Device pressure prediction method and apparatus, electronic device, and storage medium

Abstract

The application relates to the technical field of artificial intelligence, and provides a device pressure prediction method and device, electronic equipment and a storage medium, wherein the method comprises the following steps: acquiring multi-modal time series data of a target device in a historical period, wherein the multi-modal time series data comprises resource state data and business load data; performing space-time correlation analysis on the resource state data and the business load data to extract space-time correlation features representing the influence of business load on resource state; and performing device pressure prediction based on the space-time correlation features to generate a pressure prediction sequence of the target device in a future period. Through deep mining and analysis of the space-time correlation between business load and resource state, the application realizes more accurate and more forward-looking prediction of device pressure, and has the advantages of high prediction accuracy, strong robustness and effective response to business burst scenarios.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device, and storage medium for predicting equipment pressure. Background Technology

[0002] As data centers continue to expand, accurately predicting future equipment load is a key technology for achieving proactive resource scheduling and ensuring stable system operation. By anticipating changes in equipment load, such as the changing trends in CPU (Central Processing Unit) and memory usage, operations teams can adjust resource allocation in a timely manner, effectively avoiding service quality degradation or system outages caused by resource bottlenecks.

[0003] To predict equipment stress, a centralized monitoring and forecasting architecture is typically used. However, this centralized stress detection mechanism suffers from lag and cannot respond to changes in equipment stress in real time. Furthermore, forecast accuracy drops significantly when faced with sudden traffic surges caused by business anomalies. In addition, existing forecasting methods struggle to effectively capture the complex correlations in equipment stress changes. On one hand, they fail to model long-term time-series dependencies in equipment status; on the other hand, they cannot accurately reflect the differentiated impact of different services on hardware resource consumption, resulting in low forecast accuracy. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for predicting equipment pressure, in order to overcome the shortcomings of existing technologies in predicting equipment pressure due to lag, inaccuracy, and inability to cope with emergencies.

[0005] This invention provides a method for predicting equipment pressure, comprising: Acquire multimodal time-series data of the target device within a historical period, wherein the multimodal time-series data includes resource status data and service load data; Spatiotemporal correlation analysis is performed on the resource status data and the service load data to extract spatiotemporal correlation features that characterize the impact of service load on resource status; Based on the spatiotemporal correlation features, equipment pressure is predicted, and a pressure prediction sequence for the target equipment in future time periods is generated.

[0006] According to the equipment stress prediction method provided by the present invention, the steps of performing spatiotemporal correlation analysis on the resource status data and the service load data, and predicting equipment stress based on the spatiotemporal correlation features, are performed by a pre-trained stress prediction model. The stress prediction model includes an encoder and a decoder. The encoder is used to extract the spatiotemporal correlation features, and the decoder is used to generate the stress prediction sequence using an autoregressive mechanism.

[0007] According to a method for predicting equipment stress provided by the present invention, the step of performing spatiotemporal correlation analysis on the resource status data and the service load data to extract spatiotemporal correlation features characterizing the impact of service load on resource status includes: The encoder based on the stress prediction model uses a cross-modal attention mechanism to calculate the spatiotemporal correlation features between the resource status data and the business load data.

[0008] According to a method for predicting equipment stress provided by the present invention, the step of calculating the spatiotemporal correlation features between the resource status data and the service load data using a cross-modal attention mechanism includes: The resource status data is mapped to a query matrix, and the business load data is mapped to a key matrix and a value matrix; Based on the query matrix and the key matrix, calculate the attention weight matrix; The spatiotemporal correlation features are obtained by weighted summation of the value matrix based on the attention weight matrix.

[0009] According to a method for predicting equipment pressure provided by the present invention, the step of predicting equipment pressure based on the spatiotemporal correlation features and generating a pressure prediction sequence for the target equipment in future time periods includes: The spatiotemporal correlation features are input into the multi-channel output layer of the decoder; The multi-channel output layer performs parallel decoding of multiple indicators, generating pressure prediction values ​​of different dimensions at the same time step. The multi-indicator parallel decoding is performed iteratively according to time steps until the pressure prediction sequence is generated.

[0010] According to the device stress prediction method provided by the present invention, the stress prediction model is constructed based on the generator in the generative adversarial network, and the stress prediction model is trained through the following steps: Construct the generator and a discriminator for distinguishing between real sequences and generated sequences; Obtain a training sample set, which includes multimodal time-series data of the sample devices and the real pressure sequences corresponding to the sample devices; The training sample set is input into the generator to obtain the simulated pressure sequence output by the generator, and the simulated pressure sequence and the real pressure sequence are input into the discriminator to obtain the output result of the discriminator; Based on the real pressure sequence, the simulated pressure sequence, and the output of the discriminator, the generation loss and the discrimination loss are calculated. The parameters of the discriminator are updated based on the discriminant loss, and the parameters of the generator are updated based on the generation loss, with the parameter updates of the discriminator and the parameters of the generator performed alternately.

[0011] According to a method for predicting equipment pressure provided by the present invention, the step of inputting the simulated pressure sequence and the real pressure sequence into the discriminator includes: A time position code is added to both the input simulated pressure sequence and the actual pressure sequence. The time position code is used to identify the temporal continuity of data points in the sequence. The discriminator extracts the temporal features of the input sequence and outputs a probability value representing that the input sequence is a real sequence. The input sequence is the simulated pressure sequence and the real pressure sequence after adding the time position encoding.

[0012] According to the equipment stress prediction method provided by the present invention, the training process further includes a negative sample augmentation step, which includes: True negative samples representing abnormal stress states are selected from the training sample set; The real negative samples are oversampled, and the simulated abnormal stress sequence generated by the generator during adversarial training is used as the synthetic negative samples. Based on the oversampled real negative samples and the synthetic negative samples, a mixed negative sample set is constructed, which is used to train the discriminator.

[0013] According to the equipment stress prediction method provided by the present invention, the generation loss is composed of a weighted sum of a prediction loss and an adversarial loss. The prediction loss characterizes the difference between the simulated stress sequence and the real stress sequence, and the adversarial loss characterizes the probability that the simulated stress sequence is identified as a real sequence by the discriminator. The weight coefficients of the prediction loss and the adversarial loss are dynamically adjusted according to the training phase.

[0014] According to the device stress prediction method provided by the present invention, the resource status data includes at least one of processor utilization, memory occupancy, disk read / write speed and network transmission speed; The business load data includes at least one of the following: business type identifier, number of concurrent requests, and request response latency.

[0015] The present invention also provides a device for predicting equipment pressure, comprising: The data acquisition unit is used to acquire multimodal time-series data of the target device within a historical period, wherein the multimodal time-series data includes resource status data and service load data; The feature extraction unit is used to perform spatiotemporal correlation analysis on the resource status data and the service load data to extract spatiotemporal correlation features that characterize the impact of service load on resource status. The pressure prediction unit is used to predict equipment pressure based on the spatiotemporal correlation features and generate a pressure prediction sequence for the target equipment in future time periods.

[0016] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the device pressure prediction method as described above.

[0017] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the device pressure prediction method as described above.

[0018] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the device pressure prediction method as described above.

[0019] The equipment stress prediction method, apparatus, electronic device, and storage medium provided by this invention fundamentally solve the problem of predictive bias caused by the single input information in traditional technologies by acquiring multimodal time-series data including resource status data and service load data. This invention incorporates service load data into the analysis scope, providing richer contextual information for equipment stress prediction. This enables the model to understand and learn the intrinsic relationship between service load and resource consumption, thus making a more accurate response to stress fluctuations caused by specific services. Secondly, this invention performs spatiotemporal correlation analysis on resource status and service load data, extracting spatiotemporal correlation features that characterize the impact of service load on resource status. This achieves a leap from simple numerical prediction to deep causal relationship modeling. This prediction method based on correlation features allows the model not only to predict stress changes but also to uncover the service loads that cause resource stress changes, greatly enhancing the robustness and interpretability of the prediction. Finally, by using this deeply refined spatiotemporal correlation feature for equipment stress prediction, not only is high accuracy and reliability of the prediction results ensured, but accuracy is also improved in complex scenarios, long-term predictions, and responses to emergencies. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the equipment pressure prediction method provided by the present invention; Figure 2 This is a schematic diagram of the training process of the stress prediction model provided by the present invention; Figure 3 This is a schematic diagram of the deeply coupled architecture of Transformer and Generative Adversarial Network (GAN) provided by the present invention; Figure 4 This is a schematic diagram of the equipment pressure prediction device provided by the present invention; Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0023] With the deepening of digital transformation, the scale and complexity of data centers are growing at an unprecedented rate. In large data centers, the pressure (or load) of hardware resources such as servers and network equipment is constantly changing. Accurately predicting future equipment pressure is a key technology for achieving proactive resource scheduling, ensuring stable system operation, preventing service interruptions, and optimizing energy efficiency. If a surge in equipment pressure can be anticipated in advance, the operations and maintenance team can scale up or migrate services in a timely manner, avoiding service quality degradation or system crashes caused by resource bottlenecks.

[0024] To predict equipment stress, current methods typically employ time-series analysis, such as Long Short-Term Memory (LSTM) networks and Autoregressive Integrated Moving Average (ARIMA) models, or centralized monitoring and forecasting architectures. While these methods can predict future equipment stress trends to some extent, their limitations are becoming increasingly apparent in modern, large-scale, and highly dynamic data center scenarios. Specifically: First, traditional centralized control methods inherently suffer from lag in their pressure detection and response mechanisms. When equipment pressure changes, the system needs time to collect, transmit, process data, and make predictions, resulting in an inability to respond to pressure changes in real time. Furthermore, when faced with sudden surges in traffic due to weather changes, business disruptions, or other unforeseen circumstances, these models, failing to consider such external factors, often produce predictions that significantly deviate from reality, greatly reducing their accuracy.

[0025] Second, equipment stress is not an isolated change, but the result of multiple factors working together. Existing technologies have significant shortcomings in capturing these complex correlations. First, there is insufficient capture of temporal dependencies. The stress state of equipment at a future point in time is often influenced by its state sequence over a relatively long period, and traditional methods (such as LSTM relying on sliding windows) struggle to effectively capture this long-term dependency. Second, there is insufficient modeling of business-related dependencies. Different types of business (such as compute-intensive and I / O-intensive) have different consumption patterns of equipment resources (such as CPU, memory, and disk), and traditional methods struggle to accurately quantify and reflect the differentiated impact of specific business types on equipment stress.

[0026] Third, device stress is a multi-dimensional and comprehensive state, including multiple key indicators such as CPU utilization, memory usage, disk I / O, and network bandwidth. Existing methods can usually only predict a single indicator independently, and cannot predict multiple device stress indicators simultaneously. This results in an inability to fully reflect the device's stress status, making it difficult to achieve optimal resource scheduling strategies. For example, predicting only CPU utilization while ignoring memory utilization may lead to insufficient memory and affect system performance.

[0027] Fourth, in real-world operating environments, equipment is in a normal or low-load state most of the time (i.e., positive samples), while abnormal situations with high load and high pressure (i.e., negative samples) occur very infrequently. This imbalance between positive and negative samples causes the model to tend to learn normal states during training, while failing to adequately learn from the few abnormal samples, resulting in severely insufficient prediction ability for small samples. The direct consequence is that the model's prediction accuracy and recall for truly stressful abnormal events requiring early warning (such as equipment overload or precursors to hardware failure) are extremely low, rendering it ineffective in providing early warning.

[0028] Fifth, models trained using existing methods are often highly dependent on specific training scenarios or device types. When business scenarios change, or when the model is applied to new data centers or new device models, its predictive performance will significantly decline, indicating poor model generalization ability. This makes the deployment and maintenance of the model costly, as it often requires tedious data collection, model training, and parameter tuning for the new environment.

[0029] To address this issue, this invention provides a generative method for predicting equipment stress. By jointly analyzing the equipment's own resource status data and its service load data, and deeply exploring the spatiotemporal correlation between the two, it can predict the changing trends of equipment stress from the source of service changes. This allows the prediction model to not only learn the evolutionary patterns of resource usage, but more importantly, to understand that service load is the driving force behind changes in resource status. This enables faster and more accurate stress predictions in complex scenarios such as sudden increases in service volume and changes in service type, thus overcoming the aforementioned shortcomings. This method can be widely applied in scenarios such as data centers and cloud computing platforms, providing strong support for equipment stress management.

[0030] Figure 1 This is a flowchart illustrating the equipment pressure prediction method provided by the present invention, as shown below. Figure 1 As shown, the method includes: Step 110: Obtain multimodal time-series data of the target device within a historical period. The multimodal time-series data includes resource status data and service load data.

[0031] Specifically, the implementing entity of the method in this embodiment of the invention (such as a device stress prediction apparatus or system) periodically collects or receives multimodal time-series data about the target device over historical periods from the monitoring system. Here, the target device refers to the device that needs stress prediction, which can be any computing unit running in a data center, cloud computing platform, or other computing environment. In a broad sense, it can be a physical server, virtual machine, container, etc. In a narrow sense, it can specifically refer to a server that hosts a specific application or service.

[0032] Historical time periods refer to the time intervals for data collection, which are periods prior to the prediction start point. For example, a fixed-length sliding time window can be set, such as the past 60 minutes. The length of this time window (also called the time step) is configurable and can be adjusted based on the periodicity of the business or empirical values. For example, for frequently changing businesses, a shorter time window can be used to capture real-time changes.

[0033] Multimodal time series data refers to data in a time series where each point in time contains multiple types of data from different sources or with different properties. In this embodiment of the invention, multimodal time series data can be data that combines two modalities—the device's own operating status and the business load it carries—by aligning them in time. This combination enables the model to correlate business load and resource consumption.

[0034] Specifically, multimodal time-series data can include resource status data and workload data. Resource status data reflects the consumption of physical or virtual resources of a device. For example, resource status data can include at least one of the following: processor (e.g., CPU) utilization, memory (RAM) occupancy, disk read / write speed (e.g., disk I / O), and network transmission rate (e.g., network bandwidth or traffic). This data can typically be obtained directly through the monitoring interfaces of the operating system or the underlying virtualization platform.

[0035] Load metric data reflects the load generated by applications or services running on a target device. For example, load metric data may include at least one of the following: service type identifier (e.g., one-hot encoding indicating whether it is compute-intensive, I / O-intensive, or network-intensive), number of concurrent requests, and request-response latency. This data can typically be obtained from application-layer monitoring logs or middleware.

[0036] For example, at a point in time The acquired multimodal time series data can be a concatenated vector. ,in Indicates a point in time The acquired resource status data (or hardware metrics) includes dimensions such as CPU utilization, memory usage, disk I / O, and network card traffic. This represents the business load data (or business metrics) obtained at time point t, which includes dimensions such as business type, concurrent requests, and average request latency. It is obtained by continuously acquiring historical time periods (e.g., past...). The data within a given time point constitute a multimodal time series data sequence matrix, which can be represented as: In the above formula, Indicates at a point in time A multimodal time-series data sequence matrix, which contains data from the past... Time to the present A series of data at any given moment. This indicates the dimension of the sequence matrix; it has... Rows (representing historical time steps) and Columns (representing the number of features at each time point, i.e., the total feature dimension). Among them... , Dimensions representing resource status data This represents the dimension of the business load data.

[0037] Step 120: Perform spatiotemporal correlation analysis on the resource status data and the service load data to extract spatiotemporal correlation features that characterize the impact of service load on resource status.

[0038] Specifically, after acquiring the raw multimodal time-series data, directly using the raw data for prediction makes it difficult for the model to effectively learn the complex dependencies between different modalities and different time points, resulting in reduced prediction accuracy. To address this, this embodiment of the invention first performs spatiotemporal correlation analysis on resource status data and service load data to extract spatiotemporal correlation features characterizing the impact of service load on resource status. Then, prediction is performed based on these spatiotemporal correlation features, enabling the model to learn the complex dependencies between different modalities and different time points, thereby improving prediction accuracy.

[0039] Here, spatiotemporal correlation analysis includes spatial correlation analysis and temporal correlation analysis. Spatial correlation refers to the correlation between different modalities of data at the same point in time, i.e., the correspondence between business load data (such as an increase in concurrent requests) and resource status data (such as an increase in CPU utilization). Temporal correlation refers to the evolution and dependence of this correlation over time. For example, the CPU utilization at the current moment is not only related to the business load at the current moment, but may also be affected by the cumulative effect of business load at previous time points. Therefore, spatiotemporal correlation analysis aims to capture the dependence of both dimensions simultaneously.

[0040] By performing spatiotemporal correlation analysis on resource status data and business load data, the spatiotemporal correlation characteristics between the two can be obtained. Here, the spatiotemporal correlation characteristics are not simple raw data, but a feature representation obtained through calculation and abstraction that can highly condense and characterize how business load affects resource status over time. This feature can be a vector, a matrix, or a tensor. It contains the patterns learned by the model. For example, the model may learn a complex pattern such as "a video transcoding service (service type) will cause the CPU utilization (resource status) to remain above 90% for the next 5 seconds after the concurrency exceeds 1000 (number of concurrent requests)."

[0041] Understandably, spatiotemporal correlation analysis can be performed by one or more model modules, the input of which is the multimodal time series data sequence matrix obtained in step 110. The output is the spatiotemporal correlation characteristics. This analysis process can automatically discover and quantify the differentiated impacts of different business types on different hardware resources, as well as the patterns of these impacts over time.

[0042] Step 130: Based on the spatiotemporal correlation features, perform equipment pressure prediction to generate a pressure prediction sequence for the target equipment in future time periods.

[0043] Specifically, after extracting informative spatiotemporal correlation features, these features can be used to generate predictions of future equipment stress. Here, the equipment stress prediction process can be understood as a feature-based sequence generation task. It receives spatiotemporal correlation features representing historical patterns as input and infers the trend of resource status changes for the target equipment over a future period.

[0044] For example, the extracted spatiotemporal correlation features can be input into a prediction module. Based on the causal relationship encoded in the features, the module gradually generates the pressure value for the first future time point. Then, the predicted value is used as a new input to continue generating the pressure value for the second time point. This process is iterated until a pressure prediction sequence for the entire future period is generated.

[0045] Here, the pressure prediction sequence includes the target device at one or more future time points (such as prediction time windows). The predicted values ​​of one or more resource status indicators (i.e., hardware stress indicators to be predicted) on the [database name]. This stress prediction sequence can be represented as: In the above formula, Indicates the length of the prediction time window. This represents the generated pressure prediction sequence, which contains data from... Time's up Real-time forecasts of equipment pressure indicators. This indicates the dimension of the sequence matrix; it has... Rows (representing the time step of the forecast) and The column (representing the number of stress metrics to be predicted) should be understood to be either single-dimensional (e.g., predicting only CPU utilization) or multi-dimensional (e.g., predicting CPU, memory, network traffic, etc. simultaneously).

[0046] The method provided in this invention, by acquiring multimodal time-series data including resource status data and service load data, fundamentally solves the problem of predictive bias caused by the single input information in traditional technologies. This invention incorporates service load data into the analysis scope, providing richer contextual information for equipment stress prediction. This enables the model to understand and learn the intrinsic relationship between service load and resource consumption, thus allowing for a more accurate response to stress fluctuations caused by specific services. Secondly, by performing spatiotemporal correlation analysis on resource status and service load data, this invention extracts spatiotemporal correlation features that characterize the impact of service load on resource status, achieving a leap from simple numerical prediction to deep causal relationship modeling. This prediction method based on correlation features allows the model not only to predict stress changes but also to uncover the service loads that cause resource stress changes, greatly enhancing the robustness and interpretability of the prediction. Finally, by using these deeply refined spatiotemporal correlation features for equipment stress prediction, not only is high accuracy and reliability of the prediction results ensured, but accuracy is also improved in complex scenarios, long-term predictions, and responses to emergencies.

[0047] Based on the above embodiments, the resource status data includes at least one of processor utilization, memory occupancy, disk read / write speed, and network transmission speed; The business load data includes at least one of the following: business type identifier, number of concurrent requests, and request response latency.

[0048] Specifically, multimodal time-series data can include resource status data and workload data. Resource status data reflects the physical consumption status of device hardware or virtual resources, and can include one or more metrics such as processor utilization, memory occupancy, disk read / write speed, and network transmission rate. Here, processor utilization (i.e., CPU utilization) measures the CPU's activity level within a specific time period and is a core indicator of computationally intensive workloads. Memory occupancy reflects the usage of physical or virtual memory; high occupancy can lead to performance degradation or even system crashes. Disk read / write speed (e.g., disk I / O) measures the disk's input / output activity and is crucial for I / O-intensive workloads such as databases and file services. Network transmission rate (e.g., network interface card traffic) measures the data throughput of the network interface and is a key indicator of network-intensive workloads.

[0049] Load data reflects the workload of applications or services running on the device. It can include one or more metrics such as service type identifier, concurrent request count, and request response latency. Here, the service type identifier distinguishes different types of services; for example, it can be a one-hot encoded vector to identify compute-intensive or I / O-intensive services. This is fundamental for cross-modal analysis and understanding the differentiated impact of different services on resources. The number of concurrent requests refers to the number of requests processed simultaneously per unit of time, directly reflecting the load intensity of the service. Request response latency refers to the average time required to complete one service request; it is both a result of the load and an early warning signal that the system is about to reach a bottleneck.

[0050] The aforementioned metrics are standard data collection items for current data center and cloud computing monitoring systems. This means that the method provided in this invention does not require complex customized data collection modifications and can seamlessly integrate with existing operation and maintenance systems, significantly lowering the application threshold. Furthermore, these metric combinations cover multiple dimensions of resource consumption, including computing, storage, I / O, and network, as well as various aspects of business load. This provides a comprehensive and rich data foundation for the model to perform end-to-end causal inference from the source (business load) to the result (resource consumption). This is the fundamental guarantee that this invention can achieve high-precision and highly robust predictions.

[0051] Based on any of the above embodiments, the steps of performing spatiotemporal correlation analysis on the resource status data and the service load data (i.e., step 120) and predicting equipment pressure based on the spatiotemporal correlation features (i.e., step 130) are performed by a pre-trained pressure prediction model. The pressure prediction model includes an encoder and a decoder. The encoder is used to extract the spatiotemporal correlation features, and the decoder is used to generate the pressure prediction sequence using an autoregressive mechanism.

[0052] Specifically, a pre-trained stress prediction model refers to a model that has been trained offline using a large amount of historical data as a training sample set. Once trained, the model's internal parameters are fixed, enabling it to represent the general patterns inherent in the data. During the prediction phase, the model can be directly loaded and used for inference, thus ensuring real-time prediction.

[0053] This stress prediction model can employ a Transformer encoder-decoder architecture, which is well-suited for tasks where the input is a sequence and the output is also a sequence. This architecture includes an encoder for extracting spatiotemporal correlation features and a decoder for generating the prediction sequence.

[0054] The encoder receives multimodal time-series data from historical periods. This is taken as input. Through its complex internal network structure, the encoder performs deep processing and information compression on this time series containing both resource and business modalities, ultimately encoding it into a (or a set of) fixed-dimensional context vectors. This context vector is the spatiotemporal correlation feature (or spatiotemporal correlation latent variable, denoted as...). It encapsulates all the important information in historical data related to future stress prediction, including the immediate correlations between different data modalities and the evolution of these correlations over time. The encoder's processing can be represented as follows: The decoder receives spatiotemporal correlation features generated by the encoder as initial input, and its task is to generate a stress prediction sequence for future periods based on this historical summary. Specifically, it is expressed as follows: in, This represents a learnable query vector used to decode future time series.

[0055] Understandably, in order to generate a temporally continuous and reasonable prediction sequence, the decoder employs an autoregressive mechanism. Specifically, in generating the first future time point... Predicted value At that time, the decoder primarily relies on the spatiotemporal correlation features output by the encoder. Then, at the second time point... Predicted value At that time, the decoder will not only utilize the spatiotemporal correlation features of the encoder, but also the newly generated... As additional input to itself. This process iterates continuously, generating the first... Predicted values ​​at each time point At that time, all previously generated prediction values ​​will be used. For reference. Repeat this process until the entire forecast time window is generated. All sequences within.

[0056] This invention employs an encoder-decoder architecture. The encoder extracts the spatiotemporal correlation features between resource status and service load, while the decoder generates a future stress prediction sequence through autoregression. This addresses the problem of long-term time-series dependencies. Simultaneously, the introduction of an autoregressive mechanism ensures the temporal continuity and logical consistency of the generated stress prediction sequence, avoiding the generation of isolated or disjointed prediction points. This makes the prediction results closer to the real-world patterns of equipment stress changes, thereby improving the overall quality and reliability of the prediction sequence.

[0057] Based on any of the above embodiments, step 120 includes: Step 121: Based on the encoder of the pressure prediction model, a cross-modal attention mechanism is used to calculate the spatiotemporal correlation features between the resource status data and the business load data.

[0058] It's important to note that cross-modal attention mechanisms are computational mechanisms that simulate human cognitive attention. When humans analyze a complex scenario, they unconsciously focus their attention on key information. Similarly, this mechanism allows models to dynamically and selectively focus on the parts of one modality (such as workload) that are most relevant to another modality (such as resource status) when processing multimodal data. It doesn't simply concatenate the two sets of data; instead, it calculates the correlation or attention weight between them and fuses information with a specific emphasis based on this weight.

[0059] Specifically, the cross-modal attention mechanism used internally by the encoder works as follows: When the encoder processes historical time-series data sequences... Data at a specific point in time At that time, it will no longer transmit resource status data. and business load data They are not considered as equally important components. Instead, the resource status data is used as a query to find which components of the business load data are causing the current resource status.

[0060] Specifically, the model calculates an attention score between each dimension of resource status data (such as CPU utilization) and each dimension of business load data (such as concurrent requests and business type). The higher the score, the stronger the correlation between the two. For example, through learning, the model might discover that a sudden surge in concurrent requests has a very high attention score with a spike in CPU utilization, while the correlation with disk read rate is relatively low.

[0061] Through this mechanism, the encoder reweights the workload data at each time point, generating a workload feature representation that reflects resource status. This new feature representation highlights the workload factors that have the greatest impact on current resource consumption. By fusing this weighted feature with the original resource status features, the encoder obtains a feature representation that better reflects causal relationships and is richer in information; this representation is the final output spatiotemporal correlation feature.

[0062] This invention employs a cross-modal attention mechanism, enabling the model to explicitly and dynamically model the cross-modal correlation between resource states (such as CPU and memory) and business loads (such as concurrency and business type). This enhances the model's understanding and representation of complex scenarios (such as weather changes, sudden business disruptions, and other emergencies), resulting in higher quality and more interpretable spatiotemporal correlation features. This lays a solid foundation for the subsequent decoder to generate more accurate prediction sequences, significantly improving the robustness and accuracy of predictions.

[0063] Based on any of the above embodiments, step 121 specifically includes: Step 1211: Map the resource status data into a query matrix, and map the business load data into a key matrix and a value matrix; Step 1212: Calculate the attention weight matrix based on the query matrix and the key matrix; Step 1213: Based on the attention weight matrix, perform a weighted summation of the value matrix to obtain the spatiotemporal correlation features.

[0064] Specifically, when performing cross-modal attention mechanism calculations, the model first needs to assign different roles to data from different modalities in order to perform subsequent correlation calculations. For example, resource status data... Mapped to a query matrix (denoted as ) ), to transfer business load data They are respectively mapped to the key matrix (denoted as ). ) and value matrix (denoted as The mapping here is typically achieved through a learnable linear transformation (i.e., a fully connected layer), projecting the original data into a representation space more suitable for attention computation. The specific attention computation process usually employs scaled dot product attention, which can be represented by the following formula: Specifically, firstly, the query matrix AND key matrix transpose (i.e.) Perform matrix multiplication to obtain an original attention score matrix. Each element in this score matrix reflects the correlation strength between a resource state dimension and a business load dimension. To prevent the gradient from becoming too small due to an excessively large dot product, which could affect the stability of model training, this score matrix needs to be scaled, typically by dividing it by the key matrix dimension. The square root of.

[0065] Subsequently, the scaled score matrix is ​​transformed into an attention weight matrix by applying the softmax function along a specific dimension (usually the key dimension). The elements of this weight matrix are between 0 and 1, and the sum of each row (or column) is 1. This can be intuitively understood as the probability of attention allocation. A larger weight value indicates that the corresponding business load information is more important to the current resource status query.

[0066] After obtaining the attention weights, the final step is to use these weights to extract key information from the value matrix. Specifically, this involves combining the attention weight matrix calculated in the previous step with the value matrix... Perform matrix multiplication. This operation is essentially a weighted summation process. The output of this weighted summation is a new feature representation that integrates business load information and is filtered and focused by resource status data. This representation accurately captures the business load features that significantly impact resource status at the current moment; it is the spatiotemporal correlation feature to be extracted, or its core component.

[0067] Based on any of the above embodiments, traditional prediction methods typically train a separate model for each prediction indicator, which is inefficient and ignores the inherent correlation between indicators. This invention aims to address the challenge of simultaneously predicting multiple equipment stress indicators. Step 130 specifically includes: Step 131: Input the spatiotemporal correlation features into the multi-channel output layer of the decoder; Step 132: Perform parallel decoding of multiple indicators through the multi-channel output layer to generate pressure prediction values ​​of different dimensions at the same time step; Step 133: Iterate through the multi-indicator parallel decoding according to the time step until the pressure prediction sequence is generated.

[0068] Specifically, the final output stage of the decoder is designed as a multi-channel output layer. After the decoder completes the decoding operations on the spatiotemporal correlation features and the sequences generated in previous time steps, its final hidden state is passed to this special output layer. Unlike a traditional single-output layer, this output layer contains multiple parallel, independent channels or headers.

[0069] Here, the multi-channel output layer can be understood as a set of parallel processing units (e.g., multiple independent fully connected layers), each responsible for predicting a specific device stress metric. For example, channel one predicts CPU utilization, channel two predicts memory usage, and channel three predicts network traffic. All channels share the deep feature representation from the decoder body (i.e., the decoded hidden state), but each has independent parameters used to map the shared features to its own metric space.

[0070] When the hidden state of the decoder is input to the multi-channel output layer, each channel performs calculations in parallel. Parallel decoding of multiple metrics means that for a certain time step in the future (e.g., ...), ... The model does not predict each indicator sequentially one by one, but instead calculates the predicted values ​​of all target pressure indicators at a given time step simultaneously during a single forward propagation. For example, the CPU prediction channel output... CPU prediction value at any given time, memory prediction channel output The memory prediction values ​​at each time step are completed simultaneously. The resulting stress prediction values ​​of different dimensions together constitute a multidimensional stress vector for that time step.

[0071] Finally, the multi-indicator parallel decoding steps are executed iteratively according to time steps until the final complete stress prediction sequence is generated. Specifically, in the first time step (i.e. The decoder utilizes the spatiotemporal correlation features from the encoder and generates the first multidimensional stress prediction value through parallel decoding of multiple indicators. (For example, including multiple metrics such as CPU and memory). In the second time step (i.e. The decoder will As an additional input, a second multidimensional stress prediction value is generated again through parallel decoding of multiple indicators. This process is repeated until a timeframe covering the entire preset future period is generated. Complete pressure prediction sequence .

[0072] This invention, through the design of a multi-channel output layer and the implementation of parallel decoding of multiple indicators, enables efficient joint prediction of multiple equipment stress indicators within a unified model. This allows the model to learn and predict a specific indicator while utilizing information from other relevant indicators as corroborating evidence, thereby capturing the synergistic change patterns among various stress indicators. This not only significantly improves prediction efficiency but also, because the model learns a more comprehensive understanding of equipment operating modes, significantly enhances the accuracy and consistency of its predictions, effectively solving the problem of insufficient multi-indicator prediction capabilities in traditional technologies.

[0073] It should be noted that traditional model training methods usually use minimizing the error between the predicted value and the true value (such as mean squared error, MSE) as the objective. Although this method can ensure numerical closeness, it may not be able to fully capture the complexity and dynamics of the real data distribution. Especially when abnormal stress events (i.e. negative samples) are sparse, the model is prone to producing overly smooth or average predictions, and its ability to predict mutation points is insufficient.

[0074] To address this issue, this invention constructs the training process of the stress prediction model within a Generative Adversarial Network (GAN) framework. Within this framework, the stress prediction model acts as a generator, aiming to generate stress sequences that are indistinguishable from realistic ones. Simultaneously, a discriminator is introduced, aiming to distinguish between real stress sequences and simulated stress sequences generated by the generator as accurately as possible. Through mutual competition and co-evolution, the two ultimately enable the generator to produce highly realistic and accurate stress prediction sequences.

[0075] Based on any of the above embodiments Figure 2 This is a schematic diagram of the training process of the pressure prediction model provided by the present invention, as shown below. Figure 2 As shown, the stress prediction model is built based on the generator in a generative adversarial network, and the stress prediction model is trained through the following steps: Step 210: Construct the generator and the discriminator for distinguishing between the real sequence and the generated sequence.

[0076] Specifically, this step forms the foundation for building the entire GAN training framework. The generator (G) is the stress prediction model, and its structure is an encoder-decoder architecture (e.g., an Encoder-Decoder based on Transformer). Its input is multimodal time-series data from historical periods (…). The output is a simulated pressure sequence (also known as a pressure prediction sequence) that has the same dimensions and length as the real pressure sequence. ).

[0077] The discriminator (D) is a newly added structure, essentially a sequence classifier. Its task is to take a stressed sequence as input and output a probability value representing the likelihood that the input sequence is real rather than generated. To effectively distinguish between real and fake time-series data, the discriminator needs to capture the temporal continuity and inherent patterns of the sequence. Therefore, it can be constructed using a Transformer Encoder-based structure to effectively extract the temporal features of the sequence.

[0078] Step 220: Obtain a training sample set, which includes multimodal time-series data of the sample device and the real pressure sequence corresponding to the sample device.

[0079] Specifically, this step prepares data for model training. Monitoring data from a large number of sample devices in real-world operating environments, such as data centers and cloud computing platforms, is collected over different time periods. Each sample in the resulting training sample set contains two parts: one part is the historical time period. Multimodal time series data within ( The other part is the corresponding part in the future time period. The actual stress sequence that actually occurs inside ( ).

[0080] Step 230: Input the training sample set into the generator to obtain the simulated pressure sequence output by the generator, and input the simulated pressure sequence and the real pressure sequence into the discriminator to obtain the output result of the discriminator.

[0081] Specifically, this step is the forward propagation process in a single iteration of GAN. First, a batch of historical multimodal time-series data is retrieved from the training sample set ( ).Will The input is fed into generator G, which, based on its internal encoder-decoder logic, outputs a batch of simulated pressure sequences. .

[0082] Real pressure sequence and simulated pressure sequence The data are input simultaneously (or sequentially) into the discriminator D. The discriminator D evaluates each data point and outputs two sets of probability values, namely... and Ideally, It should approach 1, while It should approach 0.

[0083] Step 240: Calculate the generation loss and the discrimination loss based on the real pressure sequence, the simulated pressure sequence, and the output of the discriminator.

[0084] Specifically, this step calculates the loss based on the results of the forward propagation, providing a basis for updating the parameters during backpropagation. The loss calculated here includes the generation loss and the discrimination loss.

[0085] Generation loss (denoted as) The loss function is designed to measure the generator's performance. The generation loss is a weighted sum of the prediction loss and the adversarial loss. The prediction loss represents the difference between the simulated stress sequence and the real stress sequence, while the adversarial loss represents the probability that the discriminator will identify the simulated stress sequence as a real sequence. The weights of the prediction loss and the adversarial loss can be dynamically adjusted during the training phase.

[0086] Understandably, the generator's goal is to deceive the discriminator, that is, to make the discriminator give a high score to the simulated sequence it generates (i.e., Approaching 1). To achieve this, the loss includes an adversarial loss, which is designed to be directly related to the discriminator's output. By minimizing this loss, the generator is forced to learn the intrinsic patterns of the real data distribution to generate sequences sufficiently realistic to be indistinguishable from the real ones. Furthermore, to ensure prediction accuracy, the generation loss typically also includes a prediction loss, the purpose of which is to make the simulated stress sequence generated by the generator numerically approximate the real stress sequence as closely as possible. This loss can be calculated using the mean squared error (MSE). The final generation loss is a weighted sum of these two losses, calculated using the following formula: in, This represents the prediction loss calculated using the mean squared error (MSE), and its calculation formula is as follows: The formula for calculating the adversarial loss is as follows: .

[0087] and These represent the weighting coefficients for the prediction loss and adversarial loss, respectively, which can be dynamically adjusted based on the training phase. This is an adaptive training strategy designed to optimize the stability and efficiency of the entire training process.

[0088] A specific dynamic adjustment strategy can be designed as follows: At the beginning of training, the simulated sequences generated by the generator are of poor quality and differ greatly from the real sequences. At this time, if the weights of the adversarial loss are too large, the generator may experience gradient vanishing or training instability because it is difficult to fool the discriminator. Therefore, in the initial stage, the focus should be on predictive loss, i.e., setting a relatively large weight. and smaller This allows the generator to quickly learn the basic shape and numerical range of the stress sequence, and generate a similar sequence as soon as possible. When the training progresses to a certain stage, and the generator can generate simulated sequences that are numerically similar to the real sequence, the weight of the adversarial loss should be gradually increased, i.e., the weight should be decreased. And increase This allows the generated sequences to more closely resemble real data in terms of statistical properties. By dynamically adjusting the loss weights, the model is effectively prevented from getting stuck in local optima, and a balance is struck between prediction accuracy and generation diversity. This ensures that the final trained stress prediction model can both guarantee the accuracy of the predicted values ​​and generate more robust prediction results that conform to real dynamic patterns.

[0089] Determine the loss (denoted as) This loss function is designed to measure the performance of the discriminator. The discriminator aims to give high scores to real sequences and low scores to simulated sequences. Therefore, its loss consists of two parts: one part is the loss for identifying real sequences as fake, and the other part is the loss for identifying simulated sequences as real. Its calculation formula can be expressed as follows: Step 250: Update the parameters of the discriminator based on the discriminant loss, and update the parameters of the generator based on the generation loss. The parameter updates of the discriminator and the generator are performed alternately.

[0090] Specifically, this step is the core of GAN training, namely, the embodiment of the adversarial process. First, the generator parameters can be fixed, and only the discriminant loss can be used. The gradient is calculated using the backpropagation algorithm, and the internal parameters of the discriminator D are updated using an optimizer (such as Adam). The purpose of this step is to improve the discriminator's ability to detect falsehoods. Subsequently, the discriminator's parameters are fixed, and the generation loss is adjusted accordingly. The gradient is calculated using the backpropagation algorithm, and the internal parameters of the generator G (i.e., the stress prediction model) are updated. The purpose of this step is to improve the realism and accuracy of the sequences generated by the generator.

[0091] These two update processes alternate within a single training iteration. Through thousands of iterative games, the discriminator's ability to distinguish becomes stronger, which in turn forces the generator to produce increasingly realistic and high-quality stress sequences. Ultimately, the training reaches a dynamic equilibrium, at which point the generator becomes the desired, well-trained stress prediction model.

[0092] Figure 3 This is a schematic diagram of the deeply coupled architecture of Transformer and Generative Adversarial Network (GAN) provided by this invention, as shown below. Figure 3 As shown, this architecture is a typical Generative Adversarial Network (GAN) framework, which mainly consists of two core parts: a Transformer model as a generator and a Transformer model as a discriminator. The generator is responsible for generating future stress prediction sequences based on historical data, while the discriminator is responsible for distinguishing between the predicted sequences generated by the generator and the actual stress sequences.

[0093] Specifically, the initial input to the model is historical multimodal time series data. It is then fed into the encoder section of the generator. The generator processes... Then, the simulated future pressure sequence is output. The discriminator receives two inputs: one is the simulated pressure sequence generated by the generator. The other is the real future stress sequence. The discriminator evaluates the two sequences and assigns a probability that each is real. The discriminator's output (loss) is backpropagated to guide parameter updates for both the generator and the discriminator. The generator aims to produce sequences that the discriminator cannot distinguish between real and fake sequences, while the discriminator aims to identify fake sequences as accurately as possible. They co-evolve in this adversarial process.

[0094] like Figure 3 As shown, the generator adopts a Transformer Encoder-Decoder architecture, which is well-suited for handling sequence-to-sequence prediction tasks. The encoder consists of multiple stacked encoder layers, and its main function is to receive historical time-series data and perform deep encoding to extract spatiotemporal correlation features characterizing the impact of business load on resource status. Figure 3 The enlarged left section shows a detailed view of the internal structure of a single encoder layer. It contains a multi-head attention module and a feed-forward module, each followed by residual connections and layer normalization (add & normal). The core of the encoder lies in the application of its multi-head attention mechanism. Figure 3 The text clearly indicates that... This represents a cross-modal attention mechanism. Specifically, it uses resource state data (i.e., hardware metrics). Generate query matrix At the same time, use business load data (i.e., business metrics) Generate the key matrix Sum matrix In this way, the model can explicitly calculate the weights of the impact of various aspects of the workload on the current resource state, thereby accurately capturing the dynamic relationship between the two.

[0095] The decoder also consists of multiple stacked decoder layers. Its main function is to receive the spatiotemporal correlation features from the encoder output and generate future stress prediction sequences step by step in an autoregressive manner. When generating the prediction value at the current time step, the decoder simultaneously references the encoder output and its own output at the previous time step.

[0096] The discriminator employs an architecture consisting of multiple stacked Transformer Encoder layers. It uses a self-attention mechanism to capture long-range temporal dependencies and complex dynamic patterns in input stress sequences (whether real or simulated). The discriminator's task is to output a probability value that determines the authenticity of the input sequence.

[0097] This invention elevates model training from simple error fitting to data distribution learning by introducing a generative adversarial training framework. Under the supervision of the discriminator, the generator is forced to learn and replicate the intrinsic statistical characteristics and dynamic patterns of real stress sequences, rather than merely pursuing numerical similarity. This enables the model to generate predicted sequences that are highly consistent with reality in both structure and morphology. Furthermore, this method effectively alleviates the problem of insufficient model training caused by the sparsity of anomalous stress data (imbalanced positive and negative samples), because if the generator cannot generate realistic anomalous sequences, it will be easily detected by the discriminator, thus incurring penalties in the adversarial loss. Ultimately, the stress prediction model trained using this method yields more realistic and reliable predictions, significantly enhancing its ability to predict sudden stress events.

[0098] Based on any of the above embodiments, step 230, which involves inputting the simulated pressure sequence and the real pressure sequence into the discriminator, includes: Step 231: Add time position codes to the input simulated pressure sequence and the real pressure sequence respectively. The time position codes are used to identify the temporal continuity of data points in the sequence. Step 232: Extract the temporal features of the input sequence based on the discriminator, and output a probability value representing that the input sequence is a real sequence. The input sequence is the simulated pressure sequence and the real pressure sequence after adding the time position encoding.

[0099] It should be noted that for time-series data, the order in which data points appear is crucial. A model that simply processes a set of data points may overlook this temporal dependency, thus failing to accurately determine the authenticity of a sequence. The embodiments of this invention aim to address this problem, ensuring that the discriminator can fully understand and utilize the temporal dimension information of the sequence.

[0100] Specifically, before any stress sequence (whether real or generated) is fed into the main structure of the discriminator for feature extraction, the sequence first needs to be preprocessed to inject temporal information.

[0101] Temporal position encoding adds relative or absolute positional information about each data point in a sequence to the model input. A temporal positional encoding is typically a vector with the same feature dimension as the data points in the sequence. For example, for a sequence of length m, there will be m different positional encoding vectors, corresponding to positions 1 through m in the sequence. This encoding vector is added to the feature vector of the corresponding data point, allowing the model to determine the position of each data point within the sequence when processing it.

[0102] It is understood that, in the embodiments of the present invention, whether it is a real stress sequence taken from the training set or a simulated stress sequence generated by the generator, it will undergo the same time position encoding process when input into the discriminator, so as to ensure that the discriminator processes the real sequence and the simulated sequence in the same way in the time dimension.

[0103] The sequence after positional encoding becomes the formal input to the discriminator. Subsequently, the input sequence (i.e., the stress sequence with added temporal positional encoding) is fed into the discriminator's main network (e.g., a multi-layer Transformer Encoder). The discriminator uses its internal self-attention mechanism to analyze the entire sequence, capturing the complex dependencies between data points in the temporal dimension, such as long-term stress trends, short-term fluctuations, and periodic patterns. The output of this process is a series of highly condensed temporal features. After extracting these deep temporal features, the discriminator aggregates these features through its final classification layer (e.g., a fully connected layer followed by a sigmoid activation function), ultimately outputting a scalar value between 0 and 1, which represents the probability that the input sequence is a true sequence.

[0104] This invention significantly enhances the discriminator's temporal discrimination capability by adding time position encoding to the input sequence. The discriminator no longer merely judges whether the distribution of pressure values ​​is reasonable, but rigorously examines whether the temporal continuity and variation patterns of the pressure sequence conform to real-world physical patterns. This forces the generator (i.e., the pressure prediction model) not only to generate numerically correct prediction points, but also to learn how to organize these points in a temporally coherent and logically sound manner. Therefore, this method effectively improves the dynamic realism of the generated sequence, resulting in a better performance of the ultimately trained pressure prediction model in predicting trend changes and capturing abrupt changes.

[0105] Based on any of the above embodiments, the training process further includes a negative sample augmentation step, which includes: True negative samples representing abnormal stress states are selected from the training sample set; The real negative samples are oversampled, and the simulated abnormal stress sequence generated by the generator during adversarial training is used as the synthetic negative samples. Based on the oversampled real negative samples and the synthetic negative samples, a mixed negative sample set is constructed, which is used to train the discriminator.

[0106] It should be noted that in the equipment stress data, the equipment operates in a normal state most of the time (corresponding to positive samples), while abnormal stress states such as high load and congestion (corresponding to negative samples) are relatively rare. This data imbalance can cause the model to be biased towards normal states during training, resulting in insufficient predictive ability for key abnormal states. To address this, this embodiment of the invention introduces a negative sample augmentation step, which is specifically designed to enhance the model's learning of abnormal states.

[0107] Specifically, the first step is to identify and separate anomalous samples from the massive training data. True negative samples refer to those samples in the training data that exhibit abnormal patterns in the actual stress sequences. The definition of anomalies here can be based on specific business scenarios; for example, it can be achieved by setting threshold rules (such as CPU utilization exceeding 95% for 5 consecutive minutes) or by automatically filtering using unsupervised anomaly detection algorithms (such as Isolation Forest or cluster analysis). This step yields a small but highly valuable set of true negative samples.

[0108] Because the number of true negative samples is scarce, they alone are insufficient for adequately training the model. Therefore, it is necessary to expand the quantity and diversity of negative samples. Oversampling is a simple and effective data augmentation technique that intentionally increases the frequency with which true negative samples are sampled when organizing training batches. For example, the true negative sample set can be duplicated multiple times, or it can be given a higher weight during sampling.

[0109] Furthermore, during the adversarial training of GANs, the generator attempts to generate various stress sequences in order to deceive the discriminator, including imitations of anomalous patterns. These simulated anomalous stress sequences with anomalous characteristics generated by the generator during training can be collected and used as synthetic negative samples.

[0110] Finally, the negative samples from both sources are combined to form a more diverse negative sample library, namely the hybrid negative sample set. This set consists of oversampled real negative samples and collected synthetic negative samples. This hybrid negative sample set is specifically used to train the discriminator.

[0111] This invention enhances the learning ability of the entire GAN framework for abnormal events through a negative sample augmentation strategy, effectively solving the problems of insufficient model training and poor generalization ability caused by the sparsity of negative samples. It significantly improves the recall and accuracy of the stress prediction model for abnormal stress, making the model more robust and practical in real-world applications.

[0112] The equipment pressure prediction device provided by the present invention is described below. The equipment pressure prediction device described below and the equipment pressure prediction method described above can be referred to in correspondence.

[0113] Based on any of the above embodiments Figure 4 This is a schematic diagram of the equipment pressure prediction device provided by the present invention, as shown below. Figure 4 As shown, the device includes: The data acquisition unit 410 is used to acquire multimodal time-series data of the target device in a historical period, wherein the multimodal time-series data includes resource status data and service load data; The feature extraction unit 420 is used to perform spatiotemporal correlation analysis on the resource status data and the service load data to extract spatiotemporal correlation features that characterize the impact of service load on resource status. The pressure prediction unit 430 is used to predict equipment pressure based on the spatiotemporal correlation features and generate a pressure prediction sequence for the target equipment in future time periods.

[0114] The apparatus provided in this invention, by acquiring multimodal time-series data including resource status data and service load data, fundamentally solves the problem of predictive bias caused by the single input information in traditional technologies. This invention incorporates service load data into the analysis scope, providing richer contextual information for equipment stress prediction. This enables the model to understand and learn the intrinsic relationship between service load and resource consumption, thus allowing for a more accurate response to stress fluctuations caused by specific services. Secondly, by performing spatiotemporal correlation analysis on resource status and service load data, this invention extracts spatiotemporal correlation features that characterize the impact of service load on resource status. This achieves a leap from simple numerical prediction to deep causal relationship modeling. This prediction method based on correlation features allows the model not only to predict stress changes but also to uncover the service loads that cause resource stress changes, greatly enhancing the robustness and interpretability of the prediction. Finally, by using these deeply refined spatiotemporal correlation features for equipment stress prediction, not only is high accuracy and reliability of the prediction results ensured, but accuracy is also improved in complex scenarios, long-term predictions, and responses to emergencies.

[0115] Based on any of the above embodiments, the steps of performing spatiotemporal correlation analysis on the resource status data and the service load data, and predicting equipment stress based on the spatiotemporal correlation features, are executed by a pre-trained stress prediction model. The stress prediction model includes an encoder and a decoder. The encoder is used to extract the spatiotemporal correlation features, and the decoder is used to generate the stress prediction sequence using an autoregressive mechanism.

[0116] Based on any of the above embodiments, the feature extraction unit is specifically used for: The encoder based on the stress prediction model uses a cross-modal attention mechanism to calculate the spatiotemporal correlation features between the resource status data and the business load data.

[0117] Based on any of the above embodiments, the feature extraction unit is specifically used for: The resource status data is mapped to a query matrix, and the business load data is mapped to a key matrix and a value matrix; Based on the query matrix and the key matrix, calculate the attention weight matrix; The spatiotemporal correlation features are obtained by weighted summation of the value matrix based on the attention weight matrix.

[0118] Based on any of the above embodiments, the pressure prediction unit is specifically used for: The spatiotemporal correlation features are input into the multi-channel output layer of the decoder; The multi-channel output layer performs parallel decoding of multiple indicators, generating pressure prediction values ​​of different dimensions at the same time step. The multi-indicator parallel decoding is performed iteratively according to time steps until the pressure prediction sequence is generated.

[0119] Based on any of the above embodiments, the stress prediction model is built based on a generator in a generative adversarial network, and the device further includes a model training unit, which is used for: Construct the generator and a discriminator for distinguishing between real sequences and generated sequences; Obtain a training sample set, which includes multimodal time-series data of the sample devices and the real pressure sequences corresponding to the sample devices; The training sample set is input into the generator to obtain the simulated pressure sequence output by the generator, and the simulated pressure sequence and the real pressure sequence are input into the discriminator to obtain the output result of the discriminator; Based on the real pressure sequence, the simulated pressure sequence, and the output of the discriminator, the generation loss and the discrimination loss are calculated. The parameters of the discriminator are updated based on the discriminant loss, and the parameters of the generator are updated based on the generation loss, with the parameter updates of the discriminator and the parameters of the generator performed alternately.

[0120] Based on any of the above embodiments, the model training unit is specifically used for: A time position code is added to both the input simulated pressure sequence and the actual pressure sequence. The time position code is used to identify the temporal continuity of data points in the sequence. The discriminator extracts the temporal features of the input sequence and outputs a probability value representing that the input sequence is a real sequence. The input sequence is the simulated pressure sequence and the real pressure sequence after adding the time position encoding.

[0121] Based on any of the above embodiments, the model training unit includes a negative sample augmentation subunit, which is used for: True negative samples representing abnormal stress states are selected from the training sample set; The real negative samples are oversampled, and the simulated abnormal stress sequence generated by the generator during adversarial training is used as the synthetic negative samples. Based on the oversampled real negative samples and the synthetic negative samples, a mixed negative sample set is constructed, which is used to train the discriminator.

[0122] Based on any of the above embodiments, the generation loss is composed of a weighted sum of the prediction loss and the adversarial loss. The prediction loss represents the difference between the simulated stress sequence and the real stress sequence, and the adversarial loss represents the probability that the simulated stress sequence is identified as a real sequence by the discriminator. The weight coefficients of the prediction loss and the adversarial loss are dynamically adjusted according to the training phase.

[0123] Based on any of the above embodiments, the resource status data includes at least one of processor utilization, memory occupancy, disk read / write speed, and network transmission speed; The business load data includes at least one of the following: business type identifier, number of concurrent requests, and request response latency.

[0124] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include a processor 510, a communication interface 520, a memory 530, and a communication bus 540, wherein the processor 510, the communication interface 520, and the memory 530 communicate with each other via the communication bus 540. The processor 510 can call logical instructions in the memory 530 to execute a device stress prediction method. The method includes: acquiring multimodal time-series data of the target device in a historical period, the multimodal time-series data including resource status data and service load data; performing spatiotemporal correlation analysis on the resource status data and the service load data to extract spatiotemporal correlation features characterizing the impact of service load on resource status; and performing device stress prediction based on the spatiotemporal correlation features to generate a stress prediction sequence for the target device in future periods.

[0125] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to related technologies, 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 the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0126] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the device stress prediction method provided by the above methods. The method includes: acquiring multimodal time-series data of a target device in a historical period, the multimodal time-series data including resource status data and service load data; performing spatiotemporal correlation analysis on the resource status data and the service load data to extract spatiotemporal correlation features characterizing the impact of service load on resource status; and performing device stress prediction based on the spatiotemporal correlation features to generate a stress prediction sequence for the target device in future periods.

[0127] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the device stress prediction method provided by the above methods. The method includes: acquiring multimodal time-series data of a target device in a historical period, the multimodal time-series data including resource status data and service load data; performing spatiotemporal correlation analysis on the resource status data and the service load data to extract spatiotemporal correlation features characterizing the impact of service load on resource status; and performing device stress prediction based on the spatiotemporal correlation features to generate a stress prediction sequence for the target device in future periods.

[0128] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0129] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0130] 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting equipment pressure, characterized in that, include: Acquire multimodal time-series data of the target device within a historical period, wherein the multimodal time-series data includes resource status data and service load data; Spatiotemporal correlation analysis is performed on the resource status data and the service load data to extract spatiotemporal correlation features that characterize the impact of service load on resource status; Based on the spatiotemporal correlation features, equipment pressure is predicted, and a pressure prediction sequence for the target equipment in future time periods is generated.

2. The equipment pressure prediction method according to claim 1, characterized in that, The steps of performing spatiotemporal correlation analysis on the resource status data and the service load data, and predicting equipment stress based on the spatiotemporal correlation features, are executed by a pre-trained stress prediction model. The stress prediction model includes an encoder and a decoder. The encoder is used to extract the spatiotemporal correlation features, and the decoder is used to generate the stress prediction sequence using an autoregressive mechanism.

3. The equipment pressure prediction method according to claim 2, characterized in that, The step of performing spatiotemporal correlation analysis on the resource status data and the service load data to extract spatiotemporal correlation features characterizing the impact of service load on resource status includes: The encoder based on the stress prediction model uses a cross-modal attention mechanism to calculate the spatiotemporal correlation features between the resource status data and the business load data.

4. The equipment pressure prediction method according to claim 3, characterized in that, The step of employing a cross-modal attention mechanism to calculate the spatiotemporal correlation features between the resource state data and the service load data includes: The resource status data is mapped to a query matrix, and the business load data is mapped to a key matrix and a value matrix; Based on the query matrix and the key matrix, calculate the attention weight matrix; The spatiotemporal correlation features are obtained by weighted summation of the value matrix based on the attention weight matrix.

5. The equipment pressure prediction method according to claim 2, characterized in that, The step of predicting equipment pressure based on the spatiotemporal correlation features and generating a pressure prediction sequence for the target equipment in future time periods includes: The spatiotemporal correlation features are input into the multi-channel output layer of the decoder; The multi-channel output layer performs parallel decoding of multiple indicators, generating pressure prediction values ​​of different dimensions at the same time step. The multi-indicator parallel decoding is performed iteratively according to time steps until the pressure prediction sequence is generated.

6. The equipment pressure prediction method according to any one of claims 2 to 5, characterized in that, The stress prediction model is built based on a generator in a generative adversarial network, and is trained through the following steps: Construct the generator and a discriminator for distinguishing between real sequences and generated sequences; Obtain a training sample set, which includes multimodal time-series data of the sample devices and the real pressure sequences corresponding to the sample devices; The training sample set is input into the generator to obtain the simulated pressure sequence output by the generator, and the simulated pressure sequence and the real pressure sequence are input into the discriminator to obtain the output result of the discriminator; Based on the real pressure sequence, the simulated pressure sequence, and the output of the discriminator, the generation loss and the discrimination loss are calculated. The parameters of the discriminator are updated based on the discriminant loss, and the parameters of the generator are updated based on the generation loss, with the parameter updates of the discriminator and the parameters of the generator performed alternately.

7. The equipment pressure prediction method according to claim 6, characterized in that, The step of inputting the simulated pressure sequence and the real pressure sequence into the discriminator includes: A time position code is added to both the input simulated pressure sequence and the actual pressure sequence. The time position code is used to identify the temporal continuity of data points in the sequence. The discriminator extracts the temporal features of the input sequence and outputs a probability value representing that the input sequence is a real sequence. The input sequence is the simulated pressure sequence and the real pressure sequence after adding the time position encoding.

8. The equipment pressure prediction method according to claim 6, characterized in that, The training process also includes a negative sample augmentation step, which includes: True negative samples representing abnormal stress states are selected from the training sample set; The real negative samples are oversampled, and the simulated abnormal stress sequence generated by the generator during adversarial training is used as the synthetic negative samples. Based on the oversampled real negative samples and the synthetic negative samples, a mixed negative sample set is constructed, which is used to train the discriminator.

9. The equipment pressure prediction method according to claim 6, characterized in that, The generation loss is composed of a weighted sum of the prediction loss and the adversarial loss. The prediction loss represents the difference between the simulated stress sequence and the real stress sequence, and the adversarial loss represents the probability that the simulated stress sequence is identified as a real sequence by the discriminator. The weight coefficients of the prediction loss and the adversarial loss are dynamically adjusted according to the training phase.

10. The equipment pressure prediction method according to any one of claims 1 to 5, characterized in that, The resource status data includes at least one of processor utilization, memory usage, disk read / write speed, and network transmission speed; The business load data includes at least one of the following: business type identifier, number of concurrent requests, and request response latency.

11. A device for predicting equipment pressure, characterized in that, include: The data acquisition unit is used to acquire multimodal time-series data of the target device within a historical period, wherein the multimodal time-series data includes resource status data and service load data; The feature extraction unit is used to perform spatiotemporal correlation analysis on the resource status data and the service load data to extract spatiotemporal correlation features that characterize the impact of service load on resource status. The pressure prediction unit is used to predict equipment pressure based on the spatiotemporal correlation features and generate a pressure prediction sequence for the target equipment in future time periods.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the equipment pressure prediction method as described in any one of claims 1 to 10.

13. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the equipment pressure prediction method as described in any one of claims 1 to 10.

Images