Ultrasonic water immersion flaw detection method using intelligent sensors

Abstract

This invention discloses an ultrasonic immersion flaw detection method using intelligent sensors, belonging to the field of intelligent sensor technology. The steps of the detection method include: selecting the corresponding intelligent sensor and its operating frequency from a pre-built matching sensor library based on the specification information of the target workpiece to be inspected. The key technical points are: this invention introduces CNN into complex scenarios where ERF is prone to misjudgment, significantly improving the correction capability of key misjudgment points. CNN is only called locally when necessary, avoiding high computational consumption across the entire area, and the overall efficiency is still close to that of the pure ERF solution. Through confidence feedback, it can automatically focus on ambiguous areas, improving the detection rate of small or complex defects. The operation of the overall solution realizes dynamic collaborative operation, ensuring a certain level of accuracy while also guaranteeing detection efficiency and targeted improvement of defect type recognition, ensuring the balance and stability among the three effects.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensor technology, specifically to an ultrasonic water immersion flaw detection method using intelligent sensors. Background Technology

[0002] When intelligent sensors are used in ultrasonic water immersion testing, they typically refer to sensor units that integrate data acquisition, signal processing, and communication functions, possessing capabilities such as self-calibration and environmental compensation. The current conventional testing process involves immersing the workpiece in a water tank, transmitting ultrasonic waves via water coupling; a probe, driven by a mechanical device, scans the workpiece surface along a predetermined path; pulsed ultrasonic waves are emitted, and reflected echo signals are received; the location and size of defects are determined by analyzing echo time, amplitude, and other characteristics; and a C-scan or B-scan image is generated using an imaging system, with the nature of the defect determined manually or by software. This entire process relies on fixed parameter settings; the scanning path and resolution are usually preset, lacking real-time adjustment capabilities.

[0003] The following problems may arise when using traditional or designed flaw detection methods:

[0004] Traditional ultrasonic immersion testing mostly employs a fixed scanning strategy, using a uniform resolution regardless of the area's characteristics. This often leads to wasted resources or missed detections. For example, when inspecting thick-walled workpieces, high-resolution scanning is still performed on areas without obvious abnormalities, which is time-consuming and consumes a lot of computing resources. In complex structural areas, such as weld corners, insufficient resolution makes it difficult to identify minute cracks, resulting in misjudgments or missed detections. Currently, although there are some methods on the market that introduce automatic recognition algorithms, they mostly rely on a single model running throughout the process and cannot be dynamically adjusted according to actual needs. This results in a lack of effective verification mechanisms in areas prone to misjudgment. In addition, traditional methods lack the ability to handle regional differences and cannot achieve flexible scheduling. Therefore, it is difficult to effectively balance efficiency and accuracy. If manual verification is used, it will not only increase the burden of manual verification but also affect the overall reliability of the inspection. These problems are even more urgent in high-volume or fast-paced production scenarios. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] An ultrasonic water immersion flaw detection method using intelligent sensors includes the following steps:

[0007] Based on the specifications of the target workpiece to be inspected, the corresponding smart sensors and their operating frequencies are selected from the pre-built matching sensor library; a dynamic region division strategy is adopted to obtain the partitioning scheme of the target workpiece to be inspected: divide each scanning area, establish a probability model, output the probability of the calibrated defect type and its possible location, and the output result of the probability model is positively correlated with the priority, and execute a high or low resolution scanning strategy according to the priority.

[0008] A real-time data analysis mechanism is introduced to perform preliminary defect identification actions on scanning areas of different priorities;

[0009] When the low-resolution scanning strategy identifies a suspected defect, a strategy change action is triggered.

[0010] The scanning regions implementing the high-resolution scanning strategy are marked as critical regions, and the scanning regions implementing the low-resolution scanning strategy are marked as non-critical regions. A dual-model collaborative architecture is adopted: a main model and an auxiliary model. A scheduling management mechanism is introduced as a constraint. When using the auxiliary model, the constraint is applied to complete the adjustment action and generate the final probability.

[0011] If the final probability exceeds the judgment threshold, it is marked as a defect; otherwise, no marking is made.

[0012] Furthermore, in the matching sensor library: when detecting a target workpiece with a maximum thickness exceeding 5mm, the benchmark operating frequency is 1MHz; when detecting a target workpiece with a maximum thickness not exceeding 5mm, the benchmark operating frequency is 5MHz.

[0013] The thickness of the target workpiece is inversely proportional to the value set by the working frequency.

[0014] Furthermore, the dynamic region partitioning strategy adopted is as follows: the geometric information of the target workpiece surface is obtained through 3D scanning, and a partitioning scheme is generated using computer-aided design (CAD) software; the process of executing high or low resolution scanning strategies according to priority is as follows: when the priority exceeds the calibration value, it is marked as high priority and a high resolution scanning strategy is executed; otherwise, it is marked as low priority and a low resolution scanning strategy is executed.

[0015] Furthermore, the main model uses a lightweight convolutional neural network, i.e., CNN, to handle critical regions; the auxiliary model uses an enhanced random forest, i.e., ERF, to handle non-critical regions.

[0016] Furthermore, the operation process of the scheduling management mechanism is as follows: A1: Data complexity assessment; A2: Confidence and uncertainty handling; A3: Dynamic decision-making mechanism for the proportion of CNN participation.

[0017] Furthermore, the data complexity assessment process mentioned in A1 is as follows: analyze the signal of the corresponding non-critical area to obtain complexity assessment indicators, including at least: noise level, frequency component change and energy distribution, perform weighted calculation based on the complexity assessment indicators to obtain the complexity value; when the complexity value exceeds the preset first standard threshold, the first linear proportional adjustment strategy is triggered; otherwise, no response action is taken.

[0018] Furthermore, regarding noise level: the peak signal-to-noise ratio (SNR) of the signal is calculated as the quantized value of the noise level; frequency component variation: the Fast Fourier Transform (FFT) is used to analyze the spectral distribution of the signal. For continuous segments in non-critical regions, the difference in the spectral center frequency between adjacent segments is calculated, and the maximum value is selected as the quantized value of the frequency component variation; energy distribution: for signals in non-critical regions, the total energy is calculated, and the information entropy is obtained based on the total energy. The reciprocal of the information entropy is used to reflect the concentration of energy distribution and is used as the quantized value of energy distribution.

[0019] Furthermore, the first linear scaling adjustment strategy process is as follows: the proportion of lightweight convolutional neural networks involved is obtained based on the complexity value C, and the method is: β_C=min(1,max(0,k_C×(C-C_thd))); where β_C is the CNN proportion based on the complexity value, k_C is the adjustment coefficient with a value range of [0,2], and C_thd is the first standard threshold.

[0020] Furthermore, the confidence and uncertainty handling process mentioned in A2 is as follows: analyze the confidence of the corresponding auxiliary model output and compare the confidence with the preset second standard threshold. When the confidence does not exceed the second standard threshold, the second linear scaling adjustment strategy is triggered; otherwise, no response is taken. The triggered second linear scaling adjustment strategy process is as follows: obtain the proportion of lightweight convolutional neural network involved based on the confidence D, and the method is: β_D=max(0, 1-D / D_thd); where β_D is the CNN proportion based on confidence and D_thd is the second standard threshold.

[0021] Furthermore, the dynamic decision-making mechanism for the CNN participation ratio mentioned in A3 is as follows: when any adjustment strategy is triggered, the result of the corresponding adjustment strategy is used as the basis for the fusion probability model, and the fusion probability model is run to output the final probability; when two first and second linear ratio adjustment strategies are triggered simultaneously, the maximum value is taken as the basis for the fusion probability model, and the fusion probability model is run: P_final=a×P_CNN+(1-a)×P_ERF, and the final probability P_final is output; where the value of a is β_C or β_D.

[0022] This invention provides an ultrasonic water immersion flaw detection method using intelligent sensors, which has the following beneficial effects:

[0023] 1) This solution combines intelligent sensors with dynamic region division. During the detection process, different scanning resolutions are used for different regions. The differentiated processing not only improves the ability to identify subtle defects and effectively reduces the false alarm rate, but also saves operating resources. On the one hand, it ensures the accuracy of the detection results, and on the other hand, it ensures the efficiency and reliability of the overall detection process.

[0024] 2) This solution utilizes high and low resolution scanning strategies and introduces a real-time data analysis mechanism. When a suspected defect is initially detected, a higher resolution re-examination is triggered. At the same time, the subsequent scanning plan is dynamically adjusted according to actual needs. This not only improves the response speed and flexibility of the overall detection solution, but also reduces the computational cost to a certain extent, and realizes the rational planning or utilization of resources. It solves the problem of unreasonable allocation caused by fixed resource allocation in traditional methods.

[0025] 3) This solution adopts a dual-model collaborative architecture and incorporates an intelligent scheduling algorithm under constraints. When specific conditions are met, the collaboration between the two models will be dynamically adjusted to cope with complex and ever-changing actual working conditions. On the one hand, this enhances the robustness and adaptability of the overall solution, and on the other hand, it provides users with a flexible and efficient detection method. At the same time, the dual-model collaborative architecture, dynamic area division, and enhanced recognition solutions form a closed loop, realizing intelligent management of the entire process from data acquisition to final decision-making, ensuring the efficiency and flexibility of the ultrasonic water immersion flaw detection process.

[0026] 4) This solution introduces CNN into complex scenarios where ERF is prone to misjudgment, which significantly improves the ability to correct key misjudgment points. CNN is only called locally when necessary, avoiding high computing power consumption across the entire area. The overall efficiency is still close to that of the pure ERF solution. Through confidence feedback, it can automatically focus on the ambiguous area, improving the detection rate of small or complex defects. The operation of the overall solution realizes dynamic collaborative operation, which ensures a certain level of accuracy while also ensuring detection efficiency and targeted improvement of defect type recognition, thus ensuring a balance and stability among the three effects. Attached Figure Description

[0027] Figure 1 This is a simplified schematic diagram of the detection method in this invention. Detailed Implementation

[0028] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0029] Please see Figure 1 This embodiment provides an ultrasonic water immersion flaw detection method using intelligent sensors. This method is often used for flaw detection of steam turbine blades, and the specific steps of the method are described below:

[0030] S1. Initialization Configuration:

[0031] Based on the specifications of the target workpiece to be inspected, corresponding smart sensors and their operating frequencies are selected from a pre-built matching sensor library. The type of target workpiece can be a steam turbine blade, engine or gas turbine components, etc., and the type can be selected according to actual needs. In this embodiment, the smart sensors used are mostly cost-effective mid-range sensors. For these types of target workpieces, it is necessary to ensure basic detection accuracy while reducing costs. For example, in this embodiment, in the ultrasonic water immersion testing scenario, a sensor made of piezoelectric ceramic material with high sensitivity and wide bandwidth response is selected as the smart sensor to provide clear and detailed signal feedback, which is particularly important for detecting minute defects. In addition, sensor systems that support modular design can also be selected according to requirements, facilitating flexible configuration adjustments based on different detection needs; however, this will not be elaborated upon here.

[0032] Configure the matching sensor library as follows:

[0033] When detecting target workpieces with a maximum thickness exceeding 5mm, the benchmark operating frequency is 1MHz, which provides better penetration capability. When detecting target workpieces with a maximum thickness not exceeding 5mm, the benchmark operating frequency is 5MHz, requiring a higher frequency to achieve higher lateral resolution and better suitability for detecting minute defects. The larger the target workpiece thickness, the lower the operating frequency should be set, with an inverse relationship between the two. In actual settings, to verify this theory, a series of experiments can be conducted to compare the detection effects at different frequencies, thereby ultimately determining the optimal frequency best suited for the current application scenario. Of course, these experiments are entered into the matching sensor library after confirmation.

[0034] S2. Segmented Data Analysis:

[0035] S2.1 Dynamic Region Division:

[0036] In traditional methods, a uniform grid or a fixed-size rectangle is typically used to divide the area to be inspected. However, when dealing with workpieces with irregular shapes or significant thickness variations, this method may lead to over-scanning of some areas while neglecting other important areas. Therefore, this embodiment proposes the following specific execution scheme in S2:

[0037] The dynamic region partitioning strategy involves acquiring the geometric information of the surface of the target workpiece through 3D scanning and generating the optimal scanning path and partitioning scheme using computer-aided design (CAD) software.

[0038] For example, for aero-engine blades, the root and tip sections may require higher resolution for inspection because these areas are more prone to stress concentration and cracking; in contrast, the middle section can be scanned quickly using lower resolution. This differentiated approach can significantly improve overall efficiency without affecting the final inspection results.

[0039] S2.2 Priority ranking under the guidance of defect prior knowledge:

[0040] A probabilistic model is built using historical data and expert knowledge to identify common defect types (i.e., calibrated defect types) and their possible locations. Based on the output of this probabilistic model, each scanning area is prioritized. The output of the probabilistic model is positively correlated with the priority; the higher the probability, the more serious the potential risk, and the higher the corresponding priority.

[0041] If the priority exceeds the calibrated value, it is marked as high priority and a high-resolution scanning strategy is executed;

[0042] If the priority does not exceed the calibrated value, it is marked as low priority and a low-resolution scanning strategy is executed;

[0043] By implementing a high-resolution scanning strategy in high-priority areas and using low-resolution scanning in other high-priority areas, the overall operating strategy allows the solution to focus more on the detection of potentially serious defects, improves the identification of specific types of defects, and reduces excessive attention to non-critical areas, thus achieving the goal of accurate location and evaluation.

[0044] For example, fatigue cracks often appear near stress concentration points on aircraft engine blades; while at welded joints, problems such as porosity or lack of fusion are more likely to occur. Based on such a probability model, we can sort each scanning area according to its potential risk level and prioritize the areas most likely to have serious defects. For example, if it is known that a certain type of engine blade has failed multiple times in the past few years due to cracks in a specific location, then during the regular maintenance inspection of this type of blade, that location will be automatically set as the highest priority, and a special high-precision scanning task will be arranged even if it is located in a hard-to-reach place. Among them, the resolution exceeding the calibrated resolution index is high resolution, and the opposite is low resolution. There is a clear definition and classification for this, which will not be elaborated or illustrated here.

[0045] Effect description: By applying intelligent sensors and dynamic region division, this invention can use different scanning resolutions for different regions during the detection process. This differentiated processing not only improves the ability to identify subtle defects and effectively reduces the false alarm rate, but also saves certain operating resources. On the one hand, it ensures the accuracy of the detection results, and on the other hand, it ensures the efficiency and reliability of the overall detection process.

[0046] S2.3 Real-time feedback and adaptive adjustment:

[0047] In actual operation, a real-time data analysis mechanism is introduced to perform preliminary defect identification actions on scanning areas of different priorities. When the low-resolution scanning strategy identifies a suspected defect, a strategy change action is triggered to change the original low-resolution scanning strategy to a high-resolution scanning strategy.

[0048] Specifically, the system is allowed to dynamically adjust subsequent scanning plans based on the current situation.

[0049] For example, if a suspected major defect is unexpectedly discovered in a low-priority area, the original plan is immediately suspended and the importance of the area is reassessed. In this embodiment, however, no reassessment is performed; instead, the scanning strategy is directly changed because this is more efficient. Under other more necessary conditions, more resources would be needed to conduct a more in-depth inspection. For example, during a routine inspection, although the initial plan was to perform a comprehensive but relatively coarse scan of the entire blade, some abnormal signals were found when scanning a small area near the edge. The system then suspended the scanning of other areas and performed a higher-resolution re-examination of that area, ultimately confirming a deeply hidden small crack and avoiding potential safety hazards.

[0050] Explanation of the extended solution: Resource management and optimized scheduling operations can be added after S2.3;

[0051] The specific operational process is as follows: Based on the configuration of the smart sensors, including maximum sampling rate and processing speed, a resource allocation strategy is formulated: when the current processing speed is lower than the target setting, backup smart sensors are rented through the cloud computing platform for processing, with the number rented increasing according to the gradient setting until the current processing speed is no lower than the target setting; in the event of an emergency, the detection accuracy of some low-priority scanning areas is temporarily sacrificed to ensure that high-priority scanning areas are fully inspected; specifically, during peak periods, additional computing power can be rented through the cloud computing platform to accelerate the data analysis process; or in the event of an emergency, the detection accuracy of some non-critical areas can be temporarily sacrificed to ensure that core areas are fully inspected. Inspection; for example: In a large manufacturing plant, due to order backlog, the quality inspection of a batch of newly produced engine blades must be completed as soon as possible. In this case, the plant can choose to send a portion of the simpler and lower-risk blades to an external cloud service provider for rapid initial screening, while retaining only the most challenging samples for detailed analysis in the local laboratory, thereby maximizing efficiency. By adopting the above-mentioned extended solution, not only can work efficiency be significantly improved while maintaining high detection accuracy, but it can also provide more accurate location and assessment for specific types of defects. This flexible strategy enables the plant to handle various complex situations with ease, whether in routine maintenance or emergency repair scenarios, reflecting the demand for intelligent and personalized services.

[0052] Effect description: Combining the high and low resolution scanning strategies mentioned above, a real-time data analysis mechanism is further introduced. When a suspected defect is initially detected, a higher resolution re-examination is triggered. At the same time, the subsequent scanning plan is dynamically adjusted according to actual needs. This not only improves the response speed and flexibility of the overall detection solution, but also reduces the computational cost to a certain extent, and realizes the rational planning or utilization of resources. It can solve the problem of unreasonable allocation caused by fixed resource allocation in traditional methods.

[0053] S3. Defect Classification and Identification:

[0054] S3.1, Construction of a dual-model collaborative architecture:

[0055] The scan regions that ultimately implement the high-resolution scanning strategy are marked as critical regions, and the scan regions that ultimately only implement the low-resolution scanning strategy are marked as non-critical regions. A dual-model collaborative architecture is adopted: a primary model and a secondary model.

[0056] The conventional approach is as follows: the main model can be a lightweight convolutional neural network (CNN) to handle critical regions; the auxiliary model can be an enhanced random forest (ERF) to handle non-critical regions.

[0057] It should be noted that the lightweight convolutional neural network is used to process critical areas. Its model architecture uses depthwise separable convolution, which significantly reduces the number of parameters and computational cost, while retaining sensitivity to spatial features such as crack orientation and delamination boundaries. The input is ultrasonic A-scan and B-scan signals (i.e., the original signal), and the output is the probability of defect presence and preliminary category. The preliminary category includes cracks, pores, inclusions, and delamination. The lightweight convolutional neural network can achieve millisecond-level inference in a GPU-accelerated environment, so it can be fully adapted to these critical areas with high accuracy requirements. The enhanced random forest is used to process non-critical areas. Its input is an engineered feature vector extracted from the original signal, which includes signal peak value, rise time, energy integral, spectral centroid, and wavelet coefficient entropy. The enhanced random forest performs well when the feature dimension is low, and the training and inference speed is fast. It is suitable for deployment on edge computing devices, has a strong ability to identify specific defect patterns, such as the high-frequency attenuation features of weld pores, and is robust to noise.

[0058] S3.2, Scheduling and Management Operation:

[0059] A scheduling management mechanism is introduced as a constraint. When using the auxiliary model, the constraint is applied to complete the adjustment action. The operation process of the scheduling management mechanism is as follows:

[0060] S3.2.1 Data Complexity Assessment:

[0061] Analyze the signals in the corresponding non-critical areas to obtain complexity assessment indicators, including noise level, frequency component changes, and energy distribution. Each assessment indicator has a corresponding quantitative value. A comprehensive score is obtained by weighting the complexity assessment indicators to obtain a complexity value. If the complexity value exceeds the preset first standard threshold, the first linear proportional adjustment strategy is triggered; otherwise, no response action is taken.

[0062] The signals mentioned above are defined as follows:

[0063] The raw signals obtained during ultrasonic immersion testing are the ultrasonic A-scan and B-scan signals. The raw signals also contain information about the internal structure of the target workpiece, including the time, amplitude, and frequency components of the reflected echo. The corresponding non-critical area signals are the detection results for non-critical areas.

[0064] Noise level: Measured by the peak signal-to-noise ratio (PSNR) of the signal, which serves as the quantized value of the noise level; the definition of PSNR is: PSNR = 10 × log 10 (Ma I 2 / MSE); where PSNR is the peak signal-to-noise ratio, Ma I The maximum possible value of the signal is given by $SSE$, which represents the mean square error between the signal in the non-critical region and the smoothed background. Therefore, the lower the $PSNR, the higher the noise. Frequency component variation: The Fast Fourier Transform (FFT) is used to analyze the signal's spectral distribution. For continuous segments in the non-critical region, the difference in the spectral center frequency between adjacent segments is calculated, and the maximum value is selected as the quantized value of the frequency component variation. Specifically, the FFT is used to convert the time-domain signal to the frequency-domain signal, and the center frequency of the spectrum, i.e., the spectral centroid, is calculated, defined as: In the formula, CF represents the centroid of the spectrum, and f i For the i-th frequency component, X(f) i Let be the complex amplitude at the i-th frequency. Calculate the difference in spectral center frequencies between adjacent segments, i.e., the difference between the centroids of the corresponding spectral frequencies of adjacent frequency bands. Ensure this difference is not negative by calculating its absolute value. Select the maximum value among all differences as the quantized value of the frequency component change, denoted as FC. Energy distribution: For signals in non-critical areas, calculate their total energy and obtain the information entropy based on the total energy. Use the reciprocal of the information entropy to reflect the concentration of energy distribution and use it as the quantized value of energy distribution. The formula used to calculate the total energy is: In the formula, E represents the total energy, and x[n] represents the time series sample value of the signal. The process of obtaining information entropy based on the total energy is as follows: the signal corresponding to the non-critical region is divided into multiple sub-intervals, and the energy of each interval is calculated. For the energy E of each interval after division, j Calculate the normalized energy distribution p j =E j / E, and then calculate the information entropy Ey: Ey=﹣∑p j log(p j In the formula, j represents the number of each sub-interval. A higher information entropy Ey indicates that the energy distribution is more uniform, while a lower information entropy Ey indicates that the energy is concentrated in certain specific areas. Therefore, the final quantification value of energy distribution is -Ey, which can effectively reflect the degree of concentration of energy distribution.

[0065] When performing weighted calculations based on complexity assessment indicators, the sum of the weights used is 1. In practical applications, it is important to note that all complexity assessment indicators should be normalized to the same range, for example, [0, 1], so that different complexity assessment indicators can be directly added together. If the original range of a certain complexity assessment indicator is too large, it can be scaled to a suitable range through linear transformation or other standardization methods.

[0066] It should be noted that the preset first standard threshold is a critical indicator used to determine whether the first linear proportional adjustment strategy is triggered. Its value is set based on historical data analysis, by calculating the complexity value C of known samples and combining it with the detection results. A value that can distinguish between simple and complex data is selected, for example, 0.6. When the complexity value C exceeds this value, the corresponding strategy is triggered. The purpose is to ensure that defects can still be effectively detected in complex backgrounds. This first standard threshold needs to be adjusted and optimized according to the specific application scenario and workpiece characteristics. The second standard threshold and the judgment threshold mentioned below have similar or the same principles, so they will not be elaborated on here.

[0067] Effect Description: This invention adopts a dual-model collaborative architecture and incorporates an intelligent scheduling algorithm under constraints. When specific conditions are met, the collaboration mode between the two models is dynamically adjusted to cope with complex and ever-changing actual working conditions. On the one hand, this enhances the robustness and adaptability of the overall solution, and on the other hand, it provides users with a flexible and efficient detection method. At the same time, the aforementioned dual-model collaborative architecture, together with the dynamic region division and enhanced recognition methods mentioned in the above steps, forms a complete closed-loop system, realizing intelligent management of the entire process from data acquisition to final decision-making, and ensuring the efficiency and flexibility of the ultrasonic water immersion flaw detection process.

[0068] The process of triggering the first linear scaling adjustment strategy is as follows:

[0069] The proportion of lightweight convolutional neural networks involved is obtained based on the complexity value C. The method is as follows: β_C=min(1,max(0,k_C×(C-C_thd))); where β_C is the proportion of CNN based on the complexity value, k_C is the adjustment coefficient that controls the response slope, and the value range is [0,2], which is 2 in this embodiment; C_thd is the first standard threshold.

[0070] S3.2.2, Confidence Level and Uncertainty Handling:

[0071] The confidence level of the corresponding auxiliary model output is analyzed and compared with the preset second standard threshold. If the confidence level does not exceed the second standard threshold, it is considered that there is identification uncertainty in the corresponding non-critical area, and the second linear scale adjustment strategy is triggered; otherwise, no response action is taken.

[0072] The process of triggering the second linear scaling adjustment strategy is as follows:

[0073] The proportion of lightweight convolutional neural networks involved is obtained based on the confidence level D, using the following method: β_D = max(0, 1 - D / D_thd); where β_D is the proportion of CNNs based on confidence level, and D_thd is the second standard threshold; when the confidence level D approaches 0, β_D approaches 1; when the confidence level D does not exceed the second standard threshold, β_D = 0.

[0074] S3.2.3 Dynamic decision-making mechanism for CNN participation ratio:

[0075] If only one adjustment strategy is triggered, the result of the corresponding adjustment strategy is used as the basis for the fusion probability model, and the fusion probability model is run to output the final probability. If both adjustment strategies are triggered simultaneously, the maximum value is used as the basis for the fusion probability model, and the fusion probability model is run: P_final=a×P_CNN+(1-a)×P_ERF, and the final probability P_final is output. If the final probability exceeds the decision threshold, it is marked as a defect. If the final probability does not exceed the decision threshold, no marking is made. The value of a in the above formula can be either β_C or β_D.

[0076] If any of the triggered adjustment strategies is the first linear scaling adjustment strategy, the corresponding adjustment strategy result is the CNN scaling ratio β_C based on the complexity value; when running the fusion probability model, the method is: P_final=β_C×P_CNN+(1-β_C)×P_ERF; if any of the triggered adjustment strategies is the second linear scaling adjustment strategy, the corresponding adjustment strategy result is the CNN scaling ratio β_D based on the confidence level; when running the fusion probability model, the method is: P_final=β_D×P_CNN+(1-β_D)×P_ERF; when both the first and second linear scaling adjustment strategies are triggered simultaneously, the maximum value is taken as the basis for the fusion probability model, specifically: a=max(β_C, β_D).

[0077] Effect Description: This invention introduces a scheduling management mechanism and adopts a dual-model collaborative architecture and dynamic adjustment strategy. On the one hand, it triggers lightweight CNN fusion analysis based on complexity assessment and confidence judgment in non-critical areas, improving the accuracy of identifying potential defects. On the other hand, it avoids high computing power overhead across the entire area, significantly improving detection efficiency. The overall solution solves the problem of traditional ultrasonic flaw detection struggling to balance efficiency and accuracy, maintaining high robustness even in complex backgrounds or with ambiguous signals. Through weighted complexity scoring and linear scaling, it achieves refined control of model collaboration, demonstrating the adaptive capabilities of the intelligent detection solution in resource allocation and decision optimization. It forms a closed-loop management system from perception to decision, ensuring the efficiency and reliability of the detection process.

[0078] It should be noted that the strategy execution logic varies depending on the region:

[0079] For critical regions, CNN is used to maintain high-precision end-to-end processing without introducing the S3.2 and scheduling management mechanisms mentioned above. For non-critical regions, ERF is used: under normal circumstances, only the ERF model is run. Only when data complexity assessment or confidence constraints (corresponding to S3.2.1 and S3.2.2) are triggered will CNN be automatically called to perform supplementary analysis on the current local region and calculate the fusion ratio according to the above formula to generate the final judgment.

[0080] Example Explanation: In the preliminary ERF analysis, a non-critical area of ​​a blade or workpiece is detected with a confidence level of only 0.55, which does not exceed the second standard threshold of 0.7, and a complexity value of 0.72, which exceeds the preset first standard threshold of 0.6. Therefore, a fusion probability model needs to be run: Calculate: β_C = 2 × (0.72 - 0.6) = 0.24; β_D = 1 - 0.55 / 0.7 ≈ 0.214 (rounded to two decimal places), and take the final a = max(0.24, 0.21) = 0.24; After CNN analysis, the output probability is 0.92, and the ERF output is 0.65. Then the final probability P_final is: P_final = 0.24 × 0.92 + (1 - 0.24) × 0.65 ≈ 0.71 (rounded to two decimal places). Since 0.71 exceeds the judgment threshold of 0.7, it is marked as a defect.

[0081] By adopting the above technical solutions, a balance is achieved between ensuring accuracy, controlling efficiency, and enhancing specific recognition. Specifically, ensuring accuracy involves introducing CNNs in complex scenarios where ERF is prone to misjudgment, significantly improving the ability to correct key misjudgment points. Controlling efficiency means that CNNs are only used locally when necessary, avoiding high computational consumption across the entire area, while maintaining overall efficiency close to the pure ERF solution. Enhancing specific recognition involves automatically focusing on ambiguous areas through confidence feedback, improving the detection rate of small or complex defects. The overall solution achieves dynamic collaborative operation. For flaw detection, it's not necessary to guarantee the highest accuracy. This solution ensures a certain level of accuracy while also guaranteeing detection efficiency and targeted improvement in defect type recognition, maintaining a stable balance among the three aspects.

[0082] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0083] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0084] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An ultrasonic water immersion flaw detection method using intelligent sensors, the steps of which include: Based on the specification information of the target workpiece to be detected, the corresponding smart sensor and its operating frequency are selected from a pre-built matching sensor library; the detection method is characterized by further comprising the following steps: A dynamic region partitioning strategy is adopted to obtain the partitioning scheme of the target workpiece to be inspected: divide each scanning area, establish a probability model, output the probability of the calibrated defect type and its possible location, and the output result of the probability model is positively correlated with the priority, and execute a high or low resolution scanning strategy according to the priority. A real-time data analysis mechanism is introduced to perform preliminary defect identification actions on scanning areas of different priorities; When the low-resolution scanning strategy identifies a suspected defect, a strategy change action is triggered. The scanning regions implementing the high-resolution scanning strategy are marked as critical regions, and the scanning regions implementing the low-resolution scanning strategy are marked as non-critical regions. A dual-model collaborative architecture is adopted: a main model and an auxiliary model. The main model uses a lightweight convolutional neural network (CNN) to handle critical regions, while the auxiliary model uses an enhanced random forest (ERF) to handle non-critical regions. A scheduling management mechanism is introduced as a constraint. When using the auxiliary model, the constraint is applied to complete the adjustment action and generate the final probability. The operation process of the scheduling management mechanism is as follows: A1: Data complexity assessment; A2: Confidence and uncertainty handling; A3: Dynamic decision-making mechanism for the proportion of CNN participation. The data complexity assessment process mentioned in A1 is as follows: Analyze the signal of the corresponding non-critical region to obtain complexity assessment indicators, including at least: noise level, frequency component changes, and energy distribution. Perform weighted calculation based on the complexity assessment indicators to obtain the complexity value. When the complexity value exceeds a preset first standard threshold, the first linear proportional adjustment strategy is triggered; otherwise, no response action is taken. The first linear scaling strategy triggered is as follows: Based on the complexity value C, the proportion of lightweight convolutional neural networks involved is obtained, using the following method: β_C = min(1, max(0, k_C × (C - C_thd))); where β_C is the CNN proportion based on the complexity value, k_C is the adjustment coefficient with a value range of [0, 2], and C_thd is the first standard threshold. The confidence and uncertainty handling process mentioned in A2 is as follows: The confidence level of the corresponding auxiliary model output is analyzed, and the confidence level is compared with the preset second standard threshold. If the confidence level does not exceed the second standard threshold, the second linear scaling strategy is triggered; otherwise, no response is taken. The second linear scaling strategy triggered is as follows: Based on the confidence level D, the proportion of lightweight convolutional neural networks involved is obtained. The proportion of CNN participation is determined by the following method: β_D = max(0, 1-D / D_thd); where β_D is the CNN participation proportion based on confidence, and D_thd is the second standard threshold; the dynamic decision-making mechanism for the CNN participation proportion mentioned in A3 is as follows: when any adjustment strategy is triggered, the result of the corresponding adjustment strategy is used as the basis for the fusion probability model, and the fusion probability model is run to output the final probability; when two first and second linear proportion adjustment strategies are triggered simultaneously, the maximum value is taken as the basis for the fusion probability model, and the fusion probability model is run: P_final = a × P_CNN + (1-a) × P_ERF, and the final probability P_final is output; where a takes the value of β_C or β_D; If the final probability exceeds the judgment threshold, it is marked as a defect; otherwise, no marking is made.

2. The ultrasonic water immersion flaw detection method using intelligent sensors according to claim 1, characterized in that: In the matching sensor library: when detecting a target workpiece with a maximum thickness exceeding 5mm, the benchmark operating frequency is 1MHz; when detecting a target workpiece with a maximum thickness not exceeding 5mm, the benchmark operating frequency is 5MHz. The thickness of the target workpiece is inversely proportional to the value set by the working frequency.

3. The ultrasonic water immersion flaw detection method using intelligent sensors according to claim 1, characterized in that: The dynamic region partitioning strategy adopted is as follows: geometric information of the target workpiece surface is obtained through 3D scanning, and a partitioning scheme is generated using computer-aided design (CAD) software; the process of executing high or low resolution scanning strategies according to priority is as follows: when the priority exceeds the calibration value, it is marked as high priority and a high resolution scanning strategy is executed; otherwise, it is marked as low priority and a low resolution scanning strategy is executed.

4. The ultrasonic water immersion flaw detection method using intelligent sensors according to claim 1, characterized in that: Noise level: The peak signal-to-noise ratio (SNR) of the signal is calculated as the quantized value of the noise level; Frequency component variation: The Fast Fourier Transform (FFT) is used to analyze the spectral distribution of the signal. For continuous segments in non-critical areas, the difference in the center frequency of the spectrum between adjacent segments is calculated, and the maximum value is selected as the quantized value of the frequency component variation; Energy distribution: For signals in non-critical areas, the total energy is calculated, and the information entropy is obtained based on the total energy. The reciprocal of the information entropy is used to reflect the concentration of energy distribution and is used as the quantized value of energy distribution.

Images