Metal component flaw detection system and method

Abstract

The invention relates to the technical field of nondestructive testing, solves the problems of insufficient flaw detection precision and adaptability of a metal component caused by restriction in the prior art, and particularly discloses a flaw detection system and method for the metal component. The flaw detection system comprises a multi-modal sensor group, a cross-modal feature alignment module, a self-adaptive weight distribution module, a defect decision module and a closed-loop control unit, the flaw detection method comprises the following steps: S100, synchronously acquiring data through the multi-mode sensor group; s200, performing space-time registration on the multi-source data of the metal component by using the dynamic calibration matrix; s300, evaluating the confidence coefficient of the sensor in real time, and generating a defect feature map; s400, inputting the defect feature map into a defect classification network, matching a defect map database, and outputting a defect type and a quantization parameter; and S500, automatically adjusting the detection motion track and the sensor according to the defect parameters. According to the multi-method combined detection for the metal component, the detection precision is improved, and a detection result with higher accuracy can be obtained through comprehensive evaluation according to multi-source data.

Description

Technical Field

[0001] The present invention relates to the field of non-destructive testing, and particularly to a flaw detection system and method for metal components. Background Art

[0002] During the forging process of metal components, due to the plastic deformation of materials and the complexity of hot working processes, various internal defects are likely to occur. For example, improper control of forging temperature may lead to uneven distribution of alloy elements, forming intergranular segregation or banded carbide structures, significantly reducing the transverse plasticity and fatigue resistance of materials; residual impurities or incomplete discharge of slag during the forging process will cause non-metallic inclusion defects and damage the material continuity. In addition, thermal stress caused by uneven cooling rates may induce microcracks, and macroscopic cracks are more likely to form due to stress concentration in areas with sudden cross-section changes. Existing detection methods such as ultrasonic testing can identify defects, but they rely highly on the experience of operators and it is difficult to achieve full coverage detection of complex curved surface components. Therefore, it is necessary to accurately evaluate the forging quality through flaw detection technology and eliminate potential failure risks.

[0003] Currently, the weight allocation algorithms of current metal flaw detection systems generally adopt a method combining static weight allocation and dynamic optimization. In static weight allocation, the sensor weights are fixedly allocated according to preset types, but this method is difficult to adapt to dynamic working conditions changes. For example, when the signal-to-noise ratio of ultrasonic waves drops by more than 50% due to temperature fluctuations, the fixed weights will significantly reduce the detection reliability. Dynamic optimization algorithms such as genetic algorithms and particle swarm optimization can optimize the weight combination through search, but the single optimization takes a long time and cannot meet the real-time requirements of high-speed production lines, while Kalman filtering is prone to divergence in non-linear noise scenarios, resulting in inaccurate weight allocation.

[0004] Therefore, there are significant bottlenecks in the existing technology in complex material detection scenarios. For example, in composite materials such as carbon fiber reinforced metals, due to the anisotropic characteristics, the confidence difference of sensor data expands, and the misjudgment rate of existing algorithms is as high as 35%. Moreover, neither static weights nor dynamic optimization can effectively solve the problem of multi-modal data conflict. This technical defect severely restricts the application accuracy and adaptability of flaw detection systems in the high-end manufacturing field. Summary of the Invention

[0005] To solve the problem in the existing technology that the error rate of the flaw detection method is high, resulting in insufficient flaw detection accuracy and adaptability of metal components, the present invention provides a flaw detection system and method for metal components.

[0006] The technical solution adopted by the present invention is:

[0007] A flaw detection and inspection system for metal components, comprising: a multi-modal sensor group, a cross-modal feature alignment module, an adaptive weight allocation module, a defect decision module, and a closed-loop control unit; wherein, the output end of the multi-modal sensor group is connected to the input end of the cross-modal feature alignment module, the output end of the cross-modal feature alignment module is connected to the input end of the adaptive weight allocation module, the output end of the adaptive weight allocation module is connected to the input end of the defect decision module, the output end of the defect decision module is connected to the input end of the closed-loop control unit, and the output end of the closed-loop control unit is respectively connected to the input end of the multi-modal sensor group and a metal component detection execution device;

[0008] The multi-modal sensor group includes an ultrasonic sensor, an X-ray imaging unit, and a laser ultrasonic sensor, and is used for synchronously collecting spatial distribution data and defect feature data of metal components;

[0009] The cross-modal feature alignment module is used for spatio-temporal registration of multi-source data of metal components based on a dynamic calibration matrix, and extracting cross-modal shared features through an attention mechanism to eliminate time drift and spatial offset of sensor data;

[0010] The adaptive weight allocation module is used for real-time evaluating the confidence of each sensor by using a reinforcement learning model, dynamically generating weight coefficients in combination with the temperature of the metal component and the deformation parameters of the metal component material, and selecting data streams for fusion according to the confidence priority;

[0011] The defect decision module is used for identifying new types of metal component defects through a transfer learning model based on a dynamically updated metal component defect feature map library, and outputting the metal component defect type, size, and confidence score;

[0012] The closed-loop control unit is used for real-time adjusting the flaw detection and inspection path and sensor parameters of the metal component according to the result of the defect decision, forming a closed-loop control process from perception to decision and then to execution.

[0013] Further, the cross-modal feature alignment module includes a dynamic calibration matrix generator based on a two-stream neural network, which is used for inputting the time-frequency spectrum of ultrasonic signals and the X-ray density gradient map, and outputting spatio-temporal transformation parameters to realize the offset of sensor data caused by temperature compensation fluctuations.

[0014] Further, the cross-modal feature alignment module includes:

[0015] A multi-modal encoding unit, which is arranged in the parallel computing core of the FPGA, and is used for encoding each modal data into query vectors, key vectors, and value vectors with consistent dimensions; wherein, FPGA is Field-Programmable Gate Array, that is, field programmable gate array, and all FPGAs mentioned in the content recorded in the present invention have the same meaning;

[0016] A cross-modal attention calculation unit, which is arranged in the attention matrix accelerator of the FPGA, is used to calculate the feature similarity weights between different modalities and generate a shared feature matrix;

[0017] A time series modeling unit, which is arranged in the bidirectional LSTM hardware embedded in the FPGA, is used to perform time series modeling on the shared feature matrix, eliminate the time drift error collected by the sensor through the memory gating mechanism, and output the time series aligned features; where LSTM is Long Short-Term Memory, that is, the long short-term memory network, and the LSTM mentioned in the content recorded in the present invention has the same meaning;

[0018] A spatial compensation unit, including an interpolation compensator, a generator and a discriminator, is used to generate sub-pixel level compensation features through adversarial training for texture sparse regions, and correct the spatial offset between X-ray and laser ultrasonic data.

[0019] Further, the adaptive weight allocation module includes:

[0020] A policy network unit, which is used to output the weight coefficients of each sensor based on a lightweight reinforcement learning model by inputting the number of sensors, historical confidence and environmental parameter features;

[0021] A revenue calculation unit, which is used to calculate the revenue of the weight allocation strategy in real time based on the revenue function, obtain the confidence priority, and select the corresponding data stream to drive the parameter update of the policy network unit according to the confidence priority.

[0022] Further, the defect decision module includes:

[0023] A feature extraction unit, which is used to extract multi-scale defect features from the weighted fusion data and output a multi-dimensional feature vector;

[0024] An incremental classification unit, which is used to classify and judge the defect features. When the confidence score of the updated defect features exceeds the set threshold continuously at least three times, the corresponding defect features are marked as new defect feature types, and the data samples of the same type are loaded to record and adjust the labels of the new defect feature types. After the adjustment is completed, the defect type, size and confidence score are output;

[0025] A transfer learning unit, which is used to aggregate new defect features to form a transfer learning model and synchronize data for the detection execution devices for flaw detection of multiple metal components.

[0026] A method for flaw detection of metal components, comprising the following steps:

[0027] S100. Synchronously collect the ultrasonic time-domain signal, X-ray density distribution map, and laser ultrasonic surface deformation data of the metal component through a multi-modal sensor group;

[0028] S200. Perform spatio-temporal registration on the multi-source data of the metal component using a dynamic calibration matrix, and extract cross-modal shared features through an attention mechanism to achieve sub-pixel alignment in texture sparse regions;

[0029] S300. Based on a reinforcement learning model, evaluate the sensor confidence in real time, dynamically assign weights in combination with environmental parameters, and generate a fused high-confidence defect feature map;

[0030] S400. Input the defect feature map into a defect classification network driven by transfer learning, identify new types of defects through the transfer learning model, match the dynamically updated defect atlas library, and output the defect type and quantization parameters;

[0031] S500. Automatically adjust the detection probe movement trajectory and sensor excitation frequency of the metal component detection execution device according to the defect parameters to achieve adaptive closed-loop control of the detection process.

[0032] Furthermore, the method for extracting cross-modal shared features through the attention mechanism in S200 specifically includes:

[0033] S201. Construct a multi-head attention network, and encode the ultrasonic time-frequency spectrum, X-ray density gradient map, and laser ultrasonic surface deformation data into query vectors, key vectors, and value vectors respectively;

[0034] S202. Calculate the feature similarity weights between different modalities through a cross-modal attention layer to generate a shared feature matrix;

[0035] S203. Stack a bidirectional LSTM module in the spatio-temporal dimension to perform temporal correlation modeling on the shared feature matrix and eliminate the time drift error of sensor data;

[0036] S204. Use a generative adversarial network to interpolate and compensate the features in texture sparse regions to achieve sub-pixel spatial offset correction.

[0037] Furthermore, the method for using a reinforcement learning model to evaluate the confidence of each sensor in S300 specifically includes:

[0038] S301. Set up a reinforcement learning model based on a deep network, and the learning parameters include sensor historical confidence metrics, temperature fluctuation values, and material deformation rates;

[0039] S302. Define the action space, which is set as the dynamic adjustment interval of the weight coefficients of each sensor;

[0040] S303. Set the revenue function. Based on the weighted balance of defect prediction accuracy and computing resource consumption, the calculation formula is:

[0041] ;

[0042] where R is the rate of return, α is the accuracy adjustment coefficient, A is the defect prediction accuracy, β is the resource consumption adjustment coefficient, W i is the sensor weight, and C i is the energy consumption coefficient of the sensor;

[0043] S304. According to the rate of return obtained from the revenue function, adopt a combination of offline pre-training and online adjustment. Based on the data exceeding the preset rate of return, update the parameters of the policy network. The pre-training data is obtained from the historical defect database.

[0044] Furthermore, the method for the transfer learning model to identify new types of defects in S400 specifically includes:

[0045] S401. Based on the feature extraction unit, extract multi-scale defect features from the weighted fusion data and output a multi-dimensional feature vector;

[0046] S402. Classify and judge the defect features through the incremental classification unit. When the confidence score of the updated defect features exceeds the set threshold continuously for at least three times, mark the corresponding defect features as a new type of defect feature, and load the data samples of the same type to record and adjust the labels of the new type of defect feature. After the adjustment is completed, output the defect type, size, and confidence score;

[0047] S403. Adopt the transfer learning unit to aggregate the new defect features to form a transfer learning model, and synchronize the data of multiple metal component flaw detection execution devices.

[0048] Furthermore, when performing spatio-temporal registration on the multi-source data of metal components in S200, obtain the core parameters of sensor fusion, including spatio-temporal registration parameters and sensor confidence indicators;

[0049] Among them, the spatio-temporal registration parameters are calibrated by combining the thermal expansion coefficient of the metal component material and the spatial offset of the sensor, and the time axes corresponding to the multi-source data are registered through the temperature change curve of the metal component over time to achieve spatio-temporal registration of the multi-source data; the sensor confidence indicator evaluates the confidence of each sensor by combining the signal-to-noise ratio of the data signal and the historical detection accuracy of the sensor. The signal-to-noise ratio is the ratio of the useful component to the noise in the sensor output signal, and the historical detection accuracy is the proportion of the sensor correctly identifying metal component defects in historical data.

[0050] The beneficial effects of the present invention are:

[0051] Through the collaborative perception of the multi-modal sensor group and the dynamic calibration matrix and attention mechanism of the cross-modal feature alignment module, the present invention solves the problem of spatio-temporal registration failure caused by temperature fluctuations and material deformation in traditional methods. For example, when the temperature change rate exceeds ±5°C / s, the dynamic calibration matrix calibrates the spatio-temporal offset in real time based on the coefficient of thermal expansion. At the same time, the cross-modal attention calculation accelerated by FPGA and the sub-pixel level compensation of the generative adversarial network can achieve an alignment accuracy of more than 97% in texture sparse areas, significantly eliminating multi-modal data conflicts. The adaptive weight allocation module uses reinforcement learning to evaluate the sensor confidence and environmental parameters in real time and dynamically generates weight coefficients. Compared with static weight allocation, the false judgment rate is significantly reduced in metal component detection, and the single decision-making time is lower, meeting the real-time requirements of high-speed production lines. The defect decision module synchronizes new defect features to the dynamically updated atlas library through the transfer learning model and the federated learning mechanism. At the same time, the closed-loop control unit adjusts the probe path and sensor excitation frequency in real time according to the defect parameters, forming a closed-loop feedback from perception to execution, which significantly improves the comprehensive detection efficiency of the system under complex working conditions. Brief Description of the Drawings

[0052] Figure 1 It is a flowchart of the method for detecting flaws in metal components of the present invention. Detailed Embodiments

[0053] The present invention will be described in detail below with reference to the drawings and embodiments.

[0054] The metal component flaw detection system of the present invention consists of the following core modules:

[0055] The multi-modal sensor group, including ultrasonic sensors, X-ray imaging units, and laser ultrasonic sensors, is used to collect the ultrasonic time-domain signals, internal density distribution maps, and surface deformation data of metal components respectively. The sensor group is triggered by a synchronous clock and set with an error <1 μs to ensure the spatio-temporal consistency of multi-source data.

[0056] The cross-modal feature alignment module includes a dynamic calibration matrix generator, a multi-modal attention calculation unit, and a spatio-temporal compensation unit: The dynamic calibration matrix generator is based on a two-stream neural network, inputs the ultrasonic time-frequency spectrum and the X-ray density gradient map, and outputs spatio-temporal transformation parameters, which can be rotation matrices, translation vectors, etc., to compensate the sensor data offset caused by temperature fluctuations in real time. The multi-modal attention calculation unit is implemented in FPGA hardware, calculates the cross-modal feature similarity weights through an 8-head parallel attention mechanism, and generates a shared feature matrix with a dimension that can be set to 512×512. The spatio-temporal compensation unit uses a generative adversarial network to perform sub-pixel level interpolation compensation on low-texture areas, for example, generating compensation features with a resolution of 0.01 mm in the weld area.

[0057] The adaptive weight allocation module includes a policy network unit and a revenue calculation unit. The policy network unit uses a lightweight deep reinforcement learning model. By inputting the historical confidence of sensors, the temperature fluctuation value ΔT, and the material deformation rate ε, it can output a weight coefficient vector. The revenue calculation unit defines a revenue function:

[0058] ;

[0059] where R is the rate of return, α is the accuracy adjustment coefficient, A is the defect prediction accuracy, β is the resource consumption adjustment coefficient, W i is the sensor weight, and C i is the energy consumption coefficient of the sensor. The weight policy can be optimized by maximizing the value of R. As an optional adjustment coefficient setting scheme, α = 0.7 and β = 0.3 can be set, that is:

[0060] ;

[0061] This scheme has a higher proportion of adjusting the contribution rate of accuracy to the revenue of the data stream and is suitable for scenarios with high accuracy requirements. It can be specifically selected and set according to actual needs.

[0062] The defect decision module includes an incremental classification unit, a transfer learning unit, and a closed-loop control unit. In the incremental classification unit, when the confidence score of the new defect feature exceeds the threshold continuously for 3 times, for example, exceeds 0.85, it triggers small-sample learning to update the classification model, small samples, and marks them as new type defects. The transfer learning unit uses a federated learning framework to aggregate multi-device data and synchronously update the global defect map library. The update period can be set to <24 hours. The closed-loop control unit adjusts the detection path and sensor excitation parameters in real time according to the defect parameters. For example, the X-ray tube voltage is dynamically adjusted by ±20%, forming a closed-loop control of perception, decision-making, and execution.

[0063] As Figure 1 shown, the metal component flaw detection method of the present invention can be implemented as follows:

[0064] S100. Synchronously collect the ultrasonic time-domain signal, X-ray density distribution map, and laser ultrasonic surface deformation data of the metal component through a multi-modal sensor group. The ultrasonic sensor has a transmission frequency of 5 MHz, the X-ray imaging unit has a resolution of 0.1 mm / pixel, and the laser ultrasonic sampling rate is 100 kHz.

[0065] S200. Perform spatio-temporal registration on the multi-source data of the metal component using a dynamic calibration matrix, and extract cross-modal shared features through an attention mechanism to achieve sub-pixel alignment in the texture sparse area. The dynamic calibration matrix combines the thermal expansion coefficient α = 1.2×10⁻ 6 / °C, calculate the spatio-temporal transformation parameters, eliminate the time drift error through bidirectional LSTM, and make the error < 0.05 ms.

[0066] S300. Based on the reinforcement learning model, evaluate the sensor confidence in real time, dynamically allocate weights in combination with environmental parameters, and generate a fused high-confidence defect feature map; the weight allocation delay of the reinforcement learning model < 30 ms.

[0067] S400. Input the defect feature map into the defect classification network driven by transfer learning, identify new types of defects through the transfer learning model, match the dynamically updated defect atlas library, and output the defect type and quantization parameters; the transfer learning model uses the ResNet-50 backbone network, the feature extraction dimension is 1024, and the classification accuracy ≥ 92%. ResNet-50 is a representative backbone network architecture of the deep residual network Residual Network, which solves the problems of gradient disappearance and degradation in the training of deep neural networks by introducing the residual learning mechanism;

[0068] S500. Automatically adjust the detection probe movement trajectory and sensor excitation frequency of the metal component detection execution device according to the defect parameters to achieve adaptive closed-loop control of the detection process. Based on this detection path planning, the path redundancy can be reduced by 55% and the probe movement speed can be increased by 40%, thus significantly improving the detection effect and efficiency of the metal component flaw detection system.

[0069] Based on the above implementation manners, the present invention brings the following technical effects:

[0070] The spatio-temporal registration error of the cross-modal feature alignment module ≤ 0.1 mm, compared with about 0.25 mm of the traditional method, the error is significantly reduced, and the detection accuracy of carbon fiber composite materials is increased to 98%. By generating an adversarial network to compensate for low-texture areas, the false alarm rate is reduced to 0.5%. The adaptive weight allocation module can still maintain the detection accuracy ≥ 95% when the sensor signal-to-noise ratio drops by 50%. The resource consumption is reduced by 30%, and the X-ray usage frequency is reduced by 40%. The new defect recognition accuracy of the defect decision module ≥ 92%, compared with about 65% of the traditional method, the recognition accuracy is significantly improved, and the misjudgment rate is reduced from 35% to 5%. The update period of the defect atlas library < 24 hours, supporting rapid response to new types of defects. The closed-loop control unit reduces the detection path redundancy by 55% and increases the detection efficiency to 200 pieces / hour. The X-ray radiation dose is reduced by 50%, achieving higher detection efficiency and detection accuracy with less radiation required for detection.

[0071] As an alternative implementation, the present system can be docked with a digital twin platform to map the detection status of metal components in real time, predict the defect expansion trend through virtual simulation, generate a remaining life assessment report based on the crack growth rate, and thus perform predictive maintenance. The dynamic calibration matrix parameters can be rehearsed in the digital twin to shorten the on-site commissioning time. Multiple devices can also be set for collaborative detection, the present system can be deployed to multiple detection devices and the parameters can be synchronized, and the movement paths can be coordinated based on reinforcement learning to achieve parallel detection. For example, multiple devices can collaboratively cover large components, significantly shortening the detection time and improving the detection efficiency. During the collaborative detection process, collision avoidance can be performed, and the robot movement trajectory can be optimized based on the DDPG algorithm to reduce the collision risk. DDPG (Deep Deterministic Policy Gradient) is the Deep Deterministic Policy Gradient algorithm, which is a deep reinforcement learning method for continuous action spaces.

[0072] The specific implementation of the present invention will be described below through some embodiments.

[0073] Embodiment 1. For the implementation of the collaborative perception and data synchronization of the multimodal sensor group, first, the sensor configuration and synchronization mechanism are set. The ultrasonic sensor uses a 5MHz high-frequency pulse, with a penetration depth of 50mm, a sampling rate of 100kHz, and is equipped with a temperature compensation circuit. The X-ray imaging unit is configured with a complementary metal oxide semiconductor detector with a resolution of 0.1mm / pixel, and the tube voltage dynamic adjustment range is 80kV - 150kV, and a defect density gradient map is generated in real time. The laser ultrasonic sensor measures surface deformation based on the Doppler effect, with an accuracy of 0.01μm, and shares the FPGA clock signal with the ultrasonic sensor.

[0074] Then, preprocessing of data-level fusion is performed. The multimodal data is aligned through a spatial coordinate system to eliminate the initial installation error, so that the offset after calibration is <0.05mm. A sensor data buffer is constructed, and a dual-queue mechanism is used to balance the data flow rate difference.

[0075] Embodiment 2. For the implementation of the spatio-temporal registration optimization of the cross-modal feature alignment module, first, the dynamic calibration matrix generation and compensation are performed. A two-stream neural network architecture is established. For the input branch, for example, the ultrasonic time-frequency spectrum extracts frequency domain features through 3 layers of convolution, and the output dimension is set to 256×256. The X-ray density gradient map extracts spatial features through ResNet-18, and the output dimension is set to 256×256. The fusion output is a dynamic calibration matrix, which includes a rotation matrix, a translation vector, and compensation for temperature fluctuations. ResNet-18 is Residual Network-18, that is, a residual network, which is a deep convolutional neural network. By introducing a residual structure, the degradation problem in the training of deep networks is solved, and the accuracy of the image classification task is significantly improved.

[0076] Adversarial generative network compensation. Sub-pixel compensation features are generated based on the U-Net structure, with the resolution set to 0.01 mm. The discriminator adopts the PatchGAN architecture. Based on this compensation mechanism, data training is performed on paired samples of X-ray and laser ultrasonic in texture-sparse regions, such as weld regions. Among them, the U-Net structure is a U-shaped encoder-decoder network structure, a typical encoder-decoder architecture, which forms a U-shaped topology through a symmetric contraction path (encoder) and an expansion path (decoder); the PatchGAN architecture is a local adversarial discriminant network architecture, a generative adversarial network based on local image patch discrimination.

[0077] Example 3. In this example, the method for extracting cross-modal shared features through the attention mechanism specifically includes:

[0078] S201. Construct a multi-head attention network, and encode the ultrasonic time-frequency spectrum, X-ray density gradient map, and laser ultrasonic surface deformation data into query vectors, key vectors, and value vectors respectively; in the modal feature preprocessing, perform time-frequency analysis on the ultrasonic time-frequency spectrum to generate a time-frequency map of a preset size, and extract deep frequency-domain features through a convolutional network; perform spatial feature extraction on the X-ray density gradient map, and use a residual network to capture the gradient information of the density distribution; perform three-dimensional reconstruction on the laser ultrasonic surface deformation data, and generate geometric features of the surface deformation through a point cloud processing network. In the multi-modal vector mapping, the preprocessed ultrasonic, X-ray, and laser ultrasonic features are respectively input into independent fully connected layers, mapped into query vectors, key vectors, and value vectors, and unified to the same dimension; the multi-head attention mechanism is adopted to allocate vectors of different modalities to multiple parallel subspaces, and each subspace independently calculates the attention relationship to capture the potential cross-modal correlation.

[0079] S202. Calculate the feature similarity weights between different modalities through a cross-modal attention layer to generate a shared feature matrix; in the cross-modal attention layer, perform similarity matching between the query vector of the ultrasonic modality and the key vector of the X-ray modality to generate a feature weight matrix; through normalization processing, for example, use the Softmax function to convert the weight matrix into a probability distribution to represent the correlation strength between different modal features. In the dynamic feature fusion, perform weighted summation on the value vector of the X-ray modality according to the attention weights to generate a preliminary shared feature matrix; introduce the key-value pairs of the laser ultrasonic modality as a supplement to enhance the geometric feature weights in the weld or surface deformation significant regions and suppress noise interference.

[0080] S203. Stack a bidirectional LSTM module in the spatio-temporal dimension to perform temporal correlation modeling on the shared feature matrix and eliminate the time drift error of sensor data; stack a bidirectional long short-term memory network in the temporal dimension of the shared feature matrix to model the context dependence of the feature sequence through memory units; use forget gates and input gates to dynamically adjust the fusion ratio of historical information and current input, and eliminate the time drift error caused by sampling delay or environmental disturbance of sensor data.

[0081] S204. Use a generative adversarial network to interpolate and compensate the features in the texture sparse region to achieve sub-pixel level spatial offset correction. Construct a generative adversarial network. The generator is based on an encoder-decoder structure, inputs a low-resolution shared feature map, and outputs a high-resolution compensated feature map; the discriminator discriminates the difference between the generated features and the real annotation through local feature blocks, and drives the generator to achieve sub-pixel level feature completion in the texture sparse region, such as smooth surfaces or weld areas; adopt an alternating training strategy, and the generator and the discriminator gradually improve the compensation accuracy through adversarial learning, and combine a perceptual loss function to constrain the physical rationality of the generated features.

[0082] Based on the implementation scheme of this embodiment, through the multi-head attention mechanism and dynamic weight allocation, the complementary information of multi-modal data is effectively fused, and millimeter-level alignment accuracy is achieved on complex geometric surfaces such as curved surfaces or hole parts; the time drift error is eliminated through the bidirectional LSTM module, and the spatial offset is compensated by the generative adversarial network, significantly reducing the influence of environmental noise on the detection results, such as the influence of environmental temperature fluctuations or environmental vibrations.

[0083] Embodiment 4. In this embodiment, the method for real-time evaluating the confidence of each sensor by using a reinforcement learning model specifically includes:

[0084] S301. Set up a reinforcement learning model based on a deep network, and the learning parameters include the historical confidence index of the sensor, the temperature fluctuation value, and the material deformation rate.

[0085] S302. Define the action space and set it as the dynamic adjustment interval of the weight coefficients of each sensor.

[0086] S303. Set up a reward function, based on the weighted balance of the defect prediction accuracy and the computing resource consumption, and the calculation formula is:

[0087] ;

[0088] Among them, R is the rate of return, α is the accuracy adjustment coefficient, A is the defect prediction accuracy, β is the resource consumption adjustment coefficient, W i is the sensor weight, and C i is the energy consumption coefficient of the sensor;

[0089] S304. Based on the rate of return obtained from the profit function, the parameters of the policy network are updated by combining offline pre-training and online adjustment, using data beyond the preset rate of return. The pre-training data is obtained from the historical defect database.

[0090] In this embodiment, the confidence of multi-source sensors is evaluated online through a reinforcement learning model, and the weight allocation strategy is dynamically optimized to achieve the multi-objective balance of detection accuracy and resource efficiency; the offline pre-training based on historical defect data is combined with the online incremental policy optimization to construct a noise suppression mechanism for temperature fluctuation and deformation rate, improving the adaptability of the system under complex working conditions; by defining an extensible action space and a profit function, the plug-and-play compatibility of the heterogeneous sensor network and the robustness of low false alarm rate detection are synchronously ensured.

[0091] Embodiment 5. In this embodiment, the method for the transfer learning model to identify new types of defects specifically includes:

[0092] S401. Based on the feature extraction unit, multi-scale defect features are extracted from the weighted fusion data, and a multi-dimensional feature vector is output.

[0093] S402. The defect features are classified and judged through the incremental classification unit. When the confidence score of the updated defect features exceeds the set threshold continuously at least three times, the corresponding defect features are marked as a new type of defect feature, and the labels of the new type of defect feature are recorded and adjusted by loading data samples of the same type. After the adjustment is completed, the defect type, size, and confidence score are output.

[0094] S403. The transfer learning unit is used to aggregate the new defect features to form a transfer learning model, and data synchronization is performed on the detection execution devices for flaw detection of multiple metal components.

[0095] In this embodiment, through the cooperative mechanism of multi-scale defect feature extraction and the incremental classification unit, combined with the dynamic threshold trigger logic, it can be set to reach the confidence level continuously three times, and the adaptive optimization of the same type of samples can be realized, achieving the online incremental identification of new types of defects and the dynamic calibration of classification labels, improving the response speed and classification reliability for unknown defects; based on the feature knowledge distillation and distributed transfer learning framework of the transfer learning model, the feature representation parameters of multiple detection devices are updated synchronously, ensuring the cross-device model generalization ability and the consistency of defect determination criteria. At the same time, the false alarm and missed detection risks are reduced through the dynamic threshold mechanism and the sample self-correction strategy.

[0096] Example 6. In this example, when performing spatio-temporal registration on multi-source data of metal components, core parameters for sensor fusion are obtained, including spatio-temporal registration parameters and sensor confidence metrics. Among them, the spatio-temporal registration parameters are calibrated by combining the thermal expansion coefficient of the metal component material and the spatial offset of the sensor, and the time axes corresponding to the multi-source data are registered through the temperature change curve of the metal component over time, so as to achieve spatio-temporal registration of the multi-source data. The sensor confidence metric combines the signal-to-noise ratio of the data signal and the historical detection accuracy of the sensor to evaluate the confidence of each sensor. The signal-to-noise ratio is the ratio of the useful component to the noise in the sensor output signal, and the historical detection accuracy is the proportion of the correct identification of metal component defects in the historical data by the sensor. In this example, through the calibration of the registration parameters based on the thermal expansion coefficient of the material and the spatial offset of the sensor, thermal expansion compensation and position offset correction in the spatial dimension of the multi-source data are realized. Combined with the time-series dynamic compensation calibration of the temperature change curve, the spatio-temporal reference drift caused by the thermal deformation of the metal component and the sensor installation error is eliminated, and the spatial topology consistency and time-axis alignment accuracy of the cross-sensor data are improved. Based on the two-dimensional confidence evaluation of the signal-to-noise ratio and the historical accuracy, a dynamic weight allocation mechanism for the reliability of sensor data is constructed, and the data suppression strategy for high-noise or low-precision sensors in the multi-source fusion process is optimized, reducing the risk of misregistration and the deviation of the fusion decision, and ensuring the data collaboration quality of the detection system.

[0097] The following are the related implementation manners of some defect characteristic parameters optional in the present invention:

[0098] In terms of geometric features, the defect size is obtained through a flaw detection system, such as the defect length, width, and depth. The projection size is measured by X-ray imaging, and the depth is inverted by laser ultrasonic. The defect shape is obtained, such as morphological classifications like cracks, pores, inclusions, etc. This feature is classified by a transfer learning model based on multi-modal data fusion. In terms of physical features, the density difference of the defect can be obtained, that is, the density difference between the defect area and the matrix material, which is obtained by comparative analysis of X-ray gray values. The abnormal acoustic impedance value can be obtained, that is, the abnormal amplitude of the ultrasonic reflection coefficient at the defect interface, which is obtained by Z = ρ⋅v, that is, acoustic impedance = density × sound velocity. In terms of dynamic features, the defect growth rate can be obtained, that is, the growth rate of fatigue cracks under load, which is obtained by time series analysis of periodic detection data.

[0099] The above embodiments only represent the specific implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A flaw detection system for metal components, characterized in that, It includes a multi-modal sensor group, a cross-modal feature alignment module, an adaptive weight allocation module, a defect decision-making module, and a closed-loop control unit. Among them, the output end of the multi-modal sensor group is connected to the input end of the cross-modal feature alignment module, the output end of the cross-modal feature alignment module is connected to the input end of the adaptive weight allocation module, the output end of the adaptive weight allocation module is connected to the input end of the defect decision-making module, the output end of the defect decision-making module is connected to the input end of the closed-loop control unit, and the output end of the closed-loop control unit is respectively connected to the input end of the multi-modal sensor group and the metal component detection execution device; The multi-modal sensor group includes an ultrasonic sensor, an X-ray imaging unit, and a laser ultrasonic sensor, and is used to synchronously collect the spatial distribution data and defect feature data of metal components; The cross-modal feature alignment module is used to perform spatio-temporal registration on the multi-source data of metal components based on a dynamic calibration matrix, and extract cross-modal shared features through an attention mechanism to eliminate the time drift and spatial offset of sensor data; The adaptive weight allocation module is used to use a reinforcement learning model to evaluate the confidence of each sensor in real time, dynamically generate weight coefficients in combination with the temperature of the metal component and the deformation parameters of the metal component material, and select data streams for fusion according to the confidence priority; The defect decision-making module is used to identify new types of metal component defects through a transfer learning model based on a dynamically updated metal component defect feature map library, and output the metal component defect type, size, and confidence score; The closed-loop control unit is used to adjust the flaw detection path and sensor parameters of metal components in real time according to the results of defect decision-making, and form a closed-loop control process from perception to decision-making and then execution.

2. The flaw detection system for a metal component according to claim 1, wherein, The cross-modal feature alignment module includes a dynamic calibration matrix generator based on a two-stream neural network, which is used to input the time-frequency spectrum of ultrasonic signals and the X-ray density gradient map, and output spatio-temporal transformation parameters to realize the offset of sensor data caused by temperature compensation fluctuations.

3. A flaw detection and inspection system for metal components according to claim 1, characterized in that, The cross-modal feature alignment module includes: A multi-modal encoding unit, which is set in the parallel computing core of the FPGA, and is used to encode each modal data into query vectors, key vectors, and value vectors with the same dimension; A cross-modal attention calculation unit, which is set in the attention matrix accelerator of the FPGA, and is used to calculate the feature similarity weights between different modalities and generate a shared feature matrix; A time series modeling unit, which is set in the bidirectional LSTM hardware embedded in the FPGA, and is used to perform time series modeling on the shared feature matrix, eliminate the time drift error collected by the sensor through a memory gating mechanism, and output time series aligned features; The spatial compensation unit includes an interpolation compensator, a generator, and a discriminator, and is used to generate sub-pixel level compensation features through adversarial training for texture sparse regions to correct the spatial offset of X-ray and laser ultrasonic data.

4. A flaw detection system for metal components according to claim 1, characterized in that, The adaptive weight allocation module includes: A policy network unit, which is used to output the weight coefficients of each sensor based on a lightweight reinforcement learning model by inputting the number of sensors, historical confidence, and environmental parameter features; The revenue calculation unit is used to calculate based on the revenue function, evaluate the revenue of the weight allocation strategy in real time, obtain the confidence priority level, and select the corresponding data stream driving strategy network unit parameter update according to the confidence priority level.

5. A flaw detection system for metal components according to claim 1, characterized in that, The defect decision module includes: The feature extraction unit is used to extract multi-scale defect features from the weighted fusion data and output a multi-dimensional feature vector. The incremental classification unit is used to classify and judge the defect features. When the confidence score of the updated defect features exceeds the set threshold continuously for at least three times, the corresponding defect features are marked as a new type of defect feature, and the same type of data samples are loaded to record and adjust the labels of the new type of defect features. After the adjustment is completed, the defect type, size, and confidence score are output. The transfer learning unit is used to aggregate the new defect features to form a transfer learning model and synchronize the data of multiple metal component flaw detection execution devices.

6. A method for flaw detection and inspection of metal components, characterized in that, It includes the following steps: S100. Synchronously collect the ultrasonic time-domain signal, X-ray density distribution map, and laser ultrasonic surface deformation data of the metal component through the multi-modal sensor group. S200. Use the dynamic calibration matrix to perform spatio-temporal registration on the multi-source data of the metal component, and extract cross-modal shared features through the attention mechanism to achieve sub-pixel alignment of the texture sparse area. S300. Based on the reinforcement learning model, evaluate the sensor confidence in real time, dynamically allocate weights in combination with the environmental parameters, and generate a fused high-confidence defect feature map. S400. Input the defect feature map into the defect classification network driven by transfer learning, identify new types of defects through the transfer learning model, match the dynamically updated defect atlas library, and output the defect type and quantization parameters. S500. Automatically adjust the detection probe movement trajectory and sensor excitation frequency of the metal component detection execution device according to the defect parameters to achieve adaptive closed-loop control of the detection process.

7. A method for flaw detection of a metal component according to claim 6, characterized in that, The method for S200 to extract cross-modal shared features through the attention mechanism specifically includes: S201. Construct a multi-head attention network, and encode the ultrasonic time-frequency spectrum, X-ray density gradient map, and laser ultrasonic surface deformation data into query vectors, key vectors, and value vectors respectively. S202. Calculate the feature similarity weights between different modalities through the cross-modal attention layer to generate a shared feature matrix. S203. Stack a bidirectional LSTM module in the spatio-temporal dimension to perform temporal correlation modeling on the shared feature matrix and eliminate the time drift error of the sensor data. S204. Use the generative adversarial network to interpolate and compensate the features of the texture sparse area to achieve sub-pixel level spatial offset correction.

8. A method for flaw detection of a metal component according to claim 6, characterized in that, The method for S300 to use the reinforcement learning model to evaluate the confidence of each sensor in real time specifically includes: S301. Set up a reinforcement learning model based on a deep network, and the learning parameters include the sensor historical confidence index, temperature fluctuation value, and material deformation rate. S302. Define the action space, which is set as the dynamic adjustment interval of the weight coefficients of each sensor. S303. Set up a revenue function, based on the weighted balance of the defect prediction accuracy and the computing resource consumption, and the calculation formula is: ； Among them, R is the rate of return, α is the accuracy adjustment coefficient, A is the defect prediction accuracy, β is the resource consumption adjustment coefficient, W i is the sensor weight, and C i is the energy consumption coefficient of the sensor; S304. Based on the rate of return obtained from the revenue function, in a manner combining offline pre-training and online adjustment, update the policy network parameters based on the data exceeding the preset rate of return. The pre-training data is obtained from the historical defect database.

9. A flaw detection method for a metal component according to claim 6, characterized in that, The method for the transfer learning model to identify new types of defects in S400 specifically includes: S401. Based on the feature extraction unit, extract multi-scale defect features from the weighted fusion data and output a multi-dimensional feature vector. S402. Through the incremental classification unit, classify and judge the defect features. When the confidence score of the updated defect features exceeds the set threshold continuously for at least three times, mark the corresponding defect features as a new type of defect feature, and load data samples of the same type to record and adjust the labels of the new type of defect feature. After the adjustment is completed, output the defect type, size, and confidence score. S403. Use the transfer learning unit to aggregate the new defect features to form a transfer learning model, and synchronize the data for the detection execution devices for flaw detection of multiple metal components.

10. A method for flaw detection of metal components according to claim 6, characterized in that, When performing spatio-temporal registration on the multi-source data of the metal component in S200, obtain the core parameters of sensor fusion, including spatio-temporal registration parameters and sensor confidence indicators. Among them, the spatio-temporal registration parameters are calibrated by combining the thermal expansion coefficient of the metal component material and the spatial offset of the sensor, and the time axes corresponding to the multi-source data are registered through the temperature change curve of the metal component over time, realizing the spatio-temporal registration of the multi-source data; the sensor confidence indicator evaluates the confidence of each sensor by combining the signal-to-noise ratio of the data signal and the historical detection accuracy of the sensor. The signal-to-noise ratio is the ratio of the useful component to the noise in the sensor output signal, and the historical detection accuracy is the proportion of the correct identification of metal component defects by the sensor in the historical data.

Images