherein the sample group calculates and records the value of the perturbation degree D(t) continuously with a period of 60 seconds, and the control group records the cumulative number of acoustic emission events with an amplitude exceeding 45 dB; in order to obtain the true physical size of crack propagation, the test is paused every 50,000 cycles, and a mobile metallographic microscope equipped with a micrometer eyepiece is used to measure the crack length at the tip of the pre-cut notch, with a measurement accuracy of 0.01 mm; the measurement result is used as an objective reference to evaluate the two monitoring methods; during the entire test process, the operating parameters of each system remain unchanged until the crack propagates to the predetermined length, and the test is terminated; within the initial 100,000 cycles of fatigue loading, neither monitoring method shows a significant response, at this time the crack length observed by the microscope is only 0.55 mm, which has only a slight expansion compared to the initial notch, and the value of D(t) calculated by the sample group fluctuates around the baseline level of 0.04, while the cumulative event number of the control group is 0; when the loading proceeds to 250,000 cycles, the crack length increases to 1.21 mm, the value of D(t) of the sample group shows a sustained upward trend and reaches 0.28, while the control group only records 3 instantaneous acoustic emission events during this period; as the loading cycle continues, at the loading of 500,000 cycles, the actual crack length is 3.85 mm, the value of D(t) of the sample group has increased to 0.95, and its growth curve shows a positive correlation with the crack length expansion curve, while the cumulative event number of the control group is only 17, and the data points are sparsely distributed on the time axis, not forming a continuous sequence that can be used for trend judgment; see Table 1 for specific test data records.

[0038] Table 1: Comparison of monitoring data of the sample group and the control group at different fatigue stages.

[0039]

[0040] The test data show that the value of the perturbation degree D(t) generated by the monitoring system using the method of the present application shows a direct positive correlation with the cumulative damage length of the internal micro-cracks of the metal material; under the same noise interference conditions, compared with the traditional acoustic emission monitoring method which relies on high-energy damage events, the method of the present application provides a quantitative indicator that can continuously track the gradual change of the physical properties of the material.

[0041] In order to further verify the method claimed in the present application, compared with the traditional technical solutions mentioned in the background art, the substantial differences and significant beneficial effects are as follows.

[0042] Comparative Example 1: This comparative example aims to simulate the actual process and results of health monitoring of the same key steel box girder web butt weld joint of the cross-sea bridge using the conventional event-driven acoustic emission monitoring technology disclosed in the background section under the same engineering background and hardware conditions as Example 1; the monitoring system used in this comparative example is strictly consistent with the configuration described in Example 1 in terms of sensor type, number, layout position, and specification parameters of data acquisition hardware. The only difference between this comparative example and Example 1 is that the data processing and diagnosis logic of this comparative example does not use the method claimed in the present application based on acoustic transfer function fingerprint perturbation analysis, but uses the traditional acoustic emission event threshold monitoring method; the specific operation is as follows: the data processing unit of the monitoring system performs real-time energy amplitude analysis on the acoustic signals collected by each monitoring sensor, and only when the signal amplitude exceeds a pre-set alarm threshold, it is recorded as a valid acoustic emission event and the count is accumulated.

[0043] In the early stage of system deployment, in order to verify the feasibility of this traditional technology under real working conditions, a 24-hour field noise base test was first conducted. The test results show that during the daytime traffic peak period, the structural vibration noise caused by heavy vehicles passing through has a signal amplitude generally between 40dB and 50dB, with an instantaneous peak value of up to 55dB. To avoid the continuous triggering of false alarms by massive traffic noise, which would cause the system to malfunction, a 5dB margin was reserved based on the measured noise peak value, and the acoustic emission event alarm threshold was finally set to 60dB. This setting aims to ensure that only signals significantly higher in energy than general background noise are counted as damage events. After completing the threshold setting, the traditional monitoring system was put into continuous operation for 90 days. During this period, the operation and maintenance team tracked the physical state of the node through other non-destructive testing methods such as regular ultrasonic flaw detection. The specific monitoring results and the corresponding relationship with the physical state are shown in Table 2.

[0044] Table 2: Comparison of traditional acoustic emission monitoring method and actual state of node in Comparative Example 1.

[0045]

[0046] The test results show that the acoustic emission event monitoring method based on fixed threshold has inherent technical limitations in the task of early damage monitoring of structures in high noise background. Since the alarm threshold must be set at a high level (60 dB) sufficient to avoid environmental noise, the system has no detection capability for the acoustic emission signals released by the slow expansion of fatigue micro-cracks, which are low-energy signals. As shown in Table 2, during the 90-day monitoring period, although the physical damage (crack length) of the node has continuously expanded from 0.51 mm to 3.15 mm, the traditional monitoring system only records 2 discrete transient events without forming any regular pattern, and the output diagnostic conclusion is always that the node state is normal. This result confirms that this monitoring paradigm relying on high-energy discrete events has a huge information gap in the gradual evolution of damage, and cannot realize continuous tracking and early warning of the health state of the structure. The fundamental reason for the monitoring failure is that the technical principle itself cannot effectively separate the weak damage signal from the high-energy background noise.

[0047] Embodiment 3: This embodiment combines Figs. 1 to 3 the acoustic emission deep learning real-time diagnosis method for metal structure node damage, as shown in Fig. 1 , the overall diagnosis process starts with the environmental background sound as a passive excitation source. First, the baseline acoustic transfer function fingerprint is established, that is, the acoustic transfer function of the healthy state is calculated and stored using environmental noise. Then, the core part of continuous monitoring and comparison is entered. This part periodically acquires real-time acoustic transfer function fingerprints and compares them with the baseline to generate a perturbation degree time series. In this part, a parallel sensor network self-checking mechanism based on the acoustic reciprocity principle periodically diagnoses the sensors and coupling state to ensure the reliability of the monitoring premise. An autonomous heartbeat mode is set to automatically activate when the environmental noise is below the silence threshold, and the internally generated heartbeat sound is emitted as an alternative excitation source. In continuous monitoring, the system analyzes the trend of the perturbation degree sequence and makes a decision on whether it is continuously increasing and the growth rate exceeds the threshold. If the answer is no, it returns to the continuous monitoring and comparison part. If the answer is yes, it triggers an active sound wave interrogation. The interrogation emits an interrogation sound wave, analyzes the harmonics to identify the physical properties of the damage, and inputs the obtained physical property evidence information into the deep learning diagnosis module together with the perturbation degree sequence. The module uses an autoencoder model to analyze the abnormal evolution pattern of the perturbation degree sequence to determine the damage. Finally, the system outputs the diagnostic information and forms the final conclusion by integrating the state evolution, damage properties, and system self-checking results.

[0048] As Fig. 2As shown, the horizontal axis represents five different types of excitation sources, from left to right: narrowband vibration 50Hz, mixed noise low quality, mixed noise medium quality, broadband wind load high quality, and white noise ideal. The left vertical axis represents the normalized spectral entropy H_s(t), with a value range of 0 to 1.0, and the right vertical axis represents the diagnostic weight W(t), also with a value range of 0 to 1.0. The figure clearly shows that as the excitation source quality changes from low to high, the corresponding spectral entropy H_s(t), represented by a grid-filled bar, and the diagnostic weight W(t), represented by a diagonally filled bar, both exhibit a monotonically increasing relationship. That is, the more complex and higher the quality of the excitation source spectrum, the larger its spectral entropy H_s(t) value, and the higher the diagnostic weight W(t) assigned in the final diagnosis.

[0049] like Fig. 3 As shown, after system startup, it first enters the core passive monitoring state. In this state, the system continuously generates a micro-perturbation time series using environmental noise and analyzes its evolution trend. Depending on different triggering conditions, the system can switch between multiple states: when the environmental noise is below the silence threshold, the system switches from passive monitoring to autonomous heartbeat state, emitting internal heartbeat sound waves to replace the excitation source, thereby maintaining the continuity of monitoring data. When the environmental noise recovers to above the silence threshold, the system automatically returns to passive monitoring state. When the periodic self-calibration time is reached, the system enters the system self-calibration state, diagnosing the sensor itself and its coupling state based on the acoustic reciprocity principle, and assessing the reliability of the data source. After self-verification, the system returns to passive monitoring. When the perturbation continuously increases and its slope exceeds the degradation threshold in passive monitoring, the system switches to active interrogation. In this state, the sensor's operating mode is temporarily switched, interrogation sound waves are emitted, and the response sound waves are analyzed to identify the nature of the damage. If the analysis results show that the nonlinear response is lower than the damage threshold, it is considered that the physical evidence is insufficient, and the system returns to passive monitoring. If the nonlinear response is higher than the damage threshold, the system enters damage warning state. In this state, an abnormal evolution pattern is confirmed, diagnostic information containing physical properties is output, and an alarm is triggered to notify the operation and maintenance personnel. The system only returns to passive monitoring after receiving manual confirmation and a reset command.

[0050] Example 4: Before the monitoring system using the method of the present application is applied to the key cast steel joint of a large metal structure stadium, a standardized off-line calibration and construction procedure is required to eliminate the uncertainty of the key parameter setting and the model training process; this procedure aims to ensure that all the judgments of the system are based on the physical characteristics of the monitored object itself, and the initial state is defined as obtaining a metal material sample of the same batch, same brand and same welding process as the monitored joint, and the physical characteristics of the sample are confirmed by non-destructive testing to be free of initial defects. A laboratory test platform with the ability to prepare and measure small cracks with high precision is prepared, and the platform is required to have a fatigue load of not less than 200 kN and a mobile microscope with a measurement accuracy of 0.01 mm. To complete the off-line calibration of the nonlinear damage threshold, the first step is to generate a surface crack with a length of 0.5 mm on the sample by fatigue precasting, and confirm it by microscope. Then, on the sample with known micro-cracks, perform active interrogation operation, i.e. use the sensor to emit standardized interrogation sound waves, and receive response sound waves by another sensor, calculate the value of nonlinear response index R_nl (t) by spectral analysis of the response sound waves, repeat the operation 10 times and take the average value, and obtain a stable response value of 0.42 corresponding to a crack length of 0.5 mm. To balance the monitoring sensitivity and anti-interference, the nonlinear damage threshold of the monitoring system is set to 50% of the measured average value, i.e. 0.21.

[0051] The calibration procedure of the deterioration rate threshold is performed, which aims to establish a quantitative correlation between the expansion rate of physical damage and the disturbance D(t) growth rate. For this purpose, a healthy metal material sample of the same batch without initial defects is selected and installed on the fatigue testing machine, and the monitoring system is deployed simultaneously. During the application of a constant amplitude accelerated fatigue load on the sample, the system records the value of D(t) continuously at a cycle of 60 seconds, and the testing machine pauses loading every 10,000 cycles and measures the actual expansion length of the crack by microscope. By processing the data of the entire test process, a corresponding relationship curve between the physical crack expansion rate, i.e. the crack length increment da / dN per cycle, and the disturbance growth rate dD / dN is obtained. According to the design safety specification of the stadium structure, the maximum tolerable crack expansion rate allowed is mm / cycle, and by looking up on the corresponding relationship curve, the disturbance growth rate corresponding to this physical rate is / minute, therefore, the degradation rate threshold for triggering active interrogation is set to this value; finally, the training and construction procedure of autoencoder deep learning model for final diagnosis is performed; the input data of this procedure comes from the monitoring system which has been installed on the actual nodes of the venue and has completed the establishment of the reference acoustic transfer function fingerprint; the system continuously collects 72 hours of environmental noise data in the empty state of the venue after its completion and before its formal operation, as well as in different construction disturbance conditions such as different regional seat installation, ceiling opening and closing test, and generates a large-scale health state perturbation time series D(t) sample set covering various normal conditions; this sample set is used as the training data of the autoencoder model, and the training target of the model is set to minimize the reconstruction error of these health sequences; after the training is completed, the same set of health sample set is input into the trained model again, and the reconstruction error of each time series sample is calculated, thereby obtaining a reconstruction error statistical distribution representing the health state; finally, the decision threshold of the model for judging abnormalities is set to the sum of the mean and three times the standard deviation of the statistical distribution; by performing the above process, all key judgment thresholds and data models in the monitoring system are set based on the physical and statistical characteristics of the monitored object.

[0052] Example 5: After a set of monitoring system using the method of the present application has been continuously served on a key welded node of a large port crane for two years, the system simultaneously monitors two phenomena, one is that the perturbation time series D(t) corresponding to a certain specific monitoring sensor channel starts to deviate from the stable baseline and shows continuous growth, and the other is that a large hydraulic pump near the node starts to produce high-frequency periodic vibration due to wear, resulting in a large number of transient high-energy acoustic emission events captured by the monitoring system; in this condition, it is necessary to judge the source of the growth of D(t) sequence, i.e. whether it is caused by real structural damage.

[0053] To cope with this situation, the system performs an online evaluation procedure of the sensor coupling state based on residual signal analysis in parallel; in the process of periodically calculating the acoustic transfer function fingerprints of each channel, the processor synchronously acquires the residual signal in the monitoring signal that cannot be explained by the reference sensor signal; a well-coupled sensor has a low-energy random noise residual signal, while a sensor that has experienced physical coupling degradation will rub against the mounting base and resonate, and these non-structure-propagating signals will appear as an increase in energy and a change in spectral kurtosis in the residual signal; the system generates an index representing the coupling quality between the sensor and the structure node by calculating the residual signal energy and spectral kurtosis of the abnormal channel, and the calculation result shows that the index has exceeded the preset threshold, indicating that the physical coupling state between the sensor and the structure has changed; the system also performs a time series causal correlation analysis procedure, which acquires a continuous state sequence composed of perturbation time series and a discrete event sequence composed of monitored transient high-energy acoustic emission events in parallel; based on the transfer entropy algorithm, the system calculates the information flow from the discrete event sequence to the continuous state sequence to generate an index representing the time series causal correlation between the two; the calculation result shows that the index is close to zero, indicating that there is no causal correlation between the growth of the D(t) sequence and the high-energy events generated by the hydraulic pump; by integrating the outputs of the above two parallel procedures, the system classifies the abnormality as a sensor coupling state degradation accompanied by external environmental interference, suppresses the structure damage alarm, and issues a maintenance instruction to the operator about the physical inspection of the specified sensor.

[0054] In order to eliminate the influence of the change in the spectral characteristics of the environmental excitation source on the calculation result of the perturbation time series, before the monitoring system using the method of the present application is deployed in front of a high-rise metal structure tower in an industrial area subjected to intermittent narrowband mechanical vibration and broadband wind load, a standardized pre-calibration procedure is performed to establish a quantitative relationship between the quality of the excitation source and the weight of the diagnostic information; this procedure provides an engineering calibration method for the diagnostic weight dynamic adjustment logic based on the spectral entropy of the reference signal in the specific implementation; in order to quantitatively evaluate the quality of the environmental background noise as the excitation source, the procedure first introduces the calculation step of the normalized spectral entropy index , which quantifies the complexity of the signal spectrum using information theory; the specific operation is as follows: the system processes the environmental background noise signal collected by the reference sensor in each monitoring period; first, calculate its power spectral density P(f); second, normalize the power spectral density in the effective frequency band so that its integral is 1, and obtain the probability density function p(f); third, calculate the entropy value of the probability density function according to the following formula: In the formula, p(f) is the normalized power spectral density of the reference sensor signal; a broadband noise with a flat spectrum has a uniform p(f) distribution, and the calculated H_s value is close to 1, while a narrowband noise with energy concentrated in a few frequencies has a concentrated p(f) distribution, and the calculated H_s value is close to 1. The value is close to 0.

[0055] After defining the quantitative method for the quality of the excitation source, the next step in the procedure is to establish the quality index H_s(t) and diagnostic weights. The mapping relationship between them is established; this process is completed through an offline learning and fitting step; the system first collects long-term reference signals of the tower under healthy conditions and different excitation conditions, including the case of only broadband wind load and the case of only 50Hz narrowband vibration generated by nearby equipment, and calculates the statistical distribution of H_s(t) corresponding to these two cases respectively. The results show that H_s(t) under broadband wind load is mainly distributed around 0.85, while H_s(t) under narrowband vibration is mainly distributed around 0.85. The values ​​are mainly distributed around 0.20; based on this, the following piecewise linear weighting function is established. When H_s(t) is below 0.3, W(t) is set to 0.1 to reduce the impact of perturbation data generated by this type of excitation source on the final diagnostic results; when When the value is higher than 0.8, W(t) is set to 1.0; when H_s(t) is between 0.3 and 0.8, W(t) is determined by linear interpolation; after the system enters long-term monitoring, for the perturbation D(t) generated in each calculation cycle, the processor will synchronously calculate the H_s(t) of the excitation source for the corresponding time period, and use the weight function to calculate the corresponding weight W(t). Finally, the time series that has been dynamically adjusted by the weight is input into the deep learning model for trend analysis, thereby reducing the fluctuation of diagnostic results caused by changes in the characteristics of the excitation source.

[0056] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for real-time diagnosis of acoustic emission of a metal structure node damage using deep learning, characterized in that, The method comprises the following steps: Step S1, based on the environmental background noise collected by the reference sensor and the monitoring sensor, a baseline acoustic transfer function fingerprint representing the health state of the metal structure node is calculated and stored; Step S2, real-time acoustic transfer function fingerprints are periodically obtained, and the perturbation degree between the real-time acoustic transfer function fingerprint and the baseline acoustic transfer function fingerprint is calculated, thereby generating a perturbation time series; Step S3, it is judged whether the value of the perturbation time series continuously increases in a predetermined number of calculation periods and the growth rate exceeds a degradation rate threshold; Step S4, if the judgment of step S3 is yes, an active challenge is triggered, one of the monitoring sensors is switched to a sound wave emission mode to emit a challenge sound wave, and the remaining monitoring sensors receive a response sound wave, the physical properties of the damage are determined by analyzing the harmonic nonlinear acoustic characteristics generated in the response sound wave; Step S5, at least one pair of monitoring sensors is periodically selected as a sound wave emitter and a receiver, respectively, based on the acoustic reciprocity principle, the forward and reverse acoustic transfer functions are compared to diagnose the coupling state of the sensor itself and the metal structure node; Step S6, the perturbation time series is input into a deep learning model, and according to the physical properties of the damage and the state of the sensor and the coupling, diagnostic information about the damage state of the metal structure node is output; Wherein the harmonic nonlinear acoustic characteristics generated in the response sound wave are analyzed by calculating a nonlinear response index R_nl (t), wherein: R_nl (t)=E_harmonic / E_fundamental, wherein E_fundamental is the energy of the response sound wave at the reference frequency of the challenge sound wave, and E_harmonic is the energy of the response sound wave at the newly generated harmonic frequency; and the nonlinear response index R_nl (t) is compared with a nonlinear damage threshold preset according to the fracture mechanics properties of the metal structure material; And the perturbation is determined by calculating the integral of the difference between the logarithmic amplitude of the real-time acoustic transfer function fingerprint and the logarithmic amplitude of the baseline acoustic transfer function fingerprint in the entire frequency band; And the degradation rate threshold is preset according to the design safety level of the metal structure node and the material fatigue properties; And the method further establishes and implements the following rules: a continuous state sequence composed of the perturbation time series and a discrete event sequence composed of the monitored transient high-energy acoustic emission events are obtained in parallel; the transfer entropy algorithm is used to calculate the information flow from the discrete event sequence to the continuous state sequence to generate an index representing the time sequence causal correlation degree of the two; and according to the index, external pseudo-source storm interference working conditions are identified and false alarms are suppressed; And the deep learning model is a self-encoder model, which determines whether there is an abnormal evolution mode in the time series by identifying whether the reconstruction error of the perturbation time series exceeds a dynamic baseline determined by the reconstruction error statistical distribution generated by the model when processing the perturbation time series in the healthy state, thereby outputting the diagnostic information.

2. The acoustic emission deep learning real-time diagnosis method for damage of a metal structure node according to claim 1, characterized in that, The method further establishes and implements the following rules: continuously monitor the energy level of the ambient background noise signal, and automatically activate the autonomous heartbeat mode when the energy level is lower than a silence threshold determined based on the average energy level of a reference acoustic transfer function fingerprint calculation period; The autonomous heartbeat mode includes: controlling one of the monitoring sensors to emit an internally generated energy-stable heartbeat sound wave, and using the heartbeat sound wave as an alternative excitation source.

3. The method of claim 1, wherein the method is characterized by: The method further establishes and implements the following rules: during the calculation of the acoustic transfer function fingerprint, from the signals collected by the monitoring sensors, eliminate the components coherent with the signals collected by the reference sensors to obtain residual signals; analyze the energy and spectral kurtosis of the residual signals to generate an index representing the coupling quality between the monitoring sensors and the metal structure node; and when outputting the diagnostic information, adjust the weight of the monitoring data according to the coupling quality index.

4. The acoustic emission deep learning real-time diagnosis method for damage of a metal structure node according to claim 1, characterized in that, The method further establishes and implements the following rules: when generating the perturbation time series, calculate the power spectrum of the ambient background noise signal collected by the reference sensor, and generate a spectral entropy index representing the quality of the signal as an excitation source based on the power spectrum; and dynamically adjust the weight of the perturbation time series when outputting the final diagnostic information according to the spectral entropy index.

5. The method of claim 1, wherein the method is characterized by: The rules for diagnosing the sensor itself and its coupling state with the metal structure node also include: when the difference between the forward and reverse acoustic transfer functions exceeds a reciprocity error threshold set according to the sensor factory calibration data, it is determined that the corresponding sensor or its coupling state has failed, and when outputting the diagnostic information, the data weight from the faulty sensor is automatically reduced or it is logically isolated.

Images