Application product classification management method, system and equipment based on deep learning and dynamic and static characteristics, and medium

Abstract

The invention discloses an application product classification management method, system and device based on deep learning and dynamic and static characteristics and a medium, and belongs to the technical field of classification management, and the method comprises the steps: extracting the static characteristics of an application product after a scheduling edge cluster receives the application product; after the product runs, extracting the dynamic characteristics of the product application; performing feature extraction on the static features of the mirror image and the dynamic features of the mirror image by adopting a neural network model; classifying the application products according to the extracted features; and according to the classification result and the available domain resource state of the edge cluster, determining a deployment position and a resource allocation scheme, and completing the deployment of the application product. According to the method, after the static features and the dynamic features are obtained, feature extraction is carried out on the static features and the dynamic features of the mirror images by adopting the new neural network model, then product classification is carried out according to the extracted features, intelligent management of application product deployment is realized according to the classification result, and the deployment rationality of system function modules is improved.

Description

Technical Field

[0001] This invention relates to the field of classification management technology, specifically to a method, system, equipment, and medium for classification management of application products based on deep learning and dynamic and static features. Background Art

[0002] Edge computing, as a computing paradigm that pushes data processing and storage to the network edge, has significant advantages such as low latency, high bandwidth, and reduced network congestion. Therefore, edge computing is gradually becoming an important way to solve the needs of big data processing.

[0003] As edge clusters expand and application scenarios diversify, the management of application artifacts becomes increasingly complex. Traditional manual management methods are no longer sufficient to handle this complexity. Therefore, a more intelligent and automated management approach is needed to address this challenge. The process involves: 1) Deploying applications based on the application artifact package to obtain application instances. 2) Defining the database based on the application request to obtain the database architecture; 3) Defining microservices based on the database architecture to obtain application instance information; 4) Building the application model using the database architecture and application instance information to obtain the application model. 5) Developing code based on the application model to obtain the development results; 5) Packaging the development results, database architecture, and application instance information to obtain the application artifact package. 6) Scalabilizing the database based on the database architecture using sharding information to obtain the sharding; 7) Deploying application instances based on the sharding and application instance information to obtain the application instances. 8) Deploying applications based on pre-obtained application request by the application management platform, obtaining application instances and receiving data routing rules corresponding to the application instances from the application management platform; 8) Enabling application deployment and improving deployment efficiency.

[0004] Existing methods concatenate and fuse first and second classification feature information, and output the fused category probability result of the application artifact in a trained classification model. Based on the category probability result, the corresponding deployment strategy for the application artifact is matched in a deployment strategy template library. This enables intelligent management of application artifact deployment based on the classification result, improving the rationality of system functional module deployment. However, concatenating and fusing the application artifact and outputting the fused category probability result in a trained classification model results in a large number of errors, requiring large models and algorithms for training, which is very cumbersome. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by this invention is: how to solve the problem that there are a lot of errors in the existing splicing and fusion and output of the category probability results of the fused products in the trained classification model, which requires large models and large algorithms for training.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a method for classifying and managing application artifacts based on deep learning and dynamic / static features, comprising the following steps: after the scheduling edge cluster receives the application artifact, extracting the static features of the application artifact; after the artifact runs, extracting the dynamic features of the application artifact; using a neural network model to extract the static and dynamic features of the image; classifying the application artifact according to the extracted features; and determining the deployment location and resource allocation scheme based on the classification results and the availability domain resource status of the edge cluster, thereby completing the deployment of the application artifact.

[0008] As a preferred embodiment of the application product classification management method based on deep learning and dynamic and static features described in this invention, the static features are extracted when the product is not running, and the dynamic features are extracted by real-time monitoring during the product's operation. The static features and dynamic features are extracted by independent processing flows at different stages.

[0009] As a preferred embodiment of the product classification management method based on deep learning and dynamic and static features described in this invention, the neural network model includes a time series modeling structure for processing static features, a transformer encoding structure for processing dynamic features, and a recurrent neural network structure for fusing static and dynamic features.

[0010] As a preferred embodiment of the application artifact classification and management method based on deep learning and dynamic and static features described in this invention, the deployment location and resource allocation scheme are jointly determined based on the classification results and the availability domain resource status of the edge cluster, and are used to complete the deployment process of the application artifact.

[0011] As a preferred embodiment of the application artifact classification and management method based on deep learning and dynamic and static features described in this invention, the static features and dynamic features of the image include: filtering container application records that meet the filtering strategy from the application's container application record set, extracting the image identifier contained in the container application record, and classifying the image feature information corresponding to the image identifier to obtain at least one image subclass; extracting feature images for image testing of the application from the image subclass; detecting whether the feature image supports downloading from the image repository to the local machine in the image download dimension; if so, detecting whether the feature image supports being viewed in the image viewing dimension; if so, detecting whether the feature image supports being started in the image startup dimension; if so, detecting whether the feature image supports being deleted in the image deletion dimension; and / or detecting whether the residual information after the feature image is deleted is cleared; if so, confirming that the feature image passes the image test.

[0012] This preferred solution classifies image identifiers by filtering them from the application's container application record set and extracting image feature information. It can construct an image subclass structure oriented to application scenarios. Combined with multiple dimensions of image testing operations such as image download, viewing, starting, deleting, and residual detection, it can form dynamic features that are completely associated with specific image behavior states. These features, together with static features, participate in the feature extraction process, enhancing the model's adaptability to the multi-dimensional performance of images.

[0013] As a preferred embodiment of the product classification management method based on deep learning and dynamic and static features described in this invention, the time series modeling structure is configured to perform time series modeling on the static features and output a residual sequence; the residual sequence and the dynamic feature sequence output by the transformer encoding structure are aligned in the same time dimension according to the same time step, and the sequences are concatenated at corresponding positions; the concatenated combined sequence is used as input and sequentially fed into each time step of the recurrent neural network structure, and the recurrent neural network structure updates its state at each time step based on the current input and the hidden state of the previous time step, and outputs joint features.

[0014] This preferred scheme aligns the residual sequence obtained by time series modeling of static features with the dynamic feature sequence in the time dimension, and forms a combined sequence by positional concatenation. This helps to maintain the temporal synchronization relationship between different feature sources. By feeding the combined sequence into the recurrent neural network structure for processing in sequence, the state correlation between the original inputs can be preserved, so that the final generated joint features have temporal continuity and structural fusion capabilities, and improve the input consistency of classification processing.

[0015] As a preferred embodiment of the application artifact classification and management method based on deep learning and dynamic / static features described in this invention, the step of determining the deployment location and resource allocation scheme includes: using the classification result and the availability domain resource status of the edge cluster as deployment parameters, wherein the classification result is used to determine the set of deployment candidate nodes, and the availability domain resource status is used to provide the current resource information of the candidate nodes; matching the deployment type and node feature tags based on the classification result, and selecting nodes that meet the category requirements from the set of deployment candidate nodes; comparing the resource information of the selected candidate nodes, and determining the target deployment node according to a preset resource sorting rule; allocating resources to the application artifact based on the resource details of the target deployment node, and completing the deployment.

[0016] In determining the deployment location and resource allocation scheme, this preferred scheme uses the classification results as one of the inputs to the deployment decision to participate in the construction of the candidate node set. It also combines the availability domain resource status of the edge cluster to screen and sort nodes. This enables the establishment of association rules between the classification results and resource status of different application artifacts, realizing resource scheduling paths driven by classification information, and making the deployment behavior more in line with the artifact characteristics and resource matching relationship.

[0017] Another objective of this invention is to provide an application product classification management system based on deep learning and dynamic / static features.

[0018] To address the aforementioned technical problems, this invention provides the following technical solution: an application artifact classification and management system based on deep learning and dynamic / static features, comprising: a static feature extraction module, a dynamic feature extraction module, a fusion module, a classification module, and a deployment module; the static feature extraction module is used to extract the static features of the application artifacts after the scheduling edge cluster receives them; the dynamic feature extraction module is used to extract the dynamic features of the application artifacts after they are running; the fusion module is used to extract features from the static and dynamic features of the mirror image using a neural network model; the classification module is used to classify the application artifacts according to the extracted features; and the deployment module is used to determine the deployment location and resource allocation scheme based on the classification results and the availability domain resource status of the edge cluster, thereby completing the deployment of the application artifacts.

[0019] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the application product classification management method based on deep learning and dynamic and static features.

[0020] The present invention provides a computer-readable storage medium storing a computer program thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the application product classification and management method based on deep learning and dynamic and static features.

[0021] The beneficial effects of this invention are as follows: After obtaining static and dynamic features, this invention uses a newly built neural network model to extract the static and dynamic features of the mirror image, and then classifies the products according to the extracted features. Based on the classification results, it realizes intelligent management of the deployment of application products and improves the rationality of the system's functional module deployment. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating an application product classification and management method based on deep learning and dynamic / static features, as provided in one embodiment of the present invention.

[0024] Figure 2 This is a system block diagram of an application product classification management system based on deep learning and dynamic / static features, provided as an embodiment of the present invention. DETAILED DESCRIPTION

[0025] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0026] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a method for classifying and managing application products based on deep learning and dynamic / static features, including:

[0027] S1. After the scheduling edge cluster receives the application artifact, extract the static features of the application artifact;

[0028] S2. After the work-in-process is running, extract the dynamic characteristics of the work-in-process application;

[0029] S3. Use a neural network model to extract features from the static and dynamic features of the mirror image;

[0030] S4. Classify the applied products according to the extracted features;

[0031] S5. Based on the classification results and the availability of resources in the edge cluster, determine the deployment location and resource allocation scheme, and complete the deployment of application artifacts.

[0032] It should be noted that edge computing environments present practical challenges such as heterogeneous nodes, significant resource differences, and diverse deployment tasks. Traditional deployment solutions typically rely on static configuration, making it difficult to adjust deployment strategies based on operational behavior. This embodiment combines static and dynamic features for joint analysis, constructs a deep learning model to extract key features, and achieves adaptive deployment decision-making based on classification. It is applicable to various application artifact deployment management scenarios, and is particularly suitable for use in edge cluster environments with uneven resource distribution and dynamically changing deployment tasks.

[0033] Example 2, refer to Figure 1 This is the second embodiment of the present invention. Based on the above embodiments, an application product classification and management method based on deep learning and dynamic and static features is provided.

[0034] In this embodiment of the application, after the scheduling edge cluster receives the application artifact in step S1, the static features of the application artifact are extracted. The static features are extracted when the artifact is not running.

[0035] The scheduling edge cluster is a hyperconverged management object suite that includes storage, network, and host resources, providing the resources needed to run virtualized or containerized applications. OpenNebula's management services reside in the cloud and are responsible for scheduling, monitoring, and lifecycle management.

[0036] By dividing resources into availability domains, resources in a heterogeneous cluster can be divided into multiple logically independent but physically interconnected regions. This division makes resource management more flexible, enabling resource allocation and scheduling based on different business needs, resource types, and performance characteristics. In a heterogeneous cluster, different computing nodes have different performance, storage, and network capabilities. By dividing resources into availability domains, resource nodes with similar characteristics can be classified into the same availability domain, allowing the resource scheduler to more accurately match task requirements and resource capabilities when allocating resources. The scheduling structure of hybrid resources in a heterogeneous cluster typically consists of network slice users, mobile virtual network operators, and infrastructure providers.

[0037] Assuming there are n slice users in the network, and the required hybrid resources are A, then the virtual link information of the i-th user can be represented as:

[0038]

[0039] Among them, A j,k M represents the probability that the j-th resource is allocated to the k-th virtual machine. c This is a collection of cloud tasks. Based on the structural characteristics of heterogeneous clusters, this invention uses a five-axis graph to represent the physical network, expressed as:

[0040] G5=h×l×t×r

[0041] Where G5 is a five-way graph, l is the number of physical nodes, t is the set of physical links, and r is the attribute vector of the physical nodes. Virtual nodes in the heterogeneous cluster are mapped to physical nodes, and resource features are extracted by combining link attribute constraints. The formula is:

[0042]

[0043] Among them, s i Let α0 be the resource feature vector mapped to the i-th virtual node or application artifact, α0 be the link attribute constraint, and j be the resolution coefficient. For example, the resolution coefficient is determined by the different artifact name, artifact application type, and artifact storage requirements.

[0044] In one alternative implementation, the scheduling edge cluster extracts image description information based on a centralized metadata repository, and parses metadata fields such as environment tags, version numbers, and dependency lists during the image building process as the source of static characteristics of the artifact. This eliminates the reliance on container application records, making it suitable for brand-new artifacts without a running history.

[0045] In another alternative implementation, the system moves the static feature extraction to the scheduling request receiving stage. The scheduling controller directly extracts predefined fields, such as application ID, resource request template and image path, from the original request payload and transforms them into standardized feature vectors through the rule engine. This eliminates the need to access edge nodes or image repositories and is suitable for unified scheduling management scenarios across clusters.

[0046] This invention constructs a static feature system for artifacts based on container application records. This system can generate feature data by combining the actual operating environment and deployment structure, improving the accuracy of static features in expressing resource requirements and application forms. Furthermore, by combining a logical partitioning mechanism of resource availability domains, a mapping relationship is established between static feature extraction and resource structure. This provides a structurally consistent input basis for subsequent feature classification and deployment node selection, effectively improving the semantic consistency of feature modeling and the adaptability of deployment decisions.

[0047] In this embodiment of the application, after the work-in-process is running in step S2, the dynamic features of the work-in-process are extracted. The dynamic features are extracted by the real-time monitoring process during the operation of the work-in-process, while the static features and dynamic features are extracted by independent processing flows at different stages.

[0048] Dynamic features include data such as resource usage, inter-application interactions, application load, and application access traffic.

[0049] The core of a Transformer encoder, the Multi-Head Self-Attention (MHSA) module, captures the correlations between all feature sequences through a self-attention mechanism. A Transformer encoder also contains Multilayer Perceptron (MLP) layers and layer normalization. The output Y of the Transformer encoder... te It can be represented as:

[0050] X te =MHSA(LN(X) seq )+x seq )

[0051] Y te =MLP(LN(X) te )+X te )

[0052] Among them, X te It is an intermediate feature.

[0053] In the field of image processing, Transformer encoders are commonly used to obtain the correlation between global input image patches; the core is to extract the dynamic features of the product application based on Transformer encoders.

[0054] In one alternative implementation, dynamic features are not extracted via a neural network modeling structure. Instead, a rule engine extracts fixed-dimensional statistical features, such as average CPU utilization, request frequency, and memory peaks, from runtime logs, performance metrics, and link tracing data. These features are then mapped into feature vectors using static templates.

[0055] In another alternative implementation, the system deploys a lightweight edge proxy component to collect raw dynamic data streams locally in real time on the application running node. This data is preprocessed at the edge to generate a basic dynamic summary, which is then uploaded to the central node for further modeling, thereby reducing network bandwidth consumption and reducing the load on the central model.

[0056] This invention employs a transformer coding structure to perform high-dimensional representation modeling of runtime dynamic behavior data, effectively capturing long-distance dependencies and global structure between features while avoiding information loss issues in long sequence processing common in traditional sequence models. Furthermore, by constructing dynamic feature representations based on an attention mechanism, the system can still extract key dynamic behavior patterns when faced with highly heterogeneous or complex interactive artifacts, providing high-resolution input representations for subsequent fusion with static features, thereby enhancing the contextual adaptability of the overall classification and deployment strategies.

[0057] In this embodiment of the application, step S3 employs a neural network model to extract features from the static and dynamic features of the mirror image. The neural network model includes a time series modeling structure for processing static features, a transformer encoding structure for processing dynamic features, and a recurrent neural network structure for fusing static and dynamic features.

[0058] The static and dynamic features of the image include: filtering container application records that meet the filtering strategy from the application's container application record set; extracting image feature information corresponding to the image identifier contained in the container application record; classifying the images based on the feature similarity of the image feature information to obtain at least one image subclass; extracting feature images for image testing of the application from the image subclass; detecting whether the feature image supports downloading from the image repository to the local machine in the image download dimension; if so, detecting whether the feature image supports being viewed in the image viewing dimension; if so, detecting whether the feature image supports being started in the image startup dimension; if so, detecting whether the feature image supports being deleted in the image deletion dimension; and / or detecting whether residual information after the feature image is deleted is cleared; if so, confirming that the feature image passes the image test.

[0059] The time series modeling structure is configured to perform time series modeling on static features and output residual sequences. The residual sequences and the dynamic feature sequences output by the transformer encoding structure are aligned in the same time dimension at the same time step and concatenated at corresponding positions. The concatenated combined sequence is used as input and sequentially fed into each time step of the recurrent neural network structure. At each time step, the recurrent neural network structure updates its state based on the current input and the hidden state of the previous time step and outputs joint features.

[0060] ARIMA models are good at capturing the underlying trends and variability of features, but they are insufficient in handling nonlinear patterns and long-term dependencies. LSTM models excel at capturing complex nonlinear patterns and sequential dependencies, but may overfit relatively stable data. The ARIMA-LSTM joint model fully leverages the advantages of both ARIMA and LSTM, significantly improving the prediction accuracy and stability of features.

[0061] Construct an ARIMA regression model, which is expressed as follows:

[0062]

[0063] Among them, y t The predicted value is ε, where t is the event variable and ε is the predicted value. t It is a white noise sequence. With θ q It is a coefficient;

[0064] Based on previous experience, a product model, ARIMA, was constructed. This method can effectively handle periodic and trend factors in time series data. Data analysis was performed, calculating the sample autocorrelation coefficient (ACF) and sample partial autocorrelation coefficient (PACF) of the observed value series. An ARIMA(p, d, q)*(P, D, Q)s of appropriate order was selected for fitting. Here, p, d, and q represent the autoregression order, differencing order, and moving average order, respectively, while P, D, Q, and S represent the autoregression order, differencing order, moving average order, and period, respectively. Data from months K to K+10 in the time series were randomly selected, and the fitted model was used to predict the data for month K+11, followed by comparison. Since residual analysis was not yet performed, a low standard was set: a 30% error between the predicted and actual values ​​was considered acceptable.

[0065] Constructing an LSTM neural network: The LSTM structure employs a gating mechanism and consists of four parts: memory cells, input gates, output gates, and forget gates. The LSTM computation formula can be expressed as:

[0066] Input Gate:

[0067] i t =σ(W xi x t +W hi h t-1 +W Ci C t-1 +q)

[0068] Forgotten Gate:

[0069] f t =σ(W xf x t +W hf h t-1 +W of C t-1 +b f )

[0070] Cell state:

[0071] C t =f t ⊙C t-1 +i t ⊙tabh(W xC x t +W bC h t-1 +b t )

[0072] Output gate:

[0073] O t =σ(W x0 X t +Wh0 h t-1 +W b0 C t-1 +b0)

[0074] Hidden layer state:

[0075] h t =O t ⊙tanh(C t )

[0076] Where W is the gate weight vector, b is the bias, σ is the sig-moid activation function, and tanh is the hyperbolic tangent activation function.

[0077] The data is adapted, including normalizing variables (including input and output values). Data from months K to K+10 is used as the training set, and data from months K+11 is used as the test set. The LSTM model is initialized and training begins, with relevant parameters set.

[0078] A joint ARIMA and LSTM neural network model is constructed: The ARIMA model is used to predict the linear features of the drug consumption sequence, generating corresponding residual sequences. The nonlinear features of the original sequence are embedded in these residual sequences. The powerful nonlinear data processing capabilities of the LSTM model are used to correct these residual data sequences. The linear prediction results from the ARIMA model are combined with the residual prediction values ​​from the LSTM model to obtain the predicted values ​​of the features.

[0079] In one alternative implementation, static and dynamic features are not processed through multi-stage modeling, but are directly concatenated into a unified feature vector and input into a fully connected neural network. After feature compression, the vector is then entered into the classification layer for label prediction.

[0080] In another alternative implementation, static features are represented by sparse coding, while dynamic features are extracted by a convolutional neural network and converted into fixed-length vectors. The two vectors are then concatenated and input into a support vector machine classifier to perform a multi-class classification task.

[0081] This invention, through phased modeling and joint input processing of static and dynamic features, can more accurately capture the structural relationship and temporal dependence between the two types of features. By using residual extraction, attention modeling, and temporal recursion mechanisms in the neural network model, it achieves effective expression of the high-dimensional feature space, thereby improving the classification results' comprehensive ability to identify the structural features and operational characteristics of the applied products, and enhancing the accuracy and discriminativeness of product type classification.

[0082] In this embodiment of the application, the application articles are classified according to the extracted features in step S4.

[0083] In this embodiment of the application, step S5 determines the deployment location and resource allocation scheme based on the classification results and the availability domain resource status of the edge cluster, thereby completing the deployment of the application artifacts. The classification results and the availability domain resource status of the edge cluster are jointly determined to complete the deployment process of the application artifacts.

[0084] The classification results and the availability domain resource status of the edge cluster are used as deployment parameters. The classification results are used to determine the set of candidate deployment nodes, and the availability domain resource status is used to provide the current resource information of the candidate nodes. Based on the classification results, the deployment type and node feature tags are matched to select nodes that meet the category requirements from the set of candidate deployment nodes. The resource information of the selected candidate nodes is compared, and the target deployment node is determined according to the preset resource sorting rules. Based on the resource details of the target deployment node, resources are allocated to the application artifacts, and the deployment is completed.

[0085] By dividing resources into availability domains, resources in a heterogeneous cluster can be divided into multiple logically independent but physically interconnected regions. This division makes resource management more flexible, enabling resource allocation and scheduling based on different business needs, resource types, and performance characteristics. In a heterogeneous cluster, different computing nodes have different performance, storage, and network capabilities. By dividing resources into availability domains, resource nodes with similar characteristics can be classified into the same availability domain, allowing the resource scheduler to more accurately match task requirements and resource capabilities when allocating resources. The scheduling structure of hybrid resources in a heterogeneous cluster typically consists of network slice users, mobile virtual network operators, and infrastructure providers.

[0086] In one alternative implementation, the deployment location is selected based solely on the availability domain resource status, ignoring the classification results, and the node with the least available resources is selected from the cluster nodes for deployment based solely on the current resource utilization or load balancing strategy.

[0087] In another alternative implementation, the system does not sort the nodes, but instead uses a round-robin scheduling method to try deployment in turn among the nodes that meet the minimum resource requirements until deployment is successful.

[0088] This invention incorporates classification results into the deployment decision-making process, enabling the modeling of the correlation between artifact types and resource environments. This allows deployment node selection to consider not only resource availability but also the structural and behavioral characteristics of artifacts, thereby improving the adaptability and execution efficiency of deployment nodes. This approach is suitable for edge cluster environments with heterogeneous nodes and partitioned resource structures, and can achieve more accurate and dynamic deployment strategy execution under conditions of diverse artifact types and dynamically changing resources.

[0089] Example 3 is the third embodiment of the present invention, which differs from the previous two embodiments in that:

[0090] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0091] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0092] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0093] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented in combination with any of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0094] Example 4, refer to Figure 2 This is the fourth embodiment of the present invention. This embodiment provides an application product classification management system based on deep learning and dynamic and static features, including a static feature extraction module, a dynamic feature extraction module, a fusion module, a classification module, and a deployment module.

[0095] The static feature extraction module is used to extract the static features of the application artifacts after the scheduling edge cluster receives the application artifacts.

[0096] The dynamic feature extraction module is used to extract the dynamic features of the work-in-process after it has been running.

[0097] The fusion module is used to extract static and dynamic features from the mirror image using a neural network model.

[0098] The classification module is used to classify application products based on the extracted features.

[0099] The deployment module is used to determine the deployment location and resource allocation scheme based on the classification results and the availability of resources in the edge cluster, and to complete the deployment of application artifacts.

[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for classifying and managing application products based on deep learning and dynamic / static features, characterized in that: include, After the scheduling edge cluster receives the application artifact, it extracts the static features of the application artifact; After the work-in-process is in operation, extract the dynamic characteristics of the work-in-process application; A neural network model is used to extract features from the static and dynamic features of the mirror image. The applied products are classified according to the extracted features; Based on the classification results and the availability of resources in the edge cluster, the deployment location and resource allocation scheme are determined, and the application artifacts are deployed.

2. The application product classification and management method based on deep learning and dynamic / static features as described in claim 1, characterized in that: The static features are extracted when the work-in-process is not running, while the dynamic features are extracted during the running process of the work-in-process. The static features and dynamic features are extracted by independent processing procedures at different stages.

3. The application product classification and management method based on deep learning and dynamic / static features as described in claim 2, characterized in that: The neural network model includes a time series modeling structure for processing static features, a transformer encoding structure for processing dynamic features, and a recurrent neural network structure for fusing static and dynamic features.

4. The application product classification and management method based on deep learning and dynamic / static features as described in claim 3, characterized in that: The deployment location and resource allocation scheme are jointly determined based on the classification results and the availability domain resource status of the edge cluster, and are used to complete the deployment process of the application artifacts.

5. The application product classification and management method based on deep learning and dynamic / static features as described in claim 4, characterized in that: The static characteristics and dynamic characteristics of the image, include, By filtering container application records that meet the filtering strategy from the application's container application record set, extracting the image identifier contained in the container application record, and classifying them based on the image feature information of the image corresponding to the image identifier, at least one image subclass is obtained. Extract the feature image for image testing of the application from the image subclass. Check whether the feature image supports downloading from the image repository to the local machine in the image download dimension. If yes, check whether the feature image supports being viewed in the image viewing dimension. If yes, check whether the feature image supports being started in the image startup dimension. If yes, check whether the feature image supports being deleted in the image deletion dimension. And / or check whether the residual information after the feature image is deleted is cleared. If yes, confirm that the feature image passes the image test.

6. The application product classification and management method based on deep learning and dynamic / static features as described in claim 4, characterized in that: The time series modeling structure is configured to perform time series modeling on the static features and output a residual sequence; The residual sequence and the dynamic feature sequence output by the transformer coding structure are aligned in the same time step in the time dimension, and the sequences are concatenated at corresponding positions. The concatenated sequence is used as input and is sequentially fed into each time step of the recurrent neural network structure. At each time step, the recurrent neural network structure updates its state based on the current input and the hidden state of the previous time step, and outputs joint features.

7. The application product classification and management method based on deep learning and dynamic / static features as described in claim 4, characterized in that: The determination of deployment location and resource allocation scheme include, The classification results and the availability domain resource status of the edge cluster are used as deployment parameters, wherein the classification results are used to determine the set of deployment candidate nodes, and the availability domain resource status is used to provide the current resource information of the candidate nodes; Based on the classification results, the deployment type and node feature labels are matched to select nodes that meet the category requirements from the set of candidate deployment nodes; The resource information of the selected candidate nodes is compared, and the target deployment node is determined according to the preset resource sorting rules; Based on the resource details of the target deployment node, resources are allocated to the application artifacts, and deployment is completed.

8. An application product classification management system based on deep learning and dynamic / static features, employing the application product classification management method based on deep learning and dynamic / static features as described in any one of claims 1 to 7, characterized in that, include: The module includes a static feature extraction module, a dynamic feature extraction module, a fusion module, a classification module, and a deployment module. The static feature extraction module is used to extract the static features of the application artifacts after the scheduling edge cluster receives the application artifacts. The dynamic feature extraction module is used to extract the dynamic features of the work-in-process after it has been running. The fusion module is used to extract static and dynamic features of the image using a neural network model; The classification module is used to classify the applied products based on the extracted features; The deployment module is used to determine the deployment location and resource allocation scheme based on the classification results and the availability of resources in the edge cluster, and to complete the deployment of application artifacts.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the application product classification management method based on deep learning and dynamic and static features as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the application product classification management method based on deep learning and dynamic and static features as described in any one of claims 1 to 7.

Images