An Active Testing Method for AI Computing Clusters Based on Edge Computing

By embedding test samples into edge computing nodes and generating AI task probes, shadow reasoning and state verification are performed, solving the problem of difficulty in determining abnormal locations in AI computing cluster testing, and achieving more accurate anomaly judgment and retest result generation.

CN123068258APending Publication Date: 2026-11-10JIANGSU HAINA ELECTRONICS TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611588075.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing AI computing cluster testing methods struggle to accurately pinpoint the exact location of anomalies within the inference path, and the lack of proactive testing mechanisms results in a lack of clear path guidelines for retesting.

Method used

By embedding test samples in edge computing nodes, an initial AI task probe is generated, and shadow inference is performed in the shadow inference chain to form a state AI task probe. The test state and inference path are checked to determine the abnormal location, and retesting is performed through the retesting link to generate active test results.

Benefits of technology

It achieves centralized carrying of test samples, target edge computing nodes, and status backfill locations, reducing interference with real business requests and improving the continuity and verifiability of abnormal location determination and retest link selection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068258A_ABST
    Figure CN123068258A_ABST
Patent Text Reader

Abstract

This invention discloses an active testing method for AI computing power clusters based on edge computing, belonging to the field of edge testing technology. The method includes: embedding test samples into an inference path and marking target edge computing nodes and state backfilling positions in the inference path to generate an initial AI task probe; sending the initial AI task probe into a preliminary test shadow inference chain, driving the preliminary test shadow inference chain to call an intelligent inference model to perform shadow inference on the test samples while separated from the main business inference chain, and writing the test state formed by the shadow inference back to the initial AI task probe to form a state AI task probe; selecting a retest link based on the anomaly location, and matching the state AI task probe to the retest shadow inference chain within the retest link. This invention improves the continuity and verifiability of anomaly location determination, retest link selection, and active test result generation through collaborative testing of the backfillable AI task probe and the preliminary test shadow inference chain.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of edge testing technology, and in particular to an active testing method for AI computing power clusters based on edge computing. Background Technology

[0002] As AI inference tasks gradually shift from centralized cloud environments to the edge, edge computing has become a crucial technological foundation supporting low-latency sensing, on-site intelligent analysis, and distributed computing power collaboration. In scenarios such as intelligent manufacturing, security monitoring, vehicle-road coordination, and energy dispatching, edge computing nodes typically need to handle real-time inference tasks involving image, text, voice, and sensor-collected data locally, forming inference paths oriented towards business requests by deploying intelligent inference models. To ensure collaborative operation across multiple nodes, AI computing clusters typically manage the execution status of intelligent inference models within edge computing nodes by combining model deployment registration, node call relationships, bypass inference paths, and runtime status recording. This ensures a recordable execution relationship between test samples entering the model, the model completing inference, and the test status return, thus providing fundamental support for the operational testing of AI computing clusters in edge computing environments.

[0003] Existing AI computing cluster tests typically rely on node resource status, model call results, and runtime logs as the basis for judgment. The testing process often remains at the level of single-point status recording, lacking an active testing mechanism that continuously corresponds the transmission position of test samples in the inference path, the shadow inference position of the intelligent inference model, and the return landing point of the test status. This makes it difficult to accurately determine the specific location of the anomaly in the inference path, resulting in a lack of clear path basis for subsequent retesting. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an active testing method for AI computing power clusters based on edge computing to solve the problem of accurately determining the specific location of anomalies in the inference path.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides an active testing method for AI computing power clusters based on edge computing, comprising: Using the inference path of the intelligent inference model in the edge computing power node as the test object, the test sample is embedded in the inference path, and the target edge computing power node and the state backfilling position are marked in the inference path to generate the initial AI task probe. The initial AI task probe is sent into the initial test shadow inference chain, which drives the initial test shadow inference chain to call the intelligent inference model to perform shadow inference on the test sample in a state that is separated from the main business inference chain. The test state formed by the shadow inference is written back to the initial AI task probe to form the state AI task probe. The test state in the state AI task probe is compared with the inference path of the intelligent inference model to determine the abnormal position of the test state in the inference path. Based on the selected retest link at the abnormal location, the state AI task probe is matched to the retest shadow inference chain in the retest link, the abnormal location in the inference path is retested, and the test status in the state AI task probe after retesting is summarized to generate active test results.

[0007] As a preferred embodiment of the AI ​​computing power cluster active testing method based on edge computing described in this invention, the test sample is any one of image content, text content, voice content, and sensor-collected content used by the intelligent inference model to perform shadow inference. The inference path is formed by sequentially connecting the path entry point before the test sample enters the intelligent inference model, the shadow inference position of the intelligent inference model, and the return landing point of the test state.

[0008] As a preferred embodiment of the AI ​​computing power cluster active testing method based on edge computing described in this invention, the generation of the initial AI task probe specifically includes: Locate the path entry point of the test sample before it enters the intelligent inference model in the inference path, and use the edge computing power node associated with the path entry point as a candidate edge computing power node. Verify the calling relationship between candidate edge computing power nodes and intelligent inference models in the inference path, and mark the candidate edge computing power nodes with valid calling relationships as target edge computing power nodes; Find the return point of the test state formed after the intelligent reasoning model completes shadow reasoning in the reasoning path, and mark the return point as the state backfilling position. Write the target edge computing node into the path entry of the inference path, and write the state backfill position into the return landing point of the inference path to form a marked inference path; The test sample, target edge computing node, and state backfill location are encapsulated into the labeled inference path to generate the initial AI task probe.

[0009] As a preferred embodiment of the AI ​​computing power cluster active testing method based on edge computing described in this invention, the formation of the state AI task probe specifically includes: Using the target edge computing node as the matching object, find the initial test shadow inference chain connected to the target edge computing node, and send the initial AI task probe into the initial test shadow inference chain; The initial test of the shadow inference chain divides the shadow transmission position of the test sample and the real business request position of the main business inference chain in the marked inference path, forming a shadow execution branch; The shadow execution branch calls the intelligent inference model in the target edge computing node to perform shadow inference on the test sample and obtain the shadow inference state; Organize the sample delivery state, model invocation state, and inference completion state in the shadow inference state into the test state of the initial AI task probe; The test state is written back to the state backfill position in the initial AI task probe to form the state AI task probe.

[0010] As a preferred embodiment of the AI ​​computing power cluster active testing method based on edge computing described in this invention, the initial test shadow inference chain is a bypass inference chain formed by connecting the intelligent inference model and the state backfilling position along the marked inference path within the target edge computing power node and separating it from the main business inference chain. The main business inference chain is the business inference path in the target edge computing node that receives real business requests and calls the intelligent inference model to perform inference processing on the real business requests.

[0011] As a preferred embodiment of the AI ​​computing power cluster active testing method based on edge computing described in this invention, the construction process of the intelligent inference model specifically includes: The sample content in the edge sample training set is matched with the sample task labels to form training samples; A lightweight inference network is constructed, which includes a sample entry layer, a shared feature compression layer, a task inference branch layer, an edge consistency verification layer, and an inference output layer. The training samples are fed into the sample entry layer, which converts them into sample features and then passes them to the shared feature compression layer. By using a shared feature compression layer, edge feature compression and feature fusion are performed on sample features to form shared inference features; The shared inference features are distributed to the task inference branch layer, and the shared inference features are inferred and mapped according to the sample task labels to obtain the branch inference results. The branch inference results and sample task labels are checked for consistency through the edge consistency verification layer. When the consistency verification is successful, the branch inference results are output through the inference output layer and the current lightweight inference network is determined as the intelligent inference model. When the consistency verification is unsuccessful, the branch inference results are sent back to the shared feature compression layer for feature reorganization. The reorganized shared inference features are then redistributed to the task inference branch layer for inference mapping. The network parameters of the lightweight inference network are adjusted according to the re-formed branch inference results, and the intelligent inference model is formed after the adjustment. Deploy the intelligent inference model to edge computing nodes and record the inference path of the intelligent inference model in the edge computing nodes.

[0012] As a preferred embodiment of the AI ​​computing power cluster active testing method based on edge computing described in this invention, the edge sample training set is a training set composed of sample content and corresponding sample task labels collected by edge computing nodes and used to train intelligent inference models.

[0013] As a preferred embodiment of the active testing method for AI computing power clusters based on edge computing described in this invention, the connection verification involves matching the test states in the state AI task probes with their preceding and following positions according to the inference path of the intelligent inference model, and verifying the continuity between the test states and the inference path.

[0014] As a preferred embodiment of the active testing method for AI computing power clusters based on edge computing described in this invention, the generation of active testing results specifically includes: Centered on the abnormal location in the reasoning path, adjacent reasoning segments are extracted from the reasoning path to form the retest range; Find the bypass inference path in the target edge computing node, align the bypass inference path with the retest range, and select the retest link; Write the abnormal locations in the inference path and the test states in the state AI task probes into the retest link to form a retest probe path. Locate the retest shadow inference chain that connects to the abnormal location in the retest probe path, and match the state AI task probe to the retest shadow inference chain. The shadow inference chain is retested by calling the intelligent inference model to perform retests on abnormal locations in the inference path, thus forming a retest test state. Write the retest status into the status backfill position in the status AI task probe to form the retested status AI task probe. Summarize the test states in the AI ​​task probe after retesting and generate proactive test results.

[0015] As a preferred embodiment of the AI ​​computing power cluster active testing method based on edge computing described in this invention, the retest shadow inference chain is a bypass inference chain formed by connecting the bypass inference path, intelligent inference model and state backfill position in the retest link.

[0016] The beneficial effects of this invention are as follows: Through collaborative testing of the backfillable AI task probe and the initial test shadow inference chain, test samples, target edge computing nodes, state backfilling locations, and test states are centrally carried in the same test carrier; the state AI task probe saves the connection between the test sample entering the inference path, the intelligent inference model executing shadow inference, and the test state writing back in the edge computing environment, so that active testing is transformed from node log tracking to state-based testing that checks along the inference path; at the same time, the initial test shadow inference chain is executed separately from the main business inference chain, which can reduce the interference of active testing on real business requests, and make the test state more accurately reflect the calling status and inference completion status of the intelligent inference model in the edge computing node, thereby improving the continuity and verifiability of anomaly location determination, retest link selection, and active test result generation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0018] Figure 1 This is a flowchart of an active testing method for AI computing power clusters based on edge computing.

[0019] Figure 2 Generate a graph for the initial AI task probe.

[0020] Figure 3 A graph is generated for the state AI task probe.

[0021] Figure 4 This is a closed-loop diagram for retesting abnormal locations. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an active testing method for AI computing power clusters based on edge computing, including the following steps: S1. Using the inference path of the intelligent inference model in the edge computing node as the test object, embed the test sample into the inference path, and mark the target edge computing node and the state backfilling position in the inference path to generate the initial AI task probe.

[0026] S1.1. Taking the inference path of the intelligent inference model in the edge computing node as the test object, the test sample is embedded into the inference path, specifically, The inference paths already registered by the intelligent inference model in the edge computing nodes are used as test objects. The test samples are any one of the following: image content, text content, voice content, and sensor-collected content for the intelligent inference model to perform shadow inference. The image content comes from the screen captured by the edge camera device, the text content comes from the business request text, the voice content comes from the audio collected by the edge sound pickup device, and the sensor-collected content comes from the recorded data of the edge sensor device.

[0027] The test sample is bound to the path entry point before entering the intelligent inference model, and a continuous transmission relationship is established along the path entry point, the shadow inference position of the intelligent inference model, and the return landing point of the test state. This allows the test sample to enter the shadow inference position of the intelligent inference model from the path entry point to complete shadow inference, and then the test state formed by shadow inference is received by the return landing point, thus completing the test sample embedding in the inference path. The inference path is the inference execution path formed by sequentially connecting the path entry point before the test sample enters the intelligent inference model, the shadow inference position of the intelligent inference model, and the return landing point of the test state.

[0028] S1.2. Locate the path entry point of the test sample before it enters the intelligent inference model within the inference path, and use the edge computing power nodes associated with the path entry point as candidate edge computing power nodes. Specifically, The path entry point of the test sample before entering the intelligent inference model is found in the inference path. The path entry point records the node identifier of the edge computing power node, the model identifier of the intelligent inference model, and the starting position of the call. The node identifier of the edge computing power node comes from the registration record formed when the intelligent inference model is deployed to the edge computing power node.

[0029] The node identifier recorded at the path entry point is checked against the node identifier of the edge computing node. If the node identifiers match, the corresponding edge computing node is selected as a candidate edge computing node. If the node identifiers do not match, the corresponding edge computing node is removed, and the node identifier of the next edge computing node is checked until the consistency between the node identifier recorded at the path entry point and the node identifier of the edge computing node is completed.

[0030] S1.3. Verify the calling relationship between candidate edge computing nodes and the intelligent inference model in the inference path, and mark the candidate edge computing nodes with valid calling relationships as target edge computing nodes. Specifically, When verifying the calling relationship between candidate edge computing nodes and intelligent inference models in the inference path, the node identifier of the candidate edge computing node, the model identifier of the intelligent inference model, and the starting position of the call are used as the verification objects. First, it is confirmed whether the node identifier of the candidate edge computing node is consistent with the node identifier recorded in the path entry record. Then, it is confirmed whether the model identifier of the intelligent inference model is consistent with the registration record of the intelligent inference model deployed in the candidate edge computing node. Finally, it is confirmed whether the starting position of the call can be connected to the shadow inference position of the intelligent inference model along the inference path.

[0031] If the node identifier is consistent, the model identifier is consistent, and the call start position can be connected to the shadow inference position of the intelligent inference model, the call relationship is considered valid, and the corresponding candidate edge computing power node is marked as the target edge computing power node; if any of the following occurs, the call relationship is considered invalid, the corresponding candidate edge computing power node is removed, and the next candidate edge computing power node is verified.

[0032] S1.4. Locate the return point in the inference path where the intelligent inference model completes shadow inference and forms a test state. Mark the return point as the state backfill location. Specifically, The return landing point is located after the shadow inference position of the intelligent inference model in the inference path, and is used to take over the test state formed by the shadow inference. The return landing point records the test state write address, state write field and write trigger flag. The test state write address comes from the registration record formed when the intelligent inference model is deployed to the target edge computing node. The state write field includes the sample transfer state field, the model call state field and the inference completion state field.

[0033] When the return landing point can receive the sample delivery status, model call status, and inference completion status, and there is a continuous connection between the return landing point and the shadow inference position of the intelligent inference model, the return landing point is marked as a status backfilling position; when the return landing point cannot receive any of the sample delivery status, model call status, and inference completion status, or there is no continuous connection between the return landing point and the shadow inference position of the intelligent inference model, the return landing point is not marked as a status backfilling position, the corresponding return landing point is removed, and the search for the next return landing point in the inference path continues.

[0034] S1.5. Write the target edge computing node into the path entry of the inference path, and write the state backfill position into the return landing point of the inference path to form a marked inference path. Specifically, Write the node identifier of the target edge computing node into the path entry of the inference path, so that the path entry can point to the target edge computing node that receives the test sample and transmits the action before entering the intelligent inference model; the node identifier of the target edge computing node comes from the verification result of the call relationship between the candidate edge computing node and the intelligent inference model in the inference path.

[0035] Write the test state of the state backfill location to the address, state write field, and write trigger flag to the return landing point of the inference path, so that the return landing point can take over the sample transmission state, model call state, and inference completion state formed by shadow inference; after writing, check the continuity relationship between the path entry, the shadow inference position of the intelligent inference model, and the return landing point. If the continuity relationship exists, a marked inference path is formed. If the continuity relationship does not exist, recheck the target edge computing node recorded at the path entry and the state backfill location recorded at the return landing point.

[0036] S1.6. Encapsulate the test sample, target edge computing node, and state backfill location into the labeled inference path to generate the initial AI task probe.

[0037] S2. Send the initial AI task probe into the initial test shadow inference chain, drive the initial test shadow inference chain to call the intelligent inference model to perform shadow inference on the test sample in a state that is separated from the main business inference chain, and write back the test state formed by shadow inference to the initial AI task probe to form the state AI task probe.

[0038] S2.1. Using the target edge computing node as the matching object, find the initial test shadow inference chain connected to the target edge computing node, and send the initial AI task probe into the initial test shadow inference chain. Specifically, Based on the node identifier of the target edge computing node, the system searches for the bypass inference chain connected to the marked inference path within the target edge computing node. The bypass inference chain records the link identifier, connection start point, shadow inference position of the intelligent inference model, and state backfill position. The link identifier comes from the registration record formed when the intelligent inference model is deployed on the target edge computing node.

[0039] When the starting point of the bypass inference chain is connected to the path entry of the marked inference path, and the bypass inference chain can sequentially connect the shadow inference position and the state backfill position of the intelligent inference model, the bypass inference chain is determined as the initial test shadow inference chain; when the starting point of the bypass inference chain is not connected to the path entry of the marked inference path, or the bypass inference chain cannot sequentially connect the shadow inference position and the state backfill position of the intelligent inference model, the corresponding bypass inference chain is removed, and the search for the next bypass inference chain continues.

[0040] The initial AI task probe is sent into the initial test shadow inference chain according to the path entry in the marked inference path, so that the test sample carried by the initial AI task probe enters the shadow inference position of the intelligent inference model.

[0041] Among them, the initial test shadow inference chain is a bypass inference chain formed by connecting the intelligent inference model and the state backfilling position along the marked inference path within the target edge computing power node, and separating it from the main business inference chain.

[0042] S2.2, The initial test of the shadow inference chain divides the shadow propagation position of the test sample and the actual business request position of the main business inference chain in the marked inference path, forming a shadow execution branch, specifically, The initial shadow inference chain, based on the path entry point, the shadow inference position of the intelligent inference model, and the state backfill position in the marked inference path, maps the test samples carried by the initial AI task probe to the position segment between the path entry point and the shadow inference position of the intelligent inference model, marking it as the shadow transmission position. The main business inference chain, based on the business entry point in the target edge computing node that receives the real business request, the real inference position of the intelligent inference model, and the real business request return position, marks the real business request position. The shadow transmission position and the real business request position are distinguished by different link identifiers, which are derived from the registration records formed when the intelligent inference model is deployed on the target edge computing node.

[0043] When the shadow delivery position does not occupy the real business request position, the initial test shadow inference chain will connect the shadow delivery position, the shadow inference position of the intelligent inference model, and the state backfill position into a shadow execution branch; when the shadow delivery position occupies the real business request position, the formation of the shadow execution branch will stop, and a shadow delivery position that does not occupy the real business request position will be selected again.

[0044] The main business inference chain is a business inference path that receives real business requests in the target edge computing node and calls the intelligent inference model to perform inference processing on the real business requests.

[0045] S2.3 The shadow execution branch calls the intelligent inference model in the target edge computing node to perform shadow inference on the test sample and obtain the shadow inference state, specifically, The shadow execution branch sends the test sample to the shadow inference position of the intelligent inference model according to the shadow delivery position, and carries the node identifier of the target edge computing node in the initial AI task probe, the model identifier of the intelligent inference model, and the state backfill position.

[0046] The intelligent inference model checks whether the model identifier matches the registration record of the intelligent inference model deployed in the target edge computing node. If they match, it receives test samples and completes inference according to the test sample type. Specifically, after image content enters the image inference entry point, it undergoes image sample feature transformation and image inference branch mapping to form an image branch inference result; after text content enters the text inference entry point, it undergoes text sample feature transformation and text inference branch mapping to form a text branch inference result; after speech content enters the speech inference entry point, it undergoes speech sample feature transformation and speech inference branch mapping to form a speech branch inference result; after sensor-collected content enters the sensor inference entry point, it undergoes sensor sample feature transformation and sensor inference branch mapping to form a sensor branch inference result. The intelligent inference model outputs the corresponding branch inference result through the inference output layer to form an inference output containing the inference task category and output confidence.

[0047] When the model identifier is inconsistent, the reception of test samples is stopped, and a model call exception state is formed. The model call exception state is written into the model call state field in the state backfill position. The model call exception state is an abnormal value of the model call state. Together with the sample transmission state recorded when the test sample arrives at the shadow inference position, it is organized into the shadow inference state. When the test sample arrives at the shadow inference position of the intelligent inference model from the shadow transmission position, the sample transmission state is recorded. When the intelligent inference model completes reception and starts inference, the model call state is recorded. When the intelligent inference model forms inference output, the inference completion state is recorded. The sample transmission state, model call state and inference completion state are organized into the shadow inference state.

[0048] S2.4 The construction process of the intelligent reasoning model is as follows: S2.4.1. The sample content and sample task labels in the edge sample training set are matched and organized to form training samples. The edge sample training set is a training set composed of sample content and corresponding sample task labels collected by edge computing nodes and used to train intelligent inference models.

[0049] S2.4.2 Construct a lightweight inference network. The lightweight inference network includes a sample entry layer, a shared feature compression layer, a task inference branch layer, an edge consistency verification layer, and an inference output layer. Specifically, The lightweight inference network is built based on existing technologies such as convolutional layers, embedding layers, fully connected layers, batch normalization layers, activation layers, and depthwise separable convolutions.

[0050] The sample entry layer is set up with image entry, text entry, voice entry and sensor acquisition entry according to the sample content category. The image entry converts image content into image sample features, the text entry converts text content into text sample features, the voice entry converts voice content into voice sample features, and the sensor acquisition entry converts sensor acquisition content into sensor sample features.

[0051] The shared feature compression layer uses depthwise separable convolution, linear dimensionality reduction, batch normalization layers, and activation layers to compress sample features, making the sample features suitable for running on edge computing nodes.

[0052] The task reasoning branch layer uses a fully connected layer to generate branch reasoning results according to the task category corresponding to the sample task label; the edge consistency verification layer verifies the consistency between the branch reasoning results and the sample task label.

[0053] The inference output layer outputs the branch inference results that have passed the consistency check, thus forming a lightweight inference network that includes a sample entry layer, a shared feature compression layer, a task inference branch layer, an edge consistency check layer, and an inference output layer.

[0054] The training process is as follows: the sample content in the edge sample training set is used as the input sample, and the sample task label is used as the supervision signal to construct the training sample. The sample content includes any one of image content, text content, voice content and sensor acquisition content. The sample task label is used to label the inference task category corresponding to the sample content.

[0055] After training samples are fed into the lightweight inference network, the sample entry layer feeds in image content, text content, speech content, and sensor-acquired content according to their content categories. The image entry layer arranges image content by pixel blocks and maps it to image sample features via a convolutional layer. The text entry layer arranges text content by word number and maps it to text sample features via an embedding layer. The speech entry layer segments speech content into acoustic spectral vectors of a fixed frame length and maps them to speech sample features via a convolutional layer. The sensor-acquired content is processed according to the acquisition time. The features are arranged into temporal vectors and mapped to sensor sample features through a fully connected layer. The shared feature compression layer performs depthwise separable convolution, linear dimensionality reduction, batch normalization, and activation mapping on image sample features, text sample features, speech sample features, and sensor sample features. This process organizes the different sample features into features of the same length and then concatenates and fuses them to form shared inference features. The task inference branch layer sends the shared inference features to the corresponding task inference branch according to the inference task category corresponding to the sample task label. The branch inference result, which includes the inference task category and output confidence, is obtained through mapping by the fully connected layer.

[0056] The task inference loss term is determined based on the difference between the branch inference result and the sample task label, the edge compression loss term is determined based on the feature compression deviation of the shared inference features during the operation of the edge computing node, and the consistency verification loss term is determined based on the consistency verification result output by the edge consistency verification layer.

[0057] Among them, the task inference loss term is the category deviation value formed by comparing the inference task category and output confidence in the branch inference result with the inference task category labeled in the sample task label, and is used to represent the loss term of category deviation between the branch inference result and the sample task label; the edge compression loss term is the information retention deviation value formed by comparing the retention of the main features after aligning the shared inference features with the sample features before compression with the same feature length, and is used to represent the loss term of information retention deviation between the shared inference features after edge feature compression and the sample features before compression; the consistency verification loss term is the consistency deviation value formed by comparing the consistency verification result output by the edge consistency verification layer with the correct verification result corresponding to the sample task label, and is used to represent the loss term of consistency deviation between the consistency verification result output by the edge consistency verification layer and the correspondence between the sample task label.

[0058] The task inference loss term, edge compression loss term, and consistency verification loss term are combined according to their weights to form a joint loss function for the lightweight inference network. The weights are set based on the proportion of inference task categories in the edge sample training set and the proportion of edge computing node operation constraints. For example, the weight of the task inference loss term is 0.6, the weight of the edge compression loss term is 0.2, and the weight of the consistency verification loss term is 0.2. The weights are set according to the primary and secondary relationship between the accuracy of the inference task, the edge computing node operation compression constraints, and the output consistency verification constraints in the training objectives. The weight of the task inference loss term is 0.6 because the consistency between the branch inference result and the sample task label is the main training objective. The weights of the edge compression loss term and the consistency verification loss term are each 0.2 because feature compression adaptation and output consistency verification are used to constrain the stable operation of shared inference features in edge computing nodes, which can take the accuracy of the inference task as the main optimization direction, while retaining the constraint effect of edge operation compression adaptation and output consistency verification on the network parameter update.

[0059] The direction of network parameter adjustment is determined by the joint loss function of the lightweight inference network. Specifically, the decreasing direction of the joint loss function is used as the adjustment direction. The influence of network parameters in the sample entry layer, shared feature compression layer, task inference branch layer, edge consistency check layer, and inference output layer on the joint loss function of the lightweight inference network is evaluated. When a change in a network parameter causes the joint loss function of the lightweight inference network to decrease, the corresponding network parameter is adjusted in the decreasing direction. When a change in a network parameter causes the joint loss function of the lightweight inference network to increase, the corresponding network parameter is adjusted in the opposite direction. The network parameters in the lightweight inference network are updated using gradient backpropagation. Training stops when the change in the joint loss function of the lightweight inference network between consecutive training rounds is lower than the loss convergence threshold. The loss convergence threshold is set based on the initial change in the joint loss function of the lightweight inference network, for example, 10% of the initial change. This avoids the lightweight inference network from continuing ineffective training when the loss decrease has slowed down, while preserving sufficient space for network parameter updates. After training stops, the intelligent inference model is obtained.

[0060] S2.4.3. The training samples are fed into the sample entry layer, where they are converted into sample features. These features are then passed to the shared feature compression layer. Specifically, After the training samples enter the sample entry layer, the sample entry layer enters the corresponding entry according to the sample content category. The image content is arranged into a pixel matrix by fixed pixel blocks and mapped into image sample features by a convolutional layer. The text content is arranged by word number and mapped into text sample features by an embedding layer. The speech content is segmented into acoustic spectrum vectors by a fixed frame length and mapped into speech sample features by a convolutional layer. The sensor acquisition content is arranged into a time-series vector by acquisition time and mapped into sensor sample features by a fully connected layer.

[0061] Image sample features, text sample features, speech sample features, and sensor sample features are all organized into sample features of the same feature length. The feature length can be set according to the available storage capacity of the edge computing node. For example, the feature length can be 128-dimensional, 256-dimensional, and 512-dimensional, which can be adapted to edge computing nodes with low, medium, and high computing power, respectively. While compressing the feature occupancy, the main inference information of the sample content is retained, and the sample features are transferred to the shared feature compression layer.

[0062] S2.4.4. The sample features are compressed at the edge and fused using a shared feature compression layer to form shared inference features. Specifically, Sample features enter the shared feature compression layer. Depthwise separable convolutions perform channel separation and point-by-point fusion of local correlations in the sample features, reducing the feature load on edge computing nodes. Linear dimensionality reduction compresses the sample features to a feature length corresponding to the available storage capacity of the edge computing nodes. For example, feature lengths of 128, 256, and 512 dimensions can correspond to edge computing nodes with low, medium, and high computing power, respectively, while controlling the feature load and retaining the main features of the sample content participating in the inference mapping. The batch normalization layer normalizes the feature amplitudes corresponding to different sample contents. The activation layer performs nonlinear mapping on the normalized features, retaining the effective features participating in the inference mapping. The batch normalization layer is an existing neural network layer used to organize the sample features corresponding to a batch of training samples to a similar numerical scale, reducing the impact of differences in the feature amplitudes of different samples on network parameter updates. The compressed image sample features, text sample features, speech sample features, and sensor sample features are concatenated and fused according to the same feature length to form shared inference features.

[0063] S2.4.5. Distribute the shared inference features to the task inference branch layer, and perform inference mapping on the shared inference features according to the sample task labels to obtain the branch inference results. Specifically, After the shared inference features enter the task inference branch layer, the task inference branch layer sends the shared inference features to the corresponding branches in the image inference branch, text inference branch, speech inference branch, and sensor inference branch according to the inference task category corresponding to the sample task label. The image inference branch uses a fully connected layer to perform category mapping on the shared inference features to form the image branch inference result; the text inference branch uses a fully connected layer to perform semantic category mapping on the shared inference features to form the text branch inference result; the speech inference branch uses a fully connected layer to perform acoustic category mapping on the shared inference features to form the speech branch inference result; and the sensor inference branch uses a fully connected layer to perform temporal state mapping on the shared inference features to form the sensor branch inference result. The task inference branch layer uses the branch inference result corresponding to the sample task label as the branch inference result. The branch inference result includes the inference task category and the output confidence corresponding to the inference task category. The output confidence is used to indicate the degree of matching of the shared inference features mapped to the corresponding inference task category.

[0064] S2.4.6. The branch inference results and sample task labels are checked for consistency through the edge consistency verification layer. When the consistency verification is successful, the branch inference results are output through the inference output layer, and the current lightweight inference network is determined as the intelligent inference model. When the consistency verification is unsuccessful, the branch inference results are sent back to the shared feature compression layer for feature reorganization, and the reorganized shared inference features are redistributed to the task inference branch layer for inference mapping. The network parameters of the lightweight inference network are adjusted according to the re-formed branch inference results, and the intelligent inference model is formed after adjustment.

[0065] Among them, consistency verification is to check whether the reasoning task category in the branch reasoning result is consistent with the reasoning task category labeled in the sample task label, and to check whether the output confidence in the branch reasoning result meets the confidence requirement of the reasoning task category corresponding to the sample task label. The confidence requirement is the minimum confidence value determined by the historical output confidence distribution of the same reasoning task category in the edge sample training set, which is used to check whether the branch reasoning result can stably correspond to the sample task label.

[0066] S2.4.7 Deploy the intelligent inference model to the edge computing node and record the inference path of the intelligent inference model in the edge computing node.

[0067] S2.5 Organize the sample transfer state, model call state, and inference completion state in the shadow inference state into the test state of the initial AI task probe; that is, write the sample transfer state, model call state, and inference completion state into the state backfill position of the initial AI task probe according to the order of occurrence in the shadow execution branch to form the test state of the initial AI task probe.

[0068] S2.6 Write the test state back to the state backfill position in the initial AI task probe to form a state AI task probe; that is, write the test state into the corresponding field according to the state backfill position in the initial AI task probe, and retain the test sample, target edge computing node and marked inference path in the initial AI task probe to form a state AI task probe.

[0069] S3. Connect and verify the test state in the state AI task probe with the inference path of the intelligent inference model to determine the abnormal position of the test state in the inference path.

[0070] S3.1 When the state AI task probe enters the connection verification, the test state is broken down into sample transfer state, model call state and inference completion state, and respectively correspond to the path entry, shadow inference position and return landing point in the inference path of the intelligent inference model.

[0071] When the sample delivery status corresponds to the path entry point, the model call status corresponds to the shadow inference position, and the inference completion status corresponds to the return landing point, it is determined that the test status and the inference path are continuously connected; when the sample delivery status does not correspond to the path entry point, the path entry point is marked as an abnormal position; when the model call status does not correspond to the shadow inference position, the shadow inference position is marked as an abnormal position; when the inference completion status does not correspond to the return landing point, the return landing point is marked as an abnormal position.

[0072] When there are multiple uncorresponding positions, mark multiple abnormal positions according to the order of the path entry, shadow inference position and return landing point in the inference path, thereby determining the abnormal position of the test state in the inference path.

[0073] The connection verification involves matching the test states in the state AI task probe with the inference path of the intelligent inference model, verifying the continuity between the test states and the inference path. Specifically, it involves checking whether the sample transmission state corresponds to the transmission record between the path entry point, the shadow inference position, and the return landing point in the inference path, whether the model call state corresponds to the intelligent inference model call record at the shadow inference position, and whether the inference completion state corresponds to the output write-back record between the shadow inference position and the return landing point. If all three correspond in the correct order, the continuity between the test state and the inference path is considered to be valid. If any state is missing, misaligned, or fails to correspond to the previous or subsequent position, the corresponding position is marked as an abnormal position.

[0074] S4. Select a retest link based on the abnormal location, match the state AI task probe to the retest shadow inference chain in the retest link, perform retesting on the abnormal location in the inference path, and summarize the test status in the state AI task probe after retesting to generate active test results.

[0075] S4.1. Taking the abnormal location in the reasoning path as the center, extract adjacent reasoning segments in the reasoning path to form the retest range, specifically, Once the anomaly location in the inference path is determined, the forward and backward connection points of the anomaly location are located in the inference path of the intelligent inference model. The forward connection point is the location that directly transmits the test state before the anomaly location, and the backward connection point is the location that directly receives the test state after the anomaly location.

[0076] The forward continuation position, the abnormal position, and the backward continuation position together constitute an adjacent reasoning segment. When there is no forward continuation position, the path entry point of the reasoning path is used as the starting point of the adjacent reasoning segment. When there is a forward continuation position, the forward continuation position is used as the starting point of the adjacent reasoning segment. When there is no backward continuation position, the return point of the reasoning path is used as the ending point of the adjacent reasoning segment. When there is a backward continuation position, the backward continuation position is used as the ending point of the adjacent reasoning segment. The path content between the starting point, the abnormal position, and the ending point is organized into the retest range.

[0077] S4.2. Locate the bypass inference path in the target edge computing nodes, align the bypass inference path with the retest range, and select the retest link. Specifically, In the target edge computing node, a bypass inference path is searched according to the link identifier. The bypass inference path includes the bypass entry point, the bypass shadow inference position, and the bypass backfill position. The link identifier comes from the registration record formed when the intelligent inference model is deployed in the target edge computing node. The bypass entry point is aligned with the starting point of the retest range, the bypass shadow inference position is aligned with the abnormal position in the retest range, and the bypass backfill position is aligned with the ending point of the retest range. When the bypass entry point, bypass shadow inference position, and bypass backfill position can all be aligned with the retest range, the bypass inference path is selected as the retest link. If there are cases where the bypass entry point, bypass shadow inference position, and bypass backfill position are not aligned, the bypass inference path is excluded, and the search continues for the next bypass inference path in the target edge computing node.

[0078] S4.3. Write the abnormal locations and test states in the AI ​​task probes in the inference path into the retesting link to form a retesting probe path. Specifically, After the retest link is selected, the abnormal positions in the inference path are written into the bypass shadow inference position of the retest link, the test state in the state AI task probe is written into the bypass backfill position of the retest link, and the test samples in the state AI task probe are retained in the bypass entry of the retest link. After the bypass entry, bypass shadow inference position and bypass backfill position are written, the bypass entry, bypass shadow inference position and bypass backfill position are connected in series according to the order of the retest link to form the retest probe path.

[0079] S4.4. Locate the retest shadow inference chain that connects to the anomaly location in the retest probe path, and match the state AI task probe to the retest shadow inference chain. Specifically, Using the anomaly location in the inference path as an index, the bypass shadow inference location written to the same anomaly location is searched in the retest probe path, and the bypass entry and bypass backfill locations connected to the bypass shadow inference location are searched. When the bypass entry, bypass shadow inference location, and bypass backfill location can be sequentially connected along the retest probe path, and the bypass shadow inference location can call the intelligent inference model and the bypass backfill location can correspond to the state backfill location, the bypass inference chain between the bypass entry, bypass shadow inference location, and bypass backfill location is determined as the retest shadow inference chain. The test sample in the state AI task probe is written to the bypass entry of the retest shadow inference chain, and the test state in the state AI task probe is written to the bypass backfill location of the retest shadow inference chain, thus completing the matching between the state AI task probe and the retest shadow inference chain.

[0080] Among them, the retest shadow reasoning chain is a bypass reasoning chain formed by connecting the bypass reasoning path, intelligent reasoning model and state backfilling position in the retest link.

[0081] S4.5, Retesting the Shadow Inference Chain: The intelligent inference model is invoked to retest abnormal locations in the inference path, forming a retest test state. Specifically, The retest shadow inference chain reads test samples from the state AI task probe at the bypass entry point and sends the test samples to the bypass shadow inference position. The bypass shadow inference position calls the intelligent inference model to perform shadow inference on the test samples and limits the shadow inference to anomaly positions in the inference path. After the intelligent inference model completes shadow inference, the retest shadow inference chain records the retest sample transmission status when the test samples arrive at the bypass shadow inference position, the retest model call status when the intelligent inference model is called, and the retest inference completion status when the intelligent inference model forms inference output. The retest sample transmission status is quantified by the arrival marker, arrival time, and transmission time of the test samples arriving at the bypass shadow inference position. The retest model call status is quantified by the call trigger marker, call start time, call return time, and call time of the intelligent inference model. The retest inference completion status is quantified by the inference output formation marker, inference completion time, inference task category, and output confidence. The retest sample transmission status, retest model call status, and retest inference completion status are arranged according to the order of bypass entry point, bypass shadow inference position, and bypass backfill position to form the retest test status.

[0082] S4.6. Write the retest status into the status backfill position of the status AI task probe to form the retested status AI task probe. Specifically, Write the retest sample transmission status, retest model call status, and retest inference completion status into the status backfill position in the status AI task probe, respectively, and overwrite the original test status corresponding to the abnormal position in the status backfill position; retain the test sample, target edge computing node, marked inference path, and retest probe path in the status AI task probe; after the retest test status is written in the status backfill position, update the status AI task probe to the retested status AI task probe.

[0083] S4.7 Summarize the test states in the AI ​​task probe after retesting and generate active test results; that is, organize the original test states and retest test states in the AI ​​task probe after retesting according to the abnormal positions in the inference path, mark whether the abnormal positions still exist after retesting, and generate active test results.

[0084] This embodiment also provides a computer device applicable to the active testing method of AI computing power clusters based on edge computing, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the active testing method of AI computing power clusters based on edge computing as proposed in the above embodiment.

[0085] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0086] This embodiment also provides a storage medium on which a computer program is stored. When executed by a processor, the program implements the active testing method for AI computing power clusters based on edge computing as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0087] In summary, this invention achieves centralized hosting of test samples, target edge computing nodes, state backfilling locations, and test states within the same test carrier through collaborative testing of the backfillable AI task probe and the initial test shadow inference chain. The state AI task probe saves the connection between the test sample entering the inference path, the intelligent inference model executing shadow inference, and the test state writing back in the edge computing environment, transforming proactive testing from node log tracking to state-based testing that verifies along the inference path. At the same time, the separate execution of the initial test shadow inference chain and the main business inference chain reduces the interference of proactive testing on real business requests, making the test state more accurately reflect the calling status and inference completion status of the intelligent inference model in the edge computing nodes, thereby improving the continuity and verifiability of anomaly location determination, retest link selection, and proactive test result generation.

[0088] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended 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 proactive testing method for AI computing power clusters based on edge computing, characterized in that: include, Using the inference path of the intelligent inference model in the edge computing power node as the test object, the test sample is embedded in the inference path, and the target edge computing power node and the state backfilling position are marked in the inference path to generate the initial AI task probe. The initial AI task probe is sent into the initial test shadow inference chain, which drives the initial test shadow inference chain to call the intelligent inference model to perform shadow inference on the test sample in a state that is separated from the main business inference chain. The test state formed by the shadow inference is written back to the initial AI task probe to form the state AI task probe. The test state in the state AI task probe is compared with the inference path of the intelligent inference model to determine the abnormal position of the test state in the inference path. Based on the selected retest link at the abnormal location, the state AI task probe is matched to the retest shadow inference chain in the retest link, the abnormal location in the inference path is retested, and the test status in the state AI task probe after retesting is summarized to generate active test results.

2. The active testing method for AI computing power clusters based on edge computing as described in claim 1, characterized in that: The test sample is any one of the image content, text content, voice content, and sensor-collected content used by the intelligent reasoning model to perform shadow reasoning; The inference path is formed by sequentially connecting the path entry point before the test sample enters the intelligent inference model, the shadow inference position of the intelligent inference model, and the return landing point of the test state.

3. The active testing method for AI computing power clusters based on edge computing as described in claim 1, characterized in that: The generation of the initial AI task probe specifically includes locating the path entry point of the test sample before it enters the intelligent inference model in the inference path, and taking the edge computing power node associated with the path entry point as a candidate edge computing power node. Verify the calling relationship between candidate edge computing power nodes and intelligent inference models in the inference path, and mark the candidate edge computing power nodes with valid calling relationships as target edge computing power nodes; Find the return point of the test state formed after the intelligent reasoning model completes shadow reasoning in the reasoning path, and mark the return point as the state backfilling position. Write the target edge computing node into the path entry of the inference path, and write the state backfill position into the return landing point of the inference path to form a marked inference path; The test sample, target edge computing node, and state backfill location are encapsulated into the labeled inference path to generate the initial AI task probe.

4. The active testing method for AI computing power clusters based on edge computing as described in claim 1, characterized in that: The formation of the state AI task probe specifically includes: using the target edge computing node as the matching object, finding the initial test shadow inference chain connected to the target edge computing node, and sending the initial AI task probe into the initial test shadow inference chain. The initial test of the shadow inference chain divides the shadow transmission position of the test sample and the real business request position of the main business inference chain in the marked inference path, forming a shadow execution branch; The shadow execution branch calls the intelligent inference model in the target edge computing node to perform shadow inference on the test sample and obtain the shadow inference state; Organize the sample delivery state, model invocation state, and inference completion state in the shadow inference state into the test state of the initial AI task probe; The test state is written back to the state backfill position in the initial AI task probe to form the state AI task probe.

5. The active testing method for AI computing power clusters based on edge computing as described in claim 4, characterized in that: The initial test shadow inference chain is a bypass inference chain formed within the target edge computing node by connecting the intelligent inference model and the state backfilling position along the marked inference path and separating it from the main business inference chain. The main business inference chain is the business inference path in the target edge computing node that receives real business requests and calls the intelligent inference model to perform inference processing on the real business requests.

6. The active testing method for AI computing power clusters based on edge computing as described in claim 1, characterized in that: The construction process of the intelligent reasoning model specifically includes: organizing the sample content and sample task labels in the edge sample training set to form training samples; A lightweight inference network is constructed, which includes a sample entry layer, a shared feature compression layer, a task inference branch layer, an edge consistency verification layer, and an inference output layer. The training samples are fed into the sample entry layer, which converts them into sample features and then passes them to the shared feature compression layer. By using a shared feature compression layer, edge feature compression and feature fusion are performed on sample features to form shared inference features; The shared inference features are distributed to the task inference branch layer, and the shared inference features are inferred and mapped according to the sample task labels to obtain the branch inference results. The branch inference results and sample task labels are checked for consistency through the edge consistency verification layer. When the consistency verification is successful, the branch inference results are output through the inference output layer and the current lightweight inference network is determined as the intelligent inference model. When the consistency verification is unsuccessful, the branch inference results are sent back to the shared feature compression layer for feature reorganization. The reorganized shared inference features are then redistributed to the task inference branch layer for inference mapping. The network parameters of the lightweight inference network are adjusted according to the re-formed branch inference results, and the intelligent inference model is formed after the adjustment. Deploy the intelligent inference model to edge computing nodes and record the inference path of the intelligent inference model in the edge computing nodes.

7. The active testing method for AI computing power clusters based on edge computing as described in claim 6, characterized in that: The edge sample training set is a training set consisting of sample content and corresponding sample task labels collected by edge computing nodes and used to train intelligent inference models.

8. The active testing method for AI computing power clusters based on edge computing as described in claim 1, characterized in that: The connection verification involves matching the test states in the state AI task probe with their preceding and following positions according to the reasoning path of the intelligent reasoning model, and verifying the continuity between the test states and the reasoning path.

9. The active testing method for AI computing power clusters based on edge computing as described in claim 1, characterized in that: The generation of active test results specifically includes taking the abnormal position in the inference path as the center, extracting adjacent inference segments in the inference path, and forming a retest range; Find the bypass inference path in the target edge computing node, align the bypass inference path with the retest range, and select the retest link; Write the abnormal locations in the inference path and the test states in the state AI task probes into the retest link to form a retest probe path. Locate the retest shadow inference chain that connects to the abnormal location in the retest probe path, and match the state AI task probe to the retest shadow inference chain. The shadow inference chain is retested by calling the intelligent inference model to perform retests on abnormal locations in the inference path, thus forming a retest test state. Write the retest status into the status backfill position in the status AI task probe to form the retested status AI task probe. Summarize the test states in the AI ​​task probe after retesting and generate proactive test results.

10. The active testing method for AI computing power clusters based on edge computing as described in claim 9, characterized in that: The retest shadow inference chain is a bypass inference chain formed by connecting the bypass inference path, intelligent inference model and state backfill position in the retest link.