Acoustic emission non-destructive testing method based on acoustic emission event clustering

Abstract

The present application relates to the technical field of intelligent signal processing, and more particularly to an acoustic emission nondestructive testing method based on acoustic emission event clustering, comprising the following steps: step 1, attaching a sensor array to the outer wall of a water treatment reaction vessel, performing arrival consistency verification on candidate events to eliminate interference events, and intercepting the event waveform segment of the effective acoustic emission event; step 2, transmitting the event waveform segment to a scatter transform application specific integrated circuit (ASIC) through a field programmable gate array (FPGA), and outputting an acoustic emission event feature vector; and step 3, splicing the acoustic emission event feature vector and the acoustic emission source position coordinates into a joint description vector after normalization in a sliding time window through an edge processor. The present application realizes full-link online processing from sensing acquisition to damage early warning, effectively improving the positioning accuracy, feature stability and early warning timeliness of micro-crack monitoring of the water treatment reaction vessel.

Description

Technical Field

[0001] This invention belongs to the field of intelligent signal processing technology, specifically relating to a non-destructive testing method for acoustic emission based on acoustic emission event clustering. Background Technology

[0002] Water treatment reaction vessels endure the combined effects of internal media corrosion, pressure cycling, and temperature fluctuations during long-term service. This inevitably leads to the initiation and propagation of microcracks on the vessel walls. Failure to detect and address these cracks in time can result in leaks or even serious safety incidents such as instability and fracture. Acoustic emission detection technology, by capturing the transient elastic waves released during the initiation and propagation of microcracks, enables online monitoring of structural damage. It offers advantages such as real-time performance and the ability to perform detection while the equipment is running, and has been widely applied in pressure vessels, bridges, and storage tanks.

[0003] Existing acoustic emission detection systems typically employ discrete piezoelectric ceramic sensors for signal acquisition. These sensors rely on manual arrangement and coupling, and the sensitivity consistency and frequency response uniformity of each sensor are limited by the installation process. Synchronization accuracy between channels is also difficult to guarantee, directly impacting the source localization accuracy based on time difference of arrival (TDOA). In signal processing, existing solutions often employ traditional short-time energy threshold triggering and TDOA-based geometric localization methods, treating the propagation velocity of elastic waves in the container wall as a fixed constant. This neglects the dispersion effect of guided wave propagation in the shell structure, leading to deviations in delay compensation for different frequency components. Localization accuracy decreases significantly with increasing propagation distance. Regarding feature extraction, existing solutions mainly rely on time-domain parameters such as ring count, duration, peak amplitude, and rise time. These parameters are sensitive to minute shifts in the input waveform and differences in propagation paths. Acoustic emission events of the same damage type may produce significantly different feature values ​​under different propagation conditions, resulting in insufficient stability for subsequent classification and clustering. In event clustering, existing solutions often employ methods that require pre-specifying the number of clusters. However, in actual monitoring, the number and distribution of damaged areas are unknown beforehand, and the event density in different damaged areas often varies significantly. Clustering methods with fixed cluster numbers and density parameters struggle to accurately identify event clusters in both high-activity and low-activity areas simultaneously. Furthermore, the signal processing and feature extraction operations in existing solutions are primarily performed offline on the host computer using software, which is insufficient to meet the demands of industrial sites for real-time online monitoring and immediate early warning. Summary of the Invention

[0004] The main objective of this invention is to provide a non-destructive testing method for acoustic emission based on acoustic emission event clustering, which realizes online processing of the entire chain from sensing acquisition to damage early warning, effectively improving the positioning accuracy, feature stability and early warning timeliness of microcrack monitoring in water treatment reaction vessels.

[0005] To solve the above problems, the technical solution of the present invention is implemented as follows: An acoustic emission nondestructive testing method based on acoustic emission event clustering includes the following steps: Step 1: Attach the sensor array to the outer wall of the water treatment reaction vessel. Each element in the sensor array converts the acoustic emission elastic waves from the wall into electrical signals. The electrical signals of each element are digitized by the corresponding oversampling analog-to-digital converter and then sent to the field-programmable gate array (FPGA). The FPGA performs decimation filtering on the digitized signals to obtain multiple acoustic emission digital waveforms. The corresponding elastic wave propagation velocity is retrieved according to the main frequency band of the candidate events. The direction of the acoustic emission source is determined by scanning in each direction through delay-summation beamforming. The position coordinates of the acoustic emission source are determined according to the direction of the acoustic emission source and the propagation distance. A arrival consistency check is performed on the candidate events to eliminate interfering events. The event waveform segments of the valid acoustic emission events are extracted. Step 2: The field-programmable gate array transmits the event waveform segment to the scattering transformation ASIC. The scattering transformation ASIC performs multi-scale convolution, modulus taking, and pooling operations on the event waveform segment sequentially through a fixed multi-level filter channel, and outputs the acoustic emission event feature vector. Step 3: Within the sliding time window, the edge processor normalizes the acoustic emission event feature vector and the acoustic emission source location coordinates and then concatenates them into a joint description vector. Hierarchical density clustering is performed on the joint description vector to divide it into effective acoustic emission event clusters. The event rate changes of each effective acoustic emission event cluster are tracked. When the event rate continues to increase, a damage warning signal is output to the online monitoring system of the water treatment reaction vessel.

[0006] Furthermore, the sensor array in step 1 is a piezoelectric micromechanical ultrasonic transducer array chip, which contains multiple array elements arranged in a row and column matrix. It is attached to the approximate planar monitoring wall area on the outer wall of the water treatment reaction vessel with a local curvature lower than a preset curvature threshold through a coupling layer.

[0007] Furthermore, the oversampling analog-to-digital converter in step 1 is a continuous-time Sigma-Delta analog-to-digital converter, which performs oversampling analog-to-digital conversion at a rate not less than 64 times the upper limit of the signal bandwidth to output the oversampling digital bitstream; the decimation filtering in step 1 includes performing cascaded integral comb filtering and half-band filtering decimation on each oversampling digital bitstream.

[0008] Furthermore, step 1, which involves retrieving the corresponding elastic wave propagation velocity based on the main frequency band of the candidate event, includes: selecting one channel corresponding to a preset reference array element from the multi-channel acoustic emission digital waveforms as the reference channel waveform; calculating the short-time energy envelope for each sampling point of the reference channel waveform; marking the candidate event trigger time when the short-time energy envelope first exceeds the initial trigger threshold; extending the candidate event analysis window forward and backward by a preset half-window length centered on the candidate event trigger time; performing a short-time Fourier transform on the candidate event analysis window to determine the main frequency band; and retrieving the elastic wave propagation velocity matching the main frequency band from a pre-calibrated frequency band-propagation velocity correspondence table.

[0009] Furthermore, the delay-summation beamforming scanning in step 1 includes: using the local tangent plane at the sensor array attachment position on the outer wall of the water treatment reaction vessel as the positioning reference plane, and establishing a 2D planar coordinate system in the positioning reference plane with the projection point of the geometric center of the sensor array on the positioning reference plane as the origin; scanning direction by direction according to a preset angle step value; for the current scanning direction, determining the 2D planar offset of each array element relative to the origin in the positioning reference plane based on the row and column numbers of each array element in the row and column matrix; determining the integer sampling point delay of each acoustic emission digital waveform based on the projection length of the 2D planar offset in the current scanning direction and the obtained elastic wave propagation speed; supplementing the end of each acoustic emission digital waveform with zero-value sampling points of the same length as the maximum delay, and aligning the delay lines according to the corresponding integer sampling point delay; superimposing the aligned acoustic emission digital waveforms point by point to obtain the synthesized waveform in the current scanning direction; and squaring the synthesized waveform point by point and accumulating the sum to obtain the focused energy value in the current scanning direction.

[0010] Furthermore, in step 1, determining the position coordinates of the acoustic emission source includes: after traversing all scanning directions, selecting the scanning direction with the largest focused energy value that exceeds the adaptive energy threshold as the pointing direction of the acoustic emission source; calculating the propagation distance based on the peak time of the focused energy in the synthetic waveform corresponding to the obtained elastic wave propagation speed and the pointing direction of the acoustic emission source; and determining the position coordinates of the acoustic emission source in a 2D plane coordinate system along the pointing direction of the acoustic emission source with the origin as the starting point and according to the propagation distance.

[0011] Furthermore, the arrival consistency verification in step 1 includes: recording the first sampling point in the reference channel waveform that exceeds the trigger level as the arrival timestamp; calculating the measured channel time difference between the first sampling point in the acoustic emission digital waveform of each of the remaining array elements that exceeds the trigger level and the arrival timestamp; calculating the theoretical channel time difference of each array element according to the direction of the acoustic emission source and the obtained elastic wave propagation speed; and determining a valid acoustic emission event when the number of array elements whose difference between the measured channel time difference and the theoretical channel time difference is within the preset arrival consistency tolerance is not less than the preset number of valid array elements. The event waveform segment of the valid acoustic emission event in step 1 includes: taking the time determined by the preset number of pre-trigger sampling points backward from the arrival timestamp as the starting point of the segmentation, and continuously extracting sampling points of a preset total extraction length from the reference channel waveform as the event waveform segment.

[0012] Furthermore, in step 2, a two-stage cascaded processing pipeline is embedded within the dedicated integrated circuit for scattering transformation. The first stage pipeline sets up multiple parallel pairs of orthogonal filter channels. Each pair of orthogonal filter channels includes one real-part finite impulse response (FIR) filter and one imaginary-part FIR filter. The coefficients of the real-part FIR filter are pre-programmed as discrete sampled values ​​of the real part of the Morlet wavelet corresponding to the center frequency, and the coefficients of the imaginary-part FIR filter are pre-programmed as discrete sampled values ​​of the imaginary part of the Morlet wavelet at the same center frequency. The event waveform segment simultaneously enters all pairs of orthogonal filter channels and performs convolution operations on each. The point-by-point squared value of the real-part convolution output of the same channel pair is then compared with the imaginary-part convolution output. The modulus sequence is obtained by summing the squared values ​​point by point and taking the arithmetic square root point by point. The modulus sequence is then subjected to fixed-window-length sliding mean pooling to obtain the envelope sequence of each channel in the first stage. The second-stage pipeline feeds the envelope sequence of each channel in the first stage into the second group of orthogonal filter channel pairs to perform convolution operation. The squared values ​​of the real part convolution output and the squared values ​​of the imaginary part convolution output in the same channel pair are summed and the arithmetic square root point by point to obtain the modulus sequence of each channel in the second stage. The modulus sequence of each channel in the second stage is then subjected to fixed-window-length sliding mean pooling to obtain the envelope sequence of each channel in the second stage. The envelope sequences of all channels in the first stage and all channels in the second stage are arranged in ascending order of center frequency and then concatenated to form the acoustic emission event feature vector.

[0013] Furthermore, the normalization in step 3 includes: calculating the mean and standard deviation of each component of the acoustic emission event feature vector for all valid acoustic emission events within the sliding time window; replacing the standard deviation of a certain dimension with the preset lower limit when it is lower than the preset lower limit; subtracting the corresponding mean from each component and dividing by the corresponding standard deviation to obtain the normalized feature vector; and performing the same processing on the acoustic emission source location coordinates to obtain the normalized location coordinates. The hierarchical density clustering in step 3 includes: calculating the Euclidean distance between each joint description vector and all other joint description vectors. Distances are sorted in ascending order, and the k-th Euclidean distance value in the sort is taken as the core distance, where k is the preset number of neighborhood samples. For any two joint description vectors, the maximum value among their respective core distances and the Euclidean distance between them is taken as the cross-distance. A minimum spanning tree is constructed with all joint description vectors as nodes and cross-distance as edges. Edges in the minimum spanning tree are removed sequentially from largest to smallest cross-distance. After each removal, connected components with no less than the preset minimum cluster size are marked as valid acoustic emission event clusters.

[0014] Furthermore, step 3, which tracks the event rate changes of each effective acoustic emission event cluster, includes: using the component-wise arithmetic mean of the position coordinates of all acoustic emission sources within each effective acoustic emission event cluster as the cluster center coordinates; counting the number of new events in each fixed statistical window according to the order of arrival timestamps as the current window event rate; and determining that the corresponding effective acoustic emission event cluster has entered a rapid expansion state when the current window event rate increases window by window for a preset number of consecutive fixed statistical windows. The cluster center coordinates and the rapid expansion state identifier are then output as damage warning signals to the online monitoring system of the water treatment reaction vessel.

[0015] This invention provides a non-destructive acoustic emission (NDE) testing method based on acoustic emission event clustering, which offers the following advantages: The invention replaces traditional discrete piezoelectric ceramic sensors with a piezoelectric micromechanical ultrasonic transducer array chip. Utilizing microelectromechanical systems (MEMS) technology, multiple array elements are integrated on a single chip substrate. The spacing and sensitivity of each element are precisely controlled during the manufacturing stage, fundamentally eliminating the channel inconsistencies caused by manual assembly of discrete sensors. This provides high-quality multi-channel input signals for subsequent beamforming and localization. A continuous-time Sigma-Delta analog-to-digital converter digitizes the signals of each element at a high oversampling rate. The quantization noise shaping effect brought about by oversampling significantly improves the effective signal-to-noise ratio, while its inherent anti-aliasing characteristics reduce the design requirements of the front-end analog filter. A field-programmable gate array (FPGA) performs all operations on-chip, including decimation filtering, main frequency band identification, frequency band corresponding propagation velocity lookup, and delay-summation beamforming. The propagation velocity is retrieved from a pre-calibrated correspondence table according to the main frequency band, effectively compensating for the influence of guided wave dispersion effects on the delay accuracy and improving the adaptability of source localization under different frequency band conditions. Consistency verification compares the measured channel time difference with the theoretical channel time difference, eliminating non-crack interference events such as pump vibration, valve impact, and fluid turbulence during the beamforming stage, thus reducing the risk of false alarms in subsequent clustering. The dedicated integrated circuit for scattering transformation uses orthogonal filter channels to perform complex wavelet modulus calculations. The extracted features are stable to translations and minor deformations of the input waveform; acoustic emission events of the same damage type can still obtain similar feature representations under different propagation conditions, enhancing the reliability of event clustering. Hierarchical density clustering automatically discovers event clusters with different density distributions in the normalized joint description space without pre-specifying the number of clusters, and can simultaneously capture damage information from both high-activity and low-activity regions. The continuously increasing event rate criterion utilizes the physical law that the event rate accelerates as the crack transitions from stable to rapid propagation, issuing an early warning before crack instability, providing timely online monitoring for the safe operation of water treatment reaction vessels. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the principle of delay-summation beamforming positioning on an approximately planar monitoring wall region using a piezoelectric micromechanical ultrasonic transducer array chip provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the principle of the Morlet wavelet orthogonal filter performing convolution and modulus operations on event waveform segments according to an embodiment of the present invention, comprising 6 sub-figures; Figure 3 This is a schematic diagram illustrating the process of hierarchical density clustering for progressively segmenting acoustic emission event clusters, as provided in an embodiment of the present invention. Detailed Implementation

[0017] An acoustic emission nondestructive testing method based on acoustic emission event clustering includes the following steps: Step 1: Attach the sensor array to the outer wall of the water treatment reaction vessel. Each element in the sensor array converts the acoustic emission elastic waves from the wall into electrical signals. The electrical signals of each element are digitized by the corresponding oversampling analog-to-digital converter and then sent to the field-programmable gate array (FPGA). The FPGA performs decimation filtering on the digitized signals to obtain multiple acoustic emission digital waveforms. The corresponding elastic wave propagation velocity is retrieved according to the main frequency band of the candidate events. The direction of the acoustic emission source is determined by scanning in each direction through delay-summation beamforming. The position coordinates of the acoustic emission source are determined according to the direction of the acoustic emission source and the propagation distance. A arrival consistency check is performed on the candidate events to eliminate interfering events. The event waveform segments of the valid acoustic emission events are extracted. Step 2: The field-programmable gate array transmits the event waveform segment to the scattering transformation ASIC. The scattering transformation ASIC performs multi-scale convolution, modulus taking, and pooling operations on the event waveform segment sequentially through a fixed multi-level filter channel, and outputs the acoustic emission event feature vector. Step 3: Within the sliding time window, the edge processor normalizes the acoustic emission event feature vector and the acoustic emission source location coordinates and then concatenates them into a joint description vector. Hierarchical density clustering is performed on the joint description vector to divide it into effective acoustic emission event clusters. The event rate changes of each effective acoustic emission event cluster are tracked. When the event rate continues to increase, a damage warning signal is output to the online monitoring system of the water treatment reaction vessel.

[0018] During long-term service, water treatment reaction vessels are inevitably subject to the combined effects of internal media corrosion, pressure cycling, and temperature fluctuations, leading to the initiation and propagation of microcracks on the vessel walls. Microcracks release transient elastic waves at each stage of initiation, propagation, and even instability and fracture. These elastic waves propagate in the solid medium as stress waves and can be picked up by surface-mounted piezoelectric sensors—a phenomenon known as acoustic emission. Real-time acquisition and analysis of these acoustic emission elastic waves allows for the detection of cracks and assessment of their activity levels before they develop to the point of affecting structural integrity, thus providing online early warning for the safe operation of water treatment equipment.

[0019] The sensor array is implemented using a piezoelectric micromechanical ultrasonic transducer array chip. Compared to traditional discrete piezoelectric ceramic sensors, the piezoelectric micromechanical ultrasonic transducer array chip integrates multiple array elements on a single chip substrate using microelectromechanical systems (MEMS) technology. The spacing between each array element can be controlled at the sub-millimeter level, with typical element spacing ranging from 0.5 mm to 2 mm. Furthermore, the piezoelectric sensitivity and frequency response consistency of each array element are far superior to the manually assembled solution of discrete sensors. The array elements in the chip are arranged in a row-column matrix. A typical configuration is 8 rows and 8 columns, totaling 64 elements. Alternatively, 4 rows and 4 columns, 6 rows and 6 columns, or larger row-column configurations can be selected depending on the size of the monitoring area and the required positioning resolution. The effective sensing surface diameter of each element ranges from 50 micrometers to 500 micrometers, and the frequency response bandwidth covers 50 kHz to 1.5 MHz, fully covering the typical frequency range of acoustic emission elastic waves from the walls of water treatment reaction vessels.

[0020] When attaching a piezoelectric micromechanical ultrasonic transducer array chip to the outer wall of a water treatment reaction vessel, a coupling layer needs to be placed between the chip and the wall. The function of the coupling layer is to fill the tiny air gaps formed between the bottom surface of the chip and the wall due to surface roughness, thereby ensuring efficient acoustic transmission of elastic waves from the vessel wall to the array elements. The coupling layer material can be high-vacuum silicone grease, epoxy resin adhesive, or a special acoustic coupling paste, with the thickness controlled within the range of 10 micrometers to 100 micrometers. If the coupling layer is too thick, it will introduce additional acoustic attenuation and inconsistencies in propagation delay; if it is too thin, the interfacial air gaps cannot be completely eliminated. Therefore, during installation, the thickness of the coupling layer needs to be maintained within the above range by controlling the coating amount and the pressure applied.

[0021] The choice of attachment location is equally crucial. The outer walls of water treatment reaction vessels often exhibit curvature; for example, the side walls of cylindrical tanks, spherical heads, and conical transition sections all have different radii of curvature. When the wall curvature is too large, the fit between the chip's bottom surface and the wall decreases, and the coupling quality of edge elements is significantly worse than that of the central elements, directly affecting the accuracy of subsequent beamforming. Therefore, the attachment location is limited to an approximately planar monitoring wall area with a local curvature lower than a preset curvature threshold. The principle for selecting the curvature threshold is to ensure that within the area covered by the array chip, the maximum normal deviation of the wall from the tangent plane does not exceed the distance corresponding to half a wavelength of elastic wave propagation in the wall. Taking the propagation speed of a 100 kHz acoustic emission elastic wave in a steel plate at approximately 5000 meters per second as an example, half a wavelength is approximately 25 millimeters. For an 8x8 array chip with a 1 mm element spacing, the chip side length is approximately 8 mm. In this case, as long as the curvature radius of the monitoring wall area is not less than 1 meter, the normal deviation can be controlled within an acceptable range. For end cap areas with greater curvature, smaller array chips or smaller element spacing can be used to meet the fitting requirements.

[0022] refer to Figure 1 , Figure 1 The image, viewed from above, shows a piezoelectric micromechanical ultrasonic transducer array chip attached to the outer wall of a water treatment reaction vessel and its spatial geometric relationship with the acoustic emission source. Multiple array elements within the chip are arranged in a row-column matrix, with the physical positions of each element represented by regularly arranged square cells. The coverage area of ​​the array chip is marked by a rectangular dashed box; the entire chip is attached to an approximately planar monitoring wall region where the local curvature is below a preset curvature threshold. A two-dimensional planar coordinate system is established on the local tangent plane of the approximately planar monitoring wall region, with the origin at... Assume two coordinate axes at the geometric center of the array. and Determined along the row and column directions of the array, respectively. Figure 1 The location of a single acoustic emission source is marked, represented by a crack symbol, indicating the region in the container wall where elastic waves are generated due to the initiation and propagation of microcracks. The wavefront of the elastic wave propagating from the acoustic emission source towards the array is represented by concentric arcs. The spacing between these arcs reflects the wavelength of the elastic wave propagating through the wall, and the curvature of the arcs reflects the spherical expansion characteristic of the wavefront when the wave source is a point source. From the origin... Starting from a certain azimuth angle The arrowed straight line drawn indicates the current scanning direction and azimuth angle. by The positive axis is used as the reference for counterclockwise measurement, which is marked with an arc in the figure. The angular range. The core of beamforming lies in applying appropriate delays to each array element to achieve in-phase superposition of signals from the current scanning direction. Figure 1 The number is marked separately. The position of each element A perpendicular projection is made from this array element to the current scanning direction, and the projection point falls on the scanning direction line, with the origin at the origin. The distance between the projection point and the first point is the first point. The projection length of the 2D planar offset of each element on the current scanning direction. The projection length The decision was made The signal arrival time difference between each array element and the origin: Divide by the elastic wave propagation speed obtained from the frequency band-propagation speed correspondence table This yields the required delay compensation for the array element. From the origin... The line connecting the acoustic emission source and the location of the source is marked with a double-headed arrow, representing the propagation distance. The propagation distance is determined by the peak value of the focused energy in the synthesized waveform corresponding to the direction of the acoustic emission source and the elastic wave propagation speed. The product of the acoustic emission source and the array elements represents the actual propagation path of the elastic wave from the source to each array element. Since the distance from each array element to the acoustic emission source is different, the arrival time of the elastic wave at each array element is different. Beamforming compensates for this arrival time difference by aligning the delay lines, so that the signal from the direction pointed by the acoustic emission source is aligned in all array elements and then coherently superimposed and enhanced.

[0023] After sensing the transient strain caused by the acoustic emission elastic waves from the wall, each array element generates a charge proportional to the transient strain in its internal piezoelectric film due to the piezoelectric effect. This charge is then converted into an analog voltage signal by the built-in charge-to-voltage conversion structure within the array element. The amplitude of this analog voltage signal typically ranges from microvolts to millivolts, and the signal duration ranges from microseconds to milliseconds. Due to the extremely low signal amplitude, each array element's output is equipped with a low-noise preamplifier with a gain set between 20 dB and 60 dB, and an equivalent input noise not exceeding 2 microvolts per ohm, before being transmitted to the subsequent digitization stage. This ensures that the signal amplitude is increased to the effective input range of the analog-to-digital converter without introducing excessive noise.

[0024] Each array element's pre-amplified electrical signal is fed into its corresponding continuous-time Sigma-Delta analog-to-digital converter (ADC). The continuous-time Sigma-Delta architecture, rather than traditional successive approximation or pipelined ADCs, was chosen because it possesses inherent anti-aliasing capabilities and extremely high dynamic range. Traditional successive approximation ADCs require high-order analog anti-aliasing filters at the input to suppress out-of-band noise folding, while the continuous-time loop filter of the continuous-time Sigma-Delta ADC itself attenuates out-of-band signals, significantly reducing the requirements for front-end analog filters. Each continuous-time Sigma-Delta ADC performs oversampling analog-to-digital conversion on the electrical signal at a rate no less than 64 times the upper limit of the signal bandwidth. For example, with an acoustic emission signal bandwidth upper limit of 400 kHz, a 64-fold oversampling rate corresponds to a sampling rate of at least 25.6 MHz. A higher oversampling ratio results in lower quantization noise power density within the signal bandwidth and a higher equivalent signal-to-noise ratio (SNR). A 64x oversampling ratio provides approximately 33 dB of signal-to-noise ratio improvement in a single-order Sigma-Delta modulator, and this improvement can be further enhanced using second-order or higher modulators. Each continuous-time Sigma-Delta analog-to-digital converter outputs a 1-bit or multi-bit oversampled digital bitstream, with the bitstream rate equal to the oversampling rate. Optionally, a bandpass Sigma-Delta analog-to-digital converter can be used to directly digitize the frequency band of interest in the acoustic emission signal to further reduce the computational burden of subsequent decimation filtering.

[0025] Multiple oversampled digital bitstreams are fed in parallel into a Field-Programmable Gate Array (FPGA). The decimation filtering process within the FPGA is divided into two cascaded stages. The first stage is a cascaded integrator-comb filter, whose core operations involve only addition and delay, requiring no multipliers. Therefore, it consumes very few logic resources within the FPGA, making it suitable for high-rate primary decimation of high-rate oversampled bitstreams. The cascaded integrator-comb filter is typically set to 3 to 5 stages, with a decimation factor of 8 to 32. Taking an initial sampling rate of 25.6 MHz and a 16x decimation as an example, the bitstream rate drops to 1.6 MHz after cascaded integrator-comb filtering. The cascaded integrator-comb filter exhibits a certain droop in the passband, the amount of which depends on the filter stage and the decimation factor. The second stage is half-band filtering decimation. A characteristic of half-band filters is that approximately half of the coefficients are zero; therefore, the multiplication operation is only half that of a full-coefficient filter of the same order. Each stage of the half-band filter performs a 2x decimation. By cascading two to three stages of half-band filters, the bitstream rate can be further reduced by 4 to 8 times. Taking a single-stage half-band filter with a 2x decimation as an example, the bitstream rate is reduced from 1.6 MHz to 800 kHz. After the above two-stage decimation filtering, the sampling rate of the output multi-channel acoustic emission digital waveform has been reduced to a level suitable for subsequent real-time processing, while the frequency components within the signal bandwidth are fully preserved. Optionally, a compensation filter can be inserted between the cascaded integrator-comb filter and the half-band filter to correct the passband droop of the cascaded integrator-comb filter, making the frequency response of the final output waveform flatter.

[0026] The Field Programmable Gate Array (FPGA) selects one acoustic emission digital waveform corresponding to a preset reference element from multiple acoustic emission digital waveforms as the reference channel waveform. The principle for selecting the reference element is to choose the element with the smallest distance from the array's geometric center. When the array has an even number of rows and even numbers of columns, the geometric center does not fall on any single element. In this case, any one of the four elements closest to the geometric center can be selected as the reference element. During installation and deployment, the element's number is written into the FPGA's configuration register. The reference channel waveform serves two functions: first, it is used for initial event triggering; second, it serves as the absolute time reference for the arrival timestamp.

[0027] The short-time energy envelope of the reference channel waveform is calculated point-by-point. Specifically, it is calculated by taking the values ​​before and after the current sampling point as the center. The length of each sampling point is Sliding window, The number of half-window sampling points is preset, and the square values ​​of all sampling points within the window are summed to obtain the short-time energy envelope value of the current sampling point. The value of needs to balance temporal resolution and the smoothness of energy estimation. Too small a value will lead to drastic fluctuations in the energy envelope and frequent false triggers, while too large a value will obscure the event's start time. A typical approach to determining this value is to... The number of sampling points is set to one-tenth of the typical duration of the acoustic emission event. For example, if the duration of the acoustic emission event is approximately 200 microseconds and the sampling rate is 800 kHz, the event corresponds to approximately 160 sampling points. It can be set to 16.

[0028] When the short-time energy envelope first exceeds the initial trigger threshold, that moment is marked as a candidate event trigger moment. The initial trigger threshold is set as follows: during the quiet period of a silent emission event, the mean and standard deviation of the short-time energy envelope of the reference channel waveform are statistically analyzed. The initial trigger threshold is set as the mean plus a preset multiple of the standard deviation, typically between 3 and 6. This threshold setting is based on the fact that during quiet periods, background noise approximately follows a Gaussian distribution, and its short-time energy follows a chi-square distribution. The mean plus a certain multiple of the standard deviation provides a threshold that statistically controls the false trigger rate to a low level. Optionally, the initial trigger threshold can also be updated in real time using an adaptive approach. That is, the threshold value is continuously calculated using the moving mean and moving standard deviation of the short-time energy envelope over a longer time window (e.g., the most recent 5 to 10 seconds) to adapt to scenarios where the background noise level changes slowly with the device's operating status.

[0029] The analysis window for candidate events is truncated by extending a preset half-window length forward and backward from the trigger time of the candidate event. The selection of the half-window length needs to ensure that the window contains sufficient event energy to support subsequent frequency domain analysis, while not including too many adjacent events or background noise. Taking the typical duration of acoustic emission events of 200 microseconds as a reference, the half-window length can be set to 1.5 to 2 times the typical duration of the event, i.e., 300 to 400 microseconds, corresponding to 240 to 320 sampling points at an 800 kHz sampling rate.

[0030] A field-programmable gate array (FPGA) performs a short-time Fourier transform (SFT) on the candidate event analysis window to determine the dominant frequency band of the candidate events. The SFT transforms the time-domain waveform within the candidate event analysis window to the frequency domain, obtaining the amplitude spectrum of each frequency component. The frequency component with the largest amplitude is searched within the amplitude spectrum, and the frequency range corresponding to the point where the amplitude drops to half of the peak value is defined as the dominant frequency band of the candidate event. The dominant frequency band reflects the frequency range with the most concentrated energy in the current acoustic emission event. Determining the dominant frequency band is necessary because the acoustic emission elastic waves in the water treatment reaction vessel wall do not propagate at a single velocity. As a finite-thickness plate and shell structure, the elastic waves within the vessel wall propagate in guided wave modes, and different frequency guided wave components have different propagation velocities—a phenomenon known as dispersion. If a single fixed velocity is used indiscriminately for subsequent beamforming, the delay compensation for high-frequency and low-frequency components will deviate, and the spatial focusing effect of beamforming will be significantly degraded.

[0031] The propagation velocity of the elastic wave matching the main frequency band of the candidate event is retrieved from a pre-calibrated frequency band-propagation velocity correspondence table. This table is calibrated during system installation and deployment using either a pencil lead breaking method or a pulse excitation method: an artificial acoustic emission source is applied at a known location on the container wall, the array receives the signal, and the arrival time of different frequency band components is measured. Combined with the known source-array distance, the propagation velocity corresponding to each frequency band can be calculated. The calibration results are stored in the on-chip memory of the field-programmable gate array (FPGA) in the form of a lookup table with discrete frequency bands as keys and corresponding propagation velocity values ​​as values. For example, the propagation velocity corresponding to the 100 kHz to 150 kHz band is 5200 m / s, the propagation velocity corresponding to the 150 kHz to 200 kHz band is 4800 m / s, and so on. The granularity of frequency band division is generally set to 25 kHz to 50 kHz. Within the typical 100 kHz to 400 kHz range of acoustic emission signals, there are 6 to 12 frequency bands, each storing one velocity value. Optionally, two velocity values ​​can be stored in the frequency band-propagation velocity correspondence table, corresponding to symmetric and antisymmetric modes of guided waves respectively. During retrieval, the dominant mode is further identified based on the waveform characteristics in the event analysis window before selecting the corresponding velocity value.

[0032] Field-Programmable Gate Arrays (FPGAs) use a local tangent plane of the approximate planar monitoring wall region as the positioning reference plane. Since the attachment location is limited to a wall region with a local curvature below a preset curvature threshold, the wall surface can be approximated as a plane within the array chip's coverage area. Therefore, using this local tangent plane as the positioning reference plane does not introduce significant errors. A 2D planar coordinate system is established within the positioning reference plane with the projection point of the sensor array's geometric center onto the positioning reference plane as the origin. The two coordinate axes are defined along the array's row and column directions, respectively. In this 2D planar coordinate system, the position of each array element can be directly calculated from its row and column numbers.

[0033] The beamforming process within the field-programmable gate array (FPGA) scans the monitoring area direction-by-direction within the positioning reference plane at preset angular step values. The angular step value determines the angular resolution of the scan; a smaller value results in higher resolution but longer scanning time. For example, to cover a complete 360-degree azimuth within the positioning reference plane, a 1-degree angular step value requires scanning 360 directions, while a 2-degree angular step value requires scanning 180 directions. For acoustic emission localization applications, an angular step value between 1 and 5 degrees is typically sufficient to meet the azimuth determination requirements of the damaged area.

[0034] For the current scanning direction, the core operation of beamforming is to apply an appropriate delay to the acoustic emission digital waveforms of each element and then superimpose them. This ensures that the acoustic emission signals from the current scanning direction are coherently superimposed and enhanced after being aligned in each channel, while signals from other directions partially cancel each other out due to misalignment. The specific delay calculation method is as follows: The 2D planar offset of each element relative to the origin in the positioning reference plane is determined based on the row and column numbers of each element in the matrix. Let the first element be... The 2D plane offset of each element relative to the origin is: , Indicates the first The distance between each array element and the origin in the row direction Indicates the first The distance between each array element and the origin along the column direction, and the azimuth angle of the current scanning direction in the positioning reference plane are... Then the first The projection length of the 2D planar offset of each element on the current scanning direction. for .Will Divide by the obtained elastic wave propagation speed Get the first The signal arrival time difference between each array element and the origin , This refers to the elastic wave propagation velocity that matches the main frequency band of the current candidate event, retrieved from the frequency band-propagation velocity correspondence table. Then... Multiply by sampling rate And round down to get the first... Integer sampling point delay of each array element , To extract the sampling rate of the filtered multi-channel acoustic emission digital waveforms, zero-value sampling points of equal length to the maximum delay in all array elements are padded at the end of each channel's acoustic emission digital waveform. This zero-padding ensures that data overflow does not occur during delay line alignment. Delay line alignment is performed on each channel's acoustic emission digital waveform according to the corresponding integer sampling point delay, i.e., the [missing information - likely a specific step or step]... The overall waveform shifts backward. Each sampling point is used, and the empty head positions are filled with zero values. The effect of delay line alignment is to make the data from the current scanning direction appear as zero values. The arrival times of the signals in all array elements are uniformly aligned to the same sampling point. The aligned acoustic emission digital waveforms of each channel are then superimposed point by point to obtain the composite waveform in the current scanning direction. The superposition operation is implemented in a parallel addition tree structure within the field-programmable gate array, requiring only 6 stages of addition to complete the task for 64 signals.

[0035] The focused energy value in the current scanning direction is obtained by squaring the synthesized waveform at each sample point within a preset time window and then summing the results. The length of the time window is the same as or slightly shorter than the length of the candidate event analysis window. The summed focused energy value corresponds, in physical terms, to the total signal energy of the synthesized waveform within that time window. When the scanning direction exactly matches the true location of the acoustic emission source, the signals from each channel are aligned and coherently superimposed, resulting in a significant increase in the amplitude of the synthesized waveform and a peak in the focused energy value. When the scanning direction deviates from the true location of the acoustic emission source, the signals from each channel are not aligned, and some cancel each other out during superposition, resulting in a significant decrease in the focused energy value.

[0036] After traversing all scanning directions, the field-programmable gate array (FPGA) selects the maximum value from the focused energy values ​​of all scanning directions. When this maximum value exceeds an adaptive energy threshold, the corresponding scanning direction is determined as the direction of the acoustic emission source. The adaptive energy threshold is calculated by multiplying the median of the focused energy values ​​of all scanning directions by a preset multiple (usually set to 3 to 8 times), which serves as the criterion for distinguishing valid events from background noise. The median is chosen instead of the mean because when an acoustic emission source is present, the focused energy values ​​in a few directions may be significantly larger. The median is not sensitive to such large values ​​and can more stably reflect the background level.

[0037] After determining the pointing direction of the acoustic emission source, the propagation distance between the acoustic emission source and the origin is further determined to obtain the position coordinates of the acoustic emission source. The propagation distance is determined based on the following: in the synthetic waveform corresponding to the pointing direction of the acoustic emission source, the peak time of the focused energy corresponds to the moment when the arrival time of the acoustic emission pulses in the synthetic waveform is most concentrated and the amplitude is largest. After converting the peak time of the focused energy into a time value, multiplying it by the obtained elastic wave propagation velocity, the propagation distance is obtained. Specifically, the peak time of the focused energy refers to the time corresponding to the sampling point where the maximum value is located in the instantaneous power sequence obtained by taking the square value of each sampling point of the synthetic waveform within the time window used to calculate the focused energy value. The point determined in the 2D plane coordinate system along the pointing direction of the acoustic emission source according to the propagation distance, with the origin as the starting point, is the position coordinate of the acoustic emission source. Optionally, when the monitoring wall area is large and multiple sensor array chips are deployed, each array can independently complete beamforming positioning and report its positioning results to the host computer for multi-array intersection fusion, further improving the positioning accuracy and eliminating the ambiguity of the angle estimation of a single array.

[0038] The field-programmable gate array (FPGA) records the arrival time of the first sampling point in the reference channel waveform that exceeds the trigger level as the arrival timestamp of the acoustic emission event. The trigger level differs from the initial trigger threshold mentioned earlier: the initial trigger threshold acts on the short-time energy envelope, used to detect candidate events from the continuous data stream; the trigger level acts on the original amplitude of the reference channel waveform, used to accurately determine the absolute arrival time of the event. The trigger level is generally set to 3 to 5 times the root mean square value of the background noise amplitude of the reference channel waveform. The arrival timestamp is recorded based on the reference channel waveform and is not affected by delay compensation during beamforming; therefore, it can serve as the absolute time identifier for the acoustic emission event, used for event sorting and event rate statistics in subsequent steps.

[0039] The purpose of arrival consistency verification is to eliminate interference events caused by non-crack sources. Water treatment equipment generates various non-crack acoustic interferences during operation, such as pump vibration, valve opening and closing impacts, fluid turbulence impacting the wall, and periodic friction of the agitator. The time distribution of the elastic waves arriving at each array element from these interference sources often differs significantly from the arrival distribution of elastic waves from a single crack source. Consistency verification utilizes this difference for discrimination. Specifically, it calculates the difference between the time of the first sampling point exceeding the trigger level and the arrival timestamp in the acoustic emission digital waveform of each of the remaining array elements (i.e., all array elements except the reference element), obtaining the measured channel time difference for each array element. Simultaneously, based on the determined direction of the acoustic emission source and the obtained elastic wave propagation speed, the theoretical channel time difference that each array element should have relative to the reference element under ideal conditions (i.e., the signal indeed originates from a point source in the direction of the acoustic emission source) can be calculated. The calculation method for the theoretical channel time difference is... Projection length of each array element Divide by the elastic wave propagation speed Subtract the projected length of the reference array element and divide by the elastic wave propagation speed. If the measured channel time difference matches the theoretical channel time difference closely, it indicates that the signal indeed originates from a concentrated source near the direction of the acoustic emission source, which is highly likely to be a crack acoustic emission event. If the match is poor, it indicates that the signal may originate from a dispersed noise source or near-field interference. When the number of array elements whose measured and theoretical channel time differences are within a preset consistency tolerance is not less than the preset number of effective array elements, the current candidate event is determined to be a valid acoustic emission event. The consistency tolerance setting needs to comprehensively consider the calibration error of the propagation speed, the difference in array element coupling, and the quantization error of the sampling clock. It is generally set to 2 microseconds to 10 microseconds, corresponding to an equivalent distance error of approximately 10 mm to 50 mm in steel. The preset number of effective array elements is generally set to 60% to 80% of the total number of all non-reference array elements, rather than requiring all array elements to pass through, to tolerate a small number of edge array elements whose first threshold time shifts due to insufficient signal-to-noise ratio or coupling degradation. Candidate events that do not meet the consistency conditions are marked as interference events and eliminated, and will not proceed to subsequent feature extraction and clustering processes. Optionally, the eliminated interference events can be counted and statistically analyzed by type. If the number of interference events increases abnormally within a certain period, auxiliary prompts for abnormal equipment operation can be sent to the operation and maintenance personnel.

[0040] For valid acoustic emission events that pass the arrival consistency check, the field-programmable gate array (FPGA) extracts an event waveform segment from the reference channel waveform. The starting point for extraction is not the arrival timestamp itself, but rather a point determined by backwards from the arrival timestamp by a preset number of pre-triggered sampling points. The reason for including a waveform segment before the arrival timestamp is that although the amplitude of the leading edge of the acoustic emission elastic wave is lower than the trigger level, it contains important information about the type of acoustic emission source and the rupture mechanism. In particular, the frequency components and rise rate of the leading edge have significant discriminative value for different damage modes such as matrix cracking, fiber fracture, and interface delamination. If the extraction starting point happens to be exactly at the point where the trigger level has passed, the leading edge information will be completely lost, and the feature vector extracted by subsequent scattering transformation will lack the scale component corresponding to the low-amplitude leading edge, resulting in incomplete feature representation. The preset number of pre-triggered sampling points is generally set to 10% to 30% of the number of sampling points corresponding to the typical duration of the event. For example, when the typical duration of the event is 200 microseconds corresponding to 160 sampling points, the number of pre-triggered sampling points can be set to 16 to 48. Sampling points of a preset total extraction length are continuously extracted from the extraction starting point as the event waveform segment. The total truncation length should cover the complete event duration plus the pre-trigger segment and the aftershock segment, and is generally set to 1.5 to 2.5 times the typical event duration, corresponding to the number of sampling points. After the event waveform segment is truncated, it is transmitted in the form of a data packet to the subsequent scattering transformation dedicated integrated circuit for feature extraction.

[0041] The field-programmable gate array (FPGA) transmits the captured event waveform segments to the scattering converter application-specific integrated circuit (ASIC) via a serial interface. The serial interface uses a high-speed serial peripheral interface protocol, with a clock frequency set between 10 MHz and 50 MHz to ensure that all sampling points of the event waveform segment are transmitted before the next acoustic emission event arrives. The sampling points of the event waveform segment are sequentially fed into the input buffer memory of the scattering converter ASIC in fixed-point format. The bit width of the fixed-point number is consistent with the effective bit width of the front-end analog-to-digital converter, typically 14 to 18 bits. A cyclic redundancy check (CRC) code is added during transmission to detect transmission errors; if the check fails, a retransmission is requested.

[0042] The dedicated integrated circuit (ASIC) for scattering transformation internally integrates a two-stage cascaded processing pipeline. Its overall function is to extract multi-scale envelope features from the input event waveform segment. The reason for using an ASIC instead of implementing the scattering transformation operation on a field-programmable gate array (FPGA) or general-purpose processor is that the scattering transformation involves convolution and modulo operations across multiple parallel channels, resulting in high computational density and independent operation between channels, making it ideally suited for parallel execution via a hardware pipeline. The ASIC stores the filter coefficients in on-chip read-only memory, eliminating the need for external loading during runtime. It operates immediately upon power-up, with deterministic and predictable latency, meeting the stringent real-time requirements of online acoustic emission detection.

[0043] The basic idea of ​​scattering transform is to extract the envelope structure information of a signal layer by layer through multi-scale wavelet filtering and modulo operations. The first-stage processing passes the input signal through a set of wavelet filters covering different frequency scales, and after modulo operation, obtains the envelope at each scale, reflecting the change in energy distribution of the signal at each frequency scale over time. The second-stage processing passes the envelope output from the first stage through a set of even finer wavelet filters and performs modulo operation again, further capturing the modulation structure of the envelope itself. The features obtained after the cascaded two-stage processing are stable against translations and minor deformations of the input waveform; that is, even if acoustic emission events of the same type differ in arrival time and propagation path, the extracted feature vectors remain similar. This property is crucial for subsequent event clustering.

[0044] The first-stage pipeline uses multiple parallel pairs of orthogonal filter channels. The number of channel pairs depends on the number of frequency scales to be covered; a typical configuration is eight channel pairs, with the center frequencies of each pair spaced at octave intervals. For example, using an acoustic emission signal frequency range of 50 kHz to 400 kHz, the center frequencies of the eight channel pairs could be set to 50 kHz, 71 kHz, 100 kHz, 141 kHz, 200 kHz, 283 kHz, and 400 kHz, along with a low-pass channel covering the entire bandwidth. The center frequencies are spaced at octave intervals. The frequency spacing increases proportionally, meaning the ratio of the center frequencies of adjacent channel pairs is approximately 1.41. This spacing ensures adequate overlap between adjacent channels in the frequency domain, preventing the omission of signal components located between the two center frequencies. Optionally, the number of channel pairs can also be set to 6, 10, or 12, and the center frequency spacing can be configured in a proportional or equal bandwidth manner, depending on the trade-offs between frequency resolution and hardware resources.

[0045] Each orthogonal filter channel pair contains one real-part finite impulse response (FIR) filter and one imaginary-part finite impulse response (FIR) filter. The coefficients of the real-part FIR filter are pre-programmed with discrete samples of the real part of the Morlet wavelet at the corresponding center frequency, while the coefficients of the imaginary-part FIR filter are pre-programmed with discrete samples of the imaginary part of the Morlet wavelet at the same center frequency. The Morlet wavelet is a complex-valued wavelet; its real part is the product of a cosine function and a Gaussian window, and its imaginary part is the product of a sine function and a Gaussian window. Let the center frequency be... Sampling rate The filter length is Then the coefficient is the first one. The real part of each coefficient is The imaginary part is ,in This is the coefficient index, with values ​​ranging from 0 to... , Gaussian window function , The standard deviation of the Gaussian window controls the width of the window function in the time direction. The larger the value, the wider the time dimension and the higher the frequency resolution. The smaller the value, the narrower the time direction and the higher the time resolution. With center frequency Between usually according to Relationship settings, This is the bandwidth parameter, typically ranging from 0.5 to 1.5; a value of 1.0 corresponds to the standard Morlet wavelet. Filter length. It needs to be long enough to ensure that the Gaussian window decays to a negligible level at both ends; generally, it is taken as... to The corresponding number of sampling points. With a center frequency of 100 kHz and a sampling rate of 800 kHz... Taking version 1.0 as an example, Approximately 1.27 sampling points. Ten sampling points are sufficient. For channels with lower center frequencies, Larger filter sizes result in longer filter lengths; for example, a filter with a center frequency of 50 kHz corresponds to approximately 20 sampling points. All filter coefficients are written to on-chip read-only memory during chip manufacturing and cannot be changed during operation, ensuring the determinism and repeatability of the processing chain.

[0046] refer to Figure 2 It contains 6 sub-figures. Sub-figure (a) shows the event waveform segment of the input scattering transform application-specific integrated circuit. The time-domain waveform, with time on the horizontal axis in milliseconds and amplitude on the vertical axis in millivolts, exhibits typical acoustic emission transient pulse characteristics. Specifically, the waveform amplitude rapidly rises from the background noise level to its peak at the start of the event and then gradually decays, with the pulse duration on the order of hundreds of microseconds. Subfigure (b) shows the real part finite impulse response filter coefficients of a pair of orthogonal filter channels in the first-stage pipeline. Its waveform exhibits a cosine oscillation modulated by a Gaussian window envelope. The coefficients have the largest oscillation amplitude in the middle region, gradually decaying to zero towards both ends, reflecting the time-domain localization characteristic of the real part of the Morlet wavelet. (Coefficient index...) The filter length is obtained from 0. Subgraph (c) shows the imaginary finite impulse response filter coefficients for the same channel pair. Its waveform exhibits a sinusoidal oscillation modulated by a Gaussian window envelope, with a 90-degree phase shift compared to the real coefficients. Together, they constitute the corresponding center frequency. The discrete coefficients of the complete complex Morlet wavelet. Subgraph (d) shows the event waveform segment. With real part filter coefficients The real part of the convolution output is obtained after convolution. The horizontal axis represents time, and the vertical axis represents amplitude. The output waveform exhibits a significant oscillatory response within the time interval corresponding to the event pulse, while the amplitude approaches zero outside the pulse. Subfigure (e) shows the event waveform segment. With the imaginary part filter coefficients The imaginary part of the convolution output is obtained after convolution. Its temporal distribution is similar to that of the real part convolution output, but the oscillation phase is shifted, and the oscillation peaks and valleys of the two do not completely coincide. Subgraph (f) shows the real part convolution output. Convolution output with imaginary part The data is displayed together, and the modulus sequence calculated by both is superimposed on it. The modulus sequence is a smooth envelope curve, no longer exhibiting the rapid oscillations of the real and imaginary part convolution output, but rather smoothly reflecting the change in the instantaneous energy intensity of the signal at this center frequency scale over time. The modulus sequence monotonically rises to its peak within the event pulse interval and then smoothly declines, indicated by a filled area below the curve, visually representing the signal energy distribution captured by the channel. The modulus-taking operation of the orthogonal filter eliminates the influence of the carrier phase, allowing the modulus sequence to retain only envelope information, making it insensitive to minute shifts in the input waveform. This is precisely the key to the translational stability of the scattering transform characteristics.

[0047] The event waveform segment simultaneously enters all quadrature filter channel pairs. Within each channel pair, the real-part finite impulse response (FIR) filter and the imaginary-part finite impulse response (FIR) filter independently perform convolution operations on the event waveform segment. The convolution operation is implemented in hardware using a multiply-accumulate array, processing one output sample point per clock cycle. Let the event waveform segment be... , Indicates the first segment of the event waveform The amplitude of each sampling point For the sampling point index, the first The real part convolution output of each channel pair Event waveform segment With real part filter coefficients The linear convolution result, the imaginary part convolution output. Event waveform segment With the imaginary part filter coefficients The linear convolution result, Indexed for channel pairs.

[0048] The modulus sequence of a channel is obtained by adding the squared values ​​of the real and imaginary parts of the convolution output of the same channel and then taking the arithmetic square root point by point. Let the first... The first channel pair in the second The modulus at each sampling point is ,but , For the first The real part of the convolution output of the n channel pairs is in the th... The value at each sampling point For the first The imaginary part convolution output of the n channel pairs is in the th... The value at each sampling point. The reason for using both real and imaginary filters to form an orthogonal pair before modulo operation, rather than just using the real part filter and taking the absolute value, is that the waveform output by a single real part filter still retains the oscillating characteristics of the carrier frequency. Directly taking the absolute value results in an envelope that fluctuates rapidly and cannot smoothly reflect the energy envelope contour of the signal. Orthogonal pair modulo operation is equivalent to calculating the amplitude of the complex wavelet transform, mathematically eliminating the influence of the carrier phase. The resulting modulus sequence only reflects the instantaneous envelope strength of the signal at that frequency scale. The envelope curve is smooth and insensitive to minute shifts in the input waveform. This phase insensitivity is the physical source of the stability of the scattering transform characteristics. In hardware implementation, the squaring operation is performed using a dedicated multiplier, and the square root operation is implemented using a coordinate rotation digital calculation method or a piecewise linear approximation method to avoid introducing complex division and iteration circuits. The coordinate rotation digital calculation method gradually approximates the square root value through a fixed number of shifts and addition / subtraction operations, achieving 16-bit precision in 16 iterations.

[0049] The first-stage envelope sequence for each channel is obtained by using fixed-window-length moving average pooling on the modulus sequence. The pooling operation reduces the sequence length while preserving the overall trend of the envelope, thus reducing the amount of data processed in the subsequent second stage. The window length for moving average pooling is typically set to 4 to 16 sampling points, with the pooling step size equal to the window length (i.e., non-overlapping pooling). The length of the pooled sequence is shortened to the original modulus sequence length divided by the pooling window length. For example, with an event waveform segment length of 320 sampling points and a pooling window length of 8, the envelope sequence length of each channel output in the first stage is approximately 40 sampling points. Optionally, max pooling can be used instead of mean pooling. Max pooling preserves peak information within the window and is more sensitive to the transient peak response of pulsed acoustic emission events, but its smoothness is inferior to mean pooling.

[0050] The second-stage pipeline feeds the envelope sequences of each channel from the first stage back into the second group of quadrature filter channel pairs. The structure of the second group of quadrature filter channel pairs is identical to that of the first group; each channel pair also contains two finite impulse response filters (FIRS) with real and imaginary parts. The difference lies in the denser center frequency distribution of the second group. This denser frequency coverage is necessary because the frequency content of the envelope sequence output from the first stage has shifted from the high-frequency range of the original acoustic emission signal (50 kHz to 400 kHz) to the low-frequency range of envelope modulation (typically between several hundred Hz and tens of kHz). Within this narrower frequency range, a denser center frequency spacing is required to achieve sufficient frequency resolution. The number of channel pairs in the second group is typically set to 1 to 2 times that of the first group; for example, if the first group has 8 channel pairs, the second group can be set to 8 to 16 channel pairs. The second-stage pipeline performs convolution operations on the envelope sequences of each channel in the first stage. It adds the squared values ​​of the real and imaginary parts of the convolution output for each channel pair, and then takes the square root of each point to obtain the modulus sequence for each channel in the second stage. This modulus sequence is then subjected to fixed-window-length moving average pooling to obtain the envelope sequence for each channel in the second stage. The pooling window length for the second stage can be the same as or shorter than that for the first stage, depending on the length of the envelope sequence and the required feature dimensions.

[0051] The splicing output within the scattering transform dedicated integrated circuit arranges the envelope sequences of all channels in the first stage and the envelope sequences of all channels in the second stage in ascending order of center frequency, and then concatenates them end-to-end to form the acoustic emission event feature vector. Taking the first stage with 8 channel pairs and 40 sampling points per channel after pooling as an example, the feature component contributed by the first stage has a dimension of 320. When the second stage with 16 channel pairs processes the envelope sequences of the 8 channels in the first stage, if the second stage pools 10 sampling points per channel, the feature component contributed by the second stage has a dimension of 1280. The total dimension of the acoustic emission event feature vector after splicing the two is 1600. In actual deployment, the total dimension can be controlled by adjusting the number of channels and pooling parameters according to hardware resource constraints and feature discrimination requirements. Optionally, before splicing, L2 normalization can be performed on the envelope sequences of each channel, that is, dividing each channel envelope sequence by the arithmetic square root of the sum of squares of all sampling points of the sequence itself, in order to eliminate the scale imbalance caused by the difference in absolute energy between different channels. However, this step is not necessary, and a similar effect can be achieved in the normalization process in subsequent steps.

[0052] The scattering transformation application-specific integrated circuit outputs the stitched acoustic emission event feature vector via a serial interface. The acoustic emission event feature vector, along with the acoustic emission source location coordinates and arrival timestamps obtained in the preceding steps, is packaged by a field-programmable gate array (FPGA) and transmitted to the edge processor. The edge processor can be a computing platform with sufficient computing power, such as an embedded microprocessor, digital signal processor, or system-on-a-chip, to run subsequent normalization, clustering, and early warning algorithms.

[0053] The edge processor maintains a sliding time window with the current time as the endpoint and a preset sliding window duration as the width, retaining only valid acoustic emission events whose arrival timestamps fall within the sliding time window. The setting of the sliding window duration depends on the typical timescale of damage evolution in the monitored structure and the computing power limitations of the edge processor. For pressure equipment with low to medium stress levels, such as water treatment reaction vessels, the time span from crack initiation to the rapid propagation stage is typically on the order of several hours to several days, and the sliding window duration can be set from 1 hour to 24 hours. The longer the window, the more event samples are accumulated, and the better the statistical stability of the clustering results, but the computational load also increases accordingly. Taking an average of 200 valid acoustic emission events received per hour and a sliding window duration of 6 hours as an example, approximately 1200 valid acoustic emission events are maintained simultaneously within the window, corresponding to 1200 joint description vectors participating in the clustering operation. As time progresses, the events with the earliest arrival timestamps gradually slide out of the window and are removed, while newly arriving events are added to the window, and the clustering results are dynamically updated accordingly.

[0054] The edge processor normalizes the acoustic emission event feature vectors of all valid acoustic emission events within the sliding time window. The purpose of normalization is to eliminate numerical scale differences between the components of the acoustic emission event feature vector caused by variations in physical meaning and dimensions. Different channels in the acoustic emission event feature vector correspond to envelope sequences with different center frequencies, and the absolute amplitudes of each channel may differ by several orders of magnitude (low-frequency channels have high energy and large amplitudes, while high-frequency channels have low energy and small amplitudes). Without normalization, subsequent Euclidean distance calculations will be dominated by channels with larger amplitudes, and the discriminative information carried by channels with smaller amplitudes will be lost.

[0055] The specific normalization method is as follows: calculate the mean and standard deviation of each component of the acoustic emission event feature vector for all valid acoustic emission events within the sliding time window. Let there be a total of... One valid acoustic emission event The total number of valid acoustic emission events within the current sliding time window, the i-th The eigenvector of the acoustic emission event of the event is the th _ ... dimensional components are , For event indexing, If it is a dimension index, then the first Mean of dimensional components Standard deviation When a certain standard deviation Below the preset standard deviation lower limit At that time, with Alternative , The preset lower limit value for positive numbers is generally set to... to The magnitude. The significance of this alternative operation is to prevent the standard deviation of a certain dimension from approaching zero when exactly all events within the current window have the same or extremely similar values, leading to overflow in division operations or numerical anomalies in the normalization result. Subtracting the corresponding mean from each dimension component and then dividing by the corresponding standard deviation yields the normalized eigenvector, i.e., the [missing value]. The normalized eigenvector of the event is the first... Weiwei .

[0056] The location coordinates of the acoustic emission source also need to be normalized. These coordinates are 2D planar coordinates, and their range is directly related to the physical size of the monitoring area, typically ranging from millimeters to meters. If they are not normalized before being concatenated with the acoustic emission event feature vector, there will be a significant difference between the numerical range of the location coordinates and the normalized range of the feature vector (mean around 0, standard deviation around 1). The clustering results will then be completely dominated by the larger value. The normalized location coordinates are obtained by performing the same subtraction of the mean and division of the standard deviation on both components of the acoustic emission source location coordinates.

[0057] The normalized feature vector and the normalized position coordinates are concatenated end-to-end to form a joint descriptive vector. The concatenation order is to first arrange all dimensions of the normalized feature vector, followed by the two dimensions of the normalized position coordinates. Taking a total dimension of 1600 for the acoustic emission event feature vector as an example, the total dimension of the joint descriptive vector is 1602. The joint descriptive vector contains both waveform feature information and spatial location information of the acoustic emission event, enabling subsequent clustering to be performed jointly in the feature space and physical space: events with similar waveform features and close spatial locations tend to be grouped into the same cluster, reflecting the inherent consistency in signal characteristics and spatial distribution of acoustic emission events generated from the same damage area.

[0058] The edge processor performs hierarchical density clustering on all joint description vectors within the sliding time window. The advantage of hierarchical density clustering is that it does not require pre-specifying the number of clusters, can automatically discover event clusters with different density distributions, and is inherently robust to noise and outliers. In acoustic emission monitoring scenarios, events generated in different damage areas may form clusters with different densities in the joint description space—events are dense in high-activity areas and sparse in low-activity areas. Traditional clustering methods that require specifying a uniform density parameter struggle to accurately capture both types of clusters simultaneously. Hierarchical density clustering addresses this problem by extracting stable clusters at different density levels.

[0059] The first step in hierarchical density clustering is to calculate the core distance. For each joint description vector, calculate its Euclidean distance to all other joint description vectors and sort them in ascending order. Then, take the nth core distance from the sorted vectors. The Euclidean distance values ​​are used as the core distance of the joint descriptor vector. The preset neighborhood sampling number, which physically means the number of nearest neighbors referenced when measuring the local density around a certain event point. The larger the value, the smoother the core distance estimate and the less susceptible it is to the influence of a single outlier, but at the same time, the ability to distinguish small clusters will also decrease. The value is typically set between 5 and 30, for scenarios with approximately 1200 events within the window. Choosing a value between 10 and 20 is a reasonable option. The essential meaning of core distance is: to make the point's first... The nearest neighbor is included in its neighborhood, and the minimum radius of this neighborhood is what? The smaller the core distance, the denser the events around that point.

[0060] refer to Figure 3The horizontal and vertical axes of the three subgraphs are the projection coordinates of the normalized joint description vector onto a 2D plane. Each scatter point in the graph represents the position of a valid acoustic emission event in the joint description space. Subgraph (a) shows the complete structure of the minimum spanning tree constructed with reachability distance as the edge weight in the initial state. All event nodes are connected by edges in the minimum spanning tree to form a connected graph. The distribution of edges reflects the reachability distance relationship between events. It can be observed from the graph that the event points exhibit several naturally clustered structures in space. The edges between events within a cluster are shorter and denser, while the edges connecting different clusters are significantly longer, and these longer edges correspond to larger reachability distance values. Subgraph (b) shows the state after removing the two largest reachability distance edges from the minimum spanning tree in descending order of reachability distance. The removed edges are marked with dashed lines and crosses to indicate their original positions. After removal, the minimum spanning tree is broken into several connected components. Connected components with no fewer than the preset minimum cluster size are marked as valid acoustic emission event clusters and labeled with cluster numbers. At this point, the cross-cluster connection with the largest reach distance has been severed, and the most clearly separated event groups have become independent clusters. However, some event groups that originally belonged to different damage regions may still remain connected due to edges with medium reach distances. Subgraph (c) shows the state after removing the fifth edge with the largest reach distance. As more edges with larger reach distances are removed in sequence, the minimum spanning tree is further split into more connected components, and the previously merged event groups are separated into independent effective acoustic emission event clusters. The comparison of the three subgraphs shows the progressive segmentation process of hierarchical density clustering: the edges with the largest reach distance are removed first because these edges connect event points in low-density regions or across different high-density cores. Severing these edges naturally separates regions with significant density differences, while the tight structure within each effective acoustic emission event cluster, connected by edges with smaller reach distances, is preserved. Small connected components with fewer nodes than the preset minimum cluster size are not marked as effective acoustic emission event clusters. Events in these small connected components are considered outliers and do not participate in subsequent event rate statistics and damage status warnings. The component-wise arithmetic mean of the coordinates of all acoustic emission sources within each valid acoustic emission event cluster constitutes the cluster center coordinates, which are used to mark the spatial location of the corresponding damaged area on the wall of the water treatment reaction vessel.

[0061] The second step is to calculate the reach distance. For any two joint description vectors, the maximum value among the three values—their individual core distance and the Euclidean distance between them—is taken as the reach distance. Let the joint description vectors be... The core distance is Joint description vector The core distance is , and They represent the first The and the first The joint description vector of _ valid acoustic emission events, the Euclidean distance between them is _____. , express and The Euclidean distance across all dimensions of the joint description vector is the cross-distance. The key difference between reachability distance and ordinary Euclidean distance is that when at least one of two points is in a low-density region (large core distance), even if their actual Euclidean distance is very close, the reachability distance will still be increased to the level of the core distance. This means that points in low-density regions are not easily considered to be density-reachable from each other, thus avoiding the incorrect classification of sparse noise points into dense clusters.

[0062] Step 3 is to construct the minimum spanning tree. A complete graph is constructed using all joint description vectors as nodes and the reachability distances as edge weights. The minimum spanning tree is then extracted from the complete graph. The minimum spanning tree is an acyclic connected subgraph with the minimum sum of edge weights connecting all nodes. In actual computation, it is not necessary to explicitly construct the complete graph and then extract the minimum spanning tree. Prim's algorithm can be used to progressively expand from any node, adding the edge with the minimum reachability distance connecting to the current spanning tree to the tree each time, until all nodes are included. For With 10 nodes, the computational cost of Prim's algorithm is... Proportional to this, with a scale of 1200 events within a window, the computational power of an edge processor at a clock speed of hundreds of megahertz can be completed within seconds. Optionally, when the number of events is larger (more than 5000), spatial hashing or kd-tree indexing can be performed on the joint description vector first, and only the reachability distance between neighboring points can be calculated to reduce the overhead of full-pair full-distance calculation.

[0063] Step 4 is hierarchical partitioning. Edges in the minimum spanning tree are removed sequentially in descending order of reachability. Each removed edge splits the minimum spanning tree into more connected components. The logic behind removing edges from largest to smallest is that edges with the largest reachability connect the two nodes with the lowest density, the greatest distance between them, or those in the sparsest regions. Cutting these edges first means breaking the weakest cross-regional connections, leaving the remaining connected components with denser and more compact local structures. As the removal process progresses, each connected component continuously splits into smaller sub-clusters, forming a hierarchical cluster structure from coarse to fine. After each removal, connected components with at least a preset minimum cluster size are marked as valid acoustic emission event clusters. The minimum cluster size is generally set to... 1 to 2 times, for example When the minimum cluster size is set to 15, it can be set to 15 to 30. This setting ensures that the labeled effective acoustic emission event clusters contain a sufficient number of events to support the stability of subsequent event rate statistics. Connected components that are too small are considered outlier sets and are not labeled. Optionally, a cutoff threshold for reachability can also be set. When the reachability of the remaining largest edge is lower than the cutoff threshold, the removal stops to avoid over-splitting and breaking up the truly dense clusters.

[0064] The cluster center coordinates are calculated using the component-wise arithmetic mean of the coordinates of all acoustic emission sources within each valid acoustic emission event cluster. These cluster center coordinates represent the spatial centroid of the event cluster on the wall of the water treatment reaction vessel, physically indicating the approximate center of the damaged area. It is important to note that the cluster center coordinates use the original acoustic emission source coordinates before normalization, not the normalized coordinates. This is because the purpose of the cluster center coordinates is to mark the physical locations on the vessel wall for maintenance personnel to locate and inspect, and they should retain their true spatial meaning.

[0065] The edge processor analyzes the event rate changes for each valid acoustic emission event cluster according to their arrival timestamps to determine damage status and provide early warnings. Acoustic emission activity exhibits distinctly different event rate patterns at different stages of crack evolution: in the microcrack initiation stage, the event rate is low and approximately stable; in the stable propagation stage, the event rate increases slightly but slowly; before entering the rapid propagation stage, the event rate shows a continuously accelerating increasing trend—this increasing trend is a precursor signal of impending crack instability. Utilizing this physical law, early warnings can be issued before crack instability by tracking the increasing trend of the event rate.

[0066] The specific event rate tracking method is as follows: the time axis is divided into fixed statistical windows of equal length. The length of the fixed statistical window is generally set from 1 minute to 30 minutes, depending on the frequency of acoustic emission events and the required early warning sensitivity. The shorter the fixed statistical window, the faster the response to changes in the event rate, but the more susceptible it is to interference from short-term random fluctuations; the longer the fixed statistical window, the smoother the event rate estimation, but the timeliness of the early warning will decrease. Taking a fixed statistical window length of 10 minutes as an example, the number of new events in each valid acoustic emission event cluster within the window is counted every 10 minutes, and this number is used as the event rate for the current window.

[0067] When the event rate of the current window continuously increases within a predetermined number of fixed statistical windows, the corresponding valid acoustic emission event cluster is determined to enter a rapid expansion state. The predetermined number is typically set to 3 to 6 consecutive windows. The reason for requiring continuous increases across multiple windows, rather than just observing a sudden increase in the event rate of a single window, is that an increase in the event rate of a single window might be an illusion caused by random fluctuations or transient noise penetration. Requiring continuous increases across multiple windows effectively filters out such short-term disturbances. For example, with a predetermined number of 4, only when the event rate of the first window increases... The event rate of the first window is greater than that of the second window. The event rate of the first window, the event rate of the second window The event rate of the first window is greater than that of the second window. The event rate of the first window, the event rate of the second window The event rate of the first window is greater than that of the second window. When all three conditions of the event rate of a window are met simultaneously (i.e., the event rates of the most recent four consecutive windows form a strictly monotonically increasing sequence). The cluster is determined to be in a rapid expansion state only when it is the current fixed statistical window number. Alternatively, the criterion can be relaxed to "a number of windows with an increasing trend of not less than a preset proportion in the most recent windows" to accommodate scenarios where the event rate may experience slight pullbacks but is still accelerating overall.

[0068] Once a cluster of valid acoustic emission events is determined to be in a rapidly expanding state, the edge processor outputs the cluster center coordinates and rapid expansion status identifier as a damage warning signal to the online monitoring system of the water treatment reaction vessel. Upon receiving the damage warning signal, the online monitoring system can display the wall location corresponding to the cluster center coordinates on the human-machine interface using an alarm icon or highlighted marker. Simultaneously, it can trigger an audible and visual alarm or automatically record the warning event log. Optionally, the warning signal can also include a sequence of event rate values ​​for the valid acoustic emission event cluster within the most recent fixed statistical windows, allowing maintenance personnel to determine the rate of increase and acceleration of the event rate, thereby further assessing the urgency of crack propagation. When multiple valid acoustic emission event clusters exist simultaneously, the edge processor independently tracks the event rate changes for each cluster, and the damage warning signals for each cluster are output independently. Maintenance personnel can use this information to determine whether there are multiple simultaneously active damage areas on the vessel wall and the differences in risk levels among these areas.

[0069] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A nondestructive testing method for acoustic emission based on acoustic emission event clustering, characterized in that, Includes the following steps: Step 1: Attach the sensor array to the outer wall of the water treatment reaction vessel. Each element in the sensor array converts the acoustic emission elastic waves from the wall into electrical signals. The electrical signals of each element are digitized by the corresponding oversampling analog-to-digital converter and then sent to the field-programmable gate array (FPGA). The FPGA performs decimation filtering on the digitized signals to obtain multiple acoustic emission digital waveforms. The corresponding elastic wave propagation velocity is retrieved according to the main frequency band of the candidate events. The direction of the acoustic emission source is determined by scanning in each direction through delay-summation beamforming. The position coordinates of the acoustic emission source are determined according to the direction of the acoustic emission source and the propagation distance. A arrival consistency check is performed on the candidate events to eliminate interfering events. The event waveform segments of the valid acoustic emission events are extracted. Step 2: The field-programmable gate array transmits the event waveform segment to the scattering transformation ASIC. The scattering transformation ASIC performs multi-scale convolution, modulus taking, and pooling operations on the event waveform segment sequentially through a fixed multi-level filter channel, and outputs the acoustic emission event feature vector. Step 3: Within the sliding time window, the edge processor normalizes the acoustic emission event feature vector and the acoustic emission source location coordinates and then concatenates them into a joint description vector. Hierarchical density clustering is performed on the joint description vector to divide it into effective acoustic emission event clusters. The event rate changes of each effective acoustic emission event cluster are tracked. When the event rate continues to increase, a damage warning signal is output to the online monitoring system of the water treatment reaction vessel.

2. The method according to claim 1, characterized in that, The sensor array in step 1 is a piezoelectric micromechanical ultrasonic transducer array chip, which contains multiple array elements arranged in a row and column matrix. It is attached to the approximate planar monitoring wall area on the outer wall of the water treatment reaction vessel with a local curvature lower than a preset curvature threshold through a coupling layer.

3. The method according to claim 1, characterized in that, The oversampling analog-to-digital converter in step 1 is a continuous-time Sigma-Delta analog-to-digital converter, which performs oversampling analog-to-digital conversion at a rate not less than 64 times the upper limit of the signal bandwidth to output an oversampling digital bitstream; the decimation filtering in step 1 includes performing cascaded integral comb filtering and half-band filtering decimation on each oversampling digital bitstream.

4. The method according to claim 1, characterized in that, Step 1, which involves retrieving the corresponding elastic wave propagation velocity based on the main frequency band of a candidate event, includes: selecting one channel corresponding to a preset reference array element from the multi-channel acoustic emission digital waveforms as the reference channel waveform; calculating the short-time energy envelope for each sampling point of the reference channel waveform; marking the candidate event trigger time when the short-time energy envelope first exceeds the initial trigger threshold; extending the candidate event analysis window forward and backward by a preset half-window length centered on the candidate event trigger time; performing a short-time Fourier transform on the candidate event analysis window to determine the main frequency band; and retrieving the elastic wave propagation velocity matching the main frequency band from a pre-calibrated frequency band-propagation velocity correspondence table.

5. The method according to claim 4, characterized in that, Step 1, the delay-summation beamforming scanning in each direction, includes: using the local tangent plane at the sensor array attachment position on the outer wall of the water treatment reaction vessel as the positioning reference plane, and establishing a 2D planar coordinate system in the positioning reference plane with the projection point of the geometric center of the sensor array on the positioning reference plane as the origin; scanning in each direction according to a preset angle step value; for the current scanning direction, determining the 2D planar offset of each array element relative to the origin in the positioning reference plane based on the row and column numbers of each array element in the row and column matrix; determining the integer sampling point delay of each acoustic emission digital waveform based on the projection length of the 2D planar offset in the current scanning direction and the obtained elastic wave propagation velocity; supplementing the end of each acoustic emission digital waveform with zero-value sampling points of the same length as the maximum delay, and aligning the delay lines according to the corresponding integer sampling point delay; superimposing the aligned acoustic emission digital waveforms point by point to obtain the composite waveform in the current scanning direction; and squaring the composite waveform point by point and accumulating the results to obtain the focused energy value in the current scanning direction.

6. The method according to claim 5, characterized in that, Step 1, determining the position coordinates of the acoustic emission source, includes: after traversing all scanning directions, selecting the scanning direction with the largest focused energy value that exceeds the adaptive energy threshold as the pointing direction of the acoustic emission source; calculating the propagation distance based on the peak time of the focused energy in the synthetic waveform corresponding to the obtained elastic wave propagation velocity and the pointing direction of the acoustic emission source; and determining the position coordinates of the acoustic emission source in a 2D plane coordinate system along the pointing direction of the acoustic emission source with the origin as the starting point and according to the propagation distance.

7. The method according to claim 4, characterized in that, The arrival consistency check in step 1 includes: recording the first sampling point in the reference channel waveform that exceeds the trigger level as the arrival timestamp; calculating the measured channel time difference between the first sampling point in the acoustic emission digital waveform of each of the remaining array elements that exceeds the trigger level and the arrival timestamp; calculating the theoretical channel time difference of each array element according to the direction of the acoustic emission source and the obtained elastic wave propagation speed; and determining a valid acoustic emission event when the number of array elements whose difference between the measured channel time difference and the theoretical channel time difference is within the preset arrival consistency tolerance is not less than the preset number of valid array elements. The event waveform segment of the valid acoustic emission event in step 1 includes: taking the time determined by the preset number of pre-trigger sampling points backward from the arrival timestamp as the starting point of the segmentation, and continuously extracting sampling points of a preset total extraction length from the reference channel waveform as the event waveform segment.

8. The method according to claim 1, characterized in that, In step 2, a two-stage cascaded processing pipeline is embedded within the dedicated integrated circuit for scattering transformation. The first stage pipeline sets up multiple parallel pairs of orthogonal filter channels. Each pair of orthogonal filter channels includes one real-part finite impulse response (FIR) filter and one imaginary-part FIR filter. The coefficients of the real-part FIR filter are pre-programmed as discrete samples of the real part of the Morlet wavelet at the corresponding center frequency, and the coefficients of the imaginary-part FIR filter are pre-programmed as discrete samples of the imaginary part of the Morlet wavelet at the same center frequency. The event waveform segment simultaneously enters all pairs of orthogonal filter channels and performs convolution operations on them respectively. The pointwise squared value of the real part convolution output of the same channel pair is then compared with the pointwise squared value of the imaginary part convolution output. The magnitude values ​​are summed and the arithmetic square root is taken point by point to obtain the magnitude sequence. The magnitude sequence is then subjected to fixed-window-length sliding mean pooling to obtain the envelope sequence of each channel in the first stage. The second-stage pipeline feeds the envelope sequence of each channel in the first stage into the second group of orthogonal filter channel pairs to perform convolution operation. The point-by-point squared value of the real part convolution output and the point-by-point squared value of the imaginary part convolution output in the same channel pair are summed and the arithmetic square root is taken point by point to obtain the magnitude sequence of each channel in the second stage. The magnitude sequence of each channel in the second stage is then subjected to fixed-window-length sliding mean pooling to obtain the envelope sequence of each channel in the second stage. The envelope sequences of all channels in the first stage and all channels in the second stage are arranged from low to high center frequency and then concatenated to form the acoustic emission event feature vector.

9. The method according to claim 1, characterized in that, The normalization in step 3 includes: calculating the mean and standard deviation of each component of the acoustic emission event feature vector of all valid acoustic emission events within the sliding time window; replacing the standard deviation of a certain dimension with the preset lower limit of the standard deviation when it is lower than the preset lower limit; subtracting the corresponding mean from each component and dividing by the corresponding standard deviation to obtain the normalized feature vector; performing the same processing on the acoustic emission source location coordinates to obtain the normalized location coordinates. The hierarchical density clustering in step 3 includes: calculating the Euclidean distance between each joint description vector and all other joint description vectors and arranging them in ascending order; taking the k-th Euclidean distance value in the arrangement as the core distance, where k is the preset number of neighborhood samples; taking the maximum value among the three values ​​of the core distance and the Euclidean distance between any two joint description vectors as the cross distance; constructing a minimum spanning tree with all joint description vectors as nodes and cross distances as edges; removing edges from the minimum spanning tree in descending order of cross distance; marking the connected components with no less than the preset minimum cluster size as valid acoustic emission event clusters after each removal.

10. The method according to claim 9, characterized in that, Step 3, which tracks the event rate changes of each effective acoustic emission event cluster, includes: using the component-wise arithmetic mean of the location coordinates of all acoustic emission sources within each effective acoustic emission event cluster as the cluster center coordinates; counting the number of new events in each fixed statistical window according to the arrival timestamp as the current window event rate; and determining that the corresponding effective acoustic emission event cluster has entered a rapid expansion state when the current window event rate increases window by window for a preset number of consecutive fixed statistical windows. The cluster center coordinates and the rapid expansion state identifier are then output as damage warning signals to the online monitoring system of the water treatment reaction vessel.

Images