A machine vision-based method for detecting the etching state of crystal oscillators

Abstract

This invention discloses a crystal oscillator etching state detection method based on machine vision, belonging to the field of electronic component manufacturing inspection technology. This invention acquires crystal oscillator surface images through a high-resolution microscopic vision system; enhances etching features through anisotropic diffusion filtering and image pyramid decomposition; extracts the crystal oscillator target region through a pixel-level segmentation network; performs sub-pixel edge localization, calculates edge roughness, size width, and symmetry deviation; constructs a gray-level co-occurrence matrix to extract surface entropy values; reconstructs the three-dimensional morphology using photometric stereo vision and calculates the etching depth; inputs multi-dimensional features into a long short-term memory network and a random forest hybrid model, outputting etching state classification and predicted remaining time; compares the judgment results with a preset threshold to generate etching completion or over-etching warning signals; achieves real-time non-contact monitoring and closed-loop control of the microscopic morphology during the etching process, improving detection accuracy and yield.

Description

Technical Field

[0001] This invention belongs to the field of electronic component manufacturing and testing, and specifically relates to a crystal oscillator etching status detection method based on machine vision. Background Technology

[0002] The etching process of a quartz crystal oscillator, by controlling the amount of material removed and the surface morphology, determines its resonant frequency and stability. Existing detection technologies have two major limitations: Methods based on electrical parameter feedback (such as monitoring frequency drift) are result-oriented reverse calculations; the signal response has physical lag and cannot reflect microscopic morphological defects (such as microcracks and edge burrs) caused by uneven chemical reactions during the etching process in real time, and are easily affected by electromagnetic interference.

[0003] Spectral analysis-based methods are limited by one-dimensional perception, and the analysis logic is built on the overall energy change of the spectral envelope. They lack the ability to distinguish two-dimensional or three-dimensional spatial morphological features such as geometric distortion at the crystal oscillator edge and non-uniform etching pits. Plasma glow and background stray light under complex working conditions will severely weaken its signal-to-noise ratio.

[0004] Existing technologies suffer from generational limitations in their detection principles, particularly the lack of microscopic morphology perception, making it difficult to achieve non-contact, high-precision, real-time monitoring of sub-micron level geometric features. Therefore, there is an urgent need to introduce higher-dimensional visual perception methods to overcome the technical bottlenecks restricting the yield of high-end crystal oscillator manufacturing. Summary of the Invention

[0005] This invention solves the technical problem of the inability to monitor the microstructure in real time, non-contact, and with high precision during the crystal oscillator etching process.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: This invention provides a machine vision-based method for detecting the etching state of crystal oscillators, comprising the following steps: Step 1: Construct a high-resolution microscopic vision acquisition system and perform raw image acquisition. The high-resolution microscopic vision acquisition system includes an optical unit consisting of dual telecentric lenses, a high frame rate CMOS image sensor, and a multi-source composite controlled light source. The magnification of the dual telecentric lenses is set to 2.0X to 5.0X, and the distortion rates on both the object side and the image side are less than 0.05% to eliminate perspective errors caused by object distance fluctuations. The high frame rate CMOS image sensor has an effective resolution of no less than 20 million pixels, a pixel size of 2.4μm×2.4μm, and supports an image sampling frequency of no less than 100fps. The multi-source composite controlled light source consists of an infrared ring light source with a center wavelength of 850nm and a blue coaxial light source with a center wavelength of 450nm. The infrared ring light source consists of multiple independently controlled LED sectors, and the light intensity is synchronously triggered at the nanosecond level through a programmable constant current drive module. In the etching reaction chamber, the optical unit images the crystal oscillator in the etching environment through a high-transmittance quartz observation window to acquire raw digital image signals containing microscopic topological information of the crystal oscillator surface.

[0007] Step two: Perform image preprocessing to enhance the signal-to-noise ratio of the etching features. The acquired raw digital image is input into the preprocessing module, where anisotropic diffusion filtering is performed to filter out plasma discharge noise and etching solution bubble interference while maintaining the gradient sharpness of the crystal oscillator edges. Anisotropic diffusion filtering is achieved by solving partial differential equations, and the diffusion coefficient function is selected to suppress diffusion in large gradient regions while enhancing diffusion in smooth regions. The partial differential equation is expressed as: ; in: Indicates spatial location and time Image grayscale value function at location; This refers to the diffusion time parameter; Image gradient; Let be the magnitude of the image gradient; For divergence operators; It is a function of the diffusion conduction coefficient; in the form of ; This is the gradient threshold constant, used to control the sensitivity of edge preservation. The value of is determined based on the image gradient histogram, and the 85th percentile of the gradient histogram is taken. Furthermore, the filtered image is subjected to histogram specification processing to map its contrast to a preset standard dynamic range; using the Laplacian operator-based image pyramid decomposition technique, the image is decomposed into spatial components of different frequencies, and gain compensation is performed on the high-frequency components representing micro-texture to highlight the surface roughness change characteristics during the etching process.

[0008] Step 3: Implement crystal oscillator target region segmentation based on deep feature extraction. The preprocessed image is fed into a pre-trained pixel-level segmentation network model. The segmentation network model adopts an improved encoder-decoder architecture. In the encoding stage, it uses a residual convolution module to extract multi-scale morphological features. In the decoding stage, it restores spatial resolution through transposed convolution and introduces a spatial attention mechanism at skip connections to accurately capture the blurred boundary between the crystal oscillator edge and the background etching environment. The output of the segmentation network model is a binary mask of the same size as the input image, clearly distinguishing the crystal oscillator body region, the area to be etched, and the background region. By multiplying the mask with the original image pixel by pixel, accurate extraction of the target detection region is achieved, eliminating the interference of irrelevant background on subsequent calculation of quantization indicators. The segmentation network model uses a joint loss function during the training stage, expressed as: ; in: This is the cross-entropy loss function, used to ensure the overall accuracy of pixel classification; The Dice coefficient loss function is used to optimize the segmentation accuracy of the slender edges and small-scale feature regions of the crystal oscillator. and These are weighting coefficients, with values ​​of 0.4 and 0.6 respectively.

[0009] Step four: extract sub-pixel level micro-geometric features of the etched edge; in the segmented crystal oscillator body edge region, apply the improved Canny operator combined with Sobel gradient magnitude calculation to initially determine the edge pixel set; fit the gray-level distribution in the edge normal direction with a Gaussian surface and use an interpolation algorithm to achieve sub-pixel level edge localization.

[0010] Furthermore, the curvature distribution function of the edge profile and the edge roughness index are calculated; the edge roughness index is calculated by the arithmetic mean deviation between the measured edge profile and its fitted average line. The discretization calculation formula is as follows: ; in: This represents the number of edge sampling points; For the first The coordinates of sub-pixel edge points; and These are the slope and intercept obtained by linear fitting of the edge point set, respectively; The average line fitted to the edge contour; This metric characterizes the impact of etching processes on the edge smoothness of the crystal oscillator. The system simultaneously extracts key geometric dimensions of the crystal oscillator, including its length and width, as sub-pixel width feature vectors. This dimension is obtained by calculating the minimum bounding rectangle of the edge point set within the segmented region, and the measurement accuracy is improved to the 0.1 pixel level using sub-pixel edge localization results. Edge closure checks and contour tracking are performed, and a distance-transform-based skeleton extraction algorithm is used to obtain the geometric center axis of the crystal oscillator. The symmetry deviation of each edge point relative to the center axis is calculated using the following formula: ; in: This represents the number of edge points; For the first The perpendicular distance from each edge point to the central axis; This is the average distance from all edge points to the central axis.

[0011] Step 5: Perform feature quantization of the crystal oscillator surface morphology and etching depth; by analyzing the gray-level statistical distribution characteristics of the crystal oscillator surface in the image, construct a gray-level co-occurrence matrix to extract second-order statistical parameters such as surface energy, moment of inertia, correlation, and surface entropy; the surface entropy index is used to quantify the complexity of the surface texture during etching, and the calculation formula is as follows: ; in: The grayscale values ​​of pixel pairs in the grayscale co-occurrence matrix are... and The joint probability density; utilizing the projection relationship generated by multi-source composite light sources, photometric stereo vision technology is used to reconstruct the three-dimensional topological morphology of the crystal oscillator surface; the infrared ring light source consists of multiple independently controlled LED sectors, which are sequentially illuminated by a programmable drive module to obtain image sequences of the crystal oscillator surface under illumination from four different directions; based on the Lambertian reflection model, a reflectivity map equation is established, and the surface normal vector of each pixel is solved using the least squares method, thereby integrating and reconstructing the three-dimensional topological morphology; etching depth The height was calculated by comparing the reconstructed 3D height maps before and after etching; assuming This is a reference height map collected before etching. For the 3D height map reconstructed at the current moment, the coordinates are... Local etching depth at the location The calculation formula is: Instantaneous etch depth Take the etched area on the crystal oscillator surface All pixels The average value is calculated using the following formula: ; in for The total number of pixels within the depth measurement; this depth measurement is calibrated by a laser interferometer, and the repeatability is better than 0.08μm within the range of 0-50μm.

[0012] Step six: Construct an etching state classification and prediction model based on temporal feature fusion; and apply the edge roughness index extracted in the previous steps. Surface entropy Edge symmetry deviation Subpixel size width and instantaneous etching depth Feature vectors are arranged along the sampling time axis to form a time-series dataset reflecting the dynamic evolution of etching. This dataset is input into a hybrid model based on a Long Short-Term Memory (LSTM) network and a Random Forest Regressor. The model uses an LSM layer containing 128 hidden units to capture the nonlinear evolution trend of feature parameters over time, utilizing its forget gate, input gate, and output gate structure to capture the nonlinear growth trend of etching depth over time. The output features of the LSM layer are fed into a Random Forest Regressor composed of 200 decision trees, which outputs two key results through ensemble learning: the classification probability of the current etching state, i.e., determining whether it is currently in the initial etching period, stable etching period, or over-etching risk period; and the predicted remaining etching time required to reach the target frequency. The inference time of the Long Short-Term Memory network model is 7.2ms, and it is deployed on an industrial computer equipped with a graphics processing unit.

[0013] Step 7: Implement feedback control and process closed-loop management of the detection results; compare the etching state judgment results obtained in step 6 with the preset process threshold matrix in real time; when the predicted etching depth... The design target value was achieved, and the edge roughness index was... When the geometric dimensional deviations are all within the allowable tolerance range, the detection system generates an etching completion trigger signal; if the model determines that the etching risk period has passed, the system immediately generates an early warning signal and sends a power reduction command to the etching equipment control unit via the industrial fieldbus to implement a soft landing shutdown strategy; the signal is transmitted to the etching equipment control unit in real time via the industrial fieldbus, and the command to shut down the etching source is immediately executed; if local non-uniform etching or edge distortion is detected to exceed the preset threshold, the system issues a process abnormality warning, records the current coordinate position of the crystal oscillator, and adjusts the etching process parameters for real-time compensation.

[0014] Furthermore, in the first step, the optical axis of the high-resolution microscopic visual acquisition system is perpendicular to the horizontal stage of the crystal oscillator; the incident angle of the infrared ring light source is set to 45° to 60°, utilizing the penetrating characteristics of infrared light and the oblique illumination at a specific angle to enhance the visibility of microcracks and the bottom of deep hole etching on the crystal oscillator surface; the blue coaxial light source is used to provide uniform vertical illumination to eliminate bright spots on high reflectivity surfaces and ensure that the grayscale gradient of the crystal oscillator edge contour has high contrast.

[0015] Furthermore, in the second step, the number of iterations of the anisotropic diffusion filter is set to 5 to 15, and the time step is controlled between 0.1 and 0.2.

[0016] Furthermore, in the third step, during the deep feature extraction process, the spatial attention mechanism obtains the channel-level weight coefficients through global pooling operations and then performs weighted fusion with the feature map.

[0017] Furthermore, in the fourth step, after completing the sub-pixel edge localization, the edge geometric features also include the extraction of the etching slope. By analyzing the gray-scale gradient width of the edge transition region and combining it with the optical scattering model, the tilt angle of the etched sidewall is indirectly calculated.

[0018] Furthermore, in the fifth step, the distance step size for calculating the gray-level co-occurrence matrix is... The value ranges from 1 to 3 pixels, and the angle direction includes 0°, 45°, 90° and 135°. When the etching reaction moves from the chemical polishing stage to the crystal preferred orientation stripping stage, the surface entropy value shows a step change. This feature is extracted as a key criterion for determining the micro-state of the etching.

[0019] Furthermore, in the sixth step, the hybrid model employs a data augmentation strategy during the training phase. By randomly rotating, scaling, and contrast-changing historical detection images, as well as adding Gaussian noise, an enhanced training set is constructed to improve the model's generalization ability under different devices, different batches of wafers, and different lighting fluctuations. The system's end-to-end response latency is 18.5ms, meeting the dynamic response requirements of the etching process for real-time monitoring.

[0020] Furthermore, in the seventh step, the feedback control also includes adaptive calibration of the detection environment; the system periodically calls the standard calibration board image, and updates the transformation matrix from the pixel coordinate system to the physical coordinate system in real time by detecting feature points with known spacing on the calibration board, in order to correct the systematic error of geometric dimension measurement and ensure the long-term stability of measurement accuracy.

[0021] Furthermore, the machine vision-based crystal oscillator etching state detection method introduces lattice orientation prior constraints in the feature extraction algorithm for quartz crystals with different cuts. Based on the anisotropic etching characteristics of quartz crystals, different weight coefficients are assigned to the contour components of different crystal axis directions when calculating the edge roughness index, so as to more realistically reflect the physical relationship between the crystal oscillator vibration performance and the geometric shape.

[0022] Furthermore, the detection method establishes a cloud-based process feature database to collaboratively analyze feature data collected by multiple detection terminals; it uses big data clustering algorithms to identify consistency shifts in specific batches of raw materials during the etching process, thereby achieving cross-machine optimization management of process parameters and ensuring that the frequency accuracy distribution in mass crystal oscillator production meets the narrowband requirements of a normal distribution.

[0023] Furthermore, in step four, for the AT-cut quartz crystal oscillator, the edge profile is decomposed into components along the X-axis and along the Z'-axis, and the dominant edge roughness index along the X-axis is calculated respectively. and Z' axis dominant edge roughness index Edge roughness index Represented as a direction-weighted combination ;in This is the directional weighting coefficient.

[0024] Furthermore, in step six, the hybrid model is a long short-term memory network model based on anisotropic etching kinetics compensation. This model includes a physical information fusion layer, which receives the crystal oscillator's cutting parameters and the real-time temperature of the etching solution. ,concentration and pressure As an external input, it is mapped to an etch rate compensation factor through a fully connected network. Etching rate compensation factor The calculation formula is: ; in and These are trainable weight matrices and bias vectors; Sigmoid activation function; etch rate compensation factor The cell state update equation, injected into the Long Short-Term Memory network, is modified to... ; in Candidate cell state; This represents the updated cell state. This represents the cell state at the previous moment; Output for the forget gate; For input gate output; This indicates element-wise multiplication.

[0025] Furthermore, in step six, the hybrid model based on Long Short-Term Memory (LSTM) network and Random Forest Regressor is replaced with a Physically Augmented LSTM network model; the Physically Augmented LSTM network model includes a Physical Information Fusion Layer, an LSTM network layer, and a Random Forest Regressor layer; the Physical Information Fusion Layer receives the real-time temperature of the etching solution. ,concentration and pressure As an external input, it is mapped to an etch rate compensation factor through a multi-layer fully connected network. Etching rate compensation factor The calculation formula is: ; in: For external input vectors; and This represents the weight matrix for the first and second fully connected layers of the network. and It is the bias vector; It is a linear rectification activation function; Sigmoid activation function; etch rate compensation factor The value range is from 0.5 to 1.5.

[0026] Furthermore, the etch rate compensation factor output by the physical information fusion layer The cell state update equation of the standard long short-term memory network, injected into the long short-term memory network layer, is modified as follows: ; ; in: Candidate cell state; This represents the updated cell state. This represents the cell state at the previous moment; This represents the basic cellular state at the previous moment; Output for the forget gate; For input gate output; This is the hidden state from the previous moment; The visual feature vector input at the current moment; and The weight matrix and bias vector represent the candidate cell states. This indicates element-wise multiplication.

[0027] Furthermore, in step five, the feature quantization of the crystal oscillator surface morphology and etching depth also includes extracting multispectral fusion feature vectors. Multispectral fusion feature vector From infrared spectral surface entropy value Green spectral surface entropy Blue spectrum surface entropy and the volume of the etching pits Root mean square of surface slope Composition; In step six, the visual feature vector input at the current time. Multispectral fusion feature vector With edge geometric feature vectors The concatenated vector, i.e. ;in ; Edge geometric feature vector Edge roughness index in Directional weighted combination ; in and This refers to the directional weighting coefficient; These are the edge roughness indices for the X-axis and Z'-axis directions, respectively.

[0028] Furthermore, the training process of the physically-enhanced long short-term memory network model includes the following steps: Step A: Construct a training dataset, which contains sequences of visual feature vectors collected during historical production processes. External input vector sequence and the corresponding etching depth tag sequence and remaining etch time stamp sequence ; Step B involves inputting the visual feature vector sequence and the external input vector sequence into a long short-term memory network model based on physical information enhancement, and then performing forward computation to obtain the predicted etching depth. and the predicted remaining etching time ; Step C, employ the joint loss function Train the model. in This represents the mean square error loss for the etch depth prediction; The mean squared error loss for the remaining time prediction; For weight decay regularization; , and These are the preset weighting coefficients.

[0029] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs a high-resolution microscopic vision system composed of dual telecentric lenses and multi-source composite light sources to directly acquire microscopic images of the crystal oscillator surface. It also employs a sub-pixel-level edge positioning algorithm to extract edge roughness, size width, and symmetry deviation, thereby achieving non-contact and precise monitoring of the microscopic geometric morphology during the etching process and overcoming the physical lag problem of traditional electrical parameter feedback.

[0030] Anisotropic diffusion filtering and multispectral composite illumination are employed to filter out environmental interferences such as plasma glow and etching bubbles, while enhancing edge gradient contrast, thus ensuring the signal-to-noise ratio and stability of the detection system in complex etching environments.

[0031] A pixel-level segmentation network with an integrated spatial attention mechanism is introduced to achieve accurate extraction of the crystal oscillator body. A hybrid model based on long short-term memory network and random forest is constructed to integrate multi-dimensional temporal features such as edge roughness, surface entropy value, and etching depth, and output etching state classification and remaining time prediction, realizing a technological leap from post-detection to process prediction.

[0032] By obtaining multispectral surface entropy values ​​through dynamic multispectral synchronous illumination and combining them with physical information from anisotropic etching kinetics compensation to enhance the long short-term memory network model, the prediction accuracy and adaptability to process fluctuations are improved, the over-etching false positive rate is reduced, and the manufacturing yield of high-frequency ultrathin crystal oscillators is increased. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0034] Figure 1 This is a flowchart of the method described in this invention.

[0035] Figure 2 A system architecture diagram for implementing the method described in this invention. Detailed Implementation

[0036] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0037] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0038] See Figures 1-2This embodiment discloses a crystal oscillator etching state detection method based on machine vision: In the first step of the method, constructing a high-resolution microscopic visual acquisition system is the fundamental hardware guarantee for achieving precision detection. The system adopts a vertical optical path design in its physical architecture, with the optical axis strictly perpendicular to the crystal oscillator surface on the horizontal stage, and the perpendicularity error controlled within ±0.01°. The core of the optical unit consists of a double telecentric lens, with a magnification of 3.5X selected in the design. The double telecentric lens has extremely high object-side and image-side telecentricity, and can maintain a constant magnification over a large depth range, eliminating perspective distortion caused by liquid level fluctuations or slight displacement of the crystal oscillator during the etching process. The distortion rate of the lens was measured to be 0.032% after calibration, ensuring sub-micron level fidelity in imaging geometric accuracy.

[0039] The matching high-frame-rate CMOS image sensor adopts a back-illuminated structure with an effective pixel count of 20.48 million and a pixel size of 2.4μm × 2.4μm. In full-resolution mode, the sensor supports a sampling frequency of 120fps, enabling it to capture the instantaneous morphological evolution during the etching process. To cope with the complex optical environment inside the etching chamber, a multi-source composite controlled light source employs a specific wavelength combination strategy. An infrared ring light source with a center wavelength of 850nm illuminates the crystal oscillator at a 55° incident angle, utilizing the penetration depth of infrared light into the quartz material to reveal the microcracks inside the crystal oscillator. The deep hole structure at the bottom of the etching is also present; the infrared ring light source consists of 8 independently controlled LED sectors, which can realize the alternating illumination of light sources from different directions; the blue coaxial light source with a center wavelength of 450nm illuminates the crystal oscillator surface vertically through a beam splitter, and the scattering effect of the short blue wavelength light on the metallization layer or the etching edge can enhance the gray-scale gradient contrast of the contour; the triggering of the light source is controlled by an FPGA-based programmable constant current drive module, which realizes nanosecond-level synchronization with the camera shutter, and the pulse width jitter is less than 5ns, suppressing the dynamic blur caused by high-speed motion or fluid flow fluctuations.

[0040] In the second step of image preprocessing, to address the plasma glow noise, etching solution microbubble interference, and high-frequency random speckle in the original digital image, this invention implements anisotropic diffusion filtering based on partial differential equations; this algorithm solves the following diffusion equation: ;in: This is a function for determining the grayscale value of an image. This refers to the diffusion time parameter; Image gradient; For divergence operators; Let be the diffusion conduction coefficient function; the diffusion conduction coefficient function is set in exponential form, i.e. ;in This is the gradient threshold constant; in actual engineering implementation, this coefficient approaches 1 in smooth areas of the image (small gradient), and omnidirectional diffusion is performed to filter out noise; in edge areas (large gradient), it approaches 0, and diffusion is stopped to maintain edge sharpness; the number of iterations is set to 12, and the time step is 0.15; after processing, the equivalent signal-to-noise ratio of the image is improved from the original 28dB to more than 45dB, and the edge expansion width is controlled within 0.5 pixels.

[0041] Furthermore, to highlight the subtle changes in micro-texture, the preprocessing module performed a five-layer image pyramid decomposition based on the Laplacian operator. During the decomposition, the high-frequency components of each layer reflect surface roughness information at different scales. By applying a 1.5x gain compensation to the high-frequency components of the third and fourth layers and performing pyramid reconstruction, the visual features of the tiny pits, scratches, and surface undulations generated during the etching process are enhanced. Through histogram specification processing, the overall contrast of the image is mapped to the standard dynamic range of [0, 255], providing a normalized data input environment for subsequent deep learning processing.

[0042] In the third step, crystal oscillator target region segmentation based on deep feature extraction was implemented. The segmentation network model adopted an improved U-Net architecture, with the encoder part integrating a residual convolution module, and the gradient vanishing problem in deep networks was solved by skip connections. In the decoder part, the spatial dimension of the image was restored layer by layer by using transposed convolution. A spatial attention mechanism was introduced at each skip connection. This mechanism extracts the weight distribution map of the spatial dimension by performing global average pooling and max pooling on the feature map, and uses the Sigmoid activation function to generate weight coefficients between [0,1], which are then multiplied with the corresponding encoder feature map, thereby guiding the network to focus on the region with high information entropy at the edge of the crystal oscillator.

[0043] The segmentation network employs a joint loss function during the training phase. ; in The cross-entropy loss function; The loss function is the Dice coefficient; weights are set in the experiment. , After deep iteration on 20,000 labeled samples, the model achieved an average intersection-union ratio of 0.965 on the validation set. Through the binarized mask output by the network, the system can accurately remove interference from the etching chamber background, support fixtures, and liquid disturbances, achieving pixel-level locking of the crystal oscillator body area.

[0044] In the fourth step of edge micro-geometric feature extraction, the system applies the improved Canny operator to perform preliminary edge localization within the mask constraint region; for the extracted edge pixels, a 7×7 neighborhood window is taken in its normal direction to perform two-dimensional Gaussian surface fitting; the parameters of the Gaussian function are solved by the least squares method, and the edge localization accuracy reaches 0.1 pixels by finding the position of the function's extreme point.

[0045] Based on this, the system calculates the edge roughness index. ; at sampling length Within, extract sub-pixel edge coordinate sequences The mean centerline was obtained by fitting a first-order linear regression. The edge roughness index is calculated using the following discretization formula: ; in The number of sampling points; through subpixel edge localization and multi-frame image averaging noise reduction technology, this invention can achieve edge roughness measurement repeatability better than 50nm.

[0046] The system synchronously extracts the key geometric dimensions of the crystal oscillator, which are obtained by calculating the minimum bounding rectangle of the edge point set within the segmented region, i.e., the sub-pixel width. Furthermore, by utilizing sub-pixel edge localization results, the measurement accuracy is improved to the 0.1 pixel level; the system extracts edge symmetry deviation. The calculation formula is: ; in: This represents the number of edge points; For the first The perpendicular distance from each edge point to the central axis; This is the average distance from all edge points to the central axis.

[0047] The system synchronously extracts edge slope features; by analyzing the gray-scale gradient profile of the edge transition zone and combining the point spread function of the telecentric lenses on both sides, the tilt angle of the etched sidewall is calculated using an optical deconvolution model; for quartz crystal oscillators, the control accuracy of the sidewall tilt angle directly affects the purity of the vibration mode. This method can achieve online measurement of the tilt angle of ±0.5°.

[0048] In the fifth step, the present invention performs feature quantization of the crystal oscillator surface morphology and etching depth; by constructing a gray-level co-occurrence matrix, at a distance step size... , angle direction Texture features are calculated under the given parameters; where the surface entropy value is... As a core indicator for evaluating etching uniformity, the calculation formula is: ; in: The grayscale values ​​of pixel pairs in the grayscale co-occurrence matrix are... and The joint probability density; experimental observations show that when etching enters the preferred orientation stripping stage of the crystal, the surface entropy value will show an upward trend, reflecting the non-uniform removal of the lattice atomic level.

[0049] This invention employs photometric stereoscopic vision technology based on multi-angle illumination to reconstruct the three-dimensional topological morphology of a crystal oscillator surface. The infrared ring light source consists of eight independently controlled LED sectors, which are sequentially illuminated by a programmable drive module to acquire image sequences of the crystal oscillator surface under illumination from four different directions. Based on the Lambertian reflection model, a reflectance map equation is established. ; in and For surface gradient; and The gradient is the direction of the light source. The surface albedo is used; the surface normal vector of each pixel is calculated using the least squares method, and then the 3D topology is reconstructed by integration; the average material removal depth is calculated by comparing the current frame height map with the initial frame reference map. ;set up This is a reference height map collected before etching. For the 3D height map reconstructed at the current moment, the coordinates are... Local etching depth at the location The calculation formula is: Instantaneous etch depth Take the etched area on the crystal oscillator surface All pixels The average value is calculated using the following formula: ; in for The total number of pixels within the range; the depth measurement value is calibrated by a laser interferometer, and the repeatability is better than 0.08μm within the range of 0-50μm.

[0050] In step six, a prediction model based on temporal feature fusion was constructed; the extracted feature vector included: edge roughness index. Surface entropy Edge symmetry deviation Subpixel size width and instantaneous etching depth These features are arranged in a time series to form a multidimensional input matrix. The matrix is ​​input to a long short-term memory network, which contains 128 hidden units. By utilizing the internal forget gate, input gate, and output gate structure, it effectively captures the nonlinear growth trend of etching depth over time and the long-range correlation between edge degradation and etchant fatigue.

[0051] The output features of the Long Short-Term Memory (LSTM) network layer are fed into a Random Forest Regressor; the Random Forest Regressor consists of 200 decision trees and outputs two key results through ensemble learning: the classification decision for the current etching stage (initial, stable, over-etching risk); and the predicted remaining etching time required to reach the target frequency. The model training employs a Dropout strategy with a dropout rate of 0.3 to prevent overfitting to specific batches of samples. The inference time of the Long Short-Term Memory network model is 7.2ms, and it is deployed on an industrial computer equipped with a graphics processing unit. The end-to-end response latency of the entire system is 18.5ms, which meets the dynamic response requirements of the etching process for real-time monitoring.

[0052] The seventh step involves implementing feedback control of the detection results; the detection system communicates in real time with the programmable logic controller of the etching machine via the Profinet industrial fieldbus; when the prediction model outputs the current etching depth... The target value (e.g., 35.00 μm) is achieved, and the edge roughness index... When the value is less than the process limit (e.g., 0.05μm), the system sends an interrupt trigger signal to the programmable logic controller. If the model determines that the etching risk period has passed, the system immediately generates an early warning signal and sends a power reduction command to the etching equipment control unit through the industrial fieldbus to implement a soft landing shutdown strategy. After receiving the signal, the programmable logic controller cuts off the plasma radio frequency power supply or opens the drain valve within 15ms to achieve instantaneous braking at the physical level.

[0053] To ensure long-term accuracy and stability, the system has an adaptive calibration function. After every 1,000 testing cycles, the robotic arm automatically delivers a standard ceramic calibration plate into the field of vision. The system identifies the circular array features on the calibration plate, updates the transformation matrix from the pixel coordinate system to the physical coordinate system by solving the homography matrix, and automatically compensates for the micron-level thermal deformation of the mechanical structure caused by ambient temperature fluctuations.

[0054] As a preferred embodiment of the present invention, for AT-cut quartz crystal oscillators, since their lattice is in shaft and Differences in chemical reactivity along the crystal axis result in anisotropic morphological features at the etched edges. This method introduces prior constraints based on lattice orientation during feature extraction; the edge contour is decomposed into components along the crystal axis, and components parallel to the crystal axis are further decomposed. When calculating the roughness index, the edge component of the shaft is given a weighting coefficient of 1.2, and the perpendicular component is given a weighting coefficient of 0.8. This weighting scheme can more accurately establish a mathematical relationship between the shape index detected by vision and the final equivalent series resistance of the crystal oscillator, thereby realizing visual monitoring based on electrical performance.

[0055] As another preferred implementation, this method integrates a cloud-based process feature database; multiple detection terminals synchronously upload the extracted feature vectors and the final detected electrical performance parameters to the cloud server; the K-means++ clustering algorithm is used to analyze massive amounts of data, which can identify minute deviations in the etching rate of specific batches of quartz wafers; the cloud system issues parameter correction instructions accordingly, dynamically adjusts the bias terms of the long short-term memory network prediction model of each device, and realizes intelligent process consistency control of the group.

[0056] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further described below in conjunction with specific embodiments and comparative examples.

[0057] Example 1: The machine vision-based detection method described in this invention is used. The hardware configuration is as described above. The algorithm is deployed on an industrial computer equipped with a graphics processing unit to execute the entire process.

[0058] Comparative Example 1: The etching progress was estimated by measuring the frequency drift using a traditional detection method based on the electrical performance feedback of a quartz crystal microbalance.

[0059] Comparative Example 2: Using conventional machine vision methods, only monochrome LED ring light and simple Canny edge detection and threshold segmentation are used, without deep learning and temporal prediction models, and the system is deployed on a standard industrial computer.

[0060] The experiment monitored the wet etching of a 32.768 kHz miniature quartz crystal oscillator, with a target etching depth of 30.00 μm and a production batch size of 10,000 units. The experimental data are shown in Table 1 below: Table 1: Comparison of Key Indicators between the Embodiments of the Invention and the Comparative Examples detection indicators Example 1 Comparative Example 1 Comparative Example 2 Depth control accuracy (μm) ±0.12 ±0.45 ±0.85 Edge roughness measurement repeatability (nm) Better than 50 Unable to measure 120.6 Etching status determination end-to-end delay (ms) 18.5 450 35 Over-etching false positive rate (%) 0.08 1.55 2.1 Batch production yield (%) 99.82 97.2 95.45 Minimum identifiable defect size (μm) 0.5 Unrecognized 3.5 System continuous operation test duration (h) More than 2000 hours 1200 1800 Table 1 shows that this invention achieves a breakthrough in the detection principle. In terms of hardware architecture, the dual-sided telecentric lens and multispectral composite illumination eliminate parallax and stray light interference, improving the depth control accuracy to ±0.12μm. In terms of algorithm, sub-pixel edge localization and photometric stereo vision compress the repeatability of edge roughness measurement to within 50nm, with the smallest identifiable defect size reaching 0.5μm. The temporal prediction model reduces the feedback delay to 18.5ms, which is two orders of magnitude shorter than the electrical feedback method, and the over-etching false positive rate is reduced to 0.08%. The above technical advantages together support a mass production yield of 99.82%. The core of this invention lies in improving the detection dimension from one-dimensional electrical signals or spectral envelopes to two-dimensional geometric shapes and three-dimensional topological structures, shifting from result-oriented reverse calculations to process-oriented predictive control, and expanding from single visual features to multi-modal fusion, thus solving the technical problem of missing microscopic shape perception.

[0061] Example 2: Enhanced detection method based on dynamic multispectral synchronous illumination and multimodal feature fusion: This example further optimizes the optical acquisition system and time-series prediction model based on Example 1 to meet the higher requirements for morphological accuracy and defect identification in the manufacturing process of ultra-high frequency crystal oscillators (frequency greater than 100MHz).

[0062] In the first step of this embodiment, a green ring light source with a center wavelength of 525nm is added to the infrared ring light source and the blue coaxial light source to form a three-wavelength composite illumination system. The green ring light source consists of eight independently controlled LED sectors with an incident angle of 30° to 45°. The programmable constant current drive module is configured to execute a dynamic multispectral synchronous illumination strategy: within one image acquisition cycle, the infrared ring light source, the green ring light source, and the blue coaxial light source are lit sequentially according to a preset timing sequence, with each lighting duration controlled within 50μs and the interval between adjacent light sources being 100μs. The high frame rate CMOS image sensor synchronously acquires three original digital images corresponding to different spectral illumination conditions at a sampling frequency of 300fps, which are recorded as infrared images. Green image and blue image The eight LED sectors of the infrared ring light source are divided into four orthogonal direction groups, each containing two oppositely arranged sectors. By sequentially illuminating different direction groups, infrared oblique illumination in four orthogonal directions is achieved, which is used to enhance the three-dimensional morphological features of microcracks and the bottom of deep holes.

[0063] In the fifth step of this embodiment, the feature quantization of the crystal oscillator surface morphology and etching depth is based on the fusion of three-spectral images; for infrared images... Green image and blue image Perform gray-level co-occurrence matrix analysis separately to extract the surface entropy values ​​under their respective spectral conditions. , and Because different wavelengths of light have different penetration depths and scattering characteristics into the microstructure of quartz crystal surfaces, the surface entropy value... It reflects deep-level (approximately 5μm to 10μm) lattice defects and subsurface damage. Reflecting the etching uniformity of the intermediate layers (approximately 2μm to 5μm), It reflects the microscopic roughness of the surface layer (approximately 0.5 μm to 2 μm); the three spectral entropy values ​​are fused according to weighting coefficients to obtain the comprehensive surface quality index. ; in , , The preset weighting coefficients are set to 0.3, 0.4, and 0.3 respectively.

[0064] Based on four sets of infrared oblique illumination images from orthogonal directions, a high-precision three-dimensional topological morphology of the crystal oscillator surface is reconstructed using photometric stereo vision technology. The infrared images acquired under illumination from the four directions are input into an improved photometric stereo vision algorithm, which is based on the Lambertian reflection model and obtains the surface normal vector of each pixel by solving an overdetermined set of equations. Then, integral reconstruction of the three-dimensional height map. Based on this, the etching depth is calculated. Root mean square of surface slope and the volume of the etching pit ; Etching pit volume It is obtained by integrating the local concave regions below the reference plane in the height map, and the calculation formula is as follows: ; in It is a concave area; The height of the reference plane; This is the reconstructed 3D height map.

[0065] In the sixth step of this embodiment, the temporal feature fusion model introduces a multimodal feature fusion layer based on the Long Short-Term Memory network and the Random Forest regressor; the multimodal feature fusion layer incorporates surface entropy values ​​from the three spectra. , , Three-dimensional morphological features (etching depth) Root mean square of surface slope , Etching pit volume ) and edge geometric features (edge ​​roughness index) Edge symmetry deviation Subpixel size width The features are concatenated to form an 8-dimensional multimodal feature vector. This feature vector is then input into a Long Short-Term Memory (LSTM) network layer, which contains 256 hidden units to capture the co-evolution of multimodal features over time. The output features of the LSM layer are fed into a random forest regressor composed of 300 decision trees, which outputs the classification probability of the current etching state and the predicted remaining etching time. .

[0066] This embodiment achieves multi-dimensional morphology perception from the surface to the deep layer by using dynamic multispectral synchronous illumination and multimodal feature fusion. It can predict the arrival of the etching endpoint of the ultra-high frequency crystal oscillator 6 to 8 seconds in advance and reduce the over-etching false judgment rate to below 0.05%. It is suitable for precision etching control of ultra-thin crystal oscillators with a thickness of less than 15μm.

[0067] Example 3: Timing prediction enhancement method based on anisotropic etching kinetics compensation: Based on Example 1, this example performs physical mechanism-driven deep optimization on the timing prediction model to address the anisotropic characteristics of AT-cut and SC-cut quartz crystals during wet etching. The prior constraints of lattice orientation are upgraded from static weights to dynamic timing compensation, further improving the accuracy and physical interpretability of the prediction model.

[0068] In the fourth step of this embodiment, the edge roughness index The calculation introduces lattice orientation decomposition; for AT-cut quartz crystal oscillators, based on the anisotropic etching characteristics of quartz crystals, the edge contour is decomposed into components along the X-axis and along the Z'-axis; for each sub-pixel edge point Calculate the angle between its tangent direction and the positive X-axis direction. Edge points are classified into the dominant edge point set along the X-axis by directional filtering. and the dominant edge point set along the Z' axis Calculate the edge roughness indices in both directions. and The calculation formula is as follows: ; ; in: and They are respectively and The number of edge points in the data; and These are the fitting line parameters for the edge point set in the corresponding direction; and the edge roughness index. Further expressed as a direction-weighted combination: ; in and The direction weighting coefficient is set according to the anisotropy ratio of the etching rate of the AT-cut quartz crystal oscillator. .

[0069] In the sixth step of this embodiment, the hybrid model is a long short-term memory network model based on anisotropic etching kinetics compensation. This model adds a physical information fusion layer to the standard long short-term memory network structure. The physical information fusion layer receives the crystal oscillator's cutting parameters and the real-time temperature of the etching solution. ,concentration and pressure As an external input, the external input is mapped to an etch rate compensation factor through a fully connected network. The calculation formula is: ; in: and These are trainable weight matrices and bias vectors; The sigmoid activation function is used to constrain the compensation factor within the range of 0.5 to 1.5; the etch rate compensation factor... The cell state update formula injected into the Long Short-Term Memory (LSTM) network is modified from the standard LSM network's cell state update equation as follows: ; ; in: Candidate cell state; This represents the updated cell state. This represents the cell state at the previous moment; Output for the forget gate; For input gate output; This is the hidden state from the previous moment; The multimodal feature vector input at the current time step; and The weight matrix and bias vector represent the candidate cell states. This indicates element-wise multiplication.

[0070] Etching rate compensation factor The dynamic adjustment of the long short-term memory network model based on anisotropic etching kinetics compensation allows it to automatically adjust the etching rate prediction according to real-time process parameters, compensating for etching rate shifts caused by solution aging, temperature fluctuations, and lattice anisotropy. The output features of the long short-term memory network model based on anisotropic etching kinetics compensation are fed into a random forest regressor composed of 200 decision trees, outputting the classification probability of the current etching state and the predicted remaining etching time. .

[0071] This embodiment incorporates the anisotropic etching dynamics of quartz crystals into the cell state update process of long short-term memory networks, making the prediction model physically interpretable. In the wet etching process of AT-cut crystal oscillators, the prediction error of the remaining etching time is reduced from ±0.5s to ±0.2s, and the over-etching misjudgment rate is reduced from 0.08% to 0.03%, thereby improving the manufacturing yield and frequency consistency of high-frequency crystal oscillators.

[0072] Example 4: Etching State Prediction Method Based on Physically Augmented Long Short-Term Memory Network: Based on Example 1, this example constructs a physically augmented long short-term memory network model to address the impact of etching solution temperature fluctuations, concentration decay, and pressure changes on the etching rate during wet etching. This model deeply integrates process parameters with visual features to achieve higher accuracy in etching state prediction.

[0073] The physical information-enhanced long short-term memory network model constructed in this embodiment consists of three parts: a physical information fusion layer, a long short-term memory network layer, and a random forest regressor layer. The physical information fusion layer receives the real-time temperature of the etching solution. ,concentration and pressure As an external input, the external input is mapped to an etch rate compensation factor through a two-layer fully connected network. The calculation process for this layer is as follows: External input vector The input to the first fully connected layer is processed by a linear transformation and the ReLU activation function: ; in: This is the weight matrix of the first fully connected layer, with dimension . ; This is the bias vector of the first fully connected layer, with a dimension of 32; It is a linear rectification activation function.

[0074] The output of the first layer The input to the second fully connected network, after linear transformation and Sigmoid activation function processing, yields the etch rate compensation factor. : ; in: This is the weight matrix of the second fully connected layer, with dimension . ; This is the bias vector of the second fully connected network, with a dimension of 1; It is the Sigmoid activation function; this calculation formula ensures The value range is from 0.5 to 1.5, where This indicates that the etching rate is higher than the baseline rate under the current process conditions, and it is necessary to enhance the contribution of the current input information. This indicates that the etching rate is lower than the baseline rate, and the contribution of historical information needs to be enhanced.

[0075] The Long Short-Term Memory (LSTM) network layer receives the visual feature vector at the current time step. and the hidden state of the previous moment The etch rate compensation factor output by the physical information fusion layer is combined with the physical information fusion layer. The state is updated according to the modified cell state update equation.

[0076] The formulas for calculating the forget gate, input gate, and output gate of a standard Long Short-Term Memory (LSTM) network are as follows: Forget Gate: Input gate: Output gate: Candidate cell status: ;in: , , , This is the weight matrix; , , , It is the bias vector; This means concatenating the hidden state and the input vector; Use the Sigmoid activation function; It is the hyperbolic tangent activation function.

[0077] This embodiment modifies the cell state update equation by introducing an etching rate compensation factor. : ; in: The baseline cell state at the previous time step is the standard update value excluding compensating factors, calculated using the following formula: Initial time At that time, let the initial value of the cell state be... , initial value of basic cell state The modified update equation passed Dynamically adjust the fusion weight of historical and current information: when At that time, the weight of the current input information increases, and the model becomes more sensitive to changes in the process. At that time, the weight of historical information increases, and the model becomes more robust to noise interference.

[0078] Updated hidden status Calculated from the output gate and cell state: .

[0079] Hidden state output by Long Short-Term Memory network layer The data is fed into a random forest regressor layer; this layer consists of 200 decision trees, each performing an independent regression prediction on the input features, and the final output is the average of all the predictions from the decision trees; the random forest regressor layer outputs two key results: the classification probability of the current etched state. (Initial etching period, stable etching period, over-etching risk period) and predicted remaining etching time .

[0080] The physical information-enhanced long short-term memory network model in this embodiment uses a joint loss function for end-to-end training.

[0081] Collect production data from N batches of historical production data. Each batch contains complete time-series data from the start to the end of etching. For each batch, extract T sampling points in chronological order. Each sampling point contains: Visual feature vectors ; in For multispectral fusion feature vectors (containing , , , , ), For edge geometric feature vectors (containing , , ); External input vector ; Etching depth label (Obtained through offline measurement); Remaining etching time stamp (Calculated based on the actual etching completion time).

[0082] Model training uses a joint loss function Simultaneously optimize two tasks: etch depth prediction and remaining time prediction. ; Among them: Etching depth prediction loss ; The predicted etching depth for the t-th sampling point in the i-th batch; The corresponding true etching depth; remaining time predicted loss. ; For the predicted remaining time of the i-th batch and the t-th sampling point, This corresponds to the actual remaining time; weight decay regularization term. Used to prevent overfitting; , and The preset weighting coefficients are 0.5, 0.5, and 0.01, respectively.

[0083] The model was trained using the Adam optimizer with an initial learning rate of 0.001, which was reduced to 0.9 times every 50 epochs. The training data was divided into training and validation sets in an 8:2 ratio. During training, the model performance was evaluated on the validation set every 10 epochs, and the model parameters with the smallest loss on the validation set were retained.

[0084] During real-time detection, the system applies the trained model according to the following steps: 1. At each sampling time t, acquire the current frame image and extract the visual feature vector according to steps one to five of Example 1. 2. The temperature of the etching solution is read by the sensors of the etching equipment. ,concentration and pressure , constitute the external input vector 3. and 4. Input a pre-trained long short-term memory network model based on physical information augmentation; 5. Calculate the classification probability of the current etched state using the forward pass of the model. and the predicted remaining etching time 5. Output the classification results and remaining time to the closed-loop control system to generate an etching completion trigger signal or a power reduction command.

[0085] To verify the technical effectiveness of this embodiment, a comparative experiment was conducted. The experiment was performed on the same etching equipment, using the same crystal oscillator sample (AT cut, frequency 32.768kHz), comparing the three methods: Example 1: A hybrid model using a basic long short-term memory network and a random forest regressor is adopted, using only visual features.

[0086] Example 3: The anisotropic etching kinetics compensation method of Example 3 is adopted.

[0087] Example 4: The Long Short-Term Memory network model based on physical information enhancement in this example is adopted.

[0088] During the experiment, temperature fluctuations in the etching solution were artificially introduced (ranging from 45℃ to 55℃) using an external heating device to simulate process fluctuations in actual production. The experimental results are shown in Table 2 below: Table 2: Performance Comparison of Different Prediction Methods under Process Fluctuation Conditions detection indicators Example 1 Example 3 Example 4 (Invention) Root mean square error of remaining time prediction (s) 1.2 0.8 0.4 Over-etching false positive rate (%) 0.12 0.08 0.04 Etching depth control accuracy (μm) ±0.15 ±0.12 ±0.08 Response delay (s) to temperature fluctuations 8.5 5.2 2.1 Table 2 data reveals the predictive advantages of this invention under process fluctuation conditions. Example 4, based on Example 1, introduces a physical information fusion layer, reducing the root mean square error of the remaining time prediction from 1.2s to 0.4s, and the over-etching misjudgment rate from 0.12% to 0.04%. The technical essence lies in: by injecting process parameters such as etching solution temperature, concentration, and pressure as external inputs into the long short-term memory network, constructing an etching rate compensation factor to dynamically adjust the cell state update equation, achieving deep fusion of visual features and process parameters; the response delay to temperature fluctuations is shortened from 8.5s to 2.1s, indicating that the model has the ability to perceive process disturbances in real time and adaptively adjust the prediction strategy; this invention embeds the anisotropic etching dynamics mechanism into a deep learning architecture, enabling the prediction model to evolve from purely data-driven to physically-informed enhanced driving, solving the technical problems of low prediction accuracy and large response delay under unsteady conditions.

[0089] The deployment logic of this invention in an actual production line is further described in detail: The detection unit is installed directly above the etching reaction tank and fixed by a stainless steel vacuum flange. The observation window material is high-purity synthetic quartz with a transmittance of more than 98% in the 400nm-900nm wavelength band. After each crystal oscillator enters the field of view, the system performs global coarse positioning, identifies the reference mark of the crystal oscillator, and calculates the rotation and translation matrix. The visual algorithm then enters the local fine analysis stage.

[0090] To handle the translucent properties of quartz crystals under strong light when implementing pixel-level segmentation based on deep learning, the segmentation network adds an edge gradient weight term to the loss function. ;in For predicting masks; The term is used as a truth mask; this forces the network to produce a steeper response at edges where the gradient changes drastically, solving the problem of the difficulty in defining the boundary between the quartz edge and the acid etching solution.

[0091] When performing the fourth step of sub-pixel extraction, the system adopts an adaptive gray-scale moment algorithm. This algorithm uses the first three gray-scale moments in the edge neighborhood to calculate the precise position and orientation of the edge. For incomplete edges that are obscured by etchant bubbles, the system uses the trajectory of the time-series historical frames for Kalman filter prediction compensation to ensure that contour tracking remains continuous even under instantaneous occlusion, thus avoiding abnormal jumps in the control signal.

[0092] The fifth step, 3D reconstruction, employs photometric stereo vision technology based on multi-angle illumination. By controlling the sequential lighting of eight independent LED sectors in the ring light source, image sequences of the crystal oscillator surface under illumination from four different directions are obtained. Based on the Lambertian reflection model, the surface normal vector of each pixel is calculated. Compared to the traditional single-light source ShapefromShading technology, it has a higher tolerance for uneven surface albedo and can more realistically depict the geometric depth of micro-corrosion pits that appear on the crystal surface after etching.

[0093] In the sixth step of classification prediction, the Long Short-Term Memory (LSTM) network model not only considers the current image features, but also introduces the real-time temperature, concentration, and pressure count values ​​of the etching solution as external variables. These variables are injected into the hidden state of the LSM network in real time through a serial interface. Through this cross-modal data fusion, the prediction model can compensate for the etch rate decay caused by the aging of the etching solution, reducing the prediction error of the remaining time to within ±0.5s.

[0094] In the seventh step, the closed-loop control command generation stage, the system employs a decision-making mechanism based on fuzzy logic to predict the remaining time. When the edge roughness approaches zero, the system does not simply send a shutdown signal, but dynamically fine-tunes the plasma power according to the current edge roughness trend. For example, if accelerated edge degradation is detected, the system will reduce the power in advance to slow down the reaction speed, sacrificing a small amount of efficiency to obtain the ultimate morphological quality. This soft-landing process control strategy is the key to the invention's ability to maintain extremely high consistency in mass production.

[0095] For the manufacturing of high-frequency crystal oscillators (frequency greater than 100MHz), since the wafer thickness is usually less than 15μm, it is prone to breakage or central perforation in the later stages of etching. This invention adds a local stress anomaly early warning function to the real-time monitoring; by analyzing the distribution changes of birefringence interference fringes in the image, the stress concentration inside the wafer is evaluated in real time; once the stress gradient is found to exceed the critical threshold, the system immediately stops etching and prompts for thermal annealing, thus improving the processing safety of ultra-thin crystal oscillators.

[0096] The method of this invention employs a high-precision telecentric optical system at the physical layer, integrates nonlinear filtering, deep learning segmentation, sub-pixel geometric quantization, and long short-term memory network timing prediction at the algorithm layer, and implements high-bandwidth, low-latency closed-loop feedback at the control layer to construct a complete intelligent monitoring system for crystal oscillator etching with autonomous learning capabilities. All technical indicators can meet the manufacturing requirements of high-end crystal oscillators and are suitable for the manufacturing needs of high-performance crystal oscillators in the fields of 5G communication, aerospace, and precision measurement.

[0097] In subsequent industrial applications, the method of this invention can be further extended to the micro-etching detection of other semiconductor materials (such as silicon carbide and gallium nitride); the open algorithm architecture allows users to quickly adjust the anisotropic diffusion coefficient or train new segmentation weights according to different material properties; this high degree of flexibility and scalability makes this invention have broad application prospects in the field of modern micro-nano fabrication.

[0098] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0099] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting the etching state of a crystal oscillator based on machine vision, characterized in that, Includes the following steps: Step 1: Construct a high-resolution microscopic vision acquisition system and perform raw image acquisition; the high-resolution microscopic vision acquisition system includes an optical unit consisting of a dual telecentric lens, a high frame rate CMOS image sensor, and a multi-source composite controlled light source; in the etching reaction chamber, the optical unit images the crystal oscillator in the etching environment through a quartz observation window to acquire raw digital image signals; Step 2: The original digital image signal is input to the preprocessing module, where anisotropic diffusion filtering is performed to remove noise and maintain edge gradient sharpness. Histogram specification and image pyramid decomposition based on the Laplacian operator are then performed to enhance the signal-to-noise ratio of the etched features. Step 3: The preprocessed image is fed into a pre-trained pixel-level segmentation network model, which outputs a binary mask of the same size as the input image. The target region of the crystal oscillator is accurately extracted by multiplying the mask with the original image pixel by pixel. Step 4: Perform sub-pixel level edge positioning in the segmented edge region of the crystal oscillator body, and calculate the edge roughness index, sub-pixel size width, and edge symmetry deviation. Step 5: Perform feature quantization of the crystal oscillator surface morphology and etching depth. Extract the surface entropy value by constructing a gray-level co-occurrence matrix, and reconstruct the three-dimensional topological morphology of the crystal oscillator surface using photometric stereo vision technology to calculate the etching depth. Step 6: Arrange the extracted edge roughness index, surface entropy value, edge symmetry deviation, sub-pixel size width and etching depth according to the sampling time axis to form a time series dataset, and input it into a hybrid model based on long short-term memory network and random forest regressor to output the classification probability of the current etching state and the predicted remaining etching time. Step 7: Compare the etching status determination result with the preset process threshold matrix in real time. When the etching depth reaches the design target value and the edge roughness index and geometric dimension deviation are within the tolerance range, an etching completion trigger signal is generated. When it is determined that the etching risk period has passed, an early warning signal is generated and a power reduction command is sent.

2. The crystal oscillator etching state detection method based on machine vision according to claim 1, characterized in that, In step one, the magnification of the dual telecentric lenses is set to 2.0X to 5.0X, and the distortion rates on both the object side and the image side of the dual telecentric lenses are less than 0.05%; the high frame rate CMOS image sensor has an effective resolution of no less than 20 million pixels, a pixel size of 2.4μm×2.4μm, and supports an image sampling frequency of no less than 100fps; the multi-source composite controlled light source consists of an infrared ring light source with a center wavelength of 850nm and a blue coaxial light source with a center wavelength of 450nm, and the infrared ring light source is composed of multiple independently controlled LED sectors.

3. The crystal oscillator etching state detection method based on machine vision according to claim 1, characterized in that, In step two, the anisotropic diffusion filtering is achieved by solving a partial differential equation, which is expressed as: ； in: Indicates spatial location and time Image grayscale value function at location; This refers to the diffusion time parameter; Image gradient; Let be the magnitude of the image gradient; It is a divergence operator; Let be the diffusion conduction coefficient function, which has the following form: ; This is the gradient threshold constant.

4. The crystal oscillator etching state detection method based on machine vision according to claim 1, characterized in that, In step three, the pixel-level segmentation network model adopts an encoder-decoder architecture. During the encoding stage, residual convolution modules are used to extract multi-scale morphological features. During the decoding stage, transposed convolutions are used to restore spatial resolution, and a spatial attention mechanism is introduced at skip connections. The pixel-level segmentation network model employs a joint loss function during the training stage. ; in The cross-entropy loss function; The loss function is the Dice coefficient. and These are the weighting coefficients.

5. The crystal oscillator etching state detection method based on machine vision according to claim 1, characterized in that, In step four, the sub-pixel-level edge localization is achieved by fitting a Gaussian surface to the grayscale distribution along the edge normal direction and using an interpolation algorithm; the edge roughness index The calculation formula is: ; in: This represents the number of edge sampling points; For the first The coordinates of sub-pixel edge points; and These are the slope and intercept obtained by linear fitting of the edge point set, respectively; The fitting average line of the edge contour; the edge symmetry deviation The calculation formula is: ;in: This represents the number of edge points; For the first The perpendicular distance from each edge point to the central axis; This is the average distance from all edge points to the central axis.

6. The crystal oscillator etching state detection method based on machine vision according to claim 1, characterized in that, In step five, the calculation distance step size of the gray-level co-occurrence matrix is... The value ranges from 1 to 3 pixels, and the angular direction includes 0°, 45°, 90°, and 135°; the surface entropy value The calculation formula is: ； in: The grayscale values ​​of pixel pairs in the grayscale co-occurrence matrix are... and The joint probability density; the photometric stereo vision technology obtains image sequences of the crystal oscillator surface under illumination from multiple different directions by controlling multiple independently controlled LED sectors in an infrared ring light source to light up sequentially, and calculates the surface normal vector of each pixel according to the Lambertian reflection model, and then integrates to reconstruct the three-dimensional topological shape.

7. The crystal oscillator etching state detection method based on machine vision according to claim 1, characterized in that, In step six, the Long Short-Term Memory (LSTM) network layer contains 128 hidden units and utilizes forget gates, input gates, and output gates to capture the non-linear growth trend of the etching depth over time. The random forest regressor consists of 200 decision trees and outputs the classification probability of the current etching state and the predicted remaining etching time through ensemble learning. .

8. The crystal oscillator etching state detection method based on machine vision according to claim 1, characterized in that, In step seven, the etching completion trigger signal is transmitted in real time to the control unit of the etching equipment via the industrial fieldbus to execute the instruction to shut down the etching source; the power reduction instruction is transmitted to the control unit of the etching equipment via the industrial fieldbus to implement a soft landing shutdown strategy.

9. The crystal oscillator etching state detection method based on machine vision according to claim 1, characterized in that, In step one, the multi-source composite controlled light source also includes a green ring light source with a center wavelength of 525nm, which is composed of multiple independently controlled LED sectors; the programmable constant current drive module is configured to execute a dynamic multispectral synchronous illumination strategy, sequentially lighting up the infrared ring light source, the green ring light source, and the blue coaxial light source according to a preset timing sequence; the high frame rate CMOS image sensor synchronously acquires multiple original digital images corresponding to different spectral illumination conditions.

10. The crystal oscillator etching state detection method based on machine vision according to claim 9, characterized in that, In step five, the surface entropy value includes the infrared spectral surface entropy value. Green spectral surface entropy and blue spectrum surface entropy value The comprehensive surface quality index is obtained by fusing the three spectral entropy values ​​according to weighting coefficients. ; in , , Preset weighting coefficients; The three-dimensional topological features are reconstructed using photometric stereo vision technology based on four sets of infrared oblique illumination images in orthogonal directions, and the volume of the etched pits is calculated. The volume of the etched pit The calculation formula is: ； in It is a concave area; The height of the reference plane; This is the reconstructed 3D height map.

Images