An artificial intelligence-based multifunctional edge computing device and method

Abstract

The application discloses a kind of multifunctional edge computing device and method based on artificial intelligence, specifically related to wind power operation management technical field, including the following steps: main model and auxiliary model are deployed in edge node, main model executes real-time inference, and auxiliary model is used for credibility assessment;Set disturbance detection mechanism, inject disturbance sample with preset features to main model, and compare it with auxiliary model output, calculate response offset degree;Fusion offset degree and inference difference generate credibility score;When score is lower than threshold value, automatically select to execute parameter reload or switch auxiliary model, and rebuild inference session, guarantee edge node stable operation;The application realizes real-time perception and structural deviation judgment to the running state of edge computing device by disturbance detection, main auxiliary model collaborative verification and credibility scoring mechanism, and based on strategy tree dynamic execution parameter reload or model switching, significantly improve the stability of model in extreme environment, self-diagnosis ability and fault tolerance level.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power operation management, more particularly, to a multi-functional edge computing device and method based on artificial intelligence. BACKGROUND

[0002] With the in-depth application of artificial intelligence technology in industrial scenarios, wind power generation systems are gradually introducing edge computing and AI models to realize local perception, predictive maintenance and intelligent control of wind turbine operation status. Especially in offshore or island-type wind farms away from the main grid, edge intelligent nodes play a key role in operation judgment and decision support.

[0003] Existing edge AI models are usually deployed in embedded hardware devices and run on resource-constrained processing platforms, and need to be online for a long time. However, in the high humidity, high salt fog, and unattended operation environment specific to wind power systems, such embedded devices are easily affected by hardware micro-errors induced by environmental factors. In particular, non-fatal bit flips may occur in volatile memory units such as RAM under humid conditions, which may not immediately cause device crashes, but can cause slight drift of AI model internal parameters, and further produce imperceptible but continuously accumulated accuracy deviations in inference output.

[0004] At present, most wind power edge AI systems lack mechanisms for real-time verification and structural monitoring of model inference status, and do not establish behavior evaluation logic based on disturbance response or copy comparison. This "silent running" assumption environment makes the model continue to output normally even if the inference output has decreased in accuracy or is structurally abnormal after a long time of running, which may eventually lead to false alarms, state misidentification, or incorrect triggering of critical control logic, and in severe cases, will affect the operation efficiency and safety of the wind farm. Therefore, the present application proposes a multi-functional edge computing device and method based on artificial intelligence to solve the above problems. SUMMARY

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] A multi-functional edge computing method based on artificial intelligence, comprising the following steps:

[0007] Deploy a main model and a secondary model in the edge node, the main model is used to perform real-time inference tasks, and the secondary model is configured as a static reference model and only participates in inference during the credibility evaluation process;

[0008] Set a disturbance detection mechanism, periodically inject disturbance samples with preset features into the main model, and compare the response results of the main model to the disturbance samples with the output results of the secondary model to calculate the disturbance response offset degree;

[0009] The fusion of the disturbance response offset degree and the inference difference of the main and auxiliary models to the same real-time input data establishes a credibility scoring mechanism to judge the stability and accuracy of the current model running state;

[0010] When the credibility score is lower than the threshold, the main model parameter reload or switching to the auxiliary model is automatically selected based on the offset characteristics and difference trends, and the inference session is rebuilt to maintain the continuous and stable operation of the edge intelligent edge node.

[0011] In a preferred embodiment, the disturbance samples include two types, namely boundary samples and stability samples, which are used to stimulate the sensitivity of the main model to input boundaries and the stability of the intermediate layer inference structure respectively. The boundary sample is used to test the prediction stability under extreme input conditions, and the stability sample is used to observe the response consistency to slight input disturbance. The disturbance response offset degree is calculated by comparing the cosine similarity, Euclidean distance or Mahalanobis distance between the intermediate layer feature vectors before and after disturbance, which is used to reflect the consistency of the main model internal structure in response to disturbance.

[0012] In a preferred embodiment, the credibility score for fusing the disturbance response offset degree and the output difference of the main and auxiliary models adopts a weighted scoring strategy. The weight of the disturbance term is dynamically adjusted according to the stability performance of the disturbance sample in the historical period. The output difference of the main and auxiliary models is quantified by the root mean square error. The scoring function is constructed by one of the following ways: linear weighted combination, exponential weighted average based on time series decay or adaptive weight strategy based on information entropy distribution.

[0013] In a preferred embodiment, when the main model triggers parameter reload, it adopts a difference positioning and local parameter recovery strategy. The potential drift area is located by the credibility score, corresponding to a specific sub-layer or parameter branch in the model structure. Then the corresponding weight is extracted from the pre-stored parameter backup to replace part of the current running parameter set, realizing local reload.

[0014] In a preferred embodiment, the threshold trend evaluation mechanism with a sliding window is used to judge when the credibility score is lower than the threshold. A period of observation is defined. If the score falls into the preset warning interval for multiple consecutive observation periods, it is judged that the main model stability has degraded and the recovery behavior is triggered.

[0015] In a preferred embodiment, the auxiliary model is designed as a lightweight version optimized by knowledge distillation, which has lower inference delay and resource overhead than the main model. The auxiliary model aligns the output distribution with the main model, imitates the intermediate features and compresses the structure in the training stage, realizes the preservation of discrimination ability and the reduction of running load, keeps the input and output space consistent, and seamlessly accesses in the verification process and switching scenario.

[0016] In a preferred embodiment, the inference session reconstruction process after model reloading or switching includes releasing the current cache state, reassigning the execution context, performing a null data inference operation once to activate the internal computation graph, and regenerating event markers with timestamp and trigger information.

[0017] In a preferred embodiment, the model switching and reloading behavior is determined by a policy tree structure containing multiple trigger factors, and the evaluation conditions include disturbance sample response trend, main and auxiliary model output difference rate, deviation between model prediction and observation data, and edge node load state. Different response behaviors are selected according to the policy tree path.

[0018] In a preferred embodiment, the credibility score result is synchronized to adjust the task scheduling strategy of the edge node.

[0019] In a preferred embodiment, an artificial intelligence-based multifunctional edge computing device comprises:

[0020] A model deployment module is used to deploy a main model and an auxiliary model in an edge node, the main model is used to perform real-time inference tasks, and the auxiliary model is configured as a static reference model and only participates in inference during the credibility evaluation process;

[0021] A disturbance detection module is used to set a disturbance detection mechanism, periodically injects a disturbance sample with a preset feature into the main model, and compares the response result of the main model to the disturbance sample with the output result of the auxiliary model to calculate the disturbance response offset degree;

[0022] A score calculation module is used to fuse the disturbance response offset degree and the inference difference of the main and auxiliary models to the same real-time input data, establish a credibility scoring mechanism, and judge the stability and accuracy of the current model running state;

[0023] A behavior switching module is used to automatically select the main model parameter reloading or switching to the auxiliary model based on the offset feature and difference trend when the credibility score is lower than the threshold, and to reconstruct the inference session to maintain the continuous and stable operation of the edge intelligent edge node.

[0024] The technical effects and advantages of the present application are:

[0025] The application makes the edge AI device capable of dynamically perceiving the structural deviation risk of the model in the inference process in the environment without external network and long-time unattended, especially in the high-humidity and easy-to-occur memory bit flip extreme working conditions such as offshore wind farms, the device can capture the disturbance response abnormality and inference deviation trend in time, and effectively identify the potential precision degradation behavior of the model. Through the parameter local reloading mechanism triggered by the credibility score, the system can automatically repair the drifted model sub-layer weight, prevent the model from being silent failure, and significantly improve the stability and robustness of the long-term operation of the model.

[0026] The main and auxiliary model cooperative structure and strategy tree behavior decision mechanism proposed in the application enable the edge AI device to have the dynamic closed-loop capability of "operation self-checking-health assessment-behavior self-adaptation". Through the credibility scoring system established by fusing the disturbance response deviation and the prediction difference of the main and auxiliary models, the device can not only evaluate the current state of the model in real time, but also select the most appropriate response behavior, including delayed reloading, auxiliary model switching or continuous observation, in combination with the model state trend and system operation condition. The fault-tolerant strategy avoids the resource waste and mis-stop risk caused by the "one-size-fits-all" restart mechanism of the traditional system, and provides a more flexible and intelligent model health management scheme for key task scenarios such as wind farms. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to facilitate the understanding of those skilled in the art, the application will be further described below with reference to the accompanying drawings;

[0028] Figure 1 A multi-functional edge computing method based on artificial intelligence in the application.

[0029] Figure 2 A multi-functional edge computing device based on artificial intelligence in the application. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the application.

[0031] REFERENCE Figure 1 - Figure 2 The following embodiments are obtained:

[0032] Embodiment 1: The present application mainly aims at providing a solution to the problem of model reliability maintenance of edge computing system in extreme industrial operating environment, especially suitable for the offshore wind power island scenario deployed away from the main network environment. In such application background, high-power power equipment such as wind turbine needs to run continuously under harsh conditions such as high humidity, salt spray, wind pressure change, unstable power supply, etc. The edge AI inference node must bear the functions of local perception, prediction judgment, control scheduling of equipment state, and require long-time unattended, autonomous recovery, autonomous decision-making ability.

[0033] However, such environment has a series of problems that existing technologies cannot avoid, including implicit accuracy drift in model running process (such as parameter deviation caused by memory bit-flip), abnormal disturbance response that cannot be identified, lack of local reloadable mechanism, etc. More seriously, due to the inability to access the remote cloud model management system, the traditional model state evaluation and version update mechanism relying on cloud monitoring is difficult to implement in this environment. For example, the problem description: the high humidity in the island environment erodes the edge node equipment all year round, and some commercial embedded devices appear non-fatal RAM bit errors (such as flip) under humid conditions. Although the device does not crash, the AI model inference result appears a slight deviation. Most existing technologies do not set up a model inference self-correction mechanism, resulting in a "silent" judgment deviation of the model after a long time of running. The model output is no longer accurate, the device state prediction error accumulates, and eventually causes wind turbine miscontrol or protection mis-triggering.

[0034] To adapt to the requirements of long-term stable operation of edge AI model in extreme operating environment such as wind farm, the present application provides an edge computing device and method integrating multiple functions, which, while performing the regular inference task of the main model, integrates the following mechanisms: a secondary model cooperative verification mechanism, a disturbance response evaluation mechanism, a model credibility judgment mechanism based on scoring, a local parameter repair mechanism, and a strategy tree driven model behavior scheduling mechanism, and has the ability of task-level resource self-adaptive adjustment, forming an intelligent edge computing framework that can self-diagnose, self-recover, and self-select processing path when the model state is abnormal.

[0035] Specifically applied to AI model running management in edge devices, including the following steps:

[0036] Step one, deploy the main model and the secondary model in the edge node respectively, the main model is used to execute real-time inference task, and the secondary model is configured as a static reference model and only participates in inference in the credibility evaluation process;

[0037] Step two, set up a disturbance detection mechanism, periodically inject disturbance samples with preset features into the main model, and compare the response results of the main model to the disturbance samples with the output results of the secondary model, and calculate the disturbance response offset degree;

[0038] Step three, fuse the disturbance response offset degree with the inference difference of the main and auxiliary models to the same real-time input data, establish a credibility scoring mechanism to judge the stability and accuracy of the current model running state;

[0039] Step four, when the credibility score is lower than the threshold, automatically select the main model parameter reload or switch to the auxiliary model based on the offset characteristics and difference trend, and rebuild the inference session to maintain the continuous and stable operation of the edge intelligent edge node.

[0040] The disturbance sample includes at least two types, which are used to stimulate the sensitivity of the main model to the input boundary and the stability of the intermediate layer inference structure respectively. The boundary sample is used to test the prediction stability under extreme input conditions, and the stability sample is used to observe the response consistency to the micro input disturbance. The disturbance response offset degree is calculated by comparing the cosine similarity, Euclidean distance or Mahalanobis distance between the intermediate layer feature vectors before and after the disturbance, which is used to reflect the response consistency of the internal structure of the main model to the disturbance. The intermediate layer output is extracted by the inference stability reference layer defined in advance, which is the basis for calculating the response reliability.

[0041] The boundary sample is used to stimulate the sensitivity of the main model to the input boundary, specifically, the input data is pushed to the edge of the model training interval or beyond the normal distribution range, to simulate extreme operating conditions or irregular input signals, such as wind speed reaching an extremely high value or voltage being in a critical state. The response of the main model under such samples is used to test the stability of its prediction ability under extreme input conditions, that is, to judge whether its output remains continuous, reasonable, or whether it appears a collapsing prediction (such as output divergence, abnormal category, etc.).

[0042] The stability sample is used to observe whether the model has consistent inference results under micro disturbance input near the normal distribution. Such disturbance can be constructed by adding low-amplitude high-frequency noise, fine signal jitter or light data rotation, etc., the purpose is to simulate the influence of sensor errors or slight interference on the model, so as to detect the micro response change of the model under high-frequency operation.

[0043] In order to quantify the response difference of the main model to different types of disturbances, the invention calculates the disturbance response offset degree, that is, by comparing the similarity between the intermediate layer feature vectors before and after the disturbance to evaluate the response consistency of the model at the structure level. The intermediate layer feature vector refers to the tensor representation output by the model at a certain or several specific layers during the inference process, which usually contains high-dimensional semantic information and multi-channel activation state.

[0044] The calculation method of the offset degree includes any one of the following three methods:

[0045] Cosine similarity: measures the consistency of two vector directions, the value range is usually between [-1, 1], the closer to 1 indicates the more consistent the response direction;

[0046] Euclidean distance: reflects the geometric distance between two vectors, the smaller the distance, the closer the output before and after the perturbation;

[0047] Mahalanobis distance: introduces a covariance matrix to consider the correlation and scale influence between different features, suitable for evaluating the abnormality of samples in multi-dimensional space.

[0048] In the present application, these measurement methods are used to quantify the intermediate layer response of the main model, which is used to construct the perturbation response deviation index. This index can directly reflect the consistency of the main model's internal structure in response to perturbation, and is used as a key input for the subsequent credibility scoring mechanism. In addition, the feature vectors used for output comparison are extracted from a set of pre-defined reasoning stable reference layers. These reference layers are usually selected from the intermediate part of the model, such as intermediate convolution blocks in convolutional neural networks, intermediate attention layers in Transformer architecture, etc. By continuously monitoring the output of this layer and sampling the perturbation response, the changing characteristics of the model's structural behavior under different input perturbations can be accurately reflected.

[0049] The credibility score for fusing the perturbation response deviation and the output difference between the main and auxiliary models uses a weighted scoring strategy. The weight of the perturbation term is dynamically adjusted according to the stability performance of the perturbation sample in the historical period, and the output difference between the main and auxiliary models is quantified by the root mean square error. The scoring function can be constructed in one of the following ways: linear weighted combination; exponential weighted average based on time series decay; or adaptive weight strategy based on information entropy distribution. This fusion mechanism takes into account both instantaneous deviation and long-term stability trend, which helps to improve the judgment accuracy of the score and the rationality of the behavior trigger.

[0050] Periodically obtain two independent source measurement indicators:

[0051] Perturbation response deviation: reflects the structural response consistency of the main model under the input of the perturbation sample;

[0052] Output difference between main and auxiliary models: reflects the consistency of the prediction results of the two models under the same real-time input.

[0053] To achieve the effective integration of the two-dimensional indicators, the edge node adopts a weighted scoring strategy, in which each indicator is assigned a dynamically changing weight value, and in particular, the weight of the disturbance item is dynamically adjusted according to the stability performance of the disturbance sample in the historical period. For example, if a certain type of disturbance sample has shown high consistency in the past several periods (i.e., the model has shown stability), once the current offset value fluctuates, it will be considered as a stronger abnormal signal, and therefore the weight corresponding to the sample will be increased; otherwise, the influence of the weight on the score will be reduced. This mechanism can be regarded as a dynamic trust reconstruction based on the historical distribution of disturbance behavior.

[0054] The difference between the main model output and the auxiliary model output is quantified by the root mean square error (RMSE), which reflects the average deviation of the two models in numerical prediction. RMSE is more sensitive to large errors and can effectively amplify the inconsistency caused by prediction imbalance, making it more representative in model stability evaluation.

[0055] The credibility score function is used to integrate the above two indicators into a unified score result, which can be constructed in one of the following three ways:

[0056] Linear weighted combination: set a fixed or dynamic proportion coefficient, and linearly superimpose the two indicators, such as: Score = a · Δ1 + β · Δ2;

[0057] Exponential weighted average (EWMA) based on time series decay: introduce a time decay factor in consecutive periods, so that recent indicators have a greater impact on the score, which is suitable for responding to rapid shifts; The core idea of this strategy is to construct a weighted time series sliding value for each type of scoring indicator before merging the disturbance response offset or the difference between the main model output and the auxiliary model output, so that the latest period's data has a higher impact weight, while the contribution of historical data decreases exponentially over time, thus forming a scoring basis that is more sensitive to current state changes. Assign a time weight to each consecutive period, which decreases over time, and newer indicator data will participate in the score update with a larger weight. This processing method has stronger immediate response capability to sudden changes in model state (such as unstable reasoning, dramatic changes in disturbance response, etc.). Through this mechanism, even if the score result does not exceed the edge node threshold, if the recent fluctuation amplitude suddenly increases, it can also trigger the attention of the edge node or trigger the evaluation logic in advance. In addition, EWMA does not need to store all historical data, only the last sliding value needs to be retained, and the calculation is simple and resource occupation is low, which is particularly suitable for deployment on edge devices with limited computing power. This strategy is suitable for capturing abnormal trends in credibility in short periods, and enhances the response capability to "early signals" without causing excessive false positives.

[0058] An adaptive weight strategy based on information entropy distribution: information entropy analysis is performed on the degree of change of the input or error distribution, and the importance of each indicator in the scoring is dynamically adjusted to highlight the prediction stability or structural robustness dimension. The basic idea of this strategy is to use information entropy to measure the change distribution characteristics of the scoring indicators, so as to determine whether the indicator has enough "information" or "recognition ability" in the current period. When a certain type of indicator remains highly concentrated and narrow in fluctuation range in multiple cycles, its information entropy is low, indicating that the contribution of the indicator to state discrimination is limited, and the edge node can appropriately reduce its weight in scoring; on the contrary, if the indicator shows high uncertainty and change dispersion, it means that it may be in a sensitive stage to the edge node state change, and should be given a higher weight. This strategy periodically calculates the entropy value of each indicator's historical value to dynamically identify the "highly distinguishable" feature dimension in the current scoring signal, ensuring that the scoring mechanism focuses on "high information density and high change sensitivity" indicators, thereby avoiding interference from redundant indicators and improving the overall judgment efficiency and reliability of the scoring. This adaptive strategy is particularly suitable for handling scenarios with diverse indicator dimensions and complex state evolution, and has good scalability and edge node self-learning ability. By introducing an entropy value dynamic feedback mechanism, the edge node can achieve structural dynamic optimization of the scoring weight, and strengthen the driving ability of key abnormal signals on the model behavior.

[0059] The above scoring strategy design has strong configurability and adaptability, allowing flexible selection according to device operating environment, task type, and prediction period. The entire fusion mechanism takes into account both instantaneous deviation and long-term stability trend, enabling not only timely detection of short-term abnormalities in inference results, but also capturing the impact of long-term evolution of model internal state. The final reliability score is used to drive the behavior response of the edge node, including but not limited to model reloading, structure replacement, or inference priority adjustment, thereby improving the accuracy of the scoring judgment and the rationality of the behavior trigger, enhancing the self-regulation ability and reliable operation level of the edge intelligent edge node.

[0060] Assigning a dynamically changing weight value means adjusting its influence in the final score under different time periods, operating states, or sample behaviors. The dynamic change of the weight value can be adjusted based on the following factors:

[0061] The disturbance item weight change driven by the historical response stability of the disturbance, for the disturbance response deviation degree, its corresponding weight value not only reflects the deviation degree of the current disturbance result, but also reflects the response stability of the disturbance sample type in the past continuous running period. Specifically: when a certain disturbance type keeps its deviation degree in a stable interval (such as in a low fluctuation range) in the last N periods, it is considered to be a low-noise signal source when the model state is stable, and once the disturbance type produces a large deviation in the current period, its weight should be increased by the edge node; on the contrary, if the disturbance type shows high-frequency fluctuation or unpredictable behavior (such as frequent and severe fluctuation) in the historical period, even if its deviation degree jumps, the edge node considers that it indicates low reliability, and the contribution to the score result should be reduced, and the weight is adjusted accordingly. The change can be realized by setting the historical stability score of the disturbance sample and mapping it to the current weight adjustment factor of the disturbance item.

[0062] The error weight adjustment guided by the consistency of the current model output, the output difference of the main and auxiliary models as the second type of score index, mainly through error calculation by comparing the prediction results of the same input. The weight of this error can be adjusted based on the following two dynamic factors: if the current edge node is running in a critical control period (such as the edge device is in an extreme working condition, high-speed running state), the consistency of the output of the main and auxiliary models becomes more important, and the weight of the error term will be adjusted up; if the edge node is in a non-critical period or a fault tolerance window period (such as maintenance state, buffer period), the error weight is appropriately reduced to avoid overreaction to short-term fluctuations. This adjustment process can be combined with the current task load, running state label or control priority parameter of the edge device, and rely on existing technology for adaptive automatic adjustment.

[0063] When the main model triggers parameter reloading, the difference positioning and local parameter recovery strategy is adopted. The edge node locates the potential drift area through the credibility score, corresponding to a specific sub-layer or parameter branch in the model structure, and then extracts the corresponding weight from the pre-stored parameter backup to replace part of the current running parameter set, realizing local reloading. This process uses parameter hash mapping relationship for quick positioning and verification, thereby avoiding the running interruption caused by full model reloading, and restoring the inference consistency on the premise of minimizing the computing resources.

[0064] Firstly, the running state of the main model is evaluated through the credibility score mechanism. When the score result is lower than the set threshold, the edge node further analyzes the score details to identify the part of the model structure that shows signs of drift. These areas usually correspond to sub-layers or parameter branches in the model structure, such as the weight matrix of a certain convolution layer, normalization layer, attention submodule, fully connected layer, etc. After identification, the edge node will replace the parameters for this specific sub-area. This operation is completed by extracting the corresponding weight from the pre-stored parameter backup.

[0065] The meaning of "extracting corresponding weights from pre-stored parameter backup" is that after the initial deployment or training of the model is completed, the edge node has saved the original parameter structure of the model completely in the non-volatile storage unit (such as local Flash, eMMC or solid state disk), and the backup file not only contains the model structure information, but also stores the parameters in a sub-regional or hierarchical manner, and each parameter subset has an independent index and can be accessed and extracted independently. The backup represents the "health status" of the model in the training stage or the early stage of running, and is not affected by potential disturbances or storage abnormalities during running.

[0066] To realize fast access and verification, the edge node establishes a parameter hash mapping relationship for these parameter blocks, that is, a unique verification identifier (such as CRC verification code, SHA hash value) is calculated for each parameter partition, and a mapping index corresponding to the model structure is established. In actual operation, if a certain area is marked as "drift risk", the edge node will use the mapping to locate the corresponding weight block in the backup file, verify the integrity, and load it into the current model running memory to replace the part of the parameter set being used.

[0067] This operation is a partial reload, and its core advantage is to avoid the time delay, memory jitter and inference interruption risk brought by full model reload. After the reload is completed, the edge node will automatically rebuild the inference path, clean up the old parameter cache, and continue to complete the current inference task. Through the above difference positioning and local parameter recovery strategy, the application realizes a low-resource consumption intelligent recovery mechanism for edge devices while ensuring the overall stability of the model, significantly improving the self-maintenance ability and running continuity of the model in complex field environments.

[0068] Among them, the potential drift area refers to a local sub-area in the main model structure that is suspected to have prediction behavior abnormalities, calculation stability decline, or feature expression ability weakening risk. This area may correspond to a specific network layer (such as convolution layer, fully connected layer, normalization layer), a calculation subunit (such as attention head, activation mapping group), or a parameter substructure (such as bias vector, scale parameter, etc.). Whether the area has drift depends on the joint analysis of the edge node on the intermediate layer feature response, disturbance response behavior, and model output stability in multiple dimensions. The specific judgment process includes the following aspects:

[0069] Intermediate layer response anomaly detection: In each round of inference task, the edge node inputs part of the perturbed samples into the main model, and extracts the feature output of the intermediate layer. If the feature distribution output by the layer is obviously deviated from the historical distribution (such as mean, variance, activation map distribution), or the cosine similarity is significantly decreased, the edge node will mark the structure of the intermediate layer as a potential drift area. Obvious deviation and significant change can be compared by setting a division threshold or a normal change interval. For example, exceeding the normal change interval can be identified as obvious deviation. Similar comparison methods can be used in other parts of the application, which will not be repeated here. For example, the definition of "rapid change" in the inconsistency index of perturbation response.

[0070] Inconsistency index of perturbation response: When the model faces the same type of input sample, the response result of the perturbation changes rapidly in the short term. The edge node will locate the feature extraction path or attention distribution path involved in these abnormal perturbation responses, and further map them to the structure layer as drift candidate areas.

[0071] Structural output difference analysis of main and auxiliary models: When the main model and the auxiliary model produce high difference output under the same input condition, the edge node will trace back the calculation path of the difference, such as the output vector difference mainly concentrated in some output dimensions, combined with the model structure map, to infer the key parameter path causing the error, and then map it to a specific sub-layer.

[0072] Comprehensive sorting of credibility score and impact sensitivity: All the candidate areas obtained through the above analysis will be unified into the credibility score impact sensitivity evaluation process, combined with the impact factor sorting (such as current score decline trend, parameter position sensitivity, fault history record, etc.), and finally select the top several areas as the "potential drift area" that needs to perform local parameter recovery in this round. The sorting method can choose any method in the prior art to measure the overall impact degree, and once the potential drift area is confirmed, the parameter hash mapping table can be used to retrieve the parameter weight block corresponding to the area in the pre-stored parameter backup, and perform replacement operation to form "local reloading".

[0073] The judgment of credibility score lower than the threshold value does not depend on a single instantaneous value, but uses a threshold trend evaluation mechanism with a sliding window. The edge node defines an observation period, and if the score falls into the warning interval for multiple periods, it is judged that the stability of the main model has degraded and the recovery behavior is triggered. This strategy introduces joint judgment of abnormal count accumulation, time weighted score and interval confidence, effectively avoiding false triggering caused by short-term interference, and enhancing the tolerance of the edge node to fluctuating signals and the decision-making robustness.

[0074] The secondary model is designed as a lightweight version optimized through knowledge distillation, with lower inference delay and resource overhead compared to the primary model. During the training phase, the secondary model aligns its output distribution with the primary model, mimics intermediate features, and compresses its structure to maintain discriminative ability and reduce operational load. The input and output spaces remain consistent, allowing seamless access during validation and switching scenarios. This design allows the secondary model to operate stably under frequent calls while reducing energy consumption accumulation on edge nodes under high load.

[0075] Knowledge distillation is a model compression and transfer learning technique that transfers the knowledge learned by a large model (teacher model, primary model) to a smaller model (student model, secondary model). This process is usually completed during the training phase and includes the following three key strategies:

[0076] Output distribution alignment: By minimizing the difference between the output probability distributions of the primary model and the secondary model under the same input (such as the Kul l back-Le i b l er divergence), the secondary model mimics the judgment logic of the primary model as much as possible in the final prediction result.

[0077] Intermediate feature imitation: During the training process, the partial activation state of the intermediate layer (such as the convolution feature map or attention weight) of the primary model is used as a guide to force the secondary model to be structurally closer to the expression method of the primary model, thereby enhancing its representation ability.

[0078] Structural compression: At the model design level, model slimming is achieved by pruning channel numbers, reducing parameter dimensions, replacing complex operations with efficient operators (such as replacing standard convolution with depthwise separable convolution), etc., to ensure fast response under tight operational resource conditions.

[0079] Through the above training strategies, the secondary model has a high fitting to the discriminative logic of the primary model, while significantly reducing the parameter size and computational complexity. To ensure that the secondary model can seamlessly access the validation process and model switching scenarios, its input and output spaces remain consistent, i.e., the input data dimension, preprocessing process, output vector structure, and primary model are completely aligned. This consistency allows smooth task continuation when the edge node needs to call the secondary model to verify the primary model results or perform hot switching when the credibility fails, without additional data transformation, preprocessing reconstruction, or inference interface conversion.

[0080] Since the sub-model itself is optimized in structure, it shows lower reasoning delay and resource consumption in actual operation, and can effectively deal with frequent verification calls and high-frequency state evaluation tasks. At the same time, the structure greatly reduces the energy consumption accumulation of the edge node under high load state, reduces the chip temperature rise, power supply pressure and calculation bottleneck risk, and improves the energy efficiency ratio and reliability of the edge node under long-time operation.

[0081] The distillation design of the sub-model not only ensures that the edge node has structural complementarity and functional redundancy at the model level, but also realizes the performance and energy consumption balance in the edge computing environment through optimized structure and behavior path, which embodies the edge node innovation advantage of the present application in model design, edge node scheduling and task switching.

[0082] The reasoning session reconstruction process after model reloading or switching includes releasing the current cache state, reassigning the execution context, performing a null data reasoning operation to activate the internal computation graph, and generating event markers with timestamps and trigger information for the upper edge node to identify. The event markers contain switching trigger reasons, execution modes and model identities, which are used for log tracking and subsequent analysis. This process ensures that the response time after reasoning switching is controllable, and the edge node state is clear, which helps the continuous and stable operation of the whole.

[0083] After the model reload or switch is triggered, the edge node will immediately perform the operation of releasing the current cache state. This step includes emptying the intermediate tensor data in the running memory, gradient cache (if exists), preprocessing cache, reasoning graph execution context, etc., to avoid historical state residues interfering with new model reasoning. This operation helps to release resources, eliminate state pollution, and prepare for subsequent context reconstruction. Then the execution context is reassigned, that is, the new model (main model reload or sub-model switch) to be loaded soon is allocated a new running session environment, including model graph structure loading, weight parameter mapping, reasoning path construction and necessary hardware binding process (such as tensor RT binding, AI acceleration unit interface initialization, etc.), to ensure that the model can run correctly under the current edge node resources.

[0084] In order to activate the inference execution path of the new model and trigger the initial compilation or preheating of its internal structure (such as dynamic computation graph, control flow, cache mechanism, etc.), the edge node will perform an empty data inference operation once. So-called "empty data" refers to input samples with legal structure but invalid values (such as all-zero tensors, pseudo inputs, etc.), which are not used to obtain results, but to initialize the kernel structure of the model, cache construction and operator optimization path, to ensure that the response time of the subsequent formal inference process is controllable and the startup delay is minimized. After completing model loading and inference path preheating, the edge node generates a set of event markers with timestamp and trigger information and sends them to the upper layer control edge node or log recording module to identify this model switching event. The event marker is a structured data entity that includes at least the following three types of information:

[0085] Switch trigger reason: such as "disturbance score anomaly", "output drift", "model freezing" and the like;

[0086] Execution mode: indicates whether this is "reload original model" or "switch secondary model";

[0087] Model identity information: indicates the version number, structure summary, and verification hash of the current active model.

[0088] The event marker is used by the edge node for log tracking and subsequent analysis to achieve model lifecycle management, stability statistics, and behavior review. At the same time, if the edge node is designed with remote monitoring or OTA upgrade mechanism, the event can also be used as an input basis for external control. This process ensures that the response time after inference switching is controllable, and the edge node state is clear, which helps to maintain the stability of the overall operation. Through this design, the present invention realizes the dynamic, transparent and fine-grained management of models in the edge computing environment, improves the controllability of the model evolution process and the observability of the running behavior of the intelligent edge node, and enhances the robustness of the intelligent edge node in long-term deployment.

[0089] The credibility score result is synchronized to adjust the task scheduling strategy of the edge computing device. In the case of low model score, the edge node reduces the execution frequency of non-critical tasks, releases resources to prioritize the main inference process, and increases the core sensor sampling rate to improve environmental perception accuracy. The original task weight is gradually restored when the score returns to normal. By dynamically associating model credibility with edge node resource allocation, the present invention realizes the linkage and self-adaptation between model state and running strategy, and improves the stability and energy efficiency of the edge node.

[0090] The application introduces a task-level resource linkage mechanism based on model state judgment, so that the credibility score result can affect the task scheduling behavior of the edge device. When the device detects that the main model credibility is reduced, the execution frequency of non-critical tasks can be automatically adjusted, more resources are released for guaranteeing the inference performance of the main model, and the data sampling frequency of the critical sensor is increased to enhance the model input quality. This mechanism realizes the real-time linkage between the AI model running state and the system resource management strategy, thereby maximizing the model running efficiency under the condition of limited edge resources, prolonging the service life of the device, and improving the overall operation efficiency.

[0091] In the case of low model score, for example, lower than a preset state threshold, the edge node identifies that there is uncertainty, possible drift or performance decline risk in the current main model inference state, and the edge node will execute a series of resource allocation strategies to guarantee the stable operation of the core inference process: first, identify all the computing tasks currently running, and divide the tasks into "critical tasks" and "non-critical tasks" according to the preset task priority table. Critical tasks include main model inference, fault detection, device control, etc., while non-critical tasks include local log archiving, remote heartbeat uploading, predictive analysis, visualization updating, etc.

[0092] In the case of low score, the edge node reduces the execution frequency of non-critical tasks, such as converting the originally scheduled data visualization running once every minute to once every 10 minutes, or suspending the non-core task thread, thereby releasing CPU, memory, IO and other edge node resources, and prioritizing the main model inference process to improve the robustness and response speed of the edge node. At the same time, the edge node will increase the sampling rate of the core sensor, such as vibration, temperature, voltage, wind speed, etc., which are highly related to the model input, and will increase the data collection from the original normal frequency to high frequency sampling mode (such as from 10Hz to 50Hz). This helps to enhance the input data quality by increasing the perception accuracy when the model credibility is reduced, and externally alleviates the prediction instability of the model.

[0093] Once the model score returns to normal, the edge node will gradually restore the original task weight to "mild regression" to redistribute the resources that were previously limited, restore the execution frequency and task concurrency of non-critical tasks, and avoid running jitter or response mutation caused by instantaneous resource fluctuations. This mechanism forms a task scheduling strategy chain driven by credibility score, realizing an automatic response closed loop of "decreased prediction ability → edge node resource deviation → enhanced perception intensity". By dynamically associating model credibility and edge node resource allocation, the application realizes the linkage and adaptation between model state and running strategy, that is, the uncertainty change of the model itself can trigger the fine tuning of the overall running mode of the edge node.

[0094] The model switching and reloading behaviors are determined by a policy tree structure containing multiple trigger factors. The evaluation conditions include disturbance sample response trend, main and auxiliary model output difference rate, deviation between model prediction and observation data, and edge node load state, etc. The edge node performs different response behaviors according to the policy tree path selection, such as delay reloading, direct switching or observation extension. The policy tree is composed of preset rules and supports online parameter fine-tuning for adapting to different deployment scenarios of model health management and task fault tolerance requirements.

[0095] Each node in the policy tree structure represents an evaluation condition judgment point, each branch represents a state range and judgment result of a type of trigger factor, and the leaf node corresponds to a different response behavior policy. The edge node performs condition judgment according to the tree structure based on the current model state input during running, and finally falls into a specific behavior node.

[0096] The evaluation conditions in the policy tree include but are not limited to the following factors, each of which can independently constitute a judgment node or be used as intermediate decision logic for multi-condition combination:

[0097] Disturbance sample response trend: The edge node records the response deviation curve of the disturbance sample in consecutive periods, analyzes its trend (e.g., whether it is continuously increasing or whether there is a periodic anomaly), and judges whether the model's robustness to structural perturbation is declining. If the disturbance response deviation is observed to show an accelerating upward trend, the factor is marked as a high-risk signal, and the policy tree branches down to the "response deviation is severe" path.

[0098] Main and auxiliary model output difference rate: By comparing the output results of the main model and the auxiliary model under the same input condition in consecutive periods, the edge node can calculate the error growth rate. If the error rapidly increases in unit time, it means that the main model behavior drifts rapidly, and there may be structural abnormalities. This rate is used to evaluate the degradation rate of the main model's reasoning ability and is an important judgment index in the policy tree.

[0099] Deviation between model prediction and observation data: The edge node can compare the predicted output of the main model with the actual sensor observation value to measure the deviation. For example, the difference between the predicted value of the fan vibration and the true measured value will be considered as the model no longer having real adaptability when it exceeds a certain proportion. The deviation degree judgment node can be further superimposed with the historical deviation trend to judge whether it is a persistent deviation.

[0100] Edge node load status: including CPU occupancy, memory usage, I / O pressure, etc. of the edge device. If the current edge node running load is high, the strategy tree can delay some non-critical operations according to the rules, such as delayed execution of overload behavior; on the contrary, in low load, priority is given to model self-recovery operation. Load status as an auxiliary condition can be used with model state judgment logic to realize self-regulation of edge node resources.

[0101] Behavior selection and response path: the edge node selects different response behaviors according to the strategy tree path, mainly including but not limited to the following response modes:

[0102] Delay overload: in the case of detecting that the model state has abnormal risk but the current load of the edge node is high or the task is critical, the parameter reload is not immediately performed, but is performed in the next scheduling period or idle period;

[0103] Direct switching: when the model state is abnormal, the response deviation is severe, and the output difference rate is high, the edge node directly switches the inference task from the main model to the secondary model to ensure real-time performance;

[0104] Observation extension: if the model state abnormal signal is not significant, the edge node enters the observation window and extends the time for continuous observation of key parameter trends, and triggers the next action after the abnormal trend is established.

[0105] Embodiment 2: A multifunctional edge computing device based on artificial intelligence, comprising:

[0106] A model deployment module for deploying a main model and a secondary model in an edge node, the main model for executing real-time inference tasks, and the secondary model configured as a static reference model and only participating in inference in a credibility evaluation process;

[0107] A disturbance detection module for setting a disturbance detection mechanism, periodically injecting a disturbance sample with a preset feature into the main model, and comparing the response result of the main model to the disturbance sample with the output result of the secondary model to calculate a disturbance response deviation;

[0108] A score calculation module for fusing the disturbance response deviation and the inference difference of the main and secondary models to the same real-time input data, establishing a credibility scoring mechanism, and judging the stability and accuracy of the current model running state;

[0109] A behavior switching module for automatically selecting to perform main model parameter reload or switching to the secondary model based on the deviation feature and difference trend when the credibility score is lower than a threshold, and rebuilding the inference session to maintain the continuous and stable operation of the edge intelligent edge node.

[0110] The above formulas are all de-dimensioned to calculate the numerical values, the formulas are obtained by collecting a large amount of data to simulate the most recent real situation, and the preset parameters in the formulas are set by a person skilled in the art according to the actual situation.

[0111] It should be understood that, in various embodiments of the present application, the size of the sequence number of each process does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0112] Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. A person skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0113] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the edge node, the device and the unit described above can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

[0114] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An artificial intelligence-based multi-functional edge computing method, characterized by, The method comprises the following steps: The main model and the auxiliary model are respectively deployed in the edge node, the main model is used to perform real-time inference tasks, and the auxiliary model is configured as a static reference model and only participates in inference in the credibility evaluation process; A perturbation detection mechanism is set, perturbation samples with preset features are periodically injected into the main model, and the response results of the main model to the perturbation samples are compared with the output results of the auxiliary model to calculate a perturbation response offset degree; The perturbation response offset degree and the inference difference of the main model and the auxiliary model to the same real-time input data are fused to establish a credibility scoring mechanism to judge the stability and accuracy of the current model running state; When the credibility score is lower than a threshold value, the main model parameter reloading or switching to the auxiliary model is automatically selected based on the offset features and difference trends, and the inference session is reconstructed to maintain the continuous and stable operation of the edge intelligent edge node.

2. The multi-functional edge computing method based on artificial intelligence according to claim 1, characterized in that, The perturbation samples include two types, i.e. boundary samples and stability samples, which are respectively used to stimulate the sensitivity of the main model to input boundaries and the stability of the intermediate layer inference structure, the boundary samples are used to test the prediction stability under extreme input conditions, and the stability samples are used to observe the response consistency to slight input perturbations, the perturbation response offset degree is calculated by comparing the cosine similarity, Euclidean distance or Mahalanobis distance between the intermediate layer feature vectors before and after the perturbation, and is used to reflect the consistency of the internal structure of the main model in response to the perturbation. 3.The multi-functional edge computing method based on artificial intelligence according to claim 2, wherein, The credibility score for fusing the perturbation response offset degree and the output difference of the main model and the auxiliary model adopts a weighted scoring strategy, the weight of the perturbation term is dynamically adjusted according to the stability performance of the perturbation sample in the historical period, the output difference of the main model and the auxiliary model is quantified by the root mean square error, and the scoring function is constructed by one of the following ways: linear weighted combination, exponential weighted average based on time series attenuation or adaptive weight strategy based on information entropy distribution.

4. The multi-functional edge computing method based on artificial intelligence according to claim 3, characterized in that, When the main model triggers parameter reloading, the difference positioning and local parameter recovery strategy is adopted, the potential drift area is positioned by the credibility score, the corresponding weight is extracted from the pre-stored parameter backup to replace part of the parameter set in the current running, and local reloading is realized.

5. The multi-functional edge computing method based on artificial intelligence according to claim 4, characterized in that, The judgment that the credibility score is lower than the threshold value adopts a threshold trend evaluation mechanism with a sliding window, a period of observation is defined, and if the score falls into a preset warning interval for multiple times in continuous multiple observation periods, it is judged that the stability of the main model is degraded and the recovery behavior is triggered.

6. The multi-functional edge computing method based on artificial intelligence according to claim 5, characterized in that, The auxiliary model is designed as a light version optimized by knowledge distillation, and has lower inference delay and resource consumption than the main model. The auxiliary model aligns the output distribution with the main model, imitates the intermediate features and compresses the structure in the training stage to realize the maintenance of discrimination ability and the reduction of running load. The input and output spaces remain consistent, so that it can be seamlessly connected in the verification process and switching scene.

7. The multi-functional edge computing method based on artificial intelligence according to claim 6, characterized in that, The inference session reconstruction process after model reloading or switching includes releasing the current cache state, reallocating the execution context, performing an empty data inference operation to activate the internal computation graph, and generating an event marker with a timestamp and trigger information. 8.The multi-functional edge computing method based on artificial intelligence of claim 7, wherein, The model switching and reloading behavior is determined by a policy tree structure containing multiple trigger factors, and the evaluation conditions include the disturbance sample response trend, the output difference rate of the main and auxiliary models, the deviation between the model prediction and the observation data, and the edge node load state. Different response behaviors are selected according to the policy tree path. 9.The multi-functional edge computing method based on artificial intelligence of claim 8, wherein, The credibility score result is used to adjust the task scheduling strategy of the edge node.

10. An artificial intelligence based multi-functional edge computing device for implementing an artificial intelligence based multi-functional edge computing method according to any one of claims 1-9, characterized in that, The method comprises the following steps: a model deployment module is used to deploy a main model and an auxiliary model in an edge node, the main model is used to execute real-time inference tasks, and the auxiliary model is configured as a static reference model and only participates in inference during the credibility evaluation process; a disturbance detection module is used to set a disturbance detection mechanism, periodically inject disturbance samples with preset features into the main model, and compare the response results of the main model to the disturbance samples with the output results of the auxiliary model to calculate the disturbance response offset degree; a score calculation module is used to fuse the disturbance response offset degree and the inference difference of the main and auxiliary models to the same real-time input data, establish a credibility scoring mechanism, and judge the stability and accuracy of the current model running state; an action switching module is used to automatically select the main model parameter reloading or switch to the auxiliary model when the credibility score is lower than the threshold value, and rebuild the inference session to maintain the continuous and stable operation of the edge intelligent edge node.

Images