Ship pipeline leakage automatic detection method based on thermal imaging machine vision

Abstract

The invention discloses an automatic ship pipeline leakage detection method based on thermal imaging machine vision, and relates to the technical field of ship safety monitoring and machine vision detection. Performing gain offset correction and atmospheric transmittance correction on the original radiation value output by the thermal imaging camera to obtain a corrected thermal image; performing adaptive background vibration elimination on the acoustic emission signal to obtain a pure acoustic emission signal, extracting a time-frequency feature of the pure acoustic emission signal, calculating an acoustic emission confidence coefficient, and generating a spatial prior probability distribution diagram; performing thermal anomaly detection based on the spatial prior probability distribution diagram and calculating a joint confidence coefficient, and extracting a thermal anomaly contour sequence when a judgment condition is met; and performing thermal plume spatio-temporal evolution curve fitting on the thermal anomaly contour sequence, and executing time inversion to realize leakage source positioning and confirmation. According to the invention, the accuracy and reliability of ship pipeline leakage detection are improved.

Description

Technical Field

[0001] This invention relates to the field of ship safety monitoring and machine vision inspection technology, and in particular to an automatic detection method and system for ship pipeline leaks based on thermal imaging machine vision. Background Technology

[0002] With the increasing size and system integration of ships, the structural complexity of internal fuel lines, cooling water lines, lubricating oil lines, and hydraulic system lines has significantly increased. These lines are typically located in engine rooms, equipment rooms, and confined spaces, operating in environments characterized by high temperature, high humidity, strong vibration, and complex background noise. During long-term ship operation, factors such as aging of pipe connections, wear of seals, fatigue cracking of welds, and corrosion can all lead to minor leaks or even sudden ruptures. If leaks are not detected in time, they can not only cause equipment performance degradation, energy waste, and environmental pollution, but also potentially lead to serious safety accidents in the event of fuel or flammable media leaks. Therefore, achieving rapid, accurate, and automated detection of ship pipeline leaks has become an important research direction in the field of ship operation and maintenance safety monitoring. In recent years, with the development of machine vision technology, infrared thermal imaging technology, and intelligent image processing algorithms, visual perception-based pipeline leak detection methods have been gradually applied to the monitoring of industrial pipelines and ship equipment, achieving automatic identification by analyzing abnormal areas or leak morphological characteristics in images. However, the shipboard cabin environment is characterized by complex background interference, temperature and humidity variations, and equipment vibration and noise. Single visual information often fails to accurately distinguish between genuine leaks and false anomalies such as heat source interference, equipment radiation, or changes in lighting. Furthermore, traditional visual detection methods typically rely on static image features or target detection models, lacking constraints on the physical evolution of the leak formation process and a multi-source information collaborative judgment mechanism. This results in limited stability and reliability in detecting weak leaks, early leaks, and under complex environmental conditions. Therefore, there is an urgent need for an automatic leak detection method for ship pipelines that can integrate multi-source sensing information and combine it with physical evolution characteristics for comprehensive judgment, in order to improve the accuracy and reliability of leak detection in complex environments.

[0003] CN120543529A discloses a "Method and System for Detecting Pipeline Damage and Leakage Based on Machine Vision." This method acquires image data of the pipeline to be detected and uses a deep learning YOLOv8s network model to perform target detection on the pipeline image, thereby identifying whether there are damaged targets in the pipeline. Based on the detection results, it combines traditional target detection algorithms for auxiliary judgment to improve the detection recall rate, and finally analyzes the pipeline damage and outputs the results. This method combines deep learning detection algorithms with traditional image detection algorithms, which improves the accuracy of pipeline damage identification to a certain extent and reduces the workload of manual inspection. However, this method mainly relies on visible light image information for damage identification. Its detection process focuses on the analysis of the appearance features of pipeline structural damage, which is difficult to effectively identify small or hidden leaks that are common in ship cabin environments. At the same time, this method lacks the analysis mechanism for changes in ambient temperature, thermal radiation background, and thermal anomalies caused by internal fluid leakage in pipelines. It also does not introduce acoustic or other physical sensing information for auxiliary judgment. In complex environments, it is easily affected by factors such as changes in lighting, stains, or structural obstruction, thus affecting the reliability of detection.

[0004] CN116363584A discloses a "Machine Vision-Based Method for Monitoring Leaks in Ship Liquid Pipelines." This method acquires monitoring video images of ship liquid pipelines and preprocesses them to generate image frame sequences. It then uses the KNN algorithm to classify and identify the image frames to determine if a leak exists. Furthermore, it extracts the foreground image using a Gaussian Mixture Model (GMM) background subtraction algorithm, further extracting features from the leaking droplets and locating the leak point, thus achieving the detection and location of leaks in ship pipelines. While this method utilizes image sequence analysis and background modeling techniques to a certain extent to identify and locate leaking droplets, its core detection still relies on the droplet motion characteristics in visible light video images. When the leak volume is small or no obvious droplets have formed in the early stages of the leak, the detection sensitivity is low. In addition, this method does not consider the influence of equipment vibration, thermal radiation, and airflow disturbances in the ship's engine room environment on the image foreground extraction, nor does it establish a physical evolution constraint mechanism based on the characteristics of leak thermal plumes or energy propagation. It also lacks multi-source information fusion judgment capabilities, thus making it prone to false detections or missed detections in complex cabin environments. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0006] Given that existing ship pipeline leak detection technologies are susceptible to interference from equipment vibration, changes in ambient temperature and humidity, and the influence of local heat sources in complex cabin environments, making it difficult for a single detection method to achieve high-precision leak identification and accurate location, this invention is proposed.

[0007] Therefore, the problem to be solved by this invention is how to achieve highly reliable automatic detection and accurate location of pipeline leaks in ships under complex ship operating conditions.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide an automatic detection method for ship pipeline leaks based on thermal imaging machine vision, comprising, The ambient temperature and relative humidity parameters of the cabin area to be inspected in the pipeline of the ship are collected by an acoustic emission sensor array and a thermal imaging camera. Gain bias correction and atmospheric transmittance correction are performed on the raw radiation value output by the thermal imaging camera to obtain a corrected thermal image. Adaptive background vibration elimination is performed on the acoustic emission signals collected by each channel of the acoustic emission sensor array to obtain pure acoustic emission signals; Extract time-frequency features from the pure acoustic emission signal, calculate the acoustic emission confidence level, and trigger the thermal imaging camera to enter the high-gain sub-pixel interpolation mode when the acoustic emission confidence level is greater than or equal to the preset acoustic trigger threshold, generating a spatial prior probability distribution map. The spatial prior probability distribution map is superimposed on the thermal anomaly detection process of the calibrated thermal image as a spatial search weight. The confidence of local thermal anomalies is calculated to obtain the joint confidence. When the joint confidence is greater than or equal to the preset joint decision threshold, the thermal anomaly contour sequence of the corresponding region is extracted from the calibrated thermal image. The thermal anomaly profile sequence is fitted with the spatiotemporal evolution curve of the thermal plume over multiple consecutive frames. The fitting residual is compared with a preset physical rationality threshold. If the fitting residual is less than the preset physical rationality threshold, the thermal anomaly profile sequence is reversed using the time inversion of the spatiotemporal evolution curve to trace the source in reverse, and the location coordinates of the leak point and the leak confirmation result are output. Otherwise, the thermal anomaly profile sequence is determined to be a non-leakage interference source, and the pure acoustic emission signal is reacquired. The thermal imaging camera is reset to the standard acquisition mode and the pure acoustic emission signal is reacquired.

[0009] Secondly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of the above-described automatic detection method for ship pipeline leaks based on thermal imaging machine vision.

[0010] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the above-described automatic detection method for ship pipeline leaks based on thermal imaging machine vision.

[0011] Compared with existing technologies, the advantages of this invention are as follows: By utilizing an acoustic emission sensor array and a thermal imaging camera to collaboratively acquire cabin environmental temperature parameters, relative humidity parameters, and infrared radiation information of the pipeline area, and performing gain bias correction and atmospheric transmittance correction on the original radiation values ​​to obtain a corrected thermal image, while simultaneously performing adaptive background vibration elimination on the acoustic emission signal to obtain a pure acoustic emission signal, this invention achieves simultaneous suppression of infrared measurement errors and equipment vibration noise in the complex environment of ship cabins, making the thermal image temperature inversion results more realistic and reliable, and effectively eliminating the interference of mechanical vibration on the acoustic signal; by extracting time-frequency features from the pure acoustic emission signal and calculating the acoustic emission... When the acoustic emission confidence level exceeds a preset acoustic trigger threshold, the thermal imaging camera is triggered to enter a high-gain subpixel interpolation mode and generate a spatial prior probability distribution map. This achieves adaptive enhancement of the thermal imaging detection process using acoustic anomalies as trigger conditions. The system can improve the spatial resolution of thermal imaging when a suspected leak occurs, and utilizes the sound source localization results to form a probability distribution to limit the search area for thermal anomalies, thereby improving the observability of weak leak thermal signals and reducing unnecessary computational burden. By superimposing the spatial prior probability distribution map as spatial weights into the thermal anomaly detection process of the calibrated thermal image, and combining the local thermal anomaly confidence level with the acoustic emission confidence level to construct a joint confidence level... Threshold judgment is performed to achieve the fusion of acoustic and thermal information, enabling leak identification to move beyond relying on a single temperature anomaly. Instead, it involves comprehensive judgment under spatial probability constraints and multi-source information weighting, effectively reducing the risk of false detections caused by ambient temperature fluctuations and equipment thermal radiation, and improving the accuracy of detection results. In the detection confirmation stage, when the joint confidence level meets the judgment condition, connected pixel sets whose normalized temperature deviation exceeds a threshold are extracted and contour tracking is performed to obtain a continuous multi-frame thermal anomaly contour sequence. This achieves a continuous temporal description of the spatial boundary of the leak thermal plume, expanding the detection results from single-frame temperature anomalies to stable spatiotemporal contour information. This reduces the impact of transient noise on detection results and provides reliable data for leak diffusion behavior analysis. By fitting multi-frame spatiotemporal evolution curves of thermal plumes to the thermal anomaly profile sequence and judging based on the fitting residuals and preset physical rationality thresholds, the leak source is reversed using time inversion when the conditions are met; otherwise, it is judged as a non-leakage interference source and the signal is reacquired. This achieves physical consistency verification of whether the thermal anomaly phenomenon conforms to the real leak diffusion law, and can accurately output the location coordinates of the leak point when a leak is confirmed, thereby effectively eliminating non-leakage heat source interference and achieving accurate positioning, ultimately improving the reliability and positioning accuracy of ship pipeline leak detection. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of an automatic detection method for ship pipeline leaks based on thermal imaging machine vision. Detailed Implementation

[0013] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0014] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0015] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0016] As mentioned in the background section, existing technologies typically rely on single acoustic detection or infrared thermal imaging detection methods, which are easily affected by factors such as mechanical vibration noise, ambient temperature fluctuations, and non-leaking heat sources. This results in high false detection rates, poor detection stability, and difficulty in accurately determining the leak location. To address these problems, this invention provides an automatic leak detection method for ship pipelines based on thermal imaging machine vision.

[0017] Reference Figure 1 , Figure 1 This is a flowchart illustrating an automatic detection method for ship pipeline leaks based on thermal imaging machine vision, according to an embodiment of the present invention. Figure 1 As shown, an automatic detection method for ship pipeline leaks based on thermal imaging machine vision includes: S1: The ambient temperature and relative humidity parameters of the cabin area to be inspected in the pipeline of the ship are collected by the acoustic emission sensor array and the thermal imaging camera. The gain bias correction and atmospheric transmittance correction are performed on the raw radiation value output by the thermal imaging camera to obtain the corrected thermal image. S1.1: An acoustic emission sensor array and a thermal imaging camera are arranged in the pipeline inspection area of ​​the ship, and a temperature and humidity sensor is installed near the lens of the thermal imaging camera to collect the cabin environmental temperature and relative humidity parameters of the inspection area. Specifically, an acoustic emission sensor array is deployed along the pipeline in the inspection area of ​​the ship's pipeline at a preset installation interval, and a thermal imaging camera is fixedly installed at the field of view coverage position of the inspection area; the installation interval between adjacent acoustic emission sensors in the acoustic emission sensor array does not exceed the effective detection distance of a single acoustic emission sensor, and the installation position of the thermal imaging camera ensures that the field of view covers the entire pipeline surface of the inspection area; the coordinated deployment of the acoustic emission sensor array and the thermal imaging camera constitutes the hardware foundation of this solution, with the acoustic emission sensor array responsible for sound source localization in subsequent steps, and the thermal imaging camera responsible for thermal anomaly imaging in subsequent steps; Furthermore, the temperature and relative humidity parameters of the cabin environment in the area to be inspected are continuously collected by the temperature and humidity sensor at the same sampling period as the frame acquisition period of the thermal imaging camera. The cabin environment temperature parameters and relative humidity parameters are established in a one-to-one time index relationship with the frame sequence output by the thermal imaging camera according to the frame timestamp, so as to obtain the environmental parameter timestamp sequence. Preferably, the distance between the temperature and humidity sensor and the front end of the thermal imaging camera lens does not exceed a preset near-end distance threshold, so that the temperature and relative humidity parameters of the cabin environment are consistent with the environmental conditions within the actual working optical path of the thermal imaging camera, eliminating the deviation introduced by the uneven spatial distribution of the cabin environmental parameters to the subsequent atmospheric transmittance calculation. In an optional embodiment, the area to be inspected is an 8-meter-long high-pressure steam pipeline inside the ship's engine room, horizontally arranged along the bulkhead. Six acoustic emission sensors, numbered CH1 to CH6, are sequentially installed along the pipeline at preset installation intervals of 1.5 meters. The installation coordinates of each sensor are recorded in a sensor coordinate table with the corner of the engine room floor as the origin. The 1.5-meter spacing between adjacent sensors does not exceed the effective detection distance of 2.0 meters for a single sensor. A thermal imaging camera is fixedly installed on a bulkhead bracket at a height of 2.2 meters, with a field of view covering the entire 8-meter pipeline surface. A temperature and humidity sensor is attached 50 millimeters from the front end of the thermal imaging camera lens, not exceeding a preset near-end distance threshold of 100 millimeters. The sampling period of the temperature and humidity sensor and the frame acquisition period of the thermal imaging camera are both set to 25 milliseconds. When each frame of thermal image is output, the corresponding cabin environmental temperature parameters and relative humidity parameters are synchronously written. A one-to-one time index relationship is established according to the frame timestamps to obtain the environmental parameter timestamp sequence, where the cabin environmental temperature parameter for a certain frame is 34.2 degrees Celsius and the relative humidity parameter is 71%. S1.2: Using the cabin ambient temperature parameter as the independent variable, perform two-point radiometric calibration on each pixel in the focal plane array of the thermal imaging camera to obtain the gain bias correction radiometric value matrix. Furthermore, the current focal plane temperature value is read from the focal plane array of the thermal imaging camera, and the pixel-by-pixel gain coefficient lookup table and pixel-by-pixel offset lookup table established by the manufacturer at the local storage unit of the thermal imaging camera are retrieved. The pixel-by-pixel gain coefficient lookup table and pixel-by-pixel offset lookup table are both indexed by the focal plane temperature value as the row index and by the pixel coordinate as the column index, recording the gain coefficient and offset of each pixel under each focal plane temperature value. It should be noted that the temperature range of ship cabins is large and changes continuously with the navigation area and season. The response characteristics of the focal plane array of the thermal imaging camera will experience systematic drift as the focal plane temperature value changes. The pixel-by-pixel gain coefficient lookup table and the pixel-by-pixel offset lookup table are performed using the focal plane temperature value as a dynamic index. Specifically, based on the focal plane temperature value, the gain coefficient and offset of each pixel corresponding to the current focal plane temperature value are looked up in the pixel gain coefficient lookup table and pixel offset lookup table respectively using linear interpolation; the original radiometric value of each pixel in the original radiometric value matrix output by the thermal imaging camera in the current frame is subtracted from the offset of each pixel and then divided by the gain coefficient of each pixel to obtain the gain offset correction radiometric value matrix. Preferably, linear interpolation uses two calibrated temperature nodes adjacent to the focal plane temperature value in the pixel-by-pixel gain coefficient lookup table and the pixel-by-pixel offset lookup table as interpolation nodes. When the focal plane temperature value exceeds the calibrated temperature range of the pixel-by-pixel gain coefficient lookup table and the pixel-by-pixel offset lookup table, an abnormal flag is triggered and subsequent calculations are stopped. Data acquisition continues after recalibration. In an optional embodiment, following the environmental parameter timestamp sequence, the focal plane temperature value corresponding to the current frame is read as 38.5 degrees Celsius; the pixel-by-pixel gain coefficient lookup table and the pixel-by-pixel offset lookup table establish interpolation node pairs with 35 degrees Celsius and 40 degrees Celsius as adjacent calibration temperature nodes during factory calibration, and perform linear interpolation between the above two nodes with the focal plane temperature value of 38.5 degrees Celsius, with an interpolation weight of (38.5-35) / (40-35)=0.70; for a certain pixel in the focal plane array with coordinates (256, 128), the gain coefficient at the 35-degree Celsius node is found to be 1.024, and the offset is 0.70. With a gain of 312 and a bias of 318 at a 40°C node, the gain coefficient of this pixel after linear interpolation is 1.024 + 0.70 × (1.031 - 1.024) = 1.029, and the bias is 312 + 0.70 × (318 - 312) = 316. The original radiance of this pixel in the current frame is 3501, and the radiance after gain and bias correction is (3501 - 316) / 1.029 = 3094. This radiance is written to the corresponding position in the gain and bias correction radiance matrix. All pixels in the array are processed one by one in the same way to obtain the gain and bias correction radiance matrix. S1.3: Using the relative humidity parameter and the optical path distance from the thermal imaging camera to the surface of the pipe in the area to be inspected as inputs, the atmospheric transmittance corresponding to the current frame is calculated using the infrared band atmospheric radiation attenuation model; Furthermore, the infrared atmospheric radiation attenuation model uses relative humidity and cabin ambient temperature as environmental inputs and optical path distance as path length parameter to output atmospheric transmittance.

[0018] S1.4: Divide the gain bias correction radiance of each pixel in the gain bias correction radiance matrix by the atmospheric transmittance to obtain the atmospheric correction radiance matrix. In an optional embodiment, following the gain bias correction radiance matrix, using the cabin ambient temperature parameter of 34.2 degrees Celsius and the relative humidity parameter of 71% as environmental inputs, and the optical path distance from the thermal imaging camera installation position to the pipe surface of 3.6 meters as the path length parameter, and substituting them into the infrared band atmospheric radiation attenuation model, the atmospheric transmittance corresponding to the current frame is calculated to be 0.943; the gain bias correction radiance value of the pixel at coordinates (256, 128) of 3094 is divided by the atmospheric transmittance of 0.943 to obtain the atmospheric correction radiance value of the pixel of 3280. All pixels in the entire array are processed in the same way to obtain the atmospheric correction radiance matrix; S1.5: Based on Planck's radiation law and preset pipeline surface emissivity parameters, the atmospheric correction radiation value of each pixel in the atmospheric correction radiation value matrix is ​​converted into the corresponding apparent temperature value to obtain the corrected thermal image; It should be noted that the calibrated thermal image is stored in the frame buffer queue with the frame timestamp as the identifier, and is called by the thermal anomaly detection step after the clean acoustic emission signal triggers the thermal imaging camera to enter the high-gain sub-pixel interpolation mode. The preset pipe surface emissivity parameters are stored in the emissivity parameter table according to the pipe material category, which includes carbon steel pipes, stainless steel pipes, and copper alloy pipes. When performing Planck's radiation law conversion, the preset pipe surface emissivity parameters are automatically matched from the emissivity parameter table according to the material identifier of each pipe segment in the pipe 3D model, so as to eliminate the influence of the difference in emissivity of different pipe materials on the temperature accuracy of the calibrated thermal image. S2: Perform adaptive background vibration elimination on the acoustic emission signals collected by each channel of the acoustic emission sensor array to obtain a pure acoustic emission signal; S2.1: Read the raw acoustic emission signals of each channel within the current time window from the acquisition units of each channel of the acoustic emission sensor array; It should be noted that the sampling frequency of the original acoustic emission signal is not lower than the preset minimum sampling frequency threshold. The preset minimum sampling frequency threshold is determined based on the Nyquist sampling theorem, which is twice the highest frequency component of the acoustic emission signal from the pipeline leakage. Multiple background vibrations exist simultaneously in the ship's engine room, including main engine vibration, propulsion shaft vibration, and wave impact. The background vibration energy in the original acoustic emission signal is much higher than the leakage acoustic emission energy generated by a minor leak. If background vibration elimination is not performed and features are extracted directly, the leakage features will be submerged by the background vibration in the subsequent acoustic emission confidence calculation. An original signal buffer matrix is ​​established based on the channel number and acquisition timestamp of the original acoustic emission signal. The establishment of the original signal buffer matrix provides a unified data input basis for the adaptive processing of each channel. S2.2: Perform short-time Fourier transform on the original acoustic emission signal according to the preset frequency band division method, extract the power spectral density distribution of each channel in each frequency band, obtain the background vibration power spectrum baseline, and store the background vibration power spectrum baseline in the background baseline database with the channel number and the center frequency of the frequency band as dual indexes. Specifically, during the silent reference period when the ship's piping system is operating normally and without leakage, the selection conditions for the silent reference period are: the ship is in a constant speed straight-running condition, the main engine speed is within the preset stable speed range, and there is no planned mechanical operation in the cabin; the preset frequency band division method divides the effective frequency range of the original acoustic emission signal into several sub-frequency bands at logarithmic intervals, so that the frequency band where the low-frequency background vibration energy is concentrated and the frequency band where the high-frequency leakage acoustic emission signal is located can be distinguished in the background vibration power spectrum baseline. The background baseline database stores the background vibration power spectrum baseline of each channel according to the ship's operating condition category. In an optional embodiment, a quiet reference period is selected where the ship is sailing at a constant speed of 15 knots, the main engine speed is 85% of the rated speed, and there is no planned mechanical operation in the cabin. The raw acoustic emission signals of 6 channels are continuously collected for 60 seconds. The effective frequency range of 0 to 250 kHz is divided into 16 sub-bands at logarithmic intervals by a preset frequency band division method, with center frequencies of approximately 1 kHz, 2 kHz, 4 kHz to approximately 200 kHz, respectively. The power spectral density baseline value of channel CH3 is relatively high in the 1 kHz to 4 kHz band, corresponding to the low-frequency component of the main engine vibration, and the power spectral density baseline value is relatively low in the 100 kHz to 200 kHz band. Using channel number CH3 and the center frequency of the band 2 kHz as a dual index, the corresponding baseline power spectral density value is stored in the background baseline database. This operating condition category is marked as operating condition A, and the storage of 96 background vibration power spectral baselines for 6 channels × 16 sub-bands is completed. S2.3: Based on the current operating condition identifier of the actual ship inspection process, the background vibration power spectrum baseline is used as the reference background vibration power spectrum baseline for the current inspection cycle. If the current operating condition identifier does not completely match the existing operating condition categories in the background baseline database, the background vibration power spectrum baselines corresponding to the two operating condition categories in the background baseline database that are closest to the current operating condition identifier in Euclidean distance are linearly interpolated according to the operating condition similarity weight to obtain the interpolated background vibration power spectrum baseline. The interpolated background vibration power spectrum baseline is used to replace the reference background vibration power spectrum baseline in subsequent calculations. It should be noted that the operating conditions of a ship during actual navigation are constantly changing. Using a single baseline of background vibration power spectrum collected under a single reference operating condition as the elimination reference will result in a significant increase in the residual amount of background vibration elimination when the operating conditions are switched. The linear interpolation mechanism can maintain the continuity of background vibration elimination during the switching of operating conditions. In an optional embodiment, continuing from the background baseline database, the current operating condition identifier for the current detection period is 13 knots of speed and 78% of the rated engine speed. This operating condition identifier does not completely match the existing operating condition categories in the background baseline database. The Euclidean distances between the current operating condition identifier and operating condition A (15 knots / 85% speed) and operating condition B (10 knots / 65% speed) are calculated to be 2.24 and 3.61, respectively. Taking operating condition A and operating condition B as the two most recent operating condition categories, linear interpolation is performed on the background vibration power spectrum baselines corresponding to the two categories according to the operating condition similarity weights (0.62 for operating condition A and 0.38 for operating condition B). Taking the CH3 channel and the center frequency of the frequency band 2 kHz as an example, the power spectral density value of the operating condition A baseline is 4.21 × 10⁻⁶. -4 volt 2 / Hertz, the corresponding value for operating condition B is 3.58×10 -4 volt 2 / Hertz, the interpolation result is 4.21 × 10 -4 ×0.62+3.58×10 -4 ×0.38=3.97×10 -4 volt 2 / Hz, written into the interpolated background vibration power spectrum baseline, replacing the reference background vibration power spectrum baseline in subsequent calculations; S2.4: Perform a short-time Fourier transform on the original acoustic emission signal according to the same preset frequency band division method as the short-time Fourier transform to obtain the real-time power spectral density matrix of each channel, and subtract the real-time power spectral density value from the baseline power spectral density value of the corresponding channel and frequency band in the interpolated background vibration power spectral density baseline to obtain the remaining power spectral density matrix. Furthermore, the short-time Fourier transform uses the Hanning window function to window each time frame, and the window length is consistent with the short-time Fourier transform window length used when establishing the background vibration power spectrum baseline, so as to ensure that the real-time power spectral density matrix of each channel is consistent with the reference background vibration power spectrum baseline in terms of frequency resolution and time resolution, and eliminate the calculation error caused by inconsistent transformation parameters. S2.5: Perform adaptive thresholding on the residual power spectral density values ​​of each frequency band in the residual power spectral density matrix to obtain the sparsified residual power spectral density matrix for each channel: Furthermore, the mean and standard deviation of the residual power spectral density matrix of each channel within the sliding time window are calculated, and the result of the mean plus a preset multiple of the standard deviation is used as the adaptive retention threshold for each frequency band. The frequency band units in the residual power spectral density matrix of each channel with residual power spectral density values ​​lower than the adaptive retention threshold are set to zero, and the frequency band units with values ​​higher than the adaptive retention threshold are retained, thus obtaining the sparsified residual power spectral density matrix of each channel.

[0019] It should be noted that the preset multiplier ranges from [2,3]. The larger the preset multiplier, the stronger the suppression of residual background vibration, but the weaker the ability to retain weak leakage signals. The window length of the sliding time window is determined based on the lowest period frequency of the ship's main engine vibration to cover at least one complete background vibration cycle and ensure the statistical stability of the mean and standard deviation. In an optional embodiment, following the residual power spectral density matrix, a preset multiple of 2.5 is set, and the sliding time window length is 200 milliseconds; within the 100 kHz to 200 kHz frequency band, the mean value of the residual power spectral density within the sliding time window is 0.38 × 10⁻⁶. -4 volt 2 / Hertz, standard deviation is 0.21 × 10 -4 volt 2 / Hz, adaptive retention threshold is 0.38×10 -4 +2.5×0.21×10 -4 =0.905×10 -4 volt 2 / Hertz; The residual power spectral density value of a certain time frame in this frequency band is 1.51×10⁻⁶. -4 volt 2 / Hz is above the adaptive retention threshold and is retained; in the 2kHz low-frequency band, the remaining power spectral density value is 0.66×10 -4 volt 2 / Hertz is below the corresponding adaptive retention threshold, and is set to zero; after all frequency bands are processed in the above manner, the sparsified residual power spectral density matrix of CH3 channel is obtained, in which the frequency band unit where the high frequency leakage acoustic emission energy is concentrated is retained, and the low frequency background vibration residual component is set to zero and suppressed. S2.6: Perform short-time inverse Fourier transform on the sparsified residual power spectral density matrix to restore the sparsified residual power spectral density matrix from the time-frequency domain to the time domain, and obtain the time-domain residual signal of each channel; Furthermore, the short-time Fourier inverse transform uses the same Hanning window function and overlapping step size parameter as the short-time Fourier transform to ensure the consistency of reconstruction between the time-frequency domain transform and the inverse transform, and to eliminate the distortion of the time-domain waveform boundary caused by inconsistent windowing parameters. S2.7: Calculate the short-time energy value based on the time-domain residual signal, and mark the time period when the short-time energy value is lower than the preset silent energy threshold as a silent period; S2.8: Define the time period in which all channels are marked as silent periods within the same time period as the full array silent period, and set the time domain residual signal corresponding to the full array silent period to zero to obtain the de-silenced multi-channel time domain residual signal, where the de-silenced multi-channel time domain residual signal is the pure acoustic emission signal; It should be noted that the determination of the entire array silent segment requires all channels to meet the silent condition simultaneously, rather than a single channel. The preset silent energy threshold is determined based on the lowest frequency band energy statistics of the background vibration power spectrum baseline of each channel under each working condition in the background baseline database. It is updated synchronously with the switching of the current working condition identifier and is consistent with the working condition adaptive mechanism of the reference background vibration power spectrum baseline. In an optional embodiment, following the marking results of the silent segments for each channel, channels CH1, CH2, CH4, CH5, and CH6 are all marked as silent segments within a certain 5-millisecond time period, while the short-time energy value of channel CH3 is 3.8 × 10⁻⁶. -6 volt 2 If the energy level exceeds the preset silence threshold, it is not marked as a silence segment. This time period does not meet the criteria for a full array silence segment, and the corresponding time-domain residual signal is retained without being set to zero. In another 5-millisecond time period, all 6 channels from CH1 to CH6 meet the silence condition. This time period is defined as a full array silence segment, and the corresponding time-domain residual signal is set to zero. Finally, the de-silenced multi-channel time-domain residual signal is obtained, which is the pure acoustic emission signal. The peak value of the pure acoustic emission signal in the time domain of channel CH3 during the retention period is 0.087 volts, which can be used in the subsequent time-frequency feature extraction step of S3.1.

[0020] S3: Extract time-frequency features from the pure acoustic emission signal, calculate the acoustic emission confidence level, and when the acoustic emission confidence level is greater than or equal to the preset acoustic trigger threshold, trigger the thermal imaging camera to enter the high-gain sub-pixel interpolation mode and generate a spatial prior probability distribution map. S3.1: Extract time-frequency features from the pure acoustic emission signals corresponding to each channel, including duration features, energy envelope features, and center frequency features; Preferably, the duration feature is the continuous duration for which the amplitude of the pure acoustic emission signal exceeds the preset impact detection threshold; the energy envelope feature is the square integral value of the pure acoustic emission signal within the time period corresponding to the duration feature; and the center frequency feature is the weighted center frequency value of the power spectrum of the pure acoustic emission signal within the time period corresponding to the duration feature. S3.2: Compare the duration feature, energy envelope feature, and center frequency feature with the preset reference ranges for the duration, energy envelope, and center frequency of the leakage acoustic emission, respectively. Calculate the normalized conformity of each feature with the corresponding reference range, and sum the three normalized conformity values ​​according to preset weighting coefficients to obtain the acoustic emission confidence level. It should be noted that the preset reference range for the duration of leakage acoustic emission is determined based on the statistical distribution of the measured duration of pipeline leakage acoustic emission signals under different orifice diameters and media pressures in controlled leakage experiments. The lower and upper bounds of the statistical interval in which the amplitude of the leakage acoustic emission signal exceeds the preset impact detection threshold are used as the two ends of the reference range. The reference range for the energy envelope of leakage acoustic emission is determined based on the statistical distribution of the square integral value of the pipeline leakage acoustic emission signal in the corresponding time period under each leakage condition in the above controlled leakage experiments. The lower and upper bounds of the statistical value of the energy envelope under different combinations of leakage orifice diameters and media pressures are used as the two ends of the reference range. The reference range for the center frequency of leakage acoustic emission is determined based on the frequency band concentration characteristics generated by the pipeline leakage acoustic emission signal under the turbulent jet excitation mechanism. The lower and upper bounds of the statistical value of the weighted center frequency of the power spectrum under each leakage condition in the controlled leakage experiments are used as the two ends of the reference range. In an optional embodiment, following the time-frequency characteristics of the pure acoustic emission signal, the reference range for the duration of leakage acoustic emission is [0.5 ms, 5.0 ms], and the reference range for the energy envelope of leakage acoustic emission is [1.0 × 10⁻⁶]. -6 volt 2 ·seconds, 50.0×10 -6 volt 2 The leakage sound emission center frequency reference range is [80 kHz, 180 kHz]; the CH3 channel duration characteristic of 1.8 ms falls within [0.5 ms, 5.0 ms], with a normalization compliance of 1.0; the energy envelope characteristic is 4.32 × 10⁻⁶. -6 volt 2• The second falls within the corresponding reference range, and the normalization compliance is 1.0; the center frequency feature of 127.4 kHz falls within [80 kHz, 180 kHz], and the normalization compliance is 1.0; the weighted coefficients of the duration feature are 0.3, the weighted coefficients of the energy envelope feature are 0.4, and the weighted coefficients of the center frequency feature are 0.3. The weighted summation yields an acoustic emission confidence score of 1.0×0.3+1.0×0.4+1.0×0.3=1.0, which is used in the subsequent preset acoustic trigger threshold comparison step. S3.3: Compare the acoustic emission confidence level with the preset acoustic trigger threshold. If the acoustic emission confidence level is less than the preset acoustic trigger threshold, re-acquire the pure acoustic emission signal. If the acoustic emission confidence level is greater than or equal to the preset acoustic trigger threshold, send a mode switching command to the thermal imaging camera. After responding to the mode switching command, the thermal imaging camera extends the integration time to a preset multiple of the standard acquisition mode and performs bicubic interpolation on the current frame thermal image to increase the spatial resolution of the corrected thermal image to a preset multiple of the standard acquisition mode. The thermal imaging camera then enters the high-gain subpixel interpolation mode. Specifically, the preset acoustic trigger threshold is determined based on the statistical distribution of acoustic emission confidence under different leakage orifice diameters and different medium pressures in the controlled leakage experiment. The lower bound of the statistical distribution of acoustic emission confidence corresponding to the real pipeline leakage sample is used as the preset acoustic trigger threshold so that the acoustic emission confidence corresponding to the real pipeline leakage event falls above the preset acoustic trigger threshold with a greater than preset confidence probability. At the same time, the acoustic emission confidence corresponding to background vibration interference and non-leakage acoustic emission events is suppressed below the preset acoustic trigger threshold with a greater than preset confidence probability. In an optional embodiment, following the acoustic emission confidence level of 1.0, the preset acoustic trigger threshold is set to 0.75; when the acoustic emission confidence level of 1.0 is greater than or equal to the preset acoustic trigger threshold of 0.75, a mode switching command is sent to the thermal imaging camera; after the thermal imaging camera responds, the integration time is extended from 1 millisecond in the standard acquisition mode to 4 milliseconds, and bicubic interpolation is performed on the current frame thermal image to improve the spatial resolution of the corrected thermal image from 640×480 pixels to 1280×960 pixels. The thermal imaging camera enters the high-gain subpixel interpolation mode to provide high-resolution thermal image input for subsequent sound source localization and thermal anomaly detection steps; S3.4: Extract the arrival time difference between the pure acoustic emission signals corresponding to any two channels in the acoustic emission sensor array. Using the installation coordinates and arrival time difference of each sensor in the acoustic emission sensor array as input, use the generalized cross-correlation phase transformation algorithm to calculate the time delay estimate corresponding to each channel combination. Construct a hyperbolic equation system using the time delay estimate. Perform least squares solution on the hyperbolic equation system to obtain the sound source localization coordinates. S3.4.1: Based on the pure acoustic emission signal, generate a list of channel combinations by pairing the channel numbers of the acoustic emission sensor array. Each combination in the list consists of two different channel numbers, and the total number of combinations is the binomial combination number of the total number of channels in the acoustic emission sensor array. S3.4.2: For each channel combination in the channel combination list, extract the two pure acoustic emission signals corresponding to the channel combination from the pure acoustic emission signals, and align them according to the acquisition timestamps of the two pure acoustic emission signals to obtain the synchronization signal pair of each channel combination; S3.4.3: Perform Discrete Fourier Transform on the two pure acoustic emission signals in the synchronization signal pair of each channel combination to obtain the first frequency domain signal and the second frequency domain signal of each channel combination; divide the conjugate product of the first frequency domain signal and the second frequency domain signal by the magnitude of the conjugate product of the first frequency domain signal and the second frequency domain signal to obtain the phase transform cross power spectrum of each channel combination. S3.4.4: Perform inverse discrete Fourier transform on the phase transform cross power spectrum of each channel combination to obtain the generalized cross-correlation phase transform function of each channel combination; in the generalized cross-correlation phase transform function, take the time shift corresponding to the maximum function value as the time delay estimate of the channel combination, and obtain the set of time delay estimates of each channel combination. S3.4.5: Using the installation coordinates of each sensor in the acoustic emission sensor array and the set of time delay estimates for each channel combination as input, construct a hyperbolic equation for each channel combination in the channel combination list; It should be noted that the hyperbolic equation uses the location coordinates of the sound source to be determined as unknowns, the difference between the installation coordinates of the two sensors in the channel combination to form the hyperbolic focal coordinate pair, and the product of the estimated time delay of the channel combination and the propagation speed of the sound wave in the medium under test to form the hyperbolic real axis length parameter. The hyperbolic equation sets corresponding to each channel combination are combined into a hyperbolic equation system.

[0021] S3.4.6: Using the sound source location coordinates as the vector to be solved, the least squares method is used to iteratively solve the hyperbola equation system; the geometric center of the installation coordinates of all sensors in the acoustic emission sensor array is used as the initial iteration value of the least squares method; the Euclidean norm of the change in sound source location coordinates between two adjacent iterations is less than the preset location convergence threshold as the iteration termination condition; and the sound source location coordinates after iteration convergence are output. In an optional embodiment, following the pure acoustic emission signal, the acoustic emission sensor arrays CH1 to CH6 are combined in pairs to generate a total of 15 channel combinations. Taking the CH1 and CH3 channel combination as an example, the installation coordinates of CH1 are (0.0 m, 0.0 m) and the installation coordinates of CH3 are (3.0 m, 0.0 m). After performing discrete Fourier transform on the pure acoustic emission signals of the two channels respectively, the phase transform cross-power spectrum is calculated. Performing inverse discrete Fourier transform yields a time shift corresponding to the peak value of the generalized cross-correlation phase transform function of CH1 and CH3 of -0.00130 seconds. The estimated time delay of the channel combination is -0.00130 seconds; the hyperbola real axis length parameter is converted to 6.63 meters based on the propagation speed of sound waves in the metal wall of the steam pipeline of 5100 m / s, and the corresponding hyperbola equation is established; after all 15 channel combinations are processed in the same way, a hyperbola equation system is formed. The geometric center of the installation coordinates of all 6 sensors (2.5 m, 0.0 m) is used as the initial iteration value of the least squares method. After the iteration converges, the sound source positioning coordinates are output as (3.82 m, 0.07 m), which are used in the subsequent spatial prior probability distribution map construction step. S3.5: Using the sound source location coordinates as the mean center and the spatial location uncertainty obtained by converting the time delay estimation error of the generalized cross-correlation phase transformation algorithm through the error propagation formula as the standard deviation, a two-dimensional Gaussian distribution function is constructed. The two-dimensional Gaussian distribution function is then evaluated pixel by pixel in the plane coordinate system of the area to be inspected to obtain the spatial prior probability distribution map. In an optional embodiment, following the sound source location coordinates (3.82 m, 0.07 m), the time delay estimation error of the generalized cross-correlation phase transform algorithm under the current signal-to-noise ratio condition is 0.000025 seconds. The spatial positioning uncertainty is calculated using the error propagation formula to be 0.000025 × 5100 = 0.128 m. Using 0.128 m as the standard deviation of the two-dimensional Gaussian distribution function, the two-dimensional Gaussian distribution function is evaluated pixel-by-pixel in the 8 m × 1 m plane coordinate system of the area to be inspected at a pixel spacing of 6.25 mm to obtain a spatial prior probability distribution map. The probability peak in the spatial prior probability distribution map is located at the pixel corresponding to the sound source location coordinates (3.82 m, 0.07 m). The probability value decreases to 0.607 times the peak value at 0.128 m from the peak pixel and to 0.135 times the peak value at 0.256 m from the peak pixel, for use in the subsequent thermal anomaly search area delineation step of S4.1. S4: The spatial prior probability distribution map is superimposed on the thermal anomaly detection process of the corrected thermal image as a spatial search weight. The local thermal anomaly confidence is calculated to obtain the joint confidence. When the joint confidence is greater than or equal to the preset joint decision threshold, the thermal anomaly contour sequence of the corresponding region is extracted from the corrected thermal image. S4.1: Using the set of pixels whose probability values ​​in the spatial prior probability distribution map are greater than or equal to the preset lower probability threshold as the boundary, the thermal anomaly search area is delineated on the corrected thermal image, and the probability value corresponding to each pixel in the spatial prior probability distribution map is used as the spatial search weight of this pixel to obtain the weighted thermal anomaly search area. S4.2: Based on the weighted thermal anomaly search region, the local temperature deviation value of each pixel is obtained by subtracting the weighted average value of the inversion temperature values ​​of all pixels in the thermal anomaly search region from the inversion temperature value of each pixel in the corrected thermal image. The local temperature deviation value of each pixel is then divided by the weighted standard deviation of the inversion temperature values ​​in the thermal anomaly search region to obtain the normalized temperature deviation value of each pixel. S4.3: Multiply the normalized temperature deviation value with the spatial search weight of the corresponding pixel in the spatial prior probability distribution map pixel by pixel to obtain the weighted temperature anomaly response value of each pixel. Calculate the average of the weighted temperature anomaly response values ​​of all pixels in the thermal anomaly search area to obtain the local thermal anomaly confidence level. S4.4: Obtain the current ship operating condition parameters. Using the current ship operating condition parameters as input, look up the acoustic emission channel weight coefficient and thermal imaging channel weight coefficient according to the preset operating condition-weight mapping table. Add the product of the acoustic emission confidence and the acoustic emission channel weight coefficient to the product of the local thermal anomaly confidence and the thermal imaging channel weight coefficient to obtain the joint confidence. The current ship operating condition parameters include main engine speed, speed and cabin airflow speed. In an optional embodiment, following the acoustic emission confidence level of 1.0 and the local thermal anomaly confidence level of 0.83, the current ship operating parameters are: main engine speed 78% of rated speed, speed 13 knots, and cabin airflow speed 0.6 m / s; according to the preset operating condition-weight mapping table, under this operating condition combination, the acoustic emission channel weight coefficient is 0.55 and the thermal imaging channel weight coefficient is 0.45; the joint confidence level is 1.0×0.55+0.83×0.45=0.924, which is used in the subsequent preset joint decision threshold comparison step; S4.5: Compare the joint confidence score with the preset joint decision threshold. If the joint confidence score is less than the preset joint decision threshold, reacquire the pure acoustic emission signal and reset the thermal imaging camera to the standard acquisition mode. If the joint confidence score is greater than or equal to the preset joint decision threshold, within the thermal anomaly search area of ​​the corrected thermal image, perform a contour tracking algorithm on the set of connected pixels whose temperature deviation normalization value exceeds the preset temperature anomaly segmentation threshold, extract the outer boundary coordinate sequence of the set of connected pixels in multiple consecutive frames of corrected thermal images, and obtain the thermal anomaly contour sequence. Furthermore, the preset joint decision threshold is determined based on the statistical distribution of joint confidence under different leakage apertures, different medium pressures, and different ship operating conditions in the controlled leakage experiment. The lower bound of the statistical distribution of joint confidence corresponding to the real pipeline leakage sample is used as the preset joint decision threshold so that the joint confidence corresponding to the real pipeline leakage event falls above the preset joint decision threshold with a confidence probability greater than the preset threshold. At the same time, the joint confidence corresponding to non-leakage interference events for which neither acoustic emission confidence nor local thermal anomaly confidence exceeds the limit is suppressed below the preset joint decision threshold with a confidence probability greater than the preset threshold. In an optional embodiment, following the joint confidence level of 0.924, the preset joint decision threshold is set to 0.70; if the joint confidence level of 0.924 is greater than or equal to the preset joint decision threshold of 0.70, the preset temperature anomaly segmentation threshold is set to 2.0 within the thermal anomaly search area of ​​the corrected thermal image, and a set of 17 connected pixels with a temperature deviation normalization value exceeding 2.0 is selected. A contour tracking algorithm is then performed on this set of connected pixels to extract the outer boundary coordinate sequence corresponding to the first frame; the subsequent 12 frames of high-gain subpixel interpolated thermal images are continuously acquired and processed frame by frame in the same manner to obtain a total of 13 frames of thermal anomaly contour sequence, which is then used in the subsequent temperature centroid coordinate extraction step of S5.1; S5: Fit the spatiotemporal evolution curve of the thermal plume for multiple consecutive frames of the thermal anomaly profile sequence, and judge the fitting residual with the preset physical rationality threshold. When the fitting residual is less than the preset physical rationality threshold, the thermal anomaly profile sequence is reversed by time inversion of the spatiotemporal evolution curve, and the location coordinates of the leak point and the leak confirmation result are output. Otherwise, the thermal anomaly profile sequence is determined to be a non-leakage interference source, and the pure acoustic emission signal is re-acquired. The thermal imaging camera is reset to the standard acquisition mode and the pure acoustic emission signal is re-acquired. S5.1: For each frame of thermal anomaly contour in the thermal anomaly contour sequence, extract the area weighted centroid coordinates of the apparent temperature values ​​of all pixels within the contour to obtain the temperature centroid coordinates of each frame of thermal anomaly contour, and arrange the temperature centroid coordinates of all frames in the thermal anomaly contour sequence according to the frame timestamp order to obtain the temperature centroid coordinate time series. It should be noted that the temperature barycentric coordinate time series describes the spatial drift trajectory of the thermal plume within the pixel coordinate system of the calibrated thermal image over time. The temperature barycentric coordinates are calculated using an area-weighted method, with the ratio of the apparent temperature value of each pixel to the sum of the apparent temperature values ​​of all pixels within the thermal anomaly contour as the weight of each pixel. This biases the temperature barycentric coordinates towards the region with the highest temperature in the thermal plume, thereby more accurately tracking the location of the thermal energy concentration in the thermal plume. The temperature barycentric coordinate time series provides spatial trajectory input for subsequent fitting of the spatiotemporal evolution curve of the thermal plume, and together with the frame timestamps of each frame in the thermal anomaly contour sequence, they form a spatiotemporal data pair for fitting. S5.2: For each frame of the thermal anomaly contour sequence, extract the maximum value of the apparent temperature of all pixels within the contour to obtain the peak temperature of each frame's thermal anomaly contour; for each frame of the thermal anomaly contour sequence, calculate the ratio of the total number of thermal candidate pixels within the contour to the total number of pixels in a single frame of the high-gain subpixel interpolated heat map frame sequence to obtain the thermal anomaly contour area ratio of each frame, and arrange the peak temperature and thermal anomaly contour area ratio in the order of frame timestamps to obtain the peak temperature time series and contour area ratio time series. Preferably, the peak temperature time series, the profile area ratio time series, and the temperature centroid coordinate time series together constitute a multidimensional set of observations of the spatiotemporal evolution of the thermal plume; the trend of the change of the peak temperature of the thermal anomaly profile over time reflects the persistence of the thermal energy input of the leaking thermal plume: the peak temperature of the thermal plume generated by the actual pipeline leak shows a stable or slow upward trend during the continuous leakage stage, while the peak temperature of the transient thermal shock interference source on the pipeline surface shows a rapid decay trend; the profile area ratio time series reflects the spatial diffusion behavior of the thermal plume, the profile area of ​​the actual leaking thermal plume monotonically expands over time, while the profile area of ​​the transient thermal interference source rapidly contracts in a short period of time; S5.3: Based on the Gaussian diffusion model of leaking heat plume in fluid thermodynamics, a spatiotemporal evolution parameterization model of heat plume is established. The spatiotemporal evolution parameterization model of heat plume uses the pixel coordinates of the leakage point, the leakage start time, the heat plume diffusion coefficient and the initial peak temperature of the heat plume as parameters to be estimated, and the frame timestamp as the independent variable. It outputs the centroid coordinates of the predicted temperature of each frame, the predicted peak temperature of each frame and the predicted contour area ratio of each frame, respectively. Furthermore, in the parameterized model of the spatiotemporal evolution of the thermal plume, the centroid coordinates of the predicted temperature in each frame are calculated by the pixel coordinates of the leakage point along the preset plume drift direction using the thermal plume diffusion coefficient and the frame timestamp. The predicted peak temperature in each frame is calculated by the decay function of the initial peak temperature of the thermal plume, the thermal plume diffusion coefficient, and the frame timestamp. The predicted contour area ratio in each frame is calculated by the diffusion area function of the thermal plume diffusion coefficient and the frame timestamp. Furthermore, the preset plume drift direction is determined jointly based on the pipeline layout direction in the area to be inspected and the airflow direction of the cabin ventilation. The pipeline layout direction is extracted from the pre-stored 3D pipeline model according to the corresponding pipe segment identifier, and the airflow direction of the cabin ventilation is collected in real time by the airflow sensors deployed in the area to be inspected. The thermal plume spatiotemporal evolution parameterization model upgrades the leakage detection from static pixel temperature comparison of a single frame heat map to multi-frame temporal behavior verification based on physical diffusion laws. Non-leakage thermal interference sources that do not conform to the thermal plume diffusion laws are eliminated through physical model constraints. S5.4: Using the time series of temperature centroid coordinates, peak temperature, and contour area ratio as observations, and the output of the thermal plume spatiotemporal evolution parameterization model as the predictor, the nonlinear least squares method is used to iteratively solve the four parameters to be estimated in the thermal plume spatiotemporal evolution parameterization model: the pixel coordinates of the leakage point, the leakage start time, the thermal plume diffusion coefficient, and the initial peak temperature of the thermal plume. Specifically, the projection coordinates of the sound source coordinates in the pixel coordinate system are used as the initial iteration values ​​of the leak point pixel coordinates, the reference timestamp is used as the initial iteration value of the leak start time, and the thermal diffusion coefficient of the corresponding medium type in the pipeline medium lookup table is used as the initial iteration value of the thermal plume diffusion coefficient. The thermal plume spatiotemporal evolution curve fitting results are obtained, where the thermal plume spatiotemporal evolution curve fitting results include the optimal estimates of each parameter to be estimated. Furthermore, the iteration termination condition of the nonlinear least squares method is that the norm of the change of each parameter to be estimated between two adjacent iterations is less than the preset iteration convergence threshold, or the number of iterations exceeds the preset maximum number of iterations; the design of using the projected coordinates of the sound source coordinates as the initial iteration values ​​of the leak point pixel coordinates introduces the spatial prior information of acoustic localization into the fitting initialization process, shortens the convergence path of the nonlinear iteration, and reduces the probability of getting trapped in local minima due to the initial values ​​being far away from the real leak point coordinates, reflecting the deep synergy between acoustic and thermal imaging information in the fitting stage; In an optional embodiment, following the time series of temperature centroid coordinates, peak temperature, and contour area ratio, the projected coordinates of the sound source location coordinates (3.82 meters, 0.07 meters) in the pixel coordinate system are used as the initial iteration value of the leak point pixel coordinates. The timestamp of the first frame is used as the initial iteration value of the leak start time. The thermal diffusivity of the steam medium in the pipeline medium lookup table is 2.1 × 10⁻⁶. -5 rice 2 / second was used as the initial iterative value for the thermal plume diffusion coefficient. Through iterative solving using the nonlinear least squares method, the norm of the changes in each parameter to be estimated decreased below the preset iterative convergence threshold in the 14th iteration. The converged output was the fitting result of the thermal plume spatiotemporal evolution curve: the optimal estimated value for the leak point pixel coordinates was (614 pixels, 487 pixels), the optimal estimated value for the leak start time was 0.3 seconds before the timestamp of the first frame, and the optimal estimated value for the thermal plume diffusion coefficient was 2.6 × 10⁻⁶. -5 rice 2 / second, the optimal estimate of the initial peak temperature of the thermal plume is 47.3 degrees Celsius, which can be used in the subsequent residual calculation step of S5.5; S5.5: Based on the fitting results of the spatiotemporal evolution curve of the thermal plume, calculate the predicted temperature centroid coordinates, predicted peak temperature, and predicted contour area ratio of the parameterized model of the spatiotemporal evolution of the thermal plume at each frame timestamp; Furthermore, the centroid coordinate residual is obtained by subtracting the predicted temperature centroid coordinates of each frame from the corresponding temperature centroid coordinates in the temperature centroid coordinate time series; the peak temperature residual is obtained by subtracting the predicted peak temperature of each frame from the peak temperature of the corresponding thermal anomaly contour in the peak temperature time series; and the area ratio residual is obtained by subtracting the predicted contour area ratio of each frame from the corresponding thermal anomaly contour area ratio in the contour area ratio time series. The centroid coordinate residual, peak temperature residual, and area ratio residual are then weighted and summed according to a preset residual fusion weighting coefficient, and the root mean square is taken to obtain the fitting residual. It should be noted that the weighting coefficients of the centroid coordinate residual, the peak temperature residual, and the area ratio residual in the preset residual fusion weighting coefficients are determined after the physical dimensions and numerical magnitudes of the three types of observations are uniformly normalized to ensure that the contribution ratio of the three types of residuals in the fitting residuals matches their physical importance. The fitting residuals comprehensively reflect the overall degree of conformity between the spatiotemporal evolution behavior of the thermal anomaly profile sequence and the physical diffusion law described by the spatiotemporal evolution parameterized model of the thermal plume, and are the only input quantity for subsequent physical rationality judgment. S5.6: Compare the fitting residuals with a preset physical plausibility threshold, specifically including: When the fitting residual is less than the preset physical rationality threshold, the spatiotemporal evolution behavior of the thermal anomaly profile sequence is determined to conform to the physical diffusion law of thermal plume, and step S5.7 is executed. When the fitting residual is greater than or equal to the preset physical rationality threshold, the thermal anomaly contour sequence is determined to be a non-leakage interference source. A reset command is sent to the thermal imaging camera. After receiving the reset command, the thermal imaging camera switches the frame acquisition frequency from high-speed frame rate back to standard frame rate, resets the analog gain multiple to the standard gain multiple in standard acquisition mode, stops the spatial resolution enhancement processing of the sub-pixel interpolation algorithm, clears the cached data of the corresponding index in the thermal anomaly contour cache and prior map cache queue, and re-acquires the pure acoustic emission signal from the set of de-silenced multi-channel temporal residual signals to enter the next detection cycle. The preset physical rationality threshold is determined based on the upper bound of the statistical distribution of the fitting residuals under different leakage orifice diameters and different medium pressures in the controlled leakage experiment, so that the fitting residuals of the real leakage samples fall below the preset physical rationality threshold with a greater than preset confidence probability; the determination of non-leakage interference sources triggers the thermal imaging camera to reset to the standard acquisition mode instead of directly terminating the detection, ensuring that the detection system can automatically enter the next detection cycle without manual intervention after excluding false triggering events; In an optional embodiment, following the fitting residual, the preset physical rationality threshold is set to 0.35 based on the upper bound of the statistical distribution of the controlled leakage experiment; the current fitting residual of the thermal anomaly profile sequence is 0.18, which is less than the preset physical rationality threshold of 0.35, and it is determined that the spatiotemporal evolution behavior of the thermal anomaly profile sequence conforms to the physical diffusion law of the thermal plume, and step S5.7 is executed; if the fitting residual is 0.52, it is greater than or equal to the preset physical rationality threshold of 0.35, and it is determined that the thermal anomaly profile sequence is a non-leakage interference source, a reset command is sent to the thermal imaging camera, the cache data of the corresponding index in the thermal anomaly profile cache and the prior map cache queue are cleared, and the pure acoustic emission signal is re-acquired from the set of de-silenced multi-channel temporal residual signals, and the next detection cycle is entered; S5.7: Based on the optimal estimate of the leakage initiation time in the fitting results of the thermal plume spatiotemporal evolution curve, perform time inversion on the parameterized model of the thermal plume spatiotemporal evolution, specifically including: Time inversion involves progressively backtracking the temporal independent variable of the thermal plume spatiotemporal evolution parameterized model from the current frame timestamp towards the optimal estimate at the leak initiation time. The predicted temperature barycentric coordinates output by the thermal plume spatiotemporal evolution parameterized model are calculated at each backtracking time, resulting in a time-inverted temperature barycentric coordinate sequence. The predicted temperature barycentric coordinates corresponding to the point where the temporal independent variable in the time-inverted temperature barycentric coordinate sequence equals the optimal estimate at the leak initiation time are defined as the leak point heatmap pixel coordinates. The physical basis of time inversion is that the spatial position of the heat plume at the moment of leakage is the position of the heat radiation source of the leakage point. By pushing the spatiotemporal evolution parameterized model of the heat plume back to the moment of leakage on the time axis, the spatial drift of the heat plume caused by the influence of the cabin ventilation airflow and the heat conduction of the pipe wall is eliminated, thereby tracing the observable heat plume diffusion center position back to the position of the heat radiation source of the actual leakage point. S5.8: Based on the pre-stored thermal imaging camera extrinsic matrix and pipeline 3D model, the pixel coordinates of the leak point thermal image are back-projected from the pixel coordinate system of the thermal imaging camera to the pipeline 3D coordinate system to obtain the 3D coordinates of the leak point; the 3D coordinates of the leak point are spatially matched with the 3D coordinate range of each pipe segment in the pipeline 3D model to obtain the pipe segment identifier corresponding to the 3D coordinates of the leak point; the 3D coordinates of the leak point, the pipe segment identifier, the optimal estimate of the thermal plume diffusion coefficient in the fitting result of the thermal plume spatiotemporal evolution curve, and the reference timestamp are encapsulated together as the leak confirmation result, and together with the 3D coordinates of the leak point as the location coordinates of the leak point, the result is output to the ship pipeline monitoring system, and the thermal imaging camera is triggered to reset to the standard acquisition mode, the thermal anomaly profile cache and the prior image cache queue are cleared, and the next detection cycle begins; It should be noted that the ratio of the optimal estimate of the thermal plume diffusion coefficient to the reference value of the thermal diffusivity coefficient for the corresponding medium type in the pipeline medium lookup table reflects the relative magnitude of the leakage. The larger the optimal estimate of the thermal plume diffusion coefficient, the faster the leakage thermal plume diffuses, and the larger the corresponding leakage. Incorporating the optimal estimate of the thermal plume diffusion coefficient into the leakage confirmation results allows the ship's pipeline monitoring system to obtain a quantitative estimate of the leakage severity at the same time as receiving the location coordinates of the leak point, providing a basis for subsequent maintenance priority ranking. In an optional embodiment, following the leak point's thermal image pixel coordinates (614 pixels, 487 pixels), these pixel coordinates are back-projected onto the pipeline's three-dimensional coordinate system based on a pre-stored thermal imaging camera extrinsic matrix, yielding the leak point's three-dimensional coordinates as (3.79 m, 0.06 m, 1.43 m). The leak point's three-dimensional coordinates are then spatially matched with the three-dimensional coordinate ranges of each pipe segment in the pipeline's three-dimensional model, identifying the corresponding pipe segment as the 3rd segment of the high-pressure steam main pipe. The leak point's three-dimensional coordinates (3.79 m, 0.06 m, 1.43 m), the pipe segment identification "3rd segment of the high-pressure steam main pipe," and the optimal estimated value of the thermal plume diffusion coefficient (2.6 × 10⁻⁶) are then considered. -5 rice 2 The leak confirmation result is encapsulated together with the reference timestamp and the leak point's three-dimensional coordinates. It is then output to the ship's pipeline monitoring system, triggering the thermal imaging camera to reset to standard acquisition mode, clearing the thermal anomaly profile cache and prior image cache queue, and entering the next detection cycle.

[0022] In summary, this invention utilizes an acoustic emission sensor array and a thermal imaging camera to collaboratively acquire cabin environmental temperature parameters, relative humidity parameters, and infrared radiation information from the piping area. Gain bias correction and atmospheric transmittance correction are applied to the original radiation values ​​to obtain a corrected thermal image. Simultaneously, adaptive background vibration elimination is performed on the acoustic emission signal to obtain a pure acoustic emission signal. This achieves simultaneous suppression of infrared measurement errors and equipment vibration noise in the complex environment of ship cabins, making the thermal image temperature inversion results more realistic and reliable, and effectively eliminating the interference of mechanical vibration on the acoustic signal. By extracting time-frequency features from the pure acoustic emission signal and calculating the acoustic emission confidence level, it achieves a higher confidence level than... When a preset acoustic trigger threshold is reached, the thermal imaging camera is triggered to enter a high-gain subpixel interpolation mode and generate a spatial prior probability distribution map. This achieves adaptive enhancement of the thermal imaging detection process using acoustic anomalies as trigger conditions. The system can improve the spatial resolution of thermal imaging when a suspected leak occurs, and utilizes the sound source localization results to form a probability distribution to limit the thermal anomaly search area, thereby improving the observability of weak leak thermal signals and reducing unnecessary computational burden. Furthermore, by superimposing the spatial prior probability distribution map as spatial weights into the thermal anomaly detection process of the corrected thermal image, and combining the confidence levels of local thermal anomalies and acoustic emission to construct a joint confidence level for threshold determination, the system further enhances the detection capabilities of the thermal imaging camera. This system achieves the fusion of acoustic and thermal information, enabling leak identification to move beyond relying on a single temperature anomaly. Instead, it performs a comprehensive judgment under spatial probability constraints and multi-source information weighting, effectively reducing the risk of false detections caused by ambient temperature fluctuations and equipment thermal radiation, and improving the accuracy of detection results. In the detection confirmation stage, by extracting connected pixel sets whose temperature deviation normalization values ​​exceed a threshold when the joint confidence level meets the decision criteria and performing contour tracking, a continuous multi-frame thermal anomaly contour sequence is obtained. This achieves a continuous temporal description of the spatial boundary of the leak thermal plume, expanding the detection results from single-frame temperature anomalies to stable spatiotemporal contour information, thereby reducing... The study investigates the impact of transient noise on detection results and provides reliable data for leak diffusion behavior analysis. By fitting multi-frame spatiotemporal evolution curves of thermal plumes to the thermal anomaly profile sequence and judging based on the fitting residuals and preset physical rationality thresholds, the leak source is reversed using time inversion when the conditions are met; otherwise, it is judged as a non-leakage interference source and the signal is reacquired. This achieves physical consistency verification of whether the thermal anomaly phenomenon conforms to the real leak diffusion law, and can accurately output the location coordinates of the leak point when a leak is confirmed, thereby effectively eliminating interference from non-leakage heat sources and achieving accurate positioning, ultimately improving the reliability and positioning accuracy of ship pipeline leak detection.

[0023] This embodiment also provides a computer device applicable to the automatic detection method for ship pipeline leaks based on thermal imaging machine vision, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the automatic detection method for ship pipeline leaks based on thermal imaging machine vision as proposed in the above embodiment.

[0024] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0025] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the automatic detection method for ship pipeline leaks based on thermal imaging machine vision as proposed in the above embodiments.

[0026] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0027] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An automatic detection method for ship pipeline leaks based on thermal imaging machine vision, characterized in that: include, The ambient temperature and relative humidity parameters of the cabin environment in the pipeline inspection area of ​​the ship are collected by an acoustic emission sensor array and a thermal imaging camera. Gain bias correction and atmospheric transmittance correction are performed on the raw radiation value output by the thermal imaging camera to obtain a corrected thermal image. Adaptive background vibration elimination is performed on the acoustic emission signals collected by each channel of the acoustic emission sensor array to obtain pure acoustic emission signals; Extract time-frequency features from the pure acoustic emission signal, calculate the acoustic emission confidence level, and when the acoustic emission confidence level is greater than or equal to a preset acoustic trigger threshold, trigger the thermal imaging camera to enter the high-gain sub-pixel interpolation mode to generate a spatial prior probability distribution map. The spatial prior probability distribution map is superimposed on the thermal anomaly detection process of the corrected thermal image as a spatial search weight. The local thermal anomaly confidence is calculated to obtain the joint confidence. When the joint confidence is greater than or equal to the preset joint decision threshold, the thermal anomaly contour sequence of the corresponding region is extracted from the corrected thermal image. The thermal anomaly profile sequence is fitted with a multi-frame spatiotemporal evolution curve of the thermal plume, and the fitting residual is compared with a preset physical rationality threshold. When the fitting residual is less than the preset physical rationality threshold, the thermal anomaly profile sequence is reverse-sourced using the time inversion of the spatiotemporal evolution curve, and the location coordinates of the leak point and the leak confirmation result are output. Otherwise, the thermal anomaly profile sequence is determined to be a non-leakage interference source, and the pure acoustic emission signal is reacquired. The thermal imaging camera is reset to the standard acquisition mode, and the pure acoustic emission signal is reacquired.

2. The automatic detection method for ship pipeline leaks based on thermal imaging machine vision as described in claim 1, characterized in that: The method for obtaining the thermal anomaly contour sequence is as follows: The joint confidence level is compared with a preset joint decision threshold. When the joint confidence level is less than the preset joint decision threshold, the pure acoustic emission signal is reacquired and the thermal imaging camera is reset to the standard acquisition mode. When the joint confidence is greater than or equal to the preset joint decision threshold, a contour tracking algorithm is executed on the set of connected pixels whose temperature deviation normalization value exceeds the preset temperature anomaly segmentation threshold within the thermal anomaly search area of ​​the corrected thermal image, and the outer boundary coordinate sequence of the set of connected pixels in multiple consecutive frames of the corrected thermal image is extracted to obtain the thermal anomaly contour sequence.

3. The automatic detection method for ship pipeline leaks based on thermal imaging machine vision as described in claim 2, characterized in that: The method for obtaining the joint confidence level is as follows: Obtain the current ship operating condition parameters. Using the current ship operating condition parameters as input, look up the acoustic emission channel weight coefficient and the thermal imaging channel weight coefficient according to the preset operating condition-weight mapping table. Add the product of the acoustic emission confidence score and the acoustic emission channel weight coefficient to the product of the local thermal anomaly confidence score and the thermal imaging channel weight coefficient to obtain the joint confidence score.

4. The automatic detection method for ship pipeline leaks based on thermal imaging machine vision as described in claim 3, characterized in that: The confidence level of the local thermal anomaly includes: Using the set of pixels whose probability values ​​in the spatial prior probability distribution map are greater than or equal to a preset lower probability threshold as the boundary, a thermal anomaly search region is defined on the corrected thermal image, and the probability value of each pixel in the spatial prior probability distribution map is used as the spatial search weight of this pixel to obtain a weighted thermal anomaly search region. Based on the weighted thermal anomaly search region, the local temperature deviation value of each pixel is obtained by subtracting the weighted average of the inversion temperature values ​​of all pixels in the thermal anomaly search region from the inversion temperature value of each pixel in the corrected thermal image. The local temperature deviation value of each pixel is then divided by the weighted standard deviation of the inversion temperature values ​​in the thermal anomaly search region to obtain the normalized temperature deviation value of each pixel. The normalized temperature deviation value is multiplied pixel by pixel by the spatial search weight of the corresponding pixel in the spatial prior probability distribution map to obtain the weighted temperature anomaly response value of each pixel. The average of the weighted temperature anomaly response values ​​of all pixels in the thermal anomaly search area is calculated to obtain the local thermal anomaly confidence level.

5. The automatic detection method for ship pipeline leaks based on thermal imaging machine vision as described in claim 4, characterized in that: The method for generating the spatial prior probability distribution map is as follows: The acoustic emission confidence level is compared with a preset acoustic trigger threshold. If the acoustic emission confidence level is less than the preset acoustic trigger threshold, a clean acoustic emission signal is reacquired. If the acoustic emission confidence level is greater than or equal to the preset acoustic trigger threshold, a mode switching command is sent to the thermal imaging camera. After responding to the mode switching command, the thermal imaging camera extends the integration time to a preset multiple of the standard acquisition mode and performs bicubic interpolation on the current frame thermal image to increase the spatial resolution of the corrected thermal image to a preset multiple of the standard acquisition mode. The thermal imaging camera then enters a high-gain subpixel interpolation mode. Extract the arrival time difference between the pure acoustic emission signals corresponding to any two channels in the acoustic emission sensor array. Using the installation coordinates of each sensor in the acoustic emission sensor array and the arrival time difference as input, use the generalized cross-correlation phase transformation algorithm to calculate the time delay estimate corresponding to each channel combination. Construct a hyperbolic equation system using the time delay estimate. Perform least squares solution on the hyperbolic equation system to obtain the sound source localization coordinates. Using the sound source location coordinates as the mean center and the spatial location uncertainty obtained by converting the time delay estimation error of the generalized cross-correlation phase transformation algorithm through the error propagation formula as the standard deviation, a two-dimensional Gaussian distribution function is constructed. The two-dimensional Gaussian distribution function is then evaluated pixel by pixel in the plane coordinate system of the area to be inspected to obtain a spatial prior probability distribution map.

6. The automatic detection method for ship pipeline leaks based on thermal imaging machine vision as described in claim 5, characterized in that: The method for obtaining the corrected thermal image is as follows: An acoustic emission sensor array and a thermal imaging camera are arranged in the pipeline inspection area of ​​the ship, and a temperature and humidity sensor is installed near the lens of the thermal imaging camera to collect the cabin environment temperature and relative humidity parameters of the inspection area. Using the cabin ambient temperature parameter as the independent variable, two-point radiometric calibration is performed on each pixel in the focal plane array of the thermal imaging camera to obtain the gain bias correction radiometric value matrix. Based on Planck's radiation law, the original radiation value of each pixel output by the thermal imaging camera is multiplied by the gain correction coefficient and superimposed with the offset correction coefficient to obtain the gain-bias-corrected radiation value of each pixel. Using the relative humidity parameter and the optical path distance from the thermal imaging camera to the surface of the pipe in the area to be inspected as inputs, the atmospheric transmittance corresponding to the current frame is calculated using the infrared band atmospheric radiation attenuation model. Divide the gain bias corrected radiance value of each pixel in the gain bias corrected radiance value matrix by the atmospheric transmittance to obtain the atmospheric corrected radiance value matrix. Based on Planck's radiation law and preset pipeline surface emissivity parameters, the atmospheric correction radiation value of each pixel in the atmospheric correction radiation value matrix is ​​converted into the corresponding apparent temperature value to obtain a corrected thermal image.

7. The automatic detection method for ship pipeline leaks based on thermal imaging machine vision as described in claim 5, characterized in that: The method for obtaining the pure acoustic emission signal is as follows: Read the original acoustic emission signals of each channel within the current time window from each channel acquisition unit of the acoustic emission sensor array; The original acoustic emission signal is subjected to short-time Fourier transform according to a preset frequency band division method. The power spectral density distribution of each channel in each frequency band is extracted to obtain the background vibration power spectrum baseline. The background vibration power spectrum baseline is stored in the background baseline database with the channel number and the center frequency of the frequency band as dual indices. Based on the current operating condition identifier of the actual ship inspection process, the background vibration power spectrum baseline is used as the reference background vibration power spectrum baseline for the current inspection cycle. If the current operating condition identifier does not completely match the existing operating condition categories in the background baseline database, the background vibration power spectrum baselines corresponding to the two operating condition categories in the background baseline database that have the closest Euclidean distance to the current operating condition identifier are linearly interpolated according to the operating condition similarity weight to obtain the interpolated background vibration power spectrum baseline. The interpolated background vibration power spectrum baseline is then used to replace the reference background vibration power spectrum baseline in subsequent calculations.

8. The automatic detection method for ship pipeline leaks based on thermal imaging machine vision as described in claim 7, characterized in that: It also includes, The original acoustic emission signal is subjected to a short-time Fourier transform in the same preset frequency band division method as the short-time Fourier transform to obtain the real-time power spectral density matrix of each channel. The difference between the real-time power spectral density value and the baseline power spectral density value of the corresponding channel and frequency band in the interpolated background vibration power spectral baseline is obtained to obtain the remaining power spectral density matrix. An adaptive threshold decision is performed on the residual power spectral density values ​​of each frequency band in the residual power spectral density matrix to obtain the sparsified residual power spectral density matrix of each channel. Perform a short-time inverse Fourier transform on the sparsed residual power spectral density matrix to restore the sparsed residual power spectral density matrix from the time-frequency domain to the time domain, and obtain the time-domain residual signal of each channel; Calculate the short-time energy value based on the time-domain residual signal, and mark the time period when the short-time energy value is lower than the preset silent energy threshold as a silent period; The time period in which all channels are marked as silent segments within the same time period is defined as the full array silent segment. The time domain residual signal corresponding to the full array silent segment is set to zero to obtain the de-silenced multi-channel time domain residual signal, which is the pure acoustic emission signal.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the automatic detection method for ship pipeline leakage based on thermal imaging machine vision as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the automatic detection method for ship pipeline leakage based on thermal imaging machine vision as described in any one of claims 1 to 8.

Images