Metal component flaw detection system and method

Abstract

The present invention relates to the field of nondestructive testing technology, solving the problem of insufficient accuracy and adaptability in flaw detection of metal components due to the constraints of existing technologies. The present invention specifically discloses a flaw detection system and method for metal components. The flaw detection system includes a multimodal sensor group, a cross-modal feature alignment module, an adaptive weight distribution module, a defect decision module, and a closed-loop control unit. The flaw detection method includes S100, synchronously collecting data through the multimodal sensor group; S200, performing spatiotemporal registration of multi-source data of metal components using a dynamic calibration matrix; S300, real-time evaluation of sensor confidence and generation of a defect feature map; S400, inputting the defect feature map into a defect classification network, matching it with a defect map library, and outputting defect types and quantitative parameters; S500, automatically adjusting the detection motion trajectory and sensor according to the defect parameters. The present invention is used for multi-method combined detection of metal components and improves detection accuracy, and can obtain more accurate detection results based on comprehensive evaluation of multi-source data.

Description

Technical Field

[0001] The present invention relates to the field of nondestructive testing, and in particular to a metal component flaw detection system and method. Background Art

[0002] During the forging process, metal components are prone to various internal defects due to the plastic deformation of the material and the complexity of the hot working process. For example, improper forging temperature control may lead to uneven distribution of alloy elements, forming intergranular segregation or banded carbide structure, which significantly reduces the transverse plasticity and fatigue resistance of the material; residual impurities or incomplete discharge of slag during the forging process will cause non-metallic inclusion defects and destroy the continuity of the material. In addition, the thermal stress generated by uneven cooling rate may induce micro cracks, and the cross-sectional mutation area is more likely to form macro cracks due to stress concentration. Although existing detection methods such as ultrasonic testing can identify defects, they are highly dependent on the operator's experience 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 to eliminate potential failure risks.

[0003] Current weight allocation algorithms for metal flaw detection systems generally combine static weight allocation with dynamic optimization. In static weight allocation, sensor weights are fixed according to preset types. However, this method struggles to adapt to dynamic operating conditions. For example, when temperature fluctuations cause the ultrasonic signal-to-noise ratio to drop by more than 50%, fixed weights can significantly reduce detection reliability. While dynamic optimization algorithms such as genetic algorithms and particle swarm optimization can optimize weight combinations through search, single optimizations are time-consuming and cannot meet the real-time requirements of high-speed production lines. Kalman filters are prone to divergence in nonlinear noise scenarios, leading to inaccurate weight allocation.

[0004] Therefore, existing technologies face significant bottlenecks in complex material inspection scenarios. For example, the anisotropic properties of composite materials like carbon fiber-reinforced metals lead to wide variations in sensor data confidence, resulting in false positive rates as high as 35% for existing algorithms. Furthermore, neither static weighting nor dynamic optimization can effectively resolve multimodal data conflicts. This technical shortcoming severely restricts the accuracy and adaptability of flaw detection systems in high-end manufacturing applications. Summary of the Invention

[0005] In order to solve the problem that the existing flaw detection method has a high error rate, resulting in insufficient accuracy and adaptability in flaw detection of metal components, the present invention provides a metal component flaw detection system and method.

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

[0007] A metal component flaw detection system includes: a multimodal 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 multimodal 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 multimodal sensor group and a metal component detection execution device;

[0008] The multimodal sensor group includes an ultrasonic sensor, an X-ray imaging unit and a laser ultrasonic sensor, which are used to synchronously collect spatial distribution data and defect feature data of metal components;

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

[0010] 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 based on the temperature of the metal component and the deformation parameters of the metal component material, and select data streams for fusion based on the confidence priority;

[0011] The defect decision 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 atlas library, and output the metal component defect type, size and confidence score;

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

[0013] Furthermore, 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 the ultrasonic signal and the X-ray density gradient map, and output the spatiotemporal transformation parameters to achieve sensor data offset caused by temperature compensation fluctuations.

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

[0015] A multimodal encoding unit, located within the parallel computing core of the FPGA, is used to encode data from each modality into query vectors, key vectors, and value vectors of consistent dimensions. The term "FPGA" refers to a field-programmable gate array (FPGA), and all references to FPGAs in this disclosure have the same meaning.

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

[0017] The timing modeling unit, located within the bidirectional LSTM hardware embedded in the FPGA, is used to perform time series modeling on the shared feature matrix, eliminate time drift errors collected by sensors through a memory gating mechanism, and output timing alignment features. LSTM stands for Long Short-Term Memory, and all LSTMs mentioned in this disclosure have the same meaning.

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

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

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

[0021] The benefit calculation unit is used to evaluate the benefit of the weight allocation strategy in real time based on the benefit function calculation, obtain the confidence priority, and select the corresponding data flow driving strategy network unit parameter update according to the confidence priority.

[0022] Furthermore, the defect decision module includes:

[0023] Feature extraction unit, used to extract multi-scale defect features from weighted fusion data and output multi-dimensional feature vectors;

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

[0025] The transfer learning unit is used to aggregate new defect features to form a transfer learning model and synchronize data for multiple metal component flaw detection execution devices.

[0026] A metal component flaw detection method comprises the following steps:

[0027] S100, synchronously collecting ultrasonic time domain signals, X-ray density distribution maps, and laser ultrasonic surface deformation data of the metal component through a multimodal sensor group;

[0028] S200: Use a dynamic calibration matrix to perform spatiotemporal registration of multi-source data of metal components, and extract cross-modal shared features through an attention mechanism to achieve sub-pixel alignment in texture-sparse areas.

[0029] S300, based on the reinforcement learning model, evaluates sensor confidence in real time, dynamically assigns weights based on environmental parameters, and generates a fused high-confidence defect feature map;

[0030] 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 them with the dynamically updated defect map library, and output the defect type and quantitative parameters;

[0031] S500 , automatically adjusting the detection probe motion 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 of S200 for extracting cross-modal shared features through the attention mechanism specifically includes:

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

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

[0035] S203, superimposing a bidirectional LSTM module in the spatiotemporal dimension, performing time series correlation modeling on the shared feature matrix, and eliminating the time drift error of the sensor data;

[0036] S204: Using a generative adversarial network to interpolate and compensate for features in texture-sparse areas, thereby achieving sub-pixel spatial offset correction.

[0037] Furthermore, the S300 uses a reinforcement learning model to evaluate the confidence of each sensor in real time, specifically including:

[0038] S301, setting a reinforcement learning model based on a deep network, where the learning parameters include a sensor historical confidence index, a temperature fluctuation value, and a material deformation rate;

[0039] S302, defining an action space, and setting it as a dynamic adjustment interval for each sensor weight coefficient;

[0040] S303. Set a profit function based on the weighted balance between defect prediction accuracy and computing resource consumption. The calculation formula is:

[0041] ;

[0042] 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, and W i is the sensor weight, C i is the energy consumption coefficient of the sensor;

[0043] S304. Based on the rate of return obtained from the profit function, the policy network parameters are updated based on data exceeding the preset rate of return by combining offline pre-training and online adjustment. The pre-training data is obtained from the historical defect database.

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

[0045] S401, extracting multi-scale defect features from the weighted fusion data based on a feature extraction unit, and outputting a multi-dimensional feature vector;

[0046] S402: Defect features are classified and judged by the incremental classification unit. When the updated defect feature confidence score exceeds the set threshold for at least three consecutive times, the corresponding defect feature is marked as a new defect feature type, and data samples of the same type are loaded to record and adjust the label of the new defect feature type. After the adjustment is completed, the defect type, size, and confidence score are output;

[0047] S403: A transfer learning unit is used to aggregate new defect features to form a transfer learning model, and data synchronization is performed on multiple metal component flaw detection execution devices.

[0048] Furthermore, when performing spatiotemporal registration on multi-source data of the metal component at S200 , core sensor fusion parameters are obtained, including spatiotemporal registration parameters and sensor confidence indicators;

[0049] Among them, the spatiotemporal registration parameters are calibrated in combination with the thermal expansion coefficient of the metal component material and the sensor spatial offset, and the time axes corresponding to the multi-source data are aligned through the temperature change curve of the metal component over time to achieve spatiotemporal registration of the multi-source data; the sensor confidence index is combined with the data signal-to-noise ratio and the sensor's historical detection accuracy 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 sensor correctly identifying metal component defects in the historical data.

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

[0051] This invention addresses the spatiotemporal registration failures caused by temperature fluctuations and material deformation, often associated with traditional methods, through the dynamic calibration matrix and attention mechanism of the "multimodal sensor group collaborative perception" and "cross-modal feature alignment module." For example, when the temperature change rate exceeds ±5°C / s, the dynamic calibration matrix uses the thermal expansion coefficient to calibrate spatiotemporal offsets in real time. Simultaneously, FPGA-accelerated cross-modal attention computation and sub-pixel compensation using a generative adversarial network achieve alignment accuracy exceeding 97% in sparsely textured areas, significantly eliminating multimodal data conflicts. The "adaptive weight assignment module" uses reinforcement learning to evaluate sensor confidence and environmental parameters in real time and dynamically generates weight coefficients. Compared to static weight assignment, this significantly reduces the false positive rate in metal component inspection and reduces single-decision time, meeting the real-time requirements of high-speed production lines. The "defect decision module" utilizes a transfer learning model and federated learning mechanism to synchronize new defect features with a dynamically updated atlas. Simultaneously, a closed-loop control unit adjusts the probe path and sensor excitation frequency in real time based on the defect parameters, forming a closed-loop feedback loop from perception to execution, significantly improving the system's overall inspection efficiency under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a flow chart of the metal component flaw detection method of the present invention. DETAILED DESCRIPTION

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

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

[0055] A multimodal sensor suite, comprising an ultrasonic sensor, an X-ray imaging unit, and a laser ultrasonic sensor, is used to collect ultrasonic time-domain signals, internal density distribution maps, and surface deformation data from metal components. The sensor suite is triggered by a synchronized clock, set to an error of <1μs, ensuring spatiotemporal consistency among the multi-source data.

[0056] The cross-modal feature alignment module includes a dynamic calibration matrix generator, a multimodal attention calculation unit, and a spatiotemporal compensation unit. The dynamic calibration matrix generator, based on a two-stream neural network, takes as input the ultrasonic time-frequency spectrum and X-ray density gradient map and outputs spatiotemporal transformation parameters, such as rotation matrices and translation vectors, to compensate for sensor data drift caused by temperature fluctuations in real time. The multimodal attention calculation unit, implemented in FPGA hardware, calculates cross-modal feature similarity weights using an eight-head parallel attention mechanism to generate a shared feature matrix with a configurable dimension of 512×512. The spatiotemporal compensation unit utilizes a generative adversarial network to perform sub-pixel interpolation compensation on low-texture areas, generating compensation features with a resolution of 0.01mm in weld seams, for example.

[0057] The adaptive weight allocation module includes a policy network unit and a benefit calculation unit. The policy network unit uses a lightweight deep reinforcement learning model, which inputs the sensor's historical confidence, temperature fluctuation value ΔT, and material deformation rate ε, and outputs a weight coefficient vector. The benefit calculation unit defines the benefit function:

[0058] ;

[0059] 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, and W i is the sensor weight, C i is the energy consumption coefficient of the sensor; the weight strategy can be optimized by maximizing the R value. As an optional adjustment coefficient setting scheme, α=0.7 and β=0.3 can be set, that is:

[0060] ;

[0061] This solution aims to increase the proportion of adjustment accuracy to the data flow contribution rate, and is suitable for scenarios with high precision requirements. Specific settings can be made based on actual needs.

[0062] The defect decision module consists of an incremental classification unit, a transfer learning unit, and a closed-loop control unit. In the incremental classification unit, when the confidence score of a new defect feature exceeds a threshold (e.g., 0.85) three times consecutively, a small-sample learning update is triggered, marking the defect as a new type. The transfer learning unit utilizes a federated learning framework to aggregate data from multiple devices and synchronously update the global defect atlas, with an update cycle of less than 24 hours. The closed-loop control unit adjusts the inspection path and sensor excitation parameters in real time based on defect parameters, such as dynamically adjusting the X-ray tube voltage by ±20%, forming a closed-loop control system for perception, decision-making, and execution.

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

[0064] S100, synchronously collect ultrasonic time domain signals, X-ray density distribution maps and laser ultrasonic surface deformation data of metal components through a multimodal sensor group; the ultrasonic sensor transmission frequency is 5MHz, the X-ray imaging unit resolution is 0.1mm / pixel, and the laser ultrasonic sampling rate is 100kHz.

[0065] S200, using a dynamic calibration matrix to perform spatiotemporal registration of multi-source data of metal components, and extracting cross-modal shared features through an attention mechanism to achieve sub-pixel alignment in texture sparse areas; the dynamic calibration matrix is combined with the thermal expansion coefficient α = 1.2×10⁻ 6 / ℃, calculate the spatiotemporal transformation parameters, and eliminate the time drift error through bidirectional LSTM to make the error <0.05ms.

[0066] S300, based on the reinforcement learning model, evaluates sensor confidence in real time, dynamically assigns weights based on environmental parameters, and generates a fused high-confidence defect feature map; the reinforcement learning model weight assignment delay is <30ms.

[0067] S400 inputs the defect feature map into a transfer learning-driven defect classification network. The transfer learning model identifies new defect types, matches them to a dynamically updated defect map library, and outputs defect types and quantification parameters. The transfer learning model uses the ResNet-50 backbone network with a feature extraction dimension of 1024 and a classification accuracy of ≥92%. ResNet-50 is a representative backbone network architecture for deep residual networks. It introduces a residual learning mechanism to address the vanishing gradient and degradation issues in deep neural network training.

[0068] S500 automatically adjusts the probe motion trajectory and sensor excitation frequency of the metal component inspection actuator based on defect parameters, achieving adaptive closed-loop control of the inspection process. This inspection path planning reduces path redundancy by 55% and increases probe movement speed by 40%, significantly improving the inspection effectiveness and efficiency of the metal component flaw detection system.

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

[0070] The cross-modal feature alignment module achieves a spatiotemporal registration error of ≤0.1mm, significantly reducing the error compared to approximately 0.25mm with traditional methods, and increasing the detection accuracy of carbon fiber composite materials to 98%. By compensating for low-texture areas through a generative adversarial network, the false alarm rate is reduced to 0.5%. The adaptive weight allocation module maintains a detection accuracy of ≥95% even when the sensor signal-to-noise ratio drops by 50%. Resource consumption is reduced by 30%, and the frequency of X-ray use is reduced by 40%. The defect decision module achieves a new defect recognition accuracy of ≥92%, significantly improving the recognition accuracy compared to approximately 65% with traditional methods, and reducing the false positive rate from 35% to 5%. The defect atlas library is updated every 24 hours, enabling rapid response to new defects. The closed-loop control unit reduces detection path redundancy by 55%, increasing detection efficiency to 200 items per hour. X-ray radiation dose is reduced by 50%, achieving higher detection efficiency and accuracy with less radiation required for detection.

[0071] As an optional implementation, this system can be connected to a digital twin platform to map the inspection status of metal components in real time and predict defect expansion trends through virtual simulation. Remaining life assessment reports are generated based on the crack growth rate, allowing predictive maintenance. Dynamic calibration matrix parameters are previewed in the digital twin to shorten on-site commissioning time. Multi-device collaborative inspection can also be set up, and this system can be deployed to multiple inspection devices and parameters are synchronized. The motion paths are coordinated based on reinforcement learning to achieve parallel inspection. For example, multiple devices can collaboratively cover large components, significantly shortening the inspection time and improving inspection efficiency. During the collaborative inspection process, conflict avoidance is performed and the robot's motion trajectory is optimized based on the DDPG algorithm to reduce the risk of collision. DDPG (Deep Deterministic Policy Gradient) is a deep deterministic policy gradient algorithm, a deep reinforcement learning method for continuous action spaces.

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

[0073] In Example 1, to implement collaborative sensing and data synchronization for a multimodal sensor group, the sensor configuration and synchronization mechanism are first established. The ultrasonic sensor uses a 5MHz high-frequency pulse, has a penetration depth of 50mm, a sampling rate of 100kHz, and is equipped with a temperature compensation circuit. The X-ray imaging unit is equipped with a complementary metal oxide semiconductor detector with a resolution of 0.1mm / pixel. The tube voltage can be dynamically adjusted from 80kV to 150kV, and defect density gradient maps are 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] Data-level fusion is then preprocessed, aligning the multimodal data using a spatial coordinate system to eliminate initial installation errors and reduce the post-calibration offset to less than 0.05mm. A sensor data buffer is constructed, employing a dual-queue mechanism to balance differences in data flow rates.

[0075] Example 2, for the implementation of spatiotemporal registration optimization of the cross-modal feature alignment module, first perform dynamic calibration matrix generation and compensation, establish a two-stream neural network architecture, input branches, such as the ultrasonic time-frequency spectrum is subjected to 3 layers of convolution to extract frequency domain features, and the output dimension is set to 256×256, and the X-ray density gradient map is subjected to ResNet-18 to extract spatial features, and the output dimension is set to 256×256. The fusion output dynamic calibration matrix contains the rotation matrix, translation vector and compensation for temperature fluctuations. ResNet-18 is a Residual Network-18, that is, a residual network, a deep convolutional neural network. By introducing the residual structure, it solves the degradation problem in deep network training and significantly improves the accuracy of image classification tasks.

[0076] The Generative Adversarial Network (GAN) compensation algorithm uses a U-Net structure to generate sub-pixel compensation features at a resolution of 0.01mm. The discriminator uses the PatchGAN architecture. This compensation mechanism is trained on paired X-ray and laser ultrasound samples from texture-sparse areas, such as welds. The U-Net architecture is a U-shaped encoder-decoder network, a typical encoder-decoder architecture, forming a U-shaped topology through symmetrical contraction paths (encoder) and expansion paths (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 embodiment, the method for extracting cross-modal shared features through the attention mechanism specifically includes:

[0078] S201. Construct a multi-head attention network to 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 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. Spatial feature extraction is performed on the X-ray density gradient map, and a residual network is used to capture the gradient information of the density distribution. Three-dimensional reconstruction is performed on the laser ultrasonic surface deformation data, and geometric features of the surface deformation are generated through a point cloud processing network. In multimodal vector mapping, the preprocessed ultrasonic, X-ray, and laser ultrasonic features are input into independent fully connected layers, mapped into query vectors, key vectors, and value vectors, and unified to the same dimension. A multi-head attention mechanism is used to assign vectors of different modalities to multiple parallel subspaces, with each subspace independently calculating attention relationships to capture potential cross-modal correlations.

[0079] S202. A cross-modal attention layer calculates feature similarity weights between different modalities to generate a shared feature matrix. Within the cross-modal attention layer, the query vector of the ultrasonic modality is matched against the key vector of the X-ray modality for similarity to generate a feature weight matrix. Through normalization, such as using the Softmax function, the weight matrix is converted into a probability distribution to characterize the strength of associations between features of different modalities. In dynamic feature fusion, the value vectors of the X-ray modality are weighted and summed according to the attention weights to generate a preliminary shared feature matrix. Key-value pairs from the laser ultrasonic modality are introduced as a supplement to enhance geometric feature weights in welds or areas with significant surface deformation to suppress noise interference.

[0080] S203. Superimpose a bidirectional LSTM module on the spatiotemporal dimension to perform time series correlation modeling on the shared feature matrix to eliminate the time drift error of the sensor data; superimpose a bidirectional long short-term memory network on the time series dimension of the shared feature matrix to model the contextual dependency of the feature sequence through the memory unit; use the forget gate and the input gate to dynamically adjust the fusion ratio of historical information and current input to eliminate the time drift error of the sensor data caused by sampling delay or environmental disturbance.

[0081] S204. Use a generative adversarial network to interpolate and compensate for features in sparsely textured areas, achieving sub-pixel spatial offset correction. A generative adversarial network is constructed. The generator, based on an encoder-decoder structure, inputs a low-resolution shared feature map and outputs a high-resolution compensated feature map. The discriminator uses local feature blocks to discriminate between generated features and ground truth annotations, driving the generator to achieve sub-pixel feature completion in sparsely textured areas, such as smooth surfaces or welds. Using an alternating training strategy, the generator and discriminator gradually improve 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 plan of this embodiment, the complementary information of multimodal data is effectively integrated through the multi-head attention mechanism and dynamic weight allocation, and millimeter-level alignment accuracy is achieved on complex geometric surfaces such as curved surfaces or holes; the time drift error is eliminated through the bidirectional LSTM module, and the adversarial generation network compensates for the spatial offset, which significantly reduces the impact of environmental noise on the detection results, such as the impact of ambient temperature fluctuations or ambient vibrations.

[0083] Example 4: In this embodiment, the method of using a reinforcement learning model to evaluate the confidence level of each sensor in real time specifically includes:

[0084] S301, setting a reinforcement learning model based on a deep network, where the learning parameters include a sensor historical confidence index, a temperature fluctuation value, and a material deformation rate;

[0085] S302, defining an action space, and setting it as a dynamic adjustment interval for each sensor weight coefficient;

[0086] S303. Set a profit function based on the weighted balance between defect prediction accuracy and computing resource consumption. 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, and W i is the sensor weight, C i is the energy consumption coefficient of the sensor;

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

[0090] This embodiment uses a reinforcement learning model to online evaluate the confidence of multi-source sensors and dynamically optimize the weight allocation strategy to achieve a multi-objective balance between detection accuracy and resource efficiency. Offline pre-training based on historical defect data is combined with online incremental strategy optimization to construct a noise suppression mechanism for temperature fluctuations and deformation rates, improving the system's adaptability to complex working conditions. By defining an extensible action space and reward function, the system simultaneously ensures plug-and-play compatibility and low false alarm rate detection robustness in heterogeneous sensor networks.

[0091] Example 5: In this example, the method for identifying new defects using a transfer learning model specifically includes:

[0092] S401, extracting multi-scale defect features from the weighted fusion data based on a feature extraction unit, and outputting a multi-dimensional feature vector;

[0093] S402: Defect features are classified and judged by the incremental classification unit. When the updated defect feature confidence score exceeds the set threshold for at least three consecutive times, the corresponding defect feature is marked as a new defect feature type, and data samples of the same type are loaded to record and adjust the label of the new defect feature type. After the adjustment is completed, the defect type, size, and confidence score are output;

[0094] S403: A transfer learning unit is used to aggregate new defect features to form a transfer learning model, and data synchronization is performed on multiple metal component flaw detection execution devices.

[0095] This embodiment uses a collaborative mechanism of multi-scale defect feature extraction and incremental classification units, combined with dynamic threshold triggering logic. This can be set to three consecutive confidence levels, as well as adaptive optimization of similar samples, to achieve online incremental identification of new defects and dynamic calibration of classification labels, thereby improving the response speed and classification reliability of 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 synchronously updated to ensure the generalization capability of the model across devices and the consistency of the defect judgment criteria. At the same time, the dynamic threshold mechanism and sample self-correction strategy are used to reduce the risk of false alarms and missed detections.

[0096] Example 6. In this embodiment, when performing spatiotemporal registration on multi-source data of metal components, core sensor fusion parameters are obtained, including spatiotemporal registration parameters and sensor confidence indicators; wherein, the spatiotemporal registration parameters are combined with the thermal expansion coefficient of the metal component material and the sensor spatial offset for registration and calibration, and the time axes corresponding to the multi-source data are aligned through the temperature change curve of the metal component over time to achieve spatiotemporal registration of the multi-source data; the sensor confidence indicator is combined with the data signal-to-noise ratio and the sensor's historical detection accuracy to perform confidence assessment on 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 sensor correctly identifying metal component defects in the historical data. This embodiment achieves thermal expansion compensation and position offset correction in the spatial dimension of multi-source data through calibration of registration parameters based on the material thermal expansion coefficient and the sensor spatial offset. Combined with the time-series dynamic compensation calibration of the temperature change curve, it eliminates the spatiotemporal reference drift caused by thermal deformation of metal components and sensor installation errors, and improves the spatial topological consistency and time axis alignment accuracy of cross-sensor data. Based on the two-dimensional confidence assessment of signal-to-noise ratio and historical accuracy, a dynamic weight distribution mechanism for sensor data reliability is constructed to optimize the data suppression strategy of high-noise or low-precision sensors in the multi-source fusion process, reduce the risk of misregistration and fusion decision deviation, and ensure the data collaboration quality of the detection system.

[0097] The following are some optional implementations of defect characteristic parameters of the present invention:

[0098] In terms of geometric features, the flaw detection system acquires defect dimensions, such as length, width, and depth. Projected dimensions are measured using X-ray imaging and depth is inverted using laser ultrasonic imaging. Defect shape, such as cracks, pores, and inclusions, is acquired through morphological classification using a transfer learning model after multimodal data fusion. In terms of physical features, the defect density difference (i.e., the density difference between the defect area and the base material) can be acquired through comparative analysis of X-ray grayscale values. The acoustic impedance anomaly value (i.e., the amplitude of the ultrasonic reflection coefficient anomaly at the defect interface) is acquired using the formula Z = ρ⋅v, where acoustic impedance = density × sound velocity. In terms of dynamic features, the defect growth rate (i.e., the growth rate of fatigue cracks under load) can be acquired through time series analysis of periodic inspection data.

[0099] The above-described embodiments merely represent specific implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A metal component flaw detection system, characterized in that: The system comprises a multimodal 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 multimodal 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 multimodal sensor group and the metal component detection execution device; The multimodal sensor group includes an ultrasonic sensor, an X-ray imaging unit and a laser ultrasonic sensor, which are used to synchronously collect spatial distribution data and defect feature data of metal components; The cross-modal feature alignment module is used to perform spatiotemporal registration of multi-source data of metal components based on a dynamic calibration matrix, and extract cross-modal shared features through an attention mechanism to eliminate temporal 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 based on the temperature of the metal component and the deformation parameters of the metal component material, and select data streams for fusion based on the confidence priority; The defect decision 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 atlas library, and output the metal component defect type, size and confidence score; The closed-loop control unit is used to adjust the metal component flaw detection path and sensor parameters in real time according to the results of the defect decision, forming a closed-loop control process from perception to decision and then execution.

2. A metal component flaw detection system according to claim 1, characterized in that: 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 the ultrasonic signal and the X-ray density gradient map, and output the spatiotemporal transformation parameters to achieve sensor data offset caused by temperature compensation fluctuations.

3. A metal component flaw detection system according to claim 1, characterized in that: The cross-modal feature alignment module includes: A multimodal encoding unit, located in the parallel computing core of the FPGA, is used to encode the data of each modality into query vectors, key vectors, and value vectors of consistent dimensions; The cross-modal attention calculation unit is set in the attention matrix accelerator of the FPGA to calculate the feature similarity weights between different modalities and generate a shared feature matrix; The timing modeling unit is set in the bidirectional LSTM hardware embedded in the FPGA. It 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 timing alignment feature; The spatial compensation unit, including an interpolation compensator, a generator, and a discriminator, is used to generate sub-pixel compensation features through adversarial training for texture-sparse areas and correct the spatial offset of X-ray and laser ultrasound data.

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

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

6. A metal component flaw detection method, characterized in that: The following steps are involved: S100, synchronously collecting ultrasonic time domain signals, X-ray density distribution maps, and laser ultrasonic surface deformation data of the metal component through a multimodal sensor group; S200: Use a dynamic calibration matrix to perform spatiotemporal registration of multi-source data of metal components, and extract cross-modal shared features through an attention mechanism to achieve sub-pixel alignment in texture-sparse areas. S300, based on the reinforcement learning model, evaluates sensor confidence in real time, dynamically assigns weights based on environmental parameters, and generates 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 them with the dynamically updated defect map library, and output the defect type and quantitative parameters; S500 , automatically adjusting the detection probe motion 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 metal component flaw detection method according to claim 6, characterized in that: S200's method of extracting cross-modal shared features through the attention mechanism specifically includes: S201, construct a multi-head attention network to encode the ultrasonic time-frequency spectrum, X-ray density gradient map and laser ultrasonic surface deformation data into query vector, key vector and value vector respectively; S202, calculating feature similarity weights between different modalities through a cross-modal attention layer to generate a shared feature matrix; S203, superimposing a bidirectional LSTM module in the spatiotemporal dimension, performing time series correlation modeling on the shared feature matrix, and eliminating the time drift error of the sensor data; S204: Using a generative adversarial network to interpolate and compensate for features in texture-sparse areas, thereby achieving sub-pixel spatial offset correction.

8. A metal component flaw detection method according to claim 6, characterized in that: The S300 uses a reinforcement learning model to evaluate the confidence of each sensor in real time. The specific methods include: S301, setting a reinforcement learning model based on a deep network, where the learning parameters include a sensor historical confidence index, a temperature fluctuation value, and a material deformation rate; S302, defining an action space, and setting it as a dynamic adjustment interval for each sensor weight coefficient; S303. Set a profit function based on the weighted balance between defect prediction accuracy and computing resource consumption. 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, and W i is the sensor weight, C i is the energy consumption coefficient of the sensor; S304. Based on the rate of return obtained from the profit function, the policy network parameters are updated based on data exceeding the preset rate of return by combining offline pre-training and online adjustment. The pre-training data is obtained from the historical defect database.

9. A metal component flaw detection method according to claim 6, characterized in that: The S400 transfer learning model method for identifying new defects specifically includes: S401, extracting multi-scale defect features from the weighted fusion data based on a feature extraction unit, and outputting a multi-dimensional feature vector; S402: Defect features are classified and judged by the incremental classification unit. When the updated defect feature confidence score exceeds the set threshold for at least three consecutive times, the corresponding defect feature is marked as a new defect feature type, and data samples of the same type are loaded to record and adjust the label of the new defect feature type. After the adjustment is completed, the defect type, size, and confidence score are output; S403: A transfer learning unit is used to aggregate new defect features to form a transfer learning model, and data synchronization is performed on multiple metal component flaw detection execution devices.

10. A metal component flaw detection method according to claim 6, characterized in that: When S200 performs spatiotemporal registration on multi-source data of metal components, it obtains sensor fusion core parameters, including spatiotemporal registration parameters and sensor confidence indicators; Among them, the spatiotemporal registration parameters are calibrated in combination with the thermal expansion coefficient of the metal component material and the sensor spatial offset, and the time axes corresponding to the multi-source data are aligned through the temperature change curve of the metal component over time to achieve spatiotemporal registration of the multi-source data; the sensor confidence index is combined with the data signal-to-noise ratio and the sensor's historical detection accuracy 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 sensor correctly identifying metal component defects in the historical data.

Images