Shell manufacturing defect intelligent detection method and system

Abstract

The invention relates to the technical field of manufacturing process management, and discloses a shell manufacturing defect intelligent detection method and system, and the method comprises the steps: multi-modal data collection, dynamic space-time alignment and preprocessing, deep learning model processing, defect classification and positioning, and feedback and production process control. The system corresponds to the method. According to the invention, based on visual image acquisition, infrared thermal imaging, ultrasonic signal acquisition and laser scanning synchronous data acquisition, comprehensive acquisition of surface and internal data of the shell is carried out, then timestamp synchronization and space coordinate transformation are utilized, dynamic space-time alignment of multi-sensor data is carried out, and the data reliability is improved in combination with a redundancy check mechanism. The method comprises the following steps: firstly, performing feature-level and decision-level mixed fusion by adjusting the weight of each modal data in real time, remarkably improving the detection precision of fine defects, and finally adjusting production parameters in the production process based on a defect identification result, thereby ensuring the product quality and improving the control efficiency of the production process.

Description

Technical Field

[0001] The present application relates to the technical field of manufacturing process management, and specifically to a method and system for intelligent detection of shell manufacturing defects. Background Art

[0002] In the chassis manufacturing process of aluminum alloy equipment such as COD detection chassis, set-top box chassis, host chassis, mobile power chassis and speaker amplifier chassis, due to the complexity of material properties (such as high thermal conductivity and easy oxidation) and processing technology (such as casting, stamping, and welding), quality problems such as cold shut defects, air hole defects, and surface scratches often occur on the surface and inside of the chassis shell.

[0003] Existing aluminum alloy defect detection technologies mainly use single-sensor solutions, such as visual inspection, X-ray inspection, or ultrasonic inspection. These methods have obvious technical pain points in practical applications:

[0004] Traditional inspection methods generally rely on manual visual inspection or single-sensor testing, such as blind hole inspection and X-ray inspection. These methods are not only cumbersome and time-consuming, but also difficult to implement on-line inspection. Furthermore, existing technologies have low recognition rates for subtle defects unique to aluminum alloys, such as double-layer film defects, fine cracks, and hairline scratches. This is because traditional machine vision methods are prone to false detection when exposed to complex lighting or surface reflections. Deep learning models also face challenges in recognizing small samples and weak semantic defects in industrial scenarios. For example, even with advanced models such as RetinaNet-AACIDD, the detection accuracy of aluminum alloy casting X-ray images still fails to meet the high requirements of industrial production. Furthermore, the large grain size and highly reflective surface of aluminum alloys make it difficult for traditional inspection methods to accurately identify internal defects and minor surface defects. Furthermore, in actual production environments, inspection equipment needs to operate synchronously with the production line, providing real-time feedback on inspection results to enable timely adjustment of production parameters or rejection of substandard products. However, the inspection speed of existing technologies often becomes a bottleneck in production line efficiency, making them unable to meet this requirement. In addition, existing technologies mostly use a single sensor for defect detection, such as pure visual or pure ultrasonic solutions, which are difficult to fully cover multi-dimensional defect characteristics such as shell surface texture, internal pores and thermal stress distribution.

[0005] Therefore, existing shell defect detection technologies have problems such as single sensor capabilities, insufficient multimodal data coordination, and poor adaptability to dynamic environments. There is an urgent need for intelligent defect detection technology during shell manufacturing to improve production process control efficiency and quality. Summary of the Invention

[0006] The purpose of this application is to provide a method and system for intelligent detection of shell manufacturing defects to solve the technical problems raised in the background technology.

[0007] To achieve the above objectives, this application discloses the following technical solutions:

[0008] In a first aspect, the present application discloses a method for intelligently detecting manufacturing defects of a housing, the method comprising the following steps:

[0009] Multimodal data acquisition: Multimodal data is acquired by synchronously acquiring multidimensional data on the shell surface and interior through multiple sensors. The multidimensional data acquisition includes visual image acquisition, infrared thermal imaging, ultrasonic signal acquisition, and laser scanning data.

[0010] Dynamic spatiotemporal alignment and preprocessing: performing spatiotemporal alignment on the multimodal data, unifying the different sensor data into the same spatiotemporal coordinate system through timestamp synchronization and spatial coordinate transformation, and preprocessing the multimodal data, including noise filtering, contrast enhancement, and feature normalization;

[0011] Deep learning model processing: Feature extraction and fusion of pre-processed data based on a deep learning model. The deep learning model uses a multi-branch structure to process data of different modalities separately, and adjusts the weight of each modality data in real time through a dynamic weight allocation mechanism to perform hybrid fusion at the feature level and decision level.

[0012] Defect classification and location: The trained deep learning model is used to classify and locate the fused features, identify the defect types on the shell surface and inside, and determine the location and size of the defects.

[0013] Feedback and production process control: The recognition results are visualized through 3D reconstruction technology and fed back to the production line in real time so that the operator can adjust the production parameters of the production process based on the feedback.

[0014] Preferably, the deep learning model includes:

[0015] The attention mechanism module based on multiple convolutional blocks is configured to: enhance the attention to defect features in complex backgrounds;

[0016] A multi-scale feature fusion module using a feature pyramid network is configured to: fuse shallow texture features and deep semantic features to improve the detection accuracy of multi-scale defects;

[0017] The dynamic parameter adjustment module optimizes detection parameters based on real-time detection results and is configured to optimize detection parameters through adaptive learning algorithms to adapt to different production environments and defect types.

[0018] Preferably, the dynamic weight allocation mechanism is implemented based on a reinforcement learning algorithm and is configured to automatically adjust the weight of each modality data according to the real-time detection scenario.

[0019] Preferably, the method of optimizing detection parameters by an adaptive learning algorithm to adapt to different production environments and defect types includes:

[0020] Pre-training stage: pre-train the basic feature extraction layer of the deep learning model on a general defect dataset;

[0021] Fine-tuning stage: Fine-tune the model for shell defect samples and update the fusion weights of the feature pyramid network;

[0022] Adaptive optimization: Use online learning algorithms to adjust detection parameters and model hyperparameters in real time to adapt to the production environment and defect types.

[0023] Preferably, the attention mechanism module includes a cascade structure of a channel attention submodule and a spatial attention submodule.

[0024] Preferably, in the multimodal data acquisition step, the visual image acquisition is performed by an industrial-grade CCD camera and a high-frequency light source to perform micron-level resolution imaging, and the laser scanning data is performed by a three-dimensional camera and a robotic arm to cover the entire surface of the shell.

[0025] Preferably, the multimodal data acquisition further comprises:

[0026] The multiple sensors are dynamically scheduled through multi-sensor collaborative control, wherein the multi-sensor collaborative control includes:

[0027] Real-time monitoring of the light intensity, temperature changes and shell surface material properties of the detection environment;

[0028] The consistency of data detected by multiple sensors is verified through a redundant check mechanism to eliminate abnormal data caused by sensor failure or environmental interference.

[0029] Preferably, the identifying defect types on the surface and interior of the shell includes:

[0030] The classification threshold is dynamically adjusted according to the local statistical characteristics of the fused features to classify and identify defects, wherein the adjustment of the classification threshold includes dynamic threshold calculation based on the local standard deviation of the image and the thermal map gradient.

[0031] Preferably, the defect types include cold shut defects, pore defects and surface scratch defects of the aluminum alloy shell, and different classification thresholds are set for different types of defects.

[0032] In a second aspect, the present application discloses an intelligent detection system for shell manufacturing defects, which applies the above-mentioned intelligent detection method for shell manufacturing defects. The system includes:

[0033] A data acquisition module is configured to synchronously acquire multi-dimensional data on the surface and interior of the shell through multiple sensors to obtain multi-modal data, wherein the multi-dimensional data acquisition includes visual images, infrared thermal imaging, ultrasonic signals, and laser scanning data, wherein the visual images are used to identify surface defects, the infrared thermal imaging is used to detect abnormal internal heat distribution, the ultrasonic signals are used to analyze internal structural defects, and the laser scanning data is used to obtain high-precision three-dimensional surface topography;

[0034] a data processing module configured to: perform spatiotemporal alignment on the multimodal data, unify the different sensor data into the same spatiotemporal coordinate system through timestamp synchronization and spatial coordinate transformation, and preprocess the multimodal data, wherein the preprocessing includes noise filtering, contrast enhancement, and feature normalization;

[0035] A deep learning model is configured to: extract and fuse features from the preprocessed data based on the deep learning model, wherein the deep learning model uses a multi-branch structure to process data of different modalities separately, and adjusts the weight of each modality data in real time through a dynamic weight allocation mechanism to perform hybrid fusion at the feature level and the decision level;

[0036] The defect recognition module is configured to: classify and locate the fused features using a trained deep learning model, identify the types of defects on the shell surface and inside, and determine the location and size of the defects;

[0037] The process control module is configured to: visualize the recognition results through three-dimensional reconstruction technology and provide real-time feedback to production.

[0038] Beneficial effects: The present application provides an intelligent method and system for detecting manufacturing defects of shells, which collects data synchronously based on visual image acquisition, infrared thermal imaging, ultrasonic signal acquisition and laser scanning, and comprehensively collects surface and internal data of the shell. It then uses timestamp synchronization and spatial coordinate transformation to unify multi-sensor data into the same coordinate system, and combines a redundant check mechanism to improve data reliability. It then adjusts the weights of each modal data in real time to perform a mixed fusion of the feature level and the decision level, significantly improving the detection accuracy of micron-level defects and millimeter-level defects. Finally, it adjusts the production parameters in the production process based on the defect identification results, thereby ensuring product quality and improving the efficiency of production process control. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 A flowchart of an intelligent detection method for shell manufacturing defects provided in an embodiment of the present application. DETAILED DESCRIPTION

[0041] The following is a clear and complete description of the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0042] In this document, the term "comprising" is intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising..." does not preclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0043] In a first aspect, this embodiment provides a Figure 1 A method for intelligently detecting manufacturing defects of a shell is shown, the method comprising the following steps:

[0044] S1-Multimodal Data Acquisition: Multimodal data is acquired by synchronously collecting multi-dimensional data on the shell surface and interior using multiple sensors. This multi-dimensional data acquisition includes visual image acquisition, infrared thermal imaging, ultrasonic signal acquisition, and laser scanning. Visual images are used to identify surface defects, infrared thermal imaging is used to detect internal thermal distribution anomalies, ultrasonic signals are used to analyze internal structural defects, and laser scanning data is used to obtain high-precision three-dimensional surface topography.

[0045] S2-Dynamic Spatiotemporal Alignment and Preprocessing: Spatiotemporal alignment of the multimodal data is performed. Through timestamp synchronization and spatial coordinate transformation, the different sensor data are unified into the same spatiotemporal coordinate system to ensure the consistency of the data in time and space. The multimodal data is also preprocessed. The preprocessing includes noise filtering, contrast enhancement, and feature normalization to improve the accuracy of subsequent analysis. In this embodiment, the steps for implementing spatiotemporal alignment include:

[0046] (1) Send a synchronous trigger signal to all sensors through an external clock source (such as the IEEE 1588 protocol clock) to ensure that each sensor starts collecting data at the same time. For example, an industrial camera and a laser scanner share the same trigger pulse signal; an ultrasonic probe and an infrared thermal imager synchronize sampling through a hardware interface.

[0047] (2) Align the coordinate systems of different sensors to the same reference through cross-calibration methods (such as feature point matching). For example, match the feature points of an industrial camera and a laser scanner calibration plate and calculate the coordinate transformation matrix (rotation and translation parameters) between the two.

[0048] (3) For situations where there is no relative motion between the sensor and the shell during data acquisition (such as visual image acquisition and laser scanning), the coordinate system is aligned using the rotation matrix R and the translation vector T. For situations where the sensor viewing angles differ greatly (such as ultrasonic probes and infrared thermal imagers), a nonlinear mapping algorithm (such as thin plate spline interpolation TPS) is used to correct spatial deformation.

[0049] (4) Verify the spatial consistency of different sensor data through feature matching algorithms (such as SIFT and ICP). For example, compare the edge features of industrial camera images and laser scanning point clouds to eliminate abnormal data caused by sensor failure or environmental interference; if the spatial position deviation between infrared thermal imaging and ultrasonic signals exceeds a threshold, an alarm is triggered and recalibration is required.

[0050] S3-deep learning model processing: Feature extraction and fusion of preprocessed data based on a deep learning model, wherein the deep learning model uses a multi-branch structure to process data of different modalities respectively, and adjusts the weight of each modal data in real time through a dynamic weight distribution mechanism (such as a reinforcement learning controller) to perform hybrid fusion of feature level and decision level. It should be noted that the deep learning model with a multi-branch structure used in this embodiment can be any one of the existing technologies, and its basic framework can be, for example, the Inception module from the GoogLeNet architecture. Hybrid fusion of feature level and decision level refers to the integration of information at both the feature level and the decision level to improve the accuracy and robustness of defect detection. Among them, feature-level fusion is the direct integration of raw data or low-level features (such as image texture, thermal map gradient, ultrasonic signal intensity) from different sensors during data preprocessing and feature extraction to form a unified feature vector or tensor for use by subsequent classification models. It is achieved by fusing multimodal features (such as visual images, infrared thermal maps, ultrasonic signals) into a unified feature vector through FPN and attention mechanism and a deep learning model with shared feature vector input for global feature learning. Decision-level fusion is a process that, after each modality has independently completed feature extraction and classification, weightedly integrates the classification results (such as defect type and confidence level) of different sensors to form the final decision output. This is achieved by outputting the independent decisions of each modality through a multi-branch structure after feature fusion and weighted integration of the decision results using a dynamic weight allocation mechanism.

[0051] S4-Defect classification and location: Classify and locate the fused features through the trained deep learning model, identify the defect types on the shell surface and inside, and determine the location and size of the defects.

[0052] S5 - Feedback and Production Process Control: 3D reconstruction technology is used to visualize recognition results and provide real-time feedback to the production line, allowing operators to adjust production parameters based on this feedback. 3D reconstruction can be any existing technology, such as structured light 3D reconstruction, a mature technology widely used in industrial inspection, particularly for high-precision surface topography reconstruction. This technology projects a known pattern onto the object being measured, captures the deformed light pattern with a camera, and then calculates the 3D coordinates of the object's surface using triangulation principles. In addition, an example of the operator adjusting the production parameters of the production process based on feedback is: when a cold shut defect is detected during the casting process of an aluminum alloy chassis, a three-dimensional point cloud model of the shell surface is generated through structured light scanning and laser scanning data. The deep learning model identifies the cold shut defect and highlights the defect position and size in red in the three-dimensional model. The system transmits the defect information (such as defect depth, area, and distribution area) to the production line control interface in real time. Since cold shut defects are usually caused by excessively high mold temperature or too slow cooling rate, the operator adjusts the cooling water flow rate of the casting mold from 4L / min to 6L / min based on the feedback information to accelerate cooling and reduce the flow resistance of the molten metal. The pouring speed can also be reduced from 1.2m / s to 0.8m / s at the same time to avoid stratified flow of molten metal in the mold.

[0053] In one embodiment, the deep learning model includes:

[0054] The attention mechanism module, based on multiple convolutional blocks, is configured to enhance attention to defect features in complex backgrounds. It is a deep learning module that combines multi-scale convolution operations with an attention mechanism. This module extracts local to global multi-scale features using convolution kernels of different scales (e.g., 3×3, 5×5, 7×7). Attention weights are calculated separately in the channel and spatial dimensions, adaptively weighting the multi-scale features to highlight key features and suppress noise. This module aims to extract multi-scale features through multi-branch convolution blocks and dynamically enhance key features using attention mechanisms (e.g., channel attention and spatial attention).

[0055] The multi-scale feature fusion module, which uses a feature pyramid network, is configured to fuse shallow texture features with deep semantic features to improve the detection accuracy of multi-scale defects. It is a classic deep learning architecture for multi-scale feature fusion. The core idea is to organically fuse low-level features (high resolution, detailed information) with high-level features (low resolution, strong semantic information) through bottom-up and top-down paths and lateral connections, thereby improving the model's detection capabilities for multi-scale objects.

[0056] The dynamic parameter adjustment module optimizes detection parameters based on real-time detection results. It is configured to optimize detection parameters, including sensor scanning speed and model thresholds, to adapt to different production environments and defect types using an adaptive learning algorithm. This module dynamically adjusts detection parameters based on real-time detection results to accommodate varying production environments (e.g., varying light intensity, material properties) and defect types (e.g., cold shuts, pores, scratches). The core concept is to utilize online learning algorithms and feedback mechanisms to optimize model or sensor parameter configurations in real time, thereby improving detection accuracy and robustness.

[0057] Furthermore, the optimization of detection parameters by the adaptive learning algorithm to adapt to different production environments and defect types includes:

[0058] Pre-training phase: Pre-train the basic feature extraction layer of the deep learning model on a common defect dataset (such as ImageNet-Defect). The basic feature extraction layer refers to the initial part of the model used to extract common features (such as edges, textures, shapes, and other low-level visual patterns) from the input data. It is usually composed of multiple convolutional layers and pooling layers. The core goal is to enable the model to learn cross-domain common feature representations through pre-training on large-scale common datasets, providing efficient feature extraction capabilities for aluminum alloy shell defect detection.

[0059] Fine-tuning: The model is fine-tuned for shell defect samples (such as shell crack morphology distribution) and the fusion weights of the feature pyramid network are updated. Model fine-tuning refers to further training the model based on the existing pre-trained deep learning model for shell defect detection tasks (such as crack morphology distribution identification) using defect sample data from a specific field (such as aluminum alloy shell crack images) to optimize the model's feature extraction and fusion capabilities.

[0060] Adaptive optimization: Use online learning algorithms (such as FTRL) to adjust detection parameters and model hyperparameters in real time to adapt to the production environment and defect type. Detection parameters refer to runtime parameters that directly affect defect recognition results during the detection process, such as classification thresholds, noise filtering strength, feature extraction scale, etc. These parameters are usually dynamically adjusted during the model inference phase to adapt to changes in the production environment (such as lighting fluctuations, material differences) or changes in defect types (such as differences in detection requirements for scratches and pores). Model hyperparameters refer to parameters that need to be manually set before training a deep learning model, such as learning rate, batch size, number of network layers, activation function type, etc. These parameters directly affect the learning efficiency and performance of the model.

[0061] It is feasible that the dynamic weight allocation mechanism is implemented based on a reinforcement learning algorithm and is configured to automatically adjust the weight of each modal data according to the real-time detection scenario, such as increasing the weight of infrared data under strong light interference, and increasing the proportion of ultrasonic data in complex internal structure detection.

[0062] In this embodiment, the dynamic weight allocation mechanism belongs to the cross-modal data fusion strategy, and its core is to dynamically adjust the contribution weights of different modal data according to the real-time detection scene during the feature fusion process of multimodal data (such as vision, infrared, ultrasound, etc.). For example, the reinforcement learning controller determines which sensor data in the current scene (such as infrared thermal imaging in temperature anomaly detection, ultrasound in internal defect detection) is more critical for defect identification, and then assigns a higher weight to achieve intelligent collaboration between modalities, which is used to solve the problem of how to efficiently fuse different sensor data. It belongs to the resource allocation strategy between modalities. The implementation process of reinforcement learning is: using the fusion result of multimodal data (such as defect detection accuracy) as the feedback signal, iteratively optimizes the weight parameters of each modality. For example, when there is strong light interference in the detection environment, the system automatically reduces the weight of the visual image and increases the weight of the infrared data through reinforcement learning to ensure detection robustness.

[0063] It is feasible that the attention mechanism module includes a cascade structure of a channel attention submodule and a spatial attention submodule.

[0064] In this embodiment, the attention mechanism module belongs to a single-modal feature processing technology, which focuses on screening the features inside a single modality (such as images, point clouds, etc.), and enhances the key information of the feature map in stages through the order of channel first and space later, suppresses the interference of background noise, and thus improves the deep learning model's ability to extract target features. For example, the attention mechanism module uses the combined effect of spatial attention and channel attention to allow the model to pay more attention to the texture, shape and other features of the defective area in the image, which is used to solve the problem of how to highlight effective features in a single modality. It is a feature-level refinement processing method. It is implemented as follows: based on a convolutional neural network or a Transformer architecture, by designing an attention sub-network (such as a spatial attention sub-module and a channel attention sub-module), a weighted operation is performed on the input feature map. For example, in the processing of visual images, the MCBAM module generates an attention mask that gives higher weights to pixels in the defective area, thereby enhancing the model's sensitivity to defects.

[0065] In one embodiment, the multimodal data acquisition further includes:

[0066] The multiple sensors are dynamically scheduled through multi-sensor collaborative control, wherein the multi-sensor collaborative control includes:

[0067] Real-time monitoring of the light intensity, temperature changes and shell surface material properties of the detection environment; wherein, the shell surface material properties specifically refer to the physical, chemical and mechanical properties exhibited by the aluminum alloy material and its surface treatment process, including but not limited to: chemical composition affecting the shell strength, corrosion resistance and processing performance, density value affecting the shell lightweight, thermal conductivity affecting the shell high temperature resistance, electrical conductivity affecting the shell insulation performance, film thickness affecting corrosion resistance and wear resistance, surface roughness affecting the shell coating adhesion and imaging effect of optical detection, texture characteristics affecting visual image recognition effect, etc. These characteristics can be obtained through any corresponding existing detection technology, and the corresponding data can be obtained in combination with the production process of the production line;

[0068] The consistency of data detected by multiple sensors is verified through a redundant verification mechanism (such as Kalman filtering) to eliminate abnormal data caused by sensor failure or environmental interference.

[0069] In one embodiment, identifying the defect types on the surface and interior of the shell includes:

[0070] The classification threshold is dynamically adjusted based on the local statistical characteristics of the fused features (such as mean, variance, and distribution morphology) to classify and identify defects more accurately. The adjustment of the classification threshold includes dynamic threshold calculation based on the local standard deviation of the image and the gradient of the heat map. The classification threshold is the dividing point where the model maps the predicted probability (between 0 and 1) to the category label (such as "defect" or "normal"). The threshold is dynamically adjusted based on local statistical characteristics. The following steps are involved:

[0071] (1) Slide a window (e.g., 5×5 or 7×7) on the fused feature map, calculate the mean, variance, or distribution parameters of the features in each window, and divide the image into multiple regions (e.g., defect candidate region and background region) based on the gradient or edge information of the feature map, and calculate the statistical characteristics of each region;

[0072] (2) Based on local statistical characteristics and combined with online learning algorithms, the classification threshold is dynamically optimized according to the real-time feedback of classification results (such as false detection rate and missed detection rate).

[0073] Among them, the defect types include cold shut defects, porosity defects and surface scratch defects of aluminum alloy shells, and different classification thresholds are set for different types of defects. The purpose of this is to balance the accuracy and recall rate of the detection system and adapt to the characteristics, distribution patterns and detection targets of the defects. The setting method of different classification thresholds can be: dynamically adjusting the threshold according to real-time production data (when it is detected that the mold temperature increases and causes an increase in cold shut defects, the cold shut detection threshold is automatically lowered), and dynamically adjusting the threshold in combination with multimodal data (such as visual images, infrared thermal images) (such as for crack detection, if the infrared thermal image shows a high temperature area, the crack detection threshold of the visual image is lowered). In specific scenarios, for example, for scratch detection, the initial classification threshold is set to 0.7. When most scratches on the production line are shallow (depth <0.1mm) and the background noise is low (such as uniform lighting), if the false alarm rate is too high (such as false detection due to reflection), the threshold can be dynamically increased to 0.8; for cold shut defect detection, the initial classification threshold is set to 0.4. The cold shut feature is a weak gradient of interlayer separation (such as the low-temperature area in the thermal image), and a low threshold is required to ensure no missed detection. If the missed detection rate is too high (such as due to mold temperature fluctuations), the threshold can be further reduced to 0.35; for crack detection, the initial classification threshold is set to 0.55. Since cracks may expand, high sensitivity is required to detect early cracks. If the false alarm rate is too high (such as due to image noise), the threshold can be dynamically adjusted in combination with infrared thermal imaging data.

[0074] In summary, the intelligent shell manufacturing defect detection method of this embodiment, through multimodal data fusion, deep learning model optimization, and adaptive learning algorithms, has constructed a high-precision, highly adaptable, and real-time feedback intelligent shell defect detection technology. Specifically, through online learning algorithms, it optimizes detection parameters and model hyperparameters in real time to adapt to changes in the production environment; combines multi-source data such as vision, thermal imaging, and ultrasound to improve the comprehensiveness and robustness of defect identification; and directly feeds back detection results to the production process to achieve defect prevention and process optimization. Therefore, the intelligent shell manufacturing defect detection method of this embodiment can be widely used in industrial scenarios such as aluminum alloy chassis and precision casting, and can improve detection efficiency and product quality.

[0075] In a second aspect, this embodiment provides a system for intelligent detection of shell manufacturing defects, which applies the above-mentioned intelligent detection method for shell manufacturing defects. The system includes:

[0076] A data acquisition module is configured to synchronously acquire multi-dimensional data on the surface and interior of the shell through multiple sensors to obtain multi-modal data, wherein the multi-dimensional data acquisition includes visual images, infrared thermal imaging, ultrasonic signals, and laser scanning data, wherein the visual images are used to identify surface defects, the infrared thermal imaging is used to detect abnormal internal heat distribution, the ultrasonic signals are used to analyze internal structural defects, and the laser scanning data is used to obtain high-precision three-dimensional surface topography;

[0077] a data processing module configured to: perform spatiotemporal alignment on the multimodal data, unify the different sensor data into the same spatiotemporal coordinate system through timestamp synchronization and spatial coordinate transformation, and preprocess the multimodal data, wherein the preprocessing includes noise filtering, contrast enhancement, and feature normalization;

[0078] A deep learning model is configured to: extract and fuse features from the preprocessed data based on the deep learning model, wherein the deep learning model uses a multi-branch structure to process data of different modalities separately, and adjusts the weight of each modality data in real time through a dynamic weight allocation mechanism to perform hybrid fusion at the feature level and the decision level;

[0079] The defect recognition module is configured to: classify and locate the fused features using a trained deep learning model, identify the types of defects on the shell surface and inside, and determine the location and size of the defects;

[0080] The process control module is configured to: visualize the recognition results through three-dimensional reconstruction technology and provide real-time feedback to production.

[0081] It should be noted that the intelligent detection system for shell manufacturing defects in this embodiment corresponds to the aforementioned intelligent detection method for shell manufacturing defects. Therefore, the parts that are not described in detail in the intelligent detection system for shell manufacturing defects in this text (including but not limited to technical effects, specific technical means, etc.) can refer to the relevant descriptions in the aforementioned intelligent detection method for shell manufacturing defects, and this text will not go into details here.

[0082] In the embodiments provided herein, it should be understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, code, or any appropriate combination thereof. For hardware implementation, the processor can be implemented in one or more of the following units: an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a processor, a controller, a microcontroller, a microprocessor, other electronic units designed to implement the functions described herein, or a combination thereof. For software implementation, part or all of the processes of the embodiments can be completed by instructing the relevant hardware through a computer program. When implemented, the above program can be stored in a computer-readable storage medium or transmitted as one or more instructions or codes on a computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein the communication media include any medium that facilitates the transmission of a computer program from one place to another. The storage medium can be any available medium that a computer can access. The computer-readable storage medium can include, but is not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store the desired program code in the form of instructions or data structures and can be accessed by a computer.

[0083] Finally, it should be noted that the above is only a preferred embodiment of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent replacements for some of the technical features therein. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for intelligent detection of shell manufacturing defects, characterized in that: The method comprises the following steps: Multimodal data acquisition: Multimodal data is acquired by synchronously acquiring multidimensional data on the shell surface and interior through multiple sensors. The multidimensional data acquisition includes visual image acquisition, infrared thermal imaging, ultrasonic signal acquisition, and laser scanning data. Dynamic spatiotemporal alignment and preprocessing: performing spatiotemporal alignment on the multimodal data, unifying the different sensor data into the same spatiotemporal coordinate system through timestamp synchronization and spatial coordinate transformation, and preprocessing the multimodal data, including noise filtering, contrast enhancement, and feature normalization; Deep learning model processing: Feature extraction and fusion of pre-processed data based on a deep learning model. The deep learning model uses a multi-branch structure to process data of different modalities separately, and adjusts the weight of each modality data in real time through a dynamic weight allocation mechanism to perform hybrid fusion at the feature level and decision level. Defect classification and location: The trained deep learning model is used to classify and locate the fused features, identify the defect types on the shell surface and inside, and determine the location and size of the defects. Feedback and production process control: The recognition results are visualized through 3D reconstruction technology and fed back to the production line in real time so that the operator can adjust the production parameters of the production process based on the feedback.

2. The intelligent detection method for shell manufacturing defects according to claim 1 is characterized in that: The deep learning model includes: The attention mechanism module based on multiple convolutional blocks is configured to: enhance the attention to defect features in complex backgrounds; A multi-scale feature fusion module using a feature pyramid network is configured to: fuse shallow texture features and deep semantic features to improve the detection accuracy of multi-scale defects; The dynamic parameter adjustment module optimizes detection parameters based on real-time detection results and is configured to optimize detection parameters through adaptive learning algorithms to adapt to different production environments and defect types.

3. The intelligent detection method for shell manufacturing defects according to claim 1 or 2, characterized in that: The dynamic weight allocation mechanism is implemented based on a reinforcement learning algorithm and is configured to automatically adjust the weight of each modality data according to the real-time detection scenario.

4. The intelligent detection method for shell manufacturing defects according to claim 2, characterized in that: The adaptive learning algorithm is used to optimize the detection parameters to adapt to different production environments and defect types, including: Pre-training stage: pre-train the basic feature extraction layer of the deep learning model on a general defect dataset; Fine-tuning stage: Fine-tune the model for shell defect samples and update the fusion weights of the feature pyramid network; Adaptive optimization: Use online learning algorithms to adjust detection parameters and model hyperparameters in real time to adapt to the production environment and defect types.

5. The intelligent detection method for shell manufacturing defects according to claim 2, characterized in that: The attention mechanism module includes a cascade structure of a channel attention submodule and a spatial attention submodule.

6. The intelligent detection method for shell manufacturing defects according to claim 1, characterized in that: In the multimodal data acquisition step, the visual image acquisition is performed by an industrial-grade CCD camera and a high-frequency light source to perform micron-level resolution imaging, and the laser scanning data is performed by a three-dimensional camera and a robotic arm to cover the entire surface of the shell.

7. The intelligent detection method for shell manufacturing defects according to claim 1, characterized in that: The multimodal data collection further includes: The multiple sensors are dynamically scheduled through multi-sensor collaborative control, wherein the multi-sensor collaborative control includes: Real-time monitoring of the light intensity, temperature changes and shell surface material properties of the detection environment; The consistency of data detected by multiple sensors is verified through a redundant check mechanism to eliminate abnormal data caused by sensor failure or environmental interference.

8. The intelligent detection method for shell manufacturing defects according to claim 1, characterized in that: The identification of defect types on the shell surface and inside includes: The classification threshold is dynamically adjusted according to the local statistical characteristics of the fused features to classify and identify defects, wherein the adjustment of the classification threshold includes dynamic threshold calculation based on the local standard deviation of the image and the thermal map gradient.

9. The intelligent detection method for shell manufacturing defects according to claim 8, characterized in that: The defect types include cold shut defects, pore defects and surface scratch defects of the aluminum alloy shell, and different classification thresholds are set for different types of defects.

10. An intelligent detection system for shell manufacturing defects, characterized in that: The method for intelligent detection of shell manufacturing defects according to any one of claims 1 to 9 is applied, and the system comprises: A data acquisition module is configured to synchronously acquire multi-dimensional data on the surface and interior of the shell through multiple sensors to obtain multi-modal data, wherein the multi-dimensional data acquisition includes visual images, infrared thermal imaging, ultrasonic signals, and laser scanning data, wherein the visual images are used to identify surface defects, the infrared thermal imaging is used to detect abnormal internal heat distribution, the ultrasonic signals are used to analyze internal structural defects, and the laser scanning data is used to obtain high-precision three-dimensional surface topography; a data processing module configured to: perform spatiotemporal alignment on the multimodal data, unify the different sensor data into the same spatiotemporal coordinate system through timestamp synchronization and spatial coordinate transformation, and preprocess the multimodal data, wherein the preprocessing includes noise filtering, contrast enhancement, and feature normalization; A deep learning model is configured to: extract and fuse features from the preprocessed data based on the deep learning model, wherein the deep learning model uses a multi-branch structure to process data of different modalities separately, and adjusts the weight of each modality data in real time through a dynamic weight allocation mechanism to perform hybrid fusion at the feature level and the decision level; The defect recognition module is configured to: classify and locate the fused features using a trained deep learning model, identify the types of defects on the shell surface and inside, and determine the location and size of the defects; The process control module is configured to: visualize the recognition results through three-dimensional reconstruction technology and provide real-time feedback to production.

Images