Monitoring method and system based on abnormal cutting of quartz crystal

Abstract

The invention discloses a monitoring method and system based on quartz crystal cutting abnormity, and relates to the technical field of abnormity monitoring, and the method comprises the steps: capturing a surface image and a reflection spectrum of each section in a cutting process through cooperation of a high-speed industrial camera and a laser scanning spectrometer; preprocessing the surface image and the reflection spectrum, and then carrying out spatial registration; performing feature fusion on the image after spatial registration and the cutting path of the surface to obtain a clear input data stream; pyramid reconstruction is carried out on the data stream corresponding to each tangent plane; and performing inherent parameter simplification on the reconstructed section data stream, and analyzing the simplified section data stream to obtain a detection result of the section abnormality of the quartz crystal. The automation and intelligence level of anomaly detection in the quartz crystal cutting process can be remarkably improved, and the technical bottleneck that a traditional method is weak in small defect and low-contrast anomaly recognition capacity is broken through.

Description

Technical Field

[0001] The present invention relates to the technical field of anomaly monitoring, and particularly to a monitoring method and system for quartz crystal cutting anomalies. Background Art

[0002] Due to its excellent physical properties and chemical stability, quartz crystals are widely used in high-end manufacturing fields such as electronic components, precision instruments, and optical materials. In these application scenarios, quartz crystals usually need to undergo strict cutting, grinding, and forming processes to obtain a device structure that meets performance requirements. However, the cutting process of quartz crystals is extremely complex, limited by the anisotropy and brittleness characteristics of the material itself, as well as multiple factors such as the processing environment, tool state, and process parameters. Since quartz materials are prone to defects such as microcracks, chipping, and surface scratches, any minor processing anomaly may have a significant impact on the electrical, mechanical, and reliability properties of subsequent products, or even lead to the scrapping of an entire batch of components. Therefore, improving the monitoring and quality control capabilities of the cutting process has become a key link in the quartz device industrial chain.

[0003] Traditional quartz cutting quality control mostly relies on manual visual inspection or simple optical imaging methods, making it difficult to achieve high-sensitivity detection of minor defects and early anomalies. With the increasing requirements for product consistency and yield in high-end manufacturing, more and more enterprises and research institutions have begun to introduce intelligent sensing and automated detection means. For example, using high-speed industrial cameras to obtain cutting surface images and combining laser spectroscopy technology to analyze changes in the material surface structure has become a development trend in quality monitoring in recent years. At the same time, data-driven analysis methods such as machine vision and artificial intelligence have shown great potential in tasks such as anomaly recognition, defect grading, and process traceability in complex process flows. Nevertheless, how to further improve the automated recognition level of cutting anomalies, accurately separate material body defects from processing process defects, and provide traceable data support for subsequent process optimization remains an important technical challenge in the current quartz crystal processing field. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is that the existing monitoring methods for quartz crystal cutting anomalies have insufficient ability to identify minor defects, are difficult to distinguish process anomalies from material background anomalies, and have limited automation level.

[0006] To solve the above technical problems, the present invention provides the following technical solution: A monitoring method for quartz crystal cutting anomalies, comprising: Cooperating a high-speed industrial camera and a laser scanning spectrometer to capture the surface image and reflection spectrum of each cut surface during the cutting process; Preprocess the surface image and the reflection spectrum, and then perform spatial registration; fuse the features of the spatially registered image with the cutting path of the surface to obtain a clear input data stream; Perform pyramid reconstruction on the data stream corresponding to each cut surface; Simplify the inherent parameters of the reconstructed cut surface data stream, analyze the simplified cut surface data stream, and obtain the detection result of the quartz crystal cut surface anomaly.

[0007] As a preferred solution of the monitoring method based on quartz crystal cutting anomaly of the present invention, wherein: the preprocessing includes: for the surface image, performing noise suppression, geometric distortion correction of the cut surface image, brightness normalization, and artifact removal; for the reflection spectrum, performing spectral normalization and interference band removal.

[0008] As a preferred solution of the monitoring method based on quartz crystal cutting anomaly of the present invention, wherein: the spatial registration includes: uniformly mapping the preprocessed data into the standard coordinate system of the physical space, making the data stream correspond one by one in spatial position, and realizing pixel-level alignment.

[0009] As a preferred solution of the monitoring method based on quartz crystal cutting anomaly of the present invention, wherein: the cutting path includes: taking the fixed position of the tool tip as the reference point, and taking the tool tip lines extending from the reference point to both sides as the parameter reflection baseline during cutting, so that the feedback feature vectors at the coordinate positions on the same parameter reflection baseline are equal to the feedback feature vector of the reference point; During cutting, record the continuous path of the reference point in the two-dimensional coordinate system where the image is located through the initial position, cutting speed vector and time of the reference point, and at the same time generate the feedback feature vectors on the same parameter reflection baseline: ; wherein, represents the two-dimensional pixel coordinates in the spatially registered image, , represents the path of the reference point in the image; represents the feedback feature vector of the pixel point on the image; the cutting speed of the reference point at time t; represents the acceleration of the reference point at time t; represents the normal force of the tool tip at time t; represents the tangential force of the tool tip at time t; represents the cutting trajectory error, defined as: the Euclidean distance deviation between the actual tool tip trajectory and the control instruction; t represents the t-th moment during cutting; For each pixel point in each parameter reflection baseline, introduce a virtual dimension to represent the actual feedback feature vector of each pixel point on the parameter reflection baseline; on the virtual dimension, make random assumptions about the actual feedback feature vectors of all pixel points on both sides. The constraint of the random assumption is that the sum of the actual feedback feature vectors of all pixel points on the parameter reflection baseline is equal to the feedback feature vector of the parameter reflection baseline. Solve the fitness for the whole image, and select the combination of random assumptions with fitness higher than the preset value as the evaluation result; obtain the set of global random assumption combinations ; An represents the nth random assumption combination. According to the fitness corresponding to each element in the set A, it is proportionally converted into a probability so that the sum of the probabilities of all elements in the set is 1. The fitness is equal to the sum of the prior probability based on historical data and the rationality analysis result of the one-dimensional simplified convolutional neural network. The feature fusion includes: mapping the combination of random assumptions corresponding to each element in the set A into the spatially registered image respectively according to the corresponding relationship of pixel points, and obtaining n groups of images with increased parameter dimensions.

[0010] As a preferred scheme of the monitoring method based on quartz crystal cutting anomaly of the present invention, wherein: the pyramid reconstruction includes: for the n groups of images with increased parameter dimensions, respectively perform: multi-scale pyramid construction, perform feature enhancement and contrast learning on each layer, and through the extraction of positive samples and negative samples, amplify the loss of weak anomalies; Enhance the sensitivity of the anomaly signal by comparing the feature representations of different scales and different regions; Finally, perform upsampling and multi-layer fusion reconstruction to obtain the reconstructed image, which is used as a data stream for anomaly detection.

[0011] As a preferred scheme of the monitoring method based on quartz crystal cutting anomaly of the present invention, wherein: the inherent parameters include: the anomaly parameters that have been identified with abnormal manifestations in the data stream after pyramid reconstruction of the standard cutting surface under the same material; The simplification process of the inherent parameters includes: comparing the features of the inherent parameters in the obtained n groups of data streams, extracting the data components most similar to the inherent parameters from the n groups of data streams, screening m data streams with the content of similar data components higher than the preset value as the data streams to be analyzed; in the data streams to be analyzed, the part identical to the anomaly parameters is always regarded as the normal parameter performance, and the anomaly part caused only by cutting anomaly is obtained; where m ≤ n.

[0012] As a preferred solution of the monitoring method based on quartz crystal cutting anomalies according to the present invention, wherein: the analysis of the simplified section data stream specifically includes: Step 1: In the m groups of data streams to be analyzed, perform clustering of data features. If there is only one clustering cluster, go to Step 3; if there are more than one clustering clusters, go to Step 2; Step 2: If the ratio of the number of individuals in the cluster with the most individuals to the number of individuals in the cluster with the second most individuals is greater than the preset value, only retain the data stream corresponding to the cluster with the most individuals and go to Step 3; otherwise, go to Step 6; Step 3: According to the data streams of each individual in the same clustering cluster, divide the abnormal part and the normal part in the image, and blur the boundary of the division so that the features of each individual can be expressed in the blurred divided image; Step 4: Lock the abnormal core of the m blurred divided images and determine the inner and outer boundaries of the blurred division; Step 5: Use a pre-trained neural network to analyze the pixel features in the inner boundary to obtain the recognition result of the anomaly; Step 6: Calculate the number of individuals in each clustering cluster, and use the number of individuals in the cluster / total number of individuals × 100% as the probability coefficient; respectively enter Step 3 with the data stream corresponding to each cluster, and at the same time add the corresponding probability coefficient during output as an additional element of the evaluation result.

[0013] A monitoring system for quartz crystal cutting anomalies using the method as described in the present invention includes: a collection unit that cooperates with a high-speed industrial camera and a laser scanning spectrometer to capture the surface image and reflection spectrum of each section during the cutting process; A processing unit that preprocesses the surface image and the reflection spectrum and then performs spatial registration to obtain a registered image; fuses the features of the spatially registered image with the cutting path on the surface to obtain a clear input data stream; An optimization unit that performs pyramid reconstruction on the data stream corresponding to each section; A detection unit that simplifies the inherent parameters of the reconstructed section data stream and analyzes the simplified section data stream to obtain the detection result of the quartz crystal section anomaly.

[0014] A computer device includes: a memory and a processor; the memory stores a computer program, wherein: when the processor executes the computer program, it implements the steps of the method described in any one of the present invention.

[0015] A computer-readable storage medium stores a computer program, wherein: when the computer program is executed by a processor, it implements the steps of the method described in any one of the present invention.

[0016] Advantages of the present invention: The monitoring method based on quartz crystal cutting anomalies provided by the present invention can significantly improve the automation and intelligence level of anomaly detection during the quartz crystal cutting process, overcoming the technical bottleneck of the weak recognition ability of traditional methods for minute defects and low-contrast anomalies. By fusing multi-modal data, pyramid multi-scale analysis, and inherent parameter assimilation, precise extraction of multi-granularity and multi-type anomalies on the cutting surface is achieved, and it can effectively distinguish process-induced anomalies from material background defects, thereby significantly reducing the risks of misjudgment and missed detection. This method has high-throughput and full-process online detection capabilities, and can support real-time monitoring and quality traceability in large-scale production environments. The proposed technical framework has strong compatibility and is convenient for integration with existing manufacturing and detection equipment, which helps to improve the ex-factory yield, stability, and process controllability of quartz components, providing strong data and decision-making support for the optimization of quartz processing technologies in high-end electronics, optics, and other fields. Description of the Drawings

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for description in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0018] Figure 1 It is the overall flowchart of a monitoring method based on quartz crystal cutting anomalies provided for the first embodiment of the present invention.

[0019] Figure 2 It is the framework structure diagram of a monitoring system based on quartz crystal cutting anomalies provided for the second embodiment of the present invention. Detailed Embodiments

[0020] To make the above objects, features, and advantages of the present invention more obvious and understandable, the detailed embodiments of the present invention will be described below with reference to the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0021] Embodiment 1, referring to Figure 1 , which is an embodiment of the present invention, provides a monitoring method based on quartz crystal cutting anomalies, including: S1: Collaborating a high-speed industrial camera and a laser scanning spectrometer to capture the surface image and reflection spectrum of each cutting surface during the cutting process.

[0022] Specifically, high-speed industrial cameras are reasonably arranged along the cutting path on the cutting table, and camera models with high frame rate and high resolution are selected to achieve real-time imaging of the surface of the cutting area of the tool head movement region. During the camera acquisition process, LED or laser-assisted lighting is used to ensure image clarity and surface detail expression.

[0023] Synchronously, a laser scanning spectrometer is deployed beside the cutting area to perform point-by-point or scanning spectroscopic detection on the cutting surface passed by the tool head, and the reflection spectral curve of this area is obtained in real time.

[0024] The high-speed industrial camera and the laser scanning spectrometer are synchronized in time sequence through an industrial control computer or a data acquisition system. At each cutting moment, the corresponding two-dimensional surface image and the spectral data at this position are recorded to ensure the accurate correspondence of the image and the spectrum in spatial coordinates and cutting time sequence.

[0025] The obtained original data provides a high-quality data basis for subsequent multi-modal feature fusion and anomaly detection.

[0026] S2: Preprocess the surface image and the reflection spectrum and then perform spatial registration; fuse the features of the spatially registered image and the cutting path on the surface to obtain a clear input data stream.

[0027] In this embodiment, for the surface image, bilateral filtering is used to suppress noise; geometric correction based on a calibration plate is used to correct the geometric distortion of the cutting surface image; histogram equalization is used to normalize the brightness; and region marking based on the strong reflection threshold is used to mark artifacts.

[0028] For the reflection spectrum, maximum-minimum normalization is used to normalize the spectral signals collected under different batches and different lighting conditions to a unified interval for easy data comparison and analysis. The fixed wavelength band masking method is used to mask specific wavelength intervals (such as the multi-path artifact area unique to the laser) with abnormal instrument responses and irrelevant to physical characteristics according to preliminary experiments.

[0029] The preprocessed data is uniformly mapped to the standard coordinate system in physical space to make the data stream correspond one by one in spatial position and achieve pixel-level alignment.

[0030] Taking the fixed position of the tool head as a reference point and the tool head lines extending from the reference point to both sides as the parameter reflection baselines during cutting, the feedback feature vectors at the coordinate positions on the same parameter reflection baseline are made equal to the feedback feature vector of the reference point.

[0031] During cutting, through the initial position, cutting speed vector and time of the reference point, the continuous path of the reference point in the two-dimensional coordinate system where the image is located during the cutting process is recorded.

[0032] Parameterize the tool tip reference point path, and define the initial position of the tool tip reference point as: .

[0033] Let the cutting speed vector be: .

[0034] Then the two-dimensional coordinates of the reference point on the cutting path at any time are: ; In actual engineering, discrete integration, sensor sampling, or motion control feedback can be used to record the trajectory. Among them, represents the abscissa of the initial position of the tool tip reference point, represents the ordinate of the initial position of the tool tip reference point, represents the speed of the tool tip reference point in the x-axis direction at time; represents the speed of the tool tip reference point in the y-axis direction at time; represents the index of the time stamp. t represents the current time.

[0035] At the same time, generate a feedback feature vector on the same parameter reflection baseline: ; Among them, represents the two-dimensional pixel coordinates in the spatially registered image, , represents the path of the reference point in the image; represents the feedback feature vector of the pixel point on the image; cutting speed of the reference point at time t; represents the acceleration of the reference point at time t; represents the normal force of the tool tip at time t; represents the tangential force of the tool tip at time t; represents the cutting trajectory error, which is defined as: the Euclidean distance deviation between the actual tool tip trajectory and the control instruction.

[0036] For each pixel point on each parameter reflection baseline, introduce a virtual dimension to represent the actual feedback feature vector of each pixel point on the parameter reflection baseline; on the virtual dimension, make random assumptions about the actual feedback feature vectors of all pixel points on both sides.

[0037] The constraint of the random assumption is: on the parameter reflection baseline, the sum of the actual feedback feature vectors of all pixel points is equal to the feedback feature vector of the parameter reflection baseline.

[0038] Solve the fitness for the entire image, and select a random hypothesis combination with a fitness higher than the preset value as the evaluation result; obtain the set of global random hypothesis combinations ; A1, A2, An respectively represent the 1st, 2nd, and nth random hypothesis combinations. According to the fitness corresponding to each element in set A, convert them into probabilities proportionally so that the sum of the probabilities of all elements in the set is 1.

[0039] It should be noted that by taking the fixed reference point of the tool head as the origin of the cutting parameters, combining the cutting speed vector and time information recorded in real time, the precise parameterization of the continuous trajectory of the tool head reference point in the two-dimensional space is realized. Further, along the normal direction of the reference point path (i.e., the parameter reflection baseline), the physical feedback parameters of the reference point (such as speed, acceleration, normal force, tangential force, trajectory error, etc.) are assigned to all pixel points on the baseline, realizing the mapping and expansion of physical information in space. In this way, all pixel points on the same parameter reflection baseline are regarded as homogeneous units in the feedback feature space, uniformly expressing the physical state of the cutting process.

[0040] In order to improve the expression ability for real complex anomalies, this method introduces virtual dimensions for each pixel point on each parameter reflection baseline, generates several groups of random hypotheses for all actual feedback feature vectors, and sets constraints: under each group of random hypotheses, the sum of the feedback feature vectors of all pixel points should be equal to the overall feedback feature vector of the baseline. Through this method, the system can perform "regularized" high-dimensional modeling and inference on irregular, non-uniform, unknown noise, and multi-source anomalies under the combination of two-dimensional space and virtual random space.

[0041] Globally, select high-confidence hypothesis combinations as effective evaluation results through fitness solution, and normalize the probabilities of all random hypothesis combinations, making the analysis results have both statistical reliability and physical interpretability. Generally speaking, this solution effectively introduces the "structured randomness" of the third dimension, breaks through the description limitations of non-homogeneous and complex anomalies in two-dimensional space, and lays a solid data foundation and mathematical model foundation for the intelligent detection of cutting anomalies and process traceability.

[0042] The fitness is equal to the sum of the prior probability based on historical data and the rationality analysis result of a one-dimensional simplified convolutional neural network. Among them, the one-dimensional simplified convolutional neural network is used to determine whether each group of hypotheses (i.e., the physical parameter distribution of each pixel point) is possible, physically reasonable, and free of internal contradictions in "structural mechanics / material mechanics". Core: The distribution of physical quantities (such as normal force, tangential force, stress, displacement, velocity, etc.) needs to conform to physical laws such as force balance, continuity, and stress-strain relationship under real materials and structures. The structure of the one-dimensional simplified convolutional neural network (1D-CNN) is as follows: Input layer: The global image configured with each combination of random hypotheses, where each pixel has multi-dimensional physical features.

[0043] Convolutional layer (1D Conv): The receptive field of the convolutional kernel can cover a physically local area, equivalent to a sliding window mechanical analysis.

[0044] Pooling layer (Pooling): Aggregate local features to simulate stress continuity and overall trends.

[0045] Fully connected layer (FC): Global mechanical discrimination.

[0046] Output layer: Sigmoid activation, outputting a rationality probability score.

[0047] The feature fusion includes: mapping each combination of random hypotheses corresponding to each element in set A into the spatially registered image respectively according to the pixel correspondence, obtaining n groups of images with increased parameter dimensions.

[0048] By taking the distribution of physical quantities (such as normal force, tangential force, stress, velocity, etc.) of the parameters under each group of random hypotheses reflected on the baseline as the input, a one-dimensional simplified convolutional neural network (1D-CNN) is designed to automatically discriminate the rationality of the hypothesis at the physical and structural mechanics levels. The structure of 1D-CNN enables it to perform end-to-end modeling and feature discrimination on the continuity of physical quantities in space, mechanical equilibrium, stress-strain relationship, etc., effectively identifying those hypothesis results with internal physical contradictions, local stress mutations, or non-compliance with material mechanics laws.

[0049] In this way, the system not only considers the prior probability formed by historical data statistics but also takes complex mechanical rationality as a key element in fitness evaluation, thus significantly improving the physical authenticity and engineering reliability of anomaly detection. Hypotheses with high fitness represent parameter distributions that have both statistical support and meet mechanical rationality, being more in line with the defects and anomaly manifestations in actual engineering scenarios.

[0050] In the feature fusion stage, all combinations of hypotheses with high fitness are used as parameter mapping templates and embedded into the original image data stream after spatial registration according to the pixel-level one-to-one correspondence relationship, generating multiple new data streams with increased parameter dimensions. These new data streams retain the original spatial and optical information, while superimposing the physical feedback and rationality discrimination dimensions, providing a rich and multi-level data basis for subsequent pyramid reconstruction and intelligent detection. Finally, the system can achieve all-round high-precision identification of complex cutting anomalies and trace the physical mechanism, ensuring the scientificity, accuracy, and interpretability of the analysis results.

[0051] S3: Perform pyramid reconstruction on the data stream corresponding to each section plane.

[0052] For the n groups of images after the parameter dimension increases, the following operations are performed separately: Multi-scale pyramid construction: For the high-dimensional input feature tensor after spatial registration , through multi-level downsampling, successively generate pyramid features with decreasing resolution for ; where h is the original resolution. represents the high-dimensional input feature tensor after spatial registration, represents the original horizontal axis coordinate, represents the original vertical axis coordinate. represents the layer pyramid feature tensor, with decreasing resolution, represents the (h - 1)-th layer pyramid feature tensor; represents the layer downsampling operator, such as Gaussian downsampling, stride convolution, etc. represents the horizontal and vertical coordinates corresponding to the layer; represents the current layer number of the pyramid, ranging from 0 to . represents the total number of layers of the pyramid.

[0053] Feature enhancement and contrast learning structure for each layer: In each layer of the pyramid, an independent feature extraction and weak anomaly amplification module is adopted. A contrast learning structure and a custom loss mechanism are introduced to automatically amplify low-contrast defect signals: First, feature representations (positive and negative samples) are extracted for the normal regions and weak anomaly regions at each scale respectively: ; Construct a self-supervised contrast loss (such as InfoNCE, Triplet loss, etc.), which narrows the distance between similar features and widens the distance between different classes (normal and abnormal) features, especially increasing the weight for low-amplitude abnormal signals.

[0054] Custom weak anomaly amplification loss for each layer: ; where represents the weak anomaly amplification loss for the h-th layer, which is used to enhance low-contrast defect signals. represents the hyperparameter of the anomaly amplification weight for the h-th layer. represents the number of sampling points in the weak anomaly region for the h-th layer. represents the current sampling point number, ranging from 1 to . represents the feature vector of the s-th weak anomaly region for the h-th layer. Denote the reference feature mean vector of the normal region (positive sample) in the h-th layer.

[0055] Multi-scale contrast loss: By comparing the feature representations of different scales and different regions, enhance the sensitivity of the model to abnormal signals: ; Where, Denote the multi-scale contrast loss in the h-th layer. Denote the number of sampling points for comparison in the h-th layer. Denote the current sampling number in the contrast loss, ranging from 1 to . Denote the feature vector (positive sample) of the -th sample in the h-th layer. Denote the feature mean vector (negative sample) of the weak abnormal region in the h-th layer. Denote the temperature scaling parameter in the h-th layer. Denote the numbers of all comparison samples in the denominator. Denote the feature vector of the k-th sample in the h-th layer of the pyramid, which is the feature representation of the positive sample or all samples used for multi-scale contrast loss calculation.

[0056] Upsampling and multi-layer fusion reconstruction: The features enhanced and contrast-optimized in each layer are upsampled and uniformly regressed to the original resolution, and fused to form the final output feature tensor: ; Where, Denote the high-dimensional feature tensor of the final fusion output, with the resolution of the original input size. FG( ) denotes the multi-layer feature fusion operation (such as weighting, splicing, attention mechanism, etc.). US Denote the upsampling operator of the -th layer, which is used to restore the feature tensor to the original resolution. Denote the feature tensor of the -th layer after feature enhancement and contrast optimization. Denote the pyramid layer number, ranging from 0 to .

[0057] The total loss is the weighted sum of the weak anomaly amplification loss and the multi-scale contrast loss in each layer, as well as the global detection or segmentation loss: ; Where, Denote the overall training loss function. Denote the weak anomaly amplification loss in the -th layer. Denote the Weighting factor for layer contrast loss. Indicates the multi-scale contrast loss of the layer. Indicates the weight hyperparameter for the global detection or segmentation loss. Indicates the global detection or segmentation loss of the final output.

[0058] Finally, upsampling and multi-layer fusion reconstruction are performed to obtain the reconstructed image, which is used as a data stream for anomaly detection.

[0059] It should be noted that by using the multi-scale feature pyramid structure, all-round and multi-level anomaly information capture and intelligent enhancement are realized for the high-dimensional input data stream after the dimensionality of the cutting surface parameters is increased. Through multi-level downsampling, the feature tensor after the original spatial registration is gradually decomposed into pyramid feature layers with different resolutions. Each layer focuses on the structural features of different spatial scales, thus taking into account the detection requirements for large-range structural anomalies and tiny weak signal anomalies.

[0060] At each pyramid level, independent feature enhancement and anomaly amplification modules are respectively set, and a self-supervised contrast learning mechanism and a custom anomaly loss function are explicitly introduced, enabling the model to actively distinguish the features of normal regions and weak anomaly regions, and amplify the signals and enhance the significance of low-contrast and weak anomalies. The distribution distance between normal and abnormal features is further widened through the contrast loss, effectively improving the sensitivity and robustness of the model to complex anomalies.

[0061] The feature outputs of all pyramid layers are upsampled and fused at multiple levels to restore to the original resolution, forming a high-dimensional output tensor containing multi-scale spatial information and anomaly-enhanced features, providing a multi-angle and information-dense input basis for downstream anomaly detection. Finally, through global loss integration, the weak anomaly signals are amplified layer by layer and globally identified with high precision, breaking through the technical bottlenecks of traditional segmentation / detection models with weak low-contrast anomaly recognition ability and limited spatial scale, and greatly improving the detection sensitivity and engineering practicality of quartz crystal cutting anomalies.

[0062] S4: Simplify the inherent parameters of the reconstructed cutting surface data stream, analyze the simplified cutting surface data stream, and obtain the detection result of the quartz crystal cutting surface anomaly.

[0063] The inherent parameters include: in the data stream after pyramid reconstruction of the standard cutting surface under the same material, the abnormal parameters that have been identified as having abnormal manifestations. The simplification process of the inherent parameters includes: comparing the characteristics of the inherent parameters in the n groups of data streams obtained, extracting the data components most similar to the inherent parameters from the n groups of data streams, screening m data streams with the content of similar data components higher than the preset value as the data streams to be analyzed; in the data streams to be analyzed, the part that is the same as the abnormal parameter is always used as the normal parameter manifestation, and the abnormal part caused only by the cutting abnormality is obtained; where m ≤ n.

[0064] In the solution of the present invention, by introducing a third random dimension, multiple groups of hypothesis decompositions are performed on the feedback feature vectors of the pixels on the baseline reflected by each parameter of the cutting surface, realizing the expression and regularization of complex abnormal distributions in high-dimensional space. This mechanism greatly enhances the system's ability to capture diverse abnormalities, but also brings a potential risk: in the hypothesis space, some hypothesis combinations may have too much introduced freedom, resulting in over-correction of the original physical characteristics, that is, some data streams may be "corrected" to deviate from the true material or process background characteristics, leading to weakened physical interpretability and even loss of perception of the actual abnormal nature.

[0065] Therefore, it is necessary to actively screen those data streams in all generated data streams that are highly similar to the characteristics of the known material inherent parameters. The purpose of doing this is to ensure that the selected data streams still maintain the abnormal characteristic performance that is most consistent with the true material background to the greatest extent after introducing random hypotheses and high-dimensional mappings, and avoid algorithm drift and physical distortion caused by the "third dimension". Only these most similar data streams are the "physically reasonable representatives" reflecting the objective abnormal distribution of the material, and subsequent analysis can also be based on the true process environment and material background to ensure the scientific nature and credibility of the detection.

[0066] In addition, by setting "the similarity content is higher than the preset value" as the screening criterion, the intensity of retaining the inherent parameters can be flexibly adjusted, balancing the sensitivity of the algorithm and physical constraints, and effectively preventing the generalization out of control of abnormal detection and the increase in the false positive rate.

[0067] Further, the analysis of the simplified cutting surface data stream specifically includes: Step 1: In the m groups of data streams to be analyzed, perform clustering of data characteristics. If there is only one clustering cluster, go to Step 3; if there are more than one clustering clusters, go to Step 2. By performing feature clustering on multiple groups of data streams, the automatic classification of abnormal expression patterns is realized. This design aims to reveal the internal structural association between data, provide a hierarchical basis for subsequent abnormal identification, and reduce the deviation caused by subjective human division.

[0068] Step 2: If the ratio of the number of individuals in the family with the most individuals to the number of individuals in the family with the second most individuals is greater than the preset value, only the data stream corresponding to the family with the most individuals is retained, and go to Step 3; otherwise, go to Step 6 (normally, going to Step 3 is because the characteristic manifestations of the abnormal part are very similar and can be equivalent through fuzzy concepts. However, the content going to Step 6 is the abnormal situation, and at this time, the regional differences in abnormal manifestations may be a bit large, and at this time, it cannot be simply replaced by fuzzy concepts). By setting the judgment threshold for the ratio of the number of individuals between clustering families, the processing strategy can be dynamically adjusted in scenarios with obvious commonalities and significant differences. This shunt mechanism can avoid excessive fuzziness caused by a few outliers, enabling the mainstream features to be more powerfully expressed.

[0069] Step 3: According to the data streams of each individual in the same clustering family, divide the abnormal part and the normal part in the image, and blur the boundary of the division (in this case, it is basically caused by slight differences in the boundaries of the abnormal part. Therefore, by blurring the abnormal boundary, it can encompass the characteristics of all individuals), so that the characteristics of each individual can be expressed in the blurred image after division.

[0070] It should be noted that for the situation where there are slight differences in the abnormal feature boundaries in multiple groups of high-dimensional data streams, by introducing fuzzy processing, the boundary information between individuals is moderately fused, making the clustering results more inclusive and representative when expressing abnormal regions. This fuzzy processing can effectively alleviate the boundary uncertainty caused by noise, local perturbations, or microstructural differences, thereby more accurately identifying and locking the core abnormal regions in the data stream. In addition, when the differences between clustering families are large, by independently analyzing each family and assigning probability weights, the diversity of different abnormal patterns can be taken into account, making the subsequent abnormal identification both robust and retaining the uncertainty of the results, improving the scientificity and engineering adaptability of the overall analysis.

[0071] In this embodiment, although the concept of fuzzification is introduced, in fact, in this solution, the fuzzy concept is simplified. After blurring the boundary of the region division, only the part with consistent intermediate data is needed. Therefore, for the boundary of the fuzzy boundary band, both the inner and outer boundaries are divided, and only the inner boundary is used to extract the "part with consistent intermediate data". After simplification, the implementation of the algorithm will also be simplified: originally, the fuzzy concept was to stack the abnormal masks of each individual in the same clustering family according to pixel positions to obtain a probability map, and calculate the confidence according to the probability to achieve fuzzification. In this solution, by identifying the boundaries of the abnormal regions of each individual and stacking the identification results, an abnormal boundary identification result that stacks several data streams is finally obtained; then, in this boundary identification result, for each abnormal region, the inner and outer boundaries of the separation band are found. Thus, the purpose of fuzzy identification is simply realized.

[0072] Step 4: Lock the abnormal cores of the m images after fuzzy partitioning, and determine the inner and outer boundaries of the fuzzy partitioning.

[0073] Step 5: Use a pre-trained neural network to analyze the pixel features in the inner boundary; obtain the recognition result of the abnormality.

[0074] Step 6: Calculate the number of individuals in each clustering family. Take the number of individuals within the family / the total number of individuals × 100% as the probability coefficient; let the data stream corresponding to each family enter Step 3 respectively, and add the corresponding probability coefficient during output as an additional element of the evaluation result. By assigning probability weights to the results of different clustering families, the final output has both subjective robustness and diverse expression capabilities, can reflect the uncertainty of the data itself, and facilitates the risk quantification and traceability of subsequent decision-making.

[0075] In this embodiment, the neural network can be a convolutional neural network (CNN, Convolutional Neural Network), such as structures like U-Net, ResNet, DenseNet, EfficientNet, etc. The convolutional neural network has a powerful ability to automatically extract spatial structures and local texture features, and is suitable for processing image data streams after spatial registration, pyramid features, and fine-grained recognition tasks of abnormal regions. CNN can use pre-trained models for transfer learning, significantly improving the detection effect and generalization ability in scenarios with limited data volume or complex features. Modern CNNs support end-to-end input and output, facilitating integration with the entire automated detection process, and can efficiently and real-time complete the localization, segmentation, and classification of core abnormalities. In other embodiments, the neural network can also be an attention-based network, a graph neural network (GNN, Graph Neural Network), a recurrent neural network (RNN / LSTM / GRU), a generative adversarial network (GAN, Generative Adversarial Network), etc.

[0076] Embodiment 2, as Figure 2 shown, this embodiment also provides a monitoring system for quartz crystal cutting anomalies, which includes: An acquisition unit, which cooperates with a high-speed industrial camera and a laser scanning spectrometer to capture the surface image and reflection spectrum of each cut surface during the cutting process.

[0077] A processing unit, which preprocesses the surface image and the reflection spectrum and then performs spatial registration; fuses the features of the spatially registered image and the cutting path on the surface to obtain a clear input data stream.

[0078] Optimization unit, which performs pyramid reconstruction on the data stream corresponding to each section plane.

[0079] Detection unit, which simplifies the inherent parameters of the reconstructed section plane data stream, analyzes the simplified section plane data stream, and obtains the detection result of the quartz crystal section plane anomaly.

[0080] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this 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 for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs, etc., all kinds of media that can store program codes.

[0081] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device.

[0082] More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connection parts with one or more wirings (electronic devices), portable computer disk cartridges (magnetic devices), random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memories), fiber optic devices, and portable compact disc read-only memories (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, then editing, interpreting, or processing it in other suitable ways as necessary, and then storing it in a computer memory.

[0083] It should be understood that each part of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0084] 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 preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A monitoring method based on abnormal quartz crystal cutting, characterized in that, Including: By collaborating a high-speed industrial camera and a laser scanning spectrometer, capturing the surface image and reflection spectrum of each section during the cutting process; For the surface image and the reflection spectrum, after preprocessing, perform spatial registration to obtain the registered image; fuse the features of the spatially registered image and the cutting path on the surface to obtain a clear input data stream; Perform pyramid reconstruction on the data stream corresponding to each section; Simplify the inherent parameters of the reconstructed section data stream, analyze the simplified section data stream, and obtain the detection result of the abnormality of the quartz crystal section.

2. The monitoring method based on quartz crystal cutting anomaly according to claim 1, wherein: The preprocessing includes: for the surface image, perform noise suppression, geometric distortion correction of the section image, brightness normalization, and artifact removal; for the reflection spectrum, perform spectral normalization and interference band removal.

3. The monitoring method based on quartz crystal cutting anomaly according to claim 2, wherein: The spatial registration includes: uniformly mapping the preprocessed data into the standard coordinate system of the physical space, making the data stream correspond one by one in the spatial position, and achieving pixel-level alignment.

4. The monitoring method based on abnormal quartz crystal cutting as claimed in claim 3, wherein: The cutting path includes: taking the fixed position of the tool tip as the reference point, and taking the tool tip lines extending from the reference point to both sides as the parameter reflection baseline during cutting, so that the feedback feature vectors at the coordinate positions on the same parameter reflection baseline are equal to the feedback feature vector of the reference point; During cutting, based on the initial position of the reference point, the cutting speed vector, and time, record the continuous path of the reference point in the two-dimensional coordinate system where the image is located during the cutting process, and simultaneously generate a feedback feature vector that lies on the same parameter reflection baseline: ; Among them, represents the two-dimensional pixel coordinates in the spatially registered image, , represents the path of the reference point in the image; represents the feedback feature vector of the pixel point on the image; The cutting speed of the reference point at time t; represents the acceleration of the reference point at time t; represents the normal force of the tool tip at time t; represents the tangential force of the tool tip at time t; represents the cutting trajectory error; t represents the t-th moment in the cutting process; At each pixel point in each parameter reflection baseline, introduce a virtual dimension to represent the actual feedback feature vector of each pixel point on the parameter reflection baseline; on the virtual dimension, make random assumptions about the actual feedback feature vectors of all pixel points on both sides; The constraint of the random assumption is that the sum of the actual feedback feature vectors of all pixel points on the parameter reflection baseline is equal to the feedback feature vector of the parameter reflection baseline; Solve the fitness for the entire image, and select a random hypothesis combination with a fitness higher than the preset value as the evaluation result; obtain the set of global random hypothesis combinations ; Let An denote the nth random hypothesis combination, and convert it into a probability proportionally according to the fitness corresponding to each element in the set A; The fitness is equal to the sum of the prior probability based on historical data and the rationality analysis result of the one-dimensional simplified convolutional neural network; The feature fusion includes: mapping the random assumption combinations corresponding to each element in set A into the spatially registered image respectively according to the pixel point correspondence relationship, and obtaining n groups of images with increased parameter dimensions; 5. The monitoring method based on quartz crystal cutting anomaly according to claim 4, characterized in that: The pyramid reconstruction includes: respectively performing on the n groups of images with increased parameter dimensions: multi-scale pyramid construction, performing feature enhancement and contrast learning on each layer, and amplifying the weak anomaly loss by extracting positive and negative samples; Enhance the sensitivity of the anomaly signal by comparing the feature representations of different scales and different regions; Finally, perform upsampling and multi-layer fusion reconstruction to obtain the reconstructed image as the data stream for anomaly detection.

6. The monitoring method based on abnormal quartz crystal cutting as claimed in claim 5, wherein: The inherent parameters include: the anomaly parameters that have been identified as having abnormal manifestations in the data stream after pyramid reconstruction of the standard cutting surface under the same material; The simplification process of the inherent parameters includes: comparing the features of the inherent parameters in the obtained n groups of data streams, extracting the data components most similar to the inherent parameters from the n groups of data streams, screening m data streams with the content of similar data components higher than the preset value as the data streams to be analyzed; in the data streams to be analyzed, always regard the part same as the anomaly parameter as the normal parameter manifestation, and obtain the anomaly part caused only by the cutting anomaly; where m ≤ n.

7. The monitoring method based on quartz crystal cutting anomaly as claimed in claim 6, wherein: The analysis of the simplified cross-section data stream specifically includes: Step 1: Cluster the data features in the m groups of data streams to be analyzed. If there is only one cluster family, go to Step 3; if there are more than one cluster families, go to Step 2; Step 2: If the ratio of the number of individuals in the family with the most individuals to the number of individuals in the family with the second most individuals is greater than the preset value, only retain the data stream corresponding to the family with the most individuals and go to Step 3; otherwise, go to Step 6; Step 3: According to the data streams of each individual in the same cluster family, divide the abnormal part and the normal part in the image, and blur the boundary of the division so that the features of each individual can be expressed in the blurred divided image; Step 4: Lock the abnormal core of the m blurred divided images to determine the inner and outer boundaries of the blurred division; Step 5: Use the pre-trained neural network to analyze the pixel features in the inner boundary to obtain the abnormal recognition result; Step 6: Calculate the number of individuals in each cluster family, and use the number of individuals in the family / total number of individuals × 100% as the probability coefficient; respectively enter Step 3 with the data stream corresponding to each family, and at the same time add the corresponding probability coefficient when outputting as an additional element of the evaluation result.

8. A monitoring system based on abnormal quartz crystal cutting, which adopts the method according to any one of claims 1-7, is characterized in that: It includes an acquisition unit that cooperates with a high-speed industrial camera and a laser scanning spectrometer to capture the surface image and reflection spectrum of each cross-section during the cutting process; A processing unit that preprocesses the surface image and the reflection spectrum and then performs spatial registration to obtain the registered image; Fuse the features of the spatially registered image with the cutting path on the surface to obtain a clear input data stream; An optimization unit that performs pyramid reconstruction on the data stream corresponding to each cross-section; A detection unit that simplifies the inherent parameters of the reconstructed cross-section data stream and analyzes the simplified cross-section data stream to obtain the detection result of the quartz crystal cross-section abnormality.

9. A computer device, comprising: A memory and a processor; the memory stores a computer program, characterized in that: when the processor executes the computer program, it implements the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1-7.

Images