Generation method of software residual life prediction model, prediction method and storage medium

Abstract

The invention provides a software residual life prediction model generation method, a prediction method and a storage medium, and belongs to the technical field of software aging and regeneration, and the generation method comprises the steps: obtaining failure field data during the software operation period, and carrying out the preprocessing of the failure field data, and obtaining a training sample; training a preset network model based on the training sample to obtain a software residual life prediction model; the network model comprises the following steps: an embedded encoding layer encodes training samples to obtain high-order representation with a unified dimension; the multi-scale image convolution model converts high-order representation into multi-scale spatial feature representation; the frequency domain window fusion model obtains frequency domain and time domain fused high-dimensional feature representation based on multi-scale spatial feature representation; and the multi-scale convolution predictor predicts the residual life of the software based on the high-dimensional features. According to the method, the problem that the software restart decision is not accurate due to the fact that an existing method has defects in the aspects of extracting a complex aging factor relation, modeling a multi-scale aging mode and the like can be solved.

Description

Technical Field

[0001] This invention relates to the field of software aging and regeneration technology, specifically to a method for generating a software remaining life prediction model, a prediction method, and a storage medium. Background Technology

[0002] As software systems operate long-term in scenarios such as cloud platforms, industrial control, and smart terminals, issues like resource leaks and accumulated error states can lead to performance degradation and even system crashes. This process is often insidious and gradual, making it difficult to detect and intervene in a timely manner using traditional methods. Therefore, the concept of "remaining software lifetime" has been proposed, referring to the remaining runtime of a system before software fails due to aging. Accurately predicting the remaining software lifetime helps in the rational planning of preventative restarts and resource scheduling, improving system stability, resource utilization, and business continuity, and is a crucial foundation for achieving intelligent operation and maintenance.

[0003] In recent years, with the development of technologies such as deep learning, graph neural networks, and frequency domain modeling, software aging prediction has gradually evolved towards intelligence and data-driven approaches. By constructing a state prediction model based on deep neural networks and combining it with reinforcement learning to optimize regeneration scheduling strategies, accurate modeling and dynamic intervention of the software aging process can be achieved, significantly improving the availability and stability of the system.

[0004] However, existing methods still have shortcomings in extracting complex relationships between aging factors and modeling multi-scale aging patterns, leading to inaccurate software restart decisions. Summary of the Invention In view of this, it is necessary to provide a method for generating, predicting and storing a software remaining life prediction model, in order to solve the technical problem that existing methods are inadequate in extracting complex aging factor relationships and modeling multi-scale aging modes, which leads to inaccurate software restart decisions.

[0005] To address the aforementioned problems, in a first aspect, the present invention provides a method for generating a software remaining lifetime prediction model, comprising: Acquire failure scene data during software operation and preprocess the failure scene data to obtain training samples; A preset network model is trained based on the training samples to obtain a software remaining lifetime prediction model. The network model includes: An embedded encoding layer is used to encode the training samples to obtain a high-order representation with a unified dimension; A multi-scale graph convolution model is used to convert the higher-order representation into a multi-scale spatial feature representation; A frequency domain window fusion model is used to obtain a high-dimensional feature representation that fuses the frequency domain and time domain based on the multi-scale spatial feature representation. A multi-scale convolutional predictor is used to predict the remaining lifetime of the software based on the high-dimensional features.

[0006] In one possible implementation, the failure site data is preprocessed to obtain training samples, including: The failure site data is cleaned, and the remaining lifetime is labeled on the cleaned data to obtain labeled samples; The labeled samples are normalized to obtain samples with a uniform scale; The uniform-scale samples are divided using a sliding window strategy to obtain training samples.

[0007] In one possible implementation, the higher-order representation is converted into a multi-scale spatial feature representation, including: The higher-order representation is subjected to channel separation and feature extraction to obtain the extracted features; The extracted features are captured by multi-scale residual convolution and dynamic graph convolution to obtain the captured features; The captured features are then subjected to a displacement operation to obtain a multi-scale spatial feature representation.

[0008] In one possible implementation, the higher-order representation is subjected to channel separation and feature extraction to obtain extracted features, including: The higher-order representation is subjected to channel separation and feature extraction through conditional positional encoding to obtain extracted features; The conditional location encoding uses a 7×7 depthwise separable convolution.

[0009] In one possible implementation, each node in the multi-scale graph convolution model retains the same number of connections.

[0010] In one possible implementation, based on the multi-scale spatial feature representation, a high-dimensional feature representation fused from the frequency domain and time domain is obtained, including: Frequency domain features are extracted from the multi-scale spatial feature representation based on the encoder and employing a multi-head window self-attention mechanism and Fourier hybridization. Based on the decoder and combined with the masked multi-head window self-attention and cross-attention mechanism, the frequency domain features are fused with the spatiotemporal features in the multi-scale spatial feature representation to obtain a high-dimensional feature representation that fuses the frequency domain and time domain.

[0011] In one possible implementation, predicting the remaining software lifetime based on the high-dimensional features includes: The high-dimensional features are mapped to software aging failure time through a projection layer.

[0012] In one possible implementation, the training samples are encoded to obtain a high-order representation with a unified dimension, including: The training samples are mapped to fixed-dimensional vector representations, and relative position information is added to each time step. The temporal context is modeled to obtain a high-order representation with a unified dimension.

[0013] Secondly, the present invention also provides a method for predicting the remaining lifetime of software, comprising: Acquire the failure site data to be predicted during software operation, and preprocess the failure site data to be predicted to obtain preprocessed data; The preprocessed data is input into the software remaining lifetime prediction model obtained by the above method to obtain the predicted software remaining lifetime.

[0014] Thirdly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method for generating a software remaining lifetime prediction model as described in any of the preceding claims, or implements the steps of the software remaining lifetime prediction method as described above.

[0015] The beneficial effects of the above implementation are as follows: The software remaining lifetime prediction model generation method, prediction method, and storage medium provided by the present invention obtain a software remaining lifetime prediction model by preprocessing the failure site data during software operation to obtain training samples, and then training a preset network model. The network model includes: an embedded coding layer for encoding the training samples to obtain a high-order representation of a unified dimension; a multi-scale graph convolution model for converting the high-order representation into a multi-scale spatial feature representation; a frequency domain window fusion model for obtaining a high-dimensional feature representation fused from the frequency domain and time domain based on the multi-scale spatial feature representation; and a multi-scale convolutional predictor for predicting the software remaining lifetime based on the high-dimensional features.

[0016] The network model in this invention, by combining multi-scale graph convolution with a frequency domain window fusion model, can effectively capture the spatiotemporal correlation features among various software aging factors. The frequency domain window mechanism enhances the ability to identify aging evolution patterns at different frequency levels, comprehensively improving the accuracy and generalization ability of aging state modeling, eliminating the influence of different dimensions, improving the model's convergence efficiency and prediction accuracy, thereby improving the prediction accuracy of software remaining lifetime. This solves the technical problem of existing methods being insufficient in extracting complex aging factor relationships and modeling multi-scale aging patterns, leading to inaccurate software restart decisions. Attached Figure Description

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

[0018] Figure 1 A flowchart of an embodiment of the method for generating a software remaining lifetime prediction model provided by the present invention; Figure 2 A schematic diagram of the structure of the multi-scale graph convolution and frequency domain window fusion network provided by the present invention; Figure 3 This is a schematic diagram of the structure of the multi-scale graph convolution model provided by the present invention; Figure 4 This is a schematic diagram of the frequency domain window fusion model provided by the present invention; Figure 5 This is a schematic diagram of the prediction results provided by the present invention; Figure 6 A flowchart of an embodiment of the software remaining lifetime prediction method provided by the present invention; Figure 7 A schematic diagram of an embodiment of the electronic device provided by the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] In the description of the embodiments of this application, unless otherwise stated, "a plurality of" means two or more.

[0021] In this embodiment of the invention, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, apparatus, product or device that includes a series of steps or modules is not necessarily limited to those steps or modules that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to such process, method, product or device.

[0022] The naming or numbering of steps in the embodiments of the present invention does not mean that the steps in the method flow must be executed in the time / logical order indicated by the naming or numbering. The execution order of the named or numbered process steps can be changed according to the technical purpose to be achieved, as long as the same or similar technical effect can be achieved.

[0023] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0024] This invention provides a method for generating a software remaining lifetime prediction model, a prediction method, and a storage medium, which are described below.

[0025] This invention proposes a software remaining lifetime prediction method based on a fusion network of multi-scale graph convolution and frequency domain windowing. This method effectively captures the spatiotemporal correlation features among various software aging factors through a multi-scale graph convolution structure, and enhances the ability to identify aging evolution patterns at different frequency levels by combining a frequency domain windowing mechanism, thus comprehensively improving the accuracy and generalization ability of aging state modeling. This method is applicable to various long-term operating environments such as cloud computing, edge computing, and high-availability service systems, effectively overcoming the problems of insufficient aging modeling and inaccurate restart decisions in traditional methods, providing theoretical support and technical pathways for building highly reliable and adaptive software systems.

[0026] like Figure 1 As shown, the present invention provides a method for generating a software remaining lifetime prediction model, comprising: S101. Obtain failure scene data during software operation, and preprocess the failure scene data to obtain training samples.

[0027] Understandably, systematically collecting failure site data during software operation, including information such as free memory, device temperature, and used memory, is necessary to construct a complete raw dataset.

[0028] The original dataset comes from the software lifecycle accelerated testing experiment. This experiment can efficiently collect key field data of software before and after failure by simulating or accelerating the aging process in the software operating environment, including various performance indicators and resource usage, providing rich real samples for subsequent modeling.

[0029] S102. Train a preset network model based on the training samples to obtain a software remaining lifetime prediction model; The network model includes: An embedded encoding layer is used to encode the training samples to obtain a high-order representation with a unified dimension; A multi-scale graph convolution model is used to convert the higher-order representation into a multi-scale spatial feature representation; A frequency domain window fusion model is used to obtain a high-dimensional feature representation that fuses the frequency domain and time domain based on the multi-scale spatial feature representation. A multi-scale convolutional predictor is used to predict the remaining software lifetime based on the high-dimensional features, where the remaining software lifetime includes the software aging failure time.

[0030] Understandably, this invention provides a method for predicting the remaining lifetime of software based on a multi-scale graph convolution and frequency domain window fusion network. This method effectively captures the complex spatiotemporal features of the software aging process by combining multi-scale graph convolution and frequency domain window fusion techniques, thereby improving prediction accuracy. The specific steps are as follows: The preprocessed sliding window data (i.e., training samples) is input into the embedding encoding layer. Through numerical embedding, positional embedding, and temporal embedding, the original features are transformed into a high-order representation with a unified dimension. The embedding encoding operation provides more expressive input features for subsequent graph convolution and frequency domain modeling.

[0031] A multi-scale graph convolution (MCG) module is constructed, which includes the following operations: channel-wise partitioning and feature extraction (CPB) of the input data; capturing local and global spatial features through multi-scale residual convolution (MRConv) and dynamic weighted convolution (DW Conv); enhancing the spatial diversity of features using a shift operation (Shift k); and finally outputting a multi-scale spatial feature representation. The MGC module enhances the diversity of spatial features through dynamic graph convolution and shift operations, making it suitable for multi-scale software aging modes.

[0032] A frequency domain window fusion (Fourier-enhanced Window Encoder-Decoder, FourWinED) module is constructed, which includes an encoder and a decoder. The encoder can extract frequency domain features, and the decoder further fuses spatiotemporal features, ultimately outputting a high-dimensional feature representation that fuses the frequency and time domains.

[0033] For software remaining lifetime prediction, the features generated in the above steps are input into the prediction module of a multi-scale convolutional forecaster (MCFormer). The high-dimensional features are mapped to predicted values ​​through a projection layer, outputting the software aging failure time and generating third-party data. Then, the performance of the constructed model is evaluated using model performance evaluation metrics, and the model is saved. MCFormer is used as the prediction head, combined with residual connections and layer normalization to improve the training stability and prediction accuracy of the model. Relevant evaluation metrics such as root mean square error (RMSE), mean absolute error (MAE), and R-squared (R²) are used. 2 The performance of the model proposed in this invention is evaluated.

[0034] In some embodiments, the failure site data is preprocessed to obtain training samples, including: The failure site data is cleaned, and the remaining lifetime is labeled on the cleaned data to obtain labeled samples; The labeled samples are normalized to obtain samples with a uniform scale; The uniform-scale samples are divided using a sliding window strategy to obtain training samples.

[0035] Understandably, the data preprocessing involves cleaning and labeling the raw data (i.e., failure site data) to indicate the remaining lifetime of each sample; then, normalization operations (such as max-min normalization) are used to scale the data to a uniform scale; finally, a sliding window strategy is used to partition the data and generate model training samples.

[0036] Specifically, firstly, the collected aging indicators are standardized using the Min-MaxNormalization method to eliminate the impact of dimensional differences on model training. Then, a sliding window technique with a step size of 1 is used to slice the time series data to generate continuous and overlapping training sample sequences, thereby effectively capturing the dynamic evolution characteristics of the software aging process and improving the model's predictive ability and generalization performance.

[0037] In some embodiments, converting the higher-order representation into a multi-scale spatial feature representation includes: The higher-order representation is subjected to channel separation and feature extraction to obtain the extracted features; The extracted features are captured by multi-scale residual convolution and dynamic graph convolution to obtain the captured features; The captured features are then subjected to a displacement operation to obtain a multi-scale spatial feature representation.

[0038] It is understood that the structure diagram of the temporal relationship extraction module (corresponding to the combination of CPE module and MRConv in MGC) and the structure diagram of the spatial relationship extraction module (corresponding to sparse graph construction and adjacency modeling mechanism) of the present invention are shown.

[0039] For the temporal relation extraction module (CPE + MRConv): This module enhances the spatial sensitivity of the graph convolutional network by introducing Conditional Positional Encoding (CPE) before graph operations, thereby improving the model's ability to model the temporal and spatial structures of the graph. The main structure is as follows: CPE Module Structure: To enhance the graph neural network's ability to model implicit temporal dependencies in sequences, the module first introduces a CPE (Conditional Positional Encodings) mechanism to provide learnable positional awareness. CPE uses 7×7 depthwise separable convolutions to encode the position of the input sequence and embeds it into the message passing process of the graph convolution. Subsequently, MRConv captures the maximum feature difference between each node and its neighborhood, achieving more effective local context modeling.

[0040] In some embodiments, the higher-order representation is subjected to channel separation and feature extraction to obtain extracted features, including: The higher-order representation is subjected to channel separation and feature extraction through conditional positional encoding to obtain extracted features; The conditional location encoding uses a 7×7 depthwise separable convolution.

[0041] Understandably, to enhance the graph neural network's ability to model implicit temporal dependencies in sequences, the module first introduces a CPE mechanism to provide learnable position-aware capabilities. CPE uses 7×7 depthwise separable convolutions to encode the position of the input sequence and embeds it into the message passing process of the graph convolution. Subsequently, MRConv captures the maximum feature difference between each node and its neighborhood, achieving more effective local context modeling.

[0042] In some embodiments, each node in the multi-scale graph convolution model retains the same number of connections.

[0043] Understandably, in traditional graph construction methods, the number of connections between nodes typically increases linearly with resolution, leading to a significant increase in the computational complexity of graph structures at high resolutions. In this invention, the Multi-Scale Graph Convolutional Module (MGC) employs a fixed-connection strategy, where each node retains only 5 connections: one self-connection and four directional connections. This strategy reduces the complexity of graph construction from a resolution-dependent variable to a constant, effectively avoiding the problem of graph structures becoming overly complex in high-resolution scenes due to excessive connections. Furthermore, the fixed-connection approach can reduce the risk of over-smoothing in graph convolution while preserving necessary topological information, thus improving the robustness of spatial modeling.

[0044] In some embodiments, based on the multi-scale spatial feature representation, a high-dimensional feature representation fused from the frequency domain and time domain is obtained, including: Frequency domain features are extracted from the multi-scale spatial feature representation based on the encoder and employing a multi-head window self-attention mechanism and Fourier hybridization. Based on the decoder and combined with the masked multi-head window self-attention and cross-attention mechanism, the frequency domain features are fused with the spatiotemporal features in the multi-scale spatial feature representation to obtain a high-dimensional feature representation that fuses the frequency domain and time domain.

[0045] Understandably, constructing the Fourier-enhanced Window Encoder-Decoder (FourWinED) module involves the following operations: the encoder employs a multi-head window self-attention mechanism and Fourier Mix to extract frequency domain features; the decoder combines masked multi-head window self-attention and cross-attention mechanisms to further fuse spatiotemporal features; and finally, it outputs a high-dimensional feature representation that fuses the frequency and time domains. The FourWinED module effectively captures the correlated features between the frequency and time domains through Fourier Mix and window attention mechanisms.

[0046] In some embodiments, predicting the remaining software lifetime based on the high-dimensional features includes: The high-dimensional features are mapped to software aging failure time through a projection layer.

[0047] In some embodiments, the training samples are encoded to obtain a high-order representation with a unified dimension, including: The training samples are mapped to fixed-dimensional vector representations, and relative position information is added to each time step. The temporal context is modeled to obtain a high-order representation with a unified dimension.

[0048] In some embodiments, such as Figure 1As shown, the original multi-dimensional index data (i.e., failure site data during software operation) is normalized to eliminate the influence of different units, improve the convergence efficiency and prediction accuracy of the model, and a sliding window method is used to segment the time series to construct the input samples and labels required for training, thereby capturing the temporal dependence of the data. Then, the input data is converted into a vector form that the model can process through the embedding layer, providing a more expressive input basis for subsequent feature extraction. Next, the spatial features are extracted using the multi-scale graph convolution (MGC) module, and the spatiotemporal features are further fused through the frequency domain window fusion network (FourWinED) module. Finally, the prediction module outputs the software remaining life prediction results.

[0049] To more clearly illustrate the specific implementation of the present invention, the structure of the software remaining lifetime prediction method based on a multi-scale graph convolution and frequency domain window fusion network proposed in this embodiment is as follows: Figure 2 The specific steps are explained in detail below: Step 1: Acquisition of software lifespan data.

[0050] To efficiently acquire failure scenario data during software operation, this invention employs an accelerated lifespan testing scheme to apply stress to the target software system. This scheme accelerates the aging process by continuously simulating real user actions, such as starting / closing applications and creating / destroying virtual machines. Simultaneously with the load generation, an automated log collection tool is designed to collect and save key lifespan indicator data in real time.

[0051] Step 2, Preprocessing of lifespan indicators.

[0052] This invention employs a max-min normalization method to normalize the aging indicators of each dimension, mapping input features of different dimensions to a unified numerical range [0,1], thereby improving the stability and convergence speed of model training. The specific normalization process is shown in formula (1): (1) in, No. The sample at the th The original values ​​of each aging indicator For the normalized data, and These respectively represent the indicator (the first one). The minimum and maximum values ​​of each feature in the training set.

[0053] After normalization, a sliding window technique is used to partition the data into sequences to capture the temporal characteristics of the aging process. Specifically, the time window length is set to... , will each The aging metrics within a consecutive time step are used as input samples to predict subsequent system state metrics. A sliding window slides with a step size of 1 to generate multiple overlapping training samples from the long sequence, ensuring that the model can learn the dynamic change patterns of aging features during time evolution.

[0054] Step 3, Feature embedding encoding.

[0055] To enhance the expressive power of the original multidimensional aging index and provide a unified high-dimensional representation for subsequent model modules, this invention designs a multimodal embedding layer, which converts the preprocessed sliding window sequence into a unified embedding vector. Specifically, it includes the following three sub-modules: Numerical embedding (Token Embedding): uses one-dimensional convolution operations to map the original features at each time step into a vector representation of fixed dimensions; Positional Embedding: Introduces a learnable positional encoding mechanism to add relative positional information to each time step, enhancing the model's ability to perceive temporal order; Temporal Embedding: Models the time context using embedding layers of multiple time fields such as hour, day of the week, date, and month.

[0056] The final output embedding representation dimension is a fixed d_model, which is the feature representation dimension used uniformly in all layers of the model, and serves as the input to the subsequent graph convolution module.

[0057] Step 4: Construct the multi-scale graph convolution (MGC) module.

[0058] from Figure 3 As can be seen from the diagram, the MGC module of the present invention includes a structure diagram of a time relation extraction module (corresponding to the combination of the CPE module and MRConv in MGC) and a structure diagram of a spatial relation extraction module (corresponding to the sparse graph construction and adjacency modeling mechanism).

[0059] For the temporal relation extraction module (CPE + MRConv): This module enhances the spatial sensitivity of the graph convolutional network by introducing Conditional Positional Encoding (CPE) before graph operations, thereby improving the model's ability to model the temporal and spatial structures of the graph. The main structure is as follows: CPE Module Structure: To enhance the graph neural network's ability to model implicit temporal dependencies in sequences, the module first introduces the CPE mechanism to provide learnable position-aware capabilities. CPE uses 7×7 depthwise separable convolutions to encode the position of the input sequence and embeds it into the message passing process of the graph convolution. Subsequently, MRConv captures the maximum feature difference between each node and its neighborhood, achieving more effective local context modeling. The overall computation process is as follows: (2) in, Indicates input features, , These are the weight matrices for the linear transformation. The output of conditional position encoding, This represents the maximum relative convolution (see below for details). Residual connections are used to improve training stability and information transfer efficiency.

[0060] The calculation process for MRConv is as follows: (3) in, To represent the characteristics of a node itself, For its neighboring nodes, This represents the set of adjacencies of nodes. This structure can effectively uncover potential dependencies between nodes in a time series and capture significant dynamic changes over time.

[0061] For the spatial relationship extraction module (sparse graph construction): This module significantly reduces the computational overhead in the graph construction process by designing a static sparse graph structure, making it suitable for visual modeling at high resolution. The main contents are as follows: In traditional graph construction methods, the number of connections between nodes typically increases linearly with resolution, leading to a significant increase in computational complexity for high-resolution graph structures. In this invention, the Multi-Scale Graph Convolutional Module (MGC) employs a fixed-connection strategy, where each node retains only 5 connections: including 1 self-connection and 4 directional connections (long-range spatial connections up / down / left / right, see...). Figure 3 (As shown). This strategy reduces the complexity of graph construction from a resolution-dependent variable to a constant, effectively avoiding the problem of graph structures becoming overly complex in high-resolution scenes due to excessive connections. Furthermore, the fixed-connection approach can reduce the risk of excessive smoothing in graph convolution while preserving necessary topological information, thus improving the robustness of spatial modeling.

[0062] Step 5: Construct the frequency domain window fusion (FourWinED) module.

[0063] from Figure 4The structure and processing flow of the encoder and decoder modules of this invention can be seen, which fully integrates the local window attention mechanism and the frequency domain modeling module (Fourier TransformBlock). This structure reflects the ability of this method to jointly model local dependencies and global periodic information in temporal modeling.

[0064] For the local window attention mechanism: To reduce the computational complexity of traditional self-attention mechanisms when processing long sequences, this paper adopts a local window partitioning strategy, dividing the global sequence into several local subsequences (windows), and performing attention calculations independently within each window. Assume the original input sequence is... We divide it into There are n subsequences, each with a length of n. As shown in formula (4): (4) Accordingly, queries, keys, and values ​​are also divided, and the local attention weights are calculated as follows: (5) in, It is the first Query Matrix for each window. It is the first The key matrix of each window. It is the first Value Matrix of a Window It is the dimension of each head.

[0065] Finally, the results from all windows are concatenated to obtain the complete attention output, as shown in formula (6): (6) in: It is the first Self-attention output of each window, It represents the number of windows.

[0066] This mechanism limits the receptive field of attention through a local window, and its computational complexity is reduced by that of global attention. Reduced to This significantly improves efficiency. However, because it can only perceive local information, it struggles to capture long-range dependencies across windows. Therefore, a frequency domain modeling module is introduced: to enhance the model's ability to model global periodic features, a frequency domain representation method based on the Fast Fourier Transform (FFT) is introduced. This module transforms the input sequence from the time domain to the frequency domain by performing a Fourier transform on it. Mapping to frequency domain representation As shown in formula (7): (7) in, This indicates the Fourier transform operation. It is an implementation of the Fast Fourier Transform. It is the input time-domain signal. It is the transformed frequency domain signal.

[0067] In the frequency domain, low-frequency components correspond to the global trend and periodic structure of the sequence, while high-frequency components often represent local variations or noise. To extract useful periodic features more effectively, this paper applies complex weighted filtering operations (such as channel attention, band masking, etc.) in the frequency domain and then transforms it back to the time domain as shown in Equation (8): (8) in, This indicates the inverse Fourier transform operation. This represents the learnable frequency domain weight matrix. This represents element-wise multiplication (Hadamard product). It is an implementation of the inverse fast Fourier transform. It is the output sequence after inverse transformation, that is, the signal converted from the frequency domain back to the time domain. This process not only captures the long-period components in the time series that are difficult to model using attention mechanisms, but also improves the stability and robustness of prediction to a certain extent.

[0068] To fully integrate local contextual dependencies and global periodic information, this module combines local window attention with frequency domain modeling and embeds it into the encoder and decoder structure. The specific process is as follows: In the encoder, the spatiotemporal features that integrate temporal dynamics and structural dependencies extracted in step 3 are first used as input, and a local window attention mechanism is applied to extract local temporal correlation features. Then, the output features are transformed to the frequency domain through Fourier transform and periodic pattern enhancement is performed in the frequency domain. Finally, the inverse transform is performed to return to the time domain, and the output is fused with the original window attention output before being fed into the next layer of encoding. Similarly, the same structure is used in the decoder, and finally, multi-scale features are fused through two one-dimensional convolutional layers and layer normalization (LayerNorm).

[0069] Step 6, Software Remaining Lifetime Prediction This step takes the multi-scale fusion features extracted in step 4 as input, maps them to the target variable dimension through a linear projection layer for regression prediction, and outputs the final prediction result, such as... Figure 5 As shown. This invention employs mean absolute error (MAE), root mean square error (RMSE), and R-squared (R...2 Three indicators are used to measure the accuracy of the model's prediction results. The specific calculation formulas for the above three evaluation indicators are shown in formulas (9) to (11).

[0070] (9) (10) (11) in, For the true value, For predicted values, For the sample size, The above three evaluation indicators, from the perspective of absolute error, quantify the degree of deviation between predicted and actual values.

[0071] Compared to existing, relatively simple methods for predicting the remaining lifespan of software, this invention can more effectively address aging phenomena in complex software systems, achieving high-precision prediction of the remaining lifespan of software. Through the prediction mechanism of this invention, managers can anticipate software aging trends, take timely recovery measures, reduce failure risks, and improve system stability and availability.

[0072] This invention, by integrating graph neural networks with an improved temporal modeling structure, effectively captures the interaction relationships and dynamic evolution characteristics of multi-dimensional aging indicators during software operation, enabling accurate estimation of the remaining software lifespan. Compared to traditional methods relying on a single aging variable, the method provided by this invention possesses stronger adaptability and generalization capabilities, and can stably output failure time prediction results under complex operating conditions. This technical approach enhances the comprehensiveness of aging modeling and the proactiveness of predictive decision-making, facilitating the early deployment of fault prevention measures in the operation and maintenance system, thereby extending the software service life, reducing operation and maintenance costs, and significantly enhancing the system's operational stability and fault tolerance.

[0073] like Figure 6 As shown, the present invention also provides a method for predicting the remaining lifetime of software, comprising: S601. Obtain the failure site data to be predicted during software operation, and preprocess the failure site data to be predicted to obtain preprocessed data. S602. Input the preprocessed data into the software remaining lifetime prediction model obtained by the above method to obtain the predicted software remaining lifetime.

[0074] It is understood that the preprocessing in this embodiment differs slightly from the preprocessing in the steps described above. Specifically, the preprocessing in this embodiment does not require data annotation; the other steps are the same as those described above. Remaining software lifetime includes software aging and failure time.

[0075] like Figure 7 As shown, the present invention also provides an electronic device 700. The electronic device 700 includes a processor 701, a memory 702, and a display 703. Figure 7 Only some components of the electronic device 700 are shown, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

[0076] In some embodiments, memory 702 may be an internal storage unit of electronic device 700, such as a hard disk or memory of electronic device 700. In other embodiments, memory 702 may also be an external storage device of electronic device 700, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on electronic device 700.

[0077] Furthermore, the memory 702 may include both internal storage units of the electronic device 700 and external storage devices. The memory 702 is used to store application software and various types of data installed on the electronic device 700.

[0078] In some embodiments, processor 701 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 702 or process data, such as the software remaining lifetime prediction model generation method / software remaining lifetime prediction method in this invention.

[0079] In some embodiments, display 703 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 703 is used to display information from electronic device 700 and to display a visual user interface. Components 701-703 of electronic device 700 communicate with each other via a system bus.

[0080] In some embodiments of the present invention, when processor 701 executes the program for generating the software remaining lifetime prediction model in memory 702, the following steps can be implemented: Acquire failure scene data during software operation and preprocess the failure scene data to obtain training samples; A preset network model is trained based on the training samples to obtain a software remaining lifetime prediction model. The network model includes: An embedded encoding layer is used to encode the training samples to obtain a high-order representation with a unified dimension; A multi-scale graph convolution model is used to convert the higher-order representation into a multi-scale spatial feature representation; A frequency domain window fusion model is used to obtain a high-dimensional feature representation that fuses the frequency domain and time domain based on the multi-scale spatial feature representation. A multi-scale convolutional predictor is used to predict the remaining lifetime of the software based on the high-dimensional features.

[0081] Alternatively, when processor 701 executes the software remaining lifetime prediction program in memory 702, the following steps can be performed: Acquire the failure site data to be predicted during software operation, and preprocess the failure site data to be predicted to obtain preprocessed data; The preprocessed data is input into the software remaining lifetime prediction model obtained by the above method to obtain the predicted software remaining lifetime.

[0082] It should be understood that when the processor 701 executes the software remaining lifetime prediction model generation program / software remaining lifetime prediction program in the memory 702, in addition to the functions mentioned above, it can also perform other functions, as can be found in the description of the corresponding method embodiments above.

[0083] Furthermore, the embodiments of the present invention do not specifically limit the type of electronic device 700 mentioned. Electronic device 700 can be a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, laptop computer, or other portable electronic device. Exemplary embodiments of portable electronic devices include, but are not limited to, portable electronic devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable electronic device can also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the present invention, electronic device 700 may not be a portable electronic device, but rather a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0084] In another aspect, 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 a method for generating a software remaining lifetime prediction model provided by the methods described above, the method comprising: Acquire failure scene data during software operation and preprocess the failure scene data to obtain training samples; A preset network model is trained based on the training samples to obtain a software remaining lifetime prediction model. The network model includes: An embedded encoding layer is used to encode the training samples to obtain a high-order representation with a unified dimension; A multi-scale graph convolution model is used to convert the higher-order representation into a multi-scale spatial feature representation; A frequency domain window fusion model is used to obtain a high-dimensional feature representation that fuses the frequency domain and time domain based on the multi-scale spatial feature representation. A multi-scale convolutional predictor is used to predict the remaining lifetime of the software based on the high-dimensional features.

[0085] Alternatively, when the computer program is executed by a processor, it is implemented to perform the software remaining lifetime prediction methods provided by the methods described above, the method comprising: Acquire the failure site data to be predicted during software operation, and preprocess the failure site data to be predicted to obtain preprocessed data; The preprocessed data is input into the software remaining lifetime prediction model obtained by the above method to obtain the predicted software remaining lifetime.

[0086] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0087] The generation method, prediction method, and storage medium of the software remaining lifetime prediction model provided by this invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for generating a software remaining lifetime prediction model, characterized in that, include: Acquire failure scene data during software operation and preprocess the failure scene data to obtain training samples; A preset network model is trained based on the training samples to obtain a software remaining lifetime prediction model. The network model includes: An embedded encoding layer is used to encode the training samples to obtain a high-order representation with a unified dimension; A multi-scale graph convolution model is used to convert the higher-order representation into a multi-scale spatial feature representation; A frequency domain window fusion model is used to obtain a high-dimensional feature representation that fuses the frequency domain and time domain based on the multi-scale spatial feature representation. A multi-scale convolutional predictor is used to predict the remaining lifetime of the software based on the high-dimensional features.

2. The method for generating the software remaining lifetime prediction model according to claim 1, characterized in that, The failure site data is preprocessed to obtain training samples, including: The failure site data is cleaned, and the remaining lifetime is labeled on the cleaned data to obtain labeled samples; The labeled samples are normalized to obtain samples with a uniform scale; The uniform-scale samples are divided using a sliding window strategy to obtain training samples.

3. The method for generating the software remaining lifetime prediction model according to claim 1, characterized in that, The conversion of the higher-order representation into a multi-scale spatial feature representation includes: The higher-order representation is subjected to channel separation and feature extraction to obtain the extracted features; The extracted features are captured by multi-scale residual convolution and dynamic graph convolution to obtain the captured features; The captured features are then subjected to a displacement operation to obtain a multi-scale spatial feature representation.

4. The method for generating the software remaining lifetime prediction model according to claim 3, characterized in that, The higher-order representation is subjected to channel separation and feature extraction to obtain extracted features, including: The higher-order representation is subjected to channel separation and feature extraction through conditional positional encoding to obtain extracted features; The conditional location encoding uses a 7×7 depthwise separable convolution.

5. The method for generating the software remaining lifetime prediction model according to claim 1, characterized in that, Each node in the multi-scale graph convolution model retains the same number of connections.

6. The method for generating the software remaining lifetime prediction model according to claim 1, characterized in that, Based on the multi-scale spatial feature representation, a high-dimensional feature representation fused from the frequency and time domains is obtained, including: Frequency domain features are extracted from the multi-scale spatial feature representation based on the encoder and employing a multi-head window self-attention mechanism and Fourier hybridization. Based on the decoder and combined with the masked multi-head window self-attention and cross-attention mechanism, the frequency domain features are fused with the spatiotemporal features in the multi-scale spatial feature representation to obtain a high-dimensional feature representation that fuses the frequency domain and time domain.

7. The method for generating the software remaining lifetime prediction model according to claim 1, characterized in that, Based on the aforementioned high-dimensional features, the remaining software lifetime is predicted, including: The high-dimensional features are mapped to software aging failure time through a projection layer.

8. The method for generating a software remaining lifetime prediction model according to any one of claims 1-7, characterized in that, The training samples are encoded to obtain a high-order representation with a unified dimension, including: The training samples are mapped to fixed-dimensional vector representations, and relative position information is added to each time step. The temporal context is modeled to obtain a high-order representation with a unified dimension.

9. A method for predicting the remaining lifetime of software, characterized in that, include: Acquire the failure site data to be predicted during software operation, and preprocess the failure site data to be predicted to obtain preprocessed data; The preprocessed data is input into the software remaining lifetime prediction model obtained by the method of any one of claims 1-8 to obtain the predicted software remaining lifetime.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for generating the software remaining lifetime prediction model as described in any one of claims 1 to 8, or the steps of the method for predicting the software remaining lifetime as described in claim 9.

Images