AEMS pushing and training all-in-one machine

Abstract

The invention provides an AEMS pushing and training all-in-one machine, the AEMS pushing and training all-in-one machine comprises a box body, an acquisition gateway module, a memory module, a storage module, a CPU module and a GPU module are arranged in the box body, the acquisition gateway module is used for acquiring data from various external data sources in real time and sending the acquired data to the memory module and the storage module; the memory module is used for temporarily storing the data transmitted by the acquisition gateway module; the storage module is used for storing data for a long time; the CPU module is used for overall control and management of the AEMS pushing and training all-in-one machine; and the GPU module is used for parallel computing acceleration of the AEMS pushing and training all-in-one machine. According to the technical scheme, integration of data collection and intelligent analysis on the edge side is achieved, the training efficiency is improved, and the energy consumption is optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of energy management Internet of Things devices and artificial intelligence technology, in particular to an AEMS training all-in-one machine. BACKGROUND

[0002] The existing intelligent devices can collect and store real-time Internet of Things data such as real-time monitoring parameters of large energy equipment and real-time meter readings of sub-metering electric meters through different industrial protocols such as modbus and bacnet. However, its technical defects are: 1. Separation of data collection and intelligent analysis Traditional energy management gateways only have data collection functions, while data analysis, model training and strategy optimization usually rely on cloud servers or independent computing devices, resulting in high data delay and slow response speed, which cannot meet the real-time energy control requirements.

[0003] 2. Limitations of cloud training mode Existing AI model training mostly relies on cloud computing power, resulting in high data transmission costs and high privacy leakage risks, especially for industrial site data, enterprises are usually reluctant to upload critical energy consumption data to the cloud.

[0004] 3. Lack of localized intelligent decision-making ability Existing systems usually only support fixed strategies or simple rule control, and cannot adaptively optimize based on real-time data, resulting in inflexible energy management strategies and difficulty in coping with complex working condition changes.

[0005] 4. Weak knowledge management and interaction capabilities Traditional energy management systems lack natural language interaction capabilities, making it difficult for management personnel to quickly obtain device operating status, energy efficiency analysis results and other key information, relying on manual report analysis, which is inefficient. SUMMARY

[0006] The present application provides an AEMS training all-in-one machine to improve training efficiency and optimize energy consumption.

[0007] The present application provides an AEMS training all-in-one machine, comprising a box body, the box body is provided with a collection gateway module, a memory module, a storage module, a CPU module and a GPU module, wherein, The collection gateway module is used for collecting data from multiple external data sources in real time, and sending the collected data to the memory module and the storage module; The memory module is used for temporarily storing data transmitted by the collection gateway module; The storage module is used for long-term storage of data; The CPU module is used for overall control and management of the AEMS training and inference integrated machine. The training and inference integrated engine is used for training and inference closed loop and incremental learning. The knowledge graph constructor is used for extracting entity relationship from device operation log and constructing domain knowledge base. The multi-modal interaction interface is used for voice / text dual-mode question and answer interaction. The GPU module is used for parallel computing acceleration of the AEMS training and inference integrated machine.

[0008] In the above technical solution, the box is provided, the box is provided with a collection gateway module, a memory module, a storage module, a CPU module and a GPU module, the collection gateway module is used for collecting data from multiple external data sources in real time, and the collected data is sent to the memory module and the storage module, the memory module is used for temporarily storing the data transmitted by the collection gateway module, the storage module is used for long-term storage of data, the CPU module is used for overall control and management of the AEMS training and inference integrated machine, and the GPU module is used for parallel computing acceleration of the AEMS training and inference integrated machine, realizing data collection and intelligent analysis integration on the edge side, improving training efficiency and optimizing energy consumption.

[0009] In a specific implementable embodiment, the box is provided with a power module, wherein, The power module is used for power supply.

[0010] In a specific implementable embodiment, the box comprises a first cavity and a second cavity, wherein, The power module and the collection gateway module are arranged in the first cavity; The memory module, the storage module, the CPU module and the GPU module are arranged in the second cavity.

[0011] In a specific implementable embodiment, the memory module, the CPU module and the GPU module are integrally arranged.

[0012] In a specific implementable embodiment, the memory module, the CPU module and the GPU module are integrally arranged on the mainboard.

[0013] In a specific implementable embodiment, the collection gateway module is integrally provided with a DeepSeek14B acceleration unit, wherein, The DeepSeek14B acceleration unit is used for edge side data collection and AI training / inference integration.

[0014] In an embodiment, the DeepSeek 14B acceleration unit comprises an FPGA and NPU heterogeneous computing module, wherein, The FPGA and NPU heterogeneous computing module is configured to perform edge-side data collection and AI training / inference integration.

[0015] In an embodiment, the knowledge graph constructor adopts a GNN and Transformer hybrid architecture.

[0016] In an embodiment, the storage module comprises an edge storage array, wherein, The edge storage array is configured to perform real-time storage of data.

[0017] In an embodiment, the edge storage array is a dual-channel NVMe SSD. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A structural schematic diagram of an AEMS training and inference integrated machine provided by an embodiment of the present application is shown in FIG. 1. Figure 2 An electrical block diagram of an AEMS training and inference integrated machine provided by an embodiment of the present application is shown in FIG. 2.

[0019] In the figure, 1 is a box body, 2 is a collection gateway module, 3 is a memory module, 4 is a storage module, 5 is a CPU module, 6 is a GPU module, and 7 is a power module. DETAILED DESCRIPTION

[0020] The present application will be further described in detail by the accompanying drawings and embodiments. Through these descriptions, the features and advantages of the present application will become more apparent.

[0021] The term "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any implementation described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other implementations. Unless specifically indicated otherwise, the drawings shown in the Figures are not necessarily to scale.

[0022] In addition, the technical features involved in different embodiments of the present application described below can be combined with each other as long as there is no conflict.

[0023] To facilitate the understanding of the AEMS training integrated machine provided by the embodiments of the present application, the application scenarios thereof are first described. The AEMS training integrated machine provided by the embodiments of the present application is used to solve the problems in the prior art. The existing intelligent device can collect and store real-time Internet of Things data such as real-time monitoring parameters of large energy equipment and real-time meter readings of sub-metering electric meters through different industrial protocols such as modbus and bacnet. However, the technical defects thereof are as follows: data collection and intelligent analysis are separated: the traditional energy management gateway only has a data collection function, while data analysis, model training and strategy optimization usually rely on cloud servers or independent computing devices, resulting in high data delay and slow response speed, which cannot meet the high real-time requirement of energy control demand. The cloud training mode has limitations: the existing AI model training mostly relies on cloud computing power, resulting in high data transmission cost and high risk of privacy leakage, especially for industrial site data, enterprises are usually unwilling to upload key energy consumption data to the cloud. Lack of localized intelligent decision-making capability: the existing system usually only supports fixed strategies or simple rule control, and cannot adaptively optimize based on real-time data, resulting in inflexible energy management strategies and difficulty in coping with complex working condition changes. Knowledge management and interaction capability are weak: the traditional energy management system lacks natural language interaction capability, and management personnel have difficulty in quickly obtaining device running status, energy efficiency analysis results and other key information, and rely on manual report analysis, which is low in efficiency. Therefore, the embodiments of the present application provide an AEMS training integrated machine to solve the problems in the prior art. The embodiments will be described in detail below with reference to the specific drawings.

[0024] Reference Figure 1 and Figure 2 , Figure 1 The structure diagram of the AEMS training integrated machine provided by the embodiments of the present application is shown in FIG. 1. Figure 2 The electrical block diagram of the AEMS training integrated machine provided by the embodiments of the present application is shown in FIG. 2.

[0025] In Figure 1 and Figure 2 , the embodiments of the present application provide an AEMS training integrated machine, which comprises a box body 1, and a collection gateway module 2, a memory module 3, a storage module 4, a CPU module 5 and a GPU module 6 are arranged in the box body, wherein the collection gateway module is used to collect data from a plurality of external data sources in real time, and send the collected data to the memory module and the storage module; The memory module is used to temporarily store the data transmitted by the collection gateway module; The storage module is used to store data for a long time; The CPU module is used for overall control and management of the AEMS training integrated machine; and the CPU module comprises: A training integrated engine is used for training and reasoning closed loop and incremental learning. A knowledge graph constructor is configured to extract entity relationships from device operation logs and construct a domain knowledge base. A multi-modal interaction interface is configured to perform voice / text dual-mode question and answer interaction. The GPU module is configured to accelerate parallel computing of the AEMS training all-in-one machine.

[0026] In the above technical solution, the box is provided, the acquisition gateway module, the memory module, the storage module, the CPU module and the GPU module are arranged in the box, the acquisition gateway module is used for acquiring data from multiple external data sources in real time, and the acquired data is sent to the memory module and the storage module; the memory module is used for temporarily storing the data transmitted by the acquisition gateway module; the storage module is used for long-term storage of data; the CPU module is used for overall control and management of the AEMS training all-in-one machine; the GPU module is used for parallel computing acceleration of the AEMS training all-in-one machine; the integration of edge side data acquisition and intelligent analysis is realized, the training efficiency is improved, and the energy consumption is optimized.

[0027] Specifically, the AEMS training all-in-one machine integrates the acquisition gateway module, the memory module, the storage module, the CPU module and the GPU module in the box, realizes the integration of edge side data acquisition and intelligent analysis, and has the following beneficial effects: 1. Integration of edge side data acquisition and processing Real-time and reliability improvement: the all-in-one machine can acquire data from multiple external data sources in real time through the acquisition gateway module, and quickly transmit the data to the memory module for temporary storage and the storage module for long-term storage. This real-time acquisition and local storage mechanism effectively avoids the problems of delay, loss, etc. that may occur during network transmission, ensures the timeliness and integrity of the data, and provides a reliable data basis for subsequent intelligent analysis. For example, in an industrial production scene, device operation data can be acquired in real time, and device abnormalities can be discovered in a timely manner to ensure the continuity and stability of production.

[0028] Enhanced edge side data processing capability: the all-in-one machine realizes the integration of data acquisition and storage at the edge, reducing the need for data transmission to the cloud. This not only reduces the dependence on network bandwidth, but also avoids the problem of data processing interruption caused by network failure. In some scenes with poor network conditions or high data security requirements, such as remote mines and oil fields, the all-in-one machine can complete data acquisition and preliminary processing locally, ensuring the normal operation of the business.

[0029] 2. Strengthening of intelligent analysis function Training and inference closed loop and incremental learning: The training and inference integrated engine in the CPU module has the ability of training and inference closed loop and incremental learning. This enables the all-in-one machine to continuously optimize the model according to the real-time collected data, improving the accuracy and adaptability of the model. In practical applications, as data accumulates and business scenarios change, the all-in-one machine can automatically adjust model parameters without frequent human intervention, greatly improving the efficiency and effectiveness of intelligent analysis. For example, in the field of intelligent security, the all-in-one machine can continuously optimize the target detection and recognition model according to new monitoring data, improving the accuracy of abnormal behavior recognition.

[0030] Knowledge graph construction and application: The knowledge graph constructor can extract entity relationships from device operation logs and other data to construct a domain knowledge base. This function provides more rich semantic information and context understanding ability for intelligent analysis, enabling the all-in-one machine to more deeply mine the potential value behind the data. For example, in the medical device management scenario, by constructing a knowledge graph of device operation logs, the relationship between device failure, usage, environmental factors, etc. can be analyzed to provide more scientific decision-making basis for device maintenance and management.

[0031] Multi-modal interaction experience optimization: The multi-modal interaction interface supports voice / text dual-mode question and answer interaction, providing a more convenient and natural interaction way for users. Users can input questions through voice or text, and the all-in-one machine can quickly understand and give accurate answers. This interaction method not only improves user experience, but also broadens the application scenarios of the all-in-one machine, making it better adapt to the needs and usage habits of different users. For example, in the intelligent customer service scenario, users can interact with the all-in-one machine through voice to quickly obtain the required information.

[0032] 3. Training efficiency improvement Localized training and inference: The all-in-one machine integrates CPU and GPU modules, which can complete model training and inference tasks locally. Compared with traditional cloud training mode, localized training reduces data transmission time in the network, greatly shortening the training period. At the same time, the parallel computing acceleration capability of the GPU module further improves the training efficiency, enabling the all-in-one machine to quickly process large-scale data and accelerate model convergence. For example, in the field of image recognition, the all-in-one machine can complete the training of a large amount of image data in a short time, quickly generating a high-precision recognition model.

[0033] Incremental learning optimization training process: The incremental learning function of the push-training integrated engine enables the integrated machine to update the model incrementally based on newly collected data without the need to retrain the entire model. This not only saves computing resources and time, but also enables timely adaptation to data changes while ensuring model performance. For example, in an e-commerce recommendation system, the integrated machine can update the recommendation model incrementally based on real-time user behavior data to improve the accuracy and personalization of recommendations.

[0034] 4. Energy consumption optimization Edge-side computing reduces cloud energy consumption: Since the integrated machine completes most of the data processing and intelligent analysis tasks at the edge, it reduces the need for data transmission and processing to the cloud, thereby reducing the computing and storage energy consumption of the cloud. For large-scale Internet of Things application scenarios, this can effectively reduce the overall energy consumption cost and improve energy utilization efficiency. For example, in smart city construction, a large number of sensor data can be processed and analyzed in local integrated machines, reducing dependence on cloud servers and reducing energy consumption.

[0035] Hardware optimization reduces local energy consumption: The integrated machine uses optimized hardware architecture and energy-saving technology in its design to reduce its own energy consumption while ensuring performance. For example, CPU and GPU modules use advanced process technology and power management techniques to dynamically adjust power consumption based on task load, improving energy utilization efficiency. At the same time, the box design of the integrated machine also helps with heat dissipation, reducing additional energy consumption caused by heat dissipation problems.

[0036] In a specific implementable embodiment, the box is provided with a power module 7, wherein, The power module is used for power supply.

[0037] Specifically, the beneficial effects include: In terms of stability, the power module can provide stable and continuous power supply, effectively avoiding damage to various modules (such as acquisition gateway, memory, storage, CPU, GPU module) in the integrated machine due to external power fluctuations or sudden power outages, ensuring stable operation of the equipment and reducing the risk of failure and data loss caused by power problems.

[0038] In terms of convenience, the power module is integrated into the box, simplifying equipment deployment. There is no need for additional connection of complex external power supply equipment, reducing installation difficulty and cost, so that the integrated machine can be quickly put into use in various scenarios. In addition, the power module can be reasonably designed and optimized according to the overall power consumption of the integrated machine, improving energy utilization efficiency, reducing energy consumption, meeting energy-saving and environmental protection requirements, and also helping to reduce long-term use costs.

[0039] In a specific implementable embodiment, the box includes a first cavity and a second cavity, wherein, The power module and the collection gateway module are arranged in the first cavity; The memory module, the storage module, the CPU module and the GPU module are arranged in the second cavity.

[0040] Specifically, the beneficial effects include: In terms of heat dissipation management, the power module and the collection gateway module generate relatively less heat during operation, and are placed in the first cavity, which facilitates targeted design of heat dissipation solutions, such as the use of smaller power heat dissipation fans or heat dissipation fins, reducing heat dissipation cost and energy consumption. The memory, storage, CPU and GPU modules have heavy computing tasks and generate a large amount of heat, and are placed in the second cavity, which can be equipped with a high-efficiency heat dissipation system, such as liquid cooling or high-power air cooling devices, to ensure stable operation of each module at an appropriate temperature, avoiding performance degradation or hardware damage due to overheating.

[0041] From the perspective of electromagnetic interference protection, the power module may generate electromagnetic interference, and the collection gateway module also has certain requirements for the electromagnetic environment. Placing them in the first cavity can reduce interference with sensitive modules (such as CPU and GPU) in the second cavity through physical isolation and electromagnetic shielding design, ensuring the accuracy of data transmission and processing, and improving the overall stability and reliability of the system.

[0042] In terms of maintenance and upgrading, the divided cavity arrangement allows maintenance personnel to quickly locate the cavity where the problem module is located. If the power module or the collection gateway module fails, only the first cavity needs to be opened for maintenance, without affecting the normal operation of other modules in the second cavity, reducing the difficulty and time cost of maintenance. At the same time, when upgrading the module, the corresponding cavity can be operated accordingly, improving the upgrading efficiency and reducing the impact on the overall system.

[0043] In a specific implementable embodiment, the memory module, the CPU module and the GPU module are integrally arranged.

[0044] Specifically, the beneficial effects include: In terms of performance improvement, the integrated arrangement greatly shortens the data transmission path between modules, reduces data transmission delay, and improves data interaction efficiency. This allows the CPU and GPU to access data in the memory more quickly, accelerating the model training and inference process, and significantly improving the overall computing performance of the system.

[0045] From the perspective of space utilization, integrated design saves internal space, making the device structure more compact, facilitating deployment in limited space, especially suitable for application scenarios with strict space requirements.

[0046] In terms of energy consumption optimization, the number of inter-module connection lines and interfaces is reduced, energy loss during signal transmission is reduced, and integrated design facilitates unified power management and power consumption optimization, improving energy utilization efficiency and reducing operating costs.

[0047] In a specific implementation, the memory module, the CPU module, and the GPU module are integrated on the motherboard.

[0048] Specifically, the benefits include: In terms of performance, data interaction efficiency is greatly improved. Modules communicate directly through internal high-speed channels on the motherboard, significantly reducing data transmission delay, allowing CPU and GPU to quickly access memory data, and speeding up model training, inference, and other task processing, resulting in a leap in overall computing performance.

[0049] Space utilization is more efficient. Integrated modules reduce installation space and connection lines, making the motherboard layout more compact, saving internal space, facilitating device miniaturization design, and adapting to more space-limited scenarios.

[0050] Stability is enhanced. Module integration on the motherboard reduces external interfaces and connection components, reducing the risk of failure due to poor contact, looseness, and other issues, making the system run more stably and reliably, improving data transmission accuracy, and ensuring stable operation of the AEMS training and inference all-in-one machine.

[0051] In a specific implementation, the collection gateway module is integrated with a DeepSeek14B acceleration unit, wherein, The DeepSeek14B acceleration unit is used for edge-side data collection and AI training / inference integration.

[0052] Specifically, the benefits include: Enhance edge intelligence processing capability: Achieve data collection and AI training / inference integration on the edge side, without the need to transmit large amounts of raw data to the cloud for processing, reducing data transmission delay and bandwidth pressure. For example, in industrial scenarios, devices can collect data in real time and quickly perform AI analysis, respond to abnormalities in a timely manner, and improve production efficiency and safety.

[0053] Enhance real-time performance and response speed: DeepSeek14B acceleration unit can quickly process collected data to achieve immediate inference and decision-making. For example, in intelligent security monitoring, it can identify abnormal behavior in real time and trigger alarms, responding faster than traditional methods.

[0054] Reduce costs: Reduce dependence on cloud resources, reduce data transmission and cloud storage and computing costs, and local processing also avoids security risks and privacy leaks during data transmission.

[0055] In one specific implementation, the DeepSeek 14B acceleration unit includes an FPGA and NPU heterogeneous computing module, wherein, The FPGA and NPU heterogeneous computing module is configured to perform edge-side data collection and AI training / inference integration.

[0056] Specifically, the beneficial effects include: AEMS training and inference integrated machine DeepSeek 14B acceleration unit heterogeneous computing module beneficial effects Improve computing efficiency and flexibility FPGA has programmable characteristics, can quickly customize hardware logic circuit according to different edge-side data collection and AI task requirements, and realize efficient data processing. NPU is good at neural network calculation, can quickly execute matrix operation in AI training and inference. The combination of the two can fully exert their respective advantages, greatly improve the speed of edge-side data collection and AI task processing, and meet the needs of complex and variable edge computing scenarios.

[0057] Reduce power consumption and cost Compared with general-purpose processors, FPGA and NPU have lower power consumption in specific tasks. In edge-side devices, low power consumption means longer battery life or smaller cooling requirements, reducing the operating cost of the device. At the same time, the heterogeneous computing module reduces the dependence of the device on high-power and high-cost hardware, reducing the overall hardware cost.

[0058] Enhance real-time performance and reliability Edge-side data collection and processing require real-time response, and the heterogeneous computing module can quickly complete data processing and AI inference locally, avoiding the delay caused by data transmission to the cloud. Moreover, local processing reduces the data transmission link, reduces the risk of data loss and errors, and improves the reliability of the system.

[0059] In one specific implementation, the FPGA and NPU heterogeneous computing module is of the XCZU19EG model.

[0060] Specifically, the beneficial effects include: The module can fully utilize the programmable characteristics of FPGA and the neural network computing advantages of NPU, realizing the integration of edge data collection and AI training / inference. FPGA can customize hardware logic for specific algorithms, improving data processing efficiency and reducing latency; NPU is good at neural network computing, accelerating the AI inference process. The combination of the two can dynamically allocate computing resources according to task characteristics, improving overall computing efficiency. In edge scenarios, collected data can be quickly processed, and AI analysis and decision-making can be performed in real time, reducing data transmission delay and bandwidth pressure. At the same time, compared with general-purpose processors, this heterogeneous computing module has lower power consumption on specific tasks, which helps to reduce device operating costs, improve energy utilization efficiency, and enhance the real-time performance and reliability of the system.

[0061] In a specific implementable embodiment, the storage module comprises an edge storage array, wherein, The edge storage array is used for real-time storage of data.

[0062] Specifically, the beneficial effects include: Improving data storage real-time performance The edge storage array can realize real-time storage of data, which can be saved at the moment of data generation. Taking the industrial monitoring scenario as an example, sensors on the production line continuously generate a large amount of data, which can be immediately stored by the edge storage array, avoiding the loss of critical data or the failure to record in time due to the delay of data transmission to the cloud or central storage, ensuring the integrity and timeliness of the data.

[0063] Reducing network dependence and cost Reducing dependence on the network, without the need to transmit data to a remote storage center in real time, reducing network bandwidth requirements and transmission costs. In remote areas or environments with poor network conditions, such as field monitoring stations, the edge storage array can first store data locally, and then transmit data when the network is restored, ensuring the continuity of data collection while saving network costs.

[0064] Enhancing data security and privacy Data is stored in the local edge storage array, reducing the risk of data leakage during transmission. For data involving sensitive information, such as local monitoring data of medical institutions, edge storage can better protect patient privacy.

[0065] In a specific implementable embodiment, the edge storage array is a dual-channel NVMe SSD.

[0066] Specifically, the beneficial effects include: Significant improvement in read-write speed: The dual-channel design allows data transmission to occur in parallel, and the NVMe protocol further optimizes the storage interface, resulting in exponential growth in read-write speed compared to traditional storage. When handling large amounts of real-time data on the edge side, such as the rapid collection and storage of vehicle information in intelligent traffic monitoring, data writing can be completed instantly, avoiding data loss due to storage delays and ensuring data integrity.

[0067] Reducing latency and improving response: High-speed reading and writing enable the system to quickly acquire and store data, greatly reducing data access latency. In industrial automation scenarios, device state data can be stored in real time, and the control system can quickly read and respond, enabling precise control and improving production efficiency and quality.

[0068] Enhancing system stability and reliability: NVMe SSDs have high stability and low failure rates, and the dual-channel design provides redundancy, ensuring that even if a single channel fails, the other channel can still work normally, ensuring uninterrupted data storage and improving the overall reliability and stability of edge devices.

[0069] In one specific implementation, the AEMS push-training all-in-one machine includes: 1. Multi-protocol acquisition unit Including Modbus RTU / TCP, BACnet / IP, and other industrial protocol stacks; Function: Interface with energy devices and smart meters to achieve second-level data acquisition.

[0070] 2. DeepSeek 14B acceleration unit New FPGA + NPU heterogeneous computing module (model XCZU19EG); Function: Provides 14TOPS computing power to support model training / inference.

[0071] Technical features: Dynamic voltage and frequency adjustment technology is used to reduce power consumption.

[0072] 3. Edge storage array Upgraded to dual-channel NVMe SSD (maximum support of 8TB); Function: Stores time-series data and training data sets.

[0073] 4. Push-training integrated engine Includes: a) Model fine-tuning subsystem: supports efficient fine-tuning of LoRA parameters; b) Online inference engine: optimized inference framework based on TensorRT; Function: Achieves training and inference closed loop, supports incremental learning.

[0074] 5. Knowledge graph constructor Adopt a hybrid architecture of GNN + Transformer; Function: Extract entity relationships from device operation logs and build a domain knowledge base.

[0075] 6. Multimodal interaction interface Integrate speech recognition (ASR) and natural language understanding (NLU) modules; Function: Support voice / text dual-mode question and answer interaction.

[0076] The workflow of the AEMS training and integrated machine is as follows: 1. Data acquisition stage Obtain device operating parameters (frequency: 1 Hz) through a multi-protocol acquisition unit; Data preprocessing: Perform outlier rejection and normalization processing.

[0077] 2. Model training stage Steps: a) Automatically generate training tasks based on historical data (time window: 730 days); b) Call the DeepSeek14B base model for transfer learning; c) Generate device energy efficiency optimization strategy model (output format: ONNX).

[0078] 3. Strategy inference stage Input real-time data into the trained lightweight model (inference delay <200ms); Output device start-stop strategy suggestions and energy efficiency predictions.

[0079] 4. Knowledge interaction stage Use RAG technology combined with knowledge graph for question and answer; Support natural language queries, such as "What is the COP value of the air conditioning system yesterday?" In this embodiment, the beneficial effects include: Training efficiency improvement: Reduce 80% communication overhead compared to cloud training; Strategy response speed: Device control instruction generation time is shortened to within 300ms; Knowledge question and answer accuracy: Achieve an F1 value of 92.3% in the device operation and maintenance scenario; Energy consumption optimization: Reduce 35% of the operating power consumption through dynamic frequency adjustment.

[0080] Take the central air conditioning system control as an example: 1. Connect four chiller units (Modbus TCP) and thirty-two electric meters (RS485); 2. Collect eighteen parameters such as chilled water supply and return water temperature difference and compressor power; 3. The training cycle is set to twenty-four hours, and a start-stop strategy model is generated; 4. When the operator asks "Are there any inefficient running devices in the current system?", the system calls the knowledge base to analyze energy efficiency data and combines real-time COP calculations to provide specific device numbers and optimization suggestions.

[0081] In a specific implementation, the AEMS training and promotion all-in-one machine includes: 1. Box The box serves as the carrier of the entire device, providing physical protection and structural support for the internal modules, ensuring that each module can work normally in a stable environment.

[0082] 2. Collection gateway module function: responsible for real-time data collection from multiple external data sources, and sending the collected data to the memory module and storage module.

[0083] Function: Data collection is the basis for subsequent processing and analysis. By obtaining data from different data sources, it can provide rich information sources for the device, helping to more comprehensively understand and analyze related situations. For example, in an industrial production environment, the collection gateway module can obtain production data such as temperature, pressure, and speed from sensors, monitoring devices, and other data sources.

[0084] 3. Memory module Function: temporarily store data transmitted by the collection gateway module.

[0085] Function: The memory module has high-speed read and write capabilities, and can quickly receive and store collected data to support subsequent real-time processing. Because the read and write speed of memory is much higher than that of the storage module, temporarily storing data in memory can ensure that data can be obtained by CPU modules and other processing units in time, improving the response speed of the system. For example, in real-time data processing scenarios, the memory module can quickly store real-time data collected by sensors to allow the CPU module to analyze and process in a timely manner.

[0086] 4. Storage module Function: long-term data storage.

[0087] Function: The storage module usually has a large capacity and can store collected data for a long time to facilitate subsequent queries, analysis, and audits. Unlike the memory module, the storage module stores data more persistently, so even if the device is powered off, data will not be lost. For example, in an enterprise's data storage system, the storage module can store production data, sales data, and other historical data to support enterprise decision-making.

[0088] 5. CPU module The CPU module is the core control and management unit of the AEMS training and promotion all-in-one machine, including: Push training integrated engine Function: Conduct training inference closed loop and incremental learning.

[0089] Effect: Training inference closed loop can realize continuous optimization and update of the model, by continuously inputting new data into the model for training, improving the accuracy and generalization ability of the model. Incremental learning can integrate new learning knowledge into the existing model without retraining the entire model, improving learning efficiency. For example, in the field of image recognition, the push training integrated engine can continuously train and optimize the image recognition model according to the new image data collected, improving the recognition accuracy.

[0090] Knowledge graph constructor Function: Extract entity relationships from device operation logs and construct domain knowledge base.

[0091] Effect: By analyzing device operation logs, the knowledge graph constructor can extract the relationships between entities such as devices, faults, and operations, and construct a domain knowledge base. This knowledge base can provide support for device fault diagnosis, maintenance decision-making, etc. For example, in industrial device maintenance, the knowledge graph constructor can extract the relationship between device faults and repair methods from device operation logs, and when a device fails, it can quickly query the knowledge base to find the corresponding repair method.

[0092] Multi-modal interaction interface Function: Conduct voice / text dual-mode question and answer interaction.

[0093] Effect: The multi-modal interaction interface provides a more convenient and natural human-computer interaction method, users can interact with the device through voice or text to obtain the required information or perform corresponding operations. For example, in the intelligent customer service system, users can ask questions to the customer service through voice or text, and the system can receive the user's questions through the multi-modal interaction interface and give corresponding answers.

[0094] 6. GPU module Function: Accelerate parallel computing for AEMS push training integrated machine.

[0095] Effect: GPU module has a large number of computing cores and can handle multiple computing tasks simultaneously, greatly improving computing efficiency. In scenarios requiring large-scale data processing and model training, GPU module can significantly shorten the computing time. For example, in the training process of deep learning models, GPU module can process a large number of data samples in parallel to speed up the model training process.

[0096] In this embodiment, the AEMS push training integrated machine realizes data collection, storage, processing and analysis, and human-computer interaction functions through the collaborative work of each module.

[0097] Those skilled in the art will understand that the application can be implemented as a system, method or computer program product.

[0098] Therefore, the present disclosure can be embodied in the form of a hardware completely, a software completely (including firmware, resident software, microcode, etc.), or a combination of hardware and software, which is generally referred to as "circuitry", "module" or "system". In addition, in some embodiments, the present disclosure can also be embodied in the form of a computer program product in one or more computer readable media having computer readable program codes.

[0099] Any combination of one or more computer readable medium can be employed. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. A computer readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the above. More specific examples (a non-exhaustive list) of the computer readable storage medium include: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this document, the computer readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0100] Although the embodiments of the present application have been shown and described above, it should be understood that the above-described embodiments are exemplary, and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application. On this basis, various replacements and improvements can be made to the present application, and all of these fall within the scope of the present application.

Claims

1. An AEMS push-training integrated machine, characterized in that, The system includes a housing, which houses a data acquisition gateway module, a memory module, a storage module, a CPU module, and a GPU module. The data acquisition gateway module is used to acquire data from multiple external data sources in real time and send the acquired data to the memory module and the storage module. The memory module is used to temporarily store the data transmitted from the acquisition gateway module; The storage module is used for long-term data storage; The CPU module is used for the overall control and management of the AEMS training and push-training integrated machine; the CPU module includes: The integrated training and inference engine is used to perform training inference closed loop and incremental learning based on the data transmitted from the acquisition gateway module. A knowledge graph builder is used to extract entity relationships from device operation logs and build a domain knowledge base to provide domain knowledge support for the integrated push-training engine. A multimodal interaction interface is used to conduct voice / text dual-mode question-and-answer interaction with users using the integrated push-training engine. The GPU module is used for parallel computing acceleration of the AEMS push-training integrated machine.

2. The AEMS push-training integrated machine according to claim 1, characterized in that, The enclosure contains a power module, wherein... The power module is used to supply power.

3. The AEMS training and push-training integrated machine according to claim 2, characterized in that, The enclosure includes a first cavity and a second cavity, wherein, The power module and the data acquisition gateway module are disposed within the first cavity; The memory module, the storage module, the CPU module, and the GPU module are disposed within the second cavity.

4. The AEMS training and push-training integrated machine according to claim 3, characterized in that, The memory module, the CPU module, and the GPU module are integrated into a single unit.

5. The AEMS push-training integrated machine according to claim 4, characterized in that, The memory module, the CPU module, and the GPU module are integrated on the motherboard.

6. The AEMS push-train integrated machine according to claim 5, characterized in that, The data acquisition gateway module integrates a DeepSeek14B acceleration unit, wherein... The DeepSeek14B acceleration unit is used for integrated edge data acquisition and AI training / inference.

7. The AEMS push-training integrated machine according to claim 6, characterized in that, The DeepSeek14B acceleration unit includes a heterogeneous computing module consisting of an FPGA and an NPU. The FPGA and NPU heterogeneous computing module is used for edge-side data acquisition and AI training / inference integration.

8. The AEMS push-training integrated machine according to claim 7, characterized in that, The knowledge graph builder adopts a hybrid architecture of GNN and Transformer.

9. The AEMS push-train integrated machine according to claim 8, characterized in that, The storage module includes an edge storage array, wherein, The edge storage array is used for real-time data storage.

10. The AEMS push-train integrated machine according to claim 9, characterized in that, The edge storage array is a dual-channel NVMeSSD.

Images