Computer mainboard quality detection method and system based on artificial intelligence

Abstract

The invention relates to the technical field of computer hardware quality detection, and discloses a computer mainboard quality detection method and system based on artificial intelligence, and the method comprises the following steps: collecting multi-modal data of a to-be-detected computer mainboard; constructing a multi-modal time sequence tensor according to the timestamp; inputting the time sequence tensor into a time sequence neural network, and extracting a dynamic feature map reflecting defect evolution; performing low-rank sparse matrix decomposition to obtain an abnormal feature map; calculating the defect saliency of the abnormal feature map; when the defect enhancement trend is judged, feedback detection is executed; the system comprises an image acquisition module, a thermal imaging acquisition module, an electromagnetic induction module, a data preprocessing module, a time sequence analysis module, an anomaly detection module, a trend judgment module and a feedback control module. According to the method, the feature graph decoupling technology based on low-rank sparse matrix decomposition is introduced, and a more accurate abnormal feature graph can be extracted from the background structure and the potential defect area of the mainboard.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer hardware quality detection, in particular to a computer motherboard quality detection method and system based on artificial intelligence. BACKGROUND

[0002] In the production process of modern electronic products, the motherboard as the core component directly affects the overall performance and reliability of the product. Therefore, it is particularly important to conduct efficient and accurate quality detection on the motherboard. With the rapid development of computers, smartphones and other electronic devices, the requirements for motherboard quality are becoming higher and higher. In order to ensure that the motherboard meets the quality standards before leaving the factory, various detection technologies have emerged, among which image detection, thermal imaging detection, electromagnetic detection and other methods are widely used.

[0003] The existing motherboard quality detection mainly relies on some traditional detection methods, such as detecting the defects on the surface of the motherboard through static image analysis, detecting thermal anomalies using thermal imaging technology, and evaluating electromagnetic interference using electromagnetic field distribution data. These methods each have certain advantages, such as image detection can clearly reveal the surface defects of the motherboard, thermal imaging can quickly find overheating areas, and electromagnetic field detection can effectively identify potential electrical problems.

[0004] However, the existing motherboard quality detection methods mostly rely on single modal data, and for image data, only surface defects are detected, while thermal imaging data ignores potential defects that are difficult to display on the motherboard surface, resulting in one-sidedness of the detection results, which is difficult to fully reflect the quality state of the motherboard. The existing technology is usually difficult to track the evolution of defects in the time dimension, and by detecting motherboard defects through static images, it is difficult to determine how the defects intensify or expand over time. Therefore, the present application provides a computer motherboard quality detection method and system based on artificial intelligence to solve the deficiencies in the prior art. SUMMARY

[0005] In view of the deficiencies of the prior art, the purpose of the present application is to provide a computer motherboard quality detection method and system based on artificial intelligence, which solves the problems of insufficient comprehensive utilization of multi-modal data, lack of tracking of dynamic evolution of defects, and confusion of background noise and defect recognition in the existing motherboard quality detection technology.

[0006] To achieve the above purpose, the present application is implemented by the following technical scheme: a computer motherboard quality detection method based on artificial intelligence, comprising the following steps: Collecting multi-modal data of the motherboard to be detected; Based on the collected multimodal data, synchronous processing is performed according to the same timestamp, and a multimodal time series tensor is constructed. The constructed multimodal time series tensor is input into a temporal neural network model to extract dynamic feature maps that reflect the evolution process of motherboard defects. Based on the extracted dynamic feature map, a low-rank sparse matrix decomposition operation is performed to decouple the motherboard background structure from the potential defect region and obtain a sparse anomaly feature map. Based on the obtained abnormal feature map, calculate the defect significance index and determine whether it has a trend of increasing over time within a continuous time window; When a defect enhancement trend is detected, a feedback detection operation is performed, and finally, based on the spatial distribution, intensity change and evolution trend of the defects, the motherboard quality status is intelligently judged and classified, and the corresponding detection results are output.

[0007] Preferably, the multimodal data includes image data, thermal imaging data, and electromagnetic field distribution data: The image data is a sequence of RGB images acquired by an industrial camera under stable lighting conditions. The thermal imaging data is a temperature distribution map acquired by an infrared thermal imager; The electromagnetic field distribution data is an electromagnetic interference map collected by a near-field scanning device.

[0008] Preferably, constructing the multimodal time series tensor includes the following steps: Boundary correction and spatial registration are performed on image data, thermal imaging data, and electromagnetic field data respectively to ensure image region alignment; time interpolation based on linear interpolation is performed on multimodal data to unify the sampling frequency; A fused data stream is constructed based on the timestamps of multimodal image frames as the input tensor. The fused tensor is obtained by concatenating image data, thermal imaging data, and electromagnetic field distribution data along the channel dimension, and is denoted as M. t =[I t |T t |E t ], where I t T represents the input tensor of the image data at time t; t E represents the input tensor of the thermal imaging data at time t. t This represents the input tensor of the electromagnetic field data at time t; it is used as the input to the time-series feature extraction model.

[0009] Preferably, the temporal neural network model is a convolutional long short-term memory network, which includes an input gate, a forget gate, and an output gate. The extraction of dynamic feature maps reflecting the evolution of motherboard defects includes the following steps: Perform a 3D convolution on the input tensor to generate the input state for the current step; The recursive update of the network state is represented as: H t =ConvLSTM(I t H t-1 ;θ); Among them, H t I represents the hidden state of the network at time t; t H represents the input tensor of the image data at time t; t-1 Let θ represent the hidden state at time t-1; θ is the set of network parameters. The output time series feature map is used as a defect latent representation to show the dynamic evolution characteristics of motherboard defects over time.

[0010] Preferably, the low-rank sparse matrix factorization operation includes the following steps: Each frame of the image is modeled as a hybrid model consisting of background structure and sparse anomalous regions; The following optimization objective function is established: Where L represents a low-rank matrix; S represents a sparse matrix; and F represents the original input image frame. Denotes the Frobenius norm; λ1 and λ2 are regularization coefficients, controlling the penalty strength for low-rank matrices and sparse matrices, respectively; ||L|| * ||S||1 represents the kernel norm of a low-rank matrix; ||S||1 represents the L1 norm of a sparse matrix. The augmented Lagrange multiplier method is used to solve the optimization model algebraically to obtain the low-rank matrix and the sparse matrix.

[0011] Preferably, the calculation of the defect significance index includes the following steps: Count the number of pixels in the sparse anomaly matrix that are greater than a set threshold; The formula for calculating significance is: Among them, D t Let S(i,j) represent the defect significance at time t, m×n be the size of the image, S(i,j) represent the sparse outlier matrix element at pixel (i,j) in the image, δ be the set significance threshold, I(·) be the indicator function, m be the number of rows in the image, and n be the number of columns in the image. Whether a defect is evolving can be determined by whether the saliency sequence within a time window shows an increasing trend.

[0012] Preferably, the execution of the feedback detection operation includes the following steps: The coordinates of regions identified as abnormally enhanced are fed back to the image acquisition module for re-sampling at high local resolution. The sampled images are used as input for fine-tuning the training samples of the artificial intelligence model, and the network weights are updated using incremental learning. Finally, the area, location, and trend of the defective region are comprehensively evaluated to output the motherboard's quality grade label.

[0013] Preferably, the calculation methods for the input gate, forget gate, and output gate are as follows: i t =σ(W i *[I t H t-1 ]+b i ); f t =σ(W f *[I t H t-1 ]+b f ); o t =σ(W o *[I t H t-1 ]+b o ); Among them, i t For input gate; f t Forgotten Gate; o t The output gate is represented by *; * indicates a convolution operation; σ is the sigmoid activation function; W is the convolution kernel; b is the bias term; I t It is the input tensor at the current moment; H t-1 It is the hidden state from the previous moment.

[0014] Preferably, the formula for algebraically solving the optimization model using the augmented Lagrange multiplier method is as follows: Among them, L t Let S represent a low-rank matrix. t Let Y represent a sparse matrix, where Y is a Lagrange multiplier and F is a sparse matrix. t This represents the raw data of the image frame, where λ is the regularization coefficient, μ is the penalty coefficient, and ||L t || * L represents a low-rank matrix t The nuclear norm, ||S t ||1 represents a sparse matrix S t The L1 norm, represents the Frobenius norm, and <·,·> represents the inner product operation.

[0015] It also provides an AI-based computer motherboard quality testing system, including: The image acquisition module is used to acquire image data from the motherboard; Thermal imaging acquisition module, used to acquire thermal distribution map of the motherboard; Electromagnetic induction module, used to obtain the electromagnetic distribution map of the motherboard; The data preprocessing module is used to synchronize and fuse multimodal data to construct time series tensors; The timing analysis module includes a convolutional gated neural network for extracting dynamic features of motherboard defects; The anomaly detection module is used to perform low-rank sparse decomposition based on feature maps to extract sparse anomaly images; The trend judgment module is used to calculate the significance of defects and evaluate their evolution trend; The feedback control module is used to perform local reviews and model updates based on evolutionary trends, and output the motherboard quality inspection results.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention introduces a feature map decoupling technique based on low-rank sparse matrix factorization, which can extract more accurate anomaly feature maps from the background structure and potential defect areas of the motherboard. This innovation allows the system to identify minute defects more efficiently, avoids the problem of background noise and defect signal mixing in the prior art, significantly improves the sensitivity and accuracy of anomaly detection, and solves the shortcomings of existing methods in effectively handling complex backgrounds and defect overlap.

[0017] 2. This invention employs a synchronous acquisition and fusion method for multimodal data, combining image, thermal imaging, and electromagnetic field data into a time-series tensor, forming a multidimensional data input with complete spatiotemporal information. Through this innovation, the system can fully utilize the data characteristics of each modality during processing, significantly improving the accuracy and comprehensiveness of defect detection. Compared to existing methods that rely on a single modality for detection, this invention effectively overcomes the limitations of a single data source and enhances the ability to capture complex defects.

[0018] 3. This invention employs a convolutional long short-term memory network for temporal feature extraction, enabling dynamic tracking of the evolution of defects over time. This temporal-based deep learning method can not only identify current defects but also predict their future trends. Compared to traditional static image analysis methods, this invention effectively addresses the limitation of failing to capture the defect evolution process, achieving defect prediction and early prevention, and improving the foresight of quality inspection.

[0019] 4. This invention achieves system self-optimization and learning through a feedback detection mechanism, enabling it to adjust detection strategies according to defect evolution trends, thereby improving the adaptability and intelligence of defect detection. Unlike existing detection methods that rely solely on preset rules, the adaptive feedback mechanism of this invention solves the problem of lacking real-time learning and adjustment capabilities in existing technologies, achieving higher real-time performance and accuracy, and ensuring stable operation under different working environments. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method steps in this application; Figure 2 This is the system architecture diagram of this application. Detailed Implementation

[0021] The following is in conjunction with the appendix Figure 1 -Appendix Figure 2 This application will be described in further detail below.

[0022] Please see the appendix Figure 1 This invention provides an artificial intelligence-based method for testing the quality of computer motherboards, comprising the following steps: S1. Collect multimodal data of the motherboard of the computer under test; S2. Based on the collected multimodal data, perform synchronous processing according to the same timestamp, and construct a multimodal time series tensor; S3. Input the constructed multimodal time series tensor into the temporal neural network model to extract dynamic feature maps that reflect the evolution process of motherboard defects; S4. Based on the extracted dynamic feature map, perform low-rank sparse matrix decomposition to decouple the motherboard background structure from the potential defect area and obtain sparse anomaly feature map. S5. Based on the obtained abnormal feature map, calculate the defect significance index and determine whether it has a trend of increasing over time within a continuous time window. S6. When it is determined that there is a defect enhancement trend, a feedback detection operation is executed, and finally, based on the spatial distribution, intensity change and evolution trend of the defects, the motherboard quality status is intelligently judged and classified, and the corresponding detection results are output.

[0023] For step S1, in this embodiment, multimodal data acquisition includes the simultaneous acquisition of image data, thermal imaging data, and electromagnetic field distribution data. This process can be performed in real time on the motherboard manufacturing line or completed at an offline quality inspection station.

[0024] In some embodiments, the image data acquisition module uses an industrial camera with a global shutter function and a recommended resolution of at least 1920×1080, which can maintain exposure stability under varying ambient light conditions.

[0025] In general, to ensure that image edges and texture features are not distorted, a flexible ring light source can be used to provide stable lighting conditions, and the camera can be set to a fixed viewing angle to avoid image mismatch caused by changes in angle.

[0026] Alternatively, thermal imaging data can be acquired using an infrared thermal imager. Specifically, the thermal imager supports the 8-14μm infrared band, a frame rate of at least 9Hz, and a temperature sensitivity better than 0.05℃. Field of view (FOV) overlap calibration between the thermal imager and the industrial camera can be performed, and a preliminary spatial mapping relationship can be established using a reference calibration board.

[0027] In one possible implementation, after the motherboard is subjected to a test current, a thermal imager captures the evolution of its thermal distribution in real time. This not only helps to expose defects such as cold solder joints and microcracks, but also provides a temperature change path for subsequent timing modeling.

[0028] In this embodiment, electromagnetic field distribution data is acquired by a near-field scanning probe. This device is positioned 3-5 mm above the motherboard, and the scanning range covers the entire motherboard area.

[0029] Generally, the sampling accuracy of electromagnetic detection equipment is set to one sampling point per millimeter, with a frequency response range of 10kHz to 1GHz. To shield against external interference, the probe installation area should be covered with a flexible shield or have an electromagnetic absorption layer installed.

[0030] In some embodiments, the electromagnetic interference data is presented as a two-dimensional heatmap, where the pixel values ​​represent the instantaneous electric field intensity at each sampling point. This image is then fed into the neural network as one of the tensor channels.

[0031] As an optional technical approach, to ensure the consistency of the three modalities, each frame of acquired data is bound to a unified timestamp, with an accuracy controlled within 10ms. The timestamp records are annotated with millisecond-level system time for subsequent interpolation and synchronization operations.

[0032] Specifically, image acquisition frames, thermal imaging frames, and electromagnetic sampling frames are written into a circular queue at the software layer through a buffering mechanism, and time alignment is completed according to the system master clock.

[0033] In another implementation, a hardware synchronization triggering module can be configured, such as using a PLC system or TTL level signals to simultaneously activate multiple sensors, thereby achieving higher time consistency in data acquisition.

[0034] To facilitate subsequent processing, in this embodiment, all three types of modal data are converted into a uniform two-dimensional matrix image format, for example, 256×256 pixels. During the conversion process, the images are cropped, scaled, and have their boundaries repaired to ensure that different data modalities have consistent spatial coverage.

[0035] In step S2, this embodiment completes the acquisition of multimodal raw data, including motherboard images, thermal imaging, and electromagnetic field distribution. Due to the inconsistencies in the temporal and spatial distribution of these data, directly using them for deep model analysis would cause severe information mismatch. Therefore, a crucial preprocessing step must be performed before proceeding to deep feature modeling.

[0036] This approach addresses issues such as inconsistent sampling frequencies, alignment errors, and resolutions among different sensors, while also laying a structural foundation for the subsequent construction of channel fusion tensors. Its core function lies in spatially and temporally unifying the structural image, thermal image, and electromagnetic response image, providing them with unified input conditions for modeling.

[0037] In this embodiment, boundary correction and spatial registration are first performed on the multimodal data. Specifically, affine transformation is used for initial alignment, followed by the introduction of a homography matrix based on SIFT keypoint matching for fine registration, so that image data, thermal imaging data, and electromagnetic image data achieve region-level overlap in the two-dimensional pixel space.

[0038] In general, to avoid mismatches caused by high-contrast components on the motherboard, edge enhancement processing can be performed before matching, such as using a Laplacian filter to extract the contour map before extracting feature points.

[0039] As an alternative, the data needs to be cropped after registration, retaining the area with the largest intersection among the three types of images as the final processing range, to ensure that each pixel has a corresponding physical meaning in all three modalities.

[0040] In some embodiments, to address the issue of inconsistent frame rates in multimodal data, the system employs linear interpolation to perform time padding on the frame sequence. Specifically, let the image frame time series be... If there are missing thermal imaging data frames, linear interpolation is performed based on the data from the preceding and following frames to generate a thermal image frame T. t This interpolation can be expressed as: in, t represents the time point to be interpolated, and t1 and t2 represent the adjacent sampled times before and after it. These are heatmap frames corresponding to the time periods.

[0041] In one possible implementation, because electromagnetic image data has a finer sampling granularity, the system adopts a window averaging downsampling strategy to make it consistent with the image data in terms of frame rate, so that it can be synchronously bound to timestamps.

[0042] After spatial registration and temporal interpolation, the three types of images are stitched together along the channel dimension to form a unified fusion tensor, which represents the multimodal state of the motherboard at a certain moment. The fusion tensor can be formally represented as: M t =[I t |T t |E t ]; Among them, I t T represents the input tensor of the image data at time t; t E represents the input tensor of the thermal imaging data at time t. t This represents the input tensor of the electromagnetic field data at time t; it is used as the input to the time-series feature extraction model, where t is the time index.

[0043] It is worth noting that this tensor is usually stored as a three-dimensional array with dimensions in the order (C, H, W), where C = 3 is the number of modal channels, and H and W are the height and width of the image, respectively.

[0044] In some embodiments, the fused tensor sequence will serve as the base sequence for subsequent neural network inputs, possessing complete temporal continuity and modal synchronization.

[0045] As an extension, to further enhance the robustness of spatial alignment, the system can also integrate a deep learning-based feature reconstruction network, such as U-Net or a transform-aware alignment module, into the registration module to improve the adaptability in complex backgrounds.

[0046] In addition, to ensure storage efficiency and real-time performance, the fused tensors will be uniformly converted into floating-point tensors with a bit width of 32 bits after generation; at the same time, the frame buffer is managed in the form of a sliding window, and the length of each processing session can be set to 5 to 15 frames, which is dynamically adjusted according to the motherboard's operating cycle.

[0047] For step S3, in this embodiment, after completing the temporal synchronization and tensor construction of multimodal data, the system needs to extract its deep feature information during the time evolution process to capture the dynamic pattern of potential motherboard defects changing over time. By inputting the constructed multimodal time series tensor into the temporal neural network model, dynamic feature maps reflecting the evolution process of motherboard defects are extracted.

[0048] Unlike traditional static image classification, this invention introduces a Convolutional Long Short-Term Memory (ConvLSTM) network with temporal modeling capabilities as the backbone architecture in this step, combining spatial convolution and temporal gating mechanisms to model and analyze the dynamic motherboard state.

[0049] In this embodiment, the input is a pre-constructed three-channel fused tensor sequence, where each fused tensor contains the channel stitching result of image data, thermal imaging data, and electromagnetic map data at the same time point. This sequence is input into a ConvLSTM network for recursive modeling, and the network structure simultaneously possesses temporal memory and spatial local awareness capabilities.

[0050] Generally, a ConvLSTM structure includes three basic gating units: an input gate, a forget gate, and an output gate, which are used to control the propagation and updating of the current input, historical state, and final hidden state, respectively. Its basic recursive structure can be formally described as follows: H t =ConvLSTM(M t H t-1 ;θ); Among them, H t H represents the hidden state at the current time. t-1 Let M represent the hidden state at the previous time step, θ be the set of network parameters including all convolutional kernels and bias terms, and M be the hidden state at the previous time step. t This is the fusion tensor of the current input.

[0051] In some embodiments, to improve the response sensitivity to the evolution path of motherboard defects, the ConvLSTM network adopts a multi-layer stacked structure, with the output of each layer serving as the input of the next layer, and a skip connection or batch normalization strategy is used between layers to avoid gradient vanishing.

[0052] Specifically, the gating mechanisms in ConvLSTM are expressed as follows: i t =σ(W i *[I t H t-1 ]+b i ); f t =σ(W f *[I t H t-1 ]+b f ); o t =σ(W o *[I t H t-1 ]+b o ); Among them, i t For input gate; f t Forgotten Gate; o t For output gate; σ(·) represents the Sigmoid activation function; [I t H t-1 ] indicates the concatenation of the input tensor and the historical hidden state channels; W iW f W o These are the convolution kernel parameters corresponding to the gating; b i ,b f ,b o For the corresponding bias term; * indicates a two-dimensional or three-dimensional convolution operation, with the kernel size typically being 3×3 or 5×5; I t It is the input tensor at the current moment.

[0053] In one possible implementation, the network output hidden state is further mapped to a sequence of feature maps representing the motherboard's defect feature response at each time point, where each feature map is consistent with the input tensor in spatial dimension.

[0054] As an extension, the system can introduce attention mechanisms or gated dilated convolution modules to improve the model's ability to capture the coupling features between minute defects or multimodal heterogeneous signals.

[0055] In other embodiments, to optimize the model's convergence speed and generalization ability, a time regularization term based on KL divergence is introduced during training to penalize non-stationary sequence responses, which helps to enhance the model's stability under changes in motherboard samples.

[0056] In this embodiment, the final feature map sequence is an intermediate representation reflecting the evolution trajectory of potential defects on the motherboard at different time points, which can be used for subsequent anomaly detection and trend analysis. This representation has continuity and distribution, and has stronger temporal modeling capabilities than traditional frame-level judgment methods.

[0057] In step S4, in this embodiment, obtaining the time-series feature map based on multimodal input is still insufficient to directly support the identification and subsequent discrimination of motherboard defects. Since the feature map typically mixes normal background responses and abnormal signal regions, and may contain random noise or modal inconsistencies, further structural decoupling of the feature map is required through matrix factorization. By performing low-rank sparse matrix factorization based on the extracted dynamic feature map, the motherboard background structure and potential defect regions are decoupled to obtain a sparse abnormal feature map, which plays a crucial separation role in the entire detection method.

[0058] In this embodiment, the system first models the dynamic feature map at each time step as consisting of two main components: a low-rank background component and a sparse anomaly component, satisfying the following structural expression: F t =L t +S t +E t ; Among them, F t For the input feature map, L tThis indicates the normal structure and background response of the motherboard, S t E indicates a potential defect area or local anomaly. t This is the noise residual term.

[0059] In some embodiments, ignoring the weak perturbation term of Gaussian noise, the system constructs the following optimization objective function for the decomposition task: Where L represents a low-rank matrix; S represents a sparse matrix; and F represents the original input image frame. Denotes the Frobenius norm; λ1 and λ2 are regularization coefficients, controlling the penalty strength for low-rank matrices and sparse matrices, respectively; ||L|| * ||S||1 represents the kernel norm of a low-rank matrix, used to constrain the rank of the background matrix; ||S||1 represents the L1 norm of a sparse matrix, used to constrain the sparsity of anomalous regions.

[0060] Generally, to improve the stability and generalization ability of the solution, the regularization coefficient λ1 can be set to... Where m and n are the matrix dimensions; λ2 is usually between 0.01 and 0.1, and is fine-tuned according to the complexity of the motherboard.

[0061] Alternatively, the augmented Lagrange multiplier method is used for iterative solution of this optimization problem. This method significantly reduces sensitivity to initialization while maintaining the stability of the main optimization objective, thus improving convergence efficiency.

[0062] Specifically, the augmented Lagrange objective function is constructed as follows: Among them, L t Let S represent a low-rank matrix. t Let Y represent a sparse matrix, where Y is a Lagrange multiplier and F is a sparse matrix. t This represents the raw data of the image frame, where λ is the regularization coefficient, μ is the penalty coefficient, and ||L t || * L represents a low-rank matrix t The nuclear norm, ||S t ||1 represents a sparse matrix S t The L1 norm, represents the Frobenius norm, and <·,·> represents the inner product operation.

[0063] In one possible implementation, to improve running efficiency on the GPU, the low-rank term ||L t || * The solution employs an approximate singular value thresholding (SVT) operation, with sparse terms ||St ||1 uses a soft threshold function to achieve element-by-element stripping.

[0064] In some embodiments, the system can use the decomposition result of the previous moment as the initial value of the next moment to form a time-series recursive solution process, so as to reduce the overall solution time and maintain the continuity of the motherboard state changes.

[0065] In this embodiment, the matrix obtained after decomposition is output as the sparse anomaly feature map at the current moment, where non-zero positions indicate potential thermal, electrical, or structural anomaly regions on the motherboard, and their spatial locations are related to the original... Figure 1 To.

[0066] To further enhance the saliency of the defect region boundaries, some embodiments also introduce post-processing operations on the sparse graph, such as two-dimensional high-pass filtering, region aggregation, or morphological dilation, to strengthen region connectivity and stability.

[0067] For step S5, in this embodiment, after extracting the sparse anomaly feature map, the system still needs to further quantify the anomaly region to obtain explicit discrimination results of motherboard defects. A comprehensive defect measurement process based on multimodal dynamic features is then implemented. By performing defect measurement calculations based on the sparse anomaly feature map and outputting motherboard defect detection results, this process realizes the transformation from low-level features to high-level defect indicators, completing the final detection loop.

[0068] In this embodiment, the input is a sequence of sparse anomaly feature maps, and the system constructs a defect scoring function D for each time t. t This is used to measure the potential anomaly of the motherboard at that moment.

[0069] Generally, the defect scoring function can be defined as: Among them, S t (i,j) represents the value at pixel (i,j) in the sparse anomaly feature map; W and H are the width and height of the input image; ∑ (i,j) S t (i,j) represents the summation of the defect intensity values ​​of all pixels in the image; W×H is the total number of pixels in the image.

[0070] Specifically, to improve the spatial sensitivity of the scoring, this invention further clusters the abnormal response regions according to connected components, extracts the defect regions with the largest area and the highest intensity, and introduces a weighting factor α. k This represents the weight distribution of each cluster region. At this point, the scoring function expands to: Where K represents the number of connected regions; R kRepresents the set of pixels in the k-th connected region; |R k | represents the number of pixels in the k-th region; a k S is the weight coefficient of the k-th region, reflecting the importance of that region; t (i,j) represents the value of pixel (i,j) in the sparse anomaly feature map, which represents the defect intensity at that location; This represents the sum of defect intensities for all pixels within the k-th region.

[0071] As an alternative, to enhance the temporal consistency of defect measurements in dynamic sequences, the system employs a temporal smoothing mechanism in some embodiments to perform weighted fusion of historical frames: in, D is the smoothed defect measurement. t This represents the defect measurement at the current moment; This represents the smoothness defect measure of the previous time step; β is the smoothing factor, which is usually set between 0.6 and 0.8 to achieve a balance between response sensitivity and robustness.

[0072] In one possible implementation, the system also maps continuous scores to binary defect discrimination results based on a discrimination threshold θ: Among them, Y t =1 indicates that a defect was detected in the current frame, Y t =0 indicates a normal state. The threshold θ can be determined in the training set through cross-validation, and is generally set between 0.02 and 0.05, depending on the complexity of the motherboard and the requirements for defect sensitivity.

[0073] In some embodiments, in order to meet the requirements of high-precision detection, the system can also calculate the bounding box information of the defect area, such as the minimum bounding rectangle parameters (center coordinates, width and height, rotation angle), for use by the subsequent industrial discrimination module.

[0074] Generally, defect detection results include not only the time score D t And binary result Y t It also outputs an overlay of abnormal images and a region contour map, which facilitates human-computer collaborative review.

[0075] In this embodiment, through the above-mentioned multi-layer measurement strategy, the system effectively achieves a seamless connection from sparse abnormal features to quantitative defect results, ensuring that the detection results are interpretable and stable, and providing a complete input basis for subsequent intelligent screening or alarm mechanisms.

[0076] For step S6, in this embodiment, after completing the sparse representation and measurement of motherboard defects, in order to further classify and trace the different types of defects, the system needs to embed the aforementioned defect features into a unified embedding space so that subsequent classification and recognition based on feature similarity can be performed. By mapping the sparse anomaly feature map to a unified motherboard defect representation space, the core objective is to transform the original sparse image information into a structured vector representation through a feature embedding mechanism, thereby supporting defect type discrimination and data archiving.

[0077] In this embodiment, the input is the sparse anomaly feature map obtained after low-rank sparse decomposition. The system first performs spatial normalization on each sparse anomaly feature map obtained after low-rank sparse decomposition, standardizing it into a sparse map S of uniform size. t ′ The vector z is then fed into a feature mapping network to extract the embedded representation vector. t .

[0078] Typically, this feature mapping network is constructed as a lightweight encoder structure containing multiple convolutional and fully connected layers to compress spatial distribution information and extract high-dimensional semantic features. The overall feature mapping function can be expressed as: z t =f embed (S t ′ ;θ); in, f represents a defect feature embedding vector of dimension d; embed (·) denotes the embedded network function; θ is the set of embedded network parameters.

[0079] Specifically, in some embodiments, the embedded network adopts the following structure: Two 3×3 convolutional layers with 16 and 32 channels respectively, followed by ReLU activation; A global average pooling layer is used to compress spatial dimensions; Finally, a fully connected layer is added to output a d-dimensional embedding vector, typically d = 64 or 128.

[0080] As an alternative, to enhance the separability and discriminability of feature distribution, the system introduces the TripletLoss loss function during the training phase to optimize the embedding space structure, making the distance between defects of the same type closer and the vector distance between defects of different types farther.

[0081] In one possible implementation, the trained embedding model can be used in the online inference phase, based on the input S. t ′ Output embedding vector z tThe system then compares the results with a pre-built library of defect type prototypes. The system uses cosine similarity for discrimination, calculated using the following formula: Among them, z t For the embedding vector; z k is the prototype embedding vector of the k-th category in the defect library; sim(·) is the similarity score with a value range of [-1,1].

[0082] Specifically, the system uses the maximum similarity matching result as the classification output of the defect in the current frame, and returns the similarity value as the classification confidence.

[0083] In some embodiments, to improve the embedding consistency of the system in multimodal fusion scenarios, joint modal contrast loss can be introduced to force the sparse graph embedding vectors extracted under different modalities to remain aligned in the semantic space.

[0084] The AI-based computer motherboard quality inspection system described below can be referred to in conjunction with the AI-based computer motherboard quality inspection method described above.

[0085] Please see the appendix Figure 2 The present invention also provides an artificial intelligence-based computer motherboard quality inspection system, comprising: The image acquisition module is used to acquire image data from the motherboard; Thermal imaging acquisition module, used to acquire thermal distribution map of the motherboard; Electromagnetic induction module, used to obtain the electromagnetic distribution map of the motherboard; The data preprocessing module is used to synchronize and fuse multimodal data to construct time series tensors; The timing analysis module includes a convolutional gated neural network for extracting dynamic features of motherboard defects; The anomaly detection module is used to perform low-rank sparse decomposition based on feature maps to extract sparse anomaly images; The trend judgment module is used to calculate the significance of defects and evaluate their evolution trend; The feedback control module is used to perform local reviews and model updates based on evolutionary trends, and output the motherboard quality inspection results.

[0086] The system in this embodiment can be used to execute the above method embodiments, and its principle and technical effect are similar, so they will not be described again here.

[0087] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Identical components are represented by the same reference numerals. Therefore, all equivalent changes made to the structure, shape, and principle of this application should be covered within the scope of protection of this application.

Claims

1. A computer motherboard quality detection method based on artificial intelligence, characterized in that, The method comprises the following steps: Collecting multi-modal data of a computer motherboard to be detected; Synchronizing and processing the collected multi-modal data according to the same time stamp, and constructing a multi-modal time series tensor; Inputting the constructed multi-modal time series tensor into a time series neural network model to extract a dynamic feature map reflecting the evolution process of the motherboard defect; Based on the extracted dynamic feature map, performing a low-rank sparse matrix decomposition operation to decouple the motherboard background structure and the potential defect area, and obtaining a sparse abnormal feature map; According to the obtained abnormal feature map, calculating a defect saliency index, and judging whether there is a trend of increasing with time in the continuous time window; When it is judged that there is a defect enhancement trend, a feedback detection operation is performed, and finally the quality state of the motherboard is intelligently judged and classified according to the spatial distribution, intensity change and evolution trend of the defect, and the corresponding detection result is output. 2.The artificial intelligence-based computer mainboard quality detection method of claim 1, wherein, The multi-modal data includes image data, thermal imaging data and electromagnetic field distribution data: The image data is an RGB image sequence obtained by an industrial camera under stable lighting conditions, The thermal imaging data is a temperature distribution map collected by an infrared thermal imager; The electromagnetic field distribution data is an electromagnetic interference map collected by a near-field scanning device. 3.The artificial intelligence-based computer mainboard quality detection method of claim 1, wherein, The construction of the multi-modal time series tensor comprises the following steps: Respectively correcting the boundaries and performing spatial registration on the image data, thermal imaging data and electromagnetic field data to ensure that the image regions are aligned; time filling is performed on the multi-modal data based on linear interpolation to unify the sampling frequency; According to the timestamps of the multi-modal image frames, a fusion data stream is constructed as an input tensor, the fusion tensor is obtained by splicing image data, thermal imaging data and electromagnetic field distribution data in a channel dimension, and is denoted as M t =[I t |T t |E t ], wherein I t denotes an input tensor of image data at time t; T t denotes an input tensor of thermal imaging data at time t; E t denotes an input tensor of electromagnetic field data at time t; and the input is used for a time sequence feature extraction model. 4.The artificial intelligence-based computer mainboard quality detection method of claim 1, wherein, The time series neural network model is a convolutional long short-term memory network, which comprises an input gate, a forgetting gate and an output gate, and the extraction of the dynamic feature map reflecting the evolution process of the motherboard defect comprises the following steps: Three-dimensional convolution is performed on the input tensor to generate the input state of the current step; The recursive update of the network state is represented as: H t = ConvLSTM(I t ,H t-1 ; θ); where H t denotes the network hidden state at time t; I t denotes the input tensor of image data at time t; H t-1 denotes the hidden state at time t-1; θ is the set of network parameters; The time series feature map is output as a defect latent representation, which is used to represent the dynamic evolution characteristics of the motherboard defect over time. 5.The artificial intelligence-based computer mainboard quality detection method of claim 1, wherein, The execution of the low-rank sparse matrix decomposition operation comprises the following steps: Modeling each frame of image as a hybrid model composed of background structure and sparse abnormal area; An optimization objective function is established as follows: wherein L represents a low-rank matrix; S represents a sparse matrix; F represents an original input image frame; denotes a Frobenius norm; λ1 and λ2 are regularization coefficients, respectively controlling the penalty strength of the low-rank matrix and the sparse matrix; ||L||F * denotes a nuclear norm of the low-rank matrix; ||S||1 denotes an L1 norm of the sparse matrix; The optimization model is solved by using the augmented Lagrange multiplier method to obtain a low-rank matrix and a sparse matrix. 6.The artificial intelligence-based computer mainboard quality detection method of claim 1, wherein, The calculation of the defect saliency index comprises the following steps: The number of pixels greater than a set threshold in the sparse abnormal matrix is counted; The calculation formula of the saliency is: where D t denotes the defect saliency at time t, m x n is the size of the image, S(i,j) denotes the sparse anomaly matrix element at pixel point (i,j) in the image, δ is a set saliency threshold, I(·) is an indicator function, m is the number of rows of the image, and n is the number of columns of the image; According to whether the saliency sequence in the time window shows a growth trend, it is judged whether the defect is evolving. 7.The artificial intelligence-based computer mainboard quality detection method of claim 1, wherein, The execution of the feedback detection operation comprises the following steps: The coordinates of the region with abnormal enhancement are fed back to the image acquisition module for re-sampling at high resolution; The image obtained by sampling is input as a fine-tuning training sample of the artificial intelligence model, and the network weight is updated in an incremental learning manner; Finally, the area, position and trend of the defect area are comprehensively evaluated, and the quality grade label of the motherboard is output. 8.The artificial intelligence-based computer mainboard quality detection method of claim 4, wherein, The calculation methods of the input gate, the forgetting gate and the output gate are respectively: i t = σ(W i * [I t , H t-1 ] + b i ); f t = σ(W f *[I t ,H t-1 ]+b f ); o t = σ(W o * [I t , H t-1 ] + b o ); where i t is the input gate; f t is the forget gate; o t is the output gate; * denotes convolution operation; σ is the Sigmoid activation function; W is the convolution kernel; b is the bias term; I t is the input tensor at the current time; H t-1 is the hidden state at the previous time. 9.The artificial intelligence-based computer mainboard quality detection method of claim 5, wherein, The formula for algebraically solving the optimization model by using the augmented Lagrange multiplier method is: where L t denotes a low-rank matrix, S t denotes a sparse matrix, Y is a Lagrange multiplier, F t denotes the original data of the image frame, λ is a regularization coefficient, μ is a penalty coefficient, ||L t || * denotes the nuclear norm of the low-rank matrix L t , ||S t ||1denotes the L1 norm of the sparse matrix S t , denotes the Frobenius norm, and <·,·> denotes an inner product operation.

10. The computer motherboard quality detection system based on artificial intelligence, applied to the computer motherboard quality detection method based on artificial intelligence of any one of claims 1-9, characterized in that, It comprises: An image acquisition module for acquiring image data of a motherboard; a thermal imaging acquisition module for acquiring a thermal distribution map of the motherboard; an electromagnetic induction module for acquiring an electromagnetic distribution map of the motherboard; a data preprocessing module for synchronizing and fusing the multi-modal data and constructing a time series tensor; a time series analysis module including a convolutional gated neural network for extracting dynamic features of the motherboard defects; an anomaly detection module for performing low-rank sparse decomposition based on the feature map and extracting a sparse anomaly image; a trend judgment module for calculating defect saliency and evaluating its evolution trend; a feedback control module for performing local review and model updating based on the evolution trend and outputting a motherboard quality detection result.

Images