Low-temperature pump health monitoring platform based on deep learning algorithm

Abstract

The invention is suitable for the field of health monitoring of vacuum equipment in a semiconductor process, and provides a cryogenic pump health monitoring platform based on a deep learning algorithm, and the health monitoring platform comprises a cryogenic pump, a compressor system, a vacuum pumping system and a health monitoring system. The compressor system is composed of a helium compressor, a helium pipeline and an electric control unit. The vacuum pumping system is composed of a preceding-stage mechanical pump, a pneumatic valve and a corrugated pipe. The health monitoring system comprises an acceleration sensor, a data acquisition card, an upper computer, a data platform and a fault early warning system; measurement of vibration signals is completed by an acceleration sensor, data acquisition and uploading to an upper computer are achieved through a data acquisition card, and the upper computer carries out calculation and analysis after receiving the data. According to the method, production interruption caused by system fault shutdown or parameter adjustment can be reduced, the production efficiency is improved while the production cost is reduced, and powerful support is provided for continuous and efficient manufacturing of semiconductor integrated circuits.

Description

Technical Field

[0001] The present invention relates to the field of health monitoring and predictive maintenance of vacuum equipment in semiconductor processes, and particularly to a cryopump health monitoring platform based on deep learning algorithms. Background Art

[0002] The G-M refrigeration machine (Gifford-McMahon Refrigerator) type cryopump is a vacuum acquisition device that uses the ultra-low temperature surface (usually below 20K) generated by the refrigeration machine to condense, adsorb or trap gas molecules, thereby achieving a high vacuum environment. In the field of high-end equipment manufacturing, the G-M refrigeration machine cryopump has been deeply integrated into key links such as the manufacturing of vacuum electronic devices and the preparation of optoelectronic devices with its excellent ultra-high vacuum acquisition ability, and has become an indispensable part of core process equipment for integrated circuits such as PVD coating and ion implantation. Since this type of equipment needs to run continuously for 24 hours to ensure the stability of the production line, the long-term high-intensity work makes the cryopump system face severe reliability challenges. During operation, mechanical component wear problems such as abnormal eccentric wear of the cold head, aging of seals, and wear of bearings occur frequently, which not only leads to a significant decrease in refrigeration efficiency and pumping performance, but may also cause gas path pollution, directly threatening the acquisition accuracy and maintenance stability of the high vacuum environment. The current industry solutions are still limited to the passive maintenance mode after a failure occurs, lacking a precise capture and early warning mechanism for early failure characteristics. This lagging maintenance not only increases the downtime cost and the risk of production interruption, but also is difficult to meet the strict requirements for continuous and stable operation of equipment in fields such as semiconductor manufacturing. Developing technologies with health monitoring and intelligent early warning functions has become an urgent need to ensure the reliability of high-end equipment manufacturing processes. Summary of the Invention

[0003] The present invention provides a cryopump health monitoring platform based on deep learning algorithms, which can collect and monitor the vibration signals of the cryopump in real time to achieve predictive maintenance of the cryopump.

[0004] A cryopump health monitoring platform based on deep learning algorithms, the health monitoring platform includes: a cryopump, a compressor system, a vacuum pumping system, and a health monitoring system;

[0005] The cryopump is a two-stage G-M refrigeration machine type cryopump;

[0006] The compressor system is composed of a helium compressor, a helium pipeline, and an electronic control unit;

[0007] The vacuum pumping system is composed of a fore pump, a pneumatic valve, and a bellows;

[0008] The health monitoring system includes an acceleration sensor, a data acquisition card, a host computer, a data platform, and a fault warning system; the measurement of vibration signals is completed by the acceleration sensor, and data acquisition and uploading to the host computer are achieved through the data acquisition card. After receiving the data, the host computer performs calculation and analysis.

[0009] As a further technical solution of the present invention, the helium compressor is used to supply circulating helium to the cryogenic refrigerator system of the cryopump, so that it acts as a system refrigerant; the helium pipeline is a transmission channel for helium, maintaining the helium purity and ensuring the stability of the system sealing pressure.

[0010] As a further technical solution of the present invention, one end of the fore pump is connected to the cryopump through a pneumatic valve, and the other end is connected to the fore pump through a bellows; the pneumatic valve has the function of isolating the cryopump from the fore pump; the fore pump is used for rough vacuum extraction of the cryopump cavity to create a basic vacuum environment for the start-up of the cryopump.

[0011] As a further technical solution of the present invention, the acceleration sensor, the data acquisition card, and the host computer are physically connected in series; during the operation of the cryopump, the system monitors and records vibration signals by collecting data at different stages; the data collection at each operation stage corresponds to specific operating conditions and pump internal environment conditions; the acceleration sensor is installed on the pump housing surface close to each component to capture and record the vibration characteristic data during the operation of the cryopump; the data acquisition card is responsible for collecting the data obtained by the acceleration sensor and transmitting it to the data platform; the data platform stores the received data and conducts further analysis and processing; on the data platform, a fault warning system is run to analyze the collected data and identify potential fault characteristics and abnormal data patterns; by comparing the real-time data with the historical data and using pattern matching technology, the system generates a fault diagnosis report and prediction results for the cryopump; finally, the fault diagnosis report and prediction results are displayed on the front end to achieve health monitoring.

[0012] As a further technical solution of the present invention, the method for health monitoring of the cryopump based on the health monitoring platform includes the following steps:

[0013] First, a training data set is made. By replacing components with different damage degrees, vibration signal data of the components are obtained. The components include first and second stage pistons, gas distribution valves, drive bearings, fixed valves, seal seats, and frame bearings, etc.; the collected vibration signals not only cover detailed information of various types of faults of the cryopump, but also contain quantitative data of different damage degrees under the same fault type; during the experiment, control means are used to set the wear amount of each component to make the simulated fault conditions truly reflect the possible fault scenarios in the actual operation of the cryopump, thereby providing a reliable data basis for the research on cryopump fault diagnosis.

[0014] Fault simulation is realized by quantitatively processing parts to obtain accurate fault data. To simulate the wear state of parts under actual working conditions, a variety of precision machining technologies such as turning and grinding are adopted, and the parts are processed according to the preset wear amount gradient.

[0015] After preprocessing the vibration signals of each part collected above, such as slicing, normalizing, and randomly shuffling, they are used as the input of the Transformer framework. The encoder encodes the features of the input sequence through the multi-head self-attention mechanism, residual connection, layer normalization, and feed-forward neural network. The decoder uses the masked multi-head self-attention mechanism and cross-attention mechanism, etc., and combines the encoder output to generate the target sequence. The self-attention mechanism, as the core module, calculates the correlation between sequence elements, and the multi-head structure enhances the feature capture ability. The position encoding solves the problem of the model capturing the sequence order information, and injects the position information through sine-cosine encoding or learnable encoding. The feed-forward neural network performs a non-linear transformation on the output of the self-attention mechanism to improve the model's expression ability.

[0016] As a further technical solution of the present invention, the feed-forward neural network in the Transformer framework is replaced by a KAN network. Different from the traditional multi-layer perceptron (MLP), the KAN network fundamentally eliminates the dependence on the linear weight matrix by using a learnable function to replace the fixed activation function. The KAN is based on the Kolmogorov-Arnold representation theorem, which is expressed by the following formula:

[0017] :

[0018] For a multi-layer KAN network, a KAN layer with an input of nin dimensions and an output of nout dimensions is defined as a matrix of one-dimensional functions;

[0019] ;

[0020] There are several KAN layers between the input and the output, and they are connected through a neural network to construct a mapping from the input to the output:

[0021] ;

[0022] where is the function matrix corresponding to the l-th layer, and a general KAN network is a combination of L layers: given an input vector x0 ∈ n0, the output of the KAN is:

[0023] ;

[0024] The activation function in the KAN network is parameterized as a B-spline type univariate function, which makes each weight become a learnable parameter. Then the spline function in KAN is expressed as a linear combination of B-spline functions;

[0025] ;

[0026] represents the spline function, are the coefficients optimized during training, are the B-spline basis functions defined on the grid; during the training process, the parameters of the spline will be optimized to minimize the loss function.

[0027] Beneficial effects achieved by the present invention:

[0028] The present invention establishes a cryopump health monitoring platform, including a cryopump, a compressor system, a health monitoring system, etc. First, different components are quantitatively processed with different damage degrees through processes such as turning to simulate different degrees of faults. Subsequently, sensors are used to collect vibration signals of each component under different fault degrees, and a dataset reflecting the real working conditions of the cryopump is made. A new deep learning combined algorithm of Transformer-KAN network is established to achieve efficient and accurate diagnosis of the cryopump. A health monitoring platform is established, and all the collected vibration signals of the cryopump are stored in the database to achieve real-time health monitoring. The present invention significantly improves the safety performance and operation stability of the cryopump, and plays a key role in ensuring the quality and reliability of semiconductor integrated circuit manufacturing. By effectively maintaining the constant vacuum degree of the semiconductor integrated circuit vacuum system, the risk of generating particulate pollutants and harmful gases in the system can be reduced, the stability of process parameters during production can be ensured, and thus the consistency of product quality and the repeatability of production processes can be improved. In addition, the invention can reduce production interruptions caused by system failures or parameter adjustments, improve production efficiency while reducing production costs, and provide strong support for the continuous and efficient manufacturing of semiconductor integrated circuits. Brief Description of the Drawings

[0029] Figure 1 is a schematic structural diagram of the cryopump health monitoring platform of the present invention.

[0030] Figure 2 is a data acquisition flow chart of the cryopump health monitoring platform of the present invention.

[0031] Figure 3 is a schematic diagram of the principle of the Transformer-KAN network deep learning algorithm built in the present invention.

[0032] Figure 4This is the performance index diagram of the Transformer-KAN deep learning model proposed in the present invention. Detailed implementation manners

[0033] To make the objectives, 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 with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. Therefore, the detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed present invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0034] In the description of the present invention, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality of" means two or more unless otherwise specifically defined.

[0035] Please refer to Figure 1 , an embodiment of the present invention provides a cryopump health monitoring platform based on a deep learning algorithm. The health monitoring platform includes: a cryopump, a compressor system, a vacuum pumping system, and a health monitoring system. The compressor system is connected to the cryopump through a helium pipe to provide helium gas required for the refrigeration cycle for the cryopump; the vacuum pumping system is connected to the cryopump through a bellows and a valve to enable the cryopump to reach the operating pressure; one end of an acceleration sensor in the health monitoring system is adhered to positions such as the pump housing of the cryopump and the outer shell of the refrigerator, and the other end is connected to a data acquisition card; the data acquisition card is connected to a host computer through a network cable to upload the acquired data to the data platform;

[0036] The cryopump is a two-stage G-M refrigeration type cryopump;

[0037] The compressor system is composed of a helium compressor, a helium pipeline, and an electronic control unit;

[0038] The vacuum pumping system is composed of a fore vacuum pump, a pneumatic valve, and a bellows;

[0039] The health monitoring system includes an acceleration sensor, a data acquisition card, a host computer, a data platform, and a fault warning system; the measurement of vibration signals is completed by the acceleration sensor, and the data acquisition and upload to the host computer are realized through the data acquisition card. After receiving the data, the host computer performs calculation and analysis.

[0040] In this embodiment, the cryopump is an Edwards cryo-u6h type cryopump. In this cryopump, the driving device in the refrigerator drives the primary piston to reciprocate in the primary cylinder. The primary piston drives the secondary piston to work in the secondary cylinder. The piston and the cylinder are sealed by a sealing seat. The gas distribution valve is connected to the rotating shaft of the driving device to control the helium gas entering and leaving the cylinder. The fixed valve is installed in the pipeline or chamber to regulate the air flow at specific parts. The driving bearing supports the rotating shaft, and the frame bearing fixes the key components. During operation, the motor drives the piston to move through the driving device. The helium gas completes the refrigeration cycle of inflation, compression, deflation, and expansion in the cylinder. The primary and secondary pistons work together to achieve cryogenic refrigeration, and the cooperation of each component ensures the stable operation of the system.

[0041] In this embodiment, the helium compressor is used to supply circulating helium gas to the cryogenic refrigerator system of the cryopump, enabling it to function as the system refrigerant; the helium pipeline has the characteristics of low helium gas leakage rate and high cleanliness, serving as the transmission channel for high-purity helium gas to maintain the helium gas purity and ensure the stable sealing pressure of the system.

[0042] In this embodiment, one end of the fore-vacuum pump is connected to the cryopump through a pneumatic valve, and the other end is connected to the fore-vacuum pump through a bellows; the pneumatic valve has the function of isolating the cryopump from the fore-vacuum pump; the fore-vacuum pump is used to extract the rough vacuum of the cryopump cavity to create a basic vacuum environment for the startup of the cryopump; when the vacuum degree of the cryopump cavity drops below 10 Pa and reaches the startup condition of the cryopump, the cryopump can be started; during operation, the pneumatic valve can control the on-off between the fore-vacuum pump and the cryopump cavity.

[0043] In this embodiment, to ensure accurate and real-time data of the key components inside the cryopump, the acceleration sensor, data acquisition card, and upper computer are physically connected in series; during the operation of the cryopump, the system monitors and records the vibration signals by collecting data at different stages such as rough pumping, cooling, temperature control, and warming and regeneration; the data collection at each operation stage corresponds to specific operating conditions and pump internal environmental conditions; the acceleration sensor is precisely installed on the pump housing surface close to each component to capture and record the vibration characteristic data during the operation of the cryopump; the data acquisition card is responsible for collecting the data obtained by the acceleration sensor and transmitting it to the data platform; the data platform stores the received data and conducts further analysis and processing; on the data platform, a running fault warning system analyzes the collected data to identify potential fault characteristics and abnormal data patterns; by comparing the real-time data with the historical data and using pattern matching technology, the system generates a cryopump fault diagnosis report and prediction results including the fault cause, development trend, etc.; finally, the fault diagnosis report and prediction results are displayed on the front end to achieve health monitoring, as Figure 2 shown.

[0044] In this embodiment, the method for health monitoring of the cryogenic pump based on the health monitoring platform includes the following steps:

[0045] First, a training data set is made. By replacing components with different degrees of damage, vibration signal data of the components is obtained. The key components include the first- and second-stage pistons, gas distribution valves, drive bearings, fixed valves, seal seats, frame bearings, etc. The collected vibration signals not only cover detailed information on various types of faults of the cryogenic pump, but also contain quantitative data on different degrees of damage under the same fault type. During the experiment, precise control means are used to set the wear amount of each component, so that the simulated fault conditions can truly reflect the possible fault scenarios in the actual operation of the cryogenic pump, thus providing a reliable data basis for the research on cryogenic pump fault diagnosis;

[0046] Fault simulation is achieved by quantifying and processing the components to obtain accurate fault data. To simulate the wear state of the components under actual working conditions, various precision machining technologies such as turning and grinding are used, and the components are processed according to the preset wear amount gradient. Specifically, for 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 seal seat, and the inner and outer diameters and end faces of the frame bearing, the turning process is used, and different degrees of wear processing are achieved by controlling the cutting parameters to ensure the machining accuracy and consistency. For the gas distribution valve, a face milling process is used, and the wear situation in actual use is simulated by adjusting the milling depth and number of times. For bearing components with greater clamping difficulty, a customized processing tooling is designed and used. The end face of the inner ring of the bearing is positioned and fixed by the tooling to ensure the smooth progress of the processing process and the machining accuracy;

[0047] After preprocessing such as slicing, normalizing, and randomly shuffling the vibration signals of each component in the working state collected above, they are used as the input of the Transformer framework. The encoder encodes the features of the input sequence through the multi-head self-attention mechanism, residual connection, layer normalization, and feed-forward neural network. The decoder then uses the masked multi-head self-attention mechanism and cross-attention mechanism, etc., to generate the target sequence in combination with the encoder output. The self-attention mechanism, as the core module, calculates the correlation between sequence elements, and the multi-head structure enhances the feature capture ability. The position encoding solves the problem of the model capturing the sequence order information, and injects the position information through sine-cosine encoding or learnable encoding. The feed-forward neural network performs a non-linear transformation on the output of the self-attention mechanism to improve the model's expression ability.

[0048] In this embodiment, as Figure 3As shown, the Transormer-KAN network deep learning combination algorithm means replacing the feedforward neural network in the Transformer framework with the KAN network. Different from the traditional multi-layer perceptron (MLP), the KAN network fundamentally eliminates the dependence on the linear weight matrix by using a learnable function to replace the fixed activation function; KAN is based on the Kolmogorov-Arnold representation theorem, which asserts that any continuous multivariate function can be represented as a finite superposition of continuous univariate functions. This theorem is expressed by the following formula:

[0049] :

[0050] For a multi-layer KAN network, the entire network architecture is as Figure 3 shown; a KAN layer with nin-dimensional input and nout-dimensional output can be defined as a matrix of one-dimensional functions;

[0051] ;

[0052] There are several KAN layers between the input and the output. Connected by a neural network, a mapping from the input to the output can be constructed:

[0053] ;

[0054] where is the function matrix corresponding to the l-th layer. A general KAN network is a combination of L layers: Given an input vector x0 ∈ n0, the output of KAN is:

[0055] ;

[0056] The activation function in the KAN network is parameterized as a B-spline type univariate function, which makes each weight become a learnable parameter. Then the spline function in KAN is represented as a linear combination of B-spline functions;

[0057] ;

[0058] represents the spline function, are the coefficients optimized during training, are the B-spline basis functions defined on the grid; during the training process, the parameters of the spline will be optimized to minimize the loss function. The purpose of doing this is to adjust the shape of the spline so that it can best fit the training data. This optimization usually uses techniques such as gradient descent, and the spline parameters will be updated at each iteration to reduce the prediction error.

[0059] The above method process steps, that is, using an acceleration sensor to obtain vibration signal data of each component under different fault degrees, which are normal, minor fault, and severe fault respectively; a minor fault means that the component is worn, but the cold head temperature can reach the working temperature and the cryopump can work normally; a severe fault means that the component is worn and the cold head temperature cannot reach the working requirement, and the cryopump cannot work normally. After preprocessing the obtained data, such as slicing, normalizing, and randomly shuffling, a total of 27,000 groups of samples are finally obtained as the input of the Transformer-KAN network model. In addition, the Transformer-KAN network deep learning combination algorithm proposed in the present invention combines the powerful long-distance dependence modeling ability of Transformer, the high parameter efficiency and good interpretability of KAN, and at the same time gives full play to the advantage of the multi-head attention mechanism of Transformer in capturing complex semantics and the excellent neural expansion law brought by the unique B-spline activation function of KAN, which can ensure that the deep learning model not only has high computational performance after multiple rounds of training, but also can achieve a high accuracy rate of more than 98% and excellent generalization ability, as Figure 4 shown.

[0060] It should be noted that in this article, the term "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article, or device including a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of another identical element in the process, method, article, or device including such element.

[0061] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A cryopump health monitoring platform based on deep learning algorithms, characterized in that The health monitoring platform includes: a cryopump, a compressor system, a vacuum pumping system, and a health monitoring system; The cryopump is a two-stage G-M refrigeration type cryopump; The compressor system consists of a helium compressor, a helium pipeline, and an electronic control unit; The vacuum pumping system is composed of a fore pump, a pneumatic valve, and a bellows; The health monitoring system includes an acceleration sensor, a data acquisition card, a host computer, a data platform, and a fault warning system; the measurement of vibration signals is completed by the acceleration sensor, and data acquisition and uploading to the host computer are realized through the data acquisition card. After receiving the data, the host computer performs calculation and analysis.

2. The health monitoring platform for cryopumps based on deep learning algorithms according to claim 1, characterized in that The helium compressor is used to supply circulating helium to the cryogenic refrigerator system of the cryopump, enabling it to function as the system refrigerant; the helium pipeline serves as the transmission channel for helium, maintaining the helium purity and ensuring the stability of the system sealing pressure.

3. The cryopump health monitoring platform based on deep learning algorithm according to claim 1, characterized in that, One end of the fore pump is connected to the cryopump through a pneumatic valve, and the other end is connected to the fore pump through a bellows; the pneumatic valve has the function of isolating the cryopump from the fore pump; the fore pump is used for rough vacuum extraction of the cryopump cavity to create a basic vacuum environment for the start-up of the cryopump.

4. A cryopump health monitoring platform based on a deep learning algorithm according to claim 1, characterized in that, The acceleration sensor, the data acquisition card, and the host computer are physically connected in series; during the operation of the cryopump, the system monitors and records the vibration signals by collecting data at different stages; the data acquisition at each operation stage corresponds to specific operating conditions and pump internal environmental conditions; the acceleration sensor is installed on the pump housing surface close to each component to capture and record the vibration characteristic data during the operation of the cryopump; the data acquisition card is responsible for collecting the data obtained by the acceleration sensor and transmitting it to the data platform; The data platform stores the received data and conducts further analysis and processing; on the data platform, a fault warning system is operated to analyze the collected data and identify potential fault characteristics and abnormal data patterns therein; by comparing the real-time data with the historical data and applying pattern matching technology, the system generates a fault diagnosis report and a prediction result for the cryopump; finally, the fault diagnosis report and the prediction result are displayed on the front end to achieve health monitoring.

5. The health monitoring platform for cryopumps based on deep learning algorithms according to claim 1, characterized in that, The health monitoring method for the cryopump based on the health monitoring platform includes the following steps: First, a training data set is made. By replacing components with different degrees of damage, the vibration signal data of the components are obtained; the collected vibration signals not only cover the detailed information of various types of faults of the cryopump but also contain the quantitative data of different degrees of damage under the same fault type; during the experiment, control means are used to set the damage parameters of each component to make the simulated fault conditions truly reflect the possible fault scenarios in the actual operation of the cryopump, thus providing a reliable data basis for the research on cryopump fault diagnosis; Fault simulation is realized by quantitatively processing the components to obtain accurate fault data; to simulate the wear state of the components under actual working conditions, precision machining technology is used to process the components according to the preset wear amount gradient. The vibration signals of each component in the working state collected are preprocessed and used as the input of the Transformer framework. The encoder encodes the features of the input sequence through the multi-head self-attention mechanism, residual connection, layer normalization, and feed-forward neural network. The decoder uses the masked multi-head self-attention mechanism and cross-attention mechanism to generate the target sequence in combination with the encoder output. The self-attention mechanism, as the core module, calculates the correlation between sequence elements, and the multi-head structure enhances the feature capture ability. The position encoding solves the problem of the model capturing the sequence order information by injecting position information through sine-cosine encoding or learnable encoding. The feed-forward neural network performs a non-linear transformation on the output of the self-attention mechanism to improve the model's expression ability.

6. The cryopump health monitoring platform based on a deep learning algorithm according to claim 5, characterized in that, Replace the feed-forward neural network in the Transformer framework with a KAN network. Different from the traditional multi-layer perceptron, the KAN network fundamentally eliminates the dependence on the linear weight matrix by using a learnable function to replace the fixed activation function. KAN is based on the Kolmogorov-Arnold representation theorem, which is expressed by the following formula: ： For a multi-layer KAN network, a KAN layer with nin-dimensional input and nout-dimensional output is defined as a matrix of one-dimensional functions; ； There are several KAN layers between the input and the output, and a mapping from the input to the output can be constructed through neural network connections: ； where is the function matrix corresponding to the l-th layer. A general KAN network is a combination of L layers: Given an input vector x0 ∈ n0, the output of KAN is: ； The activation function in the KAN network is parameterized as a B-spline type univariate function, which makes each weight become a learnable parameter. Then the spline function in KAN is represented as a linear combination of B-spline functions; ； represents a spline function, is a coefficient optimized during training, is a B-spline basis function defined on a grid; During the training process, the parameters of the spline will be optimized to minimize the loss function.

Images