Dynamic balance monitoring method, system, device, storage medium and program product

Abstract

A dynamic balancing monitoring method, system, device, storage medium and program product. When the dynamic balancing monitoring method is working, a neural network model is used to train a test vibration waveform diagram to obtain a test label, and the test label is compared with a specific label under various known fault types. If the test label matches a characteristic label of a certain fault type, it can be determined that the spindle to be confirmed has the same fault type; if the test label does not match all known fault labels, it can be considered that the spindle to be confirmed is working normally. The dynamic balancing monitoring method provided by the embodiment of the present invention can monitor the dynamic balancing state of the spindle of a semiconductor device in real time.

Description

Technical Field

[0001] Embodiments of the present invention relate to the field of semiconductor equipment, and in particular, to a dynamic balance monitoring method, system, equipment, storage medium, and program product. Background Art

[0002] The cutting machine is a commonly used processing equipment. The core components of the cutting machine include the spindle and the lead screw, which play a key role in the cutting process. The spindle drives the cutting tool to rotate, while the lead screw controls the feed movement of the cutting tool, together completing the precise cutting of the workpiece.

[0003] However, as the cutting machine is used for a longer time, its core components will age and wear out. This aging and wear will cause the spindle and lead screw to produce eccentric vibration, which will affect the cutting accuracy. In particular, the vibration of the spindle will be directly transmitted to the cutting tool, causing the tool to fluctuate during the cutting process.

[0004] Due to the lack of effective real-time monitoring methods, when the saw cuts wafers, the operator cannot promptly detect the tool fluctuations caused by the spindle vibration. In this case, a variety of quality abnormalities may occur during the cutting process, such as product off-cut, edge collapse, defects or incomplete cutting. Summary of the invention

[0005] The problem solved by the embodiments of the present invention is to provide a dynamic balance monitoring method, system, device, storage medium and program product for monitoring the dynamic balance state of a spindle of a semiconductor device.

[0006] To solve the above problems, an embodiment of the present invention provides a dynamic balance monitoring method for monitoring the dynamic balance of a spindle of a semiconductor device, comprising: providing a plurality of faulty spindle samples of semiconductor devices, each faulty spindle having a different fault type, and a plurality of faulty spindles under each fault type; obtaining a fault vibration waveform diagram of the spindle under each fault type; using a neural network model to train the fault vibration waveform diagram under each fault type one by one, and obtaining a specific label of the fault vibration waveform diagram under each fault type; providing a test vibration waveform diagram of the spindle of the semiconductor device in a state to be confirmed in real time; using the neural network model to train the test vibration waveform diagram to obtain a test label; comparing the test label with the specific label under each fault type to obtain the working state of the spindle in the state to be confirmed.

[0007] An embodiment of the present invention also provides a dynamic balance monitoring system for monitoring the dynamic balance of the spindle of a semiconductor device, including: a fault sample module, used to provide faulty spindle samples of multiple semiconductor devices, each faulty spindle has a different fault type, and there are multiple faulty spindle samples under each fault type; a fault waveform acquisition module, used to obtain the fault vibration waveform of the spindle under each fault type; a specific label acquisition module, used to use a neural network model to train the fault vibration waveform under each fault type one by one, and obtain the specific label of the fault vibration waveform under each fault type; a test vibration waveform module, used to provide a test vibration waveform of the spindle of the semiconductor device in a state to be confirmed in real time; a test label acquisition module, used to use the neural network model to train the test vibration waveform to obtain a test label; a comparison module, used to compare the test label with the specific label of each fault type, and obtain the working state of the spindle in the state to be confirmed.

[0008] An embodiment of the present invention also provides an electronic device, comprising at least one memory and at least one processor, wherein the memory stores one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the dynamic balance monitoring method as described above.

[0009] An embodiment of the present invention further provides a storage medium, wherein the storage medium stores one or more computer instructions, and the one or more computer instructions are used to implement the aforementioned dynamic balance monitoring method.

[0010] An embodiment of the present invention further provides a computer program product, comprising computer instructions, which are used to implement the aforementioned dynamic balance monitoring method when executed by a processor.

[0011] Compared with the prior art, the technical solution of the embodiment of the present invention has the following advantages:

[0012] The dynamic balancing monitoring method provided by the embodiment of the present invention first provides multiple faulty spindle samples of semiconductor devices, each faulty spindle has a different fault type, and each fault type has multiple faulty spindle samples; then the fault vibration waveform of the spindle under each fault type is obtained, and the neural network model is used to train the fault vibration waveform under each fault type one by one, so as to obtain the specific label of the fault vibration waveform under each fault type. Then the neural network model is used to train the test vibration waveform to obtain a test label, and the test label is compared with the specific labels under various known fault types. If the test label matches the characteristic label of a certain fault type, it can be determined that the spindle to be confirmed has the same fault type; if the test label does not match all known fault labels, it can be considered that the spindle to be confirmed is working normally. The dynamic balancing monitoring method provided by the embodiment of the present invention can monitor the dynamic balancing state of the spindle of the semiconductor device in real time, effectively avoid the blade swing caused by the excessive amplitude of the dynamic balancing vibration of the spindle, thereby preventing the product from having quality problems such as cutting deviation, edge collapse, and incomplete cutting. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is a flow chart of an embodiment of a dynamic balance monitoring method of the present invention;

[0014] Figure 2 is a schematic structural diagram of a semiconductor device of the present invention;

[0015] Figure 3 It is a flow chart of obtaining the fault vibration waveform diagram of the main shaft under various fault types in an embodiment of the dynamic balancing monitoring method of the present invention;

[0016] Figure 4 It is a flow chart of generating fault vibration waveform diagrams under various fault types based on characteristic data of the main shaft under various fault types in an embodiment of the dynamic balancing monitoring method of the present invention;

[0017] Figure 5 A flowchart of an embodiment of the dynamic balance monitoring method of the present invention in which a neural network model is used to train the fault vibration waveform graphs under each fault type one by one to obtain specific labels of the fault vibration waveform graphs under each fault type;

[0018] Figure 6 A flowchart of using the neural network model to train the test vibration waveform diagram to obtain a test label in an embodiment of the dynamic balance monitoring method of the present invention;

[0019] Figure 7 The present invention is a flow chart of a method for monitoring dynamic balance in which the nearest neighbor method is used to compare the test tag with specific tags under various fault types to obtain the working state of the spindle to be confirmed.

[0020] Figure 8 is a functional block diagram of an embodiment of a dynamic balance monitoring system of the present invention;

[0021] Fig. 9 It is a functional block diagram of a fault waveform acquisition module in an embodiment of a dynamic balance monitoring system of the present invention;

[0022] Fig.10 It is a functional block diagram of a fault vibration waveform diagram unit in an embodiment of a dynamic balancing monitoring system of the present invention;

[0023] Fig.11 It is a functional block diagram of a specific tag acquisition module in an embodiment of a dynamic balance monitoring system of the present invention;

[0024] Fig.12 is a functional block diagram of a comparison module in an embodiment of a dynamic balance monitoring system of the present invention;

[0025] Fig.13 It is a hardware structure diagram of an embodiment of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0026] As can be seen from the background technology, due to the lack of effective real-time monitoring means, operators are unable to promptly detect tool fluctuations caused by spindle vibration. In this case, a variety of quality abnormalities may occur during the cutting process, such as product off-cut, edge collapse, defects or incomplete cutting, which not only reduces product quality, but may also lead to material waste and reduced production efficiency.

[0027] In order to solve the technical problem, the dynamic balance monitoring method provided by the embodiment of the present invention first provides multiple faulty spindle samples of semiconductor devices, each faulty spindle has different fault types, and each fault type has multiple faulty spindle samples; then obtain the fault vibration waveform of the spindle under each fault type, and use the neural network model to train the fault vibration waveform under each fault type one by one, so as to obtain the specific label of the fault vibration waveform under each fault type. For the test vibration waveform of the spindle of the semiconductor device in the state to be confirmed, the neural network model is used to train the test vibration waveform to obtain the test label, and the test label is compared with the specific labels under various known fault types. If the test label matches the characteristic label of a certain fault type, it can be determined that the spindle to be confirmed has the same fault type; if the test label does not match all known fault labels, it can be considered that the spindle to be confirmed is working normally. The dynamic balance monitoring method provided by the embodiment of the present invention can monitor the dynamic balance state of the spindle of the semiconductor device in real time, effectively avoid the blade swing caused by the excessive amplitude of the dynamic balance vibration of the spindle, thereby preventing the product from having quality problems such as cutting deviation, edge collapse, and incomplete cutting.

[0028] In order to make the above-mentioned purposes, features and advantages of the embodiments of the present invention more obvious and understandable, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0029] Figure 1 The figure is a flow chart of an embodiment of the dynamic balance monitoring method of the present invention. The dynamic balance monitoring method of the present invention is used to monitor the dynamic balance of the spindle of semiconductor equipment.

[0030] Step S1, providing a plurality of faulty spindle samples of semiconductor devices, each faulty spindle having a different fault type, and a plurality of faulty spindles under each fault type.

[0031] Providing samples of faulty spindles of multiple semiconductor devices provides the necessary data basis for the subsequent steps of obtaining fault vibration waveforms of the spindles under various fault types and training with neural network models. The fault types of each faulty spindle are different, which is conducive to covering a variety of fault types, making this monitoring method have stronger generalization ability and higher accuracy. In addition, there are multiple samples of faulty spindles under each fault type, which enhances the reliability and robustness of dynamic balancing monitoring.

[0032] In this embodiment, in the step of providing a plurality of faulty spindle samples of semiconductor devices, the fault types include: mechanical wear and aging, manufacturing and assembly errors, and external load unevenness.

[0033] Classifying fault types into mechanical wear and aging, manufacturing and assembly errors, and uneven external loads provides a clear goal for the subsequent training of the neural network model, enabling the neural network model to specifically learn the characteristics of different types of faults, which helps to improve the accuracy and efficiency of fault diagnosis.

[0034] In this embodiment, the number of faulty main axes under each fault type is greater than or equal to 5.

[0035] The number of faulty spindles under each fault type is greater than or equal to 5, so that each fault type has enough samples for training, so that the specific labels of each fault type obtained later are accurate and reliable. And the number of faulty spindles under each fault type is greater than or equal to 5. This is conducive to better distinguishing the specific labels of different fault types and reducing the subsequent misjudgment of the working status of the spindle to be confirmed.

[0036] It should be noted that the semiconductor device in this embodiment is a cutting machine, which is used to cut the processed wafer into individual chips.

[0037] Specifically, Figure 2As shown, the cutting machine includes: a main shaft 1, which is placed horizontally; a blade 2, which is installed on the main shaft of the cutting machine; a cutting disk 3, which can move in the horizontal (x) and vertical (y) directions relative to the main shaft; and a wafer 4, which is placed on the cutting disk.

[0038] When the cutting machine is working, the wafer 4 is first placed on the cutting disk 3, and then the spindle 1 rotates, and the cutting disk 3 moves horizontally (x) relative to the spindle 1, so that the cutting path on the wafer 4 is aligned with the position of the blade, and the cutting disk 3 drives the wafer 4 to move in the vertical (y) direction, approaching the blade 2, and the cutting operation of the wafer 4 is performed.

[0039] like Figure 3 As shown, in step S2, a fault vibration waveform diagram of the main shaft under each fault type is obtained.

[0040] Obtaining the fault vibration waveform of the spindle under each fault type provides basic data for the subsequent use of the neural network model to train the fault vibration waveform of each fault type one by one. By obtaining the fault vibration waveform of the spindle under each fault type, preparation is made for the subsequent accurate identification and classification of spindle faults.

[0041] Specifically, the steps of obtaining the fault vibration waveform diagram of the main shaft under each fault type include:

[0042] Step S21, setting sensors on various faulty spindles.

[0043] Sensors are set on various faulty spindles to prepare for the subsequent acquisition of the vibration data of the spindles under various fault types and the generation of fault vibration waveform diagrams under various fault types; and based on the vibration data of the spindles under various fault types, a fault database can be established to improve the accuracy and reliability of dynamic balancing monitoring.

[0044] In the step of setting sensors on various faulty spindles, the sensors include one or more of a velocity sensor, an acceleration sensor and a displacement sensor.

[0045] The speed (VEL) mm / s data can reflect the motion state of the spindle, and the acceleration (ACC) m / s 2 The data can capture sudden changes and shocks, while the displacement (DISP) mm data can show the actual position change of the spindle.

[0046] In this embodiment, a velocity sensor, an acceleration sensor and a displacement sensor are simultaneously arranged on the main shaft.

[0047] By simultaneously collecting the motion state of the spindle, sudden changes and impacts during operation, and the actual position changes of the spindle, more comprehensive fault diagnosis information can be provided, which helps to improve the accuracy and reliability of dynamic balancing monitoring.

[0048] It should be noted that in the step of setting sensors on various faulty spindles, the sensors are connected to multiple acquisition cards, and the acquisition cards are set at different positions. By setting acquisition cards at different positions, data at different positions can be collected to more comprehensively reflect the motion state of the faulty spindle.

[0049] Specifically, the number of the acquisition cards is eight. In other embodiments, the number of the acquisition cards may also be five, six, seven, or more than eight.

[0050] Step S22, collecting the vibration data of the main shaft under each fault type through the sensor.

[0051] The vibration data of the main shaft under various fault types are collected by sensors for subsequent time domain and frequency domain analysis, as well as for generating fault vibration waveform diagrams.

[0052] In this embodiment, in the step of collecting vibration data of the main shaft under various fault types through a sensor, the sensor collects vibration data of multiple positions of the faulty main shaft in parallel through multiple acquisition cards.

[0053] By collecting the vibration data of the faulty spindle in parallel using multiple acquisition cards, vibration information from multiple measurement points can be obtained simultaneously, improving the efficiency and comprehensiveness of data collection. Moreover, parallel acquisition can ensure the synchronization of data from different measurement points, providing a reliable data basis for subsequent time domain and frequency domain analysis.

[0054] Specifically, multiple acquisition cards are used to connect sensors to achieve multi-channel acquisition, which can simultaneously process signal inputs from multiple sensors to ensure data synchronization and consistency. In addition, the parallel acquisition method can greatly improve the efficiency of data acquisition, especially when a large amount of sample data needs to be collected.

[0055] Step S23, performing time domain and frequency domain analysis on the vibration data of the main shaft under each fault type, to obtain initial characteristic data under each fault type.

[0056] Time domain analysis can reflect the changing characteristics of the vibration signal over time, while frequency domain analysis can reveal the frequency composition of the signal. At the same time, time domain analysis and frequency domain analysis of the vibration data of the spindle under various fault types can provide a more comprehensive and accurate description of the fault characteristics.

[0057] In this embodiment, the time domain analysis includes but is not limited to performing time domain waveform analysis and calculating time domain high-order statistics. Specifically, the time domain waveform analysis is used to observe the amplitude change trend of the vibration data. Calculating time domain high-order statistics includes calculating skewness, kurtosis, etc. These parameters are particularly effective for detecting nonlinear and non-stationary processes.

[0058] In this embodiment, frequency domain analysis uses fast Fourier transform (FFT) to convert time domain signals into frequency domain signals, which is the basis of frequency domain analysis and can decompose vibration data into components of different frequencies. The converted frequency domain signals are analyzed to identify the main frequency components. By analyzing the frequency components of the converted frequency domain signals, the type and severity of the fault can be preliminarily determined.

[0059] Step S24, removing redundant parameters in the initial characteristic data under each fault type, and obtaining characteristic data of the spindle under each fault type.

[0060] Removing redundant parameters from the initial feature data is used to improve the efficiency and accuracy of subsequent data processing and provide more effective feature data for subsequent neural network model training. Specifically, removing redundant parameters from the initial feature data is conducive to reducing data noise and improving the signal-to-noise ratio of the signal; reducing computational complexity and speeding up subsequent neural network model training; avoiding overfitting and improving the generalization ability of the model; highlighting key features, so that subsequent neural network models can more easily capture specific labels of fault vibration waveforms.

[0061] In this embodiment, the principal component analysis (PCA) dimensionality reduction method is used to remove redundant parameters in the initial feature data under each fault type, and obtain the feature data of the main shaft under each fault type. The principal component analysis dimensionality reduction method can identify and retain the most valuable features for fault diagnosis, and can convert high-dimensional initial feature data into low-dimensional principal components, retaining the most important variation information in the data, which is particularly useful for processing complex vibration data.

[0062] It should be noted that different feature sets may need to be extracted for different fault types. For example, for mechanical wear and aging, more attention is paid to the change of the RMS value of the vibration signal; for manufacturing and assembly errors, more attention is paid to the symmetry and periodicity of the waveform.

[0063] Step S25, generating a fault vibration waveform diagram under each fault type based on the characteristic data of the main shaft under each fault type.

[0064] Based on the characteristic data of the spindle under each fault type, the fault vibration waveform diagram under each fault type is generated to convert the characteristic data into an intuitive visual representation, which helps to better understand and analyze the vibration characteristics of different fault types.

[0065] like Figure 4 As shown, based on the characteristic data of the main shaft under each fault type, the steps of generating the fault vibration waveform diagram under each fault type include:

[0066] Step S251, preprocessing the characteristic data is beneficial to improving the quality and reliability of the fault vibration waveform diagram formed subsequently.

[0067] In this embodiment, the step of preprocessing the feature data includes: data normalization, which is used to unify the feature data of different dimensions into the same scale range, usually using the minimum-maximum normalization (Min-Max Normalization) or Z-score normalization method. It also includes data interpolation, which is used to fill in missing values ​​using methods such as linear interpolation or spline interpolation to ensure the continuity of the waveform.

[0068] Step S252, generating a fault vibration waveform diagram based on the preprocessed characteristic data.

[0069] In this embodiment, a curve showing the change of vibration over time is drawn with time as the horizontal axis and the vibration amplitude as the vertical axis. The time domain signal is converted into the frequency domain through fast Fourier transform (FFT), with the horizontal axis representing the frequency and the vertical axis representing the amplitude or power spectrum density. A three-dimensional graph is generated using short-time Fourier transform (STFT) or wavelet transform to show the change of the vibration signal in the time and frequency dimensions.

[0070] In this embodiment, characteristic frequency points or abnormal peak values ​​are marked on the fault vibration waveform diagram to facilitate the location of fault characteristics.

[0071] It should be noted that the dynamic balance monitoring method further includes: after forming the fault vibration waveform diagram, before using a neural network model to train the fault vibration waveform diagram under each fault type one by one, performing a graphic enhancement operation on the fault vibration waveform diagram.

[0072] It should be noted that the fault vibration waveforms of each fault type have corresponding characteristics. For example, the fault vibration waveform of mechanical wear and aging is characterized by the fact that as time goes by, mechanical parts will wear and age, causing the amplitude in the vibration waveform to gradually increase and the frequency characteristics may shift.

[0073] The waveform characteristic is that the RMS (root mean square value) of the vibration signal continues to increase, and may increase by more than 10% compared with the new machine.

[0074] The characteristics of the fault vibration waveform of manufacturing and assembly errors are that the vibration caused by improper assembly is usually more obvious during the startup and shutdown process, and sudden peaks or irregular fluctuations may be seen in the waveform. The waveform characteristics are that this type of fault is usually manifested as a weakening of the symmetry and periodicity of the waveform, significant changes in the standard deviation and skewness coefficient, abnormal vibration at a specific frequency point, and sharp peaks will be shown in the waveform compared to the ideal dynamic balance state.

[0075] The characteristic of the vibration waveform of the external load uneven fault is that it shows periodic fluctuations. The waveform shows obvious instability, such as sudden changes in vibration amplitude. The waveform is characterized by a periodic fluctuation pattern that is different from normal operation. The amplitude and frequency may have obvious periodic changes.

[0076] like Figure 5 As shown, in step S3, a neural network model is used to train the fault vibration waveform graphs under each fault type one by one to obtain specific labels of the fault vibration waveform graphs under each fault type.

[0077] The neural network model can learn the characteristics of various fault types and has good generalization ability. The specific labels obtained through training provide a reference basis for subsequent comparison with test labels.

[0078] Specifically, the steps of using the neural network model to train the fault vibration waveform diagrams under each fault type one by one include:

[0079] Step S31, using a first convolutional neural network to process the fault vibration waveform diagram to obtain sample space feature data.

[0080] During the dynamic balancing monitoring process, spatial feature data can reflect the overall shape and trend of the fault vibration waveform, which helps to make the specific labels obtained through training more accurate.

[0081] In the embodiment of the present invention, a first convolutional neural network is processed by a convolutional neural network with a narrow channel and a deep layer to obtain a first matrix reflecting spatial information as the spatial feature data.

[0082] In this embodiment, the channels processed by the first convolutional neural network are narrow (for example, the number of channels is 32 or 64), and accordingly, the number of convolution kernels is small, which can reduce the amount of computational effort for image processing.

[0083] It should be noted that the deeper the layers of the deep convolutional network, that is, the more convolutions it undergoes, the easier it is for the gradients between layers to diverge, which makes it easier for errors to occur.

[0084] In this embodiment, spatial feature data is obtained through a convolutional neural network with a shortcut structure. The convolutional neural network with a shortcut structure reduces the error caused by network processing by making the input data have residual data, thereby optimizing the training effect.

[0085] Specifically, the first convolutional neural network can be a MobileNet V2 network. The MobileNet V2 network adopts a method of first increasing the dimension and then reducing the dimension, and sequentially performs the process of expansion, convolution feature extraction, and compression. The MobileNet V2 network is a lightweight network with narrow channels and deep layers, which can improve the network's processing speed for fault vibration waveforms. In other embodiments, the first convolutional neural network can also include ResNet 50.

[0086] Step S32: Use a second convolutional neural network to process the fault vibration waveform to obtain sample detail feature data.

[0087] The second convolutional neural network is used to process the fault vibration waveform to obtain sample detail feature data, which can capture the microscopic features and detail information in the fault vibration waveform, accurately identify and classify different types of spindle faults, such as mechanical wear and aging, manufacturing and assembly errors, and uneven external loads, thereby improving the accuracy and reliability of dynamic balancing monitoring.

[0088] In this embodiment, the second convolutional neural network adopts the VGG network structure. The VGG network is known for its simple and effective design, and is particularly suitable for extracting detailed features of fault vibration waveforms, such as peak information, amplitude changes, etc. The VGG network structure is shallower than the MobileNet V2 network, but has wider channels.

[0089] Specifically, each layer of the VGG network structure includes: convolution layer, batch normalization and activation function.

[0090] It should be noted that in practical applications, the step size and the size of the convolution kernel can be adjusted according to the requirements of computing speed and accuracy.

[0091] Step S33: Based on the initial weight, the sample space feature data and the sample detail feature data are integrated to obtain a specific label of the fault vibration waveform diagram.

[0092] By fusing sample spatial feature data and sample detail feature data, the macroscopic structure and microscopic details of the fault vibration waveform can be captured simultaneously, thereby improving the accuracy and robustness of fault identification.

[0093] In this embodiment, the initial weight determines the relative importance of the sample space feature data and the sample detail feature data in the fusion process. The initial weight can be set based on prior knowledge or can be randomly initialized and then optimized through a training process.

[0094] In this embodiment, a weighted summation or a more complex fusion algorithm (such as an attention mechanism) is used to combine the sample space feature data and the sample detail feature data. As an example, the specific steps can be expressed as:

[0095] Fusion feature = w1*sample spatial feature data + w2*sample detail feature data (1)

[0096] Wherein, w1 in formula (1) is the initial weight of the sample space feature data, w2 is the initial weight of the sample detail feature data, and both w1 and w2 are greater than 0 and less than 1.

[0097] In this embodiment, the fused features are input into the classifier to generate a specific label of the fault vibration waveform, which is usually a probability distribution, indicating the possibility that the waveform belongs to each fault type.

[0098] Specifically, the classifier uses a fully connected layer and a softmax function in combination to convert the fused features into specific labels of the fault vibration waveform image.

[0099] Step S4, providing a test vibration waveform diagram of the spindle of the semiconductor device whose status is to be confirmed in real time.

[0100] Provide the test vibration waveform of the spindle of the semiconductor device in the state to be confirmed in real time, so as to timely capture the real-time vibration state of the spindle in the state to be confirmed, reflect the dynamic balance of the spindle of the semiconductor device, and prepare for the subsequent use of the neural network model to train the test vibration waveform and obtain the test label.

[0101] It should be noted that the step of providing a test vibration waveform diagram of a spindle of a semiconductor device in a state to be confirmed is the same as the step of obtaining a fault vibration waveform diagram of the spindle under a fault type.

[0102] Specifically, the step of providing a test vibration waveform diagram of the spindle of a semiconductor device in a state to be confirmed includes: setting a sensor on the spindle in the state to be confirmed; collecting vibration data of the spindle in the state to be confirmed through the sensor; performing time domain and frequency domain analysis on the vibration data of the spindle in the state to be confirmed, respectively, to obtain initial characteristic data of the spindle in the state to be confirmed; removing redundant parameters in the initial characteristic data of the spindle in the state to be confirmed, to obtain characteristic data of the spindle in the state to be confirmed; and generating a test vibration waveform diagram of the spindle in the state to be confirmed based on the characteristic data of the spindle in the state to be confirmed.

[0103] It should be noted that the test vibration waveform of the spindle of the semiconductor device whose status is to be confirmed is provided to the server in real time, and the operator observes the test vibration waveform in real time according to the visual operation interface of the server.

[0104] like Figure 6As shown, in step S5, the test vibration waveform diagram is trained using the neural network model to obtain a test label.

[0105] The neural network model is used to train the test vibration waveform diagram to convert the real-time collected vibration data into a comparable form, so as to subsequently compare the test label with the specific label under each fault type to obtain the working state of the spindle to be confirmed.

[0106] It should be noted that the neural network model and training steps for training the test vibration waveform diagram are the same as the neural network model and training steps for training the fault vibration waveform diagram under the fault type, and will not be repeated here.

[0107] It should be noted that the training of the test vibration waveform diagram using the neural network model is also carried out on the server, and the operator obtains the test label according to the visual operation interface of the server.

[0108] like Figure 7 As shown, in step S6, the test tag is compared with the specific tag under each fault type to obtain the working state of the spindle to be confirmed.

[0109] In an embodiment of the present invention, for a test vibration waveform diagram of a spindle of a semiconductor device in a state to be confirmed, the neural network model is used to train the test vibration waveform diagram to obtain a test label, and the test label is compared with specific labels under various known fault types. If the test label matches a characteristic label of a certain fault type, it can be determined that the spindle to be confirmed has the same fault type; if the test label does not match all known fault labels, it can be considered that the spindle to be confirmed is working normally. The dynamic balancing monitoring method provided by an embodiment of the present invention can monitor the dynamic balancing state of the spindle of a semiconductor device in real time, effectively avoiding blade swing caused by excessive dynamic balancing vibration amplitude of the spindle, thereby preventing product quality problems such as off-cutting, edge collapse, and incomplete cutting.

[0110] In this embodiment, the step of comparing the test tag with the specific tag under each fault type to obtain the working status of the spindle to be confirmed includes: using the nearest neighbor method to compare the test tag with the specific tags under multiple fault types to obtain the working status of the spindle to be confirmed.

[0111] Specifically, the specific steps of using the nearest neighbor method to compare the specific label and the test label are as follows:

[0112] Step S61, using the nearest neighbor algorithm to calculate the distance between the test tag and the specific tag under each fault type. In the subsequent process, the distance between the test tag and the specific tag under each fault type is compared with a preset threshold.

[0113] It should be noted that the "distance" here can be understood as the similarity between the test tag and the specific tag. In this embodiment, the "distance" calculation can adopt Euclidean distance, Manhattan distance or other suitable measurement methods.

[0114] Step S62, setting a preset threshold value of the distance between the specific tag and the test tag under each fault type.

[0115] In this embodiment, a suitable preset threshold value can be set for each fault type based on a large amount of experimental data and experience. This threshold value should be able to effectively distinguish between a normal state and a fault state, while taking into account the characteristic differences of different fault types.

[0116] In step S63, the test tag is compared with the specific tags of all known fault types one by one to find the specific tag that is closest to the test tag. When the distance between the nearest specific tag and the test tag is less than a preset threshold, it is determined that the corresponding fault type exists in the spindle to be confirmed; when the distance between the specific tag and the test tag under all fault types is greater than the preset threshold, it can be determined that the working status of the spindle to be confirmed is normal.

[0117] By comparing the test tag with the specific tag of the known fault type, the spindle status can be accurately diagnosed, which can not only identify whether the spindle has a fault, but also further determine the specific type of fault, providing an important basis for subsequent maintenance and repair work.

[0118] It should be noted that if the working state of the spindle to be confirmed is normal, it is necessary to further verify whether the test vibration waveform is normal based on the stability trend index (Stability Trend Index), volatility index (Volatility Index), outlier frequency index (OutlierFrequency Index) and change rate index (Change Rate Index). When the test vibration waveform has no abnormality after verification of these four indicators, it is considered that the working state of the spindle to be confirmed is normal; when the test vibration waveform has corresponding abnormal conditions after verification of these four indicators, the retest label is compared with the specific label of the known fault type to confirm the fault type of the spindle to be confirmed.

[0119] Specifically, stability trend indicators are usually used to measure the stability and trend of a system or process; volatility indicators are used to measure the degree of volatility of the market or data; outlier frequency indicators focus on the frequency of outliers in the data, where outliers refer to data points that deviate significantly from other observations; and rate of change indicators measure the speed at which data changes over time.

[0120] It should also be noted that the comparison of the test tag with the specific tag under each fault type is also performed on the server. The operator compares the test tag with the specific tag under each fault type according to the visual operation interface of the server.

[0121] like Figure 8 As shown, an embodiment of the present invention further provides a dynamic balance monitoring system for monitoring the dynamic balance of a spindle of a semiconductor device, including: a fault sample module 100, for providing samples of faulty spindles of multiple semiconductor devices, each faulty spindle having a different fault type, and multiple samples of faulty spindles under each fault type; a fault waveform acquisition module 200, for acquiring a fault vibration waveform of the spindle under each fault type; a specific label acquisition module 300, for training the fault vibration waveform under each fault type one by one using a neural network model to obtain a specific label of the fault vibration waveform under each fault type; a test vibration waveform module 400, for providing a test vibration waveform of the spindle of a semiconductor device in a state to be confirmed in real time; a test label acquisition module 500, for training the test vibration waveform using the neural network model to obtain a test label; a comparison module 600, for comparing the test label with the specific label of each fault type to obtain the working state of the spindle in a state to be confirmed.

[0122] The fault sample module 100 is used to provide faulty spindle samples of a plurality of semiconductor devices, each faulty spindle has a different fault type, and each fault type has a plurality of faulty spindle samples.

[0123] The fault sample module 100 provides samples of faulty spindles of multiple semiconductor devices, which provides the necessary data basis for the subsequent acquisition of fault vibration waveforms of the spindles under various fault types and the training of neural network models. The fault types of each faulty spindle are different, which is conducive to covering a variety of fault types, so that the monitoring method has stronger generalization ability and higher accuracy. In addition, there are multiple samples of faulty spindles under each fault type, which enhances the reliability and robustness of dynamic balancing monitoring.

[0124] In this embodiment, the fault types include: mechanical wear and aging, manufacturing and assembly errors, and uneven external loads.

[0125] Classifying fault types into mechanical wear and aging, manufacturing and assembly errors, and uneven external loads provides a clear goal for the subsequent training of the neural network model, enabling the neural network model to specifically learn the characteristics of different types of faults, which helps to improve the accuracy and efficiency of fault diagnosis.

[0126] In this embodiment, the number of faulty main axes under each fault type is greater than or equal to 5.

[0127] The number of faulty spindles under each fault type is greater than or equal to 5, so that each fault type has enough samples for training, so that the specific labels of each fault type obtained later are accurate and reliable. And the number of faulty spindles under each fault type is greater than or equal to 5. This is conducive to better distinguishing the specific labels of different fault types and reducing the subsequent misjudgment of the working status of the spindle to be confirmed.

[0128] It should be noted that the semiconductor device in this embodiment is a cutting machine, which is used to cut the processed wafer into individual chips.

[0129] Specifically, Figure 2 As shown, the cutting machine includes: a main shaft 1, which is placed horizontally; a blade 2, which is installed on the main shaft of the cutting machine; a cutting disk 3, which can move in the horizontal (x) and vertical (y) directions relative to the main shaft; and a wafer 4, which is placed on the cutting disk.

[0130] When the cutting machine is working, the wafer 4 is first placed on the cutting disk 3, and then the spindle 1 rotates, and the cutting disk 3 moves horizontally (x) relative to the spindle 1, so that the cutting path on the wafer 4 is aligned with the position of the blade, and the cutting disk 3 drives the wafer 4 to move in the vertical (y) direction, approaching the blade 2, and the cutting operation of the wafer 4 is performed.

[0131] The fault waveform acquisition module 200 is used to acquire the fault vibration waveform of the main shaft under various fault types.

[0132] The fault waveform acquisition module 200 acquires the fault vibration waveform of the main shaft under each fault type, and provides basic data for the subsequent specific label acquisition module 300 to train the fault vibration waveform under each fault type one by one using a neural network model.

[0133] like Fig. 9 As shown, the fault waveform acquisition module 200 includes: a sensor setting unit 201, which sets sensors on various fault main shafts.

[0134] The sensor is used to obtain the vibration data of the spindle under various fault types, preparing for the subsequent fault waveform acquisition module 200 to obtain the fault vibration waveform under each fault type; and based on the vibration data of the spindle under each fault type, a fault database can be established to improve the accuracy and reliability of dynamic balancing monitoring.

[0135] In this embodiment, the sensor includes one or more of a velocity sensor, an acceleration sensor and a displacement sensor.

[0136] The speed (VEL) mm / s data can reflect the motion state of the spindle, and the acceleration (ACC) m / s 2The data can capture sudden changes and shocks, while the displacement (DISP) mm data can show the actual position change of the spindle.

[0137] In this embodiment, a velocity sensor, an acceleration sensor and a displacement sensor are simultaneously arranged on the main shaft.

[0138] By simultaneously collecting the motion state of the spindle, sudden changes and impacts during operation, and the actual position changes of the spindle, more comprehensive fault diagnosis information can be provided, which helps to improve the accuracy and reliability of the dynamic balancing monitoring system.

[0139] It should be noted that the sensor is connected to a plurality of acquisition cards, and each acquisition card is arranged at a different position. By arranging acquisition cards at different positions, data at different positions can be collected to more comprehensively reflect the motion state of the faulty spindle.

[0140] Specifically, the number of the acquisition cards is eight. In other embodiments, the number of the acquisition cards may also be five, six, seven, or more than eight.

[0141] The fault waveform acquisition module 200 further includes: a signal acquisition unit 202, which is used to collect the vibration data of the main shaft under various fault types through sensors.

[0142] The signal acquisition unit 202 acquires the vibration data of the main shaft under various fault types, which is used for subsequent time domain and frequency domain analysis by the time domain and frequency domain analysis unit 203 .

[0143] In this embodiment, the signal acquisition unit 202 collects vibration data of multiple positions of the faulty spindle in parallel through sensors and multiple acquisition cards.

[0144] The signal acquisition unit 202 collects the vibration data of the faulty spindle in parallel through multiple acquisition cards, and can obtain vibration information of multiple measurement points at the same time, thereby improving the efficiency and comprehensiveness of data acquisition; moreover, parallel acquisition can ensure the synchronization of data at different measurement points, providing a reliable data basis for subsequent time domain and frequency domain analysis.

[0145] Specifically, multiple acquisition cards are used to connect sensors to achieve multi-channel acquisition, which can simultaneously process signal inputs from multiple sensors to ensure data synchronization and consistency. In addition, the parallel acquisition method can greatly improve the efficiency of data acquisition, especially when a large amount of sample data needs to be collected.

[0146] The time domain and frequency domain analysis unit 203 is used to perform time domain and frequency domain analysis on the vibration data of the main shaft under each fault type, so as to obtain initial characteristic data under each fault type.

[0147] The time domain analysis performed by the time domain and frequency domain analysis unit 203 can reflect the time-varying characteristics of the vibration signal, and the frequency domain analysis performed by the time domain and frequency domain analysis unit 203 can reveal the frequency composition of the signal. At the same time, time domain analysis and frequency domain analysis of the vibration data of the main shaft under each fault type can provide a more comprehensive and accurate description of the fault characteristics.

[0148] In this embodiment, the time domain analysis includes but is not limited to performing time domain waveform analysis and calculating time domain high-order statistics. Specifically, the time domain waveform analysis is used to observe the amplitude change trend of the vibration data. Calculating time domain high-order statistics includes calculating skewness, kurtosis, etc. These parameters are particularly effective for detecting nonlinear and non-stationary processes.

[0149] In this embodiment, frequency domain analysis uses fast Fourier transform (FFT) to convert time domain signals into frequency domain signals, which is the basis of frequency domain analysis and can decompose vibration data into components of different frequencies. The converted frequency domain signals are analyzed to identify the main frequency components. By analyzing the frequency components of the converted frequency domain signals, the type and severity of the fault can be preliminarily determined.

[0150] The characteristic data unit 204 is used to remove redundant parameters in the initial characteristic data under each fault type, and obtain the characteristic data of the main shaft under each fault type.

[0151] The feature data unit 204 is used to remove redundant parameters in the initial feature data to improve the efficiency and accuracy of subsequent data processing and provide more effective feature data for subsequent neural network model training. Specifically, removing redundant parameters in the initial feature data is conducive to reducing data noise and improving the signal-to-noise ratio of the signal; reducing computational complexity and speeding up the subsequent neural network model training; avoiding overfitting and improving the generalization ability of the model; highlighting key features, so that the subsequent neural network model can more easily capture the specific labels of the fault vibration waveform diagram.

[0152] In this embodiment, the feature data unit 204 uses the principal component analysis (PCA) dimensionality reduction method to remove redundant parameters in the initial feature data under each fault type, and obtain the feature data of the main shaft under each fault type. The principal component analysis dimensionality reduction method can identify and retain the most valuable features for fault diagnosis, and can convert high-dimensional initial feature data into low-dimensional principal components, retaining the most important variation information in the data, which is particularly useful for processing complex vibration data.

[0153] It should be noted that different feature sets may need to be extracted for different fault types. For example, for mechanical wear and aging, more attention is paid to the change of the RMS value of the vibration signal; for manufacturing and assembly errors, more attention is paid to the symmetry and periodicity of the waveform.

[0154] The fault vibration waveform diagram unit 205 is used to generate a fault vibration waveform diagram under each fault type based on the characteristic data of the main shaft under each fault type.

[0155] The fault vibration waveform diagram unit 205 is used to convert the characteristic data into an intuitive visual representation, which helps to better understand and analyze the vibration characteristics of different fault types.

[0156] like Fig.10 As shown, the fault vibration waveform diagram unit 205 includes: a preprocessing component 2051, which is used to preprocess the characteristic data, which is beneficial to improving the quality and reliability of the fault vibration waveform diagram formed subsequently.

[0157] In this embodiment, the step of preprocessing the feature data includes: data normalization, which is used to unify the feature data of different dimensions into the same scale range, usually using the minimum-maximum normalization (Min-Max Normalization) or Z-score normalization method. It also includes data interpolation, which is used to fill in missing values ​​using methods such as linear interpolation or spline interpolation to ensure the continuity of the waveform.

[0158] The fault vibration waveform diagram unit 205 includes a diagram generation component 2052, which generates a fault vibration waveform diagram according to the pre-processed characteristic data.

[0159] In this embodiment, the graphic generation component 2052 uses time as the horizontal axis and the vibration amplitude as the vertical axis to draw a curve of vibration changes over time; the time domain signal is converted into the frequency domain through fast Fourier transform (FFT), with the horizontal axis representing frequency and the vertical axis representing amplitude or power spectral density; and a short-time Fourier transform (STFT) or wavelet transform is used to generate a three-dimensional graph to show the changes of the vibration signal in the time and frequency dimensions.

[0160] In this embodiment, characteristic frequency points or abnormal peak values ​​are marked on the fault vibration waveform diagram to facilitate the location of fault characteristics.

[0161] It should be noted that the fault vibration waveform unit 205 also includes an enhancement component for performing a graphic enhancement operation on the fault vibration waveform after the fault vibration waveform is formed and before training the fault vibration waveform under each fault type one by one using a neural network model.

[0162] It should be noted that the fault vibration waveforms under each fault type have corresponding characteristics. For example, the characteristic of the fault vibration waveform of mechanical wear and aging is that as time goes by, mechanical parts will wear and age, causing the amplitude in the vibration waveform to gradually increase, and the frequency characteristics may shift. The waveform characteristic is that the RMS (root mean square value) of the vibration signal continues to increase, and may increase by more than 10% compared with a new machine.

[0163] The characteristics of the fault vibration waveform of manufacturing and assembly errors are that the vibration caused by improper assembly is usually more obvious during the startup and shutdown process, and sudden peaks or irregular fluctuations may be seen in the waveform. The waveform characteristics are that this type of fault is usually manifested as a weakening of the symmetry and periodicity of the waveform, significant changes in the standard deviation and skewness coefficient, abnormal vibration at a specific frequency point, and a sharp peak will be shown in the fault vibration waveform compared to the ideal dynamic balance state.

[0164] The fault vibration waveform of the external load unevenness is characterized by periodic fluctuations. The fault vibration waveform shows obvious instability, such as sudden changes in vibration amplitude. The waveform is characterized by a periodic fluctuation pattern that is different from normal operation, and the amplitude and frequency may have obvious periodic changes.

[0165] The specific label acquisition module 300 is used to train the fault vibration waveform graphs under each fault type one by one using a neural network model to obtain specific labels of the fault vibration waveform graphs under each fault type.

[0166] The specific label acquisition module 300 can learn the characteristics of various fault types through the neural network model and has good generalization. The specific labels obtained through training provide a reference basis for subsequent comparison with test labels.

[0167] like Fig.11 As shown, the specific label acquisition module 300 includes: a sample space feature unit 301, which is used to process the fault vibration waveform diagram using a first convolutional neural network to obtain sample space feature data.

[0168] During the dynamic balancing monitoring process, spatial feature data can reflect the overall shape and trend of the fault vibration waveform, which helps to make the specific labels obtained through training more accurate.

[0169] In the embodiment of the present invention, a first convolutional neural network is processed by a convolutional neural network with a narrow channel and a deep layer to obtain a first matrix reflecting spatial information as the spatial feature data.

[0170] In this embodiment, the channel processed by the first convolutional neural network is narrow (for example, Channel is 32 or 64), and accordingly, the number of convolution kernels is small, which can reduce the amount of computational effort for image processing.

[0171] It should be noted that the deeper the layers of the deep convolutional network, that is, the more convolutions it undergoes, the easier it is for the gradients between layers to diverge, which can easily lead to errors.

[0172] In this embodiment, spatial feature data is obtained through a convolutional neural network with a shortcut structure. The convolutional neural network with a shortcut structure reduces the error caused by network processing by making the input data into data with residuals, thereby optimizing the training effect.

[0173] Specifically, the first convolutional neural network can be a MobileNet V2 network. The MobileNet V2 network adopts a method of first increasing the dimension and then reducing the dimension, and sequentially performs the process of expansion, convolution feature extraction, and compression. The MobileNet V2 network is a lightweight network with narrow channels and deep layers, which can improve the network's processing speed for fault vibration waveforms. In other embodiments, the first convolutional neural network can also include ResNet 50.

[0174] The specific label acquisition module 300 also includes: a sample detail feature unit 302, which is used to process the fault vibration waveform diagram using a second convolutional neural network to obtain sample detail feature data.

[0175] The sample detail feature unit 302 uses a second convolutional neural network to process the fault vibration waveform diagram to obtain sample detail feature data, which can capture the microscopic features and detail information in the fault vibration waveform diagram, accurately identify and classify different types of spindle faults, such as mechanical wear and aging, manufacturing and assembly errors, and uneven external loads, thereby improving the accuracy and reliability of dynamic balancing monitoring.

[0176] In this embodiment, the second convolutional neural network adopts the VGG network structure. The VGG network is known for its simple and effective design, and is particularly suitable for extracting detailed features of fault vibration waveforms, such as peak information, amplitude changes, etc. The VGG network structure is shallower than the MobileNet V2 network, but has wider channels.

[0177] Specifically, each layer of the VGG network structure includes: convolution layer, batch normalization and activation function.

[0178] It should be noted that in practical applications, the step size and the size of the convolution kernel can be adjusted according to the requirements of computing speed and accuracy.

[0179] The specific label unit 303 is used to fuse the sample space feature data and the sample detail feature data based on the initial weight to obtain a specific label of the fault vibration waveform diagram.

[0180] By fusing sample spatial feature data and sample detail feature data, the macroscopic structure and microscopic details of the fault vibration waveform can be captured simultaneously, thereby improving the accuracy and robustness of fault identification.

[0181] In this embodiment, the initial weight determines the relative importance of the sample space feature data and the sample detail feature data in the fusion process. The initial weight can be set based on prior knowledge or can be randomly initialized and then optimized through a training process.

[0182] In this embodiment, a weighted summation or a more complex fusion algorithm (such as an attention mechanism) is used to combine the sample space feature data and the sample detail feature data. As an example, the specific steps can be expressed as:

[0183] Fusion feature = w1*sample spatial feature data + w2*sample detail feature data (1)

[0184] Wherein, w1 in formula (1) is the initial weight of the sample space feature data, w2 is the initial weight of the sample detail feature data, and both w1 and w2 are greater than 0 and less than 1.

[0185] In this embodiment, the fused features are input into the classifier to generate a specific label of the fault vibration waveform, which is usually a probability distribution, indicating the possibility that the waveform belongs to each fault type.

[0186] Specifically, the classifier uses a fully connected layer and a softmax function in combination to convert the fused features into specific labels of the fault vibration waveform image.

[0187] The test vibration waveform diagram module 400 is used to provide a test vibration waveform diagram of a spindle of a semiconductor device in a state to be confirmed in real time.

[0188] Provide the test vibration waveform of the spindle of the semiconductor device in the state to be confirmed in real time, so as to timely capture the real-time vibration state of the spindle in the state to be confirmed, reflect the dynamic balance of the spindle of the semiconductor device, and prepare for the subsequent use of the neural network model to train the test vibration waveform and obtain the test label.

[0189] It should be noted that the steps executed by the fault waveform acquisition module 200 and the test vibration waveform module 400 are the same, the only difference is that the input data of the fault waveform acquisition module 200 is the vibration data of the spindle under each fault type, and the input data of the test vibration waveform module 400 is the spindle vibration data of the semiconductor device in the confirmed state.

[0190] Specifically, the test vibration waveform diagram module 400 includes: a module for setting a sensor on the spindle in the state to be confirmed; a module for collecting vibration data of the spindle in the state to be confirmed through the sensor; a module for performing time domain and frequency domain analysis on the vibration data of the spindle in the state to be confirmed, respectively, to obtain initial characteristic data of the spindle in the state to be confirmed; a module for removing redundant parameters in the initial characteristic data of the spindle in the state to be confirmed, to obtain characteristic data of the spindle in the state to be confirmed; and a module for generating a test vibration waveform diagram of the spindle in the state to be confirmed based on the characteristic data of the spindle in the state to be confirmed.

[0191] It should be noted that the test vibration waveform of the spindle of the semiconductor device whose status is to be confirmed is provided to the server in real time, and the operator observes the test vibration waveform in real time according to the visual operation interface of the server.

[0192] The test label acquisition module 500 is used to train the test vibration waveform diagram using the neural network model to obtain a test label.

[0193] The test label acquisition module 500 uses the neural network model to train the test vibration waveform diagram, so as to convert the vibration data collected in real time into a comparable form, so as to subsequently compare the test label with the specific label under each fault type to obtain the working status of the spindle to be confirmed.

[0194] It should be noted that the steps executed by the specific label acquisition module 300 and the test label acquisition module 500 are the same, the only difference is that the specific label acquisition module 300 inputs the fault vibration waveform diagram of the spindle under each fault type, and the test label acquisition module 500 inputs the test vibration waveform diagram of the spindle of the semiconductor device in the state to be confirmed.

[0195] It should be noted that the training of the test vibration waveform diagram using the neural network model is also carried out on the server, and the operator obtains the test label according to the visual operation interface of the server.

[0196] The comparison module 600 is used to compare the test tag with the specific tag under each fault type to obtain the working state of the spindle to be confirmed.

[0197] The dynamic balancing monitoring system provided by the embodiment of the present invention uses the neural network model to train the test vibration waveform diagram to obtain a test label, and compares the test label with specific labels under various known fault types. If the test label matches the characteristic label of a certain fault type, it can be determined that the spindle to be confirmed has the same fault type; if the test label does not match all known fault labels, it can be considered that the spindle to be confirmed is working normally. The dynamic balancing monitoring system provided by the embodiment of the present invention can monitor the dynamic balancing state of the spindle of semiconductor equipment in real time, effectively avoiding blade swing caused by excessive vibration amplitude of the spindle dynamic balancing, thereby preventing product quality problems such as off-cut, edge collapse, and incomplete cutting.

[0198] In this embodiment, the comparison module 600 is used to compare the test tag with specific tags under multiple fault types by using the nearest neighbor method to obtain the working state of the spindle to be confirmed.

[0199] like Fig.12 As shown, the comparison module 600 includes: a distance acquisition unit 601, which is used to calculate the distance between the test label and the specific label under each fault type using the nearest neighbor algorithm. In the subsequent process, the distance between the test label and the specific label under each fault type is compared with a preset threshold.

[0200] It should be noted that the "distance" here can be understood as the similarity between the test tag and the specific tag. In this embodiment, the "distance" calculation can adopt Euclidean distance, Manhattan distance or other suitable measurement methods.

[0201] The comparison module 600 further includes: a preset threshold unit 602, which is used to set a preset threshold of the distance between the specific tag and the test tag under each fault type.

[0202] In this embodiment, a suitable preset threshold value can be set for each fault type based on a large amount of experimental data and experience. This threshold value should be able to effectively distinguish between a normal state and a fault state, while taking into account the characteristic differences of different fault types.

[0203] The comparison module 600 also includes: a comparison unit 603, which is used to compare the test tag with the specific tags of all known fault types one by one to find the specific tag that is closest to the test tag. When the distance between the nearest specific tag and the test tag is less than a preset threshold, it is determined that the corresponding fault type exists in the spindle to be confirmed; when the distance between the specific tag and the test tag under all fault types is greater than the preset threshold, it can be determined that the working status of the spindle to be confirmed is normal.

[0204] By comparing the test tag with the specific tag of the known fault type, the spindle status can be accurately diagnosed, which can not only identify whether the spindle has a fault, but also further determine the specific type of fault, providing an important basis for subsequent maintenance and repair work.

[0205] It should be noted that if the working state of the spindle to be confirmed is normal, it is necessary to further verify whether the test vibration waveform is normal based on the stability trend index (Stability Trend Index), volatility index (Volatility Index), outlier frequency index (OutlierFrequency Index) and change rate index (Change Rate Index). When the test vibration waveform has no abnormality after verification of these four indicators, it is considered that the working state of the spindle to be confirmed is normal; when the test vibration waveform has corresponding abnormal conditions after verification of these four indicators, the retest label is compared with the specific label of the known fault type to confirm the fault type of the spindle to be confirmed.

[0206] Specifically, stability trend indicators are usually used to measure the stability and trend of a system or process; volatility indicators are used to measure the degree of volatility of the market or data; outlier frequency indicators focus on the frequency of outliers in the data, where outliers refer to data points that deviate significantly from other observations; and rate of change indicators measure the speed at which data changes over time.

[0207] It should also be noted that the comparison of the test tag with the specific tag under each fault type is also performed on the server. The operator compares the test tag with the specific tag under each fault type according to the visual operation interface of the server.

[0208] Fig.13 It is a hardware structure diagram of an embodiment of an electronic device provided by the present invention.

[0209] An embodiment of the present invention further provides an electronic device, comprising at least one memory 03 and at least one processor 01, wherein the memory 03 stores one or more computer instructions, wherein the one or more computer instructions are executed by the processor 01 to implement the dynamic balance monitoring method of the embodiment of the present invention.

[0210] An optional hardware structure of the electronic device provided by the embodiment of the present invention can be as follows Fig.13 As shown, it includes: at least one processor 01, at least one communication interface 02, at least one memory 03 and at least one communication bus 04.

[0211] In this embodiment, the number of the processor 01 , the communication interface 02 , the memory 03 , and the communication bus 04 is at least one, and the processor 01 , the communication interface 02 , and the memory 03 communicate with each other through the communication bus 04 .

[0212] The communication interface 02 may be an interface of a communication module for network communication, such as an interface of a GSM module.

[0213] The processor 01 may be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the dynamic balancing monitoring method according to the embodiment of the present invention.

[0214] The memory 03 may include a high-speed RAM memory, and may also include a non-volatile memory (NVM), such as at least one disk memory.

[0215] The memory 03 stores one or more computer instructions, and the one or more computer instructions are executed by the processor 01 to implement the dynamic balance monitoring method of the embodiment of the present invention.

[0216] It should be noted that the above-mentioned electronic device may also include other devices (not shown) that may not be necessary for understanding the contents disclosed in the embodiments of the present invention; since these other devices may not be necessary for understanding the contents disclosed in the embodiments of the present invention, the embodiments of the present invention will not introduce them one by one.

[0217] An embodiment of the present invention further provides a storage medium, including a storage medium storing one or more computer instructions, wherein the one or more computer instructions are used to implement the dynamic balance monitoring method of the embodiment of the present invention.

[0218] An embodiment of the present invention further provides a computer program product, including computer instructions, which are used to implement the dynamic balance monitoring method of the embodiment of the present invention when executed by a processor.

[0219] The above-mentioned embodiments of the present invention are combinations of elements and features of the present invention. Unless otherwise mentioned, elements or features may be considered as optional. Each element or feature may be put into practice without being combined with other elements or features. In addition, embodiments of the present invention may be constructed by combining some elements and / or features. The order of operations described in the embodiments of the present invention may be rearranged. Some constructions of any one embodiment may be included in another embodiment, and may be replaced by the corresponding construction of another embodiment. It is obvious to those skilled in the art that claims that do not have a clear reference relationship to each other in the attached claims may be combined into embodiments of the present invention, or may be used as new claims in the amendment after submitting this application.

[0220] The embodiments of the present invention can be implemented by various means such as hardware, firmware, software or a combination thereof. In a hardware configuration, the method according to the exemplary embodiment of the present invention can be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), processors, controllers, microcontrollers, microprocessors, etc.

[0221] In a firmware or software configuration, the embodiments of the present invention may be implemented in the form of modules, procedures, functions, etc. The software code may be stored in a memory unit and executed by a processor. The memory unit is located inside or outside the processor and may send data to and receive data from the processor via various known means.

[0222] Although the present invention is disclosed as above, the present invention is not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the scope defined by the claims.

Claims

1. A dynamic balance monitoring method, characterized in that: Used to monitor the dynamic balance of the spindle of semiconductor equipment, including: Providing samples of faulty spindles of multiple semiconductor devices, each faulty spindle has a different fault type, and the number of faulty spindles under each fault type is multiple; Obtain the fault vibration waveform of the spindle under each fault type; A neural network model is used to train the fault vibration waveforms of each fault type one by one to obtain specific labels of the fault vibration waveforms of each fault type; Providing a test vibration waveform of the spindle of the semiconductor device whose status is to be confirmed in real time; Using the neural network model to train the test vibration waveform diagram to obtain a test label; The test tag is compared with the specific tag under each fault type to obtain the working state of the spindle to be confirmed.

2. The dynamic balance monitoring method according to claim 1, characterized in that: The steps of obtaining the fault vibration waveform of the spindle under each fault type include: Set up sensors on various fault spindles; The vibration data of the spindle under various fault types are collected through sensors; The vibration data of the main shaft under each fault type are analyzed in the time domain and frequency domain respectively to obtain the initial characteristic data under each fault type; Removing redundant parameters from the initial characteristic data under each fault type to obtain characteristic data of the spindle under each fault type; Based on the characteristic data of the main shaft under each fault type, a fault vibration waveform diagram under each fault type is generated.

3. The dynamic balance monitoring method according to claim 2, characterized in that: In the step of setting sensors on various faulty spindles, the sensors are connected to a plurality of acquisition cards, and the setting positions of the acquisition cards are different; In the step of collecting the vibration data of the main shaft under each fault type one by one through the sensor, the sensor collects the vibration data of the faulty main shaft in parallel through the multiple acquisition cards.

4. The dynamic balance monitoring method according to claim 2, characterized in that: In the step of setting sensors on various faulty spindles, the sensors include one or more of a velocity sensor, an acceleration sensor and a displacement sensor.

5. The dynamic balance monitoring method according to claim 2, characterized in that: The principal component analysis dimensionality reduction method is used to remove redundant parameters in the initial characteristic data under each fault type, so as to obtain the characteristic data of the main shaft under each fault type.

6. The dynamic balance monitoring method according to claim 2, characterized in that: Based on the characteristic data of the main shaft under each fault type, the steps of generating the fault vibration waveform diagram under each fault type include: preprocessing the characteristic data; generating the fault vibration waveform diagram according to the preprocessed characteristic data; The dynamic balance monitoring method further includes: after forming the fault vibration waveform diagram, before using a neural network model to train the fault vibration waveform diagrams under each fault type one by one, performing a graphic enhancement operation on the fault vibration waveform diagram.

7. The dynamic balance monitoring method according to claim 1, characterized in that: The steps of using the neural network model to train the fault vibration waveform diagrams under each fault type one by one include: Using a first convolutional neural network to process the fault vibration waveform diagram to obtain sample space feature data; Using a second convolutional neural network to process the fault vibration waveform diagram to obtain sample detail feature data; Based on the initial weight, the sample space feature data and the sample detail feature data are fused to obtain a specific label of the fault vibration waveform diagram.

8. The dynamic balance monitoring method according to claim 7, characterized in that: The first convolutional neural network is a network with a direct connection structure.

9. The dynamic balance monitoring method according to claim 7, characterized in that: The second convolutional neural network includes a VGG network, and the first convolutional neural network includes a MobileNet V2 network or a ResNet50.

10. The dynamic balance monitoring method according to claim 1, characterized in that: The steps of using the neural network model to train the test vibration waveform diagram and obtaining a test label include: Using a first convolutional neural network to process the test vibration waveform diagram to obtain test sample spatial feature data; Using a second convolutional neural network to process the test vibration waveform diagram to obtain detailed feature data of the test sample; Based on the initial weight, the test sample spatial feature data and the test sample detail feature data are fused to obtain the test label of the test vibration waveform diagram.

11. The dynamic balance monitoring method according to claim 1, characterized in that: The step of comparing the test tag with the specific tag under each fault type to obtain the working state of the spindle to be confirmed includes: The test tag is compared with specific tags under various fault types by using the nearest neighbor method to obtain the working state of the spindle to be confirmed.

12. The dynamic balance monitoring method according to claim 11, characterized in that: The steps of using the nearest neighbor method to compare the test tag with specific tags under multiple fault types to obtain the working state of the spindle to be confirmed include: The nearest neighbor algorithm is used to calculate the distance between the test label and the specific label under each fault type; Setting preset thresholds for the distance between a specific tag and a test tag under each fault type; Compare the test tag with the specific tags of all known fault types one by one to find the specific tag closest to the test tag. When the distance between the nearest specific tag and the test tag is less than the preset threshold, it is determined that the corresponding fault type exists in the spindle to be confirmed; when the distance between the specific tag and the test tag under all fault types is greater than the preset threshold, it can be determined that the working status of the spindle to be confirmed is normal.

13. The dynamic balance monitoring method according to claim 1, characterized in that: In the step of providing a plurality of failed spindle samples of semiconductor devices, the failure types include: mechanical wear and aging, manufacturing and assembly errors, and external load unevenness.

14. A dynamic balance monitoring system, characterized in that: Used to monitor the dynamic balance of the spindle of semiconductor equipment, including: A fault sample module is used to provide samples of faulty spindles of multiple semiconductor devices, each faulty spindle has a different fault type, and each fault type has multiple samples of faulty spindles; A fault waveform acquisition module is used to obtain the fault vibration waveform of the main shaft under various fault types; A specific label acquisition module is used to train the fault vibration waveform diagrams under each fault type one by one using a neural network model to obtain specific labels of the fault vibration waveform diagrams under each fault type; A test vibration waveform diagram module, used for providing a test vibration waveform diagram of a spindle of a semiconductor device whose status is to be confirmed in real time; A test label acquisition module, used to train the test vibration waveform diagram using the neural network model to obtain a test label; The comparison module is used to compare the test tag with the specific tag of each fault type to obtain the working state of the spindle to be confirmed.

15. An electronic device, characterized in that: It comprises at least one memory and at least one processor, wherein the memory stores one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the dynamic balance monitoring method according to any one of claims 1 to 13.

16. A storage medium, characterized in that: The storage medium stores one or more computer instructions, and the one or more computer instructions are used to implement the dynamic balance monitoring method according to any one of claims 1 to 13.

17. A computer program product, characterized in that It comprises computer instructions, which are used to implement the dynamic balance monitoring method as claimed in any one of claims 1 to 13 when executed by a processor.

Images