Method and device for monitoring quality of service of edge composite service and electronic equipment

Abstract

This invention provides a method, apparatus, and electronic device for monitoring the service quality of edge composite services. For each atomic service in the edge composite service, its service quality time-series data is collected and preprocessed. The processed data is then input into a prediction model to output the predicted service quality distribution within a future prediction window. The call chain topology of multiple atomic services is obtained, and based on their corresponding distribution operators, the predicted service quality distributions of each atomic service are aggregated to output the overall service quality distribution of the edge composite service. This method can output the predicted service quality distribution of each atomic service within a future prediction time window through a prediction model, and simultaneously combine the call chain topology of multiple atomic services to output the overall service quality distribution of the edge composite service. In edge computing environments with high uncertainty, this can improve the accuracy of monitoring the service quality of edge composite services.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, and electronic device for monitoring the quality of service of edge composite services. Background Technology

[0002] With the deep integration of 5G, the Internet of Things (IoT), and the Industrial Internet, the modern computing paradigm is undergoing a fundamental shift from centralized cloud computing to distributed edge computing. By pushing computing power down to the network edge (such as base stations, gateways, or terminal devices), edge computing effectively solves the bottlenecks of traditional cloud computing in terms of bandwidth limitations, transmission latency, and data privacy protection, becoming a key infrastructure supporting digital transformation. Meanwhile, to adapt to the heterogeneity and agile development needs of edge environments, microservice architecture has become the mainstream development model for edge applications. In this architecture, large monolithic applications are decoupled into composite services composed of multiple independent services. These independent services are deployed on geographically dispersed, heterogeneous edge nodes, working collaboratively through complex call chains (such as serial, parallel, and selection structures).

[0003] In the aforementioned mission-critical scenarios, ensuring high Quality of Service (QoS) for these composite services, particularly ensuring that end-to-end latency, throughput, and reliability strictly meet Service Level Agreements (SLAs), is the lifeline for guaranteeing system high availability and security. However, edge computing environments are inherently highly dynamic and resource-constrained. Unlike cloud data centers with controlled environments and abundant resources, edge nodes face multiple challenges, including drastic fluctuations in network connectivity, extremely uneven spatial and temporal load distribution, limited device performance, and frequent cold starts.

[0004] In this context, traditional QoS monitoring and management models are proving inadequate. Traditional monitoring methods mainly rely on static threshold judgments or simple statistical rules based on mean and variance. These methods suffer from significant lag and false positive rates when faced with the complex randomness of edge environments. In summary, existing technologies struggle to accurately monitor the service quality of composite services in highly uncertain edge computing environments. Summary of the Invention

[0005] The purpose of this invention is to provide a service quality monitoring method, apparatus, and electronic device for edge composite services, so as to improve the accuracy of service quality monitoring for edge composite services in edge computing environments with high uncertainty.

[0006] This invention provides a service quality monitoring method for edge composite services, which include multiple atomic services. The method includes: collecting service quality time series data for each atomic service; preprocessing the service quality time series data and outputting the processed data; inputting the processed data into a pre-trained prediction model to output the service quality prediction distribution of the atomic service within a future prediction window; obtaining the call chain topology of multiple atomic services; and aggregating the service quality prediction distribution of each atomic service according to the distribution operator corresponding to the call chain topology to output the overall service quality distribution of the edge composite service.

[0007] Furthermore, the steps for preprocessing service quality time series data and outputting processed data include: cleaning the service quality time series data and outputting the cleaned data; standardizing the cleaned data using pre-saved standard data and outputting the processed data; wherein, the standard data includes: the mean and standard deviation calculated based on the historical training dataset.

[0008] Furthermore, the prediction model includes an encoder and a decoder; the step of inputting the processed data into the pre-trained prediction model to output the service quality prediction distribution of the atomic service within the future prediction window includes: encoding the processed data through the encoder to output a context vector; and decoding the context vector through the decoder to output the service quality prediction distribution of the atomic service within the future prediction window.

[0009] Furthermore, the encoder includes a first network layer and a second network layer; the step of encoding the processed data and outputting a context vector by the encoder includes: extracting features from the processed data in the forward time dimension through the first network layer and outputting a first hidden state vector; extracting features from the processed data in the reverse time dimension through the second network layer and outputting a second hidden state vector; concatenating the first hidden state vector and the second hidden state vector to output a first vector; and performing regularization on the first vector to output a context vector.

[0010] Furthermore, the decoder includes a multilayer perceptron; the step of decoding the context vector through the decoder to output the service quality prediction distribution of the atomic service within the future prediction window includes: mapping the context vector through the multilayer perceptron to output dynamic kernel parameters corresponding to multiple time points within the future prediction window; wherein, the dynamic kernel parameters at each time point include: a preset number of dynamic kernel center values ​​and an adaptive bandwidth value; based on the dynamic kernel center value and adaptive bandwidth value corresponding to each time point, a Gaussian mixture probability density function corresponding to that time point is constructed as the service quality prediction distribution of the atomic service within the future prediction window.

[0011] Furthermore, the prediction model is trained using a composite loss function. The composite loss function is obtained by weighted summation of the negative log-likelihood loss term, the mean anchoring loss term, and the heteroscedasticity tail regularization term. The negative log-likelihood loss term is used to maximize the likelihood probability of the observed data under the service quality prediction distribution. The mean anchoring loss term is used to constrain the error between the mean of the service quality prediction distribution and the true label value. The heteroscedasticity tail regularization term, based on the principle of heteroscedastic regression, is used to construct a continuous penalty relationship between the residual and the prediction bandwidth value. When the true label value deviates from the mean of the service quality prediction distribution, causing the residual to increase and the prediction bandwidth value to decrease, a penalty value with non-linear amplification is generated to force the prediction model to actively increase the prediction bandwidth value in the specified area, thereby covering the potential tail fluctuation risk.

[0012] Furthermore, the call chain topology includes: serial structure, parallel structure, and selection structure. The steps for aggregating the service quality prediction distributions corresponding to each atomic service and outputting the overall service quality distribution of the edge composite service, based on the distribution operators corresponding to the call chain topology, include: if the call chain topology is a serial structure, performing convolution operations on the service quality prediction distributions corresponding to each atomic service to output the overall service quality distribution of the edge composite service; if the call chain topology is a parallel structure, calculating the cumulative distribution function corresponding to each atomic service based on its service quality prediction distribution; multiplying each cumulative distribution function to obtain the product result; differentiating the product result to obtain the probability density function of the edge composite service, and determining the overall service distribution of the edge composite service based on the probability density function; if the call chain topology is a selection structure, obtaining the weight parameters corresponding to each atomic service; and weighting the service quality prediction distributions corresponding to each atomic service according to the weight parameters to obtain the overall service quality distribution of the edge composite service.

[0013] Furthermore, the overall service quality distribution includes: sub-probability density functions corresponding to multiple times within the future prediction window; the method also includes: integrating the sub-probability density function at each time based on a preset service level threshold to obtain the single-step spatial satisfaction probability corresponding to each time; performing a time-series multiplication of the single-step spatial satisfaction probabilities corresponding to all times within the future prediction window to obtain the joint satisfaction probability corresponding to the edge composite service; if the joint satisfaction probability is lower than a preset confidence threshold, an early warning signal is generated.

[0014] This invention provides a service quality monitoring device for edge composite services, which include multiple atomic services. The device includes: a collection module for collecting service quality time series data for each atomic service; a preprocessing module for preprocessing the service quality time series data and outputting the processed data; an output module for inputting the processed data into a pre-trained prediction model to output the service quality prediction distribution of the atomic service within a future prediction window; an acquisition module for acquiring the call chain topology of multiple atomic services; and an aggregation processing module for aggregating the service quality prediction distribution of each atomic service according to the distribution operator corresponding to the call chain topology, and outputting the overall service quality distribution of the edge composite service.

[0015] The present invention provides an electronic device, including a processor and a memory, wherein the memory stores machine-executable instructions that can be executed by the processor, and the processor executes the machine-executable instructions to implement the service quality monitoring method for edge composite services described above.

[0016] This invention provides a service quality monitoring method, apparatus, and electronic device for edge composite services. For each atomic service in the edge composite service, time-series service quality data is collected and preprocessed. The processed data is then input into a prediction model to output the predicted service quality distribution within a future prediction window. The call chain topology of multiple atomic services is obtained, and the predicted service quality distribution of each atomic service is aggregated according to its corresponding distribution operator to output the overall service quality distribution of the edge composite service. This method can output the predicted service quality distribution of each atomic service within a future prediction time window through a prediction model, and simultaneously combine the call chain topology of multiple atomic services to output the overall service quality distribution of the edge composite service. In edge computing environments with high uncertainty, this can improve the accuracy of service quality monitoring for edge composite services. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart of a service quality monitoring method for edge composite services provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the polymerization principle of a serial structure provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the aggregation principle of a parallel structure provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the polymerization principle of a selective structure, provided as an embodiment of the present invention. Figure 5 This is an overall flowchart of a service quality monitoring method for edge composite services provided in an embodiment of the present invention; Figure 6 A schematic diagram of a network architecture for a prediction model provided in an embodiment of the present invention; Figure 7 A schematic diagram illustrating QoS aggregation of an edge composite service provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of distribution prediction under a sliding window provided in an embodiment of the present invention; Figure 9 A schematic diagram illustrating performance gain at the decision-making level, provided as an embodiment of the present invention; Figure 10 This is a schematic diagram of a service quality monitoring device for edge composite services provided in an embodiment of the present invention; Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Currently, to adapt to the heterogeneity and agile development needs of edge environments, microservice architecture has become the mainstream development model for edge applications. In this architecture, large monolithic applications are decoupled into composite services composed of multiple independent services. For example, a typical intelligent video analytics application may include multiple microservices such as video acquisition, preprocessing, inference analysis, and result uploading. These services are deployed on geographically dispersed and performance-heterogeneous edge nodes, working collaboratively through complex call chains (such as serial, parallel, and selection structures).

[0021] In the aforementioned mission-critical scenarios, ensuring high Quality of Service (QoS) for these composite services, particularly ensuring that end-to-end latency, throughput, and reliability strictly meet Service Level Agreements (SLAs), is the lifeline for guaranteeing system high availability and security. However, edge computing environments are inherently highly dynamic and resource-constrained. Unlike cloud data centers with controlled environments and abundant resources, edge nodes face multiple challenges, including drastic fluctuations in network connectivity, extremely uneven spatiotemporal distribution of load, limited device performance, and frequent cold starts. Measured data and theoretical analysis show that the QoS indicators of edge services often exhibit significant non-stationarity and non-Gaussian characteristics. Specifically, the data generation process is modulated by multiple random factors, exhibiting two dominant random characteristics: first, multimodality, reflecting the probabilistic branching of heterogeneous processing paths (such as local execution and cloud offloading); and second, heavy-tailedness, reflecting the extreme latency risk caused by occasional congestion or partial failures.

[0022] Against this backdrop, traditional QoS monitoring and management models are proving inadequate. Traditional monitoring methods primarily rely on static threshold judgments or simple statistical rules based on mean and variance. These methods suffer from severe lag and high false positive rates when facing the complex randomness of edge environments. To address this issue, academia has introduced Signal Temporal Logic (STL) and its probabilistic extensions, such as PrSTL (Probabilistic Signal Temporal Logic) and BSTL (Bayesian Signal Temporal Logic), attempting to quantify and monitor the spatiotemporal constraints of QoS through formal methods. However, existing technologies face four major technical bottlenecks when applied to edge composite service scenarios, which constitute the main technical problems this solution aims to address: First, the structural blind spot of the parametric model, also known as the Gaussian trap.

[0023] Existing technologies (such as PrSTL and BSTL) typically assume, for computational convenience and the existence of closed-form solutions, that QoS prediction errors or state distributions follow a Gaussian (normal) distribution. However, edge services often experience upward-tailed latency due to network congestion, resource contention, or garbage collection mechanisms. Because of its thin-tailed nature, the probability density function of a Gaussian distribution decays exponentially to zero in regions far from the mean. This means that when extreme latency events, though low in probability, have severe consequences in the physical system, models based on the Gaussian assumption may incorrectly predict the probability of those events as zero or extremely low. This model misspecification leads to severe biases in risk assessment by monitoring systems near SLA thresholds, thus masking actual violation risks beyond the SLA line. In industrial practice, this means that while a system may be displaying a green light of safety, it may be on the verge of serious violations. This structural blind spot is unacceptable for safety-critical applications.

[0024] Second, there are deceptive mean traps and false safe zones during the transient load mitigation phase.

[0025] During the dynamic adjustment process of edge computing, a temporary drop in service load and a decrease in average response time (mean) often occur. In traditional parametric models, the decrease in the mean directly leads to a downward shift in the entire confidence interval. Because the variance estimate of the Gaussian model is often lagging or averaged, its upper bound of the confidence interval will quickly fall back to within the SLA safety line, forming a deceptive false safety zone. At this time, the monitor based on the Gaussian model will output a safety confidence level close to 100%. However, the actual non-Gaussian distribution at the edge often has skewed characteristics. Even if the physical mean decreases, its long tail may still be enhanced due to network jitter, even exceeding the SLA threshold. Existing technology cannot identify this complex situation where the mean decreases but the risk does not decrease, leading to false alarms at the most critical moments, resulting in serious decision-making misguidance, such as incorrect resource scaling down, which in turn triggers service avalanche.

[0026] Third, monitoring based on snapshot probability lacks early warning capabilities from a time-series cumulative perspective.

[0027] Most existing predictive monitoring methods rely on marginal probability at a single moment, checking for violations at each time step. This approach is essentially a snapshot-based monitoring method, ignoring the cumulative effect of uncertainty in multi-step prediction. In long-term monitoring scenarios, even if the probability of a single violation at each future moment is low (e.g., 1%), the joint probability of the system trajectory remaining free of violations over a longer time window can decrease significantly due to cumulative effects. For example, the probability of safety at each of the ten consecutive moments may be 99%, but the overall safety probability may drop below 90%. Ignoring this temporal dependency leads to a severe underestimation of long-term system operational risks, preventing monitoring systems from providing sufficient lead time. Alerts are often only issued when a violation is about to occur or has already occurred, leaving maintenance personnel with insufficient time for resource scheduling or degradation before a failure occurs.

[0028] Fourth, there is a lack of distributed aggregation mechanisms for composite service topologies.

[0029] Existing monitoring research largely focuses on atomic services (monolithic services), lacking built-in support for composite service structures. In a microservice architecture, a user request often requires the coordinated processing of multiple atomic services, involving various logical structures such as serial, parallel, and selective processing. The connection relationships (topology) between atomic services determine the overall QoS distribution. When the QoS of atomic services exhibits a complex non-Gaussian distribution, current technologies lack effective mathematical operators to accurately derive the overall SLA satisfaction probability of composite services. The inability to handle these complex distribution transformations makes it impossible to guarantee system reliability from an end-to-end perspective, resulting in monitoring systems only seeing the status of local components and failing to gain insight into the overall health of the business process.

[0030] Based on this, embodiments of the present invention provide a service quality monitoring method, apparatus, and electronic device for edge composite services. This technology can be applied to scenarios that require quality monitoring of edge composite services.

[0031] To facilitate understanding of this embodiment, a service quality monitoring method for edge composite services disclosed in this invention will first be introduced. Edge composite services include multiple atomic services; that is, edge composite services are typically composed of a complex topology consisting of multiple atomic services, which may involve various logical structures such as serial and parallel processing. Figure 1 As shown, the method includes the following steps: Step S102: For each atomic service, collect the service quality time series data of that atomic service; The aforementioned service quality time series data may include, but is not limited to, end-to-end response time, throughput, CPU utilization, memory usage, and energy consumption readings. In actual implementation, service quality time series data of each atomic service in the edge composite service can be collected in real time from edge computing nodes, gateways, or microservice orchestration platforms (such as Kubernetes).

[0032] Step S104: Preprocess the service quality time series data and output the processed data; Due to the instability of the edge environment, the collected service quality time series data often contains problems such as noise and missing values. Therefore, in order to ensure the accuracy of subsequent data processing, it is usually necessary to preprocess the collected service quality time series data, such as removing outliers and filling in missing values, to obtain the processed data.

[0033] Step S106: Input the processed data into the pre-trained prediction model so that the prediction model can output the service quality prediction distribution of the atomic service in the future prediction window. The time length corresponding to the aforementioned future prediction window can be set according to actual needs, such as the next 24 hours. In actual implementation, the processed data corresponding to the atomic service can be input into a pre-trained prediction model. The prediction model processes the processed data and outputs the service quality prediction distribution of the atomic service within the future prediction window. This service quality prediction distribution can reflect all possible values ​​of the atomic service within the future prediction window and their corresponding probabilities.

[0034] Step S108: Obtain the call chain topology of multiple atomic services; In practical implementation, since edge composite services are usually composed of complex topologies consisting of multiple atomic services, the call chain topology of multiple atomic services can be obtained. For example, the call chain dependency graph (such as DAG, Directed Acyclic Graph) or orchestration file of multiple atomic services can be read, and the logical relationships between multiple atomic services can be parsed to identify the call chain topology of multiple atomic services. For example, it can include three basic structures: serial, parallel, and selection.

[0035] Step S110: Based on the distribution operator corresponding to the call chain topology, aggregate the service quality prediction distribution corresponding to each atomic service and output the overall service quality distribution of the edge composite service.

[0036] The aforementioned distribution operators can be understood as mathematical rules associated with the call chain topology, used to calculate the overall service quality distribution of edge composite services. Different call chain topologies typically correspond to different distribution operators. In practical implementation, after obtaining the call chain topologies of the multiple atomic services and the service quality prediction distribution corresponding to each atomic service, the distribution operators corresponding to the call chain topology can be used to aggregate the service quality prediction distributions of each atomic service to obtain the overall service quality distribution of the edge composite services.

[0037] The aforementioned service quality monitoring method for edge composite services collects and preprocesses the service quality time series data of each atomic service within the edge composite service. The processed data is then input into a prediction model to output the predicted service quality distribution within a future prediction window. The method also acquires the call chain topology of multiple atomic services and aggregates their predicted service quality distributions based on their corresponding distribution operators, outputting the overall service quality distribution of the edge composite service. This approach, by using a prediction model to output the predicted service quality distribution of each atomic service within a future prediction time window, and combining the call chain topology of multiple atomic services to output the overall service quality distribution of the edge composite service, can improve the accuracy of service quality monitoring for edge composite services in highly uncertain edge computing environments.

[0038] This invention also provides another service quality monitoring method for edge composite services. This method is implemented based on the method in the above embodiments. In this method, the prediction model includes an encoder and a decoder. That is, in this embodiment, the prediction model adopts an encoder-decoder architecture. The call chain topology includes a serial structure, a parallel structure, and a selection structure. The serial structure can be understood as the tasks corresponding to multiple atomic services being executed sequentially. For example, atomic service B can only execute its corresponding task after atomic service A has completed its corresponding task. The parallel structure can be understood as the tasks corresponding to multiple atomic services being executed independently at the same time. The selection structure can be understood as selecting one task execution path based on preset judgment conditions, instead of selecting another task execution path.

[0039] The method includes the following steps: Step 1: For each atomic service, collect the service quality time series data for that atomic service; Step 2: Clean the service quality time series data and output the cleaned data; In practical implementation, due to the instability of the edge environment, the collected service quality time series data often contains noise and missing values. Therefore, it is necessary to clean the service quality time series data. For example, outlier detection algorithms can be executed to remove physically impossible values ​​(such as negative latency) or extreme outliers caused by sensor malfunctions. The above outlier detection algorithms can be implemented using relevant technologies, such as statistical methods and distance-based methods. For missing data, linear interpolation based on nearest neighbors can be used to fill in the gaps and ensure the continuity of the time series. This linear interpolation method can be implemented using relevant technologies, which will not be elaborated here.

[0040] Step 3: Standardize the cleaned data using pre-saved standard data and output the processed data; the standard data includes the mean and standard deviation calculated based on the historical training dataset.

[0041] The mean and standard deviation can be pre-calculated based on the historical training dataset. These mean and standard deviation are not only used for forward input but also frozen and persistently stored as an independent data block. In this embodiment, to eliminate the influence of different units on data processing, the Z-Score standardization method can be used to normalize the cleaned data. Specifically, the cleaned data can be standardized based on the pre-saved mean and standard deviation to obtain the processed data. The Z-Score standardization method is a standardization method that converts the cleaned data into a standard with a mean of 0 and a standard deviation of 1, which is used to eliminate the differences in units and magnitudes between different variables and make the processed data comparable.

[0042] Step four: Encode the processed data using an encoder to output a context vector; The encoder includes a first network layer and a second network layer; this step can be achieved through steps 40 to 43: Step 40: Extract features from the processed data in the positive time dimension through the first network layer and output the first hidden state vector; Step 41: Extract features from the processed data in the inverse time dimension through the second network layer and output the second hidden state vector; Step 42: Concatenate the first hidden state vector and the second hidden state vector to output the first vector; Step 43: Perform regularization on the first vector and output the context vector.

[0043] Considering that the evolution of service quality for edge composite services is influenced by both long-term historical trends (such as periodic load changes) and recent sudden disturbances (such as instantaneous network congestion), this embodiment selects a Bidirectional Recurrent Neural Network (BRNN) or a Bidirectional Long Short-Term Memory (Bi-LSTM) network as the backbone encoder. The input to the encoder is the data processed as described above. The encoder contains two independent network layers, namely the first network layer and the second network layer, which scan and extract features from the processed data in the forward and backward time dimensions, respectively. By extracting features in the forward time dimension, the causal dependencies of historical evolution can be captured, and by extracting features in the backward time dimension, the smoothing and completion of recent context can be enhanced. Finally, the first hidden state vector output by the first network layer and the second hidden state vector output by the second network layer are directly concatenated to output the first vector. To enhance the robustness of the prediction model to abnormal conditions, the first vector can be regularized using a dropout mechanism to form a fixed-dimensional context vector. This context vector highly aggregates all key temporal features within the input window, providing sufficient conditional information for subsequent distribution generation.

[0044] Step 5: Decode the context vector using a decoder to output the service quality prediction distribution of the atomic service within the future prediction window; The decoder includes a multilayer perceptron; this step can be achieved through the following steps 50 to 51: Step 50: The context vector is mapped using a multilayer perceptron to output the dynamic kernel parameters corresponding to multiple time points within the future prediction window; wherein, the dynamic kernel parameters at each time point include: a preset number of dynamic kernel center values ​​and an adaptive bandwidth value; Step 51: Based on the dynamic kernel center value and adaptive bandwidth value corresponding to each time moment, construct the Gaussian mixture probability density function corresponding to that time moment, so as to serve as the service quality prediction distribution of that atomic service in the future prediction window.

[0045] To achieve efficient non-parametric distribution fitting on edge devices, this embodiment designs a sparse KDE (Kernel Density Estimation) decoder. This decoder head abandons the dependence of non-parametric methods on the entire dataset and uses a multilayer perceptron (MLP) to directly map the context vector to a set of dynamic kernel parameters. Specifically, within a future prediction window of length H, for each prediction step in the future, the MLP outputs K dynamic kernel center values ​​(K is recommended to be between 8 and 16) and an adaptive bandwidth value. The length H of the future prediction window can be understood as the total number of future prediction moments. For example, H=3 indicates predicting 3 moments within the future prediction window, meaning the MLP outputs dynamic kernel parameters corresponding to the 3 moments, each containing K dynamic kernel center values ​​and an adaptive bandwidth value.

[0046] The generation process of dynamic kernel center values ​​is further explained below. For each time step within the future prediction window, the multilayer perceptron outputs K scalar values, representing the center positions of the K Gaussian kernels in the standardized QoS numerical space (i.e., the K dynamic kernel center values). To prevent the mode collapse problem common in mixed-density networks, this scheme forces all dynamic kernel center values ​​to have equal mixing weights, all of which are 1 / K. This design compels the multilayer perceptron in the decoder to physically move the positions of these dynamic kernel center values ​​to adapt to the high-density regions of the data. For example, if the QoS distribution exhibits a bimodal characteristic, the multilayer perceptron in the decoder will automatically move some dynamic kernel center values ​​to the vicinity of the first peak and another part to the vicinity of the second peak, thereby naturally fitting the multimodal distribution.

[0047] The following section further explains the process of generating the adaptive bandwidth value. The multilayer perceptron in the decoder first outputs a scalar value as the shared bandwidth value. To ensure the positive definiteness and numerical stability of the bandwidth, a sigmoid activation function is used to scale this shared bandwidth value and linearly map it to a preset positive real number range (e.g., 0.01 to 1.0). Bandwidth, in a physical sense, corresponds to the aleatoric uncertainty of the system. When the processed data input to the encoder exhibits high fluctuations or strong noise, the multilayer perceptron in the decoder automatically predicts a larger adaptive bandwidth value, making the generated probability density function flatter and wider, thereby covering potential long-tail regions.

[0048] In this embodiment, during the training of the prediction model, in order to enable it to still keenly capture abnormal risks when facing imbalanced data, the prediction model is trained using a composite loss function; the composite loss function is obtained by weighted summation of the negative log-likelihood loss (NLL Loss), the mean anchoring loss, and the heteroscedastic tail regularization. The negative log-likelihood loss term is the main loss term, used to maximize the likelihood probability of the observed data under the service quality prediction distribution. It takes the negative logarithm of this likelihood probability, transforming it into a loss function that is as small as possible. By minimizing the negative log-likelihood loss term, the prediction model during training is driven to generate a Gaussian mixture distribution that approximates the true label value in shape. The calculation employs the Log-Sum-Exp (log-sum-exponential) technique to ensure numerical stability. It can also be understood that the negative log-likelihood loss term is responsible for "drawing the general outline," ensuring that the predicted service quality distribution approximates the true label value in its overall shape.

[0049] The mean-anchoring loss term constrains the error between the mean of the service quality prediction distribution and the true label value. This term calculates the mean squared error between the mean of the service quality prediction distribution and the true label value during training. This is equivalent to imposing a strong prior constraint on the prediction model during training, requiring the center of gravity of the distribution to closely follow the actual physical trend, thus preventing distribution drift in the early stages of training and accelerating convergence. It can also be understood as the mean-anchoring loss term "grasping the center of gravity," forcing the mean of the predicted service quality distribution to follow the physical trend and preventing the model from deviating from its intended path.

[0050] The heteroscedasticity tail regularization term is based on the principle of heteroscedastic regression and is used to construct a continuous penalty relationship between the residual and the predicted bandwidth value. When the true label value deviates from the mean of the service quality prediction distribution, causing the residual to increase and the predicted bandwidth value to decrease, a penalty value with non-linear amplification is generated to force the prediction model to actively increase the predicted bandwidth value in the specified area, thereby covering the potential tail fluctuation risk.

[0051] The heteroscedasticity tail regularization term is a key technique in this scheme to combat the mean trap. Based on the principle of heteroscedastic regression, this term has two mathematical parts: the first is the squared residual divided by the squared bandwidth, and the second is the logarithm of the bandwidth. When the true label value deviates significantly from the mean of the predicted service quality distribution (i.e., a tail-burst event occurs, resulting in a large residual), if the bandwidth predicted by the prediction model during training is small, the first part will incur a huge penalty. To reduce the total loss, the prediction model is forced to increase the bandwidth. While increasing the bandwidth increases the value of the second part (the logarithmic term), it significantly reduces the value of the first part. Through this game-theoretic mechanism, the prediction model learns to proactively increase the uncertainty range in areas of uncertainty or high volatility. This ensures that even if the mean of the predicted service quality distribution seems safe, as long as there is potential for drastic fluctuations, the prediction model will expand the tail of the distribution, extending it into dangerous areas. It can also be understood that the heteroscedasticity tail regularization term is responsible for "fallback protection against missed reports," specifically designed to deal with extreme long-tail risks that cannot be fitted by a Gaussian distribution.

[0052] It's important to note that during the training of the prediction model, it's not required that the negative log-likelihood loss term, mean-anchored loss term, and heteroscedasticity tail regularization term converge individually. Only the total loss of the weighted composite loss function needs to converge. The heteroscedasticity tail regularization term forces a larger bandwidth to cover risk, but this may slightly sacrifice the accuracy of the negative log-likelihood loss term. Therefore, the ultimate goal of model training is the convergence of the composite loss function, i.e., finding the optimal balance point among these three loss terms, rather than requiring the loss value of each term to decrease to zero.

[0053] Step 6: Obtain the call chain topology of multiple atomic services; After obtaining the service quality prediction distribution for each atomic service within the future prediction window, the call chain topology of multiple atomic services can be obtained. Based on this call chain topology, the service quality prediction distribution for each atomic service is aggregated into the overall service quality distribution of the edge composite service using its corresponding distribution operator. Specifically, the call chain dependency graph or orchestration file of the microservices can be read first to parse the logical relationships between atomic services and identify three basic structures: serial, parallel, and selection. That is, this embodiment can support three basic combination modes.

[0054] Step 7: If the call chain topology is a serial structure, perform convolution operation on the service quality prediction distribution corresponding to each atomic service, and output the overall service quality distribution of the edge composite service. Specifically, such as Figure 2The diagram illustrates the aggregation principle of a serial structure. For a service chain with serial calls (e.g., service A calling service B), the total latency of the edge composite service is the algebraic sum of the latency of each atomic service. In probability theory, the probability density function of the sum of two independent random variables is equal to the convolution of their individual probability density functions. This scheme utilizes this principle to perform convolution operations on the predicted service quality distribution corresponding to each predicted atomic service. Since this scheme uses a Gaussian mixture form, the convolution of two Gaussian mixture distributions is still a Gaussian mixture distribution (its kernel center is the sum of the original kernel centers, and its variance is the sum of the original variances). This analytical solution characteristic makes the calculation of serial aggregation very efficient. This step can accurately reflect the long-tail amplification effect, that is, after two services with slightly long tails are superimposed, the tail of their composite distribution will become significantly longer, and the risk will increase non-linearly.

[0055] Step 8: If the call chain topology is a parallel structure, calculate the cumulative distribution function corresponding to each atomic service based on the service quality prediction distribution corresponding to each atomic service. Step nine: Multiply each cumulative distribution function to obtain the product result; Step 10: Differentiate the product result to obtain the probability density function of the edge composite service, and determine the overall service distribution of the edge composite service based on the probability density function. like Figure 3 The diagram illustrates the aggregation principle of a parallel structure. For a service chain with parallel calls, the latency of edge composite services typically depends on the slowest branch, i.e., the maximum latency of each atomic service. This corresponds to the order statistics problem in probability theory. The system, based on extreme value theory, calculates the distribution of the maximum value using the cumulative distribution function (CDF) of each atomic service. Specifically, the service quality prediction distributions corresponding to each atomic service are integrated to obtain the CDF of each atomic service. The CDF of the edge composite service (corresponding to the product above) is equal to the product of the CDFs of the atomic services. By differentiating this product, the probability density function of the edge composite service can be obtained, thus determining the overall service distribution of the edge composite service. This algorithm can accurately assess the constraint of system bottlenecks on overall performance and reflects the filtering or amplification effect of parallel redundancy structures on extreme latency.

[0056] Existing edge monitoring technologies, when handling parallel microservices, typically only compare a few "deterministic single-point values" (e.g., predicting A's latency to be 100ms and B's to be 150ms, assuming the overall latency is 150ms). Current technologies lack a mechanism for topological aggregation of highly uncertain "probability distributions." This solution innovatively introduces order statistics into the performance evaluation of microservice call chains, integrating the underlying Gaussian mixture probability density function (PDF) into a cumulative distribution function (CDF) to calculate the parallel bottleneck effect.

[0057] Step 11: If the call chain topology is a selection structure, obtain the weight parameters corresponding to each atomic service. Step 12: Perform a weighted calculation on the service quality prediction distribution corresponding to each atomic service according to the weight parameters to obtain the overall service quality distribution of the edge composite service.

[0058] like Figure 4 The diagram illustrates the aggregation principle of a selection structure. For selection structures based on conditional judgment (such as atomic service A / atomic service B testing or content-based routing), the overall service quality distribution of edge composite services is a weighted mixture of the service quality prediction distributions corresponding to each atomic service. Based on historical call logs or preset routing probabilities, the weight parameters corresponding to each atomic service are calculated, and the service quality prediction distributions corresponding to each atomic distribution are linearly superimposed according to the weight parameters to generate the overall service quality distribution of edge composite services.

[0059] Step 13: Based on the preset service level threshold, integrate the sub-probability density function at each time step to obtain the single-step spatial satisfaction probability at each time step. The aforementioned service level thresholds can be defined by operations personnel using Signal-Time Logic (STL) to define SLA requirements, for example, "The service response time must always be less than 380 milliseconds within the next H time steps." The system transforms this natural language description into a formalized STL formula. This formula includes timing operators (such as the Always global operator and the Eventually existent operator) and atomic propositions (such as numerical comparisons). The overall service quality distribution includes: sub-probability density functions corresponding to multiple time points within the future prediction window; in actual implementation, for a given SLA threshold (such as the response time must be less than 380 milliseconds), the sub-probability density function PDF of each time step can be integrated to obtain the corresponding cumulative distribution function CDF value. This CDF value represents the single-step spatial satisfaction probability of meeting the SLA safety requirements at that single time point.

[0060] Step 14: Perform a time-series multiplication of the single-step spatial satisfaction probabilities corresponding to all times within the future prediction window to obtain the joint satisfaction probability corresponding to the edge composite service. To calculate the temporal joint probability of operators such as "Always" in the STL formula, this scheme employs an efficient analytical aggregation mechanism. Based on the assumption of temporal independence, the system mathematically multiplies the single-step spatial satisfaction probabilities of all discrete time steps within the aforementioned future prediction window. This multiplication mechanism yields the joint satisfaction probability corresponding to edge composite services, strictly penalizing uncertainties in the temporal dimension: as long as the risk of violation increases at any point in the future long-term time interval (i.e., the single-step spatial satisfaction probability decreases), the joint satisfaction probability obtained by multiplication will be significantly affected and drop sharply. This allows the monitoring system to keenly capture the cumulative effect of risk on the time axis.

[0061] Step 15: If the probability of joint satisfaction is lower than the preset confidence threshold, an early warning signal is generated.

[0062] The scheme first calculates the single-step spatial satisfaction probability at each independent time point in the future, and then directly multiplies these single-step spatial satisfaction probabilities mathematically. The result of the product is the joint satisfaction probability that the system will not violate any rules throughout a continuous future period. This method retains extreme sensitivity to the risk of accumulated time (if the probability decreases at any point, the multiplied joint satisfaction probability will drop sharply), thus achieving extremely sensitive early warning.

[0063] Specifically, the calculated joint satisfaction probability can be compared with a preset confidence threshold (e.g., 0.95). When the joint satisfaction probability falls below the confidence threshold, an early warning signal can be triggered immediately. Due to the use of joint satisfaction probability and non-Gaussian modeling, this signal often experiences a significant drop (decisive plungge) several time steps before the physical QoS metric actually exceeds the SLA threshold. After the system detects this characteristic, it sends an alarm to the operations and maintenance platform, along with a predicted risk distribution map and key cause analysis, prompting operations and maintenance personnel or the automatic controller to take intervention measures such as traffic scrubbing, service degradation, or resource expansion.

[0064] For ease of understanding, see Figure 5 The diagram illustrates an overall flowchart of a service quality monitoring method for edge composite services. This solution can be implemented through an end-to-end intelligent monitoring pipeline, which mainly consists of four parts: a data preprocessing and feature engineering module, a deep sparse probability prediction module, a composite service distribution aggregation module, and an online time-series logic monitoring module. The following sections will provide a more detailed explanation of this solution, focusing on the execution steps of each component.

[0065] Step 1: Data Preprocessing and Feature Engineering: This step aims to transform the raw edge QoS stream data into a normalized tensor suitable for deep learning model processing through the data preprocessing and feature engineering module.

[0066] 1. Data Acquisition and Anomaly Cleaning: QoS time-series data for each atomic service is collected in real time from edge computing nodes, gateways, or microservice orchestration platforms (such as Kubernetes). Due to the instability of the edge environment, the collected data often contains noise and missing values. Anomaly detection algorithms are first executed through data preprocessing and feature engineering modules to remove physically impossible values ​​(such as negative latency) or extreme outliers caused by sensor malfunctions. For missing data, linear interpolation based on nearest neighbors is used to fill in the gaps, ensuring the continuity of the time series.

[0067] 2. Z-Score Standardization and Physical Scale Inverse Restoration Mechanism: To eliminate the influence of different scales on model training, this scheme uses the Z-Score standardization method to normalize the feature matrix X within the sliding window. It is particularly important to emphasize that this step not only calculates the mean and standard deviation of the historical training dataset for forward input, but also freezes and persistently stores these statistical parameters as an independent data block. The key to this design is that since the subsequent deep sparse KDE prediction model (corresponding to the prediction model mentioned above) outputs the dynamic kernel center and adaptive bandwidth of the probability distribution, rather than specific numerical values, the system must recall these frozen statistics at the end of the inference phase to perform inverse standardization (inverse normalization) on the predicted probability density function, thereby accurately restoring it to the QoS distribution scale with actual physical meaning (corresponding to the service quality prediction distribution mentioned above). Furthermore, when standardizing on the validation and test sets, only uniform statistics calculated from the training set must be used to strictly prevent future information leakage.

[0068] 3. Sliding Window Sample Construction: To transform the time series prediction problem into a supervised learning problem, this scheme uses the sliding window technique to construct sample pairs. The input historical window length is set to L. For each time t, the system extracts the historical observation sequence from t-L+1 to time t as the input feature matrix X, which is then used by the subsequent network to extract time series features.

[0069] Step 2: Construction and training of the deep sparse KDE prediction model; This step aims to train a deep neural network capable of accurately outputting sparse KDE parameters, which is the core engine for uncertainty perception. The network adopts an encoder-decoder architecture.

[0070] 1. Construction of a bidirectional temporal feature encoder: This invention selects a bidirectional recurrent neural network (BRNN) or a bidirectional long short-term memory network (Bi-LSTM) as the backbone encoder. For example... Figure 6 The diagram shows a network architecture for a prediction model. The encoder's input is the normalized temporal sequence output from the previous processing step. The encoder contains two independent network layers that scan and extract features from the input sequence along the forward and backward time dimensions, respectively, ultimately forming a fixed-dimensional context vector.

[0071] 2. Sparse KDE Decoding Header Design: This scheme utilizes a multilayer perceptron (MLP) to directly map the context vector to dynamic kernel parameters corresponding to multiple time points within the future prediction window. The dynamic kernel parameters at each time point include a preset number of dynamic kernel center values ​​and an adaptive bandwidth value to construct the Gaussian mixture probability density function corresponding to that time point, which serves as the service quality prediction distribution of the atomic service within the future prediction window.

[0072] 3. Composite Loss Function Optimization: In order to train the above network so that it can still keenly capture the risk of anomalies when faced with imbalanced data, this scheme designs a composite loss function with three sub-objectives, namely the negative log-likelihood loss term, the mean anchoring loss term, and the heteroscedasticity tail regularization term.

[0073] Step 3: Spatial Aggregation of Composite Service QoS Distribution 1. Topology analysis: This aims to analyze the logical relationships between atomic services and identify three basic structures: serial, parallel, and selection.

[0074] 2. Distribution Aggregation Operator: Based on different topologies, the corresponding distribution operator is used to aggregate the service quality prediction distribution corresponding to each atomic service, and output the overall service quality distribution of the edge composite service.

[0075] Step 4: Online timing logic monitoring and early warning based on KDESTL 1. STL specification formalization: After obtaining the overall service quality distribution of edge composite services, this step performs online temporal logic monitoring, calculates the probability of satisfying the temporal logic specification, and makes early warning decisions.

[0076] 2. Spatial Satisfaction Calculation: Following the overall service quality distribution output in step three, this step first performs spatial probability calculations for each independent time step within the future prediction window. Specifically, for a given SLA threshold (e.g., response time must be less than 380 milliseconds), the sub-probability density function for each time step is integrated to obtain the corresponding cumulative distribution function (CDF) value. This CDF value represents the single-step spatial satisfaction probability of the system meeting the SLA safety requirements at that single moment.

[0077] 3. Temporal multiplication aggregation and joint probability calculation: The single-step spatial satisfaction probabilities of all discrete time steps within the above future prediction window are mathematically multiplied to obtain the joint satisfaction probability corresponding to the edge composite service.

[0078] When the probability of joint satisfaction is lower than the preset confidence threshold, an early warning signal can be generated proactively.

[0079] like Figure 7 The diagram illustrates a QoS aggregation of edge composite services. It shows the true distribution of atomic services when combined. While the independent distributions of atomic services A (blue dashed line) and B (purple dotted line) appear secure, the overall service quality distribution (red solid line) obtained through aggregation using the KDE-based distribution operator employed in this solution exhibits significant multimodal and long-tail characteristics. The tail of this composite distribution structurally penetrates the 350ms SLA threshold (shaded area), demonstrating that this solution can accurately capture the overall composite risks missed by traditional single-atom service monitoring.

[0080] like Figure 8 The diagram illustrates a distribution prediction using a sliding window, showcasing the future time-series evolution. Traditional parametric Gaussian models (dashed line distribution) are too rigid and may mistakenly identify a period of absolute safety when the mean briefly decreases. In contrast, the non-parametric KDE distribution predicted by this scheme (red solid line distribution) independently and accurately characterizes the true probability density at each time step (green filled area), avoiding the misleading effect of "false safety zones."

[0081] like Figure 9 The diagram illustrates a performance gain at the decision-making level, comparing the final monitoring satisfaction probability curve. The existing PrSTL method (yellow dashed line, marked with squares) relies on independent evaluation at each discrete time point, reacting extremely slowly under the illusion of risk mitigation. In contrast, this approach calculates the joint satisfaction probability within the future prediction window (red solid line, marked with circles). The curve decisively falls below the confidence threshold several time steps before the system actually violates the rules, providing invaluable early warning time for system scalability.

[0082] The above Figures 7-9 This visually demonstrates the core advantages of this solution in handling composite service aggregation, dealing with the Gaussian mean trap, and achieving early warning of joint probability.

[0083] In summary, this solution provides an uncertainty-aware QoS timing logic monitoring method and system for edge composite services. It aims to overcome the theoretical bottleneck of the traditional Gaussian assumption by constructing an intelligent monitoring framework that integrates deep learning, nonparametric statistics, and formal methods. This will solve the problem of accurate prediction and early warning under non-Gaussian and heavy-tailed distributions, and achieve perception and distribution aggregation of composite service topology, thus realizing a leap from passive response to proactive early warning.

[0084] This solution primarily addresses the problem that existing predictive monitoring methods cannot provide non-parametric probabilistic guarantees for edge composite services over continuous time intervals in highly uncertain edge computing environments. Specifically, the following issues exist: 1. Deficiencies in system monitoring dimensions: Existing technologies lack the ability to quantify "temporal joint probability" and "combined joint probability".

[0085] The pain point of the temporal dimension: Existing probabilistic STLs (such as PrSTL) are semantically based only on the extreme values ​​of probabilities at a single moment (instantaneous), meaning they can only monitor whether the system is safe at a specific point. This leads to a fatal problem: it ignores the "joint temporal probability" over the entire time interval. In other words, it cannot guarantee the probability that the system will remain safe "as a whole over the entire future period".

[0086] Pain points in the spatial dimension: Existing monitoring technologies mainly target atomic services and cannot handle composite services with complex topologies such as serial, parallel, and conditional selection. They cannot simply add up the monitoring results of atomic services to obtain the overall service violation probability.

[0087] The solution proposed in this paper is JPSTL (Joint Probability Signal Temporal Logic), which extends the evaluation of temporal joint probability and combinatorial joint probability with new metrics.

[0088] 2. Deficiencies in underlying distribution modeling: Existing models cannot support non-parametric QoS distribution prediction.

[0089] The pain point of underlying prediction: To quantify the randomness of edge environments, existing methods typically model QoS as a probability distribution. However, they generally simplify or assume that these distributions are "parameterized" (e.g., forced to assume a Gaussian distribution), which is unsuitable for real-world edge scenarios where environmental prediction uncertainty is unknown or that exhibit multimodal, heavy-tailed characteristics. If the underlying distribution assumptions are incorrect, the upper-layer logical monitoring will fail.

[0090] The solution proposed in this invention is to introduce a deep Bayesian neural network combined with sparse kernel density estimation (SparseKDE) to generate a non-parametric QoS future distribution sequence without relying on strict parameterization assumptions.

[0091] In summary, this technical solution mainly includes the following four core innovations: Innovation Point 1: End-to-end depth probabilistic prediction architecture based on sparse parameterization To address the challenge of fitting arbitrarily complex non-Gaussian distributions within the limited storage and computing resources of edge devices, this solution abandons the traditional kernel density estimation (KDE) method, which requires storing the entire historical data, and innovatively designs an end-to-end deep sparse KDE prediction network. This architecture mainly consists of a bidirectional temporal feature encoder and a sparse KDE decoder head. The sparse KDE decoder head is the core, utilizing a multilayer perceptron (MLP) as the decoder to directly map a set of minimally simplistic dynamic kernel parameters (including dynamic kernel centers and adaptive bandwidth) from the context vector. The dynamic kernel centers are designed to automatically cluster in high-density regions of the data, naturally capturing multimodal features; the adaptive bandwidth mechanism, as a measure of random uncertainty, automatically adjusts the smoothness of the distribution according to environmental fluctuations. This model can accurately fit arbitrarily complex multimodal distributions in constant-time complexity using a mixture of Gaussian kernel functions, completely solving the structural blind spot problem caused by the inability of parameterized Gaussian models to fit long-tailed distributions. This solution overcomes the limitation of existing technologies (such as PrSTL and BSTL) that strongly rely on parameterized Gaussian distributions. By combining a bidirectional time encoder with an MLP decoder, the dynamic kernel center and adaptive bandwidth are directly output, achieving accurate fitting of extreme long-tail and multimodal distributions under conditions of limited edge resources.

[0092] Innovation Point Two: Tail-Enhanced Composite Loss Function Resistant to Mean Trap To address the issue of deep learning models tending to fit most normal samples while ignoring a few extreme outliers during training, this solution innovatively designs a composite loss function that includes a heteroscedasticity tail regularization term. This loss function not only focuses on negative log-likelihood (NLL) to improve distribution fit but also introduces a mean-anchored loss term to stabilize trend predictions. Most importantly, it introduces a tail penalty term. Based on the principle of heteroscedasticity regression, this term imposes a high penalty when the true label value deviates significantly from the mean of the service quality prediction distribution (i.e., a tail event occurs) and the model's output prediction bandwidth is small. This forces the model to proactively increase bandwidth to cover potential risks when facing inaccurate predictions. This mechanism allows the model to maintain sensitivity to potential violations by expanding the distribution tail during the transient mitigation period of declining physical mean, effectively breaking the mean trap and eliminating false safe zones.

[0093] Innovation Point 3: Joint Probabilistic Proactive Early Warning Mechanism Based on Temporal Multiplication Aggregation This scheme overcomes the limitation of existing PrSTL methods that only focus on single-step spatial satisfaction probabilities, proposing a monitoring semantic based on joint satisfaction probabilities. Instead of calculating the violation probability at a single moment, this scheme focuses on the probability that the system state will remain below the safety threshold within a future prediction window. By performing a temporal multiplication of the single-step spatial satisfaction probabilities at each step within the future prediction window, a large number of future evolution trajectories are generated from the predicted probability density function. The proportion of trajectories satisfying STL reductions is then calculated, enabling this scheme to accurately quantify the probability that the system will remain safe within a future time window. Compared to single-step spatial satisfaction probabilities, joint satisfaction probabilities can capture the cumulative effect of risk over time and are more sensitive to trend deterioration. Experiments demonstrate that this mechanism can issue high-confidence warning signals based on structural deterioration of probabilistic trends before physical violations occur, significantly extending the system's early warning time and providing a valuable operational window for fault avoidance. This scheme abandons the lagging approach of traditional methods that only focus on single-point marginal probabilities (snapshot-style) and innovatively calculates temporal joint probabilities within continuous time windows. By multiplying the smoothed single-step spatial satisfaction probabilities of each independent time step within the future prediction window, the system becomes extremely sensitive to the cumulative risk on the time axis.

[0094] Innovation Point 4: Support for Topology-Aware Composite Service Distribution Aggregation Operators Addressing the structural characteristics of microservice call chains, this solution establishes a complete probabilistic aggregation theory for composite service QoS. Instead of relying on simple numerical addition and subtraction, this solution derives the transmission rules of QoS from a probability distribution perspective. For serial structures, a distributed convolution operator accelerated by Fast Fourier Transform (FFT) is designed to accurately calculate the long-tail amplification effect after latency superposition. For parallel structures, a distributed aggregation operator is designed using extremum theory and order statistics to assess the impact of system bottlenecks on overall performance. For selection structures, a hybrid model operator is employed. These operators achieve a precise bottom-up derivation from the non-Gaussian distribution of sub-services to the overall reliability of composite services, filling the gap in formal monitoring methods under microservice architectures. This solution innovatively introduces low-level mathematical operators such as distributed convolution, achieving a precise derivation from the non-Gaussian distribution of atomic services to the end-to-end overall distribution of composite services.

[0095] This solution involves the interdisciplinary technologies of edge computing, distributed system reliability engineering, service computing, and artificial intelligence. Specifically, it relates to a method and system for predicting Quality of Service (QoS), quantifying uncertainty, distributive aggregation, and proactive monitoring and early warning based on signal timing logic (STL) for microservice architectures (i.e., edge composite services) in edge computing environments. This solution primarily addresses how to provide end-to-end Service Level Agreement (SLA) guarantees and early fault warnings for edge composite services composed of multiple atomic services in highly dynamic, resource-constrained edge network environments with significant non-Gaussian noise characteristics. It is particularly suitable for critical mission scenarios such as autonomous driving, industrial internet, telemedicine, and smart grids, which are sensitive to latency and have extremely high reliability requirements.

[0096] The aforementioned service quality monitoring method for edge composite services represents a significant improvement and has beneficial effects compared to existing technologies, specifically in the following aspects: First, the advance warning time has been significantly extended.

[0097] By employing non-parametric modeling based on sparse KDE and joint satisfaction probability calculation, this scheme can penetrate the random fluctuations of data and keenly capture the early accumulation of risk at the distribution level. Experimental data show that, under the same false alarm rate, this scheme can detect potential violations 2 to 3 time steps earlier than traditional PrSTL and BSTL methods based on Gaussian assumptions. Especially when facing deceptive mean trap scenarios, where traditional methods misjudge safety due to a decrease in the mean, this scheme's monitoring signal can decisively drop and lock into the negative interval, thereby eliminating false sense of security, greatly reducing the false alarm rate, and buying valuable buffer time for the system's elastic scaling.

[0098] Second, excellent robustness in non-Gaussian environments.

[0099] This scheme abandons the Gaussian distribution assumption and adaptively fits QoS distributions of arbitrary shapes through a data-driven approach. It demonstrates exceptional robustness when handling edge data with significant multimodal and heavy-tailed characteristics (such as wind power data or response time under high concurrency). When the SLA threshold is set in the tail region of the distribution, the performance of existing Gaussian methods often drops sharply, with the ROC-AUC (Area Under the Receiver Operating Characteristic Curve) falling to near-random guess levels, resulting in numerous false negatives. However, this scheme maintains high accuracy and recall, proving its applicability in complex statistical environments.

[0100] Third, it fills the technological gap in the accurate aggregation of probability distributions for edge composite services.

[0101] This innovative solution introduces aggregation operators based on distributed convolution and order statistics, establishing a clear monitoring link from components to the system in a microservice architecture. Explicitly modeling the distributed aggregation of composite services provides a higher safety margin than simply monitoring the average of individual atomic services or their simple summation. This solution can accurately quantify the long-tail amplification effect caused by the superposition of fluctuations in sub-services, avoiding the blind spot of "seemingly safe locally, but actually violating the overall system," and providing solid theoretical support for the end-to-end reliability of microservice architectures.

[0102] Fourth, it balances the need for algorithm accuracy with the lightweight requirements of edge deployment.

[0103] Although this solution employs a non-parametric statistical approach, it transforms the inference process into efficient matrix operations through an innovative deep sparse parametric network design. The inference phase no longer requires retrieving massive amounts of historical samples; only a very small number of kernel parameters (e.g., 8 sets) need to be generated, resulting in constant-time computation. This enables the system to be lightweightly deployed on resource-constrained edge gateways or servers, meeting millisecond-level real-time monitoring requirements, and possesses significant engineering application value and broad industrialization prospects.

[0104] Fifth, it provides a transparent and interpretable risk measurement.

[0105] The joint satisfaction probability output by this solution has a clear physical meaning, namely, the confidence level of the system's continued reliability over a future period. This provides operations and maintenance personnel with an intuitive view of risks, avoiding the problems of abstract robustness values ​​being difficult to understand and threshold setting in traditional methods, and improving the efficiency of human-machine collaboration.

[0106] This invention provides a service quality monitoring device for edge composite services, which include multiple atomic services; such as Figure 10As shown, the device includes: a data acquisition module 1001, used to acquire service quality time series data for each atomic service; a preprocessing module 1002, used to preprocess the service quality time series data and output the processed data; an output module 1003, used to input the processed data into a pre-trained prediction model, so as to output the service quality prediction distribution of the atomic service within a future prediction window through the prediction model; an acquisition module 1004, used to acquire the call chain topology structure of multiple atomic services; and an aggregation processing module 1005, used to aggregate the service quality prediction distribution corresponding to each atomic service according to the distribution operator corresponding to the call chain topology structure, and output the overall service quality distribution of the edge composite service.

[0107] The aforementioned service quality monitoring device for edge composite services can output the predicted service quality distribution of each atomic service within a future prediction time window through a predictive model. At the same time, it combines the call chain topology of multiple atomic services to output the overall service quality distribution of edge composite services. In edge computing environments with high uncertainty, this can improve the accuracy of monitoring the service quality of edge composite services.

[0108] Furthermore, the preprocessing module is also used to: clean the service quality time series data and output the cleaned data; and to standardize the cleaned data using pre-saved standard data and output the processed data; wherein the standard data includes the mean and standard deviation calculated based on the historical training dataset.

[0109] Furthermore, the prediction model includes an encoder and a decoder; the output module is also used to: encode the processed data through the encoder to output a context vector; and decode the context vector through the decoder to output the service quality prediction distribution of the atomic service within the future prediction window.

[0110] Furthermore, the encoder includes a first network layer and a second network layer; the output module is also used to: extract features from the processed data in the forward time dimension through the first network layer and output a first hidden state vector; extract features from the processed data in the reverse time dimension through the second network layer and output a second hidden state vector; concatenate the first hidden state vector and the second hidden state vector to output a first vector; and perform regularization on the first vector to output a context vector.

[0111] Furthermore, the decoder includes a multilayer perceptron; the output module is also used to: map the context vector through the multilayer perceptron and output the dynamic kernel parameters corresponding to multiple time points within the future prediction window; wherein, the dynamic kernel parameters at each time point include: a preset number of dynamic kernel center values ​​and an adaptive bandwidth value; based on the dynamic kernel center value and adaptive bandwidth value corresponding to each time point, construct the Gaussian mixture probability density function corresponding to that time point, as the service quality prediction distribution of the atomic service within the future prediction window.

[0112] Furthermore, the prediction model is trained using a composite loss function. The composite loss function is obtained by weighted summation of the negative log-likelihood loss term, the mean anchoring loss term, and the heteroscedasticity tail regularization term. The negative log-likelihood loss term is used to maximize the likelihood probability of the observed data under the service quality prediction distribution. The mean anchoring loss term is used to constrain the error between the mean of the service quality prediction distribution and the true label value. The heteroscedasticity tail regularization term, based on the principle of heteroscedastic regression, is used to construct a continuous penalty relationship between the residual and the prediction bandwidth value. When the true label value deviates from the mean of the service quality prediction distribution, causing the residual to increase and the prediction bandwidth value to decrease, a penalty value with non-linear amplification is generated to force the prediction model to actively increase the prediction bandwidth value in the specified area, thereby covering the potential tail fluctuation risk.

[0113] Furthermore, the call chain topology includes: serial structure, parallel structure, and selection structure; the aggregation processing module is also used for: if the call chain topology is a serial structure, performing convolution operations on the service quality prediction distribution corresponding to each atomic service to output the overall service quality distribution of the edge composite service; if the call chain topology is a parallel structure, calculating the cumulative distribution function corresponding to each atomic service based on the service quality prediction distribution corresponding to each atomic service; multiplying each cumulative distribution function to obtain the product result; differentiating the product result to obtain the probability density function of the edge composite service, and determining the overall service distribution of the edge composite service based on the probability density function; if the call chain topology is a selection structure, obtaining the weight parameters corresponding to each atomic service; and performing weighted calculations on the service quality prediction distribution corresponding to each atomic service according to the weight parameters to obtain the overall service quality distribution of the edge composite service.

[0114] Furthermore, the overall service quality distribution includes: sub-probability density functions corresponding to multiple times within the future prediction window; the device is also used to: integrate the sub-probability density function at each time based on a preset service level threshold to obtain the single-step spatial satisfaction probability corresponding to each time; perform a time-series multiplication of the single-step spatial satisfaction probabilities corresponding to all times within the future prediction window to obtain the joint satisfaction probability corresponding to the edge composite service; if the joint satisfaction probability is lower than a preset confidence threshold, an early warning signal is generated.

[0115] The service quality monitoring device for edge composite services provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned service quality monitoring method embodiment for edge composite services. For the sake of brevity, any parts not mentioned in the service quality monitoring device embodiment for edge composite services can be referred to the corresponding content in the aforementioned service quality monitoring method embodiment for edge composite services.

[0116] This invention also provides an electronic device, see [link to relevant documentation]. Figure 11 As shown, the electronic device includes a processor 130 and a memory 131. The memory 131 stores machine-executable instructions that can be executed by the processor 130. The processor 130 executes the machine-executable instructions to implement the above-described service quality monitoring method for edge composite services.

[0117] Furthermore, Figure 11 The electronic device shown also includes a bus 132 and a communication interface 133, with the processor 130, the communication interface 133 and the memory 131 connected via the bus 132.

[0118] The memory 131 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 133 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 132 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 11 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0119] Processor 130 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 130 or by instructions in software form. Processor 130 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 131, and processor 130 reads the information in memory 131 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0120] This invention also provides a machine-readable storage medium storing machine-executable instructions. When these machine-executable instructions are invoked and executed by a processor, they cause the processor to implement the aforementioned service quality monitoring method for edge composite services. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0121] The computer program product of the service quality monitoring method, device and electronic device for edge composite services provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0122] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for quality of service monitoring for edge composite services, characterized in that, The edge composite service includes multiple atomic services; the method includes: For each of the atomic services, collect the service quality time series data for that atomic service; The service quality time series data is preprocessed, and the processed data is output. The processed data is input into a pre-trained prediction model, so that the prediction model can output the service quality prediction distribution of the atomic service within the future prediction window. Obtain the call chain topology of the multiple atomic services; Based on the distribution operator corresponding to the call chain topology, the service quality prediction distribution corresponding to each atomic service is aggregated to output the overall service quality distribution of the edge composite service.

2. The method of claim 1, wherein, The steps for preprocessing the service quality time series data and outputting the processed data include: The service quality time series data is cleaned, and the cleaned data is output. The cleaned data is standardized using pre-saved standard data, and the processed data is output. The standard data includes the mean and standard deviation calculated based on historical training datasets.

3. The method according to claim 1, characterized in that, The prediction model includes an encoder and a decoder; the step of inputting the processed data into the pre-trained prediction model to output the service quality prediction distribution of the atomic service within a future prediction window includes: The encoder encodes the processed data and outputs a context vector. The context vector is decoded by the decoder to output the service quality prediction distribution of the atomic service within the future prediction window.

4. The method according to claim 3, characterized in that, The encoder includes a first network layer and a second network layer; the step of encoding the processed data using the encoder and outputting a context vector includes: The first network layer extracts features from the processed data in the positive time dimension and outputs a first hidden state vector. The second network layer extracts features from the processed data in the inverse time dimension and outputs a second hidden state vector. The first hidden state vector and the second hidden state vector are concatenated to output the first vector; The first vector is regularized, and the context vector is output.

5. The method according to claim 3, characterized in that, The decoder includes a multilayer perceptron; the step of decoding the context vector using the decoder to output the service quality prediction distribution of the atomic service within a future prediction window includes: The context vector is mapped by the multilayer perceptron to output dynamic kernel parameters corresponding to multiple time points within the future prediction window; wherein, the dynamic kernel parameters at each time point include: a preset number of dynamic kernel center values ​​and an adaptive bandwidth value; Based on the dynamic kernel center value and adaptive bandwidth value corresponding to each time step, a Gaussian mixture probability density function corresponding to that time step is constructed as the service quality prediction distribution of that atomic service within the future prediction window.

6. The method according to claim 1, characterized in that, The prediction model is trained using a composite loss function; the composite loss function is obtained by weighted summation of the negative log-likelihood loss term, the mean anchoring loss term, and the heteroscedasticity tail regularization term. The negative log-likelihood loss term is used to maximize the likelihood probability of the observed data under the service quality prediction distribution; The mean anchoring loss term is used to constrain the error between the mean of the service quality prediction distribution and the true label value; The heteroscedasticity tail regularization term is based on the heteroscedasticity regression principle and is used to construct a continuous penalty relationship between the residual and the predicted bandwidth value. When the true label value deviates from the mean of the service quality prediction distribution, causing the residual to increase and the predicted bandwidth value to decrease, a penalty value with non-linear amplification is generated to force the prediction model to actively increase the predicted bandwidth value in the specified area, thereby covering the potential tail fluctuation risk.

7. The method according to claim 1, characterized in that, The call chain topology includes: serial structure, parallel structure, and selection structure; The steps of aggregating the service quality prediction distributions of each atomic service according to the distribution operator corresponding to the call chain topology and outputting the overall service quality distribution of the edge composite service include: If the call chain topology is the serial structure, perform a convolution operation on the service quality prediction distribution corresponding to each atomic service, and output the overall service quality distribution of the edge composite service. If the call chain topology is the parallel structure, the cumulative distribution function corresponding to each atomic service is calculated based on the service quality prediction distribution corresponding to each atomic service; Multiply each of the cumulative distribution functions to obtain the product result; The product result is differentiated to obtain the probability density function of the edge composite service, and the overall service distribution of the edge composite service is determined based on the probability density function. If the call chain topology is the selection structure, obtain the weight parameters corresponding to each atomic service; The service quality prediction distribution corresponding to each atomic service is weighted according to the weight parameters to obtain the overall service quality distribution of the edge composite service.

8. The method according to claim 1, characterized in that, The overall service quality distribution includes: sub-probability density functions corresponding to multiple time points within the future prediction window; the method further includes: Based on a preset service level threshold, the sub-probability density function at each time point is integrated to obtain the single-step spatial satisfaction probability corresponding to each time point. The joint satisfaction probability corresponding to the edge composite service is obtained by performing a time-series multiplication on the single-step spatial satisfaction probability corresponding to each moment within the future prediction window. If the probability of the joint satisfaction is lower than a preset confidence threshold, an early warning signal is generated.

9. A service quality monitoring device for edge composite services, characterized in that, The edge composite service includes multiple atomic services; the device includes: The acquisition module is used to acquire the service quality time series data of each of the atomic services. The preprocessing module is used to preprocess the service quality time series data and output the processed data. The output module is used to input the processed data into a pre-trained prediction model, so as to output the service quality prediction distribution of the atomic service in the future prediction window through the prediction model. The acquisition module is used to acquire the call chain topology structure of the multiple atomic services; The aggregation processing module is used to aggregate the service quality prediction distribution corresponding to each atomic service according to the distribution operator corresponding to the call chain topology, and output the overall service quality distribution of the edge composite service.

10. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to implement the service quality monitoring method for edge composite services as described in any one of claims 1-8.

Images