Method for detecting surface micro-defects of car-gauge semiconductor devices based on machine haptics

Abstract

The application discloses a kind of based on machine haptics's car gauge level semiconductor device surface microdefect detection method, belong to semiconductor device quality detection technical field.The application is by establishing the quantitative coupling relationship of microdefect geometric parameter and haptics signal characteristic parameter, design high-density flexible piezoelectric haptics sensor array, and using signal stable enhancement under high-speed scanning and multi-force level signal fusion technology, construct the car gauge level microdefect intelligent discrimination model based on improved C3D-Attention, realize to submicron level microcrack, nanometer scratch, glue surface wrinkle and other surface microdefects High-precision online detection.The application can overcome the bottleneck of traditional machine vision in high light, heterogeneous surface is susceptible to light interference, high miss rate, cannot obtain three-dimensional mechanical information, etc., false detection rate ≤100PPM, satisfy AEC-Q100 car gauge reliability requirement, can be adapted to automobile electronic production line high-speed scanning and batch detection, provide technical support for car gauge level semiconductor advanced manufacturing quality control and device reliability improvement.

Description

Technical Field

[0001] This invention relates to the field of semiconductor device quality inspection technology, and specifically to an online detection method for micro-defects on the surface of automotive-grade semiconductor devices based on machine tactile sensing. Background Technology

[0002] Automotive-grade semiconductor devices (such as MCUs, power IGBTs, and sensor chips) are the core of automotive electronic control systems. Micron-level defects on their surfaces, such as microcracks, chip scratches, and adhesive wrinkles, can directly lead to the failure of electrical performance of the devices, reduced reliability, and potential safety hazards in automobiles.

[0003] Current surface defect detection primarily relies on machine vision technology, but it faces the following technical bottlenecks: automotive-grade components are mostly made of highly reflective silicon-based or metal-coated surfaces, making optical imaging susceptible to illumination interference. Submicron-level defects exhibit extremely low contrast, resulting in a high rate of missed detections. Visual inspection depends on two-dimensional image analysis, which cannot acquire three-dimensional information such as defect depth and mechanical properties, making it difficult to distinguish between real defects and surface texture interference. On-line inspection requires a balance between speed and accuracy, but the sampling frequency of vision systems is limited, failing to meet the demands of rapid full-surface scanning for high-density components. Existing machine tactile inspection technologies are mostly applied to macroscopic components. Key technologies for nanoscale tactile perception, multi-force signal fusion, and automotive-grade defect discrimination for automotive-grade semiconductor devices are not yet mature and cannot meet the stringent requirements of on-line inspection. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide an online detection method for micro-defects on the surface of automotive-grade semiconductor devices based on machine tactile sensing. This method aims to overcome the limitations of visual inspection, achieve high-precision and interference-free identification of micro-defects at the micrometer level, construct a machine tactile recognition model adapted to automotive-grade standards, reduce false detection rates, meet AEC-Q100 reliability requirements, realize high-speed online scanning on production lines, balance detection accuracy and production cycle time, and adapt to batch inspection of high-density semiconductor devices.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for detecting micro-defects on the surface of automotive-grade semiconductor devices based on machine tactile sensing includes: Multi-force level contact tactile signal acquisition: Control the flexible tactile array sensor to scan the surface of the device under test along a preset path, apply multi-level normal forces to each detection point on the device surface in sequence, and simultaneously acquire the three-dimensional signals of normal force, tangential force, and deformation displacement of each sensing unit in the flexible tactile array sensor to generate a position-multi-force level mechanical feature dataset; Micro-defect mechanical fingerprint extraction and feature fusion: Time-domain and frequency-domain analysis is performed on the multi-level force signals of each detection point in the location-multi-force level mechanical feature dataset to extract multi-dimensional core features and construct the multi-force level mechanical fingerprint vector of the detection point; and the mechanical fingerprint vectors of adjacent sensing units are spatially correlated and fused to obtain a fused mechanical feature vector that highlights the abnormal mechanical features of the defect area. Intelligent identification of automotive-grade micro-defects: The fused mechanical feature vector is input into a pre-trained improved C3D-Attention neural network model, which outputs the defect type and size of the corresponding detection point; The improved C3D-Attention neural network model includes a three-stage cascaded feature extractor and classifier. Each feature extractor includes two neural network branches with different convolutional kernel sizes. One branch applies a spatial attention mechanism, and the other branch applies a channel attention mechanism. The output features of the two branches are fused and used as the output of the feature extractor at that stage. The output of the previous stage feature extractor is used as the input of the next stage feature extractor, and the output feature maps of the three stage feature extractors are fused and then input into the classifier.

[0006] Furthermore, in the multi-force level contact tactile signal acquisition, three levels of normal force are applied sequentially, each force is held for a preset time, and continuous mechanical time-series data is acquired at a preset sampling frequency. After removing environmental vibration and electromagnetic interference noise, the position-multi-force level mechanical feature dataset is generated.

[0007] Furthermore, the step of extracting multi-dimensional core features and constructing a multi-force level mechanical fingerprint vector for the detection point specifically involves: extracting the peak force, force change rate, deformation gradient, and frequency domain energy spectrum features of the multi-level force signal for each detection point, and constructing a multi-force level mechanical fingerprint vector.

[0008] Furthermore, in the two neural network branches of the feature extractor, one branch uses a convolutional kernel with a receptive field of 1 in the time dimension, and the other branch uses a convolutional kernel with a receptive field greater than 1 in the time dimension; and each branch structure contains multiple 3D convolutional blocks, and all max pooling layers are not reduced in the spatial length dimension.

[0009] Furthermore, the spatial attention mechanism employs three-dimensional convolution and three-dimensional pooling operations to derive a three-dimensional spatial attention feature map; the channel attention mechanism employs three-dimensional pooling operations combined with a multilayer perceptron to derive a one-dimensional channel attention feature map; the feature maps enhanced by the two branches are then fused element-wise.

[0010] Furthermore, the classifier includes a three-dimensional global average pooling layer, a discard layer, a fully connected layer, and an output layer.

[0011] This invention also provides a machine-tactile-based system for detecting micro-defects on the surface of automotive-grade semiconductor devices, comprising: The signal acquisition module is used to control the flexible tactile array sensor to scan the surface of the device under test along a preset path, apply multi-level normal forces to each detection point on the surface of the device in sequence, and simultaneously acquire the three-dimensional signals of normal force, tangential force, and deformation displacement of each sensing unit in the flexible tactile array sensor to generate a position-multi-force level mechanical feature dataset. The feature fusion module is used to perform time-domain and frequency-domain analysis on the multi-level force signals of each detection point in the location-multi-level mechanical feature dataset, extract multi-dimensional core features, construct the multi-level mechanical fingerprint vector of the detection point, and spatially correlate and fuse the mechanical fingerprint vectors of adjacent sensing units to obtain a fused mechanical feature vector that highlights the abnormal mechanical features of the defect area. The intelligent discrimination module is used to input the fused mechanical feature vector into a pre-trained improved C3D-Attention neural network model and output the defect type and size of the corresponding detection point. The improved C3D-Attention neural network model includes a three-level cascaded feature extractor and classifier. Each level of feature extractor includes two neural network branches with different convolutional kernel sizes. One branch applies a spatial attention mechanism, and the other branch applies a channel attention mechanism. The output features of the two branches are fused and used as the output of the feature extractor at that level. The output of the previous level of feature extractor is used as the input of the next level of feature extractor, and the output feature maps of the three level feature extractors are fused and then input into the classifier.

[0012] The present invention also provides a flexible tactile array sensor, comprising a plurality of piezoelectric sensing units. Each piezoelectric sensing unit includes a substrate, a lower electrode, an interelectrode film, an upper electrode, and a contact head arranged sequentially. Each piezoelectric sensing unit is provided with a lower electrode and four upper electrodes arranged in an array. The contact head is truncated pyramidal in shape and covers the four upper electrodes, used to decompose and transmit external three-dimensional forces to the interelectrode film regions corresponding to the four upper electrodes.

[0013] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the machine tactile-based method for detecting micro-defects on the surface of automotive-grade semiconductor devices as described in any of the preceding claims.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method for detecting micro-defects on the surface of automotive-grade semiconductor devices based on machine tactile sensing as described in any of the preceding claims.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the machine tactile-based method for detecting micro-defects on the surface of automotive-grade semiconductor devices as described in any of the preceding claims.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention employs multi-level contact tactile signal acquisition. By applying multi-level normal forces to the detection point and simultaneously acquiring three-dimensional signals of normal force, tangential force, and deformation displacement, based on the quantitative coupling relationship between micro-defect geometric parameters and tactile signal characteristic parameters, it can obtain three-dimensional information such as defect depth and mechanical properties. This effectively overcomes the technical bottlenecks of traditional machine vision in automotive-grade devices with high reflectivity, heterogeneous surfaces susceptible to light interference, high false negative rates, and inability to acquire three-dimensional mechanical information. It achieves high-precision and interference-free identification of surface micro-defects such as submicron-level microcracks, nano-level scratches, and adhesive surface wrinkles.

[0017] 2. This invention extracts and fuses mechanical fingerprints of micro-defects, performs time-domain and frequency-domain analysis on multi-level force signals to extract multi-dimensional core features to construct a mechanical fingerprint vector, and spatially correlates and fuses the mechanical fingerprint vectors of adjacent sensing units. Based on the principle of spatial correlation to suppress interference from normal surface textures, it can effectively suppress interference from normal surface textures (such as photolithography patterns), highlight the abnormal mechanical characteristics of defect areas, and significantly improve the signal-to-noise ratio and the accuracy of feature representation.

[0018] 3. This invention employs an improved C3D-Attention neural network model for intelligent discrimination. Through a three-level cascaded feature extractor with dual-branch convolutional kernels of different sizes (different receptive fields in the temporal dimension), based on the principle of multi-scale temporal receptive field feature extraction, it can simultaneously capture local details and global temporal dependencies. Combining spatial attention and channel attention mechanisms, based on the principle of adaptive feature refinement, it effectively avoids redundant features and enhances the extraction capability of key defect features. Furthermore, the feature maps output from the three-level feature extractor are fused and input into the classifier. Based on the principle of multi-level feature fusion, this further improves the model's accuracy in classifying and locating complex micro-defects, achieving a false detection rate ≤100PPM, meeting the AEC-Q100 automotive-grade reliability requirements. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the method for detecting micro-defects on the surface of automotive-grade semiconductor devices based on machine tactile feedback, as described in this invention. Figure 2 This is a schematic diagram of the flexible tactile array sensor structure of the present invention; Figure 3 This is an exploded schematic diagram of the piezoelectric sensing unit in the flexible tactile array sensor of the present invention; Figure 4 This is a top view of the piezoelectric sensing unit of the present invention after removing the contact head; Figure 5 This is a schematic diagram of the improved C3D-Attention neural network model of this invention; Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention.

[0021] Figure label: 100-Flexible tactile array sensor, 110-Piezoelectric sensing unit, 101-Contact head, 102-Substrate, 103-Inter-electrode film, 104-Upper electrode, 105-Lower electrode. Detailed Implementation

[0022] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0023] In the description of this invention, it should be noted that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0024] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0025] Example 1 like Figure 1 As shown, this embodiment provides a method for detecting micro-defects on the surface of automotive-grade semiconductor devices based on machine tactile sensing, including: S1. Multi-force level contact tactile signal acquisition: Control the flexible tactile array sensor to scan the surface of the device under test along a preset path, apply multi-level normal forces to each detection point on the surface of the device in sequence, and simultaneously acquire the three-dimensional signals of normal force, tangential force, and deformation displacement of each sensing unit in the flexible tactile array sensor to generate a position-multi-force level mechanical feature dataset. S2. Micro-defect mechanical fingerprint extraction and feature fusion: Perform time-domain and frequency-domain analysis on the multi-level force signal of each detection point in the location-multi-level mechanical feature dataset, extract multi-dimensional core features, and construct the multi-level mechanical fingerprint vector of the detection point; and spatially correlate and fuse the mechanical fingerprint vectors of adjacent sensing units to obtain a fused mechanical feature vector that highlights the abnormal mechanical features of the defect area. S3, Intelligent identification of automotive-grade micro-defects: The fused mechanical feature vector is input into a pre-trained improved C3D-Attention neural network model, which outputs the defect type and size of the corresponding detection point; The improved C3D-Attention neural network model includes a three-stage cascaded feature extractor and classifier. Each feature extractor includes two neural network branches with different convolutional kernel sizes. One branch applies a spatial attention mechanism, and the other branch applies a channel attention mechanism. The output features of the two branches are fused and used as the output of that level of feature extractor. The output of the previous level of feature extractor is used as the input of the next level of feature extractor, and the output feature maps of the three level feature extractors are fused and then input into the classifier.

[0026] This invention employs multi-force-level contact tactile signal acquisition, micro-defect mechanical fingerprint extraction and feature fusion, and automotive-grade micro-defect intelligent discrimination techniques. Based on the quantitative coupling relationship between micro-defect geometric parameters and tactile signal feature parameters, it solves the technical problems of traditional visual inspection in automotive-grade devices with high reflectivity, heterogeneous surfaces susceptible to light interference, high false negative rates, and inability to obtain three-dimensional mechanical information. It achieves high-precision online detection of surface micro-defects such as submicron-level microcracks, nano-level scratches, and adhesive surface wrinkles.

[0027] Specifically, in step S1, in the multi-force level contact tactile signal acquisition, three levels of normal force are applied sequentially, each force is held for a preset time, and continuous mechanical time series data is acquired at a preset sampling frequency. After removing environmental vibration and electromagnetic interference noise, the position-multi-force level mechanical feature dataset is generated.

[0028] In practice, the specific process of step S1 is as follows: (1) The detection system consists of a flexible tactile array sensor, a three-axis motion platform, a detection platform, a signal acquisition unit, and an automotive-grade industrial control computer. The detection platform (with a vacuum suction fixture) is used to place the device under test. The three-axis motion platform is used to carry the flexible tactile array sensor to scan the surface of the device under test along a preset path and apply multi-level normal forces to each detection point on the surface of the device in sequence. The signal acquisition unit is used to collect signals. The automotive-grade industrial control computer is used to control the three-axis precision motion platform and to receive the signals collected by the signal acquisition unit.

[0029] like Figures 2 to 4 As shown, the flexible tactile array sensor 100 includes several piezoelectric sensing units. Each piezoelectric sensing unit includes a substrate 102, a lower electrode 105, an inter-electrode film 103, an upper electrode 104, and a contact head 101 arranged sequentially. The substrate 102 is made of polyimide (PI), PDMS, or PET, with a thickness ranging from 50 μm to 150 μm. In this embodiment, 75 μm PI is preferred, used to provide flexible support. The lower electrode 105 is a common electrode, with only one lower electrode in each piezoelectric sensing unit. It is square in shape, with a side length ranging from 400 μm to 800 μm. In this embodiment, it is 600 μm. The material is silver nanowires or a Cr / Au bilayer structure, with a thickness ranging from 100 nm to 300 nm. In this embodiment, a 200 nm silver nanowire conductive layer is preferred. The inter-electrode film 103 is made of P(VDF-TrFE) (polyvinylidene fluoride-trifluoroethylene copolymer), PI / PZT nanofiber composite material, or AlN film, with a thickness ranging from 1 μm to 5 μm. In this embodiment, a 3 μm P(VDF-TrFE) film is preferred. This film completely covers the lower electrode and is polarized along the thickness direction (polarization electric field 80 MV / m). There are four upper electrodes 104, which are distributed in a 2×2 array on the upper surface of the inter-electrode film 103. Each upper electrode is square with a side length ranging from 200 μm to 300 μm. In this embodiment, it is 250 μm. The gap between the four upper electrodes in the same unit ranges from 30 μm to 50 μm. In this embodiment, it is 40 μm. The upper electrode material is a Cr / Au bilayer (Cr 10 nm, Au 80 nm) or graphene. The contact head 101 is made of PDMS and is shaped like a regular square frustum (it can be solid or hollow). The side length of its lower base ranges from 500μm to 700μm (to cover the four upper electrode areas), and in this embodiment, it is 580μm. The side length of its upper base ranges from 300μm to 500μm, and in this embodiment, it is 400μm. The height ranges from 200μm to 400μm, and in this embodiment, it is 300μm. The angle between the side surface and the bottom surface of the frustum is approximately 75°. The contact head 101 is used to directly contact the surface of the semiconductor device, decomposing and transmitting external three-dimensional forces to the four upper electrode areas.

[0030] All piezoelectric sensing units share a common substrate 102. Different piezoelectric sensing units are isolated from each other by etched trenches on the inter-electrode film 103. The trench width ranges from 10μm to 30μm, and in this embodiment, it is 20μm. The depth is approximately 80% of the thickness of the inter-electrode film, and in this embodiment, it is 2.5μm. The contact heads 101 between adjacent piezoelectric sensing units are connected by thin layers of the same material to form an integral structure. Within each piezoelectric sensing unit, one lower electrode 105 and four upper electrodes 104 are independently led out. Figure 2 (Leads not shown): The lower electrode is connected to the peripheral pad via silver paste leads (40μm-60μm linewidth, 50μm in this embodiment) on the substrate surface and connected to the virtual ground of the subsequent charge amplifier; the four upper electrodes are led out to the multiplexer switch via Au microstrip lines (20μm-40μm linewidth, 30μm in this embodiment) printed on the surface of the inter-electrode thin film. This embodiment uses a 4×4 array as an example, with a total of 16 sensing units, therefore there are 16 lower electrode leads and 64 upper electrode leads. The external signal processing circuit first selects the lower electrode of each sensing unit row by row through the analog switch (to ground it), and at the same time reads the charge signals of the four upper electrodes in that unit. The reading time for each unit is about 20μs, and the full frame scan time is about 1.28ms.

[0031] The measurement principle is explained in detail below: The inter-electrode film 103 is a piezoelectric material. When external pressure is transmitted to the inter-electrode film through the contact head 101, polarization charges are generated inside the film. The piezoelectric effect follows the following relationship: ,in The piezoelectric constant (P(VDF-TrFE) in this embodiment) is (Approximately -30 pC / N) This is the normal force component acting perpendicularly on this local area. This represents the amount of charge generated. For each interelectrode thin film region covered by the upper electrode, the output charge is collected by the corresponding upper electrode and converted into a voltage signal by a charge amplifier: ,in The feedback capacitor is 10pF. Since the contact head 101 is designed as a regular square frustum, its sides have a certain tilt angle. When subjected to a tangential force (parallel to the sensor surface), the tilted sides of the frustum decompose the tangential force into normal components that act on the upper electrode regions of the four quadrants, and the direction of the normal force increment differs in different quadrants. Specifically, as... Figure 4 As shown, assume that the four upper electrodes within a sensing unit are numbered A (top left), B (top right), C (bottom right), and D (bottom left) in counter-clockwise order. When only normal force is present... When pressed vertically downwards, the four upper electrodes experience uniform pressure, generating equal amounts of charge, resulting in essentially consistent output voltages. However, when a tangential force exists along the positive X-axis (from left to right)... At this time, the two upper electrodes (B and C) on the right side of the frustum experience increased additional normal compressive stress, while the two upper electrodes (A and D) on the left side experience decreased normal compressive stress or even slight tensile stress (depending on the shape of the frustum and friction), thus creating a differential signal in the four outputs. Similarly, the tangential force along the positive Y-axis (from front to back) This will increase the output of the two rear upper electrodes (C and D) and decrease the output of the two front upper electrodes (A and B). Based on the above physical mechanism, this sensor uses the following linear decoupling formula to calculate the three-dimensional force components:

[0032] in The calibration coefficients are obtained by applying known normal and tangential forces (e.g., using a microforce gauge with a precision displacement stage) and performing linear fitting. For the sensor in this embodiment, the typical calibration value is... The linear range of the normal force is 0–500 mN, and the linear range of the tangential force is ±100 mN. It should be noted that in practical applications, due to the nonlinearity of the inter-electrode film and the slight influence of contact head deformation, a quadratic term correction can be added to the above linear model. However, in this embodiment, by optimizing the frustum pyramid angle and material hardness, the nonlinear error has been controlled within 2%. Therefore, the linear model is sufficient to meet the accuracy requirements for micro-defect detection in automotive-grade semiconductor devices.

[0033] Through the aforementioned structural combination, this flexible tactile array sensor can convert minute protrusions, depressions, or cracks on the surface of semiconductor devices into three-dimensional force distribution electrical signals. Experimental tests show that the sensor's normal force sensitivity can reach 0.01μN, tangential force resolution 0.02μN, and response frequency 10kHz. Furthermore, after 500 temperature cycles within the range of -40℃ to 125℃, the piezoelectric constant change is less than 5%, and the sensitivity decay is less than 3%, making it fully suitable for automotive-grade operating environments. During manufacturing, PI is spin-coated onto a glass substrate and then peeled off to obtain the substrate. Next, the lower electrode and leads are inkjet-printed, a piezoelectric thin film is spin-coated and polarized, isolation trenches are created through reactive ion etching, and the upper electrode and its leads are fabricated through electron beam evaporation and lift-off processes. Finally, a PDMS contact head is molded and assembled using oxygen plasma bonding. After overall encapsulation, a parylene conformal coating is applied to enhance moisture resistance and voltage withstand performance, enabling online detection of micro-defects on the surface of automotive-grade IGBTs, SiC MOSFETs, and other semiconductor devices.

[0034] (2) Benchmark calibration: Place standard, defect-free automotive-grade components on the testing platform, control the tactile array to contact the surface with three preset normal forces of 0.5mN, 1mN, and 2mN, collect benchmark mechanical signals, establish a standard mechanical fingerprint library, and correct the zero-point drift and temperature drift of the sensor. (3) Positioning and clamping of the device under test: The automotive-grade semiconductor device under test is fixed by vacuum adsorption fixture to ensure that the surface level of the device is ≤0.1μm; the motion platform calibrates the device coordinate system and sets the scanning path: scans line by line along the X / Y axis (with the two adjacent vertical sides in the horizontal plane where the detection platform is located as the X axis and Y axis respectively) to cover the entire surface of the device; (4) Multi-force level signal acquisition: Control the three-axis motion platform to drive the tactile array to scan along the preset path, and apply three levels of normal force (F1=0.5mN, F2=1mN, F3=2mN) to each detection point in sequence. Each level of force is held for 1ms. The normal force, tangential force and deformation displacement three-dimensional signals of each sensing unit are collected synchronously through the signal acquisition device (signal acquisition card). Continuous mechanical time series data are obtained at a sampling frequency of 20kHz. Environmental vibration and electromagnetic interference noise are removed. Signal filtering, interpolation and normalization are completed to generate position-multi-force level mechanical feature dataset.

[0035] It should be noted that removing environmental vibrations and electromagnetic interference noise, and performing signal filtering, interpolation, and normalization are all conventional methods in the prior art, and will not be elaborated upon in this invention.

[0036] Specifically, in step S2, the extraction of multi-dimensional core features and the construction of a multi-force level mechanical fingerprint vector for the detection point are as follows: extract the peak force, force change rate, deformation gradient, and frequency domain energy spectrum features of the multi-level force signal for each detection point, and construct a multi-force level mechanical fingerprint vector.

[0037] In practice, the specific process of step S2 is as follows: (1) Perform time-domain analysis on multi-level force signals to extract peak force and force change rate; perform spatial differentiation to extract deformation gradient; perform FFT frequency domain transformation to extract frequency domain energy spectrum; each class contains 3 sub-features, and a total of 12-dimensional multi-level force fingerprint vectors are constructed.

[0038] The constructed 12-dimensional multi-force mechanical fingerprint vector consists of 3 levels of force × 4 major categories of basic features. Each level of force extracts 4-dimensional features, as detailed below: Level 1 force F1 = [V1 peak force, V2 rate of change of force, V3 deformation gradient, V4 frequency domain energy spectrum] Level 2 force F2 = [V5 peak force, V6 rate of change of force, V7 deformation gradient, V8 frequency domain energy spectrum] Level 3 force F3 = [V9 peak force, V] 10 Rate of change of force, V 11 Deformation gradient, V 12 Frequency domain energy spectrum Among them, the first dimension peak force is the maximum value of the time-series signal acquired during each force holding period; the second dimension force change rate is calculated by the formula: (peak force - initial force) / holding time; the third dimension deformation gradient is obtained by spatially differentiating the deformation displacements of three adjacent sensing units at the same detection point; the third dimension frequency domain energy spectrum is obtained by performing an FFT transform on the time-series signal, taking the total energy and normalizing it.

[0039] (2) A hierarchical feature fusion algorithm is adopted to spatially correlate the mechanical fingerprints of adjacent sensing units, suppress the interference of normal textures (such as photolithography patterns) on the device surface, highlight the abnormal mechanical features of the defect area, and obtain the fused mechanical feature vector.

[0040] Specifically, in step S3, in the two neural network branches of the feature extractor, one branch uses a convolutional kernel with a temporal receptive field of 1, and the other branch uses a convolutional kernel with a temporal receptive field greater than 1; and each branch structure contains multiple 3D convolutional blocks, and all max pooling layers are not reduced in the spatial length dimension. The spatial attention mechanism uses three-dimensional convolution and three-dimensional pooling operations to derive a three-dimensional spatial attention feature map; the channel attention mechanism uses three-dimensional pooling operations combined with a multilayer perceptron to derive a one-dimensional channel attention feature map; the enhanced feature maps of the two branches are fused element-wise. The classifier includes a three-dimensional global average pooling layer, a discard layer, a fully connected layer, and an output layer.

[0041] In practice, the specific process of step S3 is as follows: (1) Load the pre-trained automotive-grade defect discrimination model (based on the improved C3D-Attention algorithm). The model uses the fused mechanical feature vector of automotive-grade defect samples as the training set and meets the AEC-Q100 standard requirement of false detection rate ≤100PPM. Among them, such as Figure 5The diagram shows the structure of the improved C3D-Attention neural network model. In this model, the feature extractor is primarily used to identify and extract surface features of the image, generating feature maps. The classifier's input is the feature map output from the feature extractor, which is used to output the neural network's classification results. The feature extractor has two branches with the same structure, but each branch uses different kernel sizes (1×3×3 and 3×3×3, respectively). This is primarily to address complex surface defects by using different kernel sizes to extract differentiated feature information, thus enhancing the neural network's feature extraction capabilities. Each branch of the C3D-Attention neural network contains three 3D convolutional blocks of size s×h×w, where s represents the spatial length dimension, h represents the image length, and w represents the image width. Due to the different kernel sizes in the two branches, to ensure subsequent feature fusion in the two-branch structure, the number of channels and shape of the output feature map from the last convolutional block in each branch are consistent; therefore, the spatial length dimension of each branch is not reduced. The first and second convolutional blocks each have two convolutional layers with 64 and 128 channels respectively, and a max-pooling layer. The third convolutional block has three convolutional layers and one max-pooling layer with 256, 512, and 512 channels respectively. The ReLU activation function is used as the activation function in the feature extractor in each branch. To ensure that each convolutional layer can better learn and pass on spatial feature information, the spatial length dimension (s-dimension) of all max-pooling layers is set to 1.

[0042] To further enhance the different feature extraction capabilities of the two branches, channel attention and spatial attention mechanisms are applied to the two-branch structure, respectively. The application of attention mechanisms enables adaptive feature refinement of the input feature map, effectively avoiding redundant features in the two-branch structure to a certain extent. Assume that the output feature map of the last convolutional block in branch 1 is... As the input to the spatial attention mechanism (where c is the number of channels), the spatial attention mechanism can derive a three-dimensional spatial attention feature map as follows: The enhanced feature map is obtained according to the following formula. : (1) (2) in, σ This represents the sigmoid activation function. This indicates a convolution operation with a 3×3×3 kernel. This represents the generated 3D average pooling feature map. This represents the generated 3D max-pooled feature map. This indicates element-wise multiplication.

[0043] Suppose the output feature map of the last convolutional block on branch 2 is As the input to the channel attention mechanism, the channel attention mechanism can derive a one-dimensional channel attention feature map. The enhanced feature map is obtained according to the following formula. : (3) (4) in, Represents the weights of the MLP (multi-layer perceptron), W0∈R c / r×c W1∈R c×c / r , r The reduction rate. This represents the generated average pooling feature map. This represents the generated max-pooling feature map.

[0044] Finally, the final output feature information of the dual-branch structure is fused to obtain the final output feature map F of the feature extractor. out for: (5) Here, ⊕ represents element-wise addition.

[0045] The three-level feature extractor adopts a cascaded structure, with the output of the previous level serving as the input of the next level. The feature maps of the three levels are spliced ​​and fused along the channel dimension to obtain multi-level features containing details, intermediate layers, and global semantics, which are then input into the classifier to complete defect identification.

[0046] The classifier mainly consists of a 3D global average pooling layer, a dropout layer, a fully connected layer, and an output layer. The classifier's head is a 3D global average pooling layer, which can replace the fully connected layer, reducing the storage space required by the large weight matrix in the fully connected layer. The dropout layer is a simple technique to avoid overfitting; it temporarily discards some neurons from the neural network with a certain probability. Here, the dropout rate of the dropout layer is set to 0.2. Furthermore, the number of neurons in the fully connected layer is set to 256. Finally, the number of neurons in the output layer depends on the number of categories that need to be distinguished in the image classification task.

[0047] (2) Based on the fused mechanical feature vector of the detection point, output the defect type and defect size (length / depth / width), and use the discrimination threshold to adaptively match the quality standards of automotive-grade components; (3) The industrial control computer generates an inspection report, marks the location, type and size of defects, and generates a defect heat map; it communicates with the production line digital twin system to automatically mark and sort unqualified parts, and uploads the inspection data to the quality control platform to achieve full-process traceability.

[0048] Example 2 This embodiment provides a machine-tactile-based system for detecting micro-defects on the surface of automotive-grade semiconductor devices, including: The signal acquisition module is used to control the flexible tactile array sensor to scan the surface of the device under test along a preset path, apply multi-level normal forces to each detection point on the surface of the device in sequence, and simultaneously acquire the three-dimensional signals of normal force, tangential force, and deformation displacement of each sensing unit in the flexible tactile array sensor to generate a position-multi-force level mechanical feature dataset. The feature fusion module is used to perform time-domain and frequency-domain analysis on the multi-level force signals of each detection point in the location-multi-level mechanical feature dataset, extract multi-dimensional core features, construct the multi-level mechanical fingerprint vector of the detection point, and spatially correlate and fuse the mechanical fingerprint vectors of adjacent sensing units to obtain a fused mechanical feature vector that highlights the abnormal mechanical features of the defect area. The intelligent discrimination module is used to input the fused mechanical feature vector into a pre-trained improved C3D-Attention neural network model and output the defect type and size of the corresponding detection point. The improved C3D-Attention neural network model includes a three-level cascaded feature extractor and classifier. Each level of feature extractor includes two neural network branches with different convolutional kernel sizes. One branch applies a spatial attention mechanism, and the other branch applies a channel attention mechanism. The output features of the two branches are fused and used as the output of the feature extractor at that level. The output of the previous level of feature extractor is used as the input of the next level of feature extractor, and the output feature maps of the three level feature extractors are fused and then input into the classifier.

[0049] Figure 6 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other through the communication bus 840. The processor 810 can call the logic instructions in the memory 830 to execute a method for detecting micro-defects on the surface of automotive-grade semiconductor devices based on machine tactile sensing. The method includes: S1, multi-force-level contact tactile signal acquisition: controlling a flexible tactile array sensor to scan the surface of the device under test along a preset path, sequentially applying multi-level normal forces to each detection point on the device surface, and simultaneously acquiring the three-dimensional signals of normal force, tangential force, and deformation displacement of each sensing unit in the flexible tactile array sensor to generate a position-multi-force-level mechanical feature dataset; S2. Micro-defect mechanical fingerprint extraction and feature fusion: Perform time-domain and frequency-domain analysis on the multi-level force signal of each detection point in the location-multi-level mechanical feature dataset, extract multi-dimensional core features, and construct the multi-level mechanical fingerprint vector of the detection point; and spatially correlate and fuse the mechanical fingerprint vectors of adjacent sensing units to obtain a fused mechanical feature vector that highlights the abnormal mechanical features of the defect area. S3, Intelligent identification of automotive-grade micro-defects: The fused mechanical feature vector is input into a pre-trained improved C3D-Attention neural network model, which outputs the defect type and size of the corresponding detection point; The improved C3D-Attention neural network model includes a three-stage cascaded feature extractor and classifier. Each feature extractor includes two neural network branches with different convolutional kernel sizes. One branch applies a spatial attention mechanism, and the other branch applies a channel attention mechanism. The output features of the two branches are fused and used as the output of that level of feature extractor. The output of the previous level of feature extractor is used as the input of the next level of feature extractor, and the output feature maps of the three level feature extractors are fused and then input into the classifier.

[0050] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the above-mentioned method for detecting micro-defects on the surface of automotive-grade semiconductor devices based on machine tactile feedback. The method includes: S1, multi-force level contact tactile signal acquisition: controlling a flexible tactile array sensor to scan the surface of the device under test along a preset path, sequentially applying multi-level normal forces to each detection point on the surface of the device, and simultaneously acquiring the three-dimensional signals of normal force, tangential force, and deformation displacement of each sensing unit in the flexible tactile array sensor to generate a position-multi-force level mechanical feature dataset; S2. Micro-defect mechanical fingerprint extraction and feature fusion: Perform time-domain and frequency-domain analysis on the multi-level force signal of each detection point in the location-multi-level mechanical feature dataset, extract multi-dimensional core features, and construct the multi-level mechanical fingerprint vector of the detection point; and spatially correlate and fuse the mechanical fingerprint vectors of adjacent sensing units to obtain a fused mechanical feature vector that highlights the abnormal mechanical features of the defect area. S3, Intelligent identification of automotive-grade micro-defects: The fused mechanical feature vector is input into a pre-trained improved C3D-Attention neural network model, which outputs the defect type and size of the corresponding detection point; The improved C3D-Attention neural network model includes a three-stage cascaded feature extractor and classifier. Each feature extractor includes two neural network branches with different convolutional kernel sizes. One branch applies a spatial attention mechanism, and the other branch applies a channel attention mechanism. The output features of the two branches are fused and used as the output of that level of feature extractor. The output of the previous level of feature extractor is used as the input of the next level of feature extractor, and the output feature maps of the three level feature extractors are fused and then input into the classifier.

[0052] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the above-described method for detecting micro-defects on the surface of automotive-grade semiconductor devices based on machine tactile feedback. The method includes: S1, multi-force-level contact tactile signal acquisition: controlling a flexible tactile array sensor to scan the surface of the device under test along a preset path, sequentially applying multi-level normal forces to each detection point on the surface of the device, and simultaneously acquiring the three-dimensional signals of normal force, tangential force, and deformation displacement of each sensing unit in the flexible tactile array sensor to generate a position-multi-force-level mechanical feature dataset; S2. Micro-defect mechanical fingerprint extraction and feature fusion: Perform time-domain and frequency-domain analysis on the multi-level force signal of each detection point in the location-multi-level mechanical feature dataset, extract multi-dimensional core features, and construct the multi-level mechanical fingerprint vector of the detection point; and spatially correlate and fuse the mechanical fingerprint vectors of adjacent sensing units to obtain a fused mechanical feature vector that highlights the abnormal mechanical features of the defect area. S3, Intelligent identification of automotive-grade micro-defects: The fused mechanical feature vector is input into a pre-trained improved C3D-Attention neural network model, which outputs the defect type and size of the corresponding detection point; The improved C3D-Attention neural network model includes a three-stage cascaded feature extractor and classifier. Each feature extractor includes two neural network branches with different convolutional kernel sizes. One branch applies a spatial attention mechanism, and the other branch applies a channel attention mechanism. The output features of the two branches are fused and used as the output of that level of feature extractor. The output of the previous level of feature extractor is used as the input of the next level of feature extractor, and the output feature maps of the three level feature extractors are fused and then input into the classifier.

[0053] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0054] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A machine-tactile based automotive-grade semiconductor device surface micro-defect detection method, characterized by, include: Multi-force level contact tactile signal acquisition: Control the flexible tactile array sensor to scan the surface of the device under test along a preset path, apply multi-level normal forces to each detection point on the device surface in sequence, and simultaneously acquire the three-dimensional signals of normal force, tangential force, and deformation displacement of each sensing unit in the flexible tactile array sensor to generate a position-multi-force level mechanical feature dataset; Micro-defect mechanical fingerprint extraction and feature fusion: Time-domain and frequency-domain analysis is performed on the multi-level force signal of each detection point in the position-multi-level mechanical feature dataset to extract multi-dimensional core features and construct the multi-level mechanical fingerprint vector of the detection point; The mechanical fingerprint vectors of adjacent sensing units are spatially correlated and fused to obtain a fused mechanical feature vector that highlights the abnormal mechanical characteristics of the defect area. Intelligent identification of automotive-grade micro-defects: The fused mechanical feature vector is input into a pre-trained improved C3D-Attention neural network model, which outputs the defect type and size of the corresponding detection point; The improved C3D-Attention neural network model includes a three-stage cascaded feature extractor and classifier. Each feature extractor includes two neural network branches with different convolutional kernel sizes. One branch applies a spatial attention mechanism, and the other branch applies a channel attention mechanism. The output features of the two branches are fused and used as the output of that level of feature extractor. The output of the previous level of feature extractor is used as the input of the next level of feature extractor, and the output feature maps of the three level feature extractors are fused and then input into the classifier.

2. The machine tactile based car-grade semiconductor device surface micro-defect detection method of claim 1, wherein, In the multi-force level contact tactile signal acquisition, three levels of normal force are applied sequentially, each force is held for a preset time, and continuous mechanical time series data is acquired at a preset sampling frequency. After removing environmental vibration and electromagnetic interference noise, the position-multi-force level mechanical feature dataset is generated.

3. The machine tactile based car-grade semiconductor device surface micro-defect detection method of claim 1, wherein, The extraction of multi-dimensional core features and the construction of a multi-force level mechanical fingerprint vector for the detection point specifically involves: extracting the peak force, force change rate, deformation gradient, and frequency domain energy spectrum features of the multi-level force signal for each detection point, and constructing a multi-force level mechanical fingerprint vector.

4. The machine tactile based car-grade semiconductor device surface micro-defect detection method of claim 1, wherein, In the two neural network branches of the feature extractor, one branch uses a convolutional kernel with a receptive field of 1 in the time dimension, and the other branch uses a convolutional kernel with a receptive field of more than 1 in the time dimension; and each branch structure contains multiple 3D convolutional blocks, and all max pooling layers are not reduced in the spatial length dimension.

5. The machine tactile based, car-grade, semiconductor device surface micro-defect detection method according to claim 1 or 4, wherein The spatial attention mechanism employs 3D convolution and 3D pooling operations to derive a 3D spatial attention feature map; the channel attention mechanism employs 3D pooling operations combined with a multilayer perceptron to derive a 1D channel attention feature map; the feature maps enhanced by the two branches are then fused element-wise.

6. The method for detecting micro-defects on the surface of automotive-grade semiconductor devices based on machine tactile sensing according to claim 1, characterized in that, The classifier includes a three-dimensional global average pooling layer, a discard layer, a fully connected layer, and an output layer.

7. A machine tactile based automotive grade semiconductor device surface micro-defect detection system, comprising: include: The signal acquisition module is used to control the flexible tactile array sensor to scan the surface of the device under test along a preset path, apply multi-level normal forces to each detection point on the surface of the device in sequence, and simultaneously acquire the three-dimensional signals of normal force, tangential force, and deformation displacement of each sensing unit in the flexible tactile array sensor to generate a position-multi-force level mechanical feature dataset. The feature fusion module is used to perform time-domain and frequency-domain analysis on the multi-level force signal of each detection point in the location-multi-level mechanical feature dataset, extract multi-dimensional core features, and construct the multi-level mechanical fingerprint vector of the detection point. The mechanical fingerprint vectors of adjacent sensing units are spatially correlated and fused to obtain a fused mechanical feature vector that highlights the abnormal mechanical characteristics of the defect area. The intelligent discrimination module is used to input the fused mechanical feature vector into a pre-trained improved C3D-Attention neural network model and output the defect type and size of the corresponding detection point. The improved C3D-Attention neural network model includes a three-level cascaded feature extractor and classifier. Each level of feature extractor includes two neural network branches with different convolutional kernel sizes. One branch applies a spatial attention mechanism, and the other branch applies a channel attention mechanism. The output features of the two branches are fused and used as the output of the feature extractor at that level. The output of the previous level of feature extractor is used as the input of the next level of feature extractor, and the output feature maps of the three level feature extractors are fused and then input into the classifier.

8. A flexible tactile array sensor, comprising a plurality of piezoelectric sensing units, characterized in that: Each piezoelectric sensing unit includes a substrate, a lower electrode, an interelectrode film, an upper electrode, and a contact head arranged sequentially. Each piezoelectric sensing unit has one lower electrode and four upper electrodes arranged in an array. The contact head is truncated pyramidal in shape and covers the four upper electrodes. It is used to decompose and transmit external three-dimensional forces to the interelectrode film regions corresponding to the four upper electrodes. 9.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for detecting micro-defects on the surface of automotive-grade semiconductor devices based on machine tactile feedback as described in any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for detecting micro-defects on the surface of automotive-grade semiconductor devices based on machine tactile feedback as described in any one of claims 1 to 6.

Images