A method and system for detecting container corrosion by magnetic flux leakage based on machine learning

Abstract

The present application relates to the technical field of container corrosion detection, and discloses a container corrosion magnetic leakage detection method and system based on machine learning. Real-time magnetic leakage signals and lift-off value change data of container corrosion are obtained from a sensor array, and filtered signal sequences are obtained through adaptive filtering. A lift-off distortion identification model is called to identify distortion components and influence degree, and when the threshold is exceeded, a compensation model is used to output a compensation amount for correction, obtaining a compensated sequence. For the compensated sequence, material magnetic difference data of the detection area is obtained, compared with the preset material magnetic database reference value, the signal reference stability is evaluated, and after extracting the deviation characteristics, the calibrated sequence is obtained through dynamic calibration. The corrosion depth characteristics are extracted, and if there is deviation, iterative optimization is performed, and the final corrosion depth value and corrosion defect position information are output after analyzing the optimized sequence. The method adapts to the characteristics of the magnetic leakage signal, accurately eliminates multi-source interference, improves detection reliability, and provides high-quality data support for container safety maintenance and corrosion evaluation.

Description

Technical Field

[0001] This invention relates to the field of container corrosion detection technology, and in particular to a container corrosion magnetic flux leakage detection method and system based on machine learning. Background Technology

[0002] Container corrosion detection is a core area for ensuring the safe operation of industrial equipment, especially in industries such as petrochemicals, shipping, and tank management. Accurately detecting the degree of corrosion on container walls directly affects equipment lifespan and operational safety. Magnetic flux leakage (MFL) testing, as a mainstream non-destructive testing technology, is widely used in this field due to its high efficiency and non-contact nature. However, its accuracy and reliability are limited by complex working environments. The core of research in this area lies in improving the accuracy of detection results to avoid equipment failure risks caused by misjudgments—such as safety accidents caused by corrosion leaks in petrochemical storage tanks, or unnecessary costs incurred due to over-maintenance. Therefore, optimizing MFL testing technology and overcoming environmental and material interference bottlenecks have become important issues in the field of industrial equipment safety monitoring.

[0003] Current methods for detecting magnetic flux leakage (MF) corrosion in containers have significant limitations. Traditional techniques often rely on fixed calibration parameters, which cannot adapt to dynamic changes caused by environmental or material variations during the testing process. The distance between the testing equipment and the container surface (lift-off value) can fluctuate due to operational deviations or equipment vibrations, directly leading to distortion in the intensity and waveform of the MF leakage signal. Simultaneously, differences in the magnetic properties of the container material (such as fluctuations in the permeability of different batches of steel) and temperature variations in the working environment further interfere with the stability of the detection signal. In complex industrial environments, these problems are compounded and amplified. For example, when inspecting large storage tanks, uneven surfaces can easily cause fluctuations in lift-off values, and the same area may exhibit different signal characteristics due to differences in material magnetic properties. This can lead to the detection system misinterpreting interference as corrosion defects or overlooking minor corrosion. Improved methods that involve manually adjusting compensation parameters also struggle to balance signal fidelity and interference suppression, failing to meet the demands of high-precision testing.

[0004] Therefore, how to dynamically calibrate the magnetic flux leakage signal under complex and variable detection conditions to eliminate deviations caused by changes in lift-off values ​​and differences in material magnetic properties has become a key issue in improving the accuracy of magnetic flux leakage detection for container corrosion. Solving this problem not only concerns the reliability upgrade of magnetic flux leakage detection technology itself, but also directly affects the safety, stability, and economic benefits of industrial production, and is of great significance for promoting the large-scale application of non-destructive testing technology in the field of industrial equipment monitoring. Summary of the Invention

[0005] This invention provides a machine learning-based method and system for detecting magnetic flux leakage in container corrosion, which adapts to the characteristics of magnetic flux leakage signals, accurately eliminates multi-source interference, improves detection reliability, and provides high-quality data support for container safety maintenance and corrosion assessment.

[0006] Firstly, to address the aforementioned technical problems, this invention provides a machine learning-based method for detecting magnetic flux leakage in container corrosion, comprising: Real-time magnetic flux leakage signal data and corresponding lift-off value change data of container corrosion detection are acquired from the sensor array, and environmental noise is filtered out by an adaptive filtering algorithm to obtain the filtered signal intensity sequence. Based on the filtered signal strength sequence, a pre-built machine learning model for lift-off distortion identification is invoked to identify signal distortion components caused by changes in lift-off value, and the degree of influence of the signal distortion components is calculated. When the degree of influence exceeds the preset influence threshold, a pre-built machine learning model for outputting compensation is invoked. The filtered signal strength sequence and the signal distortion components are input to obtain the corresponding compensation amount. Based on the compensation amount, the filtered signal strength sequence is corrected and adjusted to obtain the compensated signal strength sequence. For the compensated signal intensity sequence, the material magnetic difference data of the detection area is obtained and compared with the reference magnetic value in the preset material magnetic database. The stability of the signal reference is evaluated by the comparison results, and the signal stability evaluation result is obtained. The signal deviation features caused by material magnetic differences are extracted from the signal stability assessment results. The signal deviation features are then fused with the compensated signal intensity sequence using a dynamic calibration algorithm to obtain the calibrated signal sequence. Corrosion depth-related feature parameters are extracted from the calibrated signal sequence. If measurement deviation is found, the deviation part is refined and adjusted through an iterative optimization algorithm to obtain the optimized signal sequence. The final corrosion depth value and the corresponding corrosion defect location information of the container are parsed from the optimized signal sequence, and the final corrosion depth value and corrosion defect location information are output as the container corrosion detection result.

[0007] According to some embodiments of the present invention, the step of acquiring real-time leakage magnetic field signal data and corresponding lift-off value change data of container corrosion detection from the sensor array, and filtering out environmental noise through an adaptive filtering algorithm to obtain a filtered signal intensity sequence includes: Real-time leakage magnetic field signal data and lift-off value change data of container corrosion detection are collected synchronously by sensor array to form raw dataset; The original dataset is subjected to outlier removal processing to remove abnormal jump data caused by sensor momentary failure or sudden external electromagnetic interference, and retain the valid data. The dynamic correlation between the change characteristics of the lift-off value and the intensity of the leakage magnetic field signal in the effective data is analyzed, and the filtering parameters of the adaptive filtering algorithm are dynamically adjusted based on the dynamic correlation. An adaptive filtering algorithm with adjusted parameters is used to process the real-time leakage magnetic field signal data in the effective data, separate and filter out environmental noise, and obtain the filtered signal strength sequence.

[0008] According to some embodiments of the present invention, the step of calling a pre-built machine learning model for lift-off distortion identification based on the filtered signal strength sequence, identifying signal distortion components caused by changes in lift-off values, and calculating the degree of influence of the signal distortion components includes: The filtered signal intensity sequence and the corresponding lift-off value change data are input into a machine learning model for lift-off distortion identification. Based on the built-in recognition rules, the fluctuation characteristics and amplitude change trends of the signal are analyzed to identify the signal components related to the lift-off value change, and the signal characteristics generated by corrosion itself and the distortion components caused by lift-off interference are distinguished simultaneously. From the distortion components caused by the removal of interference, the signal amplitude offset, waveform distortion degree, and signal strength fluctuation range are extracted as quantization feature parameters. Based on the quantization feature parameters and combined with the built-in influence degree quantification standard of the model, the influence degree of the signal distortion component is calculated; The machine learning model used for lift-off distortion identification is trained using historical magnetic flux leakage signal samples, corresponding lift-off value change data, and labeled lift-off distortion samples. It incorporates identification rules for lift-off-related distortion features and quantification standards for the degree of influence. The degree of influence is characterized by the deviation ratio between the feature parameters and the normal signal benchmark.

[0009] According to some embodiments of the present invention, the step of calling a pre-built machine learning model for outputting compensation amounts, inputting the filtered signal intensity sequence and the signal distortion components to obtain the corresponding compensation amount, and correcting and adjusting the filtered signal intensity sequence based on the compensation amount to obtain a compensated signal intensity sequence includes: The filtered signal strength sequence, the signal distortion components, and the corresponding lift-off value change data are input into a machine learning model that outputs the compensation amount. Based on the current distortion characteristics and lift-off fluctuation patterns, the model outputs the specific compensation amount for the corresponding signal sequence. Based on the specific compensation amount, the filtered signal strength sequence is reversed to offset the signal amplitude shift, waveform distortion and reference drift caused by the change in lift-off value. The corrected signal intensity sequence was verified to confirm that the distortion caused by the change in lift-off value had been eliminated and that the effective signal features related to container corrosion were retained, and finally the compensated signal intensity sequence was obtained. The machine learning model used to output the compensation amount is trained using historical extracted distorted signal samples, corresponding compensation parameters, and adjustment effect data. It includes compensation amount calculation rules and signal correction logic under different degrees of distortion. The compensation amount includes signal amplitude correction value, waveform calibration coefficient, and time dimension compensation parameter.

[0010] According to some embodiments of the present invention, the step of acquiring material magnetic difference data of the detection area for the compensated signal intensity sequence, comparing it with the reference magnetic value in a preset material magnetic database, evaluating the stability of the signal reference through the comparison result, and obtaining a signal stability evaluation result includes: For the detection area corresponding to the compensated signal intensity sequence, the magnetic data of the detection area is collected in real time, and the measured value of magnetic permeability and magnetic uniformity are extracted as material magnetic difference data to characterize the magnetic characteristics of the material. The magnetic difference data of the material is compared with the reference magnetic value of the corresponding material in the preset material magnetic database. The deviation between the measured magnetic permeability and the reference magnetic value and the fluctuation range of magnetic uniformity are calculated as the quantitative index of magnetic difference. The magnetic difference quantification index is compared with the preset signal reference stability judgment threshold in the material magnetic database. If it exceeds the threshold, the signal reference is determined to be unstable. If it does not exceed the threshold, the signal reference is determined to be stable, and the final signal stability evaluation result is formed. The material magnetic database includes magnetic permeability benchmark values, magnetic characteristic parameters, and signal benchmark stability judgment thresholds for container materials of different materials and batches.

[0011] According to some embodiments of the present invention, the step of extracting signal deviation features caused by material magnetic differences from the signal stability assessment results, and fusing the signal deviation features with the compensated signal intensity sequence using a dynamic calibration algorithm to obtain a calibrated signal sequence includes: Based on the information regarding material magnetic differences in the signal stability assessment results, the signal deviation characteristics caused by magnetic differences are extracted. Based on the aforementioned signal deviation characteristics, a correlation mapping relationship between magnetic differences and signal deviation is constructed to determine the signal calibration parameters corresponding to different degrees of magnetic differences. A dynamic calibration algorithm is used to fuse the calibration parameters corresponding to the correlation mapping relationship with the compensated signal intensity sequence to correct the reference offset caused by the difference in material magnetic properties. The compensated signal intensity sequence is processed by the aforementioned correction to eliminate the deviation components caused by differences in material magnetic properties, and finally a calibrated signal sequence is obtained. The signal deviation characteristics include signal amplitude offset, waveform reference drift, and signal strength reference fluctuation range.

[0012] According to some embodiments of the present invention, the step of extracting corrosion depth-related feature parameters based on the calibrated signal sequence, and if measurement deviation is found, refining and adjusting the deviation portion through an iterative optimization algorithm to obtain an optimized signal sequence, includes: Based on the correlation between container corrosion depth and leakage magnetic field signal, characteristic parameters reflecting corrosion depth are extracted from the calibrated signal sequence. The feature parameters are compared with a preset standard corrosion sample feature library, and the deviation between the actual measured value and the standard value of the feature parameters is calculated. If the deviation exceeds the preset accuracy allowable threshold, it is determined that there is a measurement deviation. An iterative optimization algorithm is used to gradually adjust the signal components related to the deviation in the calibrated signal sequence, with the deviation amount as the correction target. After each iteration, the deviation amount of the characteristic parameters is recalculated and the adjustment effect is verified. When the deviation of the feature parameters drops to within the accuracy allowable threshold during the iteration process, the optimization is stopped, and the optimized signal sequence is obtained. The characteristic parameters include signal peak intensity, waveform attenuation rate, and signal duration.

[0013] According to some embodiments of the present invention, the step of parsing the final corrosion depth value and the corresponding corrosion defect location information of the container from the optimized signal sequence, and outputting the final corrosion depth value and corrosion defect location information as the container corrosion detection result, includes: Based on the correlation between the characteristics of container corrosion leakage magnetic field signals and corrosion parameters, a mapping relationship between the characteristic parameters and corrosion depth and defect location in the optimized signal sequence is established in advance; wherein, the mapping relationship includes the correspondence rules between signal peak intensity and corrosion depth, and the conversion standard between signal start and end positions and defect spatial coordinates; Key feature parameters related to corrosion are extracted from the optimized signal sequence; wherein, the key feature parameters include signal peak intensity, spatiotemporal coordinates of the duration interval and waveform-related key feature parameters steepness, which correspond to the corrosion depth quantification index and the defect location positioning basis, respectively. Based on the mapping relationship, the key feature parameters are converted into the final corrosion depth value and defect spatial location coordinates, and verified by a preset signal feature validity threshold. The final corrosion depth value that has passed verification is structurally integrated with the spatial coordinates of the defect to form a container corrosion detection result that includes corrosion parameters, location coordinates and confidence level, and then outputs it.

[0014] Secondly, the present invention also provides a container corrosion magnetic flux leakage detection system based on machine learning, comprising: Signal filtering and acquisition module: acquires real-time leakage magnetic field signal data and corresponding lift-off value change data of container corrosion detection from sensor array, and filters out environmental noise through adaptive filtering algorithm to obtain the filtered signal intensity sequence; Lift-off distortion identification module: Based on the filtered signal strength sequence, it calls a pre-built machine learning model for lifting-off distortion identification to identify signal distortion components caused by changes in lifting-off value and calculate the degree of influence of the signal distortion components; Lift-off compensation adjustment module: When the degree of influence exceeds the preset influence threshold, a pre-built machine learning model for outputting compensation amount is invoked, the filtered signal strength sequence and the signal distortion component are input to obtain the corresponding compensation amount, and the filtered signal strength sequence is corrected and adjusted based on the compensation amount to obtain the compensated signal strength sequence; Magnetic reference evaluation module: For the compensated signal intensity sequence, it acquires the material magnetic difference data of the detection area and compares it with the reference magnetic value in the preset material magnetic database. The stability of the signal reference is evaluated by the comparison result, and the signal stability evaluation result is obtained. Dynamic signal calibration module: Extracts signal deviation features caused by material magnetic differences from the signal stability assessment results, and uses a dynamic calibration algorithm to fuse the signal deviation features with the compensated signal intensity sequence to obtain the calibrated signal sequence; Signal Iterative Optimization Module: Extracts corrosion depth-related feature parameters based on the calibrated signal sequence. If measurement deviation is found, it refines and adjusts the deviation part through an iterative optimization algorithm to obtain the optimized signal sequence. Corrosion Result Output Module: Extracts the final corrosion depth value and corresponding corrosion defect location information of the container from the optimized signal sequence, and outputs the final corrosion depth value and corrosion defect location information as the container corrosion detection result.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) By synchronously acquiring real-time magnetic flux leakage signal data and corresponding lift-off value change data of container corrosion detection, and using an adaptive filtering algorithm to filter out environmental noise, the original detection data and key interference factors (lift-off values) were synchronously acquired, while environmental noise interference was specifically suppressed. This combination of synchronous acquisition and adaptive filtering not only preserves the corrosion characteristic information in the magnetic flux leakage signal, but also lays the foundation for subsequent accurate processing of lift-off related distortions, thereby improving the signal-to-noise ratio and effectiveness of the initial signal.

[0016] (2) By calling a pre-built machine learning model for lift-off distortion recognition, the model accurately identifies the distortion components in the filtered signal caused by changes in lift-off value and quantifies their impact, overcoming the limitations of traditional methods that rely on fixed rules and cannot distinguish between corrosive signals and lift-off interference. The model's targeted identification of lift-off distortion features not only clarifies the impact of distortion on the detection results but also provides accurate quantitative basis for subsequent compensation processing, reducing the risk of misjudgment caused by lift-off fluctuations.

[0017] (3) When the degree of influence of lift-off distortion exceeds the threshold, the corresponding compensation amount is output through the lift-off compensation machine learning model to correct the signal, thereby realizing dynamic adaptive compensation based on distortion features. Compared with the traditional method of manually adjusting compensation parameters, this model can output a compensation amount that matches the real-time distortion components and signal features, effectively offsetting the signal amplitude shift and waveform distortion caused by changes in lift-off value, restoring the signal's ability to truly represent erosion, and improving the accuracy of the compensated signal.

[0018] (4) By acquiring the magnetic difference data of the material in the detection area and comparing it with the reference value in the preset material magnetic database to evaluate the stability of the signal reference, the problem of signal reference drift caused by neglecting the magnetic differences of materials in traditional methods is solved. The preset database covers magnetic references of different materials and batches. Combined with the comparison and evaluation method, the stability of the signal reference is clearly understood, providing reliable magnetic feature support for subsequent calibration and ensuring the consistency of the signal reference.

[0019] (5) Deviation features caused by material magnetic differences are extracted from the signal stability assessment results, and a dynamic calibration algorithm is used to fuse and correct the signal, thereby achieving accurate dynamic calibration for material magnetic fluctuations. This dynamic calibration method avoids the limitation of traditional fixed calibration parameters being unable to adapt to material magnetic differences, effectively eliminates signal reference offset caused by magnetic differences, ensures the comparability of signal characteristics in different material regions, and improves the stability of signal sequences.

[0020] (6) By extracting the corrosion depth feature parameters of the calibrated signal, if there is a measurement deviation, an iterative optimization algorithm is used to refine the adjustment, forming a closed-loop optimization mechanism of "feature extraction - deviation judgment - iterative correction". Iterative optimization can gradually reduce the deviation between the feature parameters and the standard value, ensuring that the optimized signal sequence can accurately characterize the corrosion depth. This overcomes the shortcomings of traditional single calibration in eliminating subtle measurement deviations and improves the accuracy of the signal in characterizing corrosion features.

[0021] (7) The final corrosion depth value and corrosion defect location information are extracted from the optimized signal sequence, integrating the results of previous multi-step interference removal, calibration and optimization. The output detection results have both the quantification of corrosion degree and the location of defect. This result provides reliable data support for container safety maintenance (such as targeted maintenance) and life assessment, enhances the guiding value of the detection results for industrial production safety decisions, and reduces the cost of equipment failure or over-maintenance caused by inaccurate detection. Attached Figure Description

[0022] Figure 1 This is a schematic flowchart of a machine learning-based magnetic flux leakage detection method for container corrosion provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a container corrosion leakage magnetic flux detection system based on machine learning provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Reference Figure 1 This invention provides a machine learning-based method for detecting magnetic flux leakage in container corrosion, comprising the following steps: S11. Obtain real-time leakage magnetic field signal data and corresponding lift-off value change data of container corrosion detection from the sensor array, and filter out environmental noise through an adaptive filtering algorithm to obtain a filtered signal intensity sequence; S12. Based on the filtered signal intensity sequence, call a pre-built machine learning model for lift-off distortion identification to identify the signal distortion components caused by lift-off value changes, and calculate the degree of influence of the signal distortion components. S13, when the degree of influence exceeds the preset influence threshold, a pre-built machine learning model for outputting compensation amount is invoked, the filtered signal strength sequence and the signal distortion component are input to obtain the corresponding compensation amount, and the filtered signal strength sequence is corrected and adjusted based on the compensation amount to obtain the compensated signal strength sequence. S14. For the compensated signal intensity sequence, acquire the material magnetic difference data of the detection area and compare it with the reference magnetic value in the preset material magnetic database. Evaluate the stability of the signal reference through the comparison result and obtain the signal stability evaluation result. S15, extract the signal deviation features caused by material magnetic differences from the signal stability assessment results, and use a dynamic calibration algorithm to fuse the signal deviation features with the compensated signal intensity sequence to obtain the calibrated signal sequence; S16. Extract corrosion depth-related feature parameters based on the calibrated signal sequence. If measurement deviation is found, refine and adjust the deviation part through iterative optimization algorithm to obtain the optimized signal sequence. S17, the final corrosion depth value and the corresponding corrosion defect location information of the container are parsed from the optimized signal sequence, and the final corrosion depth value and corrosion defect location information are output as the container corrosion detection result.

[0025] In step S11, real-time leakage magnetic field signal data and corresponding lift-off value change data of container corrosion detection are obtained from the sensor array, and environmental noise is filtered out by an adaptive filtering algorithm to obtain the filtered signal intensity sequence.

[0026] In one embodiment, this example uses a 3000m³ vertical steel crude oil storage tank (made of Q345R steel, 12mm thick wall, in use for 8 years, with electrochemical corrosion risk on the inner wall, and the testing environment affected by vibration from surrounding oil pump units, including 50Hz power frequency electromagnetic interference and 5-10Hz equipment vibration interference) in the petrochemical industry as the testing object. Specifically, the experimental equipment used is a linear sensor array consisting of 16 LM-200 leakage magnetic field sensors (sensitivity 0.1mV / mT, sampling frequency 1kHz, measurement amplitude range -10~+10V, resolution 0.01mV) and a KeyenceIL-600 laser displacement sensor. The instrument (measuring range 0-50mm, accuracy ±0.005mm, sampling frequency 500Hz) and NIUSB-6363 data acquisition card (16-bit resolution, maximum sampling rate 2MS / s, supports multi-channel synchronous acquisition) were used. During implementation, the magnetic flux leakage sensor array was uniformly arranged along the axial direction of the tank wall (adjacent spacing 0.5m, covering a 10-20m high-corrosion area). The laser displacement sensor was coaxially fixed with each magnetic flux leakage sensor. Data acquisition was started using the acquisition card's clock synchronization function (magnetic flux leakage signal sampling at 1kHz, lift-off value sampling at 500Hz, ensuring that every two lift-off value data points correspond to one magnetic flux leakage signal data point). A total of 1.8 × 10⁻⁶ data points were acquired over 30 minutes. 6 One leakage magnetic signal data (16 channels) and 9×10 5The lift-off value data (16 channels) was used to form the original dataset by associating channels with timestamps. Then, the 3σ criterion was used to remove outliers. Taking the No. 1 leakage magnetic field sensor as an example, its signal mean was calculated to be 0.8V and standard deviation to be 0.3V. Sensor fault jumps and electromagnetic interference spikes exceeding the range of [-0.1V, 1.7V] were removed. Similarly, outliers exceeding the range of [-0.3mm, 2.7mm] corresponding to a mean of 1.2mm and a standard deviation of 0.5mm were removed from the lift-off value data. Finally, 98.2% of the valid data was retained. Subsequently, the valid data was analyzed using a 0.1s time window (100 leakage magnetic field signal points corresponding to 50 lift-off value points). It was found that when the lift-off value increased from 0.8mm to 2.0mm, the leakage magnetic field signal amplitude decreased from 1.2V to 0.5V (correlation coefficient R = -0.92), and the lift-off change... The faster the signal rate, the more obvious the signal distortion. Accordingly, the forgetting factor of the RLS adaptive filtering algorithm was adjusted from 0.9 to 0.95, and the filtering window was adjusted from 20 points to 50 points. Finally, the RLS algorithm with adjusted parameters was used to process the effective leakage magnetic field signal, separating and filtering out 50Hz power frequency noise (amplitude reduced from 0.2V to 0.03V) and 5-10Hz vibration noise (amplitude reduced from 0.15V to 0.02V), while retaining 98% of the peak characteristics of the corrosion signal. The filtered signal strength sequence with a signal-to-noise ratio improved from 18dB to 35dB and an amplitude range of 0.4-1.3V was obtained. It was verified that the peak values ​​of corrosion signals with a depth of more than 0.1mm in this sequence can be stably identified, and the correlation between the signal amplitude and the lift-off value was reduced to R=-0.1, which fully meets the signal quality requirements of the subsequent lift-off distortion identification step.

[0027] In step S12, based on the filtered signal strength sequence, a pre-built machine learning model for lift-off distortion identification is invoked to identify signal distortion components caused by changes in lift-off value, and the influence of the signal distortion components is calculated.

[0028] In one implementation, this embodiment continues to advance based on the filtered signal strength sequence (amplitude range 0.4-1.3V, signal-to-noise ratio 35dB) obtained in the previous embodiment. The core is to distinguish the corrosion signal from the lift-off interference and quantify the distortion effect through a specially constructed lift-off distortion recognition machine learning model. A CNN-LSTM hybrid neural network was selected as the recognition model—it combines the ability of convolutional layers to extract local fluctuation features of signals with the ability of long short-term memory layers to capture the temporal correlation between lift-off values ​​and signals. The model has been trained with 500 sets of historical samples: the training samples cover leakage magnetic signal data of Q345R steel (corrosion depth 0.1-2.0mm, lift-off value 0.5-3.0mm, sampling frequency 1kHz), corresponding lift-off value change data, and 120 sets of manually annotated lift-off distortion samples (the annotation includes the start and end time of distortion, distortion type, and matching lift-off fluctuation amplitude). At the same time, the influence degree quantification standard is built in (the standard corrosion signal without lift-off interference is used as the normal signal benchmark, and the benchmark value is based on laboratory calibration: the signal amplitude corresponding to a corrosion depth of 1.0mm is 1.0V, the waveform similarity is 1.0, and the intensity fluctuation range is 0.2V).

[0029] In implementation, the filtered signal intensity sequence is first divided into 3000 samples (each sample contains 100 signal data points) in a 0.1s time window, and the corresponding lift-off value change data (each sample contains 50 lift-off data points) is synchronously input into the lift-off distortion recognition model. The model extracts the signal fluctuation features (such as amplitude change slope and number of peaks) and waveform details (such as inflection point angle) of each sample through the CNN layer, and the LSTM layer captures the temporal correlation between lift-off value change and signal features. It identifies the lift-off distortion component when the lift-off value increases from 1.0mm to 1.8mm within a certain time window (signal amplitude decreases from 1.0V to 0.88V, without sharp peaks), and simultaneously distinguishes the effective corrosion signal (amplitude 1.1V, containing clear sharp peaks) in adjacent time windows.

[0030] Subsequently, quantization feature parameters are extracted from the distortion components: signal amplitude offset A. offset =0.12V (absolute value of the deviation between 0.88V and 1.0V), waveform distortion degree S wave =0.75 (waveform similarity calculated by dynamic time warping algorithm), signal strength fluctuation range R fluct =0.06V (0.88V-0.82V). Based on the built-in influence degree calculation standard of the model, the influence degree is calculated using formula (1): In the formula: I represents the degree of influence of lift-off distortion (%); A base =1.0V is the normal signal reference amplitude; R base=0.2V is the normal signal baseline fluctuation range. Substituting the data, we get: I=(0.12 / 1.0)×100%+(1-0.75)×50%+(0.06 / 0.2)×50%=12%+12.5%+15%=39.5%. This result intuitively represents the impact of lift-off distortion on the current signal. Verification shows that the model's accuracy in identifying lift-off distortion components reaches 97.2%, providing a reliable quantitative basis for subsequent compensation.

[0031] In step S13, when the degree of influence exceeds a preset influence threshold, a pre-built machine learning model for outputting compensation is invoked, the filtered signal strength sequence and the signal distortion components are input to obtain the corresponding compensation amount, and the filtered signal strength sequence is corrected and adjusted based on the compensation amount to obtain the compensated signal strength sequence.

[0032] In one implementation, this embodiment builds upon the achievements of the previous embodiment—the previous embodiment had identified lift-off distortion components (amplitude offset 0.12V, waveform similarity 0.75, influence degree 39.5%) from the filtered signal strength sequence (amplitude range 0.4-1.3V, signal-to-noise ratio 35dB). The core of this step is to generate correction parameters through a specially constructed compensation output machine learning model to eliminate lift-off interference and retain the effective erosion signal. A backpropagation (BP) neural network was selected as the compensation model, which was trained on 300 sets of historical samples. The training samples included lift-off distortion signals of Q345R steel (lift-off value fluctuation range of 0.5-3.0mm, corresponding to a distortion degree of 10%-60%), laboratory-calibrated compensation parameters (such as amplitude correction values ​​and waveform calibration coefficients under different lift-off fluctuations), and correction effect feedback data (the deviation rate between the corrected signal and the standard undistorted signal). The model has built-in compensation rules for different distortion degrees. For example, for every 0.1mm increase in lift-off value, an amplitude correction value of +0.015V needs to be output. The flat top part of the waveform needs to be restored to sharpness through a calibration coefficient of 0.9-1.0. At the same time, it can output time dimension compensation parameters according to the lift-off change rate to align the signal timing.

[0033] During implementation, the filtered signal strength sequence (3000 samples divided into 0.1s time windows), the identified lift-off distortion components (including amplitude offset of 0.12V, waveform distortion data, and intensity fluctuation range of 0.06V), and the corresponding lift-off value change data (fluctuation of 1.0-1.8mm, including fluctuation rate of 0.8mm / 0.1s) are simultaneously input into the compensation model. The model operates through two hidden layers (containing 64 and 32 neurons respectively), combining the current distortion characteristics and lift-off fluctuation patterns: as the lift-off value increases from 1.0mm to 1.8mm (fluctuation rate of 0.8mm / 0.1s), the corresponding signal amplitude attenuation is 0.12V and the waveform flattens out. Finally, three sets of specific compensation values ​​are output: signal amplitude correction value +0.11V (to offset amplitude attenuation), waveform calibration coefficient 0.92 (to restore waveform sharpness), and time dimension compensation parameter 8ms (to correct signal timing offset caused by lift-off fluctuation).

[0034] Subsequently, the filtered signal strength sequence was reverse-corrected based on these three sets of compensation values: In the amplitude dimension, +0.11V was added to each data point, increasing the amplitude of the original 0.88V distorted signal to 0.99V, close to the normal signal reference (1.0V); in the waveform dimension, the flat-top portion was nonlinearly adjusted by a waveform calibration coefficient of 0.92—correcting the signal slope of the original flat-top segment from 0.02V / ms to 0.05V / ms, restoring the sharp peak shape unique to the corrosion signal; in the time dimension, the corrected signal was shifted by 8ms to align with the signal timing before the lift-off fluctuation, avoiding signal position shift caused by lift-off changes.

[0035] The corrected signal requires dual verification: First, using a dynamic time warping algorithm, the corrected signal is compared with the laboratory-calibrated "standard signal of 1.0mm deep corrosion without lift-off interference." The waveform similarity improves from 0.75 to 0.98, the amplitude deviation decreases from 0.12V to 0.01V, and the proportion of lift-off distortion is confirmed to have decreased from 39.5% to 2.1%. Second, the retention of effective corrosion features is checked—the originally identified 1.1mm deep corrosion signal (amplitude 1.1V, sharp peak) still clearly exists, with a peak deviation of only 0.02V, indicating that no key information was lost due to the correction. The resulting compensated signal intensity sequence has a stable amplitude range of 0.9-1.2V and a signal-to-noise ratio of 35dB, which can be directly used for subsequent signal benchmark evaluation related to material magnetic differences.

[0036] In step S14, for the compensated signal intensity sequence, the material magnetic difference data of the detection area is obtained and compared with the reference magnetic value in the preset material magnetic database. The stability of the signal reference is evaluated by the comparison result, and the signal stability evaluation result is obtained.

[0037] In one implementation, this embodiment is based on the compensated signal intensity sequence (amplitude range 0.9-1.2V, signal-to-noise ratio 35dB) obtained in the aforementioned embodiments. The core is to collect material magnetic data for the detection area of ​​the storage tank corresponding to this signal, and evaluate the signal reference stability by comparing it with a preset database, providing a basis for subsequent calibration. The detection area focuses on the axial and circumferential 90-120° range of the storage tank from 10-20m in diameter—this area completely corresponds to the acquisition location of the compensated signal, and previous testing showed a risk of batch material differences. An FD-M1 portable permeability meter is used to collect magnetic data; this device has a measurement range of 1×10⁻⁶. -4 -1×10 - ²H / m, accuracy ±0.5%, supports real-time on-site readings, and can be paired with magnetic uniformity analysis software to automatically calculate the differences in magnetic distribution at multiple measurement points.

[0038] During implementation, five magnetic sampling points were evenly distributed within the testing area (with a spacing of 0.8m between adjacent points, covering an area of ​​approximately 2m²), ensuring that the sampling points avoided weld seams (to prevent welding from affecting the magnetic data). Magnetic permeability data was collected three times at each sampling point, and the average value was taken. The final measured magnetic permeability value for each sampling point was 3.25 × 10⁻⁶. - ³H / m, 3.32×10 - ³H / m, 3.41×10 - ³H / m, 3.28×10 - ³H / m, 3.35×10 - ³H / m; then, the magnetic uniformity of the region is calculated using magnetic uniformity analysis software—based on the average permeability of 5 measuring points (3.32 × 10⁻⁶ H / m). - Based on ³H / m, (maximum value 3.41×10) - ³H / m - minimum value 3.25×10 - (³H / m) / average value × 100% = 4.8%, and the measured value of magnetic permeability and magnetic uniformity are used as data to characterize the magnetic differences of materials.

[0039] The system then calls a pre-set material magnetic database, which is stored locally in SQLite and contains magnetic parameters for different production batches of Q345R steel. The target batch (matching the tank wall material) has a baseline magnetic permeability value of 3.3 × 10⁻⁶. - The permeability is 3H / m, with an allowable deviation range of ±5%; the baseline threshold for magnetic uniformity is ≤3%; simultaneously, the signal reference stability is determined by the stored criteria—a single-point deviation of magnetic permeability ≤5% and regional magnetic uniformity ≤3% are considered stable; otherwise, there is a risk of instability. The collected magnetic difference data is compared with the database baseline values: the permeability deviation at each measurement point is calculated, with a maximum deviation of 3.41 × 10⁻⁶. -The difference between ³H / m and the baseline value is (3.41-3.3)×10 - ³ / 3.3×10 - 3×100%≈3.3% (less than 5%); however, the regional magnetic uniformity of 4.8% exceeds the benchmark threshold of 3%, thus obtaining the quantitative index of magnetic difference: the maximum deviation of magnetic permeability is 3.3%, and the fluctuation range of magnetic uniformity is 4.8%.

[0040] Finally, the quantitative indicators are compared with the signal reference stability judgment threshold in the database. Because the magnetic uniformity fluctuation exceeds the threshold, it is determined that the detection area corresponding to the compensated signal has "slight instability in the signal reference, and magnetic difference interference needs to be eliminated through subsequent dynamic calibration". At the same time, the details of the magnetic permeability deviation at each measurement point are recorded to form a signal stability evaluation result containing "magnetic data details, deviation analysis, and stability conclusion". This result will be directly used for setting the signal dynamic calibration parameters in the next stage.

[0041] In step S15, the signal deviation features caused by material magnetic differences are extracted from the signal stability assessment results, and the signal deviation features are fused with the compensated signal intensity sequence using a dynamic calibration algorithm to obtain the calibrated signal sequence.

[0042] In one implementation, this embodiment is based on the signal stability evaluation results of the aforementioned embodiments—the results show that there are material magnetic differences in the tank detection area (10-20m axial, 90-120° circumferential) corresponding to the compensated signal: the maximum permeability deviation is 3.3% (3.41×10 at a single measurement point). - ³H / mvs baseline 3.3×10 - The magnetic uniformity of the region fluctuated by 4.8% (³H / m), exceeding the 3% threshold, indicating slight instability in the signal baseline. The core of this step is to extract the signal deviation characteristics caused by magnetic differences, eliminate the deviation through dynamic calibration, and obtain a stable signal sequence.

[0043] During implementation, based on the magnetic difference determination information in the signal stability assessment results, and combined with the characteristic analysis of the compensated signal strength sequence, three types of signal deviation characteristics are extracted: one is the signal amplitude offset, compared with the measurement point with the highest permeability (3.41×10). -The signal amplitudes in the magnetic permeability region (³H / m) and the reference magnetic permeability region were compared. It was found that the former signal amplitude was 0.18V higher than the latter (1.18V vs 1.00V), which is due to the amplitude shift caused by magnetic differences. Secondly, the waveform reference drift was calculated, and the offset between the signal baseline in the magnetic difference region and the signal baseline in the reference magnetic region was calculated, which accounted for 4.5% of the average signal amplitude (drift amount 0.05V / average amplitude 1.1V). Thirdly, the signal intensity reference fluctuation range was analyzed. The amplitude fluctuation range of the signals corresponding to the five magnetic measurement points was statistically analyzed, and it expanded from 0.2V (0.9-1.1V) after compensation to 0.35V (0.95-1.3V), which is due to the intensity fluctuation caused by insufficient magnetic uniformity.

[0044] Based on these deviation characteristics, a correlation mapping relationship between "magnetic difference" and "signal deviation" was constructed: through linear regression analysis of the permeability and signal amplitude data of five measurement points, the mapping formula "signal amplitude offset (V) = 0.056 × permeability deviation rate (%)" was obtained (e.g., a 3.3% deviation corresponds to 0.056 × 3.3 ≈ 0.18V, consistent with the actual measurement); at the same time, the correlation between waveform baseline drift and magnetic uniformity was determined: "drift (%) = 0.94 × magnetic uniformity fluctuation (%)" (4.8% uniformity fluctuation corresponds to 0.94 × 4.8 ≈ 4.5% drift). Based on this, signal calibration parameters were set as follows: an amplitude correction value of -0.18V for the measurement point with the highest permeability (to offset the positive offset), a waveform baseline calibration coefficient of 0.96 for the entire region (to pull back the drifted baseline), and an intensity fluctuation suppression coefficient of 0.6 (to compress the fluctuation range from 0.35V to 0.21V).

[0045] Kalman filtering is used as the dynamic calibration algorithm to fuse the calibration parameters corresponding to the above correlation mapping relationship with the compensated signal intensity sequence: the algorithm takes the compensated signal as the initial input and updates the calibration parameters in real time through the state equation. When the permeability deviation is detected to drop from 3.3% to 2.5%, the amplitude correction value is automatically adjusted from -0.18V to -0.14V. At the same time, the waveform calibration coefficient is continuously optimized by comparing the deviation between the corrected signal and the reference magnetic region signal through the observation equation (and finally stabilizing at 0.97).

[0046] The corrected signal sequence was then validated: the signal amplitude at the point with the highest permeability was corrected from 1.18V to 1.00V, with a deviation from the baseline signal of ≤0.02V; the waveform baseline drift decreased from 4.5% to 1.2%; and the signal intensity fluctuation range was compressed from 0.35V to 0.2V, matching the initial stability of the compensated signal. More importantly, the signal peak value corresponding to a 1.0mm corrosion depth (originally 1.0V) was fully preserved, with a peak deviation of only 0.01V, indicating no loss of corrosion characteristics due to calibration. The final calibrated signal sequence has a stable amplitude range of 0.9-1.1V and a flat waveform baseline, making it directly usable for subsequent corrosion depth feature extraction and optimization.

[0047] In step S16, corrosion depth-related feature parameters are extracted based on the calibrated signal sequence. If a measurement deviation is found, the deviation is refined and adjusted using an iterative optimization algorithm to obtain an optimized signal sequence.

[0048] In one implementation, this embodiment is based on the calibrated signal sequence (amplitude range 0.9-1.1V, flat waveform baseline, signal intensity fluctuation range 0.2V) obtained in the aforementioned embodiments. The core is to extract characteristic parameters related to corrosion depth from this sequence, compare it with standard samples to determine if there are any deviations, and if so, refine and adjust it iteratively to finally obtain a signal sequence that accurately reflects the corrosion state. The detection object is still the axial region of 10-20m and the circumferential region of 90-120° in the storage tank. This region has been pre-processed to eliminate the effects of lift-off interference and magnetic differences, focusing on optimizing the measurement accuracy of the corrosion signal itself.

[0049] During implementation, based on the correlation between the corrosion characteristics of Q345R steel and the leakage magnetic field signal (laboratory calibration: for every 0.1 mm increase in corrosion depth, the peak intensity of the leakage magnetic field signal increases by an average of 0.1 V; the waveform attenuation rate (the proportion of the time it takes for the signal to drop from the peak to half the amplitude to the total duration) is stable at 0.3 ± 0.02; the signal duration (the time it takes for the corrosion signal to fall from the peak to the baseline) increases linearly with depth, with 1.0 mm depth corresponding to 150 ms), three types of characteristic parameters are extracted from the calibrated signal sequence: for a signal segment (time window 0.3 s) in a suspected corrosion area, the peak signal intensity is extracted to be 1.08 V, the waveform attenuation rate is 0.35, and the signal duration is 162 ms.

[0050] Subsequently, a pre-defined standard corrosion sample feature library was invoked. This library stores standard feature parameters corresponding to different corrosion depths (0.1-2.0 mm). The standard parameters for 1.0 mm corrosion depth are: peak intensity 1.0 V, attenuation rate 0.3, and duration 150 ms. The preset accuracy allowable threshold is 5% (i.e., peak deviation ≤ 0.05 V, attenuation rate deviation ≤ 0.015, and duration deviation ≤ 7.5 ms). The actual extracted feature parameters were compared with the standard values, and the deviations were calculated: peak intensity deviation 0.08 V (8%, exceeding 5%), attenuation rate deviation 0.05 (33.3%, exceeding 5%), and duration deviation 12 ms (8%, exceeding 5%). It was determined that the signal segment had measurement deviations and iterative optimization was required.

[0051] The gradient descent iterative optimization algorithm is adopted with the goal of reducing the deviation of all feature parameters to within 5%. The adjustment step size for each iteration is set as follows: peak intensity is adjusted by -0.02V each time (towards the standard value of 1.0V), waveform attenuation rate is adjusted by -0.01 each time (towards 0.3), and signal duration is adjusted by -2ms each time (towards 150ms). After the first iteration, the parameters were updated to a peak value of 1.06V, a decay rate of 0.34, and a duration of 160ms, with deviations of 6%, 26.7%, and 6.7%, respectively, still exceeding the threshold. After the second iteration, the parameters were a peak value of 1.04V, a decay rate of 0.33, and a duration of 158ms, with deviations of 4%, 20%, and 5.3%, respectively. The peak value met the standard, but the others still needed adjustment. After the third iteration, the parameters were a peak value of 1.02V, a decay rate of 0.32, and a duration of 156ms, with deviations of 2%, 13.3%, and 4%, respectively, and the duration met the standard. After the fourth iteration, the parameters were a peak value of 1.01V, a decay rate of 0.31, and a duration of 154ms, with deviations of 1%, 3.3%, and 2.7%, respectively. All parameter deviations were reduced to within the 5% threshold, and the iteration was stopped.

[0052] After optimization, the signal sequence was verified: the peak value of the signal in the eroded region (1.01V) deviated by 1% from the standard 1.0V, the attenuation rate (0.31) deviated by 3.3% from the standard 0.3, and the duration (154ms) deviated by 2.7% from the standard 150ms, all meeting the accuracy requirements. Simultaneously, the waveform inflection points, baseline stability, and other characteristics of the signal were not distorted by the optimization, and the signal characteristics corresponding to a 1.0mm depth of erosion were fully preserved. The final optimized signal sequence showed significantly improved consistency between its characteristic parameters and the standard sample, and can be directly used for subsequent analysis of erosion depth and location.

[0053] In step S17, the final corrosion depth value and the corresponding corrosion defect location information of the container are parsed from the optimized signal sequence, and the final corrosion depth value and corrosion defect location information are output as the container corrosion detection result.

[0054] In one implementation, this embodiment is based on the optimized signal sequence obtained in the aforementioned embodiments (characteristic parameters: peak intensity 1.01V, attenuation rate 0.31, duration 154ms, deviation ≤5%). The core is to parse the final corrosion depth and defect location of the container from this sequence to form a structured detection result. The detection object is still the axial region of 10-20m and the circumferential region of 90-120° of the storage tank. After multiple processing steps in the previous stage, the signal in this region can accurately reflect the corrosion characteristics.

[0055] During implementation, a mapping relationship was first established based on the correlation law between "leakage magnetic signal characteristics and corrosion parameters" calibrated in the laboratory in the early stage: First, the correspondence rule between signal peak intensity and corrosion depth - based on 200 groups of leakage magnetic signal samples of Q345R steel with different corrosion depths (0.1-2.0mm), the mapping formula (2) was obtained by linear regression fitting: In the formula: D is the corrosion depth (mm); A peak The peak intensity (V) of the optimized signal is given; the fitting result is k=1.0 (mm / V), b=0 (intercept, since 0 depth corrosion corresponds to 0V signal), and the goodness of fit R is given. 2 =0.99, therefore the formula simplifies to D=A peak Secondly, the conversion standard between the signal start and end positions and the spatial coordinates of the defect is as follows: one sensor array is deployed every 0.5m along the axial direction (numbered 1-20, corresponding to axial distances of 10-20m). The circumferential positioning is achieved through a laser displacement sensor and an angle encoder (accuracy ±0.5°). The midpoint of the signal duration interval corresponds to the center coordinates of the defect.

[0056] Then, key feature parameters were extracted from the optimized signal sequence: signal peak intensity A. peak =1.01V, the midpoint of the continuous interval corresponds to an axial distance of 15.007m, and the circumferential angle is 95°. According to formula (2), the final corrosion depth D = 1.01mm is obtained. After verification by the preset signal feature validity threshold (correlation coefficient ≥ 0.8), the correlation coefficient between the optimized signal and the standard corrosion signal with a depth of 1.0mm is 0.96 (≥ 0.8), the judgment result is reliable, and a 99% confidence level is given.

[0057] Finally, the results were structured and integrated to form an inspection report containing "corrosion depth 1.01mm (accuracy ±0.02mm), defect location (axial 15.007m, circumferential 95°), confidence level 99%, corrosion type: pitting corrosion". On-site ultrasonic re-inspection showed that the actual corrosion depth at this location was 1.02mm, with a deviation of 0.01mm from the inspection result, verifying the accuracy of the analysis.

[0058] In summary, this invention achieves accurate detection of magnetic flux leakage in container corrosion through multi-source data fusion and intelligent end-to-end processing: First, magnetic flux leakage signals and lift-off value changes are simultaneously acquired. Environmental noise is dynamically suppressed through adaptive filtering, and a CNN-LSTM model is used to accurately identify and quantify the impact of lift-off distortion. Then, a BP neural network outputs compensation to offset lift-off interference. Simultaneously, material magnetic data is acquired and compared with a preset database to evaluate signal baseline stability and eliminate magnetic difference deviations based on a dynamic calibration algorithm. Finally, corrosion feature parameters are extracted, and deviations are refined and adjusted through iterative optimization. Ultimately, the corrosion depth and location are analyzed through feature-parameter mapping relationships. This process forms a closed-loop mechanism of "noise suppression - distortion compensation - magnetic calibration - feature optimization - result analysis," effectively solving interference problems such as lift-off fluctuations and material magnetic differences. This reduces the corrosion depth measurement error to ±0.05mm and the defect location accuracy to ±0.01m, providing highly reliable data support for container safety assessment and maintenance, and significantly improving detection accuracy and industrial application value.

[0059] refer to Figure 2 The second embodiment of the invention provides a machine learning-based container corrosion magnetic flux leakage detection system, comprising: Signal filtering and acquisition module: acquires real-time leakage magnetic field signal data and corresponding lift-off value change data of container corrosion detection from sensor array, and filters out environmental noise through adaptive filtering algorithm to obtain the filtered signal intensity sequence; Lift-off distortion identification module: Based on the filtered signal strength sequence, it calls a pre-built machine learning model for lifting-off distortion identification to identify signal distortion components caused by changes in lifting-off value and calculate the degree of influence of the signal distortion components; Lift-off compensation adjustment module: When the degree of influence exceeds the preset influence threshold, a pre-built machine learning model for outputting compensation amount is invoked, the filtered signal strength sequence and the signal distortion component are input to obtain the corresponding compensation amount, and the filtered signal strength sequence is corrected and adjusted based on the compensation amount to obtain the compensated signal strength sequence; Magnetic reference evaluation module: For the compensated signal intensity sequence, it acquires the material magnetic difference data of the detection area and compares it with the reference magnetic value in the preset material magnetic database. The stability of the signal reference is evaluated by the comparison result, and the signal stability evaluation result is obtained. Dynamic signal calibration module: Extracts signal deviation features caused by material magnetic differences from the signal stability assessment results, and uses a dynamic calibration algorithm to fuse the signal deviation features with the compensated signal intensity sequence to obtain the calibrated signal sequence; Signal Iterative Optimization Module: Extracts corrosion depth-related feature parameters based on the calibrated signal sequence. If measurement deviation is found, it refines and adjusts the deviation part through an iterative optimization algorithm to obtain the optimized signal sequence. Corrosion Result Output Module: Extracts the final corrosion depth value and corresponding corrosion defect location information of the container from the optimized signal sequence, and outputs the final corrosion depth value and corrosion defect location information as the container corrosion detection result.

[0060] It should be noted that the machine learning-based container corrosion magnetic flux leakage detection system provided in this embodiment of the invention is used to execute all the process steps of the machine learning-based container corrosion magnetic flux leakage detection method described in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0061] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A machine learning-based method for detecting magnetic flux leakage in container corrosion, characterized in that, include: Real-time magnetic flux leakage signal data and corresponding lift-off value change data of container corrosion detection are acquired from the sensor array, and environmental noise is filtered out by an adaptive filtering algorithm to obtain the filtered signal intensity sequence. Based on the filtered signal strength sequence, a pre-built machine learning model for lift-off distortion identification is invoked to identify signal distortion components caused by changes in lift-off value, and the degree of influence of the signal distortion components is calculated. When the degree of influence exceeds the preset influence threshold, a pre-built machine learning model for outputting compensation is invoked. The filtered signal strength sequence and the signal distortion components are input to obtain the corresponding compensation amount. Based on the compensation amount, the filtered signal strength sequence is corrected and adjusted to obtain the compensated signal strength sequence. For the compensated signal intensity sequence, the material magnetic difference data of the detection area is obtained and compared with the reference magnetic value in the preset material magnetic database. The stability of the signal reference is evaluated by the comparison results, and the signal stability evaluation result is obtained. The signal deviation features caused by material magnetic differences are extracted from the signal stability assessment results. The signal deviation features are then fused with the compensated signal intensity sequence using a dynamic calibration algorithm to obtain the calibrated signal sequence. Corrosion depth-related feature parameters are extracted from the calibrated signal sequence. If measurement deviation is found, the deviation part is refined and adjusted through an iterative optimization algorithm to obtain the optimized signal sequence. The final corrosion depth value and the corresponding corrosion defect location information of the container are parsed from the optimized signal sequence, and the final corrosion depth value and corrosion defect location information are output as the container corrosion detection result.

2. The machine learning-based magnetic flux leakage detection method for container corrosion according to claim 1, characterized in that, The process involves acquiring real-time magnetic flux leakage signal data and corresponding lift-off value changes from the sensor array, and filtering out environmental noise using an adaptive filtering algorithm to obtain a filtered signal intensity sequence, including: Real-time leakage magnetic field signal data and lift-off value change data of container corrosion detection are collected synchronously by sensor array to form raw dataset; The original dataset is subjected to outlier removal processing to remove abnormal jump data caused by sensor momentary failure or sudden external electromagnetic interference, and retain the valid data. The dynamic correlation between the change characteristics of the lift-off value and the intensity of the leakage magnetic field signal in the effective data is analyzed, and the filtering parameters of the adaptive filtering algorithm are dynamically adjusted based on the dynamic correlation. An adaptive filtering algorithm with adjusted parameters is used to process the real-time leakage magnetic field signal data in the effective data, separate and filter out environmental noise, and obtain the filtered signal strength sequence.

3. The machine learning-based magnetic flux leakage detection method for container corrosion according to claim 1, characterized in that, The step of calling a pre-built machine learning model for lift-off distortion identification based on the filtered signal intensity sequence, identifying signal distortion components caused by changes in lift-off values, and calculating the degree of influence of the signal distortion components includes: The filtered signal intensity sequence and the corresponding lift-off value change data are input into a machine learning model for lift-off distortion identification. Based on the built-in recognition rules, the fluctuation characteristics and amplitude change trends of the signal are analyzed to identify the signal components related to the lift-off value change, and the signal characteristics generated by corrosion itself and the distortion components caused by lift-off interference are distinguished simultaneously. From the distortion components caused by the removal of interference, the signal amplitude offset, waveform distortion degree, and signal strength fluctuation range are extracted as quantization feature parameters. Based on the quantization feature parameters and combined with the built-in influence degree quantification standard of the model, the influence degree of the signal distortion component is calculated; The machine learning model used for lift-off distortion identification is trained using historical magnetic flux leakage signal samples, corresponding lift-off value change data, and labeled lift-off distortion samples. It incorporates identification rules for lift-off-related distortion features and quantification standards for the degree of influence. The degree of influence is characterized by the deviation ratio between the feature parameters and the normal signal benchmark.

4. The machine learning-based magnetic flux leakage detection method for container corrosion according to claim 1, characterized in that, The process involves calling a pre-built machine learning model for outputting compensation amounts, inputting the filtered signal strength sequence and the signal distortion components to obtain the corresponding compensation amounts, and then correcting and adjusting the filtered signal strength sequence based on these compensation amounts to obtain the compensated signal strength sequence, including: The filtered signal strength sequence, the signal distortion components, and the corresponding lift-off value change data are input into a machine learning model that outputs the compensation amount. Based on the current distortion characteristics and lift-off fluctuation patterns, the model outputs the specific compensation amount for the corresponding signal sequence. Based on the specific compensation amount, the filtered signal strength sequence is reversed to offset the signal amplitude shift, waveform distortion and reference drift caused by the change in lift-off value. The corrected signal intensity sequence was verified to confirm that the distortion caused by the change in lift-off value had been eliminated and that the effective signal features related to container corrosion were retained, and finally the compensated signal intensity sequence was obtained. The machine learning model used to output the compensation amount is trained using historical extracted distorted signal samples, corresponding compensation parameters, and adjustment effect data. It includes compensation amount calculation rules and signal correction logic under different degrees of distortion. The compensation amount includes signal amplitude correction value, waveform calibration coefficient, and time dimension compensation parameter.

5. The machine learning-based magnetic flux leakage detection method for container corrosion according to claim 1, characterized in that, For the compensated signal intensity sequence, material magnetic difference data of the detection area is acquired and compared with the reference magnetic value in the preset material magnetic database. The stability of the signal reference is evaluated based on the comparison result to obtain the signal stability evaluation result, including: For the detection area corresponding to the compensated signal intensity sequence, the magnetic data of the detection area is collected in real time, and the measured value of magnetic permeability and magnetic uniformity are extracted as material magnetic difference data to characterize the magnetic characteristics of the material. The magnetic difference data of the material is compared with the reference magnetic value of the corresponding material in the preset material magnetic database. The deviation between the measured magnetic permeability and the reference magnetic value and the fluctuation range of magnetic uniformity are calculated as the quantitative index of magnetic difference. The magnetic difference quantification index is compared with the preset signal reference stability judgment threshold in the material magnetic database. If it exceeds the threshold, the signal reference is determined to be unstable. If it does not exceed the threshold, the signal reference is determined to be stable, and the final signal stability evaluation result is formed. The material magnetic database includes magnetic permeability benchmark values, magnetic characteristic parameters, and signal benchmark stability judgment thresholds for container materials of different materials and batches.

6. The machine learning-based magnetic flux leakage detection method for container corrosion according to claim 1, characterized in that, The signal deviation features caused by material magnetic differences are extracted from the signal stability assessment results. A dynamic calibration algorithm is then used to fuse these signal deviation features with the compensated signal intensity sequence to obtain the calibrated signal sequence, including: Based on the information regarding material magnetic differences in the signal stability assessment results, the signal deviation characteristics caused by magnetic differences are extracted. Based on the aforementioned signal deviation characteristics, a correlation mapping relationship between magnetic differences and signal deviation is constructed to determine the signal calibration parameters corresponding to different degrees of magnetic differences. A dynamic calibration algorithm is used to fuse the calibration parameters corresponding to the correlation mapping relationship with the compensated signal intensity sequence to correct the reference offset caused by the difference in material magnetic properties. The compensated signal intensity sequence is processed by the aforementioned correction to eliminate the deviation components caused by differences in material magnetic properties, and finally a calibrated signal sequence is obtained. The signal deviation characteristics include signal amplitude offset, waveform reference drift, and signal strength reference fluctuation range.

7. The machine learning-based magnetic flux leakage detection method for container corrosion according to claim 1, characterized in that, The step involves extracting corrosion depth-related feature parameters based on the calibrated signal sequence. If a measurement deviation is detected, an iterative optimization algorithm is used to refine and adjust the deviation portion, resulting in an optimized signal sequence, including: Based on the correlation between container corrosion depth and leakage magnetic field signal, characteristic parameters reflecting corrosion depth are extracted from the calibrated signal sequence. The feature parameters are compared with a preset standard corrosion sample feature library, and the deviation between the actual measured value and the standard value of the feature parameters is calculated. If the deviation exceeds the preset accuracy allowable threshold, it is determined that there is a measurement deviation. An iterative optimization algorithm is used to gradually adjust the signal components related to the deviation in the calibrated signal sequence, with the deviation amount as the correction target. After each iteration, the deviation amount of the characteristic parameters is recalculated and the adjustment effect is verified. When the deviation of the feature parameters drops to within the accuracy allowable threshold during the iteration process, the optimization is stopped, and the optimized signal sequence is obtained. The characteristic parameters include signal peak intensity, waveform attenuation rate, and signal duration.

8. The machine learning-based magnetic flux leakage detection method for container corrosion according to claim 1, characterized in that, The process of parsing the final corrosion depth value and corresponding corrosion defect location information of the container from the optimized signal sequence, and outputting the final corrosion depth value and corrosion defect location information as the container corrosion detection result, includes: Based on the correlation between the characteristics of container corrosion leakage magnetic field signals and corrosion parameters, a mapping relationship between the characteristic parameters and corrosion depth and defect location in the optimized signal sequence is established in advance; wherein, the mapping relationship includes the correspondence rules between signal peak intensity and corrosion depth, and the conversion standard between signal start and end positions and defect spatial coordinates; Key feature parameters related to corrosion are extracted from the optimized signal sequence; wherein, the key feature parameters include signal peak intensity, spatiotemporal coordinates of the duration interval and waveform-related key feature parameters steepness, which correspond to the corrosion depth quantification index and the defect location positioning basis, respectively. Based on the mapping relationship, the key feature parameters are converted into the final corrosion depth value and defect spatial location coordinates, and verified by a preset signal feature validity threshold. The final corrosion depth value that has passed verification is structurally integrated with the spatial coordinates of the defect to form a container corrosion detection result that includes corrosion parameters, location coordinates and confidence level, and then outputs it.

9. A machine learning-based container corrosion magnetic flux leakage detection system, used to implement the machine learning-based container corrosion magnetic flux leakage detection method as described in any one of claims 1 to 8, characterized in that, include: Signal filtering and acquisition module: acquires real-time leakage magnetic field signal data and corresponding lift-off value change data of container corrosion detection from sensor array, and filters out environmental noise through adaptive filtering algorithm to obtain the filtered signal intensity sequence; Lift-off distortion identification module: Based on the filtered signal strength sequence, it calls a pre-built machine learning model for lifting-off distortion identification to identify signal distortion components caused by changes in lifting-off value and calculate the degree of influence of the signal distortion components; Lift-off compensation adjustment module: When the degree of influence exceeds the preset influence threshold, a pre-built machine learning model for outputting compensation amount is invoked, the filtered signal strength sequence and the signal distortion component are input to obtain the corresponding compensation amount, and the filtered signal strength sequence is corrected and adjusted based on the compensation amount to obtain the compensated signal strength sequence; Magnetic reference evaluation module: For the compensated signal intensity sequence, it acquires the material magnetic difference data of the detection area and compares it with the reference magnetic value in the preset material magnetic database. The stability of the signal reference is evaluated by the comparison result, and the signal stability evaluation result is obtained. Dynamic signal calibration module: Extracts signal deviation features caused by material magnetic differences from the signal stability assessment results, and uses a dynamic calibration algorithm to fuse the signal deviation features with the compensated signal intensity sequence to obtain the calibrated signal sequence; Signal Iterative Optimization Module: Extracts corrosion depth-related feature parameters based on the calibrated signal sequence. If measurement deviation is found, it refines and adjusts the deviation part through an iterative optimization algorithm to obtain the optimized signal sequence. Corrosion Result Output Module: Extracts the final corrosion depth value and corresponding corrosion defect location information of the container from the optimized signal sequence, and outputs the final corrosion depth value and corrosion defect location information as the container corrosion detection result.

Images