Industrial edge gateway and industrial time series data processing method

Abstract

The invention provides an industrial edge gateway and an industrial time series data processing method, and relates to the technical field of the Internet of Things, and the industrial edge gateway comprises a real-time operation system which is used for carrying out the real-time processing of first industrial data provided by industrial equipment, and obtaining second industrial data; the universal operating system is used for performing artificial intelligence (AI) reasoning according to the second industrial data; wherein the real-time operating system and the universal operating system share a memory area through a bus; and the real-time operating system triggers the general operating system to adjust the task priority by adopting a hardware interruption mode. Therefore, low-delay processing of the industrial data is realized.

Description

Technical Field

[0001] This application relates to the field of Internet of Things (IoT) technology, and in particular to an industrial edge gateway and an industrial time-series data processing method. Background Technology

[0002] In the field of industrial time-series data processing, existing technologies mainly include time-series data processing systems based on Linux edge gateways and industrial data analysis systems accelerated by DeepSeek's cloud-based Multi-head Latent Attention (MLA). The former uses a single Linux system and suffers from insufficient real-time performance (unable to meet the μs-level latency requirements of hard real-time tasks), edge computing power bottlenecks (high memory consumption of traditional Multi-Head Attention (MHA) mechanisms), and resource scheduling conflicts (competition for resources between Artificial Intelligence (AI) inference and protocol conversion processes). The latter, while utilizing cloud-based MLA technology, suffers from high latency in cloud collaboration, high edge deployment costs, and fragmented multimodal data. Overall, existing technologies face challenges such as the conflict between real-time performance and versatility, edge computing power bottlenecks, and architectural fragmentation. Summary of the Invention

[0003] This application provides an industrial edge gateway and an industrial time-series data processing method, which solves the problems of contradiction between real-time performance and versatility, edge computing power bottleneck, and architectural fragmentation in current industrial time-series data processing.

[0004] Firstly, to achieve the above objectives, embodiments of this application provide an industrial edge gateway, comprising:

[0005] A real-time operating system is used to process the first industrial data provided by industrial equipment in real time to obtain the second industrial data;

[0006] A general-purpose operating system for performing artificial intelligence (AI) inference based on the second industrial data;

[0007] The real-time operating system and the general-purpose operating system share a memory area via a bus; the real-time operating system uses hardware interrupts to trigger the general-purpose operating system to adjust task priorities.

[0008] The real-time operating system is used to process the first industrial data provided by the industrial equipment in real time, including:

[0009] The system collects the first industrial data in real time, performs protocol conversion on the first industrial data to obtain the second industrial data, and writes the second industrial data into the shared memory area via a bus.

[0010] The real-time operating system is also used to receive control commands and execute a first operation corresponding to the control commands;

[0011] In this process, the first operation has a higher priority than the data acquisition, and the data acquisition has a higher priority than the protocol conversion.

[0012] The general-purpose operating system runs an AI inference model, which is used to perform AI inference. The AI ​​inference model is an FP8 quantized deployment model, and the AI ​​inference model is supported by a multi-head potential attention (MLA) acceleration engine.

[0013] The general-purpose operating system also runs a dynamic resource scheduling engine, which uses a group relative strategy optimization algorithm to schedule the resources of the industrial edge gateway in real time.

[0014] Secondly, to achieve the above objectives, embodiments of this application provide an industrial time-series data processing method applied to the industrial edge gateway as described in the first aspect, the method comprising:

[0015] Collect first-hand industrial data from industrial equipment;

[0016] The first industrial data is processed in real time, and the processed second industrial data is written to the shared memory area via a bus.

[0017] AI inference is performed based on the second industrial data.

[0018] The process includes real-time processing of the first industrial data and writing the processed second industrial data into a shared memory area via a bus, including:

[0019] The first industrial data is converted to a protocol to obtain second industrial data that meets the requirements of a preset protocol.

[0020] The second industrial data is written to the shared memory area via the bus.

[0021] The method further includes:

[0022] Receive control commands, which correspond to AI inference results;

[0023] Execute the first operation corresponding to the control command;

[0024] In this process, the first operation has a higher priority than the data acquisition, and the data acquisition has a higher priority than the protocol conversion.

[0025] Among these, AI inference based on the second industrial data includes:

[0026] Read the second industrial data from the shared memory area;

[0027] The second industrial data is preprocessed to obtain the third industrial data, wherein the preprocessing includes normalization and / or sliding window segmentation;

[0028] The third industrial data is subjected to KV compression and FPB quantization;

[0029] Based on the data stream load corresponding to the third industrial data, at least some expert modules in the AI ​​inference model are activated, and the at least some expert modules are used to perform fault location inference based on the processed third industrial data.

[0030] The method further includes:

[0031] Real-time monitoring of resource utilization corresponding to the AI ​​inference process;

[0032] The resources of the industrial edge gateway are dynamically scheduled based on the resource utilization rate.

[0033] The beneficial effects of the above technical solution in this application are as follows:

[0034] In embodiments of this application, the industrial edge gateway includes a real-time operating system for processing first industrial data provided by industrial equipment in real time to obtain second industrial data; and a general-purpose operating system for performing artificial intelligence (AI) inference based on the second industrial data. This enables deep collaboration between hard real-time control and complex AI inference in industrial edge scenarios, resolving the contradiction between real-time performance and general-purpose capabilities, achieving coexistence of hard real-time performance and high computing power, and addressing edge computing bottlenecks and architectural fragmentation. The real-time operating system and the general-purpose operating system share a memory area via a bus; the real-time operating system uses hardware interrupts to trigger the general-purpose operating system to adjust task priorities, thus enabling low-latency processing of industrial time-series data. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the structure of an industrial edge gateway according to an embodiment of this application;

[0036] Figure 2 This is one of the flowcharts illustrating the industrial time-series data processing method according to an embodiment of this application;

[0037] Figure 3 This is a schematic diagram of the system architecture where the industrial edge gateway of this application is located;

[0038] Figure 4 This is a second schematic flowchart of the industrial time-series data processing method according to an embodiment of this application;

[0039] Figure 5 This is a schematic diagram of the resource scheduling process in the embodiments of this application;

[0040] Figure 6 This is the third flowchart illustrating the industrial time-series data processing method according to an embodiment of this application. Detailed Implementation

[0041] To make the technical problems, technical solutions and advantages of this application clearer, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0042] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0043] In the various embodiments of this application, it should be understood that the sequence number of each process described below does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0044] In addition, the terms "system" and "network" are often used interchangeably in this article.

[0045] In the embodiments provided in this application, it should be understood that "B corresponding to A" means that B is associated with A, and B can be determined based on A. However, it should also be understood that determining B based on A does not mean that B is determined solely based on A; B can also be determined based on A and / or other information.

[0046] Before describing the embodiments of this application, the relevant technical points will first be described by way of example:

[0047] I. Time-series data processing system based on Linux edge gateway

[0048] Hardware architecture: The industrial gateway adopts a single Linux system (such as Ubuntu), equipped with an x86 / ARM processor and an AI acceleration card, and supports multi-protocol conversion (such as PROFIBUS, EtherCAT, etc.).

[0049] Data processing flow: Collect sensor time-series data using tools such as Telegraf and store it in a local database (such as KaiwuDB); extract time-series features using the traditional MHA mechanism and deploy a lightweight model (such as DeepSeek-r1 with 1.5B parameters) for predictive maintenance; interact with the cloud platform via REST API.

[0050] Limitations: 1) Insufficient real-time performance: A single Linux system cannot meet the hard real-time requirements of industrial control tasks (such as PLC instruction response delay > 1ms); 2) Low computational efficiency: The traditional MHA mechanism has high KV cache usage, resulting in high memory pressure on edge devices and inability to process long-cycle time-series data streams; 3) Resource conflicts: AI inference tasks and real-time control tasks share CPU resources, causing process scheduling delays.

[0051] The above system has the following defects:

[0052] 1) The conflict between real-time performance and versatility: Because a single Linux system lacks the priority preemption mechanism of a real-time operating system (RTOS), it cannot isolate real-time and non-real-time tasks. Therefore, a single Linux system cannot meet the μs-level latency requirements of hard real-time tasks in industrial scenarios (such as PLC instruction response and emergency equipment status control). For example, in predictive maintenance scenarios, AI inference tasks and real-time control tasks share the central processing unit (CPU) resources, resulting in latency exceeding 1ms for critical control instructions. Therefore, a single system (such as pure Linux) cannot simultaneously handle hard real-time control and complex data processing.

[0053] 2) Edge computing bottleneck: Because the traditional MHA / MTP mechanism does not introduce a low-rank compression attention mechanism and is not adapted to the quantization deployment capability of edge devices, the memory usage is too high in long-cycle industrial time-series data processing (e.g., KV cache usage > 8GB when processing 24-hour vibration data), which makes it impossible for edge devices to support continuous inference, resulting in low efficiency on resource-constrained edge devices.

[0054] 3) Resource scheduling conflict: AI inference processes (such as TensorFlow Lite) and protocol conversion processes (such as Modbus to MQTT) compete for computing resources within the Linux system, causing the overall gateway response latency to fluctuate by more than 200% during sudden data surges.

[0055] II. In-depth exploration (e.g., DeepSeek): Cloud-based MLA-accelerated industrial data analysis systems

[0056] Architecture design: Upload industrial data to the cloud and use the MLA technology of the DeepSeek-V3 model for time series data analysis. Reduce the number of activated parameters through the Mixture of Experts (MoE) architecture (e.g., only activate 37 billion / 671 billion parameters).

[0057] Core optimizations: FP8 mixed-precision training reduces GPU memory usage by 30%; Multi-Token Prediction (MTP) improves training efficiency by 40%.

[0058] Limitations: 1) Lack of edge collaboration: Cloud processing results in high network latency (>100ms), which cannot meet the real-time control requirements; 2) High cost: The MoE architecture requires 24GB+ of video memory (such as RTX 4090), which greatly increases the deployment cost of edge devices; 3) Data security risks: Sensitive industrial data needs to be transmitted through the public network, which is easily intercepted or leaked.

[0059] Analysis of the technical shortcomings of the above system:

[0060] 1) Because MLA technology has not been brought down to the edge, the reliance on cloud computing power leads to a break in the data closed loop, resulting in high latency in cloud collaboration. Specifically, industrial data needs to be transmitted to the cloud for processing via the public network (such as the DeepSeek-V3 model), with a network round-trip latency of >100ms, which cannot meet the real-time requirements of the device control closed loop (such as the emergency stop command requiring a response of <10ms).

[0061] 2) The lack of integration with lightweight models (such as DeepSeek-r1:1.5b) and dynamic scheduling of heterogeneous edge computing power results in high edge deployment costs. Specifically, the MoE architecture of DeepSeek-V3 requires 24GB+ of video memory (such as RTX 4090), while typical industrial edge devices only support 8GB of video memory. Forced deployment requires additional hardware upgrade costs.

[0062] 3) Multimodal data fragmentation: Cloud-based solutions rely on ETL links to synchronize relational data and time-series data (such as KaiwuDB, which requires additional data pipeline configuration), resulting in a 30% increase in end-to-end inference latency.

[0063] Therefore, the existing solution has the following shortcomings:

[0064] The conflict between real-time performance and versatility: A single system (such as pure Linux) cannot simultaneously handle hard real-time control and complex data processing;

[0065] Edge computing bottleneck: Traditional MHA / MTP mechanisms are inefficient on resource-constrained edge devices;

[0066] Fragmented architecture: Insufficient collaboration between cloud and edge computing leads to a break in the data loop (e.g., prediction results cannot be fed back to the PLC in real time).

[0067] In view of this, embodiments of this application provide an industrial edge gateway, such as... Figure 1 As shown, the industrial edge gateway includes:

[0068] A real-time operating system is used to process the first industrial data provided by industrial equipment in real time to obtain the second industrial data; a general-purpose operating system is used to perform artificial intelligence (AI) inference based on the second industrial data.

[0069] This creates an industrial edge gateway with a hybrid architecture based on a real-time operating system and a general-purpose operating system, enabling the coordinated operation of hard real-time tasks (such as PLC instruction response) and complex data processing (such as protocol conversion and AI inference) in industrial scenarios. The real-time operating system is responsible for μs-level real-time control, while the general-purpose operating system provides protocol stack support and AI framework deployment capabilities.

[0070] For example, the first industrial data is, for instance, industrial time-series data.

[0071] For example, the real-time operating system is an IoT operating system, specifically a one operating system (OS); the general-purpose operating system is a Linux OS. Based on this, the industrial edge gateway can include two domains: a one OS real-time domain and a Linux general-purpose domain, where the one OS runs in the one OS real-time domain and the Linux OS runs in the Linux general-purpose domain.

[0072] For example, the one OS real-time domain includes a real-time control unit, such as an ARM Cortex-R52 processor, which runs One OS and supports hard real-time protocols, such as PROFINET / Ethernet for Control Automation Technology (EtherCAT). The response latency of this real-time control unit is <50μs.

[0073] For example, the Linux OS includes an AI inference unit, such as an AI processor, which runs the Linux OS. This AI processor is equipped with 8GB of dedicated video memory and supports FP8 mixed-precision computing.

[0074] For example, the AI ​​inference results of a general-purpose operating system can be used to assist in the generation of control signaling to control industrial equipment.

[0075] The real-time operating system and the general-purpose operating system share a memory area via a bus; for example, the two systems share 4GB of DDR4 memory via a PCIe 3.0 bus, enabling zero-copy data transfer. This can improve the real-time performance of data processing to a certain extent.

[0076] The real-time operating system (RTOS) triggers the general-purpose operating system (GOS) to adjust task priorities via hardware interrupts; for example, the ROS's real-time tasks trigger the GOS's inference task priority adjustment via GPIO interrupts. This ensures low-latency processing of industrial data, thereby further improving the real-time performance of data processing.

[0077] In the embodiments of this application, the industrial edge gateway includes a real-time operating system for processing first industrial data provided by industrial equipment in real time to obtain second industrial data; and a general-purpose operating system for performing artificial intelligence (AI) inference based on the second industrial data. The real-time operating system and the general-purpose operating system share a memory area via a bus. The real-time operating system triggers the general-purpose operating system to adjust task priorities via hardware interrupts. Thus, through this hybrid architecture, real-time control tasks and AI inference tasks can be isolated, achieving coexistence of hard real-time performance and high computing power. This solves the problems of the contradiction between real-time performance and general-purpose capabilities, the edge computing bottleneck of a single Linux system, and architectural fragmentation.

[0078] As an optional implementation, the real-time operating system is used to process the first industrial data provided by the industrial equipment in real time, including:

[0079] The system collects the first industrial data in real time, performs protocol conversion on the first industrial data to obtain the second industrial data, and writes the second industrial data into the shared memory area via a bus.

[0080] For example, the first industrial data is industrial equipment sensor data (such as vibration and temperature signals), and the first industrial data can be acquired in real time by reading the industrial equipment sensor data at a 1ms cycle.

[0081] For example, protocol conversion can be used to convert data from serial protocols used for communication between industrial devices, such as Modbus RTU, into Message Queuing Telemetry Transport (MQTT) JSON format (conversion latency <200μs).

[0082] The above optional implementation is executed periodically every 1ms; the execution method is to directly read the sensor register via DMA and use a lightweight JSON parsing library (such as Jansson) to complete the protocol conversion.

[0083] Furthermore, as an optional implementation, the real-time operating system is also used to receive control commands and execute a first operation corresponding to the control commands;

[0084] In this process, the first operation has a higher priority than the data acquisition, and the data acquisition has a higher priority than the protocol conversion.

[0085] For example, the control command may be a command related to the result of AI inference. For instance, if the result of AI inference is an industrial equipment malfunction, then the control command may be a command to control the industrial equipment to shut down.

[0086] For example, setting the priority of the first operation to the highest priority enables the priority execution of control instructions (such as PLC instructions).

[0087] Based on the above implementation, as an optional implementation, the general operating system runs an AI inference model, which is used to execute the AI ​​inference. The AI ​​inference model is an FP8 quantized deployment model, and the AI ​​inference model is supported by a multi-head potential attention MLA acceleration engine.

[0088] Among the above-mentioned optional implementation methods, the AI ​​inference model is, for example, the DeepSeek-r1 model (1.5B parameters). By performing 8-bit floating-point quantization on this model, the inference latency can be reduced by 40%, enabling low-latency and high-precision processing of industrial time-series data. In addition, the video memory requirement is reduced from 24GB to 4GB, thereby reducing the video memory usage of the industrial edge gateway by 87.5%. This allows the industrial edge gateway to support 24-hour continuous time-series data inference and is compatible with edge devices with 8GB of video memory, reducing the high cost requirements of hardware.

[0089] The aforementioned AI inference model is supported by the MLA acceleration engine, which enables accelerated processing of industrial time-series data, achieving low-latency and high-precision processing of industrial time-series data.

[0090] As an optional implementation, the general-purpose operating system also runs a dynamic resource scheduling engine. This engine utilizes Group Relative Policy Optimization (GRPO) to schedule the resources of the industrial edge gateway in real time. This allows for real-time adjustment of the resource allocation weights between the Neural Network Processing Unit (NPU) and the CPU; for example, CPU redundancy can be enabled when the NPU utilization exceeds 80%. This ensures that the dynamic resource utilization rate is increased to at least 90%.

[0091] The industrial edge gateway in this application embodiment is based on the One OS-Linux hybrid operating system and DeepSeek MLA (multi-head latent attention) technology to build an industrial edge gateway that coordinates real-time control and AI inference, thereby achieving low-latency and high-precision processing of industrial time-series data.

[0092] Embodiments of this application also provide an industrial time-series data processing method utilizing the aforementioned industrial edge gateway, such as... Figure 2 As shown, the method includes:

[0093] Step 201: Collect the first industrial data from the industrial equipment.

[0094] For example, step 201 above can be implemented by the real-time operating system of the industrial edge gateway. For instance, the real-time operating system can periodically read the sensor registers of the industrial equipment through DMA at a period of 1ms to obtain the first industrial data.

[0095] Step 202: Process the first industrial data in real time, and write the processed second industrial data into the shared memory area via the bus.

[0096] For example, step 202 above can also be implemented by the real-time operating system of the industrial edge gateway. That is, the real-time operating system periodically reads the sensor registers of the industrial equipment directly through DMA to obtain the first industrial data, and processes the first industrial data in real time so as to store the second industrial data obtained in real time into the shared memory area of ​​the industrial edge gateway through the bus, thereby realizing zero-copy data transmission and providing a foundation for the high real-time requirements of industrial time-series data.

[0097] Step 203: Perform AI inference based on the second industrial data.

[0098] For example, step 203 above can be implemented by the general operating system of the industrial edge gateway. The general operating system can read the second industrial data stored in the shared memory area through the bus and use the acquired second industrial data to perform AI inference, so as to control the industrial equipment based on the results of AI inference.

[0099] In the embodiments of this application, firstly, first industrial data from industrial equipment is collected; secondly, the first industrial data is processed in real time, and the processed second industrial data is written into a shared memory area via a bus; and thirdly, AI inference is performed based on the second industrial data. In this way, the collaborative operation of hard real-time tasks and complex data processing in industrial scenarios is realized using a heterogeneous industrial edge gateway, achieving low-latency and high-precision processing of industrial time-series data, solving the problem of the contradiction between universality and real-time performance, as well as the problem of high latency in cloud collaboration.

[0100] As an optional implementation, step 201 above, which involves real-time processing of the first industrial data and writing the processed second industrial data into a shared memory area via a bus, includes:

[0101] The first industrial data is converted to a second industrial data that meets the requirements of a preset protocol. For example, this step may be to convert Modbus RTU protocol data into MQTT JSON format, wherein the conversion delay between the two protocols is less than 200μs.

[0102] The second industrial data is written to the shared memory area via the bus. This achieves zero-copy data transmission.

[0103] Furthermore, as an optional implementation, the method also includes:

[0104] The system receives control commands, which correspond to AI inference results. For example, the control commands can be sent from the cloud; in other words, the industrial edge gateway sends the AI ​​inference results to the cloud, which then determines the status of the industrial equipment based on the AI ​​inference results and sends control commands to the industrial edge gateway accordingly. The status of the industrial equipment may include, for example, fault, normal, or abnormal. Alternatively, the control commands can be generated by the industrial edge gateway based on the AI ​​inference results, specifically, control commands generated by the general-purpose operating system of the industrial edge gateway. In this case, the general-purpose operating system can send the control commands to the real-time operating system.

[0105] The first operation corresponding to the control command is executed; for example, when the control command instructs the industrial equipment to stop working, the industrial edge gateway, specifically the real-time operating system of the industrial edge gateway, controls the industrial equipment to shut down.

[0106] In this configuration, the first operation has a higher priority than data acquisition, and the data acquisition has a higher priority than the protocol conversion. Thus, upon receiving a control command, the industrial edge gateway can immediately execute the first operation corresponding to that control command, enabling the industrial edge gateway to prioritize the execution of PLC instructions.

[0107] As an optional implementation, step 203 involves performing AI inference based on the second industrial data, including:

[0108] Sub-step one: Read the second industrial data from the shared memory; for example, this step is for a general operating system to read the second industrial data from the shared memory area via a bus.

[0109] Sub-step two: Preprocess the second industrial data to obtain the third industrial data. The preprocessing includes normalization and / or sliding window segmentation. For example, normalization may be performed using Z-Score, and the window length in the sliding window segmentation operation may be, for example, 256, and the step size may be, for example, 64.

[0110] Sub-step 3: Perform KV compression and FPB quantization on the third industrial data.

[0111] For example, sub-step 3 above can be implemented by the DeepSeek MLA acceleration engine. DeepSeek's MLA mechanism reduces KV cache usage by 50% through low-rank compression of the key-value matrix.

[0112] KV compression, also known as KV cache compression, specifically compresses the original key-value matrix (d=1024) into a low-dimensional latent space (d_c=128), thus reducing memory usage by 87.5%. For example, the KV cache compression process can be represented as:

[0113] c_{kv,t} = W_{dkv} h_t (compression);

[0114] k_{c,t} = W_{uk} c_{kv,t} (Reconstruct Key);

[0115] v_{c,t} = W_{uv} c_{kv,t} (Refactoring Value).

[0116] Where h_t represents the hidden state at time step t, W_{dkv} represents the compression (dimensionality reduction) matrix, c_{kv,t} represents the compressed latent state, W_{uk} and W_{uv} represent the reconstruction (dimensionality increase) matrix, and k_{c,t} and v_{c,t} represent the reconstructed key and value.

[0117] Among them, FP8 quantization can convert the data in the KV cache from a high-precision format to an 8-bit floating-point number for storage, which can reduce inference latency by 40%.

[0118] Sub-step four: Based on the data stream load corresponding to the third industrial data, activate at least some expert modules in the AI ​​inference model, wherein the at least some expert modules are used to perform fault location inference based on the processed third industrial data.

[0119] For example, the data stream load can be represented by the number of packets per second. If the number of packets per second is greater than 1000, the expert module size is automatically switched, and the kernel startup latency is optimized by a graph optimization method that reduces the processor kernel startup latency.

[0120] It's important to note that in the MoE model, the massive neural network is broken down into multiple smaller subnetworks, each acting as an "expert." Normally (95% of the time): only a small number of experts are active (5% of parameters activated), allowing the gateway to handle the workload easily with low power consumption and stable latency. In critical moments (such as when fault signs are detected): all experts are instantly activated (100% of parameters) for the most precise analysis, ensuring no cases are missed.

[0121] The above implementation combines low-rank compression and FP8 quantization, enabling the AI ​​model to support 24-hour continuous time-series data inference on an 8GB video memory device, achieving low-latency and high-precision processing of industrial time-series data.

[0122] Furthermore, as an optional implementation, the method also includes:

[0123] Real-time monitoring of resource utilization during the AI ​​inference process; for example, resource utilization metrics such as NPU utilization and memory bandwidth can be calculated every 5 seconds.

[0124] Based on the resource utilization rate, the resources of the industrial edge gateway are dynamically scheduled. For example, dynamic scheduling of industrial edge gateway resources can be achieved by updating scheduling weights through a policy gradient algorithm. As a special case, when the NPU is overloaded, for example, when the NPU utilization rate is greater than 95%, computing tasks can be automatically migrated to the CPU. In this way, although the performance may decrease by 50%, the availability of the system can be guaranteed.

[0125] Among the above implementation methods, in the MoE architecture, expert modules are dynamically activated according to load, and resource utilization is monitored in real time. Resource scheduling is performed based on the GRPO strategy to optimize resource allocation. In this way, the resource utilization rate of industrial edge gateways is increased from 60% in traditional solutions to 90%, while ensuring that latency fluctuation is less than 10% under high load scenarios.

[0126] It should be noted here that, as an example, the industrial time-series data processing method described above in this application can be implemented by, for example, by, Figure 3 The system architecture shown consists of an industrial edge gateway layer, a cloud platform / control center, and an industrial equipment layer; the data processing flow of the three is as follows: Figure 4 As shown, it includes the following steps:

[0127] Step 1: Industrial equipment sends sensor data (Modbus RTU) to the one OS real-time domain.

[0128] Step 2: The protocol conversion module in the one OS real-time domain performs protocol conversion, converting Modbus RTU data into MQTT protocol data;

[0129] Step 3: The one OS real-time domain writes the transformed data to the shared memory area; the data written to the shared memory area is in JSON format.

[0130] Step 4: Trigger the Linux general domain's MLA engine to execute the following steps: First, use the KV compression algorithm to compress the data read from the shared memory area, reducing the dimensions from 1024 to 128. Second, process each batch of data in a loop. In low-load scenarios, such as when CPU utilization is less than 60%, activate the 5% expert module to perform FP8 quantization inference. In high-load scenarios, such as when CPU utilization is greater than or equal to 60%, activate the 100% expert module and optimize resource scheduling through GRPO parallel optimization.

[0131] Step 5: The Linux general domain uploads the prediction results, such as the predicted device health status, to the cloud platform;

[0132] Step 6: The cloud platform sends control commands to industrial equipment through the one OS real-time domain, such as shutdown maintenance and parameter adjustment;

[0133] Step 7: The industrial equipment sends an execution status acknowledgment (ACK) to the cloud platform through the one OS real-time domain.

[0134] The specific implementation of parallel resource scheduling optimization via GRPO in step 4 is as follows: Figure 5 As shown, it includes the following steps:

[0135] Step 501: Continuous data stream input;

[0136] Step 502: The system status monitor continuously collects data such as NPU / CPU utilization.

[0137] Step 503: The monitoring data is sent to the GRPO decision engine;

[0138] Step 504: Every 5 seconds, decide whether the current state is a normal state or an emergency state. If it is a normal state, proceed to step 505; if it is an emergency state, proceed to step 507.

[0139] Step 505: Fine-tune the scheduling weights using the policy gradient algorithm;

[0140] Step 506: Update the scheduler configuration based on the fine-tuning results of step 506;

[0141] Step 507: Trigger the fault rollback mechanism;

[0142] Step 508: Immediately migrate some NPU tasks to the CPU;

[0143] Step 509: The scheduler assigns the next computation task according to the new weights;

[0144] Step 510: The data processing pipeline executes using the updated resources.

[0145] Below, in conjunction with Figure 6 Another embodiment of the industrial time-series data processing method of this application will be described below, which specifically includes the following steps:

[0146] Step 601: One OS performs real-time domain data acquisition;

[0147] Step 602: The one OS real-time domain performs protocol conversion on the acquired data; specifically, for example, converting Modbus data to MQQTT data;

[0148] Step 603: Data transfer to shared memory, specifically including: the one OS real-time domain writes the protocol-converted data to the shared memory area via the bus, and the Linux general domain reads the data from the shared memory area via the bus;

[0149] Step 604: AI intelligent analysis, specifically: The Linux general domain's MLA engine performs AI intelligent analysis on the data read from the shared memory area;

[0150] Step 605: The Linux general domain queries the pre-configured rule base for the mapping logic corresponding to the AI ​​intelligent analysis results and generates control commands; wherein, the control commands are related to the queried mapping logic;

[0151] Optionally, the mapping logic in the rule base can be updated based on the execution of control commands;

[0152] Step 606: The Linux general domain sends the control command transmission results back to the one OS real-time domain; specifically: the cloud generates control commands based on the decision results and sends the control commands to the one OS real-time domain;

[0153] Step 607: Instruction execution; specifically: One OS controls the PLC.

[0154] This enables the generation and distribution of control commands by the industrial edge gateway without the involvement of the cloud platform, and achieves an end-to-end control loop with a latency of less than 5ms.

[0155] The implementation process of the industrial time-series data processing method of this application embodiment will be described below for different scenarios.

[0156] Example 1: Predictive Maintenance Scenario

[0157] Hardware configuration: Huawei Ascend Atlas 500 + ARM Cortex-R52

[0158] Data flow:

[0159] OneOS acquires motor vibration signals at a 1ms cycle (Modbus RTU).

[0160] The Linux-side MLA engine detected a sudden change in vibration amplitude (>5g) and activated the 100% expert module for fault location.

[0161] The GRPO strategy adjusts the NPU computing resource allocation weight from 0.5 to 0.9 within 2 seconds, reducing inference latency from 15ms to 8ms.

[0162] The predicted result (bearing wear > 90%) is pushed to the cloud via MQTT, triggering a PLC stop command.

[0163] Example 2: Protocol Conversion Optimization

[0164] When processing complex OPC UA tags (nested structures), the MLA engine automatically enables a multi-head potential attention mechanism, reducing protocol parsing time from 20ms to 5ms.

[0165] Embodiments of this application also provide an industrial time-series data processing apparatus, which is, for example, the aforementioned industrial edge gateway or a module within the industrial edge gateway, the apparatus comprising:

[0166] The data acquisition module is used to collect initial industrial data from industrial equipment.

[0167] The processing module is used to process the first industrial data in real time and write the processed second industrial data into the shared memory area via the bus.

[0168] The inference module is used to perform AI inference based on the second industrial data.

[0169] The processing module includes:

[0170] The protocol conversion submodule is used to convert the first industrial data into a protocol to obtain second industrial data that meets the preset protocol requirements.

[0171] The data writing submodule is used to write the second industrial data into the shared memory area via the bus.

[0172] The device further includes:

[0173] A receiving module is used to receive control commands, which correspond to AI inference results;

[0174] The execution module is used to execute the first operation corresponding to the control command;

[0175] In this process, the first operation has a higher priority than the data acquisition, and the data acquisition has a higher priority than the protocol conversion.

[0176] The reasoning module includes:

[0177] A reading submodule is used to read the second industrial data from the shared memory area;

[0178] The first processing submodule is used to preprocess the second industrial data to obtain the third industrial data. The preprocessing includes normalization and / or sliding window segmentation.

[0179] The second processing submodule is used to perform KV compression and FPB quantization on the third industrial data;

[0180] An activation submodule is used to activate at least some expert modules in the AI ​​inference model based on the data stream load corresponding to the third industrial data. The at least some expert modules are used to perform fault location inference based on the processed third industrial data.

[0181] The device further includes:

[0182] The monitoring submodule is used to monitor the resource utilization rate corresponding to the AI ​​inference process in real time.

[0183] The scheduling submodule is used to dynamically schedule the resources of the industrial edge gateway based on the resource utilization rate.

[0184] It should be noted that the industrial time-series data processing apparatus provided in this application embodiment can implement all the method steps implemented in the above industrial time-series data processing method embodiment and can achieve the same technical effect. Here, the parts that are the same as those in the method embodiment and the beneficial effects will not be described in detail.

[0185] An embodiment of this application also provides an industrial edge gateway, including a transceiver, a processor, a memory, and a program or instructions stored in the memory and executable on the processor; when the processor executes the program or instructions, it implements the above-described industrial time-series data processing method.

[0186] The transceiver is used to receive and send data under the control of the processor.

[0187] The bus architecture can include any number of interconnected buses and bridges, specifically linking various circuits of one or more processors (represented by processors) and memories (represented by memory). The bus architecture can also link various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. The bus interface provides the interface. A transceiver can be multiple elements, including transmitters and receivers, providing a unit for communicating with various other devices over a transmission medium.

[0188] The processor is responsible for managing the bus architecture and general processing, while the memory stores the data used by the processor during operation.

[0189] This application provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the steps in the industrial time-series data processing method described above and achieve the same technical effect. To avoid repetition, further details are omitted here.

[0190] The processor mentioned above is the processor in the industrial edge gateway described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0191] In this embodiment, the module can be implemented in software so that it can be executed by various types of processors. For example, an identified executable code module may include one or more physical or logical blocks of computer instructions, which may be constructed as objects, procedures, or functions. Nevertheless, the executable code of the identified module does not need to be physically located together, but may include different instructions stored in different bits, which, when logically combined, constitute the module and achieve the module's intended purpose.

[0192] Through the above description of the implementation methods, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, disk, optical disk) and includes several instructions for executing the methods described in the various embodiments of this application.

[0193] Therefore, embodiments of this application also provide a computer program product, including computer instructions, which, when executed by a processor, implement the industrial time-series data processing method described above and achieve the same technical effect. To avoid repetition, further details are omitted here.

[0194] In practice, an executable code module can be a single instruction or many instructions, and can even be distributed across multiple different code segments, different programs, and across multiple memory devices. Similarly, operational data can be identified within the module and can be implemented in any suitable form and organized within any suitable data structure. This operational data can be collected as a single dataset or distributed across different locations (including different storage devices), and can exist, at least in part, solely as electronic signals within the system or network.

[0195] When a module can be implemented using software, considering the current level of hardware technology, modules that can be implemented in software can be implemented using hardware circuits by those skilled in the art to achieve the corresponding functions, without considering cost. These hardware circuits include conventional very-large-scale integrated circuits (VLSI) or gate arrays, as well as existing semiconductors such as logic chips and transistors, or other discrete components. Modules can also be implemented using programmable hardware devices, such as field-programmable gate arrays, programmable array logic, and programmable logic devices.

[0196] The exemplary embodiments described above are illustrated with reference to the accompanying drawings. Many different forms and embodiments are feasible without departing from the spirit and teachings of this application, and therefore, this application should not be construed as limiting the exemplary embodiments set forth herein. Rather, these exemplary embodiments are provided to make this application complete and convey the scope of this application to those skilled in the art. In these drawings, component dimensions and relative dimensions may be exaggerated for clarity. The terminology used herein is for the purpose of describing particular exemplary embodiments only and is not intended to be limiting. As used herein, unless explicitly stated otherwise, the singular forms “a,” “an,” and “the” are intended to include all such forms. It will be further understood that the terms “comprising” and / or “including”, when used in this specification, indicate the presence of the stated features, integers, steps, operations, components, and / or elements, but do not exclude the presence or addition of one or more other features, integers, steps, operations, components, elements, and / or groups thereof. Unless otherwise indicated, a range of values ​​is stated to include the upper and lower limits of the range and any subranges therebetween.

[0197] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An industrial edge gateway, characterized in that, The industrial edge gateway includes: A real-time operating system is used to process the first industrial data provided by industrial equipment in real time to obtain the second industrial data; A general-purpose operating system for performing artificial intelligence (AI) inference based on the second industrial data; The real-time operating system and the general-purpose operating system share a memory area via a bus; the real-time operating system uses hardware interrupts to trigger the general-purpose operating system to adjust task priorities.

2. The industrial edge gateway according to claim 1, characterized in that, The real-time operating system is used to process the first industrial data provided by the industrial equipment in real time, including: The system collects the first industrial data in real time, performs protocol conversion on the first industrial data to obtain the second industrial data, and writes the second industrial data into the shared memory area via a bus.

3. The industrial edge gateway according to claim 2, characterized in that, The real-time operating system is also used to receive control commands and execute a first operation corresponding to the control commands; In this process, the first operation has a higher priority than the data acquisition, and the data acquisition has a higher priority than the protocol conversion.

4. The industrial edge gateway according to any one of claims 1 to 3, characterized in that, The general-purpose operating system runs an AI inference model, which is used to perform AI inference. The AI ​​inference model is an FP8 quantized deployment model, and the AI ​​inference model is supported by a multi-head potential attention (MLA) acceleration engine.

5. The industrial edge gateway according to claim 1, characterized in that, The general-purpose operating system also runs a dynamic resource scheduling engine, which uses a group relative strategy optimization algorithm to schedule the resources of the industrial edge gateway in real time.

6. An industrial time-series data processing method, characterized in that, Applied to an industrial edge gateway as described in any one of claims 1 to 5, the method comprises: Collect first-hand industrial data from industrial equipment; The first industrial data is processed in real time, and the processed second industrial data is written to the shared memory area via a bus. AI inference is performed based on the second industrial data.

7. The method according to claim 6, characterized in that, The process involves real-time processing of the first industrial data and writing the processed second industrial data into a shared memory area via a bus, including: The first industrial data is converted to a protocol to obtain second industrial data that meets the requirements of a preset protocol. The second industrial data is written to the shared memory area via the bus.

8. The method according to claim 7, characterized in that, The method further includes: Receive control commands, which correspond to AI inference results; Execute the first operation corresponding to the control command; In this process, the first operation has a higher priority than the data acquisition, and the data acquisition has a higher priority than the protocol conversion.

9. The method according to claim 7, characterized in that, AI inference based on second industrial data includes: Read the second industrial data from the shared memory area; The second industrial data is preprocessed to obtain the third industrial data, wherein the preprocessing includes normalization and / or sliding window segmentation; The third industrial data is subjected to KV compression and FPB quantization; Based on the data stream load corresponding to the third industrial data, at least some expert modules in the AI ​​inference model are activated, and the at least some expert modules are used to perform fault location inference based on the processed third industrial data.

10. The method according to any one of claims 6 to 9, characterized in that, The method further includes: Real-time monitoring of resource utilization during the AI ​​inference process; The resources of the industrial edge gateway are dynamically scheduled based on the resource utilization rate.

Images