Introduction

Information technology is one of the key driving forces behind modern scientific and technological development. It encompasses multiple domains, including computer science, software engineering, communication technologies, and network systems1,2,3. With ongoing social progress and technological advancements, information technology has been widely adopted across various industries, significantly accelerating digital transformation and industrial upgrading4,5. For instance, in the financial sector, information technology has revolutionized payment systems, electronic payment methods such as bank cards, mobile payments, and online banking have become mainstream, greatly improving transaction efficiency and security6,7,8. In manufacturing, applications of information technology such as the Industrial Internet, Internet of Things (IoT), and intelligent manufacturing have substantially enhanced production efficiency and product quality9,10,11. Sensor-based systems enable real-time equipment monitoring and predictive maintenance, reducing failure rates and minimizing downtime. Additionally, intelligent production lines reduce manual operations, increase automation levels, lower costs, and shorten delivery cycles12,13.

However, with the increasing complexity of systems, the explosive growth of data, and the dynamic nature of operating environments, traditional optimization methods for information systems are becoming insufficient to meet the demands for high real-time performance, accuracy, and adaptability14,15,16. The emergence of artificial intelligence (AI), particularly the advancement of deep learning, offers new pathways for upgrading and modernizing information systems17,18,19. Convolutional Neural Network (CNN), a core architecture of deep learning, has played a vital role in the field of computer vision20. Due to their powerful feature extraction capabilities and end-to-end modeling properties, CNN has achieved remarkable success in image recognition, speech processing, and time-series analysis21,22,23. Furthermore, their architecture is well-suited for handling system-level data flows, mining behavioral patterns, and analyzing multi-dimensional metrics, thus providing a strong theoretical foundation for building intelligent information systems and decision support models24.The construction of information systems often requires significant financial and human resources, and it can lead to substantial changes in organizational management concepts and workflows. Without sufficient foresight and compatibility during the system planning stage, and without reserving room for future scalability, organizations may face serious challenges and limitations later on25.

In practical applications, task loads in cloud environments exhibit high dynamism, with occurrences of data loss, sudden peaks, and resource bottlenecks. These conditions make it difficult for traditional resource scheduling methods to balance accuracy, real-time performance, and robustness, leading to degraded system performance, increased scheduling failure rates, and significant fluctuations in quality of service. Although AI methods such as CNN have been widely applied in time-series classification and state recognition tasks, existing studies mostly focus on single-task processing or model optimization. They lack system-level integration and real-time strategy generation methods for large-scale cloud workflows. In summary, current methods have the following research gaps: (1) Insufficient state recognition accuracy under dynamic loads and sudden anomalies. (2) Lack of an end-to-end strategy generation and closed-loop feedback mechanism. (3) Absence of system-level studies that verify model robustness and scheduling optimization effects in actual cloud environments.

To address the above issues, this study proposes a CNN-based decision support model. It realizes the organic integration of system state recognition and strategy generation through an end-to-end mechanism, and constructs a closed-loop feedback optimization framework to achieve resource scheduling optimization and bottleneck mitigation for complex information systems in cloud environments. The main contributions of this paper are as follows: (1) It proposes a system-level integrated CNN-driven decision support method, which unifies state recognition and strategy generation in an end-to-end framework. (2) It verifies the model performance under conditions of dynamic loads, data loss, and sudden peaks, demonstrating its advantages in high accuracy, fast response, and robustness. (3) Through experimental results from Azure VM Workload Traces and Google Cluster Data, it quantitatively demonstrates the model’s improvements in scheduling success rate, strategy response time, and overall system optimization, highlighting the practical application value of the method in cloud workflow optimization.

Literature review

The optimization of information systems and intelligent decision support has become a research hotspot across various fields in recent years. For example, Waqar26 conducted a comprehensive analysis of AI and machine learning methods in the construction engineering sector, highlighting their practical applications, advantages, and limitations. Settembre-Blundo et al.27 developed an information-based theoretical model linking risk management with sustainability management. They also proposed a multidimensional risk assessment framework, which supported analysis across different levels, including business operations, competitive scenarios, and system environments.

CNN has gradually become a vital tool for information system modeling and decision support due to its powerful feature extraction and pattern recognition capabilities. For example, Ehtisham et al.28 proposed an innovative approach that leveraged CNN and image processing techniques to revolutionize the evaluation of wooden structure damage using digital images. Narmadha and Vijayakumar29 introduced hybrid neural network algorithms, combining CNN and Long Short-Term Memory (LSTM), for short-term traffic flow prediction based on multivariate analysis. Nizam et al.30 developed a hybrid end-to-end deep anomaly detection framework capable of accurately identifying anomalies and rare events in sensitive IoT stream data on information systems in real-time or near real-time. Athar et al.31 and Siyal et al.32 focused on deep learning-based models using 1D CNN for early fault detection in machining centres, demonstrating the potential of CNN in predictive maintenance and intelligent system monitoring. Although these studies demonstrate the effectiveness of CNN in different fields, most methods focus on specific tasks or single application scenarios, and lack system-level integration for multi-dimensional state recognition and resource scheduling in complex cloud workflows. Meanwhile, when facing challenges such as dynamic loads, data loss, and sudden peaks, the verification of their robustness and real-time performance remains insufficient. This leaves room for further exploration of end-to-end optimization frameworks for practical applications.

Additional studies have explored workflow scheduling and resource optimization in cloud environments. For example, Chandrasiri & Meedeniya33 proposed an energy-efficient dynamic workflow scheduling method based on deep learning to improve energy efficiency and resource utilization in cloud environments. Aditi et al.34 used CNN to analyze resource allocation bottlenecks in cloud ecosystems, demonstrating the feasibility of deep learning models in complex system optimization. Simaiya et al.35 proposed a hybrid method integrating deep learning and optimization techniques for load balancing and host utilization prediction in cloud environments, further enhancing overall system performance. Although these studies have achieved valuable results in energy efficiency optimization, resource bottleneck identification, and load balancing, most methods remain limited to single optimization objectives and lack an end-to-end integrated framework for state recognition and strategy generation. Meanwhile, the verification of robustness in complex scenarios such as dynamic loads, data loss, or sudden peaks is relatively insufficient, which to a certain extent restricts their application potential in actual large-scale cloud workflows. Table 1 presents the research objects, core technologies, objectives, and limitations of different methods:

Table 1 Literature comparison.
Full size table

In summary, although existing studies have made important progress in information system optimization, intelligent decision support, and task scheduling and load balancing in cloud environments, there are still several obvious limitations: (1) Most studies focus on single tasks or optimization objectives, lacking system-level integrated design for state recognition and resource scheduling. (2) For challenging scenarios such as data loss and sudden loads in complex dynamic environments, the verification of robustness and real-time performance of existing methods is insufficient. (3) End-to-end closed-loop optimization and feedback mechanisms have not been fully explored in relevant studies. To address these shortcomings, the proposed CNN-based decision support model aims to integrate system state recognition and strategy generation through an end-to-end approach and construct a closed-loop optimization framework. This enables resource scheduling optimization and bottleneck mitigation with high accuracy, low latency, and strong robustness in complex cloud workflows, effectively filling the gaps in existing research.

Research methodology

The overall framework and core module design of the model

To realize efficient optimization and intelligent decision-making of information system in dynamic environment, this study designs a decision support model driven by CNN. The decisions in question mainly include three categories: task rescheduling, resource reallocation, and early anomaly warning. These are aimed at improving the overall performance, resource utilization, and stability of information systems. This model forms a complete closed-loop structure with the framework of “data perception-state modeling-policy generation-system feedback”. The specific structure is shown in Fig. 1.

Fig. 1
figure 1

Schematic diagram of the overall architecture of the model.

Full size image

In the model design shown in Fig. 1, the CNN-driven framework is tailored to handle large volumes of heterogeneous data commonly found in information systems. These include structured performance metrics, time-series operational logs, and semi-structured sensor status data. A unified data acquisition and integration module is implemented to standardize these diverse data sources into a consistent input stream. To meet the input format requirements of CNN modeling, a sliding window mechanism is applied to segment the raw time-series data. Suppose the original time series is:

$$\:X=\{{x}_{1},{x}_{2},...,{x}_{T}\}$$
(1)

Every time point \(\:{x}_{t}\in\:{\mathbb{R}}^{d}\) represents a d-dimensional feature vector. X represents the original observation sequence of the system in the time dimension, with a total length of T. This representation organizes the discrete time-series data into a unified format, facilitating subsequent sliding window segmentation and convolutional modeling. On the basis of setting the window length L, the input sample tensor is constructed:

$$\:{\stackrel{\sim}{X}}_{i}=\{{x}_{i},{x}_{i+1},...,{x}_{i+L-1}\}$$
(2)

Time segments with a length of L are formed and used as inputs for network training. The dimension of the input tensor is L×d, where L denotes the number of time steps and d represents the feature dimension. This method preserves local temporal dependency information, enabling the convolutional layer to capture state change patterns in the time series.

Its dimension is \(\:{\mathbb{R}}^{L\times\:d}\), which is used for subsequent network training. In the part of feature modeling, the model adopts multi-layer one-dimensional convolution structure to extract the local change pattern in the running state of the system. Convolution operation can effectively capture the evolution characteristics between successive time slices, and this process is expressed by the equation:

$$\:{h}_{t}=\sigma\:\left(\sum\:_{i=0}^{k-1}\:{W}_{i}^{\left(1\right)}\cdot\:{x}_{t+i}+b\right)$$
(3)

\(\:{h}_{t}\) represents the eigenvector output by convolution operation at time step t, and represents the state mode of local time segment. where k is the width of convolution kernel. σ(⋅) represents the nonlinear activation function. W(1) and b are the parameters of convolution layer. The convolution operation is used to capture local patterns, while the pooling layer is employed for dimensionality reduction and enhancement of salient features. By stacking multiple layers of convolution and pooling, the model can learn multi-scale feature representations of system states. Subsequently, the convolutional features are mapped to a state representation vector, which integrates the time-series information within the entire window and serves as the input for subsequent classification.

$$\:z=\text{R}\text{e}\text{L}\text{U}({W}^{\left(2\right)}\cdot\:h+{b}^{\left(2\right)})$$
(4)

H denotes the feature vector obtained from the convolution-pooling operation of the previous layer, which contains multi-scale local information about the system state. \(\:{W}^{\left(2\right)}\) represents the weight matrix of the fully connected layer, used to map the input feature vector to the state representation space. \(\:{b}^{\left(2\right)}\) denotes the bias vector of the fully connected layer, which adjusts the mapped output and enhances the model flexibility. ReLU(⋅) stands for the rectified linear unit activation function, which introduces non-linearity and enables the model to capture non-linear relationships between complex features. Z represents the state representation vector output by the fully connected layer, serving as the input to the classifier. With a dimension of p, it summarizes the global features of the current system state.

Then, the state vector is input into the Softmax classifier to judge the state category or operation mode of the current system:

$$\:\widehat{y}=\text{S}\text{o}\text{f}\text{t}\text{m}\text{a}\text{x}({W}^{\left(3\right)}\cdot\:z+{b}^{\left(3\right)})$$
(5)

\(\:\widehat{y}\in\:{\mathbb{R}}^{C}\). C indicates the number of categories of system status or events. z is the output vector of the fully connected layer from Eq. (4), which contains the global representation of the system state. \(\:{W}^{\left(3\right)}\) denotes the weight matrix of the classification layer, mapping the state representation vector to the score space of each category. \(\:{b}^{\left(3\right)}\) represents the bias vector of the classification layer, used to adjust the category scores. Softmax(⋅) stands for the multi-class normalization function, which converts the score of each category into a probability distribution, ensuring the sum of probabilities of all categories equals 1.

In the training stage, the model is optimized by cross entropy loss function to minimize the deviation between the prediction label and the real state. However, the model not only outputs the classification results of the system, but also maps the state result \(\:\widehat{y}\) into specific optimization behaviors or control instructions through the strategy mapping mechanism. This policy mapping function is defined as:

$$\:\pi\::{\mathbb{R}}^{C}\to\:\mathcal{A},a=\pi\:\left(\widehat{y}\right)$$
(6)

\(\:\mathcal{A}\) represents the optional action space of the system, and a represents the recommended decision-making behavior at the current moment.

This process demonstrates the model’s complete transformation capability from state recognition to strategy output. To enable closed-loop optimization in real-world systems, the system’s new state, after executing the current strategy, is re-entered into the data acquisition process. This serves as input for subsequent iterations of model training, forming a continuous learning loop. The CNN model consists of two convolutional layers, two pooling layers, two fully connected layers, and a Softmax output layer. The structure is illustrated in Fig. 2. The input layer receives time-series data processed by sliding windows, ensuring each sample contains multi-dimensional features of L consecutive time steps. Two convolutional layers are used to extract low-level and high-level local features respectively: the first convolutional kernel captures the basic patterns of the original input, while the second further abstracts more complex temporal dependencies. The convolution operation can effectively capture the evolutionary features of continuous time segments. A pooling layer is equipped after each convolutional layer. Through max-pooling or average-pooling, it compresses the feature dimension, highlights salient patterns, reduces computational load, and simultaneously improves robustness against local disturbances. Two fully connected layers are responsible for mapping the features after convolution-pooling to the state representation space, integrating multi-scale local features into a global representation to provide input for classification. The Softmax output layer is used to output the probability distribution of the system’s current state or event category, realizing state classification. The classification result is used to evaluate the system’s operating status and serves as the input for the strategy mapping module.

Fig. 2
figure 2

Schematic diagram of CNN model structure.

Full size image

Optimizing strategy generation and feedback mechanism

After completing system state modeling, the model must transform prediction outputs into concrete optimization strategies or scheduling instructions to guide real-time adjustments and resource reallocation in the information system. To achieve this, a strategy mapping mechanism is introduced following the CNN output, establishing a dynamic linkage between “state recognition” and “action decision-making”. Combined with the system feedback loop, this enables a data-driven adaptive optimization cycle. This process is illustrated in Fig. 3.

Fig. 3
figure 3

Closed-loop diagram of optimization strategy generation and system feedback.

Full size image

In Fig. 3, in the policy generation phase, the state classification result \(\:\widehat{y}\in\:{\mathbb{R}}^{C}\) output by the model represents the current operation category or abnormal type of the system. This output is used as the input of the policy generation function π(⋅) and mapped to the executable action space \(\:\mathcal{A}\). The process of policy generation can be formalized as:

$$\:a=\pi\:(\widehat{y},\theta\:),\pi\::{\mathbb{R}}^{C}\to\:\mathcal{A}$$
(7)

where a represents the generated optimization operations (such as task migration, resource allocation, priority rearrangement, etc.), and θ is an adjustable policy parameter. To enhance the practical adaptability of the strategy, the π function not only depends on the classification results, but also can be dynamically adjusted according to the current system context environment (such as load pressure, operating cost, delay index, etc.), which is embodied as follows:

$$\:a=\text{a}\text{r}\text{g}\underset{a\in\:\mathcal{A}}{min}\:{\mathcal{L}}_{opt}(a,\mathcal{E})$$
(8)

\(\:{\mathcal{L}}_{opt}\) represents the optimization objective function, such as minimum delay, maximum throughput and minimum energy consumption. \(\:\mathcal{E}\) represents the description vector of the current system environment state. In the implementation process of the Policy Mapping Module and Decision Execution Block, the predicted state vector is first converted into an action embedding through a fully connected network. The similarity matching is performed with a predefined action prototype library to generate a set of candidate actions. The candidate actions undergo rule-based filtering based on system constraints and load conditions, including resource capacity limits, priority constraints, and migration costs. For cases where multiple optional actions exist for the same state, the module introduces a reinforcement learning-based online fine-tuning strategy to rank and select actions. The reinforcement learning reward is driven by quantitative indicators of policy execution effects (such as scheduling success rate, average response time, and improvement in resource utilization), aiming to enhance the adaptability and stability of policies in dynamic environments. The actions output by the policy mapping module include the recommended operation a, and the action confidence and cost evaluation, providing guidance for subsequent execution.

Subsequently, the decision execution block converts high-level actions into low-level system operations. In the dataset experiments of this study: Task rescheduling is realized by migrating containerized tasks between available nodes based on a priority queue mechanism. Resource reallocation is achieved by dynamically adjusting CPU and memory quotas through the Kubernetes controller. Anomaly warning is completed by triggering threshold alarms and recording them in the monitoring dashboard. This module ensures that policies can actually drive the system to perform optimization operations while maintaining real-time responsiveness.

After execution, the new state of the system re-enters the data collection process and serves as input for the next round of model training and policy fine-tuning, thereby completing the closed-loop cycle of “state recognition – policy generation – action execution – feedback learning”. This mechanism enables the model to possess high-precision state classification capabilities and effectively convert classification results into executable policies. Moreover, it maintains robustness and adaptive optimization capabilities in complex environments such as dynamic loads, data loss, or sudden peaks.

In this process, the system feedback mechanism not only involves collecting post-execution state data, but also incorporates quantitative evaluation metrics of execution effectiveness, such as strategy benefit scores, task completion efficiency, and changes in resource utilization. For instance, if a specific strategy repeatedly results in negative feedback (e.g., persistent resource bottlenecks), the model can automatically update the parameters of the mapping function π(⋅), or trigger a retraining process of the strategy layer.

To enhance long-term stability and adaptability of the system, a dynamic strategy update mechanism is also designed. When the system detects a distribution shift in input data beyond a predefined threshold or a significant decline in strategy effectiveness, the model enters a retraining phase, extracting features from the latest samples to reconstruct the strategy mapping network. This feedback-update interaction ensures the model’s robustness and continuous optimization capability in non-static environments.

The model integration and system deployment scheme

After completing the model design and closed-loop mechanism construction, a key measure of its practical value lies in whether the model can be efficiently deployed into real-world information systems and operate reliably. To this end, a deployable system integration solution is proposed, covering deployment architecture, model integration interfaces, runtime optimization, and system adaptability.

In terms of deployment architecture, the proposed model supports both modular and distributed deployment strategies. Modular deployment suits small- to medium-scale information systems, where the model can be embedded into the system’s existing scheduling or monitoring layer, enabling rapid integration and strategy control. For large-scale and high-concurrency systems, a distributed architecture is recommended: data acquisition and preprocessing, model inference, and strategy execution modules are separately deployed across edge and central nodes. This design balances real-time performance with efficient system resource utilization. For resource-constrained edge devices or embedded systems, the model can be compressed via lightweight techniques such as model pruning and parameter quantization to improve inference speed and reduce hardware dependency.

To facilitate smooth integration with existing systems, a standardized data interface and strategy invocation interface are designed. On the data intake side, the model interfaces with system logs, databases, and sensor platforms through unified data formatting specifications, ensuring consistency and completeness of input. On the output side, the model’s scheduling or control decisions are translated into system-recognizable opcodes or scripts, which are then transmitted via interfaces to the operating system, resource scheduler, or intelligent control platforms to enable automated execution. Additionally, several optimization techniques are employed to improve runtime efficiency. These include limiting convolutional layer depth and parameter volume, using 1D convolutions instead of conventional 2D ones, and applying ReLU activation functions alongside Batch Normalization to reduce computational overhead.

Experimental design and performance evaluation

Datasets collection

To evaluate the applicability and generalizability of the proposed model, two publicly available datasets are selected for experimentation: Azure VM Workload Traces (https://github.com/Azure/AzurePublicDataset) and Google Cluster Data (ClusterData2019) (https://github.com/google/cluster-data). The Azure VM Workload Traces dataset records various indicators from virtual machines, including resource utilization, workload fluctuations, instance running states, and task execution logs. These records reflect the dynamic nature of resource changes in real-world cloud platforms. In contrast, the Google Cluster Data provides multi-dimensional system information, such as node operational status, task scheduling outcomes, priority levels, CPU/memory allocation, and alert logs, making it suitable for constructing system state samples and strategy feedback mechanisms.

To better support the model’s learning of temporal features, the raw data undergoes standardization and sliding window processing. Each input sample is generated using a fixed-length window of L = 30, resulting in a tensor sequence. The input feature dimension is set to d = 12, covering key system resource metrics such as CPU utilization, memory usage, task completion time, and scheduling status. The processed samples are then split into training and testing sets in an 8:2 ratio, which are used for model training and performance evaluation.

Experimental environment

The experiment is conducted on a high-performance computer, and the specific environment settings are shown in Table 2:

Table 2 Experimental environment settings.
Full size table

Parameters setting

To ensure strong model performance while balancing representation capability and inference efficiency, this study carefully configures key parameters related to the convolutional structure, classification layers, and optimization process. The detailed settings are presented in Table 2. The model adopts a 3-layer one-dimensional convolutional structure to accommodate time-series input data. The convolution kernel size is set to 5, with increasing channel numbers across layers to enhance feature abstraction. Each convolution layer is followed by a ReLU activation function and a max pooling layer, which compresses feature dimensions and highlights significant variations. The output is passed through fully connected layers and a Softmax function to perform system state classification. During training, the Adam optimizer is employed to accelerate convergence, together with a cross-entropy loss function for multi-class supervised learning. Additionally, Dropout and BatchNorm are used jointly to effectively mitigate overfitting. Parameter details are summarized in Table 3.

Table 3 Model structure and training parameter configuration table.
Full size table

To ensure the fairness of comparative experiments, this study also describes the main parameters and training processes of the four baseline models: Support Vector Machine (SVM): The Radial Basis Function (RBF) kernel is adopted, and the one-vs-one strategy is used for multi-class classification. The regularization parameter (C = 1.0) and the kernel coefficient (γ = 0.1). Input features are standardized before training, and parameters are optimized through cross-validation. Long Short-Term Memory (LSTM) Network: The network consists of three layers, with 64 hidden units in each layer. The time step is set to 30, and the output is connected to a fully connected layer for state classification. Training uses the Adam optimizer and cross-entropy loss function, with a batch size of 64 and 50 training epochs. Meanwhile, Dropout (p = 0.3) is applied to prevent overfitting. Random Forest: It is composed of 100 decision trees, where each tree has a maximum depth of 10 and the maximum number of features is the square root of the total number of features. Out-of-bag error is used for performance verification. Logistic Regression (LR): A Softmax output layer is used to realize multi-class classification. The regularization type is L2, the regularization coefficient (C = 1.0), and the solution method is the Newton conjugate gradient method. Input features are standardized, and regularization parameters are optimized through cross-validation.

Performance evaluation

To evaluate the classification performance of the proposed CNN model in the context of information system state recognition, this study conducts comparative experiments against several representative baseline methods. These include Support Vector Machine (SVM), Long Short-Term Memory (LSTM) network, Random Forest (RF), and Logistic Regression (LR), as well as the recent state-of-the-art methods in References36,37. Specifically, Li et al.36 proposed a cloud task scheduling framework based on multiple deep neural networks, while Patel & Kushwaha37 introduced a hybrid CNN-LSTM model for server load prediction. The evaluation metrics used in the comparison are: Accuracy, Precision, Recall, F1-score, Specificity, Area Under the Curve (AUC). The calculation equations for these metrics are as follows:

$$\:\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}=\frac{TP+TN}{TP+TN+FP+FN}$$
(9)
$$\:\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}=\frac{TP}{TP+FP}$$
(10)
$$\:\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}=\frac{TP}{TP+FN}$$
(11)
$$\:\text{F}1-\text{s}\text{c}\text{o}\text{r}\text{e}=\frac{2\times\:\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}\times\:\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}{\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}+\text{R}\text{e}\text{c}\text{a}\text{l}\text{l}}$$
(12)
$$\:\text{Specificity}=\frac{TN}{TN+FP}$$
(13)
$$\:\text{A}\text{U}\text{C}={\int\:}_{0}^{1}\:TPR\left(FPR\right)d\left(\text{F}\text{P}\text{R}\right)$$
(14)

True Positive (TP) refers to the number of samples that are actually positive and correctly classified as positive. True Negative (TN) represents the number of samples that are actually negative and correctly classified as negative. False Positive (FP) and False Negative (FN) indicate the number of incorrectly classified positive and negative samples, respectively. The output categories of the information system state recognition task are clearly divided into five categories, which correspond to the typical operating states or abnormal conditions of the system respectively. Each category is supported by sufficient samples in the training and test data to ensure that the model’s learning and evaluation are representative, as shown in Table 4:

Table 4 Output category division.
Full size table

The test results are illustrated in Fig. 4.

Fig. 4
figure 4

Performance comparison of state classification.

Full size image

In Fig. 4, the proposed model achieves an accuracy of 96.2%, indicating a strong overall classification capability. The F1-score reaches 0.961, reflecting a well-balanced performance between precision and recall. The specificity is 95.9%, effectively reducing false alarms by avoiding the misclassification of normal states as anomalies. The AUC reaches 0.978, demonstrating the model’s robustness and generalizability in multi-class state recognition tasks. In comparison, traditional shallow models such as LR and SVM perform reasonably well on structured data but exhibit limited generalization when faced with complex interdependencies between state features. While LSTM offers advantages in temporal modeling, it suffers from slower training and higher sensitivity to parameter settings, making it less effective than CNN in local feature extraction and structural alignment. A comparison is made with the methods in References36,37. The multi-deep neural network task scheduling framework proposed by Li et al.36 has achieved good results in optimizing resource utilization in cloud environments, but it mainly focuses on energy efficiency and scheduling performance, with limited verification on multi-dimensional state recognition and end-to-end closed-loop optimization. The hybrid CNN-LSTM model by Patel & Kushwaha (2022) demonstrates high accuracy in server load prediction, yet it lacks system-level integration in the closed loop of strategy generation and execution. Therefore, the proposed model outperforms the methods in References36,37 significantly in Accuracy, F1-score, Specificity, and AUC, demonstrating advantages in state recognition accuracy, strategy execution reliability, and adaptability to dynamic environments.

Furthermore, a set of simulated tasks under representative scenarios is designed to evaluate the model’s response time performance. These tasks include event detection, state transition handling, resource rescheduling, and strategy rollback. Event detection involves monitoring system logs and state data to identify abnormal or delayed events. State transition processing refers to recognizing and adjusting changes in system states. Resource rescheduling involves migrating containerized tasks across multiple nodes to alleviate loads. Policy rollback refers to rolling back failed or inefficiently executed policies and re-scheduling them. Each task is executed via system interfaces, with the response time measured from state input to action execution. The results are shown in Fig. 5.

Fig. 5
figure 5

Evaluation of strategy response efficiency.

Full size image

In Fig. 5, the proposed model maintains low processing latency across various types of system tasks, with an average response time of 135.5 ms and the maximum latency not exceeding 203 ms, fully meeting the sub-second performance requirements of information systems. Among them, the “Strategy Rollback and Recovery” task had the shortest latency, at only 125 ms, indicating the model’s faster judgment when recalling previously known states. The relatively longer response time for the “Resource Preemption Handling” task is mainly due to the need for strategy weight reconstruction and multi-node interaction. By leveraging the Softmax output mapping mechanism and a lightweight network architecture, the model significantly reduces the time overhead of inference and strategy generation. Compared to traditional rule-based or statistical response mechanisms, the proposed model learns a nonlinear mapping between system states and optimal actions through end-to-end training, thereby offering stronger generalization and real-time adaptability.

To further evaluate the model’s optimization effectiveness in real-world operational environments, this study compares system states before and after the model’s intervention. The results are presented in Fig. 6.

Fig. 6
figure 6

System resource optimization effect.

Full size image

In Fig. 6, the system resource utilization significantly decreases after the deployment of the model, effectively relieving operational pressure. Specifically, CPU usage drops from 86.5% to 67.3%, and memory usage declines from 91.2% to 70.5%, with an average reduction of over 20%. This indicates that the model possesses strong real-time control capabilities in identifying system states and generating corresponding strategies. Additionally, task queue length is reduced by nearly 48%, suggesting that the model effectively alleviate system congestion and improved task processing efficiency. Most notably, the scheduling failure rate decreases from 6.1% to 2.2%, an improvement of 63.9%, demonstrating the model’s high stability and practical applicability in strategy generation.

To further validate the adaptability and robustness of the proposed model under different operational scenarios, multiple test environments are constructed. These include scenarios such as sudden anomaly surges, missing data, and abrupt high-load conditions, simulating fluctuations in real-world systems under non-ideal circumstances. The evaluation results are shown in Fig. 7.

Fig. 7
figure 7

Multi-scene robustness evaluation results.

Full size image

In Fig. 7, the model continues to demonstrate strong performance under various disturbance scenarios. During instances of concentrated anomalies, the accuracy remains at 93.4%, with a strategy success rate of 95.2%, indicating the model’s capability to identify abnormal behavior and output effective response strategies. In the scenarios of “data missing segments” and “sudden high load”, although accuracy slightly decreased (to 90.1% and 88.7%, respectively), the model still maintains a strategy success rate exceeding 89% and automatically triggers the retraining mechanism to adapt to the new data distribution. These results demonstrate that the proposed CNN-based decision-making model achieves high-precision recognition. It also maintains robustness and adaptability in the presence of input noise, partial data loss, or abrupt workload fluctuations, proving its feasibility for deployment in real-world environments.

Discussion

The above research demonstrates that the proposed CNN-driven optimization and decision support model exhibits superior overall performance in system state recognition, response efficiency, resource scheduling optimization, and multi-scenario adaptability. Leveraging its efficient capability to extract local patterns in time series data, the CNN model significantly outperforms traditional models such as SVM and RF. As noted by Yu et al.38, the advantages of convolutional structures in modeling structured operational data lied not only in accuracy but also in stable discrimination across high-dimensional state spaces. In terms of strategy response, the model’s average response time is considerably shorter than that of traditional rule-based methods, confirming the feasibility of end-to-end neural network structures in time-sensitive system tasks. As Vitente et al.39 suggested, lightweight neural network model could partially replace expert-rule-based control modules in task scheduling and anomaly feedback, thereby enhancing system responsiveness and automation. The resource optimization experiments also indicate the model’s effectiveness in reducing system load. As emphasized in the study by Han et al.40, system-level strategy generation mechanisms should balance local peak load reduction with global operational stability. In such tasks, deep learning models can improve decision-making efficiency by learning latent correlations.

Regarding the requirements for model interpretability and policy transparency, this study has incorporated a policy mapping mechanism and a closed-loop feedback process into its design. This enables each predicted system state to be mapped to specific optimization actions, which can then be executed through system interfaces. The predicted state vector is converted into an action embedding via a fully connected layer and matched against a predefined action space. Simultaneously, rule-based filtering and policy effect feedback are integrated to ensure that recommended actions align with current load and resource constraints and allow for traceability of their decision sources. The closed-loop feedback module further collects post-execution state data and policy effectiveness metrics, such as task completion efficiency, changes in resource utilization, and policy benefit scores. These metrics provide a basis for dynamically adjusting the policy mapping function. This design enhances the transparency of model decisions and strengthens its robustness and adaptability in non-stationary environments. Through the analysis of quantifiable indicators, researchers and system administrators can understand the logic behind the model’s recommended policies, thereby making reasonable evaluations and optimizations of system behavior.

Although this study mainly conducts experimental verification on the data and tasks of specific information systems, the CNN-driven closed-loop optimization and decision support framework exhibits strong scalability. Future work can extend this model to other types of computing environments, such as cloud computing platforms, edge computing nodes, or industrial control systems. This will verify its adaptability and robustness in different scenarios, and further enhance the model’s wide application value.

Conclusion

Research contribution

This study proposes an intelligent decision support model driven by CNN to address state awareness and optimization decision-making in information systems. By constructing a lightweight convolutional architecture combined with a strategy output module, the model achieves high-precision classification of system states and dynamic policy inference for resource optimization. The integration of a system feedback mechanism and a retraining module further enhances the model’s adaptability and stability in complex environments. Comprehensive performance evaluations are conducted using two representative public datasets, Azure VM Workload Traces and Google Cluster Data, covering aspects such as state recognition accuracy, decision response efficiency, system resource load optimization, and multi-scenario robustness. The results demonstrate that the proposed model outperforms benchmark methods across all metrics and possesses strong potential for practical deployment in real-world information systems.

Future works and research limitations

Despite the preliminary achievements in methodological design and empirical results, this study still has several limitations. For instance, the experimental scenarios are currently limited to single-type task cluster environments. In future work, it is necessary to extend the evaluation to real-world systems with heterogeneous resources and diverse workloads to further validate the model’s adaptability. Additionally, the current strategy generation mechanism is implemented through static classification mappings, without incorporating reinforcement learning to continuously improve policy quality. Future research could explore integrating deep reinforcement learning algorithms to enhance the model’s intelligence. Moreover, the input states in this study are modeled using fixed-length temporal windows, which may overlook long-term dependencies or multi-scale state variations. Introducing Transformer architectures or hybrid attention mechanisms could be a promising direction to strengthen the model’s temporal awareness and dynamic modeling capabilities.