CN116824337A

CN116824337A - Method and system for generating roughness prediction model based on feature transfer learning

Info

Publication number: CN116824337A
Application number: CN202310821158.4A
Authority: CN
Inventors: 刘坚; 朱理; 索鑫宇; 周飞滔; 路恩会
Original assignee: Jiangsu Upna Technology Co ltd; Hunan University
Current assignee: Jiangsu Upna Technology Co ltd; Hunan University
Priority date: 2023-07-05
Filing date: 2023-07-05
Publication date: 2023-09-29

Abstract

The application discloses a method and a system for generating a roughness prediction model based on feature transfer learning. The method for generating the roughness prediction model comprises the following steps: collecting images of machined surfaces machined by different machining modes and having different roughness, and generating a sample set, wherein the sample set comprises: a plurality of first sample images corresponding to the first processing mode, a plurality of second sample images corresponding to the second processing mode, and roughness metric values respectively corresponding to each of the first sample images and each of the second sample images; inputting a first sample image and a second sample image with the same roughness metric value into a deep transfer learning model, determining an objective function by learning the correlation of the features of the first sample image and the second sample image, and adjusting network parameters of the deep transfer learning model by solving the objective function; and generating a roughness prediction model by using the deep migration learning model.

Description

Method and system for generating roughness prediction model based on feature transfer learning

Technical Field

The application relates to the technical field of computers, in particular to a method and a system for generating a roughness prediction model based on feature transfer learning.

Background

In the machining process, cutting and friction occur between a cutter or a grinding wheel and the surface of a part to separate chips, and factors such as plastic deformation of surface metal, vibration of a machine tool in the machining process and the like can cause the machined surface to form fine peaks and valleys. The unevenness of the minute peaks and valleys in a certain observation range is called surface roughness. The surface roughness has a very important influence on the service performance and service life of the mechanical equipment.

In recent years, with the continuous improvement of industrial automation level, the productivity of an automatic production line is also increasing, and the defects of large labor workload, incomplete detection, untimely feedback and the like of a machining quality monitoring mode based on spot check are not matched with the high-speed production rhythm. On the other hand, with the continuous improvement of computer power and the continuous development of image analysis and processing technology, a roughness measurement method based on machine vision begins to become a research hotspot in the field of intelligent manufacturing, and the method has the advantages of high accuracy, low cost and real-time detection. This has a very important role in improving production efficiency and quality, reducing cost, and reducing human error. The roughness measurement method based on machine vision is mainly a classification or prediction method based on a machine learning or deep learning principle, and requires that the number of samples involved in fitting or training is sufficient, the representative value of the roughness is real and correct, the sample diversity degree is high, otherwise, an accurate mapping relation cannot be constructed, and the generalization of a final prediction model is poor or the confidence of a prediction result is low. In practical applications, the process of collecting samples is time and effort intensive and the cost of obtaining a highly enriched data set is high.

In view of the above, it is important and valuable to study how to rely on a small amount of samples to construct an accurate roughness prediction model in practical engineering application.

Disclosure of Invention

The present application provides methods and systems for generating roughness prediction models based on feature migration learning in an effort to solve or at least alleviate at least one of the problems presented above.

According to one aspect of the present application, there is provided a method of generating a roughness prediction model based on feature migration learning, comprising: collecting images of machined surfaces machined by different machining modes and having different roughness, and generating a sample set, wherein the sample set comprises: a plurality of first sample images corresponding to the first processing mode, a plurality of second sample images corresponding to the second processing mode, and roughness metric values respectively corresponding to each of the first sample images and each of the second sample images; inputting the first sample image and the second sample image with the same roughness metric value into a deep transfer learning model, determining an objective function by measuring the difference of the characteristics of the first sample image and the second sample image, and adjusting network parameters of the deep transfer learning model by solving the objective function; and generating a roughness prediction model by using the deep migration learning model, wherein the roughness prediction model is used for predicting the roughness of the surface of the workpiece.

Optionally, in the method according to the present application, the deep migration learning model includes: a first convolutional neural network component and a second convolutional neural network component; and an objective function calculation component coupled to the first convolutional neural network component and the second convolutional neural network component, respectively, wherein the first convolutional neural network component and the second convolutional neural network component have the same network structure and network parameters, and each of the first convolutional neural network component and the second convolutional neural network component at least comprises a plurality of convolutional processing layers.

Optionally, in the method according to the present application, inputting the first sample image and the second sample image having the same roughness metric value into the deep migration learning model includes: inputting the first sample image into a first convolutional neural network component to extract the characteristics of the first sample image as first characteristics and obtain a roughness predicted value of the first sample image; inputting the second sample image into a second convolutional neural network component to extract features of the second sample image as second features; the first and second features, and the roughness prediction value are input to an objective function calculation component to solve the objective function.

Optionally, in the method according to the application, the objective function comprises: a predicted loss value based on the roughness predicted value and the corresponding roughness metric value, and a distance metric value based on the first feature and the second feature output by the corresponding convolution processing layer.

Optionally, in the method according to the present application, the objective function calculation component includes a prediction loss calculation module and at least one inter-domain difference calculation module, and the inter-domain difference calculation module is correspondingly coupled to the partial convolution processing layers of the first convolution neural network component and the second convolution neural network component, the prediction loss calculation module being coupled to an output of the first convolution neural network component.

Optionally, in the method according to the present application, the inter-domain difference calculation module is adapted to measure a distance between the first feature and the second feature output by the corresponding convolution processing layer; the prediction loss calculation module is suitable for calculating the prediction loss according to the roughness prediction value and the roughness measurement value of the first sample image; the objective function calculation component is adapted to determine an objective function based on the distance and the predicted loss.

Optionally, in the method according to the present application, the first convolutional neural network component and the second convolutional neural network component are implemented by a depth-adaptive network, and the inter-domain difference calculation module measures the distance of the first feature and the second feature using a multi-core maximum mean difference.

Optionally, in the method according to the present application, the first convolutional neural network component and the second convolutional neural network component are implemented by deep coral, and the inter-domain difference calculation module measures the distance of the first feature and the second feature using a correlation alignment loss.

Optionally, in the method according to the application, the first convolutional neural network component and the second convolutional neural network component are implemented by a depth-sub-domain adaptive network, and the inter-domain difference calculation module is adapted to measure the distance of the first feature and the second feature using a local maximum average difference.

Optionally, in the method according to the present application, the first processing mode and the second processing mode are similar processing technologies, and the workpiece surface processed by the first processing mode and the second processing mode has a strong similarity in texture information and gray scale distribution. Wherein the first machining mode is turning machining and the second machining mode is end milling machining.

Optionally, in the method according to the application, capturing images of machined surfaces machined by different machining modes with different roughness, generating a sample set, comprising: collecting images of the cutting machining surfaces with different roughness machined in a first machining mode, and taking the images as a first original image set; collecting images of the cutting machining surfaces with different roughness machined in the second machining mode, and taking the images as a second original image set; and respectively preprocessing the first original image and the second original image to obtain a plurality of first sample images and second sample images with fixed sizes, wherein the preprocessing comprises cutting, downsampling, rotating, overturning transformation and gray scale overturning.

According to still another aspect of the present application, there is provided a system for generating a roughness prediction model, comprising: the data acquisition module is suitable for acquiring images of cutting surfaces which are processed in different processing modes and have different roughness, and generating a sample set, wherein the sample set comprises: a plurality of first sample images corresponding to the first processing mode, a plurality of second sample images corresponding to the second processing mode, and roughness metric values respectively corresponding to each of the first sample images and each of the second sample images; the model training module is suitable for inputting the first sample image and the second sample image with the same roughness measurement value into the deep transfer learning model, determining an objective function by measuring the difference of the characteristics of the first sample image and the second sample image, and adjusting network parameters of the deep transfer learning model by solving the objective function until a roughness prediction model is generated.

According to yet another aspect of the present application, there is provided a computing device comprising: one or more processor memories; one or more programs, wherein the one or more programs are stored in memory and configured to be executed by the one or more processors, the one or more programs comprising instructions for performing any of the methods described above.

According to yet another aspect of the present application, there is provided a computer readable storage medium storing one or more programs, the one or more programs comprising instructions, which when executed by a computing device, cause the computing device to perform any of the methods described above.

In summary, according to the scheme of the application, turning and end milling image data sets are used as two mutually migrated field sample sets, and a depth migration learning model based on depth feature alignment type migration learning is provided. And constructing an objective function based on the distance measurement, and obtaining a final optimized deep migration learning model by solving the objective function. Based on the deep migration learning model, a roughness prediction model is determined, so that the problems of complex and expensive sample collection in a similar processing mode are solved.

In addition, according to the application, experiments prove that under the condition that the sample quantity of the source domain is sufficient, the scarce target domain sample can establish a high-accuracy roughness prediction model through the method.

The foregoing description is only an overview of the present application, and is intended to be implemented in accordance with the teachings of the present application in order that the same may be more clearly understood and to make the same and other objects, features and advantages of the present application more readily apparent.

Drawings

To the accomplishment of the foregoing and related ends, certain illustrative aspects are described herein in connection with the following description and the annexed drawings, which set forth various ways in which the principles herein may be practiced, and all aspects and equivalents thereof are intended to fall within the scope of the claimed subject matter. The above, as well as additional objects, features, and advantages of the present application will become more apparent from the following detailed description when read in conjunction with the accompanying drawings. Like reference numerals generally refer to like parts or elements throughout the present application.

FIG. 1 illustrates a schematic diagram of a system 100 for generating a roughness prediction model in accordance with some embodiments of the application;

FIG. 2 illustrates a schematic diagram of a computing device 200 according to some embodiments of the application;

FIG. 3 illustrates a flow diagram of a method 300 of generating a roughness prediction model based on feature transfer learning, according to some embodiments of the application;

FIG. 4 illustrates a basic architecture diagram of a deep-migration learning model 400 according to some embodiments of the application;

FIG. 5 illustrates a block diagram of a deep migration learning model employing a DAN;

FIG. 6 shows a block diagram of a deep-migration learning model employing deep coral;

FIG. 7 illustrates a block diagram of a deep migration learning model employing a deep sub-domain adaptive network.

Detailed Description

Exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the application to those skilled in the art.

The surface of a workpiece subjected to cutting processing often has texture characteristics with high identification degree, certain relation exists between texture information and surface roughness, but the value of the surface roughness is a specific value calculated by a statistical method, and a clear mapping relation cannot be established between a roughness surface image and the roughness value, so that a computer vision method is generally used for only predicting the grade of roughness formulated according to national standards. The samples of different cutting processes have a certain similarity in surface texture characteristics, so that according to the embodiment of the application, the roughness prediction model of one type of processing mode is used for adapting to the roughness sample of another type of cutting process, thereby avoiding tedious sample collection work and enabling the roughness prediction model to be more rapidly generalized to similar processing modes. According to the application, a roughness prediction scheme based on feature alignment type migration learning is provided, and roughness of cutting surfaces in different processing modes can be predicted.

It should be noted that the above-mentioned different processing manners should be similar processing manners, for example, similar processing manners have similar processing processes, and the texture information of the cut surface processed by the different processing manners has a strong similarity with the gray distribution.

According to the scheme of the application, firstly, a deep migration learning model is constructed based on a deep convolutional neural network; meanwhile, two roughness sample images with different cutting processing modes are collected, and two groups of surface roughness image data sets are constructed; and training a deep migration learning model by using the two sets of image data sets to obtain a final roughness prediction model. The roughness prediction model can predict the roughness of a machined surface machined in a similar machining manner. According to the scheme, the problems of complex and expensive sample collection in a similar processing mode can be solved.

FIG. 1 illustrates a schematic diagram of a system 100 for generating a roughness prediction model according to some embodiments of the application. As shown in fig. 1, the system 100 includes a data acquisition module 110 and a model training module 120.

The data acquisition module 110 is used for acquiring images of machined surfaces machined by different machining modes and having different roughness, and generating a sample set.

According to the embodiment of the application, two similar processing modes are selected to construct a sample set, and subsequent transfer learning is performed. In some embodiments, the applicant has found that, from the physical mechanism of machining, turning and milling are both effected by the cooperation of a rotary motion with a feed motion, the main difference being that the workpiece is rotated during turning and the tool is fed; while during milling, the rotary motion is generated by the tool and the feed motion is generated by the workpiece. Turning and milling (e.g., end milling) are thus two relatively similar machining processes from the standpoint of both the primary cutting motion and the feed motion, and the present application selects to acquire images of the machined surface that is machined by both the turning and milling modes (regardless of planing, grinding). In addition, milling is divided into horizontal milling and vertical milling according to the cutting edges of the milling cutter, the cutting edges and the workpiece surface are greatly different in microcosmic cutting modes, and the applicant confirms through research and comparison that the cutting parameters of the vertical milling and the surface textures after the machining are similar to those of turning. Thus, in some embodiments, two processing modes are employed: the first machining mode is turning machining, and the second machining mode is vertical milling machining. Of course, the first machining method may be end milling machining, and the second machining method may be turning machining, without being limited thereto.

It should be noted that, the turning and milling are only used as examples, and the mode of selecting the sample according to the scheme of the present application is shown, and the present application is not limited to the two modes of turning and end milling. Those skilled in the art will be able to select two similar machining modes from the angles of the main cutting motion and the feed motion, the similarity of the surface textures of the machined workpiece, etc., based on the description of the embodiments of the present application. In short, the first processing mode and the second processing mode are similar in processing technology and cutting parameters, and the surfaces of the workpieces processed by the first processing mode and the second processing mode have strong similarity in texture information and gray distribution.

Specifically, the data acquisition module 110 selects turning and milling (end milling) roughness standard blocks for acquisition of the original image. As a sample pair, 1 turning swatch and 1 milling swatch with the same roughness grade were used. Preferably, the coupons can be from different manufacturers, 1 turning coupon and 1 milling coupon with the same roughness grade from the same manufacturer as one sample pair. In some embodiments, the roughness comprises 4 grades, ra0.8, ra1.6, ra3.2 and ra6.3, respectively.

Acquisition of the original image is generally accomplished by a system of optical imaging hardware and image acquisition software. According to some embodiments of the present application, after the original image is acquired, the data acquisition module 110 further performs preprocessing on the original image, including: the original image is cropped to a predetermined size (e.g., 448 x 448) and then downsampled to the network required size (e.g., 224 x 224). In some embodiments, the original image corresponding to the first processing mode is denoted as a first original image, the original image corresponding to the second processing mode is denoted as a second original image, and the first original image and the second original image after the preprocessing are denoted as a first sample image and a second sample image, respectively.

In addition, in order to improve the generalization capability of the network, the data acquisition module 110 may further perform preprocessing such as data enhancement on each sample image, so as to obtain more sample images. The data enhancement operations include, for example: rotation, inversion transformation, gray scale inversion, contrast adjustment, brightness adjustment, etc., as the present application is not limited in this regard.

Thus, the final generated sample set includes: the first sample images corresponding to the first processing mode, the second sample images corresponding to the second processing mode, and roughness metric values corresponding to the first sample images and the second sample images, respectively.

The model training module 120 is configured to complete the transfer learning of the deep transfer learning model to finally generate a roughness prediction model for predicting the roughness of the machined surface in a similar machining mode.

According to the application, a first sample image and a second sample image with the same roughness metric value are input into a deep migration learning model to determine an objective function by measuring the difference of the features of the first sample image and the second sample image, and network parameters of the deep migration learning model are adjusted by solving the objective function until a roughness prediction model is generated.

According to the system 100 of the application, two groups of sample images are acquired and generated through the data acquisition module 110, the distance between the corresponding features of the two groups of sample images is measured through the model training module 120 by utilizing the feature-based migration learning thought, and the deep migration learning model is optimized according to the distance, so that the roughness prediction model is finally obtained. In other words, according to the scheme of the application, the features of the first sample image and the second sample image are respectively extracted by using the deep convolutional neural network, and the migration loss is calculated by a distance metric function based on feature statistics, so that the distance between the distributions of the two fields is reduced. In the present embodiment, the first sample image is taken as the source domain data, and the second sample image is taken as the target domain data. The depth feature alignment-based migration learning method can obviously improve the prediction performance of the target field data by using source field knowledge.

The system 100 according to the present application may be implemented by one or more computing devices to perform the method 300 of generating a roughness prediction model. Fig. 2 illustrates a block diagram of a computing device 200 according to some embodiments of the application. It should be noted that the computing device 200 shown in fig. 2 is only an example, and in practice, the computing device used to implement the present application may be any type of device, and the hardware configuration of the computing device may be the same as the computing device 200 shown in fig. 2 or may be different from the computing device 200 shown in fig. 2. In practice, the hardware components of computing device 200 shown in FIG. 2 may be added or subtracted from a computing device used to implement the present application, which is not limited by the specific hardware configuration of the computing device.

As shown in FIG. 2, in a basic configuration 202, a computing device 200 typically includes a system memory 206 and one or more processors 204. A memory bus 208 may be used for communication between the processor 204 and the system memory 206.

Depending on the desired configuration, processor 204 may be any type of processor including, but not limited to: a microprocessor (μp), a microcontroller (μc), a digital information processor (DSP), or any combination thereof. Processor 204 may include one or more levels of cache, such as a first level cache 210 and a second level cache 212, a processor core 214, and registers 216. The example processor core 214 may include an Arithmetic Logic Unit (ALU), a Floating Point Unit (FPU), a Digital Signal Processing (DSP) core, or any combination thereof. The example memory controller 218 may be used with the processor 204, or in some implementations, the memory controller 218 may be an internal part of the processor 204.

Depending on the desired configuration, system memory 206 may be any type of memory including, but not limited to: volatile memory (such as RAM), non-volatile memory (such as ROM, flash memory, etc.), or any combination thereof. Physical memory in a computing device is often referred to as volatile memory, RAM, and data in disk needs to be loaded into physical memory in order to be read by processor 204. The system memory 206 may include an operating system 220, one or more applications 222, and program data 224. In some implementations, the application 222 may be arranged to execute instructions on an operating system by the one or more processors 204 using the program data 224. The operating system 220 may be, for example, linux, windows or the like, which includes program instructions for handling basic system services and performing hardware-dependent tasks. The application 222 includes program instructions for implementing various user desired functions, and the application 222 may be, for example, a browser, instant messaging software, a software development tool (e.g., integrated development environment IDE, compiler, etc.), or the like, but is not limited thereto.

When the computing device 200 starts up running, the processor 204 reads the program instructions of the operating system 220 from the memory 206 and executes them. Applications 222 run on top of operating system 220, utilizing interfaces provided by operating system 220 and underlying hardware, to implement various user-desired functions. When the user launches the application 222, the application 222 is loaded into the memory 206, and the processor 204 reads and executes the program instructions of the application 222 from the memory 206.

Computing device 200 also includes storage device 232, storage device 232 including removable storage 236 (e.g., CD, DVD, U disk, removable hard disk, etc.) and non-removable storage 238 (e.g., hard disk drive HDD, etc.), both removable storage 236 and non-removable storage 238 being connected to storage interface bus 234.

Computing device 200 may also include a storage interface bus 234. Storage interface bus 234 enables communication from storage devices 232 (e.g., removable storage 236 and non-removable storage 238) to base configuration 202 via bus/interface controller 230. At least a portion of operating system 220, applications 222, and program data 224 may be stored on removable storage 236 and/or non-removable storage 238, and loaded into system memory 206 via storage interface bus 234 and executed by one or more processors 204 when computing device 200 is powered up or application 222 is to be executed.

Computing device 200 may also include an interface bus 240 that facilitates communication from various interface devices (e.g., output devices 242, peripheral interfaces 244, and communication devices 246) to basic configuration 202 via bus/interface controller 230. The example output device 242 includes a graphics processing unit 248 and an audio processing unit 250. They may be configured to facilitate communication with various external devices, such as a display or speakers, via one or more a/V ports 252. The example peripheral interface 244 may include a serial interface controller 254 and a parallel interface controller 256, which may be configured to facilitate communication via one or more I/O ports 258 and external devices such as input devices (e.g., keyboard, mouse, pen, voice input device, touch input device) or other peripherals (e.g., printer, scanner, etc.). The example communication device 246 may include a network controller 260 that may be arranged to facilitate communication with one or more other computing devices 262 over a network communication link via one or more communication ports 264.

The network communication link may be one example of a communication medium. Communication media may typically be embodied by computer readable instructions, data structures, program modules, and may include any information delivery media in a modulated data signal, such as a carrier wave or other transport mechanism. A "modulated data signal" may be a signal that has one or more of its data set or changed in such a manner as to encode information in the signal. By way of non-limiting example, communication media may include wired media such as a wired network or special purpose network, and wireless media such as acoustic, radio Frequency (RF), microwave, infrared (IR) or other wireless media. The term computer readable media as used herein may include both storage media and communication media.

Computing device 200 may be implemented as a personal computer including desktop and notebook computer configurations. Of course, computing device 200 may also be implemented as part of a small-form factor portable (or mobile) electronic device such as a cellular telephone, digital camera, personal Digital Assistant (PDA), personal media player device, wireless web-watch device, personal headset device, application specific device, or hybrid device that may include any of the above functions. And may even be implemented as servers, such as file servers, database servers, application servers, WEB servers, and the like. The embodiments of the present application are not limited in this regard.

In an embodiment according to the application, the computing device 200 is configured to perform the method 300 of generating a roughness prediction model according to the application. Wherein the application 222 disposed on the operating system contains a plurality of program instructions for performing one or more of the methods described above, which may instruct the processor 204 to perform the method 300 of the present application described above.

FIG. 3 illustrates a flow diagram of a method 300 of generating a roughness prediction model according to some embodiments of the application. As shown in fig. 3, method 300 begins at 310.

At 310, images of machined surfaces machined by different machining modes with different roughness are acquired, creating a sample set.

As described above, the workpiece surface machined by the two cutting machining modes of turning and end milling has a strong similarity in texture information and gray scale distribution, so the first machining mode is selected to be turning machining, and the second machining mode is end milling machining.

According to some embodiments of the present application, images of machined surfaces machined by a first machining mode with different roughness are acquired as a set of first raw images. Meanwhile, images of the machined surfaces with different roughness machined in the second machining mode are collected and used as a second original image set. And then, respectively preprocessing the first original image and the second original image to obtain a plurality of first sample images and second sample images with fixed sizes. Wherein preprocessing includes clipping, downsampling, rotation, inversion transformation, gray scale inversion, and the like.

Thus, the final generated sample set includes: the first sample images corresponding to the first processing mode, the second sample images corresponding to the second processing mode, and roughness metric values corresponding to the first sample images and the second sample images, respectively. In some embodiments, the first and second sample images from the same processing manufacturer having the same roughness level may be used as a sample pair, each sample pair having corresponding label data, i.e., a roughness metric value, which in some embodiments may be represented by a roughness level. According to the present embodiment, the roughness grade includes: ra0.8, ra1.6, ra3.2 and Ra6.3 total 4 grades.

At 320, a first sample image and a second sample image having the same roughness metric value are input into the deep-migration learning model, and an objective function is determined by measuring the variability of the features of the first sample image and the second sample image to adjust network parameters of the deep-migration learning model by solving the objective function.

The transfer learning application problem according to the present application can be formatted as follows: the source domain sample space (i.e., the sample space composed of the first sample image) and the target domain sample space (i.e., the sample space composed of the second sample image) can be defined as X, respectively _s ，X _t The corresponding tag space (i.e., the corresponding roughness metric) may defineIs Y _s ，Y _t The source domain can thus be defined as a sample space X in the source domain _s Data distribution D of (2) _s And a tag function fX _s →[0,1,2,3]And (3) forming a combination, wherein when the prediction result of the label is uncertain, the function can obtain a label expected value corresponding to one input sample instance. The task of the source domain is defined as T _s ＝{y _s ,f(x _s ) X, where x _s Is X _s One example of (a), x _s ∈X _s ，f(x _s ) Is x _s Is the predictive label of y _s Is x _s True tags, y _s ∈Y _s The alike target domain is T _t ＝{y _t ,f(x _t ) }. In this problem, only mutual migration between two domains, one source domain and one target domain, is considered. In terms of real machining, the turning sample and the milling sample can achieve consistent micro surface texture under the condition of a certain feeding amount and a certain cutting amount, and the roughness grade is the same, so that we assume that the sample spaces of the two fields are not independent of each other, namelyAnd Y is _s ＝Y _t From a domain perspective, there is a probability distribution of non-perfect alignment between two domains but with the same predicted objective, namely D _s ≠D _t ，T _s ＝T _t . The transfer learning problem having such characteristics is generally referred to as a isomorphic transfer learning problem.

According to the embodiment of the application, a deep transfer learning model is firstly constructed, the deep transfer learning model is used for carrying out transfer learning on a sample image, and a roughness prediction model is determined based on the deep transfer learning model finally determined after multiple times of optimization. The following describes the process of the transfer learning in detail.

The general idea of solving the above problems by using a feature-based transfer learning method can be roughly divided into the following steps: first, domain correlation characteristic indexes are selected or designed according to the characteristics of the sample, and domain characteristics are generally required to be performedIf a plurality of domain correlation indexes exist in the correlation analysis of the index and the domain distribution difference, different weights are given to the domain correlation characteristic indexes by using methods such as principal component analysis and the like so as to ensure that probability distribution of the two domains is correctly and efficiently aligned. Then a pair of mapping functions needs to be learnedThe domain correlation feature index from two domains is mapped to a common ground feature space, and the difference between the two domains is measured in the common feature space by a specific mapping method, and the problem of narrowing down the inter-domain difference can be generally regarded as a convex optimization problem (i.e., objective function). And then solving the convex optimization problem, namely training a source domain classifier and a target domain classifier, and finally obtaining a roughness prediction model. And finally, in the test stage, selecting a test data set from a target domain, extracting domain characteristic indexes of the test data set, mapping the domain characteristic indexes into a new characteristic space, and inputting the obtained characteristic indexes into a trained classifier for prediction, so that the accuracy performance of the transfer learning method can be judged.

In the field of migration learning, features extracted by a deep convolutional neural network can be generally used as domain correlation feature indexes, and compared with manually designed feature indexes, the indexes do not depend on the professional knowledge of related fields of designers, and have strong domain correlation. Therefore, when the deep transfer learning model is constructed, a feature extraction network component is constructed to extract features of the sample image as domain correlation feature indexes.

In feature-based migration learning, the key to solving domain adaptation is how to learn domain invariance features, i.e., how to learn inter-domain feature mappingTo map samples of different domains to a common space constituted by domain invariance features. The key to obtaining such domain invariance features is how to measure such domain invariance, and related studies have proposed several kinds of domain invarianceDenaturation metrics, the most important of which are maximum mean difference embedding (Maximum Mean Discrepancy Embedding, MMDE) to measure and migrate component analysis.

In some embodiments according to the application, deep neural networks are used instead of kernel-based feature mapping. Optionally, the distance between deep features learned in convolutional neural networks is measured using coded MMD.

Fig. 4 illustrates a basic architecture diagram of a deep-migration learning model 400 according to some embodiments of the application.

As shown in fig. 4, the deep migration learning model 400 includes: the first convolutional neural network component 410 and the second convolutional neural network component 420, and an objective function calculation component 430 coupled to the first convolutional neural network component 410, the second convolutional neural network component 420, respectively. Wherein the first convolutional neural network component 410 and the second convolutional neural network component 420 have the same network structure and network parameters, and each comprises at least a plurality of convolutional processing layers. The convolution processing layers in fig. 4 are merely examples, and the present application does not limit the number of convolution processing layers.

In some embodiments, the first convolutional neural network component 410 is exemplified as comprising a plurality of convolutional processing layers that in turn process an input image to extract image features. Meanwhile, the last convolution processing layer is generally configured as a classification processing layer (e.g., softmax layer) for processing the image features output by the previous convolution processing layer to predict the roughness level corresponding to the image, i.e., obtain a roughness prediction value. Furthermore, the convolution processing layer may include, for example, convolution kernels, pooling, activation, full concatenation, softmax, etc., which the present application does not impose excessive limitations.

Specifically, inputting a first sample image and a second sample image having the same roughness metric value into a deep migration learning model, comprising: inputting the first sample image into a first convolutional neural network component to extract the characteristics of the first sample image as first characteristics and obtain a roughness predicted value of the first sample image; inputting the second sample image into a second convolutional neural network component to extract features of the second sample image as second features; the first and second features, and the roughness prediction value are input to an objective function calculation component to solve the objective function. In the embodiment of the present application, the features output by the other convolution processing layers except the last convolution processing layer in the first convolution neural network component are all called a first feature; the features output by the other convolution processing layers except the last convolution processing layer in the second convolution neural network component are all called second features.

It should be noted that the second convolutional neural network component 420 also contains a classification processing layer that classifies image features to output a roughness prediction value for the second sample image, which is not shown here because it is not required to be processed during the transition learning phase.

According to some embodiments, the objective function will typically include: a predicted loss value based on the roughness predicted value and the corresponding roughness metric value, and a distance metric value based on the first feature and the second feature output by the corresponding convolution processing layer.

In some embodiments, as in fig. 4, the objective function computation component 430 includes: a prediction loss calculation module 432 and at least one inter-domain difference calculation module 434. In some embodiments, the inter-domain difference calculation module 434 is coupled to the partial convolution processing layers of the first and second convolution neural network components 410, 420, respectively, and the prediction loss calculation module 432 is coupled to the output of the first convolution neural network component 410. As shown in fig. 4, the objective function computation component 430 includes 1 inter-domain difference computation module 434. It should be appreciated that the present application is not limited to the number of inter-domain difference calculation modules.

The inter-domain difference calculating module 434 is configured to measure a distance between the first feature and the second feature output by the corresponding convolution processing layer. The predictive loss calculation module 432 calculates a predictive loss from the roughness predictive value and the roughness metric value of the first sample image. In summary, the objective function calculation component 430 determines an objective function based on the measured distance and the predicted loss. By using this distance metric approach, convolutional neural networks can minimize the differences between domains while maximizing tag correlation, thereby automatically learning features with cross-domain representation capabilities.

In particular, the empirical estimation of distance can be expressed in the form:

wherein x is _s ∈X _s 、x _t ∈X _t Respectively representing the input data of the current convolution processing layer, namely, the output of the first sample image and the second sample image after the conversion of the previous layers of convolution neural networks,indicating the convolution operation of the convolution processing layer on the input data, < >>And->And respectively representing a first characteristic and a second characteristic corresponding to the current convolution processing layer.

In order for the entire network to require high prediction accuracy in the source domain and to be able to reduce the distribution distance between the two domains, one way to meet both of these conditions is to use a combined loss function as an objective function, expressed as follows:

wherein, the liquid crystal display device comprises a liquid crystal display device,to predict loss, the data is tagged with a predictive tag (X _s ) And a true label (y) as input, outputting the result by a cross entropy function. The latter term is a distance metric function, also known as migration loss. The hyper-parameter lambda is a balance index between two loss functions, the intensity degree of a desired confusion domain is determined, lambda setting is generally judged according to training results of a model, if lambda setting is too small, MMD regularization terms have no effect on model learning domain invariance characteristics, but lambda setting is too large, regularization is too serious, the model cannot learn class difference characteristics well, and all data points are gathered and cannot be classified.

According to the embodiment of the application, corresponding super parameters (such as an optimizer, a learning rate, a batch size, a migration loss weight and the like) are set for a sample set, and an objective function is calculated through multiple iterations until the value of the objective function meets the condition, and the training is finished, so that a corresponding deep migration learning model is obtained.

At 330, a roughness prediction model is generated using the deep transfer learning model.

According to the application, the first convolutional neural network component or the second convolutional neural network component in the trained deep migration learning model is used as a roughness prediction model. The roughness prediction model can be used for processing images of the machined surface processed in a similar machining mode to predict the roughness of the machined surface.

The architecture is a typical architecture of a deep migration learning model, and various network models are adopted on the basis of the architecture to realize the convolutional neural network component according to the application. The following table lists several models that perform well and their performance.

TABLE 1 depth feature based migration learning model and performance on OFFICE-31 dataset

Wherein A, W, D respectively represent three different fields of data in the OFFICE-31 dataset, namely cutting surface images processed by different processing modes, "A- & gt W" represents taking data A as a first sample image, W as a second sample image, and so on.

From the above table, it can be seen that the deep migration learning model of the method 300 according to the present application can significantly improve the predictive performance of the target domain data by using the knowledge of the source domain. Specifically, according to the scheme, an objective function is constructed based on the distance measurement of the first feature and the second feature, and the final optimized deep migration learning model can be obtained by solving the objective function. Based on the deep migration learning model, a roughness prediction model is determined, so that the problems of complex and expensive sample collection in a similar processing mode are solved.

The method 300 will be used hereinafter to verify the mobility of surface roughness between different cut samples. Considering that there is a certain similarity between models, only three representative models of DAN, deepCoral and DSAN are selected here.

In some embodiments, the first convolutional neural network component and the second convolutional neural network component are implemented by a depth-adaptive network (DAN).

The DAN adopts different modes for training network parameters according to the mobility of the features extracted by different deep convolutional layers of the neural network, can learn the mobility features under statistical assurance, and utilizes the characteristic that mean embedding registration is sensitive to kernel selection, and an inter-domain difference calculation module adopts Multi-core maximum mean difference (Multi Kernel-Maximum Mean Discrepancy, MK-MMD) to measure the distance between the first feature and the second feature, so that compared with a single kernel, the domain adaptation capability is remarkably improved.

FIG. 5 shows a block diagram of a deep migration learning model employing DANs. The source domain data set and the target domain data set in fig. 5 are the set of the first sample image and the set of the second sample image, respectively. Where conv1-conv5, fc6, fc7, fc8 represent convolutional processing layers in the convolutional neural network component, conv is a convolutional layer, and fc is a fully-connected layer.

Since the features extracted by the convolution processing layers are transited from general to specific along with the increase of the depth, different parameter adjustment strategies are adopted for the convolution processing layers with different depths according to the application. In particular, the full link layers fc6-fc8 are designed to extract features that are suitable for a particular classification task, so they are not migratable and are tuned using MK-MMD. As shown in FIG. 5,3 inter-domain difference calculation modules are denoted by "MK-MMD" and measure the distances of fc6-fc8, respectively. The prediction loss calculation module is denoted by "classification loss".

The optimized objective function for the depth feature layer is as follows:

wherein, the liquid crystal display device comprises a liquid crystal display device,representing predicted loss->MK-MMD, l calculated from the first and second features of the first layer ₁ And l ₂ The starting layer sequence number and the ending layer sequence number of MK-MMD are calculated respectively.

The parameter θ optimization strategy of the architecture is as follows:

Since the temporal complexity of computing MK-MMD using core skills is O (n ² ) For quite non-ideal training of deep learning, the method directly carries out unbiased estimation on MK-MMD, and the calculation formula is as follows:

in which quaternions are usedBy the function:

the multidimensional kernel function on each quadruple is calculated, and the calculation time complexity of the method is O (n). The objective function is then adjusted for θ by a random gradient descent algorithm (SGD).

The optimization strategy for parameter β is as follows:

wherein the method comprises the steps ofIs the estimated variance.

In summary, it can be seen that the objective function of a DAN is essentially a maximum and minimum problem:

in some embodiments, the first convolutional neural network component and the second convolutional neural network component are implemented by deep coral.

Deep Coral innovations are to use correlation alignment (CORelation Alignment, coral) loss to minimize differences in feature covariance across domain learning, similar to minimizing MMD with polynomial kernels. According to the application, the inter-domain difference calculation module measures the distance of the first feature and the second feature using the correlation alignment loss.

Fig. 6 shows a structural diagram of a deep-migration learning model using deep coral. The source domain data set and the target domain data set in fig. 6 are the set of the first sample image and the set of the second sample image, respectively. Where conv1-conv5, fc6, fc7, fc8 represent convolutional processing layers in the convolutional neural network component, conv is a convolutional layer, and fc is a fully-connected layer. Fig. 6 contains 1 inter-domain difference calculation module, denoted by "correlation alignment Loss Coral Loss", for measuring the distance of the feature output by the full link layer fc8, i.e., coral Loss.

Specifically, the Coral loss is defined as the distance between the second order statistics (covariance) of the first feature (i.e., the source feature) and the second feature (i.e., the target feature):

wherein the method comprises the steps ofIs the matrix Frobenius norm. The covariance matrix of the source data and the target data is:

the gradient with respect to the input features can be calculated using the chain law:

the loss function of deep Coral is very simple and efficient compared to DAN, demonstrating that Coral loss can better observe the correlation between the two distributions.

In some embodiments, the first convolutional neural network component and the second convolutional neural network component are implemented by a depth subdomain adaptation network (Deep Subdomain Adaptation Network, DSAN).

The depth subdomain adaptive network is also a transfer learning method based on feature distance measurement, and is different from the previously mentioned global domain invariance feature alignment method such as DAN, deep Coral and the like, the method focuses research on the relationship between two subdomains in the same category of different domains, and the network is trained by aligning the relevant subdomain distribution features of specific layers of different domains. The method has the advantages that on the premise of guaranteeing integral distribution alignment, fine granularity information of all categories of the subdomains is considered, and the phenomenon that the integral distribution alignment but confusion of all data in two domains causes inaccurate classification is avoided. According to the present embodiment, the inter-domain difference calculation module measures the distance of the first feature and the second feature using a local maximum average difference (Local Maximum Mean Discrepancy, LMMD) based.

FIG. 7 illustrates a block diagram of a deep migration learning model employing a deep sub-domain adaptive network. The source domain data set and the target domain data set in fig. 7 are the set of the first sample image and the set of the second sample image, respectively. Where the feature extraction network represents a convolutional neural network component, as shown in fig. 7 with 2 inter-domain difference computation modules (not limited thereto), denoted by "LMMD", each requiring four inputs (for example, the 1 st LMMD): depth feature z extracted by the first convolution processing layer ^sl 、z ^tl Real tag y of source domain sample ^s (i.e., roughness metric value of first sample image) and predictive label of target domain sample(i.e., the roughness prediction value of the second sample image), the inter-domain difference calculation module can be extended according to the set number of depth feature extraction layers.

According to the present embodiment, the distance metric formula of the inter-domain difference calculation module is as follows:

/>

wherein x is ^s And x ^t Respectively X _s And X _t Examples of (p) ^(c) And q ^(c) Respectively X _s And X _t Unlike MMD, which only focuses on the overall variability of the distribution, the above equation can measure the inter-domain variability of the local distribution. LMMD then assumes that each sample is according to the weight ω ^c It is determined to which category.

Unbiased estimation was performed on the above equation, in the form:

wherein omega ^c The calculation formula of (2) is as follows:

wherein y is _ic Is the vector y _i For source domain samples, the weights of the respective samples are known because real labels are used, and for target domain samples, the labels of the target domain are not available because the model learning pattern is an unsupervised domain adaptation, and therefore the label prediction value is used to calculate the weights that the samples assign to a class. CMMD can therefore be considered a special case of LMMD (CMMD assumes that each sample has the same weight).

DSAN is a very simple and efficient method of depth domain adaptation. The method is easy to train, is different from a domain adaptation method based on countermeasure, has high convergence rate, and experiments show that DSAN (digital storage area network) is used as a non-countermeasure method, and can obtain obvious standard domain adaptation effect in a target identification task and a digital classification task of a standard migration data set. However, in terms of a method for assigning a class weight to a target domain sample, the method is susceptible to abnormal samples or samples with poor class discrimination.

Further, the above 3 network models were evaluated in terms of prediction accuracy, recall, and feature distribution visualization.

The model training hardware GPU used in the task is RXT 2080Ti single graphics card, and the video memory is 11GB; the CPU is 12vCPU Intel (R) Xeon (R) Platinum 8255C CPU@2.50GHz; the deep learning environment is PyTorrch1.8.1+TorrchVision 0.9.1+Cuda11.1. The same super parameters used for the three methods are: a feature extraction network ResNet50; bottleneck feature dimension 256 dimensions; optimizer SGD { momentum:0.9,weight decay:5e-4}; learning rate adjustment strategy inv { gamma:3e-4, decay:0.75}; batch size 32; training round 200. Some of the different settings super parameters are shown in table 2.

Table 2 experimental hyper-parameters table

And performing mutual migration experiments between two data sets on each model respectively, and setting three random seeds for each experiment to obtain three groups of repeated experiment results.

(1)DAN

Based on the experimental requirements described above, the final effect of the DAN method is shown in table 3 (wherein mAP (mean Average Precision) represents the average of all class accuracy):

TABLE 3DAN prediction accuracy results

(2)DeepCoral

Based on the above experimental requirements, the final effects of the deep coral process are shown in table 4:

TABLE 4DeepCora l prediction accuracy results

(3)DSAN

Based on the above experimental requirements, the final effects of the DSAN method are shown in table 5:

TABLE 5DSAN prediction accuracy results

Since the experimental results are affected by random seeds and fluctuate within a certain range, the best experimental results are considered to be used as final accuracy evaluation indexes of the model, and comprehensive experimental results are set as shown in table 6, wherein ResNet50 is used as a control experimental group.

TABLE 6 summary of best prediction accuracy results table

In terms of prediction accuracy, since DAN, deepCoral focuses mainly on migration effect of the overall distribution, although the overall prediction accuracy is improved to some extent relative to the control group, the accuracy improvement of each subfield is not large. The DSAN-based feature transfer learning method can better capture fine grain information in the sub-fields, so that distribution alignment effects among the sub-fields can be ensured while distribution is integrally aligned, the accuracy and recall rate of each sub-field are improved, the integral prediction accuracy is obviously improved, and the prediction effect of the method on transfer learning models of different cutting surface roughness is best. However, from the experimental results, the repeated experimental results of the method under different random number generators have larger fluctuation, mainly because DSAN concerns about fine granularity characteristics affect the trade-off between migration loss and classification loss, so that the overall loss value is larger, and the model training process has larger fluctuation.

The recall rate is also one of important indexes for evaluating the classification effect of the model, and can measure the capability of the model to check all real samples, so that the discrimination capability of the model to each class is indirectly reflected. Table 7 lists the recall results for the best experimental group of the three models described above.

TABLE 7 recall result summary table

When the data of two fields are subjected to characteristic alignment, if the intra-field variability is not considered, inter-field variability is reduced, but the intra-field authenticability is not further improved. In the two groups of migration tasks, the sample of Ra1.6 is worst in identifiability, the fitting capability of the model to the subdomain characteristics can be reflected from the recall effect of the sample, wherein the recall rate of Ra1.6 in the two groups of migration tasks is respectively improved by 30% and 41.6% relative to ResNet50 by the DSAN method, and the recall rate of other categories is also obviously improved, so that the multi-subdomain migration learning task can be effectively solved by the DSAN method.

In addition, the optimal result set of each method is reduced to 2 dimensions of the depth features by using a t-SNE method, and the depth features are visualized through a chart so as to obtain the distribution of the two-dimensional feature points. In the turning-end milling task (i.e., taking a turning sample as a source domain and an end milling sample as a target domain), compared with DAN, deepCoral, DSAN realizes the integral distribution alignment of the two domains and simultaneously aligns all sub-domains, most of characteristic points of the source domain and the target domain form an aggregation characteristic point cluster with clear boundaries, but partial abnormal samples are difficult to identify for Ra0.8 and Ra1.6 samples. In the end mill-turning task (i.e., using an end mill sample as a source domain and a turning sample as a target domain), it is difficult to determine a clear boundary between the classes from the two-dimensional feature distribution of the 3 models, and the feature scatter diagram of the four kinds of roughness samples should be in an aggregated state from the prediction result. Therefore, taking DAN and DSAN as examples, three-dimensional characteristic visual analysis is carried out, and whether the visual result accords with an ideal state is further judged. Common to both models is: the distribution alignment effects of Ra3.2 and Ra6.3 are ideal, and the sub-field boundaries can be intuitively observed. For the two types Ra1.6 and Ra0.8, the discrimination is poor, and the boundaries between the categories cannot be found intuitively. The two models differ in that: the DSAN can grasp fine granularity characteristic information of subclasses, the aggregation degree among similar samples is obviously stronger than that of DAN, and the distribution gap among the subclasses is large, so that the classification boundary is more definite.

In summary, based on the method 300 of the present application, a high-accuracy cross-domain prediction model can be trained by using the source domain data set and the target domain data set, and the model can significantly reduce the distribution distance calculated based on the statistical index of the example feature points in the two fields.

According to the scheme of the application, turning and end milling image data sets are used as two mutually-migrated field sample sets, and a cutting surface roughness prediction model based on depth feature alignment migration learning is provided, so that the problems of complex and expensive sample collection in a similar processing mode are solved. In addition, according to the application, experiments prove that under the condition that the sample quantity of the source domain is sufficient, the scarce target domain sample can establish a high-accuracy roughness prediction model through the method.

The various techniques described herein may be implemented in connection with hardware or software or, alternatively, with a combination of both. Thus, the methods and apparatus of the present application, or certain aspects or portions of the methods and apparatus of the present application, may take the form of program code (i.e., instructions) embodied in tangible media, such as removable hard drives, U-drives, floppy diskettes, CD-ROMs, or any other machine-readable storage medium, wherein, when the program is loaded into and executed by a machine, such as a computer, the machine becomes an apparatus for practicing the application.

In the case of program code execution on programmable computers, the computing device will generally include a processor, a storage medium readable by the processor (including volatile and non-volatile memory and/or storage elements), at least one input device, and at least one output device. Wherein the memory is configured to store program code; the processor is configured to execute the inventive approach of generating a roughness prediction model according to instructions in said program code stored in the memory.

The application discloses a method for preparing a composite material, which comprises the following steps:

the method of any one of A4-6, wherein the first convolutional neural network component and the second convolutional neural network component are implemented by deep coral, and the inter-domain difference calculation module measures the distance of the first feature and the second feature using a correlation alignment loss. The method of any of A4-6, wherein the first and second convolutional neural network components are implemented by a depth subdomain adaptation network, the interdomain difference calculation module adapted to measure the distance of the first and second features using a local maximum average difference based. A10, the method of any one of A1-9, wherein the first processing mode and the second processing mode are similar processing technologies, and the surface of the workpiece processed by the first processing mode and the surface of the workpiece processed by the second processing mode have strong similarity in texture information and gray level distribution. A11, the method of any of A1-10, wherein the first machining mode is turning and the second machining mode is end milling. A12, the method of any of A1-11, wherein the acquiring images of machined surfaces machined in different machining modes having different roughness, generating a sample set, comprises: collecting images of the cutting machining surfaces with different roughness machined in a first machining mode, and taking the images as a first original image set; collecting images of the cutting machining surfaces with different roughness machined in the second machining mode, and taking the images as a second original image set; and respectively preprocessing the first original image and the second original image to obtain a plurality of first sample images and second sample images with fixed sizes, wherein the preprocessing comprises cutting, downsampling, rotating, overturning transformation and gray scale overturning.

By way of example, and not limitation, readable media comprise readable storage media and communication media. The readable storage medium stores information such as computer readable instructions, data structures, program modules, or other data. Communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media. Combinations of any of the above are also included within the scope of readable media.

In the description provided herein, algorithms and displays are not inherently related to any particular computer, virtual system, or other apparatus. Various general-purpose systems may also be used with examples of the application. The required structure for a construction of such a system is apparent from the description above. In addition, the present application is not directed to any particular programming language. It should be appreciated that the teachings of the present application as described herein may be implemented in a variety of programming languages and that the foregoing descriptions of specific languages are provided for disclosure of preferred embodiments of the present application.

In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the application may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

Similarly, it should be appreciated that in the foregoing description of exemplary embodiments of the application, various features of the application are sometimes grouped together in a single embodiment, figure, or description thereof for the purpose of streamlining the disclosure and aiding in the understanding of one or more of the various inventive aspects. However, the method of this application should not be interpreted as reflecting the intent: i.e., the claimed application requires more features than are expressly recited in each claim. Thus, the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate embodiment of this application.

Those skilled in the art will appreciate that the modules or units or components of the devices in the examples disclosed herein may be arranged in a device as described in this embodiment, or alternatively may be located in one or more devices different from the devices in this example. The modules in the foregoing examples may be combined into one module or may be further divided into a plurality of sub-modules.

Those skilled in the art will appreciate that the modules in the apparatus of the embodiments may be adaptively changed and disposed in one or more apparatuses different from the embodiments. The modules or units or components of the embodiments may be combined into one module or unit or component and, furthermore, they may be divided into a plurality of sub-modules or sub-units or sub-components. Any combination of all features disclosed in this specification (including any accompanying claims, abstract and drawings), and all of the processes or units of any method or apparatus so disclosed, may be used in combination, except insofar as at least some of such features and/or processes or units are mutually exclusive. Each feature disclosed in this specification (including any accompanying claims, abstract and drawings), may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise.

Furthermore, those skilled in the art will appreciate that while some embodiments described herein include some features but not others included in other embodiments, combinations of features of different embodiments are meant to be within the scope of the application and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination.

Furthermore, some of the embodiments are described herein as methods or combinations of method elements that may be implemented by a processor of a computer system or by other means of performing the functions. Thus, a processor with the necessary instructions for implementing the described method or method element forms a means for implementing the method or method element. Furthermore, the elements of the apparatus embodiments described herein are examples of the following apparatus: the apparatus is for performing functions performed by elements for purposes of this disclosure.

As used herein, unless otherwise specified the use of the ordinal terms "first," "second," "third," etc., to describe a general object merely denote different instances of like objects, and are not intended to imply that the objects so described must have a given order, either temporally, spatially, in ranking, or in any other manner. Furthermore, the number word "plurality" means "two" and/or "more than two".

While the application has been described with respect to a limited number of embodiments, those skilled in the art, having benefit of the above description, will appreciate that other embodiments are contemplated within the scope of the application as described herein. Furthermore, it should be noted that the language used in the specification has been principally selected for readability and instructional purposes, and may not have been selected to delineate or circumscribe the inventive subject matter. Accordingly, many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the appended claims. The disclosure of the present application is intended to be illustrative, but not limiting, of the scope of the application, which is defined by the appended claims.

Claims

1. The method for generating the roughness prediction model based on the feature transfer learning comprises the following steps:

collecting images of machined surfaces machined by different machining modes and having different roughness, and generating a sample set, wherein the sample set comprises: a plurality of first sample images corresponding to the first processing mode, a plurality of second sample images corresponding to the second processing mode, and roughness metric values respectively corresponding to each of the first sample images and each of the second sample images;

Inputting the first sample image and the second sample image with the same roughness metric value into a deep transfer learning model, determining an objective function by measuring the difference of the characteristics of the first sample image and the second sample image, and adjusting network parameters of the deep transfer learning model by solving the objective function;

and generating a roughness prediction model by using the deep migration learning model, wherein the roughness prediction model is used for predicting the roughness of the surface of the workpiece.

2. The method of claim 1, wherein the deep-migration learning model comprises:

a first convolutional neural network component and a second convolutional neural network component; and

an objective function calculation component coupled to the first convolutional neural network component and the second convolutional neural network component, respectively,

the first convolutional neural network component and the second convolutional neural network component have the same network structure and network parameters and at least comprise a plurality of convolutional processing layers.

3. The method of claim 2, wherein inputting the first sample image and the second sample image having the same roughness metric value into a deep-migration learning model comprises:

Inputting the first sample image into the first convolutional neural network component to extract the characteristics of the first sample image as first characteristics and obtain a roughness prediction value of the first sample image;

inputting the second sample image into the second convolutional neural network component to extract features of the second sample image as second features;

the first feature, the second feature, and the roughness prediction value are input to the objective function calculation component to solve the objective function.

4. A method as claimed in claim 3, wherein the objective function comprises:

and a predicted loss value based on the roughness predicted value and the corresponding roughness measurement value, and a distance measurement value based on the first feature and the second feature output by the corresponding convolution processing layer.

5. The method of any of claims 2-4, wherein the objective function computation component comprises a prediction loss computation module and at least one inter-domain difference computation module, and the inter-domain difference computation module is coupled to the partial convolution processing layers of the first and second convolution neural network components, respectively, the prediction loss computation module being coupled to an output of the first convolution neural network component.

6. The method of claim 5, wherein,

the interdomain difference calculation module is suitable for measuring the distance between the first feature and the second feature output by the corresponding convolution processing layer;

the prediction loss calculation module is suitable for calculating the prediction loss according to the roughness predicted value and the roughness measurement value of the first sample image;

the objective function calculation component is adapted to determine an objective function based on the distance and the predicted loss.

7. The method of any of claims 4-6, wherein the first convolutional neural network component and the second convolutional neural network component are implemented by a deep adaptation network,

the inter-domain difference calculation module measures a distance between the first feature and the second feature using a multi-core maximum mean difference.

8. A system for generating a roughness prediction model, comprising:

the data acquisition module is suitable for acquiring images of cutting surfaces which are processed in different processing modes and have different roughness, and generating a sample set, wherein the sample set comprises: a plurality of first sample images corresponding to the first processing mode, a plurality of second sample images corresponding to the second processing mode, and roughness metric values respectively corresponding to each of the first sample images and each of the second sample images;

The model training module is suitable for inputting the first sample image and the second sample image with the same roughness measurement value into a depth migration learning model, determining an objective function by measuring the difference of the characteristics of the first sample image and the second sample image, and adjusting network parameters of the depth migration learning model by solving the objective function until a roughness prediction model is generated.

9. A computing device, comprising:

one or more processors;

a memory;

one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs comprising instructions for performing the method of any of claims 1-7.

10. A computer readable storage medium storing one or more programs, the one or more programs comprising instructions, which when executed by a computing device, cause the computing device to perform the method of any of claims 1-7.