Ai inference method, system, and computer-readable storage medium

Abstract

The application discloses an AI inference method, system and computer readable storage medium, the AI inference system comprises a system master control and a distributed line card, the system master control obtains the inference result of the distributed line card, the inference result is obtained by the first level inference of the performance data obtained by the distributed line card; the second level inference is carried out according to the inference result, and the global inference result is obtained. Based on this, by adopting the two-level AI inference architecture of the system master control and the distributed line card, the AI application can be flexibly extended like a plug-in, the distributed line card has the preliminary first level AI inference function, the system master control has the global second level AI inference function, so that the distributed AI computing capacity of the whole AI inference system is improved, the AI inference system is decoupled from the AI application, the system integration capacity of various AI applications is effectively improved, therefore, the application can support the flexible expansion of communication network element equipment in different AI applications.

Description

Technical Field

[0001] The embodiments of the present invention relate to, but are not limited to, the field of AI artificial intelligence in network devices, and in particular to an AI inference method, system and computer-readable storage medium. Background Technology

[0002] In the 5G era, the Internet of Everything is becoming increasingly complex, and the original static manual management methods are no longer applicable. Operators and network equipment manufacturers are promoting the application of AI in the field of communications, making full use of big data information to achieve intelligent network management, such as intelligent energy saving, intelligent operation and maintenance, and network security. Among these, the intelligentization of communication network elements has become a key focus of development.

[0003] Communication network elements generate massive amounts of dynamic data every moment, including dynamic performance data, alarms, logs, and other types of information. From the perspective of data source, they can be divided into different hardware units such as network processors, TMs, line cards, switching networks, and CPUs. From the perspective of the time characteristics of information, they can be divided into time-variable information and static invariant information (such as some fixed information of hardware). Time-variable performance statistics are mainly traffic statistics generated in real time by forwarding chips, which have typical time series characteristics. Different AI applications are all based on the analysis and prediction of data, and all involve common processes such as data sample collection, fusion management, AI inference, training, and learning.

[0004] Currently, the industry lacks AI inference applications for communication network elements. AI applications are relatively limited, and implementation requires customized hardware and software architectures for specific smart terminal devices, along with proprietary data acquisition channels and data management models. The AI ​​inference architecture is strongly coupled with data features. However, for communication network elements, the data that different AI applications focus on cannot be obtained directly. It is necessary to fuse feature data from different spatial and temporal dimensions. Therefore, the application of proprietary AI inference architectures has scalability issues and cannot meet the flexible expansion requirements of different AI applications on communication network elements. Summary of the Invention

[0005] The following is an overview of the subject matter described in detail herein. This overview is not intended to limit the scope of the claims.

[0006] This invention provides an AI inference method, system, and computer-readable storage medium that can support the flexible expansion of communication network element devices in different AI applications.

[0007] In a first aspect, embodiments of the present invention provide an AI inference method applied to the system master controller of an AI inference system, wherein the system master controller is connected to a distributed line card, and the method includes:

[0008] The inference result of the distributed line card is obtained, which is obtained by the distributed line card performing a first-level inference on the acquired performance data;

[0009] Based on the reasoning results, a second level of reasoning is performed to obtain the global reasoning result.

[0010] Secondly, embodiments of the present invention provide an AI inference method applied to a distributed line card in an AI inference system, wherein the distributed line card is connected to the system main controller, and the method includes:

[0011] Obtain performance data from different data collectors;

[0012] The performance data is subjected to first-level inference to obtain the inference result;

[0013] The inference result is pushed to the system master controller, so that the system master controller can perform a second-level inference based on the inference result to obtain a global inference result.

[0014] Thirdly, embodiments of the present invention provide a system master controller, the system master controller being connected to distributed line cards, the system master controller comprising:

[0015] The main control AI data management unit is used to obtain the inference results of the distributed line card, wherein the inference results are obtained by the distributed line card performing a first-level inference on the acquired performance data;

[0016] The main AI inference unit is used to perform second-level inference based on the inference results to obtain the global inference results.

[0017] Fourthly, embodiments of the present invention provide a distributed line card, the distributed line card being connected to a system main controller, the distributed line card comprising:

[0018] The line card AI data management unit is used to acquire performance data from different data collectors;

[0019] The line card AI inference unit is used to perform first-level inference on the performance data, obtain inference results, and push the inference results to the system master controller, so that the system master controller can perform second-level inference based on the inference results to obtain global inference results.

[0020] Fifthly, embodiments of the present invention provide an AI inference system, including the system controller as described in the third aspect above and the distributed line card as described in the fourth aspect above.

[0021] Fifthly, embodiments of the present invention provide a computer-readable storage medium storing a computer-executable program for causing a computer to perform the AI ​​inference method as described in the first aspect above, or the AI ​​inference method as described in the second aspect above.

[0022] This invention includes an AI inference system comprising a system master controller and distributed line cards. The system master controller acquires the inference results from the distributed line cards, which are obtained by performing a first-level inference on the acquired performance data. A second-level inference is then performed based on the inference results to obtain a global inference result. Based on this, by adopting a two-level AI inference architecture of system master controller and distributed line cards, AI applications can be flexibly expanded like plug-ins. The distributed line cards have the function of performing preliminary first-level AI inference, while the system master controller has the function of performing global second-level AI inference, thereby improving the overall distributed AI computing capability of the AI ​​inference system. Simultaneously, it achieves an AI inference system decoupled from AI applications, effectively improving the system integration capability of various AI applications. Therefore, this invention can support the flexible expansion of communication network element equipment for different AI applications.

[0023] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the description, claims, and drawings. Attached Figure Description

[0024] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0025] Figure 1 This is a main flowchart (system control side) of an AI inference method provided in one embodiment of the present invention;

[0026] Figure 2 This is a schematic diagram of the AI ​​inference system structure of a communication network element device provided in one embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of the AI ​​training and learning structure of a communication network element device and a cloud management and control server provided in one embodiment of the present invention;

[0028] Figure 4 This is a sub-flowchart of an AI inference method provided in one embodiment of the present invention;

[0029] Figure 5 This is another sub-flowchart of an AI inference method provided in one embodiment of the present invention;

[0030] Figure 6 This is a main flowchart of an AI inference method provided in one embodiment of the present invention (distributed line card side);

[0031] Figure 7 This is a sub-flowchart of an AI inference method provided in one embodiment of the present invention;

[0032] Figure 8 This is another sub-flowchart of an AI reasoning method provided in one embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0034] It should be understood that in the description of the embodiments of the present invention, "multiple" (or "amounts") means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. If "first," "second," etc., are used in the description, they are only for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0035] In the 5G era, the Internet of Everything is becoming increasingly complex, and the original static manual management methods are no longer applicable. Operators and network equipment manufacturers are promoting the application of AI in the field of communications, making full use of big data information to achieve intelligent network management, such as intelligent energy saving, intelligent operation and maintenance, and network security. Among these, the intelligentization of communication network elements has become a key focus of development.

[0036] Communication network elements generate massive amounts of dynamic data every moment, including dynamic performance data, alarms, logs, and other types of information. From the perspective of data source, they can be divided into different hardware units such as network processors, TMs, line cards, switching networks, and CPUs. From the perspective of the time characteristics of information, they can be divided into time-variable information and static invariant information (such as some fixed information of hardware). Time-variable performance statistics are mainly traffic statistics generated in real time by forwarding chips, which have typical time series characteristics. Different AI applications are all based on the analysis and prediction of data, and all involve common processes such as data sample collection, fusion management, AI inference, training, and learning.

[0037] Currently, the industry lacks AI inference applications for communication network elements. AI applications are relatively limited, and implementation requires customized hardware and software architectures for specific smart terminal devices, along with proprietary data acquisition channels and data management models. The AI ​​inference architecture is strongly coupled with data features. However, for communication network elements, the data that different AI applications focus on cannot be obtained directly. It is necessary to fuse feature data from different spatial and temporal dimensions. Therefore, the application of proprietary AI inference architectures has scalability issues and cannot meet the flexible expansion requirements of different AI applications on communication network elements.

[0038] To address the shortcomings of existing technologies, which focus on specific AI sub-fields and lack a unified AI data acquisition and management solution at the system architecture level, thus failing to support flexible expansion for different AI applications, this invention provides an AI inference method, system, and computer-readable storage medium. The AI ​​inference system includes a system controller and distributed line cards. The system controller acquires the inference results from the distributed line cards, which perform first-level inference on the acquired performance data. Second-level inference is then performed based on the inference results to obtain the global inference result. Based on this, by adopting a two-level AI inference architecture of system controller and distributed line cards, AI applications can be flexibly expanded like plug-ins. The distributed line cards have initial first-level AI inference capabilities, while the system controller has global second-level AI inference capabilities, thereby improving the overall distributed AI computing power of the AI ​​inference system. Simultaneously, it achieves an AI inference system decoupled from AI applications, effectively enhancing the system integration capabilities of various AI applications. Therefore, this invention can support the flexible expansion of communication network elements for different AI applications.

[0039] like Figure 1 As shown, Figure 1 This is a flowchart illustrating an AI inference method according to an embodiment of the present invention. The AI ​​inference method is applied to the system master controller of an AI inference system, which is connected to distributed line cards. The AI ​​inference method includes, but is not limited to, the following steps:

[0040] Step 101: Obtain the inference results of the distributed line card. The inference results are obtained by the distributed line card performing first-level inference on the acquired performance data.

[0041] Step 102: Perform second-level reasoning based on the reasoning results to obtain the global reasoning results.

[0042] Understandably, the distributed line cards perform first-level inference on the acquired performance data to obtain the inference result. The system master controller then acquires the inference result from the distributed line cards and performs second-level inference based on this result to obtain the global inference result. Based on this, by adopting a two-level AI inference architecture of system master controller and distributed line cards, AI applications can be flexibly expanded like plug-ins. The distributed line cards have the function of performing initial first-level AI inference, while the system master controller has the function of performing global second-level AI inference, thereby improving the overall distributed AI computing capability of the AI ​​inference system. Simultaneously, it achieves an AI inference system decoupled from AI applications, effectively improving the system integration capability of various AI applications. Therefore, this invention can support the flexible expansion of communication network element equipment for different AI applications.

[0043] It is understandable that an AI inference system includes a system master controller and distributed line cards. The distributed line cards have initial first-level AI inference capabilities, while the system master controller has global second-level AI inference capabilities. Therefore, in terms of architecture, first-level inference units can be set up on the distributed line cards, and second-level inference units can be set up on the system master controller. In addition, line card AI data management units can be set up on the distributed line cards, and master control AI data management units can be set up on the system master controller to handle the collection and management of data for different types of AI applications, respectively.

[0044] It is understandable that performance data for communication network elements is diverse and originates from numerous sources. The AI ​​data management unit can collect performance data using a subscription method. This invention primarily collects the following performance data:

[0045] (1) Network processor statistics: including port transmit and receive rates, packet rates of different byte lengths on the port, packet change rates of TCP, UDP and ICMP, and NP abnormal packet loss statistics.

[0046] (2) TM queue statistics: including queue depth change statistics, queue forwarding rate, and queue discarding rate.

[0047] (3) Flow statistics: TCP SYN, ACK, FIN and RST packet change rate.

[0048] (4) Alarm log statistics: including software write forwarding table error alarms, NP chip register abnormal alarms, protocol interaction abnormal alarms and key event logs.

[0049] The characteristics of the acquisition performance data are shown in the table below:

[0050] Subscribing to datasets illustrate Data ID Data type ID registered in the AI ​​data acquisition management unit Date Data collection date Time Data acquisition time, 1588 synchronized PTP time format Period Data sampling period, supports seconds and milliseconds. Slot Single board, slot information SourceType Data source type (NP, TM, ACL, alarm logs, etc.) Entity ID Data entity ID (interface, queue, ACL, alarm log ID, etc.) feature_1 Eigenvalues, in TLV format ... feature_n Eigenvalues, in TLV format

[0051] Understandably, a master AI application management unit can be set up in the system master control for AI application registration. AI application requirements adopt a registration process. The master AI application management unit registers the data processing requirements of the AI ​​application with the master AI data management unit. The master AI data management unit determines the registration source, spatiotemporal data feature set, periodic fusion method, and templated pipeline feature engineering processing flow based on the AI ​​application's registration information. Detailed AI application registration information is shown in the table below:

[0052]

[0053]

[0054] The Action Type primarily specifies the type of preprocessing action to be performed on the data, which may include the following feature engineering preprocessing actions:

[0055] (1) Missing value handling operations: Set actions for missing data for specific sample features, such as discarding, averaging, and filling with neighboring values.

[0056] (2) Data reduction operation: perform data reduction on features at different scales, such as 0-1 method or variance reduction.

[0057] (3) Data subset action: For specific collected data, only select a specified feature set to meet the requirements of AI applications.

[0058] (3) Data union action: Merge multiple different types of sample data according to time series to meet the requirements of AI applications.

[0059] (4) Feature mathematical operations: Perform mathematical transformation operations on specific features to meet the data requirements of AI applications.

[0060] (5) Data statistics operation: Perform analog statistical calculations on the data according to a specific period.

[0061] (6) Time fusion method: For data fusion actions of different periods, the main issue is how to align the time when merging different data, including downward time alignment (taking the previous data of the larger period) and upward alignment (taking the next data of the larger period).

[0062] It is understandable that, such as Figure 2 As shown, this invention realizes a distributed AI inference architecture based on a general data management platform through an AI application registration mechanism and a data subscription mechanism. The detailed steps are described below:

[0063] Step 1: The distributed line cards and main controller on the communication network element device enable the data management platform and access different types of performance data collectors through subscription. Different performance collectors push real-time data to the data management platform according to their own periodic characteristics.

[0064] Step 2: The main control AI application management unit determines the inference mode based on the characteristics of the AI ​​application, namely the system main control global AI inference and line card AI inference, and at the same time determines the subscription and registration process of AI application data.

[0065] Step 3: The data management unit receives the application processing requirements registered by the AI ​​application with the data management platform, including the scope of the feature set of interest, the fusion method, feature preprocessing, feature engineering, etc. The data management platform arranges the processing flow according to the pipeline mode.

[0066] Step 4: The data management platform stores the received performance data locally according to the time series, and processes it through a pipeline process according to the data processing requirements of different AI applications. The processed real-time data is then pushed to the line card AI inference unit for AI inference.

[0067] Step 5: The line card AI inference unit performs relevant actions based on the AI ​​inference results and in conjunction with the AI ​​application, including pushing the line card AI inference results to the main control system for secondary inference.

[0068] Understandably, when the master AI data management unit meets the inference conditions, it pushes real-time data to the master AI inference unit for inference. In this way, the AI ​​application can be flexibly expanded like a plug-in. The AI ​​inference only focuses on its own algorithm and performance, while the data-related parts are provided by the platform.

[0069] Understandably, when AI applications require the system master controller to integrate data from all distributed line cards for global inference, in order to reduce the data processing load of the master controller and improve the real-time performance of the system master controller's AI inference, this invention adopts a two-level AI inference architecture. The distributed line cards complete the initial first-level AI inference function. Only when the global inference conditions are met are the first-level inference results of the distributed line cards sent to the system master controller for global second-level inference.

[0070] Understandably, the specific master AI data management unit subscribes to data from the line card AI inference unit, the distributed line cards complete the first level of AI inference, and the result serves as the data input for the system master's second level of AI inference.

[0071] It is understandable that, in order to improve the main control data processing capability, the main control AI data management unit can also subscribe to data from the line card AI data management unit. That is, the distributed line card will not directly send the raw data to the system main control, but will first perform feature engineering preprocessing on the collected data, and then send the processed data to the main control AI data management unit for corresponding processing.

[0072] It should be noted that, as Figure 3 As shown, the communication network element mainly realizes AI inference. The training and learning of the AI ​​model can also be realized by the cloud management server. The communication network element sends real-time raw data to the cloud management server through the telemetry network telemetry channel. The cloud management server performs AI training and learning, and then sends the trained model parameters to the communication network element to realize AI inference.

[0073] like Figure 4 As shown, the system main control includes a main control AI application management unit, and the following sub-steps may be included, but are not limited to, before step 101:

[0074] Step 401: The main AI application management unit releases the AI ​​application;

[0075] Step 402: The main control AI application management unit registers the application with the distributed line card to determine the application registration information.

[0076] Understandably, the master AI application management unit determines the inference mode based on the characteristics of the AI ​​application, namely, system master-controlled global AI inference and line card AI inference, while also determining the subscription and registration process for AI application data. For example, the master AI application management unit releases an AI anti-DDoS attack application and registers it with the line card AI data management unit. Anti-DDoS attack mainly identifies abnormal attack types by analyzing network traffic statistics changes through AI. Therefore, the data types in the registration information include the port's uplink and downlink TCP SYN / ACK / FIN / RST packet change rate, the port's new source IP change rate, the port's total uplink rate, and the traffic period that needs to be aligned, data cleaning and preprocessing methods, etc. The registration source is the line card AI inference unit.

[0077] like Figure 5 As shown, the system master control includes a master control AI data management unit and a master control AI inference unit. After step 101, the following sub-steps may be included, but are not limited to:

[0078] Step 501: The main control AI data management unit queries the application registration information from the distributed line card;

[0079] Step 502: The main control AI data management unit preprocesses the inference results according to the application registration information to obtain preprocessed data, wherein the preprocessing includes feature engineering preprocessing.

[0080] Step 503: The master AI data management unit pushes the preprocessed data to the master AI inference unit.

[0081] Understandably, the main control AI data management unit queries application registration information from the distributed line card. Based on this information, the main control AI data management unit preprocesses the inference results to meet the data requirements of the AI ​​application. This preprocessing includes feature engineering preprocessing, which may include the following actions:

[0082] (1) Missing value handling operations: Set actions for missing data for specific sample features, such as discarding, averaging, and filling with neighboring values.

[0083] (2) Data reduction operation: perform data reduction on features at different scales, such as 0-1 method or variance reduction.

[0084] (3) Data subset action: For specific collected data, only select a specified feature set to meet the requirements of AI applications.

[0085] (3) Data union action: Merge multiple different types of sample data according to time series to meet the requirements of AI applications.

[0086] (4) Feature mathematical operations: Perform mathematical transformation operations on specific features to meet the data requirements of AI applications.

[0087] (5) Data statistics operation: Perform analog statistical calculations on the data according to a specific period.

[0088] (6) Time fusion method: For data fusion actions of different periods, the main issue is how to align the time when merging different data, including downward time alignment (taking the previous data of the larger period) and upward alignment (taking the next data of the larger period).

[0089] When the preprocessed data meets the global inference conditions, the master AI data management unit pushes the preprocessed data to the master AI inference unit for corresponding processing in order to perform global secondary inference.

[0090] Understandably, after obtaining the global inference result, the system master controller can execute actions corresponding to the global inference result.

[0091] Understandably, the master AI data management unit can receive inference results from the distributed line cards via a subscription method.

[0092] like Figure 6 As shown, Figure 6This is a flowchart illustrating an AI inference method according to an embodiment of the present invention. The AI ​​inference method is applied to a distributed line card in an AI inference system. The distributed line card is connected to the system main controller. The AI ​​inference method includes, but is not limited to, the following steps:

[0093] Step 601: Obtain performance data from different data acquisition devices;

[0094] Step 602: Perform first-level inference on the performance data to obtain the inference results;

[0095] Step 603: Push the inference results to the system master controller so that the system master controller can perform second-level inference based on the inference results and obtain the global inference results.

[0096] It is understandable that, such as Figure 2 As shown, the distributed line card acquires performance data from different data collectors, including but not limited to user flow statistics collectors, TM queue statistics collectors, PORT traffic data collectors, NP chip status data collectors, and chip anomaly data collectors. The distributed line card performs first-level inference on the performance data to obtain the inference result. The distributed line card then pushes the inference result to the system master controller, enabling the system master controller to perform second-level inference based on the result, obtaining the global inference result. Based on this, by adopting a two-level AI inference architecture of system master controller and distributed line card, AI applications can be flexibly expanded like plug-ins. The distributed line card has the function of performing preliminary first-level AI inference, while the system master controller has the function of performing global second-level AI inference, thereby improving the overall distributed AI computing capability of the AI ​​inference system. Simultaneously, it achieves an AI inference system decoupled from AI applications, effectively improving the system integration capability of various AI applications. Therefore, this invention can support the flexible expansion of communication network element equipment for different AI applications.

[0097] like Figure 7 As shown, step 601 may include, but is not limited to, the following sub-steps:

[0098] Step 701: Collect performance data periodically from different data collectors via subscription;

[0099] Step 702: Cache the performance data according to quantity, type, and time series.

[0100] Understandably, the line card AI data management unit subscribes to push notifications of real-time performance data from different data acquisition units on the line card via a subscription method, thereby achieving the collection and management of raw data. Different data acquisition units collect data at regular intervals according to their respective settings and synchronously send the real-time data to the line card AI data management unit, which then caches the data based on its quantity, type, and time series.

[0101] It is understandable that an AI inference system includes a system master controller and distributed line cards. The distributed line cards have initial first-level AI inference capabilities, while the system master controller has global second-level AI inference capabilities. Therefore, in terms of architecture, first-level inference units can be set up on the distributed line cards, and second-level inference units can be set up on the system master controller. In addition, line card AI data management units can be set up on the distributed line cards, and master control AI data management units can be set up on the system master controller to handle the collection and management of data for different types of AI applications, respectively.

[0102] It is understandable that before the line card AI data management unit performs the first-level inference on the performance data and obtains the inference result, it also includes querying application registration information, which is obtained by the system master controller registering applications with the distributed line card.

[0103] like Figure 8 As shown, step 602 may include, but is not limited to, the following sub-steps:

[0104] Step 801: The line card AI data management unit preprocesses the performance data according to the application registration information to obtain preprocessed data, wherein the preprocessing includes feature engineering preprocessing.

[0105] Step 802: The line card AI data management unit pushes the preprocessed data to the line card AI inference unit;

[0106] Step 803: The line card AI inference unit performs the first level of inference on the preprocessed data to obtain the inference result.

[0107] It is understandable that the line card AI data management unit preprocesses the performance data based on the application registration information to obtain preprocessed data. This preprocessing includes feature engineering preprocessing, which may include the following actions:

[0108] (1) Missing value handling operations: Set actions for missing data for specific sample features, such as discarding, averaging, and filling with neighboring values.

[0109] (2) Data reduction operation: perform data reduction on features at different scales, such as 0-1 method or variance reduction.

[0110] (3) Data subset action: For specific collected data, only select a specified feature set to meet the requirements of AI applications.

[0111] (3) Data union action: Merge multiple different types of sample data according to time series to meet the requirements of AI applications.

[0112] (4) Feature mathematical operations: Perform mathematical transformation operations on specific features to meet the data requirements of AI applications.

[0113] (5) Data statistics operation: Perform analog statistical calculations on the data according to a specific period.

[0114] (6) Time fusion method: For data fusion actions of different periods, the main issue is how to align the time when merging different data, including downward time alignment (taking the previous data of the larger period) and upward alignment (taking the next data of the larger period).

[0115] Once the preprocessed data meets the inference conditions, the line card AI data management unit pushes the preprocessed data to the line card AI inference unit. The line card AI inference unit then performs the first-level inference on the preprocessed data to obtain the inference result.

[0116] The AI ​​inference method provided by the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0117] like Figure 2 As shown, taking the implementation of DDoS attack mitigation through AI as an example, the corresponding steps for AI inference are as follows:

[0118] a. During system initialization, the line card and main control data acquisition unit, the line card AI data management unit and the main control AI data management unit, the main control AI application management unit, the line card AI inference unit and the main control AI inference unit are enabled, and each unit module operates independently.

[0119] b. The line card AI data management unit subscribes to push messages of real-time performance data from different data acquisition devices on the line card, thereby enabling the collection and management of raw data.

[0120] c. The AI ​​application management unit releases the AI ​​anti-DDoS attack application and registers it with the line card AI data management unit. Anti-DDoS attack mainly identifies abnormal attack types by analyzing network traffic statistics changes through AI. Therefore, the data types in the registration information include the port uplink and downlink TCP SYN / ACK / FIN / RST packet change rate, the port new source IP change rate, the port uplink total rate, as well as the traffic period to be aligned, data cleaning and preprocessing methods, etc. The registration source is the line card AI inference unit.

[0121] d. The master control AI data management unit subscribes to the line card AI inference unit for the line card inference results. The line card completes the first level of preliminary AI inference. When the line card AI inference indicates DDoS attack traffic, the line card AI inference unit sends the inference results to the master control data management unit, which then integrates the inference results from different line cards to make the final AI inference.

[0122] e. Different data collectors collect data at regular intervals according to their respective settings and send the real-time data synchronously to the line card AI data management unit. The AI ​​data management unit caches the data according to its quantity, type, and time series.

[0123] f. After receiving the real-time data update sent by the collector, the line card AI data management unit queries the registration information of the AI ​​anti-DDoS attack, performs pipeline processing on the collected raw data, and sends the data to the line card AI inference unit for inference after the data meets the inference conditions.

[0124] g. The line card AI inference unit executes corresponding actions based on the inference results and sends the inference results to the main control AI data management unit via real-time messages.

[0125] h. The main control AI data management unit receives the AI ​​inference results from the line card, queries the registration information of the AI ​​anti-DDoS attack, performs corresponding pipeline processing on the collected raw data, and pushes the processed real-time data to the main control AI inference unit for secondary global inference.

[0126] i. The main AI inference unit determines the type of abnormal attack based on the inference results and executes the corresponding actions.

[0127] This invention also provides a system master controller, which is connected to distributed line cards. The system master controller includes:

[0128] The main control AI data management unit is used to obtain the inference results of the distributed line cards. The inference results are obtained by the distributed line cards through first-level inference of the acquired performance data.

[0129] The master AI inference unit is used to perform second-level inference based on the inference results to obtain the global inference result. It can be understood that the distributed line cards perform first-level inference on the acquired performance data to obtain the inference result. The system master controller then acquires the inference result from the distributed line cards and can perform second-level inference based on that result to obtain the global inference result. Based on this, by adopting a two-level AI inference architecture of system master controller and distributed line cards, AI applications can be flexibly expanded like plug-ins. The distributed line cards have the function of performing preliminary first-level AI inference, while the system master controller has the function of performing global second-level AI inference, thereby improving the overall distributed AI computing capability of the AI ​​inference system. Simultaneously, it achieves an AI inference system decoupled from AI applications, effectively improving the system integration capability of various AI applications. Therefore, this invention can support the flexible expansion of communication network element equipment for different AI applications.

[0130] This invention also provides a distributed line card, which connects to the system main controller. The distributed line card includes:

[0131] The line card AI data management unit is used to acquire performance data from different data collectors;

[0132] The line card AI inference unit is used to perform first-level inference on performance data, obtain inference results, and push the inference results to the system master controller so that the system master controller can perform second-level inference based on the inference results to obtain global inference results.

[0133] It is understandable that the distributed line card acquires performance data from different data collectors, including but not limited to user flow statistics collectors, TM queue statistics collectors, PORT traffic data collectors, NP chip status data collectors, and chip anomaly data collectors. The distributed line card performs first-level inference on the performance data to obtain the inference result. The distributed line card then pushes the inference result to the system master controller, so that the system master controller can perform second-level inference based on the inference result to obtain the global inference result. Based on this, by adopting a two-level AI inference architecture of system master controller and distributed line card, AI applications can be flexibly expanded like plug-ins. The distributed line card has the function of performing preliminary first-level AI inference, and the system master controller has the function of performing global second-level AI inference, so as to improve the overall distributed AI computing capability of the AI ​​inference system, while realizing an AI inference system decoupled from AI applications, effectively improving the system integration capability of multiple AI applications. Therefore, this invention can support the flexible expansion of communication network element equipment in different AI applications.

[0134] like Figure 2 As shown, this embodiment of the invention also provides an AI inference system, including the aforementioned system controller and the aforementioned distributed line cards. The distributed line cards perform first-level inference on the acquired performance data to obtain the inference result. The system controller acquires the inference result from the distributed line cards and can then perform second-level inference based on the inference result to obtain the global inference result. Based on this, by adopting a two-level AI inference architecture of system controller and distributed line cards, AI applications can be flexibly expanded like plug-ins. The distributed line cards have the function of performing preliminary first-level AI inference, while the system controller has the function of performing global second-level AI inference, thereby improving the overall distributed AI computing capability of the AI ​​inference system. Simultaneously, it achieves an AI inference system decoupled from AI applications, effectively improving the system integration capability of various AI applications. Therefore, this invention can support the flexible expansion of communication network element devices in different AI applications.

[0135] Furthermore, embodiments of the present invention also provide a computer-readable storage medium storing a computer-executable program, which is executed by one or more control processors, for example, by one processor, causing the one or more processors to execute the AI ​​inference method described in the embodiments of the present invention, for example, to perform the above-described method. Figure 1Steps 101 to 102 in the method are as follows. Figure 4 Method steps 401 to 402, Figure 5 Method steps 501 to 503, or perform the above-described steps. Figure 6 Method steps 601 to 603, Figure 7 Method steps 701 to 702, Figure 8 In steps 801 to 803 of the method, the distributed line card performs first-level inference on the acquired performance data to obtain the inference result. The system master controller obtains the inference result from the distributed line card and can then perform second-level inference based on the inference result to obtain the global inference result. Based on this, by adopting a two-level AI inference architecture of system master controller and distributed line card, AI applications can be flexibly expanded like plug-ins. The distributed line card has the function of performing preliminary first-level AI inference, and the system master controller has the function of performing global second-level AI inference, so as to improve the overall distributed AI computing capability of the AI ​​inference system. At the same time, it realizes an AI inference system decoupled from AI applications, effectively improving the system integration capability of various AI applications. Therefore, this invention can support the flexible expansion of communication network element equipment in different AI applications.

[0136] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable programs, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable programs, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0137] The above provides a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. An AI inference method applied to a system master of an AI inference system, the system master comprising a master AI application management unit, the system master being connected to a distributed line card, the method comprising: the master AI application management unit publishing an AI application; the master AI application management unit registering the application to the distributed line card to determine application registration information; obtaining an inference result of the distributed line card, wherein the inference result is obtained by the distributed line card performing first-level inference on preprocessed data, the preprocessed data being obtained by the distributed line card preprocessing performance data obtained from different types of data collectors according to the application registration information, the preprocessing including feature engineering preprocessing; performing second-level inference according to the inference result to obtain a global inference result.

2. The method of claim 1, wherein, The system master comprises a master AI data management unit and a master AI inference unit, and after the step of obtaining the inference result of the distributed line card, the inference result being obtained by the distributed line card performing first-level inference on the obtained performance data, the method further comprises: the master AI data management unit querying the application registration information from the distributed line card; the master AI data management unit preprocessing the inference result according to the application registration information to obtain preprocessed data, wherein the preprocessing includes feature engineering preprocessing; the master AI data management unit pushing the preprocessed data to the master AI inference unit.

3. The method of claim 2, wherein, The step of obtaining the inference result of the distributed line card comprises: the master AI data management unit receiving the inference result from the distributed line card by subscription.

4. An AI inference method applied to a distributed line card of an AI inference system, the distributed line card being connected to a system master, the method comprising: querying application registration information of an AI application, the application registration information being obtained by the system master registering the application to the distributed line card; obtaining performance data from different types of data collectors; preprocessing the performance data according to the application registration information to obtain preprocessed data, wherein the preprocessing includes feature engineering preprocessing; performing first-level inference on the preprocessed data to obtain an inference result; pushing the inference result to the system master to enable the system master to perform second-level inference according to the inference result to obtain a global inference result.

5. The method of claim 4, wherein, The step of obtaining performance data from different types of data collectors comprises: timely collecting performance data from different types of data collectors by subscription; performing cache processing on the performance data according to quantity type and time sequence.

6. The method of claim 4, wherein, The distributed line card comprises a line card AI data management unit and a line card AI inference unit, and the steps of preprocessing the performance data according to the application registration information to obtain preprocessed data, and performing first-level inference on the preprocessed data to obtain an inference result, comprise: the line card AI data management unit preprocessing the performance data according to the application registration information to obtain preprocessed data; The line card AI data management unit pushes the preprocessed data to the line card AI inference unit. The line card AI inference unit performs first-level inference on the preprocessed data to obtain inference results. 7.A system master connected with distributed line cards, comprising: a master AI application management unit configured to publish an AI application, and to perform application registration to the distributed line cards to determine application registration information; a master AI data management unit configured to obtain inference results of the distributed line cards, wherein the inference results are obtained by the distributed line cards performing first-level inference on preprocessed data, the preprocessed data is obtained by preprocessing performance data collected from different types of data collectors according to the application registration information, and the preprocessing includes feature engineering preprocessing; a master AI inference unit configured to perform second-level inference according to the inference results to obtain global inference results. 8.A distributed line card connected with a system master, comprising: a line card AI data management unit configured to query application registration information of an AI application, to collect performance data from different types of data collectors, and to preprocess the performance data according to the application registration information to obtain preprocessed data, wherein the application registration information is obtained by the system master performing application registration to the distributed line card, and the preprocessing includes feature engineering preprocessing; a line card AI inference unit configured to perform first-level inference on the preprocessed data to obtain inference results, and to push the inference results to the system master to enable the system master to perform second-level inference according to the inference results to obtain global inference results. 9.An AI inference system comprising the system master of claim 7 and the distributed line card of claim 8.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer executable program for causing a computer to execute the AI inference method of any one of claims 1 to 3, or the AI inference method of any one of claims 4 to 6.

Images