G-M refrigerator type cryopump fault diagnosis platform based on deep learning
By designing a deep learning-based fault diagnosis platform and combining it with the CNN-LSTM algorithm to analyze vibration signals, the problem of gas path contamination caused by mechanical wear of the GM refrigeration cryopump was solved, efficient and accurate fault diagnosis was achieved, and system stability and semiconductor manufacturing reliability were improved.
Patent Information
- Application Number
- CN202510920691.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-04
- Publication Date
- 2025-09-26
AI Technical Summary
Existing technologies are unable to provide early warning for GM refrigeration cryopumps, resulting in mechanical wear and failure leading to gas line contamination and reduced refrigeration efficiency, affecting the reliability and stability of high vacuum acquisition.
A deep learning-based fault diagnosis platform was designed, which included a cryogenic pump fault reproduction system, vibration sensors, data acquisition equipment, edge computing equipment, and a host computer. The CNN-LSTM algorithm was used to analyze vibration signals to achieve efficient and accurate fault diagnosis.
It achieves efficient and accurate diagnosis of cryogenic pump failures, improves system safety and stability, reduces production interruptions, reduces production costs, and ensures the quality and reliability of the semiconductor manufacturing process.
Smart Images
Figure CN120701558A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of vacuum technology for semiconductor process equipment, and in particular to a deep learning-based low-GM refrigerator-type cryopump fault diagnosis platform. Background Art
[0002] GM refrigerator cryopumps have been widely used in the manufacture of high-end equipment such as vacuum electronic device manufacturing, optoelectronic device preparation, optical thin film preparation, and solar panel preparation. It has become one of the core components of integrated circuit manufacturing process equipment such as PVD and ion implantation. Under normal circumstances, the cryopump system needs to run continuously around the clock and 24 hours a day. Therefore, in its daily operation, various faults are very likely to occur. Mechanical wear faults such as cold head eccentric wear, seal aging, and bearing wear can easily cause cryopump gas path contamination, thereby adversely affecting its refrigeration efficiency and pumping characteristics, and thus affecting the reliability of high vacuum acquisition and the stability of high vacuum maintenance. Among its many faults, the wear of moving mechanical parts in the GM refrigerator has attracted much attention from researchers. The existing technology can only take the approach of inspecting or maintaining the faulty cryopump after the fault occurs to solve the problem of cryopump failure, and cannot provide early warning to avoid accidents. The fault diagnosis method based on deep learning has the advantages of automation and high accuracy. A fault diagnosis method based on deep learning has been established. C The NN-LSTM deep learning combination algorithm characterizes and forms a theoretical model support for the fault diagnosis and detection scheme of the GM refrigerator cryogenic pump gas valve drive bearing. Summary of the Invention
[0003] The present invention provides a GM cryogenic pump fault diagnosis platform based on deep learning to solve the technical problems mentioned in the background technology.
[0004] A deep learning-based GM refrigeration machine cryopump fault diagnosis platform includes a cryopump fault reproduction system, a vibration sensor, a data acquisition device, an edge computing device, a switch, a data platform, and a host computer connected in sequence.
[0005] The fault reproduction system includes a cryogenic pump and replaceable parts, which are used to simulate the mechanical wear of the cryogenic pump parts during operation; the replaceable parts include: primary and secondary pistons, gas distribution valves, drive bearings, fixed valves, sealing seats and frame bearings;
[0006] Furthermore, the precise identification of component failures is achieved through quantitative processing of the faulty components. By setting a reasonable wear gradient, the wear degree of each component is gradually increased;
[0007] Furthermore, the inner ring of the bearing is subjected to different degrees of wear processing by turning to simulate the bearing wear state under actual use;
[0008] Furthermore, by using specially made tooling, the bearing is mounted on the customized tooling to effectively fix the end face of the bearing inner ring so that the turning process can be carried out smoothly;
[0009] Furthermore, the vibration sensor is a piezoelectric IEPE-type vibration sensor that utilizes the piezoelectric effect. The PZT ceramic piezoelectric material within it generates an electric charge when subjected to mechanical stress, converting vibration or acceleration into an electrical signal. This output signal is proportional to the perceived vibration acceleration, with a sensitivity of 100mV / g, effectively sensing weak vibration signals and ensuring accurate measurement. The vibration sensor is glued to the pump housing near the components to collect vibration signals from different parts.
[0010] Furthermore, the core of the data acquisition device is an eight-channel acquisition card that uses IEPE sensors to collect vibration signals. It uses a standard BNC interface to connect to the IEPE sensor and uses Ethernet to communicate with the host computer. The sampling rate is adjustable from 1 ksps to 200 ksps. The LabVIEW host computer system used in conjunction can achieve integrated acquisition, display, and storage. The edge computing device uses an eight-core 64-bit processor, specifically: a quad-core A76 and a quad-core A55, designed using an 8nm process, with a main frequency of up to 2.4 GHz, an integrated ARM Mali-G610 MP4 GPU, an embedded high-performance 3D and 2D image acceleration module, and a built-in AI accelerator NPU with a computing power of 6 TOPS. The reference interfaces include: HDMI output, GPIO interface, Type-C, and Gigabit network port.
[0011] Furthermore, the vibration signals of replaceable components such as the first and second stage pistons, valves, and drive bearings under working conditions are collected through acceleration sensors, and the vibration signals are preprocessed to extract the characteristic information of the cryogenic pump. The preprocessing includes slicing, data conversion, random disorder, normalization, etc.
[0012] Furthermore, the CNN-LSTM fault diagnosis algorithm is used in the edge computing device analysis process. The CNN-LSTM fault diagnosis algorithm adopts a time series feature analysis method that combines a convolutional neural network (CNN) and a long short-term memory network (LSTM). CNN extracts the spatial features and local features of the input data, and uses LSTM to process the long-term dependencies of the sequence data to achieve feature fusion in the spatiotemporal domain. This method combines the advantages of CNN in feature extraction and the expertise of LSTM in time series modeling, and can better adapt to the complexity of data with time series characteristics; it effectively captures key information and long-distance dependencies in the sequence, thereby reducing the impact of redundant information.
[0013] As a further technical solution of the present invention, the CNN-LSTM fault diagnosis algorithm realizes efficient extraction and fusion of spatiotemporal features through the collaborative work of convolutional neural networks (CNN) and long short-term memory networks (LSTM); based on the convolutional layer and pooling layer structure of CNN, the spatial features and local features of the input data are automatically extracted, and the long-term dependencies and temporal features in the sequence data are effectively processed through the gating mechanism (input gate, forget gate and output gate) of LSTM.
[0014] As a further technical solution of the present invention, the spatiotemporal feature extraction process of the CNN-LSTM fault diagnosis algorithm can be expressed as:
[0015] ;
[0016] Where: Represent the input data, and Respectively represent the convolution kernel weight and bias, represents the convolution operation, Representation activation function, Characterize the max pooling operation, Characterize the extracted spatial features.
[0017] As a further technical solution of the present invention, the LSTM time series modeling stage is expressed as:
[0018] ;
[0019] Where: Characterize the input features at time t (i.e. ), and Respectively represent the hidden state and unit state of the previous moment, 、 、 Represent the input gate, forget gate and output gate respectively, Characterize the sigmoid activation function, represents element-wise multiplication, and Characterize the trainable weight and bias parameters, Represents the candidate unit state at time t, represents the cell state at time t, Represents the hidden state at time t.
[0020] Beneficial effects achieved by the present invention:
[0021] This invention utilizes a cryopump fault reproduction system, a compressor system, a vacuum pumping system, and a cryopump diagnostic system, all working together to analyze the damage type of actual faulty components. Using processes such as turning, the system accurately captures the actual fault. Subsequently, quantitative machining with varying degrees of damage is performed to simulate real-world operating conditions. Furthermore, a vibration data acquisition and measurement platform is established to collect vibration signals from key components at varying degrees of wear, generating a dataset that reflects actual GM refrigerator cryopump faults. Together, these systems form a cryopump fault diagnosis system, providing data support for characterizing cryopump fault diagnosis. This system also develops a novel CNN-LSTM deep learning combination algorithm, enabling efficient and accurate diagnosis of cryopump faults. This invention not only improves the safety and stability of cryopumps but also significantly impacts the quality and reliability of semiconductor integrated circuit manufacturing processes. This invention helps semiconductor integrated circuit vacuum systems effectively maintain a constant vacuum level, reduces the risk of particulate matter and hazardous gases generated within the vacuum system, and ensures stable process parameters during operation, thereby improving product consistency and repeatability. This invention helps reduce production interruptions caused by system downtime or adjustments, lowering production costs and improving efficiency, ensuring smooth semiconductor manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] Figure 1 This is a structural schematic diagram of a GM refrigeration machine cryogenic pump fault diagnosis platform based on deep learning in the present invention.
[0023] Figure 2 This is a schematic diagram of the CNN-LSTM fault diagnosis algorithm constructed in this invention.
[0024] Figure 3 This is a flow chart of a GM refrigeration machine cryogenic pump fault diagnosis platform based on deep learning in the present invention.
[0025] Figure 4 This is the confusion matrix diagram of the CNN-LSTM fault diagnosis algorithm constructed in this invention.
[0026] Figure 5 This is a graph showing the loss function decline of the CNN-LSTM fault diagnosis algorithm constructed in the present invention. DETAILED DESCRIPTION
[0027] Characterization makes the purpose, technical solutions and advantages of the embodiments of the present invention clearer. The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0028] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.
[0029] See also Figure 1 , an embodiment of the present invention provides a GM refrigerant cryogenic pump fault diagnosis platform based on deep learning, the platform comprising a cryogenic pump fault reproduction system, a vibration sensor, a data acquisition device, an edge computing device, a switch, a data platform and a host computer connected in sequence;
[0030] The fault reproduction system includes a cryopump and replaceable components used to simulate the mechanical wear of the cryopump components during operation. The replaceable components include primary and secondary pistons, a distribution valve, a drive bearing, a fixed valve, a sealing seat, and a frame bearing. The cryopump is a GM refrigerator cryopump. In GM refrigerator cryopumps, a drive unit drives the primary piston to reciprocate within the primary cylinder, which then drives the secondary piston to operate within the secondary cylinder. The primary piston and the secondary piston are sealed against the cylinder by a sealing seat. The distribution valve is connected to the drive unit's rotating shaft to control the flow of helium in and out of the cylinder. The fixed valve is fixed to a pipe or chamber to control airflow at specific locations. The drive bearing supports the rotating shaft, and the frame bearing secures key components. During operation, the motor drives the piston to cause the helium to undergo inflation and deflation processes to complete the refrigeration cycle. The primary piston drives the secondary piston to further cool the gas. The valves, bearings, and sealing seats work together to ensure stable gas flow and reliable system operation.
[0031] The fault reproduction system simulates the actual fault condition of the cryogenic pump. The faulty parts are replaced first, and then the cryogenic pump goes through the stages of evacuation, cooling, and reheating. During operation, the vibration sensor collects vibration data, which is converted by the data acquisition equipment and transmitted to the host computer. The host computer uses software algorithms to deeply analyze the data and finally obtains the diagnosis results of the cryogenic pump operating status, which provides a reference for characterization and maintenance. Figure 2 As shown in the figure, this system acquires fault data for various key components, including primary and secondary pistons, valve trains, drive bearings, fixed valves, seal seats, and frame bearings, by replacing components with varying degrees of damage. This data not only covers detailed information on different types of cryopump failures, but also provides quantitative data for the same type of failure at varying degrees of severity. During the experiment, the system precisely controls the damage level of each component to ensure that the simulated failure conditions reflect the full range of scenarios that may occur in real-world operation. This provides reliable data support for characterizing cryopump fault diagnosis.
[0032] The precise emergence of component failures is achieved through quantitative processing of the faulty parts. The system uses a series of sophisticated processing technologies to characterize the wear state under effective simulation of actual use conditions. By setting a precise wear gradient, the degree of wear of each component is gradually increased. For example, the outer diameter of the first and second stage pistons, the inner ring of the drive bearing, the thickness of the fixed valve, the inner diameter of the sealing seat, and the inner and outer diameters and end faces of the frame bearing are processed to varying degrees using a turning process to ensure the accuracy and consistency of the processing. When processing the gas distribution valves, the system uses a plane milling process to perform different degrees of wear treatment on the valves to simulate their wear conditions in actual use. For bearing parts that are difficult to clamp, the system has specially made customized processing fixtures. By installing the bearings on these fixtures, the end face of the bearing inner ring is effectively fixed, allowing for smooth and precise processing operations.
[0033] This vibration sensor is a piezoelectric IEPE-type sensor that utilizes the piezoelectric effect. The PZT ceramic piezoelectric material within it generates an electric charge when subjected to mechanical stress, converting vibration or acceleration into an electrical signal. This output signal is proportional to the vibration acceleration sensed, and with a sensitivity of 100mV / g, it can effectively sense weak vibration signals and ensure accurate measurements.
[0034] The core of the data acquisition device is an eight-channel acquisition card that uses IEPE sensors to collect vibration signals. It uses a standard BNC interface to connect to the IEPE sensor and Ethernet to communicate with the host computer. The sampling rate is adjustable from 1ksps to 200ksps. The LabVIEW host computer system used in conjunction can realize integrated acquisition, display, and storage. The edge computing device uses an eight-core 64-bit processor, specifically: a quad-core A76 and a quad-core A55, designed using an 8nm process, with a main frequency of up to 2.4GHz, an integrated ARM Mali-G610 MP4 GPU, an embedded high-performance 3D and 2D image acceleration module, and a built-in AI accelerator NPU with a computing power of 6TOPS. The reference interfaces include: HDMI output, GPIO interface, Type-C, and Gigabit network port.
[0035] One end of the vibration sensor is glued to the cryopump casing, closest to the replaceable component, and the other end is connected to the BNC port of the data acquisition device to collect vibration signals. The data acquisition device, edge computing device, and host computer are connected to the switch via a network cable and are located on the same network segment. This allows the signals to be collected by the data acquisition device and sent to the edge computing device for storage and analysis. This calculates the signs and abnormal patterns that may lead to cryopump failures, and generates cryopump fault diagnosis results based on historical data. The final analysis results are presented to relevant personnel through the data platform, enabling human-computer interaction.
[0036] In this embodiment, vibration sensors are used to collect vibration signals of components such as the primary and secondary pistons, valves, and drive bearings in working conditions, and the original vibration signals are preprocessed. The preprocessing includes slicing, data conversion, random ordering, and normalization.
[0037] like Figure 2 As shown in the figure, the CNN-LSTM fault diagnosis algorithm is used in the edge computing device analysis process. The CNN-LSTM fault diagnosis algorithm adopts a time series feature analysis method that combines convolutional neural network (CNN) and long short-term memory network (LSTM). CNN is used to extract the spatial and local features of the input data, and LSTM is used to process the long-term dependencies of the sequence data to achieve feature fusion in the spatiotemporal domain. This method combines the advantages of CNN in feature extraction and the expertise of LSTM in time series modeling, and can better adapt to the complexity of data with time series characteristics. It can effectively capture key information and long-distance dependencies in the sequence, thereby reducing the impact of redundant information.
[0038] In this embodiment, the CNN-LSTM fault diagnosis algorithm realizes efficient extraction and fusion of spatiotemporal features through the collaborative work of convolutional neural networks (CNN) and long short-term memory networks (LSTM); based on the convolutional layer and pooling layer structure of CNN, the spatial features and local features of the input data are automatically extracted, and the gating mechanism (input gate, forget gate and output gate) of LSTM is used to effectively process the long-term dependencies and temporal features in the sequence data.
[0039] The spatiotemporal feature extraction process of the CNN-LSTM fault diagnosis algorithm can be expressed as:
[0040] CNN feature extraction stage:
[0041] ;
[0042] Where: Represent the input data, and Respectively represent the convolution kernel weight and bias, represents the convolution operation, Representation activation function, Characterize the max pooling operation, Characterize the extracted spatial features.
[0043] LSTM time series modeling stage:
[0044] ;
[0045] ;
[0046] ;
[0047] ;
[0048] ;
[0049] ;
[0050] Where: Characterize the input features at time t (i.e. ); and Respectively represent the hidden state and unit state of the previous moment; 、 、 Represent the input gate, forget gate and output gate respectively, Characterize the sigmoid activation function; represents element-wise multiplication; and Representing trainable weight and bias parameters; Represent the candidate unit state at time t; Represents the cell state at time t; Represents the hidden state at time t.
[0051] like Figure 3 As shown, the vibration sensor is used to systematically extract the vibration signals of the four components of the cryogenic pump, and the vibration signals of the normal working condition and the vibration signals of the faulty working condition are respectively characterized. After labeling, 3000 sets of measured experimental data are finally obtained as the input of the CNN-LSTM deep learning model / CNN-LSTM fault diagnosis algorithm, and the trained model is then implanted into the edge computing device. Finally, the data is sent to the edge computing device through the data acquisition device, and the trained model is used for fault diagnosis. In addition, the algorithm proposed in the present invention combines the local feature extraction ability of the convolutional neural network, the key information focusing ability of the attention mechanism, and the global parameter optimization ability of the particle swarm optimization algorithm to ensure that the accuracy of the deep learning model is above 99% after multiple rounds of training, as shown in FIG. Figures 4 and 5 As shown in the figure, training and testing output the results, and based on the results, corresponding maintenance strategies are formulated to reduce maintenance and operating costs, while also avoiding unplanned downtime caused by damage to cryogenic pumps in widely used semiconductor processes.
[0052] It should be noted that, in this document, the term "comprises" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.
[0053] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. A deep learning-based GM refrigeration cryopump fault diagnosis platform, comprising a cryopump fault reproduction system, a vibration sensor, a data acquisition device, an edge computing device, a switch, a data platform, and a host computer connected in sequence; The fault reproduction system includes a cryogenic pump and replaceable components, and is used to simulate the mechanical wear of the components during the operation of the cryogenic pump.
2. A GM refrigeration machine cryopump fault diagnosis platform based on deep learning according to claim 1, characterized in that: One end of the vibration sensor is pasted and fixed on the cryopump casing, and the other end is connected to the BNC interface of the data acquisition device to collect vibration signals; the data acquisition device, edge computing device, and host computer are connected to the switch via a network cable and are in the same network segment. The signal is collected by the data acquisition device and sent to the edge computing device for storage and analysis to generate the cryopump fault diagnosis results. The final diagnosis results are presented to relevant personnel through the data platform to realize human-computer interaction.
3. A GM refrigeration machine cryopump fault diagnosis platform based on deep learning according to claim 1, characterized in that: The vibration signals of replaceable components in working state are collected by vibration sensors, and the vibration signals are preprocessed to extract the characteristic information of the cryopump.
4. A GM refrigeration machine cryopump fault diagnosis platform based on deep learning according to claim 3, characterized in that: The CNN-LSTM fault diagnosis algorithm is used in the edge computing device analysis process. The CNN-LSTM fault diagnosis algorithm adopts a time series feature analysis method that combines convolutional neural network / CNN and long short-term memory network / LSTM. CNN extracts spatial features and local features of input data, and uses LSTM to process long-term dependencies of sequence data to achieve feature fusion in the spatiotemporal domain.
5. A GM refrigeration machine cryopump fault diagnosis platform based on deep learning according to claim 4, characterized in that: The feature extraction process of the CNN-LSTM fault diagnosis algorithm is expressed as: ； Where: Represent the input data, and are the convolution kernel weight and bias respectively, represents the convolution operation, is the activation function, is the maximum pooling operation, is the extracted spatial feature.
6. A GM refrigeration machine cryopump fault diagnosis platform based on deep learning according to claim 4, characterized in that: The LSTM timing modeling stage is expressed as: ； ； ； ； ； ； Where: Characterize the input features at time t, that is ; and Respectively represent the hidden state and unit state of the previous moment; 、 、 Represent the input gate, forget gate and output gate respectively; Characterize the sigmoid activation function; represents element-wise multiplication, and Respectively represent the trainable weight and bias parameters; Represents the candidate unit state at time t, represents the cell state at time t, Represents the hidden state at time t.