Intelligent energy management software quality adaptive assurance method based on deep learning

Abstract

This invention discloses a deep learning-based adaptive quality assurance method for intelligent energy management software, relating to the interdisciplinary fields of intelligent energy management and software quality assurance. The method comprises five steps: S1, collecting multimodal operational quality data on system resources, software services, and energy management operations via probes and APIs; S2, quantitatively assessing software health and risk through deep time-series feature extraction; S3, generating adaptive optimization decisions such as resource allocation based on deep reinforcement learning; S4, executing decisions and updating the model through closed-loop feedback; and S5, constructing a dynamic knowledge graph and mining association rules to optimize decisions. This invention addresses the problem of lagging quality monitoring through multimodal data collection and deep model quantitative assessment; it also utilizes reinforcement learning to achieve multi-dimensional adaptive optimization, balancing software quality and energy consumption; and it combines knowledge graph rule injection to improve decision robustness and adaptability, reducing operational costs.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of intelligent energy management and software quality assurance, specifically to an adaptive assurance method for intelligent energy management software quality based on deep learning. Background Technology

[0002] With the deep integration of industrial intelligence and energy decarbonization, intelligent energy management software has become a core support for achieving efficient energy scheduling and reducing energy consumption per unit product in industrial production, building parks, and other scenarios. This type of software not only needs to ensure its own operational stability (such as meeting key indicators like API call response latency and transaction throughput), but also needs to accurately adapt to dynamic business scenarios such as energy load fluctuations and changes in the operating conditions of energy-consuming equipment, achieving synergistic optimization of software operation quality and energy utilization efficiency.

[0003] However, existing quality assurance methods for smart energy management software still have many technical limitations and are difficult to meet the actual needs of complex scenarios:

[0004] First, data collection is limited in scope and quality assessment lacks quantitative basis. Traditional methods often focus on system resource data such as CPU utilization and memory usage, or only on indicators such as API call success rate at the software service level. They fail to incorporate core energy management business data (such as real-time equipment power, load forecast deviation rate, and unit product energy consumption) into the monitoring system, resulting in an inability to comprehensively reflect the correlation between software operating status and business efficiency. Simultaneously, quality assessment often relies on manual experience or simple threshold alarms, lacking in-depth fusion analysis of multi-dimensional data. This makes it difficult to quantify the overall health of the software, and risk identification often lags behind the occurrence of faults, easily leading to problems such as energy dispatching errors and equipment malfunctions.

[0005] Secondly, the optimization decisions lack specificity and fail to consider both quality and energy consumption objectives. Existing adaptive optimization solutions are mostly limited to static adjustments of basic resources such as JVM memory and thread pools, failing to deeply integrate with energy management business scenarios and thus unable to dynamically optimize management strategies based on business indicators such as load forecast accuracy and energy consumption changes. Furthermore, optimization objectives often focus solely on software operational stability, neglecting energy consumption optimization in decision-making, leading to an imbalance of prioritizing quality assurance over energy efficiency improvement, making it difficult to achieve a balance between technological value and economic and environmental benefits.

[0006] Third, existing methods suffer from poor adaptability to different scenarios and a lack of experience reuse capabilities. Their optimization logic largely relies on preset rules or fixed model parameters. When scenarios change, such as the addition or removal of energy-consuming equipment types or changes in load fluctuation patterns, these preset rules easily become ineffective, requiring frequent manual parameter adjustments to adapt to the new scenarios, resulting in high operation and maintenance costs. Furthermore, the experience accumulated during historical operation regarding the correlation between "quality risks and optimization actions" has not been effectively preserved and reused. The decision-making process often involves repeated and ineffective explorations, leading to low optimization efficiency, insufficient robustness, and difficulty in adapting to energy management needs of different scales and operating conditions.

[0007] In summary, existing technologies have significant shortcomings in terms of comprehensive data collection, accurate assessment, collaborative optimization, and scenario adaptability. There is an urgent need for an adaptive quality assurance method for intelligent energy management software that can achieve multi-dimensional data fusion perception, accurate quality assessment, collaborative optimization decision-making, and experience-driven evolution. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a deep learning-based adaptive quality assurance method for intelligent energy management software. This invention solves the problem of lagging quality monitoring through multimodal data acquisition and deep model quantitative evaluation; simultaneously, it utilizes reinforcement learning to achieve multi-dimensional adaptive optimization, balancing software quality and energy consumption; and by incorporating knowledge graph rule injection, it improves decision robustness and adaptability, and reduces operation and maintenance costs.

[0009] To address the aforementioned technical problems, this invention provides the following technical solution: Firstly, a deep learning-based adaptive quality assurance method for intelligent energy management software. This method constructs a closed-loop quality assurance system through five core steps, the specific implementation process of which is as follows:

[0010] Step S1: Perception and Fusion Acquisition of Multimodal Quality Data in Operational State

[0011] This step establishes a low-intrusion, high-coverage multimodal data acquisition mechanism to provide a comprehensive and reliable data foundation for subsequent quality assessment.

[0012] Data Acquisition Deployment Scheme: Deploy a lightweight acquisition probe based on eBPF technology in the host environment of the energy management software instance. This probe does not require modification of the software source code, consumes ≤3% of CPU resources and ≤100MB of memory resources, and achieves non-intrusive data acquisition. At the same time, it calls the monitoring API natively provided by the software instance, adopting a parallel acquisition mode of probe and API to ensure both the comprehensiveness and real-time performance of data acquisition.

[0013] Multimodal data definition and acquisition specifications:

[0014] System resource modal data: CPU core utilization (sampling frequency 1 time / second), memory usage (MB) and garbage collection frequency (times / minute, sampling frequency 1 time / 10 seconds), disk I / O throughput (MB / s, sampling frequency 1 time / 2 seconds), and network inbound / outbound bandwidth utilization (sampling frequency 1 time / 2 seconds) are collected to form a time series. , , , .

[0015] Software service modal data includes: API call chain response latency (ms) and success rate (percentage) between microservices, message queue backlog depth (number of messages), active database connection pool connections (number of connections), and key business transaction throughput (number of transactions per second). The data is collected once every 5 seconds to form a time-series sequence. , , , .

[0016] Energy management business modal data includes: real-time power of energy-consuming equipment (unit: kW, collection frequency: once / 10 seconds), unit product energy consumption (calculated in real-time based on equipment power and output data, unit: kWh / unit, collection frequency: once / 30 seconds), load forecast deviation rate (the ratio of the difference between the forecast value and the actual load to the actual load, collection frequency: once / 30 seconds), and energy-saving dispatch command execution success rate (percentage, collection frequency: once / 30 seconds), forming a time series. , , , .

[0017] Data preprocessing process: The collected raw data is processed in three steps: outlier cleaning, time alignment, and normalization. The 3σ criterion is used to remove outlier data that exceed the mean ± 3 times the standard deviation. Using 1 second as a unified timestamp, linear interpolation is used to complete the data at different collection frequencies to ensure consistency in the time dimension. Min-max normalization is used to map all data to the [0, 1] interval to eliminate dimensional differences.

[0018] Step S2: Quantitative assessment of quality status based on deep temporal feature extraction

[0019] This step utilizes a customized neural network model to achieve feature fusion and accurate quantification of quality status in multimodal data, solving the technical problems of fragmented single-modal features and difficulty in quantifying quality status.

[0020] Data organization method: The preprocessed multimodal data is reorganized into a multi-channel time series tensor according to a fixed time window (window length 30 seconds, sliding step size 30 seconds). The tensor dimension is [time step size (30) × number of modal channels (12)], of which there are 4 channels for system resource mode, 4 channels for software service mode, and 4 channels for energy business mode.

[0021] Neural network model for quality status assessment:

[0022] The overall model architecture includes three parallel feature extraction branches, a cross-modal feature fusion layer, and a joint output layer for regression and classification, and adopts an end-to-end training mode.

[0023] Feature extraction branch design:

[0024] The first branch (system resource modality processing) adopts a 1D-CNN+GRU combined structure. The 1D-CNN layer contains 3 convolutional blocks (each convolutional block contains 32 3×1 convolutional kernels, stride 1, padding=same), used to extract local dependency features of data (such as the short-term coupling relationship between CPU utilization and memory usage); the GRU layer has 2 layers, each with 64 hidden units, used to capture long-term periodic patterns (such as the resource usage pattern during daily peak hours).

[0025] The second branch (software service modality processing) adopts a Transformer encoder based on a multi-head attention mechanism (4 attention heads, 64 hidden layer dimensions, and 256 feedforward network dimensions). By calculating the attention weights of different microservice call chain characteristics, it accurately captures the correlation effects between services (such as the amplification effect of insufficient database connections on API response latency).

[0026] The third branch (energy business modal processing) adopts a 3-layer fully connected network with residual connections, with 128, 64 and 32 neurons in each layer, respectively. The activation function is ReLU, which is used to extract static performance benchmarks (such as the industry benchmark value of energy consumption per unit product) and dynamic trends (such as the cumulative growth characteristics of load forecast deviation).

[0027] Feature fusion layer: Employs a multi-head cross-modal attention mechanism to calculate the mutual attention weights of the three branch features (e.g., when API response delays are abnormal, the weight ratio of the software service modal features is automatically increased), and merges the 64-dimensional feature vectors output by the branches into a unified 64-dimensional comprehensive quality feature vector.

[0028] Joint Output Layer: The regression head outputs the overall quality score through a single fully connected network. (The score ranges from 1 to 10, with higher scores indicating better software health); the classification header is output through a multi-label classification network. =5-dimensional risk vector (These correspond to memory risk, network transmission risk, service call risk, prediction failure risk, and instruction execution risk, respectively. The values ​​for each dimension are risk confidence levels, ranging from [0,1]).

[0029] Model training process:

[0030] Training dataset construction: 1,051,200 time window samples were extracted from the historical operation logs of an industrial energy management system over one year. Three domain experts with more than five years of relevant experience were invited to label each sample with a comprehensive quality level label, combining system alarm records and fault handling reports. (1-10 points) and risk type label (A 5-dimensional binary vector, where 1 indicates the existence of a corresponding risk).

[0031] Loss function design: A composite loss function is used to balance the regression and classification tasks. The formula is as follows:

[0032] ,in, Mean squared error loss (MSE). =0.5, For multi-label cross-entropy loss, =0.3, It is an L1 norm, used to implement sparsity constraints on the risk vector (highlighting the main risks). =0.2.

[0033] Training parameter configuration: The Adam optimizer is used (initial learning rate 1e-4, decaying by 10% every 20 rounds), batch size 64, iterative training for 100 rounds, training is stopped when the validation set loss is below 0.05 for 3 consecutive rounds, the final model has an overall quality score prediction error of ≤0.3 on the test set, and a risk identification accuracy of ≥92%.

[0034] Step S3: Adaptive Optimization Decision Generation Based on Deep Reinforcement Learning

[0035] This step takes the quality assessment results as input and uses a reinforcement learning agent to generate adaptive optimization decisions, thus solving the technical problem of multi-objective optimization decision-making in dynamic environments.

[0036] Markov Decision Process (MDP) definition: quintuple ( , , , , ):

[0037] state space : Based on the overall quality score With risk vector The resulting structure is formed by piecing together elements, resulting in a dimension of 1 + 5 = 6. = .

[0038] Action Space A: Hybrid action space, containing three categories and a total of eight specific optimized actions:

[0039] Software resource configuration actions: 1. Increase JVM heap memory by 10%; 2. Decrease JVM heap memory by 10%; 3. Adjust the core thread count of the thread pool by ±2; 4. Adjust the maximum number of database connection pool connections by ±5 (one of these actions will be executed at a time).

[0040] Data processing steps: 1. Switch to stream processing mode; 2. Switch to micro-batch processing mode; 3. Enable data downsampling (downsampling rate is dynamically adjusted according to the current computing load, ranging from 30% to 70%); 4. Adjust the parallelism of real-time computing tasks by ±1 (one task is selected to be executed at the same time).

[0041] Energy management strategy actions: 1. Switch the load forecasting model to an online reinforcement learning model; 2. Use a fast heuristic algorithm for degraded energy-saving scheduling (select one or not to execute based on triggering conditions).

[0042] State transition probability The simulation is determined by the dynamic characteristics of the actual operating environment of the energy management software. It is conducted by building a simulation test platform that is consistent with the real environment (including simulation of 200+ energy-consuming devices and deployment of 12 microservices).

[0043] reward function : Using the definition in step S4 Used to balance the goals of improving software quality and reducing energy consumption; discount factor =0.9, used to balance immediate rewards and long-term rewards.

[0044] Construction and training of reinforcement learning decision-making agents:

[0045] Intelligent agent architecture: based on the actor-critic framework.

[0046] The actor network consists of one shared feature extraction layer (64 neurons, ReLU activation) and three parallel action output heads, which output the probability distribution of discrete actions (such as model switching actions) and the parameter values ​​of continuous actions (such as resource adjustment magnitude).

[0047] The critic network employs a 2-layer fully connected network (128 neurons → 64 neurons, ReLU activation) to output the long-term expected reward of state-action pairs, which guides the policy updates of the actor network.

[0048] Training Algorithm: The Proximal Policy Optimization (PPO) algorithm is used to avoid training instability by constraining the policy update magnitude. The objective function is:

[0049] ,in, For actor network parameters, This is the estimated value of the dominance function. =0.2 is a hyperparameter.

[0050] Training process: Ten typical scenarios (normal load, high load, data burst, equipment failure, etc.) are simulated in the simulation test platform, allowing the agent to interact with the environment for 5 million steps. A model snapshot is saved every 100,000 steps, and the network parameters are updated every 1,000 steps by sampling experience data. The final model has an average reward value ≥1.1 and a decision response time ≤200ms in the test scenario.

[0051] Step S4: Decision Execution and Closed-Loop Feedback Learning

[0052] This step enables online iterative optimization of the agent through decision-making, effect evaluation, and experience storage, thus constructing a closed-loop mechanism of decision-making, execution, and feedback.

[0053] Decision execution mechanism: The decision generation engine uses standardized interfaces to execute optimization actions. Distributed to target objects: JVM heap memory and thread pool parameter adjustments are executed via JMX interface; data processing mode switching is implemented through the RESTful API of the software configuration center; prediction model and scheduling algorithm switching are implemented through the service routing interface of the microservice registry; all actions include legality verification (such as resource adjustments not exceeding hardware limits) and retry mechanism (retrying after 1 second after execution failure, up to 3 times).

[0054] Effect observation and data collection: Setting the observation period =60 seconds (to ensure the optimization actions are fully effective). After the observation period, repeat step S1 for data acquisition and step S2 for quality assessment to obtain the new status. New Comprehensive Quality Score New unit product energy consumption .

[0055] Reward calculation: The reward function balances software quality improvement and energy consumption reduction, and the formula is as follows:

[0056] ,in, =0.6 (Software quality weight) =0.4 (Energy optimization weight). = - (The decrease in unit product energy consumption during the observation period; if energy consumption increases, then...) If the value is negative, the reward value will be reduced accordingly.

[0057] Experience replay and model update: The experience quadruple ( , , , The system stores 100,000 experience replay data points in a buffer and employs a random sampling strategy (to avoid sample correlation). After every 10 decision-making processes, 64 experience data points are sampled. The parameters of the Actor network and Critic network are updated using gradient descent to ensure that the agent can continuously adapt to environmental changes.

[0058] Step S5: Dynamic Construction and Evolutionary Reasoning of Quality Knowledge Graph

[0059] This step uses knowledge graphs to achieve the structured accumulation and reuse of historical experience, solving the technical problems of fragmented optimization experience and low decision-making efficiency.

[0060] Quality Knowledge Graph Construction: Storage medium: Neo4j graph database is used, and the relationships between graph nodes and edges are defined as follows:

[0061] Node types: Software component nodes (e.g., "Load Prediction Microservice", "Database Service", attributes: name, version, deployment node), hardware resource nodes (e.g., "Industrial Server A", "Edge Gateway", attributes: model, number of CPU cores, memory capacity), quality indicator nodes (e.g., "API Response Latency", "Overall Quality Score", attributes: name, threshold range, data type), optimization action nodes (e.g., "Increase JVM Heap Memory", "Switch Micro-Batch Processing Mode", attributes: action type, execution parameters, effective time), energy consumption indicator nodes (e.g., "Unit Product Energy Consumption", "Load Prediction Deviation Rate", attributes: name, baseline value, calculation method).

[0062] Edge relationship types: "Running on" (software component → hardware resource), "Affecting" (hardware resource → quality indicator), "Causing" (quality indicator → risk type), "Mitigating" (optimization action → risk type), "Optimizing for" (optimization action → energy consumption indicator).

[0063] Dynamic updates to the graph: After each closed-loop feedback is completed (step S4), the current state is automatically updated. ,action ,award The software configuration context (such as current JVM parameters and data processing mode) is converted into graph node and edge relationships and written into the graph database to achieve incremental updates of the graph.

[0064] Relational reasoning based on graph neural networks:

[0065] Node feature initialization: Generate initial feature vectors based on node type and attributes. For example, hardware resource nodes use “normalized value of CPU core count + normalized value of memory capacity + number of deployed software components” to form a 3-dimensional initial feature, while quality indicator nodes use “current value + upper threshold + lower threshold” to form a 3-dimensional initial feature. All initial features are mapped to the [0,1] interval.

[0066] Message passing and feature aggregation: A graph attention network (GAT) is used for three rounds of message passing. Each GAT layer contains 64 hidden units. By calculating the attention coefficients between nodes (such as the correlation strength between the "memory risk" node and the "increase JVM heap memory" action node), the information of neighboring nodes is weighted and aggregated to update the 64-dimensional embedding vector of each node.

[0067] Link prediction and rule extraction: Calculating the probability of the existence of a "mitigation" relationship between quality risk indicator nodes and optimization action nodes. (Calculated via the inner product of node embedding vectors) >0.85 (preset threshold) The association relationship with a confidence level of 0.85 was extracted as a valid rule, and the rule format is: IF [risk mode condition] THEN [suggested action (confidence level: 0.85)] For example, “IF [Memory risk confidence > 0.7 and remaining physical memory > 5GB] THEN [Increase JVM heap memory by 10% (confidence: 0.93)]”, the extracted rules are stored in the rule base.

[0068] Rule Injection and Policy Evolution: Effective rules from the rule base are injected into the reinforcement learning agent through the following two methods to improve decision-making efficiency and reliability:

[0069] Action masking mechanism: When the actor network outputs action probabilities, actions that are explicitly prohibited by the rules (such as "increase JVM heap memory" when physical memory is insufficient) have their probabilities set to zero; actions that are strongly recommended by the rules (confidence > 0.9) have their probabilities weighted and amplified by 2 times to reduce invalid exploration.

[0070] Internal reward mechanism: setting rule reward coefficients =0.2, when the agent's decision action conforms to the valid rules of the current state, an additional [value] is given. Intrinsic rewards; conversely, giving The penalty is adjusted, and the total reward function is modified as follows: .

[0071] Secondly, a deep learning-based intelligent energy management software quality adaptive assurance system is provided. This system serves as the hardware and software implementation platform for the aforementioned method, ensuring its stable and efficient operation. The system architecture is as follows:

[0072] 1. Data Sensing and Acquisition Module

[0073] Deployment location: The host environment of the energy management software instance (physical server, virtual machine, container platform);

[0074] Hardware components: Lightweight acquisition probe based on eBPF technology (supports x86 / ARM dual architecture, power consumption ≤2W, operating temperature -20℃~60℃).

[0075] Software functions: Enables parallel acquisition, preprocessing (outlier cleaning, time alignment, normalization), and encrypted transmission (using AES-128 encryption) of multimodal data.

[0076] Communication interface: Supports MQTT / HTTP protocol, communicates with cloud intelligent analysis platform, transmission latency ≤100ms, supports batch data upload and breakpoint resume.

[0077] 2. Cloud-based intelligent analysis platform

[0078] Deployed on public or private cloud servers, it adopts a distributed architecture design, supports horizontal scaling, and includes three core engines:

[0079] Quality assessment engine: Built-in trained quality status assessment neural network model, receives preprocessed data uploaded by data perception and acquisition module, outputs comprehensive quality score and risk vector, single data processing latency ≤500ms;

[0080] Decision generation engine: It has a built-in reinforcement learning decision agent based on the actor-critic framework, loads the trained network parameters (weight file format is .h5), receives the output of the quality evaluation engine, and generates adaptive optimization decisions with a decision generation latency of ≤200ms;

[0081] Knowledge Graph Engine: Includes Neo4j graph database and GAT inference model, responsible for storing, dynamically updating, learning node embeddings and mining association rules of quality knowledge graph, supporting CRUD operations and batch injection of rule base, with inference latency ≤1 second.

[0082] 3. Strategy Execution and Feedback Module

[0083] Core functions: Receive optimization action instructions from the decision generation engine and distribute them to the energy management software instance or its host environment for execution via a standardized interface; collect operational data (status) after decision execution. Energy consumption data The feedback is sent to the cloud-based intelligent analysis platform, forming a closed-loop feedback circuit;

[0084] Reliability design: Supports instruction execution status verification (success / failure feedback), failure retry (up to 3 times), and configuration rollback (restoring to the configuration before the decision when execution fails) to ensure stable system operation.

[0085] Thirdly, a computer-readable storage medium storing a computer program thereon is provided. The storage medium is a non-volatile storage medium, including a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a hard disk, an optical disc (CD-ROM, DVD-ROM), etc. The computer program is stored thereon. When the computer program is executed by a processor, it implements all the steps of the above-mentioned deep learning-based intelligent energy management software quality adaptive assurance method, and supports multiple operating systems such as Windows, Linux, and Unix.

[0086] Compared with existing technologies, this deep learning-based adaptive quality assurance method for intelligent energy management software has the following advantages:

[0087] I. This invention addresses the technical problems of traditional energy management software quality monitoring, such as single data collection dimensions, vague quality status assessment without quantitative basis, and lagging risk identification. It utilizes a deep neural network model with three dedicated parallel feature extraction branches, combined with cross-modal attention weight fusion technology, to generate accurate comprehensive quality scores and multi-dimensional risk vectors. This solves the technical problems of single data collection dimensions, vague quality status assessment without quantitative basis, and lagging risk identification in traditional energy management software quality monitoring. It achieves accurate quantitative assessment and early warning of the quality status of the entire software operation chain, effectively reducing energy dispatching errors and equipment malfunctions caused by quality deterioration, and ensuring the continuous and stable operation of the energy management system.

[0088] II. This invention constructs a hybrid action space encompassing three types of actions: software resource configuration, data processing flow, and energy management strategy. It integrates a reward function that incorporates both the improvement in overall quality score and the reduction in energy consumption per unit product. Utilizing a deep reinforcement learning decision-making agent based on an actor-critic framework, it generates adaptive optimization decisions and continuously updates the agent's policy and value network parameters through a closed-loop feedback mechanism. This solves the technical problems of traditional software optimization schemes, which are limited to single resource adjustments, lack deep integration with energy management operations, and fail to balance software performance and energy efficiency. It achieves synergistic optimization of high-quality software operation and efficient energy utilization. While ensuring that core quality indicators such as API response latency and transaction throughput meet standards, it significantly reduces energy consumption per unit product, enhancing the economic value and environmental benefits of the energy management system.

[0089] Third, this invention stores historical and real-time states, actions, rewards, and software configuration context information in a graph structure, constructs and dynamically updates a software quality-configuration-performance knowledge graph, and uses graph neural networks for multi-round message passing and relational reasoning to mine strong correlation rules between "quality risk patterns and effective optimization actions." These rules are then injected into a deep reinforcement learning decision-making agent through action masking and intrinsic reward correction. This solves the technical shortcomings of traditional adaptive methods, such as reliance on preset rules, lack of historical experience reuse capabilities, poor adaptability to changes in energy-consuming equipment types and load fluctuation patterns, and the need for frequent manual intervention. It enables optimization decision-making to possess the autonomous evolution capability of "experience accumulation - rule sedimentation - strategy upgrade," significantly improving the robustness and transferability of the method, reducing long-term operation and maintenance costs, and making it suitable for energy management systems of different scales and operating conditions.

[0090] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0091] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0092] Figure 1 The main flowchart is shown for the adaptive quality assurance method of intelligent energy management software based on deep learning.

[0093] Figure 2 A flowchart for training a neural network model for quality status assessment in a deep learning-based adaptive quality assurance method for intelligent energy management software.

[0094] Figure 3 This is a flowchart illustrating the decision-making, closed-loop feedback, and knowledge graph evolution linkage of a deep learning-based intelligent energy management software quality adaptive assurance method. Detailed Implementation

[0095] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0096] This embodiment uses an intelligent energy management system for an automotive parts manufacturing plant as an application scenario. The system, deployed in an industrial environment using a microservice architecture, needs to simultaneously achieve the dual goals of stable software operation and optimized energy utilization efficiency. The plant's energy management involves over 200 energy-consuming devices (including stamping machines, injection molding machines, air compressors, etc.). The software system comprises 12 core microservices (load forecasting service, energy-saving scheduling service, data acquisition service, etc.), processing approximately 500GB of equipment data daily. The original manual operation and maintenance model suffered from delayed response to quality risks and insufficient targeted energy consumption optimization. Therefore, the method described in this invention is adopted to achieve adaptive software quality assurance.

[0097] The system deployment environment for this embodiment is as follows:

[0098] (a) Hardware environment

[0099] Host servers: 3 Dell PowerEdge R750 physical servers, each configured with an Intel Xeon 8375C 24-core CPU, 128GB DDR4 memory, and 2TB NVMe SSD disk, supporting containerized deployment;

[0100] Data Acquisition Probe: The NetisNS-IPC100 lightweight probe based on eBPF technology supports x86 architecture, consumes 1.8W, operates at a temperature of -20℃ to 60℃, and consumes ≤2.5% of CPU resources and ≤80MB of memory resources.

[0101] Edge devices: 10 industrial edge gateways (Huawei IE4000) are responsible for data collection and command issuance from energy-consuming devices and support the MQTT protocol.

[0102] (ii) Software Environment

[0103] Operating system: The server uses CentOS 7.9, and the edge gateway uses a Linux embedded system;

[0104] Containerization platform: Docker 20.10.17 + Kubernetes 1.23.5;

[0105] Deep learning frameworks: TensorFlow 2.8 (model training), PyTorch 1.11 (reinforcement learning agent);

[0106] Databases: InfluxDB 2.1 (time series database, used to store runtime quality data) and Neo4j 4.4 (knowledge graph storage);

[0107] Development languages: Python 3.9 (data processing, model development), Java 11 (microservice development).

[0108] II. Specific Implementation Steps

[0109] Step S1: Perception and Fusion Acquisition of Multimodal Quality Data in Operational State

[0110] Data collection and deployment: NetisNS-IPC100 probes are deployed in Docker containers on three host servers. Without modifying the energy management software source code, system resource data is captured through eBPF technology. At the same time, the software's native RESTful API (port 8080) and JMX interface (port 9999) are called to collect data in parallel with the probes, ensuring both comprehensiveness and real-time performance of the data collection.

[0111] Multimodal data acquisition execution:

[0112] System resource modal data: CPU core utilization is collected once per second, memory usage (in MB) and garbage collection frequency (in times per minute) are collected once every 10 seconds, and disk I / O throughput (in MB / s) and network inbound / outbound bandwidth utilization are collected once every 2 seconds to form a time series. , , , ;

[0113] Software service modal data: API call chain response latency (ms) and success rate (percentage) between microservices, message queue (RabbitMQ) backlog depth (number of messages), active connections in the database (MySQL 8.0) connection pool (number of connections), and key business transaction processing throughput (transactions / second) are collected every 1 time / 5 seconds, forming a time-series data. , , , ;

[0114] Energy management business modal data: Real-time power of energy-consuming equipment (unit: kW) is collected once every 10 seconds via the edge gateway; unit product energy consumption (calculated in real time based on equipment power and output data, unit: kWh / unit) is collected once every 30 seconds; load prediction deviation rate (the percentage of the difference between the predicted value and the actual load to the actual load); and energy-saving dispatch command execution success rate (percentage) are collected to form a time series. , , , .

[0115] Data preprocessing: The preprocessing workflow is performed using the Python Pandas library.

[0116] Outlier cleaning: Remove outlier data that exceeds the mean ± 3 times the standard deviation using the 3σ criterion (such as power outliers caused by sudden equipment failures).

[0117] Time alignment: Using 1 second as a unified timestamp benchmark, linear interpolation is used to complete data from different collection frequencies to ensure consistency in the time dimension;

[0118] Normalization: Min-max normalization maps all data to the interval [0, 1], eliminating dimensional differences. The normalization formula is: .

[0119] Step S2: Quantitative assessment of quality status based on deep temporal feature extraction

[0120] Data organization: The preprocessed multimodal data is reorganized into a multi-channel time series tensor according to a fixed time window (window length 30 seconds, sliding step size 30 seconds). The tensor dimension is [30×12], including 4 channels for system resource mode, 4 channels for software service mode, and 4 channels for energy business mode.

[0121] Deployment and operation of the quality status assessment neural network model:

[0122] Model architecture implementation:

[0123] The first branch (system resource modality processing) adopts a 1D-CNN+GRU combined structure. The 1D-CNN layer contains 3 convolutional blocks, each containing 32 3×1 convolutional kernels with a stride of 1 and padding=same, used to extract the short-term coupling relationship between CPU utilization and memory usage; the GRU layer has 2 layers, each with 64 hidden units, used to capture the long-term periodic patterns of resource usage during daily peak periods.

[0124] The second branch (software service modality processing) adopts a Transformer encoder based on a multi-head attention mechanism, with 4 attention heads, 64 hidden layer dimensions, and 256 feedforward network dimensions. By calculating the attention weights of different microservice call chain characteristics, it captures the amplification effect of insufficient database connection number on API response latency.

[0125] The third branch (energy management business modal processing) adopts a 3-layer fully connected network with residual connections, with 128, 64 and 32 neurons in each layer, respectively. The activation function is ReLU, and the cumulative growth characteristics of the unit product energy consumption industry benchmark (static efficiency) and load forecast deviation (dynamic change) are extracted.

[0126] Feature fusion layer: Employs a multi-head cross-modal attention mechanism to calculate the mutual attention weights of the three branch features. For example, when API response latency is abnormal, the weight ratio of the software service modal features is automatically increased, and the 64-dimensional feature vectors output by the branches are fused into a unified 64-dimensional comprehensive quality feature vector.

[0127] Joint Output Layer: The regression head outputs a comprehensive quality score through a single fully connected network. (Value range 1-10), the classification head outputs a 5-dimensional risk vector through a multi-label classification network. (These correspond to memory risk, network transmission risk, service call risk, prediction failure risk, and instruction execution risk, respectively. The value for each dimension is the risk confidence level, ranging from [0,1]).

[0128] Model Training Process: Training Dataset Construction: 1,051,200 time window samples were extracted from the manufacturing plant's energy management system's historical operation logs over one year. Three domain experts with more than five years of experience in industrial energy management software operation and maintenance were invited to label each sample with a comprehensive quality level tag, combining system alarm records and fault handling reports. (1-10 points) and risk type label (A 5-dimensional binary vector, where 1 indicates the existence of a corresponding risk);

[0129] Loss function configuration: A composite loss function is used, where =0.5、 =0.3、 =0.2, Mean squared error loss (MSE). For multi-label cross-entropy loss, The L1 norm is used to implement the sparsity constraint of the risk vector;

[0130] Training parameter configuration: The Adam optimizer was used, with an initial learning rate of 1e-4, decaying by 10% every 20 epochs, a batch size of 64, and 100 epochs of iterative training. Training was stopped when the validation set loss was below 0.05 for 3 consecutive epochs. Model performance: The final model had a comprehensive quality score prediction error of 0.23 on the test set and a risk identification accuracy of 95.3%.

[0131] Quality status assessment execution: The reorganized multi-channel temporal tensor is input into the trained model. The model has a single data processing latency of 420ms and outputs a real-time comprehensive quality score. With risk vector .

[0132] Step S3: Adaptive Optimization Decision Generation Based on Deep Reinforcement Learning

[0133] Markov Decision Process (MDP) definition:

[0134] state space : Based on the overall quality score With risk vector Formed by splicing, it has 6 dimensions, that is... = ;

[0135] Action Space A: Hybrid action space, containing three categories and a total of eight specific optimized actions:

[0136] Software resource configuration actions: 1. Increase JVM heap memory by 10%; 2. Decrease JVM heap memory by 10%; 3. Adjust the core thread count of the thread pool by ±2; 4. Adjust the maximum number of database connection pool connections by ±5 (one of these actions will be executed at a time).

[0137] Data processing steps: 1. Switch to stream processing mode; 2. Switch to micro-batch processing mode; 3. Enable data downsampling (downsampling rate is dynamically adjusted according to the current computing load, ranging from 30% to 70%); 4. Adjust the parallelism of real-time computing tasks by ±1 (one task is selected to be executed at the same time).

[0138] Energy management strategy actions: 1. Switch the load forecasting model to an online reinforcement learning model; 2. Use a fast heuristic algorithm for degraded energy-saving scheduling (select one or not to execute based on triggering conditions).

[0139] State transition probability Simulation was conducted by building a simulation testing platform consistent with the real environment (including simulation of 200+ energy-consuming devices and deployment of 12 microservices, implemented using JMeter + DockerCompose);

[0140] reward function : Using the definition in step S4 Discount factor =0.9.

[0141] Construction and operation of deep reinforcement learning decision-making agents:

[0142] Intelligent agent architecture: based on the actor-critic framework.

[0143] The actor network consists of one shared feature extraction layer (64 neurons, ReLU activation) and three parallel action output heads, which output the probability distribution of discrete actions (such as model switching actions) and the parameter values ​​of continuous actions (such as resource adjustment magnitude).

[0144] The critic network employs a 2-layer fully connected network (128 neurons → 64 neurons, ReLU activation), outputting the long-term expected reward of state-action pairs to guide the policy updates of the actor network.

[0145] Training algorithm: Proximity Policy Optimization (PPO) algorithm is used, with hyperparameters... =0.2, the objective function is configured as required;

[0146] Training process: Ten typical scenarios (normal load, high load, data burst, equipment failure, network fluctuation, etc.) are simulated in the simulation test platform, allowing the agent to interact with the environment for 5 million steps. A model snapshot is saved every 100,000 steps, and the network parameters are updated every 1,000 steps by sampling experience data.

[0147] Agent performance: After training, the model has an average reward of 1.3 and a decision response time of 180ms in the test scenario, and is deployed on a distributed intelligent quality assurance platform.

[0148] Adaptive optimization decision generation: The output of step S2 and splice to the current state Input the decision-making agent, and the agent dynamically outputs optimized decision actions based on the state. For example, when Medium memory risk confidence level =0.75 (above the threshold) When the JVM heap memory is 0.7 and the remaining physical memory is 8GB (higher than the safe threshold of 5GB), output the action instruction "Increase JVM heap memory by 10%"; when... Exceeding the threshold for 5 consecutive cycles =15% and predicted failure risk confidence level =0.8 (above the threshold) When the value is 0.7, the action command "Switch the load prediction model to the reinforcement learning online model" will be output.

[0149] Step S4: Decision Execution and Closed-Loop Feedback Learning

[0150] Decision execution: The decision generation engine sends optimization actions At to the target objects through standardized interfaces: JVM heap memory and thread pool parameter adjustments are executed through the JMX interface. Before execution, it is verified whether the adjustment range exceeds the hardware resource limit (e.g., the maximum JVM heap memory does not exceed 60% of the server memory).

[0151] Data processing mode switching takes effect via the RESTful API in the software configuration center;

[0152] The switching of prediction models and scheduling algorithms is achieved through the service routing interface of the microservice registry.

[0153] All actions include a retry mechanism. If an action fails, it will be retried after 1 second, up to 3 times. If a retry fails, the configuration will be rolled back (restored to the configuration before the decision).

[0154] Effect observation and data collection: Setting the observation period =60 seconds (to ensure the optimization actions are fully effective). After the observation period, repeat step S1 for data acquisition and step S2 for quality assessment to obtain the new status. New Comprehensive Quality Score New unit product energy consumption .

[0155] Reward value calculation: according to the formula Calculate the reward value, where =0.6 (Software quality weight) =0.4 (Energy optimization weight). = - .For example =7.2, =8.5, =5.2kWh / unit =4.4kWh / unit, then =0.8Wh / piece 0.6×(8.5−7.2)+0.4×0.8=1.1.

[0156] Experience replay and model update: The experience quadruple ( , , , The system stores 100,000 experience replay data points in a buffer and employs a random sampling strategy (to avoid sample correlation). After every 10 decision-making processes, 64 experience data points are sampled. The parameters of the Actor network and Critic network are updated using gradient descent to ensure that the agent can continuously adapt to environmental changes.

[0157] Step S5: Dynamic Construction and Evolutionary Reasoning of Quality Knowledge Graph

[0158] Quality Knowledge Graph Construction:

[0159] Storage medium: Neo4j4.4 graph database is used and deployed on one of the host servers, supporting distributed deployment and horizontal scaling;

[0160] Definition of the relationship between nodes and edges:

[0161] Software component nodes: such as "load forecasting microservice" and "database service", with attributes including name, version, and deployment node;

[0162] Hardware resource nodes: such as "Industrial Server A" and "Edge Gateway 1", with attributes including model, number of CPU cores, and memory capacity;

[0163] Quality indicator nodes: such as "API response latency" and "overall quality score", with attributes including name, threshold range, and data type;

[0164] Optimize action nodes: such as "increase JVM heap memory" and "switch to micro-batch processing mode", with attributes including action type, execution parameters, and effective time;

[0165] Energy consumption indicator nodes: such as "energy consumption per unit product" and "load forecast deviation rate", with attributes including name, benchmark value, and calculation method;

[0166] Edge relationship types: "Running on" (software component → hardware resource), "Affecting" (hardware resource → quality indicator), "Causing" (quality indicator → risk type), "Mitigating" (optimization action → risk type), "Optimizing for" (optimization action → energy consumption indicator);

[0167] Dynamic graph update: After each closed-loop feedback is completed (step S4), the current state S, action A, reward Reward, and software configuration context (such as current JVM parameters and data processing mode) are automatically converted into graph node and edge relationships and written to the graph database to realize incremental graph update.

[0168] Relational reasoning based on graph neural networks:

[0169] Node feature initialization: Generate initial feature vectors based on node type and attributes. Hardware resource nodes use “normalized value of CPU core count + normalized value of memory capacity + number of deployed software components” to form a 3-dimensional initial feature. Quality indicator nodes use “current value + upper threshold + lower threshold” to form a 3-dimensional initial feature. All initial features are mapped to the [0,1] interval.

[0170] Message passing and feature aggregation: Graph Attention Network (GAT) is used for three rounds of message passing. Each GAT layer contains 64 hidden units. By calculating the attention coefficients between nodes (such as the correlation strength between "memory risk" nodes and "increase JVM heap memory" action nodes), neighborhood node information is weighted and aggregated to update the 64-dimensional embedding vector of each node.

[0171] Link prediction and rule extraction: Calculating the probability of the existence of a "mitigation" relationship between quality risk indicator nodes and optimization action nodes. >0.85 (preset threshold) The association relationship with a confidence level of 0.85 was extracted as a valid rule, and the rule format is: IF [risk mode condition] THEN [suggested action (confidence level: 0.85)] For example, “IF [Memory risk confidence > 0.7 and remaining physical memory > 5GB] THEN [Increase JVM heap memory by 10% (confidence: 0.93)]”, the extracted rules are stored in the rule base.

[0172] Rule Injection and Policy Evolution: Two methods are used to inject effective rules from the rule base into the reinforcement learning agent:

[0173] Action masking mechanism: When the actor network outputs action probabilities, actions explicitly prohibited by the rules (such as "increase JVM heap memory" when physical memory is insufficient) have their probabilities set to zero; actions strongly recommended by the rules (confidence > 0.9) have their probabilities weighted and amplified by 2 times to reduce invalid exploration;

[0174] Internal reward mechanism: setting rule reward coefficients =0.2, when the agent's decision action conforms to the valid rules of the current state, an additional intrinsic reward of +0.2 is given; otherwise, a penalty of -0.2 is given. The total reward function is then modified as follows: .

[0175] Implementation effect verification

[0176] The method described in this embodiment was continuously run in the intelligent energy management system of the automotive parts manufacturing plant for 3 months. The effectiveness of the method was verified by comparison with the traditional manual operation and maintenance mode. The specific results are as follows:

[0177] Software quality improvement: The average overall software quality score increased from 7.1 to 8.6, an increase of 21.1%; the risk identification accuracy rate reached 95.3%, which is 30.3% higher than the traditional manual identification (accuracy rate of 65%), realizing early warning and accurate identification of quality risks;

[0178] Energy consumption optimization effect: The average energy consumption per unit product is reduced by 8.2%. The factory's average monthly output is 100,000 units, and the average energy consumption per unit is reduced by 0.41 kWh. Calculated at an industrial electricity price of 0.75 yuan / kWh, the monthly electricity cost is reduced by about 30,800 yuan, and the annual electricity cost is reduced by about 370,000 yuan.

[0179] Improved operational efficiency: The average time to repair software faults was reduced from 45 minutes to 15 minutes, an improvement of 66.7%; 92% of quality optimization decisions can be completed without manual intervention, reducing repetitive operational work by 80%;

[0180] Real-time performance and stability: The average decision response time is 180ms, meeting the real-time requirements of industrial scenarios; the system has operated continuously for 3 months without failure, with an availability of 99.9%.

[0181] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for adaptive quality assurance of intelligent energy management software based on deep learning, characterized in that, The method includes: Step S1, Multimodal operational quality data perception and fusion acquisition: Multimodal operational quality data is acquired in parallel using a lightweight probe deployed in the energy management software instance host environment and the monitoring application programming interface provided by the instance; the multimodal operational quality data includes: System resource modal data: including time-series data of CPU core utilization. Memory usage and garbage collection frequency timing sequence Disk I / O throughput timing sequence Network interface inbound and outbound bandwidth utilization time series ; Software service modal data: including time series sequences of response latency and success rate of API call chains between various microservices within the energy management software. Message queue backlog depth time series Time series of active connections in database connection pool Time series of throughput for critical business transaction processing ; Energy management business modal data: including real-time power time-series sequences of energy-consuming devices controlled by energy management software instances. Time series of unit product energy consumption calculated in real time based on equipment power and output data. The time series of load forecast results and actual load deviation rate based on historical and real-time data rolling execution. The timing sequence of the ratio of energy-saving dispatch instructions successfully executed by the field programmable logic controller. ; Step S2: Quality Status Quantitative Assessment Based on Deep Temporal Feature Extraction: The multimodal operational quality data collected and processed in Step S1 is organized into a multi-channel temporal tensor according to time windows and input into a pre-trained quality status assessment neural network model. This model processes different modal data through three parallel feature extraction branches, calculates cross-modal attention weights through a feature fusion layer, generates a comprehensive quality feature vector, and inputs it into the joint output layer of regression and classification, simultaneously outputting a comprehensive quality score representing the overall health of the software. and a by A risk vector consisting of quality degradation risk indicators and their confidence levels in each dimension. ; Step S3: Adaptive optimization decision generation based on deep reinforcement learning: The comprehensive quality score output in step S2 is used to generate the overall optimization decision. With risk vector The splicing, as the current state of the Markov decision process. A deep reinforcement learning decision-making agent based on an actor-critic framework, according to the state... Output a set of adaptive optimization decision actions The decision-making action The mixed action space includes: Software resource configuration actions: Dynamically adjust the JVM heap memory size limit, thread pool core thread count, and database connection pool maximum number of connections for the energy management software instance; Data processing workflow actions: switch high-frequency sensor data processing mode to stream processing or batch processing, dynamically start and stop data downsampling algorithm, and adjust the parallelism of real-time computing tasks; Energy management strategy action: in the load forecast deviation rate sequence continuous The number of cycles exceeded the threshold At that time, the load forecasting model is switched from a time series model to an online learning model based on real-time reinforcement learning; in the comprehensive quality score Below the threshold In such cases, the globally optimal energy-saving scheduling algorithm is downgraded to a fast heuristic algorithm that ensures critical loads. Step S4, Decision Execution and Closed-Loop Feedback Learning: The adaptive optimization decision action generated in step S3... The command is distributed to the corresponding energy management software instance or its host environment for execution; an observation period is then waited after the action is executed. Repeat steps S1 and S2 to obtain a new state. and the corresponding overall quality score Calculate reward value The calculation formula is as follows: ,in, Observation period Internal unit product energy consumption The decrease and The preset positive weighting coefficients; the quadruples ( , , , The data are stored as experience samples in the experience replay buffer and periodically sampled to update the policy network and value network parameters of the deep reinforcement learning decision agent. Step S5: Dynamic Construction and Evolutionary Reasoning of the Quality Knowledge Graph: Continuously integrating historical and real-time states. ,action ,award The corresponding software configuration context information is stored in a graph database in the form of a graph structure to construct and dynamically update a software quality-configuration-performance knowledge graph. The node types in this graph include software components, hardware resources, quality indicator entities, optimization action entities, and energy consumption indicator entities. The edge relationship types include running on, affecting, causing, mitigating, and optimizing. The knowledge graph is periodically embedded and reasoned using a graph neural network to mine strong association rules between frequently occurring quality risk patterns and effective optimization actions. These rules are then injected into the deep reinforcement learning decision agent in step S3 as new constraints or initialization strategies.

2. The method for adaptive quality assurance of intelligent energy management software based on deep learning according to claim 1, characterized in that, The training method for the quality status assessment neural network model in step S2 specifically includes: Step S201: Extract multimodal operational quality data samples spanning a long period from historical operation logs and monitoring databases, and invite domain experts to assign a comprehensive quality level label to each time window based on system alarm records and fault reports. and specific risk type labels ; Step S202: Construct an initial neural network model, which includes three parallel feature extraction branches, a feature fusion layer, and a joint output layer for regression and classification; the feature extraction branches include: The first branch uses a one-dimensional convolutional neural network combined with a gated recurrent unit network to extract the local dependencies and long-term periodic patterns of the system resource modal data. The second branch employs an attention-based temporal encoder network to extract the correlation and influence features between different microservice call chains in the software service modality data. The third branch employs a fully connected network with residual connections to extract the static and dynamic performance characteristics of the energy management business modal data; Step S203: Use a composite loss function The model undergoes end-to-end supervised training, and the formula for the composite loss function is as follows: ,in, The mean squared error loss for the overall quality score prediction. The multi-label cross-entropy loss is used for risk vector prediction. The L1 norm is used to impose sparsity constraints on the risk vector. , and To balance the hyperparameters; Step S204: Using the backpropagation algorithm and adaptive moment estimation optimizer, train the model on the labeled dataset until the loss function converges, and obtain the trained quality status assessment neural network model.

3. The method for adaptive quality assurance of intelligent energy management software based on deep learning according to claim 1, characterized in that, The method for constructing and training the deep reinforcement learning decision agent based on the actor-critic framework in step S3 specifically includes: Step S301: Define the Markov decision process quintuple ( , , , , ): State space The overall quality score With risk vector Concatenated vectors; action space The mixed action space defined in step S3; state transition probabilities Simulated from a real software environment; reward function Use the definition in step S4 Discount factor ; Step S302: Construct the actor-critic network structure, where: The actor network consists of a shared feature extraction layer and multiple parallel action output heads, which output the probability distribution of discrete actions and the parameter values ​​of continuous actions, respectively. Critics network, used to evaluate the state Take action below The expected long-term return that can be obtained; Step S303: In a simulation test environment connected to the energy management software instance, the agent interacts with the environment to collect experience data, and iteratively updates the network parameters using a proximal policy optimization algorithm; the objective function of the PPO algorithm... for: ,in, For actor network parameters, This is the estimated value of the dominance function. For hyperparameters; Step S304: Deploy the trained agent on the distributed intelligent quality assurance platform for online decision generation.

4. The method for adaptive quality assurance of intelligent energy management software based on deep learning according to claim 1, characterized in that, The specific execution logic for dynamically adjusting the upper limit of JVM heap memory size in the software resource configuration action is as follows: when the risk vector The confidence level of the risk dimension of frequent garbage collection in the middle indicates the risk. Above the threshold When the remaining physical memory in the system resource modal data is higher than the safety threshold, a decision action is taken. It will generate an instruction to increase the JVM heap memory limit by a preset ratio; when memory usage remains above the threshold and garbage collection frequency increases... When the value is below the threshold, the decision action is... Generate an instruction to reduce the upper limit of heap memory.

5. The method for adaptive quality assurance of intelligent energy management software based on deep learning according to claim 1, characterized in that, The specific execution logic for switching data processing modes in the data processing flow is as follows: when the average response latency of the API call chain in the software service modal data is... A sudden increase exceeding the threshold Furthermore, when the energy management business modal data shows an abnormally high frequency of sensor data reporting, a decision-making action is initiated. It will generate instructions to temporarily switch the processing mode of the corresponding data source from latency-sensitive stream processing to throughput-priority micro-batch processing, and at the same time enable the data downsampling algorithm, wherein the downsampling rate is dynamically determined according to the current computing load.

6. The method for adaptive quality assurance of intelligent energy management software based on deep learning according to claim 1, characterized in that, The specific execution logic for switching the load forecasting model in the energy management strategy action is as follows: when the load forecasting deviation rate sequence in the energy management business modality data... continuous The number of cycles exceeded the first threshold. And the risk vector The confidence level of the risk of prediction algorithm failure is indicated in the middle. Above the threshold At that time, decision-making actions The system generates an instruction to pause the use of a long-term prediction model based on historical data and activate a short-term online prediction model based on deep reinforcement learning. This online prediction model uses recent real-time load sequences and environmental parameters as its state to minimize the prediction bias at the next time step. To reward the target, its prediction policy network is updated through online learning.

7. The method for adaptive quality assurance of intelligent energy management software based on deep learning according to claim 1, characterized in that, The specific steps for relational reasoning using graph neural networks in step S5 include: Step S501: For each node in the knowledge graph, initialize feature encoding according to its type and attributes; Step S502: Use a graph attention network to perform multi-round message passing and feature aggregation, and update the embedding vector of each node to fuse the information of its multi-hop neighboring nodes; Step S503: Based on the updated node embedding representation, predict the existence of a specific relationship between any two nodes by calculating the inner product score between any two nodes or training a dedicated link prediction network. probability ; Step S504: Predict the probability Higher than the preset threshold Furthermore, the relationship edges connecting the quality risk indicator nodes and the optimization action nodes are extracted into new association rules, formalized as: IF [risk mode condition] THEN [suggested action (confidence level: ...] )]; Store these rules in the rule base.

8. The method for adaptive quality assurance of intelligent energy management software based on deep learning according to claim 7, characterized in that, Step S5 further includes the steps of rule injection and policy evolution: injecting the rules in the rule base into the deep reinforcement learning decision agent through at least one of the following methods: Method 1: As an action mask, when the actor network outputs action probabilities, the rules in the rule base for the current state are used. The probability of unrecommended actions is set to zero, or the probability of actions strongly recommended by the rules is weighted and amplified. Method 2: Transform the rules into additional intrinsic rewards When the intelligent agent makes a decision Conforms to the current state When establishing a valid rule, provide positive intrinsic rewards. Conversely, negative intrinsic rewards are given. ,in The rule reward coefficient is used; the total reward function is modified as follows: .

9. A deep learning-based intelligent energy management software quality adaptive assurance system, applicable to the deep learning-based intelligent energy management software quality adaptive assurance method described in any one of claims 1-8, characterized in that, The system includes: Data sensing and acquisition module: Deployed in the host environment of the energy management software instance, it is used to execute step S1 and collect multimodal operating quality data; Cloud-based intelligent analysis platform: Communicatively connected to the data sensing and acquisition module, the platform includes: Quality assessment engine: Built-in quality status assessment neural network model, used to execute step S2, outputting a comprehensive quality score and risk vector; Decision generation engine: Built-in deep reinforcement learning decision agent based on actor-critic framework, used in step S3 to generate adaptive optimization decisions; Knowledge graph engine: Used to store, maintain and run graph neural networks, execute step S5, and perform knowledge graph construction and reasoning; Strategy Execution and Feedback Module: This module distributes the decision actions output by the decision generation engine to the target system for execution, and collects the effect data after execution to form a closed-loop feedback loop. This loop drives the online learning of the decision generation engine and the rule updates of the knowledge graph engine.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a deep learning-based intelligent energy management software quality adaptive assurance method as described in any one of claims 1 to 8.

Images