Lightweight intelligent terminal and control system for real-time data processing

Abstract

The application discloses a kind of lightweight intelligent terminal and control system of real-time data processing, it is related to lightweight intelligent terminal real-time data processing technical field, including terminal node and two core components of cloud management and control platform;Terminal node is equipped with the low-power simplified hardware architecture of scene adaptation pruning, built-in quantization compression lightweight inference model, can complete the whole process of original perception data preprocessing, feature extraction and abnormal inference locally, only upload abnormal associated features and event data on demand to cloud end;Terminal built-in dynamic computing power scheduling and weak network breakpoint transmission module, can adjust computing power distribution according to real-time load, automatically cache data to be transmitted when network quality is poor.The application greatly reduces the hardware cost and operating power consumption of edge terminal, and can adapt to field deployment scenarios without external power supply.The local inference mode can effectively reduce bandwidth occupation, improve the real-time performance of abnormal response, and the weak network adaptation mechanism can ensure data transmission stability.

Description

Technical Field

[0001] This invention relates to the field of lightweight smart terminal real-time data processing technology, and in particular to a lightweight smart terminal and control system for real-time data processing. Background Technology

[0002] With the large-scale implementation of the Industrial Internet and the ubiquitous Internet of Things, the demand for real-time data processing capabilities on the edge side continues to increase in scenarios such as industrial equipment status monitoring, operation and maintenance of new energy power stations in the field, and deployment of IoT sensing nodes in cities. The core industry needs focus on four dimensions: low deployment cost, low operating power consumption, low transmission bandwidth usage, and high real-time response. At the same time, it is necessary to adapt to complex deployment environments with no external power supply and unstable network quality.

[0003] One of the current mainstream technical approaches is a general-purpose edge terminal paired with a centralized cloud processing solution. This solution works by having the edge terminal only handle raw data acquisition, uploading all raw data directly to the cloud platform. All data preprocessing, anomaly detection, and analysis / decision-making operations are handled by cloud computing power. This solution utilizes general-purpose edge terminals with ARM or x86 architectures, retaining all general-purpose peripherals and computing units, supporting multiple types of sensor access, and is complemented by a cloud-based big data processing platform and distributed inference cluster. The advantages of this solution are low development difficulty on the terminal side, sufficient cloud computing power to support complex algorithm models, and suitability for small-batch, stable network indoor deployment scenarios. However, this solution has significant drawbacks: full raw data transmission consumes extremely high bandwidth, resulting in high bandwidth operating costs; data loss and analysis delays exceeding several minutes are prone to occur when the network is unstable, failing to meet real-time requirements; and the general-purpose edge terminal has many redundant peripherals and high power consumption, making it unsuitable for outdoor deployment scenarios without external power supply.

[0004] Another mainstream technical approach is the high-performance edge intelligent terminal local full-data processing solution. This solution works by having the terminal equipped with a high-performance main control chip to process and infer all data locally, uploading only the final analysis results. This is achieved by having the edge terminal integrate a GPU or NPU unit with a complete deep learning inference framework, running a high-precision floating-point inference model locally. The advantages of this solution are that it does not rely on cloud computing power, has low response latency, and does not require full data upload, thus meeting privacy protection requirements. However, this solution also has significant drawbacks: the hardware cost of a high-performance edge terminal is 5 to 10 times that of a regular data acquisition terminal, and it lacks model lightweighting and computing power scheduling optimization, resulting in high power consumption. It also cannot adapt to low-power deployment requirements, leading to excessively high overall investment costs and significant maintenance difficulties during large-scale deployment.

[0005] Neither of the two existing technologies can simultaneously meet the multiple requirements of low power consumption, low cost, low bandwidth, and high real-time performance. They are not well adapted to large-scale distributed IoT and industrial monitoring scenarios. The industry urgently needs targeted solutions to overcome these technical contradictions. Summary of the Invention

[0006] This invention proposes a lightweight intelligent terminal and control system for real-time data processing to solve the problems mentioned in the prior art, such as the inability of existing edge data processing solutions to simultaneously meet the requirements of low power consumption, low cost, low bandwidth and high real-time performance, and poor adaptability to complex scenarios with weak networks in the wild.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a lightweight intelligent terminal for real-time data processing, comprising a hardware layer, an edge data processing layer running on the hardware layer, and a data transmission layer; The hardware layer is a low-power, streamlined architecture designed for industrial data acquisition scenarios, which removes redundant peripherals and general-purpose computing units that are not necessary for industrial scenarios. The hardware cost and power consumption are less than one-third of those of general-purpose edge terminals with the same computing power. The edge data processing layer has a built-in real-time data preprocessing module, a dynamic computing power scheduling module, and a local lightweight inference module. The real-time data preprocessing module is used to process the multi-source real-time data, including industrial sensor time series data, equipment operating status data, and structured feature data output by front-end acquisition equipment, by removing invalid values ​​and pre-extracting features in sequence, which greatly reduces the data dimensionality of subsequent processing. The dynamic computing power scheduling module is used to dynamically allocate hardware computing power resources based on real-time data access throughput and the length of the task queue to be processed, giving priority to ensuring the computing power quota of high-priority data processing tasks and avoiding increased processing delays caused by task backlog. The local lightweight inference module adopts a quantized and compressed lightweight neural network model, which can complete data anomaly identification and state prediction inference operations locally without relying on cloud computing power, and the inference latency meets the requirements of industrial real-time control. The data transmission layer is pre-configured with data upload rules, and only uploads the valid feature data and abnormal alarm association data after preprocessing and anomaly identification to the upper control system. It does not need to upload the full amount of original collected data, which greatly reduces the transmission bandwidth usage.

[0008] Preferably, the hardware layer adopts an industrial-grade main control chip with a RISC-V reduced instruction set architecture, cutting out unnecessary general-purpose floating-point arithmetic units, multimedia codec engines, consumer-grade display interfaces, and audio input / output peripherals. It integrates a dedicated time-series data acceleration processing core and a low-power solid-state storage unit that supports power-down protection. The dedicated time-series data acceleration processing core is specifically designed for hardware-level acceleration of filtering, sampling, and aggregation operations of time-series data, with a computational efficiency far exceeding that of general-purpose processing cores. The overall power consumption of the machine is far lower than that of general-purpose edge terminals with the same computing power, supporting low-power deployment modes such as battery power and solar power, and adapting to industrial field deployment scenarios without external power supply.

[0009] Preferably, the real-time data preprocessing module adopts a configurable sliding time window verification mechanism, which supports custom configuration of window length from milliseconds to minutes. It performs outlier detection and duplicate data verification on the collected data within a unit time window. Noise data exceeding the preset confidence interval and redundant data that appears more than a preset number of times are directly removed from the local cache. Only the original data that meets the data validity rules is retained for subsequent feature pre-extraction operations. The feature pre-extraction process prioritizes the extraction of statistical features and time-series features that are strongly correlated with anomaly identification, further compressing the data volume of subsequent processing.

[0010] Preferably, the dynamic computing power scheduling module has a built-in pre-calibrated computing power mapping table. The computing power mapping table is generated by calibrating the computing power requirements corresponding to different data access throughputs in a simulation environment. During the scheduling process, three core indicators are monitored in real time: the amount of data accessed per second, the length of the preprocessing queue, and the backlog of inference tasks. The corresponding computing power allocation strategy is matched, and idle processing cores and unnecessary peripheral interfaces are automatically shut down during idle periods, so that the terminal enters a low-power sleep mode. The corresponding processing core is only woken up when the data access is interrupted, which greatly reduces the standby power consumption of the terminal.

[0011] Preferably, the lightweight neural network model used in the local lightweight inference module is a lightweight classification or time series prediction derivative model that has undergone structured pruning and INT8 weight quantization. First, the basic floating-point model is trained in the cloud, then redundant convolutional kernels and fully connected layer neurons in the model are pruned, and then the 32-bit floating-point weights are quantized into 8-bit integers. The model size is much smaller than that of general inference models, the computational complexity is greatly reduced, it is adapted to low-computing-power simplified architecture hardware, and the local single inference time meets the millisecond-level real-time processing requirements. The inference conclusion can be directly output locally without waiting for the cloud to return the result.

[0012] Preferably, the data transmission layer has a built-in breakpoint resume and fragmented caching unit, which monitors the current network's packet loss rate, transmission latency, and available bandwidth in real time. When the network transmission quality is detected to be lower than a preset threshold, the real-time upload operation is paused, and the valid feature data to be uploaded and the abnormal alarm data are stored in the local non-volatile storage unit in the order of timestamps. Each data block is added with an integrity check code. When the network quality recovers to above the qualified threshold, the cached data is automatically re-uploaded according to the priority from oldest to newest time. After the re-upload is completed, the local cache copy is automatically deleted, which avoids data loss and does not occupy too much local storage resources.

[0013] Preferably, a lightweight intelligent terminal control system for real-time data processing is used to manage the lightweight intelligent terminal, including a cloud management platform, a terminal configuration management module, a multi-source data aggregation module, and a global visualization module; The cloud-based management platform supports multiple deployment modes, including public cloud, private cloud, and hybrid cloud, to meet the deployment needs of industrial scenarios of different scales. The terminal configuration management module supports two modes: batch configuration and single-terminal custom configuration. It sends data processing rules, inference model judgment thresholds, computing power scheduling strategies, and sliding window length configuration parameters to each connected lightweight smart terminal. The multi-source data aggregation module supports multiple commonly used industrial transmission protocols such as MQTT, HTTP, and CoAP. It receives valid feature data and abnormal alarm data uploaded by various lightweight smart terminals, and stores them into a dedicated time-series database after unifying the format and deduplicating the data. The global visualization module is used to provide multi-dimensional visualization of the operating status, computing load, data processing volume, and abnormal alarm information of the access terminals. It supports filtering and querying by region, terminal type, and time dimension, providing operation and maintenance personnel with an intuitive reference for the system's operating status.

[0014] Preferably, it also includes a global computing power scheduling module. The global computing power scheduling module is used to collect multi-dimensional operating data of all access terminals in real time, such as computing power utilization, memory utilization, network upload speed, queue length of pending tasks, and predicted value of regional data access volume. It dynamically adjusts the task allocation strategy according to the real-time data processing needs of each region. When the load of a single terminal is detected to exceed a preset threshold, some of the non-real-time processing tasks undertaken by that terminal are migrated to nearby low-load terminals to balance the computing power load of each terminal and avoid the problem of increased processing latency and data loss caused by single terminal overload.

[0015] Preferably, the system also includes a model iteration and update module. This module periodically retrains and optimizes the lightweight inference model based on the aggregated effective feature data and abnormal sample data. It then performs a differential comparison between the optimized new version model and the old version model currently running on the terminal, extracting only the weight difference between the two versions to generate a differential update package. The size of the differential update package is much smaller than the full model package. The differential update package is sent to the terminal using a breakpoint resume method. After receiving the update package, the terminal completes the model merging locally without restarting the terminal, significantly reducing bandwidth usage during model updates without affecting the normal operation of the terminal.

[0016] Preferably, the system also includes a security management module. This module employs a two-way authentication mechanism based on X.509 certificates for all access terminals. Only terminals holding valid certificates can access the management platform. All data transmitted between the terminal and the cloud is encrypted end-to-end using the national cryptographic SM4 symmetric encryption algorithm. The module monitors the terminal's operating status in real time. When an abnormal operating status is detected, such as unauthorized configuration modifications, uploading data in abnormal formats, or abnormal computing power usage, the module automatically isolates the abnormal terminal, stops receiving uploaded data from that terminal, and does not issue any configuration commands to it, thus preventing security risks from spreading to the entire system.

[0017] Compared with existing technologies, the beneficial effects of this invention are: This invention significantly reduces terminal hardware costs and power consumption by adopting a scenario-oriented simplified architecture for terminal hardware, coupled with a quantized and compressed lightweight inference model and a dynamic computing power scheduling mechanism. It effectively solves the shortcomings of existing general-purpose edge terminals, such as excessive redundant peripherals and high power consumption, and high investment costs for high-performance edge terminal hardware, which are unable to adapt to field deployment scenarios without external power supply. The terminal can operate stably for a long time through low-power power supply, which greatly reduces the deployment threshold of edge nodes.

[0018] This invention completes the entire process of data preprocessing and anomaly inference locally on the terminal side, uploading only valid features and anomaly-related data. Combined with a network quality-adaptive breakpoint resume caching mechanism, it significantly reduces data transmission bandwidth usage and achieves real-time response with extremely low latency at the edge. It effectively solves the shortcomings of existing cloud-based centralized processing solutions, such as high cost of full raw data transmission bandwidth, high inference latency, and data loss due to network instability. It can operate stably in weak network environments and meet the core requirements of industrial real-time control and real-time anomaly alarm.

[0019] The accompanying management and control system of this invention supports multi-mode deployment, batch configuration, differential model updates, and global computing power scheduling. Coupled with an end-to-end security management and control mechanism, it significantly reduces the operation and maintenance costs after large-scale deployment, improves the utilization rate of computing power resources of the entire system, and effectively solves the shortcomings of existing solutions such as high operation and maintenance difficulty, easy interruption of business during model updates, and low resource utilization when deployed on a large scale. It can adapt to deployment needs of different scales, such as single-site small factories and multi-site cross-regional projects. It can be widely used in multiple fields such as industrial equipment condition monitoring, new energy power station operation and maintenance, and field IoT sensing, and has extremely high promotion and application value. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the overall operation of the lightweight intelligent terminal real-time data processing system proposed in this invention. Figure 2 This is a flowchart of the local data processing process on the terminal side proposed in this invention; Figure 3 This is a flowchart of the operation process of the weak network adaptation module proposed in this invention; Figure 4 This is a flowchart of the cloud-based differential model update process proposed in this invention; Figure 5 This is a flowchart of the terminal dynamic computing power scheduling process proposed in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Reference Figures 1 to 5 This invention discloses a lightweight smart terminal for real-time data processing, comprising a hardware layer, an edge data processing layer running on the hardware layer, and a data transmission layer.

[0023] The hardware layer is a low-power, streamlined architecture specifically designed for scenarios such as industrial data acquisition, equipment status monitoring, and field IoT monitoring, with redundant peripherals removed. All non-essential consumer electronics peripherals and general-purpose computing units are removed, retaining only industrial-grade data transmission interfaces such as RS485, LoRa, Industrial Ethernet, and NB-IoT. The hardware material cost and rated operating power consumption are less than 30% of those of general-purpose edge terminals with the same computing power.

[0024] The edge data processing layer has a built-in real-time data preprocessing module, a dynamic computing power scheduling module, and a local lightweight inference module. The real-time data preprocessing module is used to process the multi-source real-time data, including industrial sensor time series data, motor current and voltage operation data, and structured feature data output by front-end acquisition devices, by removing invalid values ​​and pre-extracting features in sequence, which can compress the data dimension of subsequent processing by more than 90%.

[0025] The dynamic computing power scheduling module is used to dynamically allocate hardware computing power resources such as CPU cores and memory based on real-time data access throughput and the length of the task queue to be processed. It prioritizes the computing power quota of high-priority data processing tasks, with the computing power ratio of high-priority tasks such as anomaly alarm inference not less than 70%, which can avoid the increase in processing latency caused by task backlog. The end-to-end processing latency is stably controlled within 10ms.

[0026] The local lightweight inference module adopts a quantized compressed lightweight neural network model, which can complete data anomaly identification and equipment status prediction inference operations locally without relying on cloud computing power. The inference latency fully meets the requirements of industrial real-time control and real-time alarm for field events.

[0027] The data transmission layer is pre-configured with data upload rules, which only uploads the valid feature data and abnormal alarm association data after preprocessing and anomaly identification to the upper control system. It does not need to upload the full amount of original collected data, which can reduce the transmission bandwidth usage by more than 99%.

[0028] In this invention, the hardware layer adopts an industrial-grade main control chip with a RISC-V reduced instruction set architecture. Specifically, it can be an industrial-grade low-power main control chip with the Nuclei Technology RV32IMAC architecture. Unnecessary general-purpose floating-point arithmetic units, multimedia codec engines, consumer-grade display interfaces, and audio input / output peripherals are removed, and a dedicated timing data acceleration processing core and a low-power solid-state storage unit that supports power-down protection are integrated.

[0029] The dedicated time-series data acceleration processing core is specifically designed for hardware-level acceleration of sliding filtering, sampling, Fourier transform, and feature aggregation operations on time-series data. Its computational efficiency is more than five times that of a general-purpose processing core with the same computational load. The rated power consumption of the entire device is only 0.3W, far lower than the 2W or more power consumption of general-purpose edge terminals with similar computing power. It supports low-power deployment modes such as 3.7V lithium battery power supply and solar power supply. A single 2000mAh lithium battery can support continuous operation for more than 6 months, making it fully suitable for industrial and field deployment scenarios without external power supply.

[0030] In this invention, the real-time data preprocessing module employs a configurable sliding time window verification mechanism. The sliding window length supports a custom configuration from 10ms to 3600s, which can be adjusted according to different scenario requirements. For example, it can be set to 100ms for industrial motor fault monitoring scenarios and 1min for field environment monitoring scenarios. The module first performs outlier detection on the collected data within a unit time window, using the 3σ criterion to determine noisy data. The corresponding calculation formula is as follows: ; in the formula Let $i$ be the $i$-th sampled data point within the window. The mean of all sampled data within the window. The standard deviation of all sampled data within the window is given. Data points that satisfy the above formula are directly identified as noise data and discarded.

[0031] Subsequently, duplicate data verification is performed. Data with a numerical error of less than 0.1% uploaded from the same sensor at the same timestamp are considered redundant, and only one valid data point is retained. Finally, only the original data that meets the data validity rules is retained for subsequent feature pre-extraction operations. The feature pre-extraction process prioritizes extracting six statistical features strongly correlated with anomaly identification: maximum value, minimum value, mean, variance, peak-to-peak value, and kurtosis. It is not necessary to retain the original full sampling points, which can further compress the data volume by more than 94%. The pseudocode of the module's core logic is as follows: import numpy as np import scipy.stats def sliding_window_preprocess(window_data, sigma_threshold=3): # 3σ criterion for noise removal data_mean = np.mean(window_data) data_std = np.std(window_data) valid_data = window_data[np.abs(window_data - data_mean) <sigma_threshold * data_std] # Redundant data deduplication valid_data = np.unique(np.round(valid_data, decimals=3)) # Extracting strongly correlated features features = [ np.max(valid_data), np.min(valid_data), np.mean(valid_data), np.std(valid_data), np.ptp(valid_data), scipy.stats.kurtosis(valid_data) ] return features In this invention, the dynamic computing power scheduling module has a pre-calibrated computing power mapping table built-in. This computing power mapping table is generated by measuring and calibrating the computing power requirements corresponding to different data access throughputs under a simulation environment. During the scheduling process, the data access volume per second is monitored in real time. Preprocessing queue length Reasoning task backlog Three core metrics, matching corresponding computing power allocation strategies, and computing power allocation ratios. The calculation formula is: ; in the formula This represents the maximum amount of data that the terminal can support. This is the maximum length of the preprocessing queue. This represents the maximum backlog for the reasoning task. The value ranges from 0 to 1, corresponding to the percentage of processing cores that need to be enabled, for example... Only one low-power core is activated at a time. Two cores are enabled at the same time. All 3 cores are activated during off-peak hours. It automatically shuts down idle processing cores and unnecessary peripheral interfaces, putting the terminal into a low-power sleep mode. The corresponding processing core is only woken up when a data access interrupt is triggered, which can reduce the terminal's standby power consumption by more than 90%.

[0032] In this invention, the lightweight neural network model used in the local lightweight inference module is a lightweight classification or temporal prediction derivative model that has undergone structured pruning and INT8 weight quantization. The model generation process is as follows: First, a 32-bit floating-point precision base model is trained in the cloud using the PyTorch or TensorFlow framework. For classification scenarios, MobileNetV2 can be used as the base skeleton, and for temporal prediction scenarios, TCN temporal convolutional network can be used as the base skeleton. Then, an L1 regularized pruning strategy is used to prune weights with absolute values ​​less than 1. Redundant convolutional kernels and fully connected neurons are pruned at a rate of no less than 70%, and the model size can be compressed by more than 70%. Finally, the pruned model is subjected to INT8 weight quantization, and the quantization calculation formula is as follows: ; in the formula The original weights are 32-bit floating-point types. This is the quantization scaling factor. To quantize the zero point, The quantized 8-bit integer weights further compress the model size by 75%, resulting in a final model size of only about 7.5% of the original floating-point model. This significantly reduces computational complexity, making it perfectly compatible with low-computing-power, streamlined hardware architectures. Local single-inference time is less than 5ms, and inference accuracy is only within 0.2% of the original floating-point model, fully meeting the scene's accuracy requirements. The pseudocode for the core logic of module quantization is as follows: def int8_quantize(fp32_weights): # Calculate quantization parameters r_min = np.min(fp32_weights) r_max = np.max(fp32_weights) s = (r_max - r_min) / 255 z = round(-r_min / s) # Execution Quantification q_weights = np.round(fp32_weights / s + z).astype(np.int8) return q_weights, s, z In this invention, the data transmission layer incorporates a breakpoint resume and fragmented caching unit to detect the current network packet loss rate in real time. Transmission delay Available bandwidth Three network quality parameters, when detected , , If any one of the conditions is met, the network transmission quality is determined to be below a preset threshold. Real-time upload is paused, and the valid feature data and abnormal alarm data to be uploaded are fragmented into 1KB units. Each fragment is then stored in a local non-volatile storage unit after an MD5 integrity checksum is added. When a fault is detected... , , When all three conditions are met, the network quality is deemed to have recovered to acceptable levels. Cache data is automatically retransmitted according to time priority from oldest to newest. Integrity is checked after each segment is uploaded. Once the check passes, the local cache copy is automatically deleted to avoid consuming too much local storage resources. The maximum cache capacity of the terminal is 64MB. When storage space is insufficient, the oldest non-alarm data is overwritten first to ensure that abnormal alarm data is not lost.

[0033] This invention discloses a lightweight intelligent terminal control system for real-time data processing, used to manage and control the lightweight intelligent terminals described above, including a cloud-based management and control platform, a terminal configuration management module, a multi-source data aggregation module, and a global visualization module. The cloud-based management and control platform supports multiple deployment modes, including public cloud, private cloud, and hybrid cloud, and can adapt to the needs of industrial scenarios of different scales, such as single-site deployment in small factories and multi-site deployment across regional groups.

[0034] The terminal configuration management module supports both batch configuration and single-terminal custom configuration modes. It can send data processing rules, inference model judgment thresholds, computing power scheduling strategies, and sliding window length configuration parameters to each connected lightweight intelligent terminal. A single batch can support the simultaneous configuration of more than 1,000 terminals. The configuration is transmitted using the MQTT protocol, and the configuration takes effect within 1 second after the terminal receives it. The multi-source data aggregation module supports multiple commonly used industrial transmission protocols such as MQTT, HTTP, CoAP, and Modbus. It receives valid feature data and abnormal alarm data uploaded by each lightweight intelligent terminal, performs format unification and deduplication processing, and stores it in the InfluxDB dedicated time-series database. The data query efficiency is more than 10 times higher than that of ordinary MySQL relational databases.

[0035] The global visualization module is developed using the ECharts framework and is used to provide multi-dimensional visualization of the operating status, computing load, data processing volume, and abnormal alarm information of the access terminals. It supports filtering and querying by region, terminal type, and time dimension, and can provide operation and maintenance personnel with an intuitive reference for the system's operating status, enabling them to complete the entire system's operating status inspection without going to the site.

[0036] In this invention, the control system further includes a global computing power scheduling module. This module collects multi-dimensional operational data every 10 seconds, including computing power utilization, memory utilization, network upload speed, queue length of pending tasks, and predicted data access volume for the next hour for each region. It dynamically adjusts the task allocation strategy based on the real-time data processing needs of each region. When a single terminal's computing power utilization exceeds 80% or its queue length exceeds 50, it is determined that the terminal is overloaded. Non-real-time processing tasks undertaken by this terminal, such as historical feature statistics and report generation, are migrated to nearby low-load terminals with computing power utilization below 30%. This balances the computing power load across all terminals, preventing increased processing latency and data loss caused by single-terminal overload. This can increase the average computing power utilization of the entire system from 40% to over 70%.

[0037] In this invention, the control system further includes a model iteration update module. Every 7 days, the model iteration update module uses newly gathered effective feature data and abnormal sample data to retrain and optimize the lightweight inference model. The optimized new version model is compared with the old version model currently running on the terminal. Only the weight difference between the two versions of the model is extracted to generate a differential update package. The size of the differential update package is less than 10% of the full model package.

[0038] The differential update package is sent to the terminal using a breakpoint resume method. After receiving the update package, the terminal completes the model merging operation in the background without restarting the terminal. This does not affect the terminal's normal data processing business, significantly reduces bandwidth usage during model updates, and avoids network congestion caused by large-scale terminal synchronous updates.

[0039] In this invention, the control system further includes a security management module. The security management module adopts a two-way identity authentication mechanism based on X.509 certificates for all access terminals. When a terminal accesses the platform for the first time, it needs to submit a certificate issuance application. The management platform issues a unique identity certificate for the legitimate terminal through the built-in CA root certificate. Each subsequent connection requires two-way certificate verification. Only terminals holding legitimate certificates can access the management platform.

[0040] All data transmitted between the terminal and the cloud is encrypted end-to-end using the national standard SM4 symmetric encryption algorithm. The encryption key is automatically updated every 24 hours to prevent data theft and tampering. The module monitors the terminal's operating status in real time. When abnormal operating conditions are detected, such as unauthorized configuration modifications, uploading data in abnormal formats, or abnormal computing power usage exceeding 5 minutes, the abnormal terminal is automatically isolated, and uploading data from that terminal is stopped, and no configuration commands are issued to it, preventing security risks from spreading to the entire system.

[0041] Example 1: Industrial motor equipment condition monitoring scenario Application scenario description: A car parts production workshop has 200 three-phase asynchronous motors. The original solution used a general edge terminal to collect the raw current and voltage data of the motors at a sampling rate of 1kHz and directly uploaded them to the cloud for fault identification. This solution had problems such as high bandwidth consumption, fault identification delay of more than 1 minute, high power consumption of the general terminal, and the need to lay additional power supply lines, resulting in high deployment and operation costs.

[0042] Each motor is equipped with the lightweight intelligent terminal of this invention. The hardware layer adopts the RV32 architecture low-power main control, which directly connects to the current transformer and voltage transmitter of the motor to collect data. The preprocessing module has a sliding window configured with 100ms, uses the 3σ criterion to remove noise data caused by power grid fluctuations, and extracts 6 types of time-series features. The dynamic computing power scheduling module starts 2 processing cores according to the real-time data access volume of 1000 data points / second. The local lightweight inference module uses the pruned and quantized TCN model to identify bearing faults and rotor bar breakage faults, with a single inference time of 3ms. The data transmission layer only uploads feature data and fault alarm data.

[0043] The control system is deployed on the factory's local private cloud. The terminal configuration management module issues fault judgment thresholds in batches, the global visualization module displays the real-time operating status of all motors, the global computing power scheduling module balances the computing power load of 200 terminals, and the model iteration update module updates the model every 7 days with newly collected fault samples. The differential update package is only 80KB in size, and background updates do not affect production.

[0044] The end-to-end fault identification latency has been reduced from 1 minute to less than 10ms, the transmission bandwidth usage has been reduced from 100Mbps to 1Mbps, the bandwidth cost has been reduced by 99%, the terminal deployment cost has been reduced by 60%, no additional dedicated power supply lines are required, and the fault identification accuracy has been improved from 92% to 98%.

[0045] Example 2: Monitoring Scenario of Photovoltaic Power Station Modules in the Gobi Desert A 100MW photovoltaic power station in the Gobi Desert region of Northwest China has 5,000 photovoltaic module strings. The original solution used a general-purpose data acquisition terminal to transmit raw current and voltage data to the cloud. This solution had problems such as high terminal power consumption, insufficient power supply on cloudy or rainy days leading to power outages, unstable 4G network in the Gobi Desert causing a data packet loss rate of over 15%, and anomaly identification delay of over 5 minutes, resulting in extremely high operation and maintenance costs.

[0046] Each photovoltaic module string is equipped with the lightweight smart terminal of this invention. The hardware layer adopts a solar + lithium battery power supply scheme, which can support continuous operation for 7 days in cloudy and rainy weather when fully charged. The preprocessing module has a sliding window of 1 minute to remove noise data caused by sudden changes in light intensity and extract 5 types of features related to power generation. The local lightweight inference module uses a pruned and quantized classification model to identify component shadow and hotspot anomalies. The data transmission layer caches data to local Flash when the 4G signal is below a threshold and automatically retransmits it after the signal recovers. The control system is deployed on a public cloud. The global visualization module displays the real-time power generation of the entire power plant and the location of abnormal components. The security management module uses two-way certificate authentication to prevent unauthorized terminals from accessing and tampering with data.

[0047] Terminal battery life is increased by 6 times, data packet loss rate is reduced from 15% to below 0.1%, anomaly identification response time is reduced from 5 minutes to less than 1 second, and power plant operation and maintenance costs are reduced by 70%.

[0048] refer to Figure 1 This diagram illustrates the entire system's workflow, covering the complete transition logic from the terminal sensing end to the cloud-based management end. It clearly presents the core design principle of prioritizing local processing and uploading abnormal data on demand, clarifies the division of labor between the terminal and the cloud, and intuitively distinguishes this system from traditional full-data upload architectures. This helps users quickly understand the overall system operation logic, clarify the input-output relationships of each stage, and provides an overall framework for understanding the processes of subsequent sub-modules. It also visually demonstrates the system's core design direction of reducing bandwidth consumption and improving response speed.

[0049] refer to Figure 2 This diagram illustrates the local data processing flow on the terminal side, showcasing the entire process of local data processing that can be completed by the terminal without relying on cloud computing power. All processing steps presented in this diagram have been specifically optimized for low-computing-power, low-power terminals. The entire process, from raw data to inference results, can be completed without high-performance hardware support. It intuitively demonstrates the core implementation logic of the terminal's lightweight design, clearly defining the sequence and function of each processing step. This helps users understand the operational logic of the terminal completing local perception analysis without uploading the entire dataset, and also provides a process-level demonstration of the terminal's low-power, low-cost deployment characteristics.

[0050] refer to Figure 3 This diagram illustrates the workflow of the weak network adaptation module, showcasing the system's data transmission assurance logic in scenarios with weak or no network coverage. The diagram clearly presents the complete flow path of network quality detection, caching, and retransmission. Unlike traditional direct transmission, which is prone to packet loss, interruptions, and data loss, this diagram intuitively demonstrates the system's data reliability assurance mechanism in unstable network environments, such as in the field. It helps users understand the design logic for adapting to complex environments, clarifying the triggering conditions and operating rules of each step, and providing a clear, intuitive workflow demonstration for system deployments without stable public network coverage.

[0051] refer to Figure 4 This diagram illustrates the cloud-based differential model update process, showcasing the entire iterative update process of the terminal inference model in the cloud. The differential update mode presented in this diagram differs from the traditional full model update method, significantly reducing bandwidth consumption and terminal computing power overhead required for model updates. It intuitively demonstrates the convenient design for batch system operation and maintenance, helping users understand the low-overhead model iteration logic in large-scale terminal deployment scenarios, clarifying the interaction rules of each step, and providing process-level support for the system's adaptability to deployment needs of different scales and reduced operation and maintenance costs.

[0052] refer to Figure 5 This diagram illustrates the dynamic computing power scheduling process for terminals, showcasing the operational logic of dynamic computing power allocation. It clearly presents the complete path for adjusting computing power allocation based on the terminal's real-time load. Unlike traditional fixed computing power allocation, which is prone to issues like insufficient computing power leading to lag and wasted idle computing power, this diagram intuitively demonstrates the system's low-power operation guarantee mechanism. It helps users understand the design logic for terminals adapting to scenarios without external power supply, clarifies the triggering conditions and allocation rules for computing power scheduling, and provides a clear, process-level visualization of the terminal's long battery life and stable operation characteristics.

[0053] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A lightweight intelligent terminal for real-time data processing, comprising a hardware layer, an edge data processing layer running on the hardware layer, and a data transmission layer, characterized in that: The hardware layer is a low-power, streamlined architecture hardware with redundant peripherals removed. The edge data processing layer has a built-in real-time data preprocessing module, a dynamic computing power scheduling module, and a local lightweight inference module. The real-time data preprocessing module is used to perform invalid value removal and feature pre-extraction processing on the accessed multi-source real-time data. The dynamic computing power scheduling module is used to dynamically allocate hardware computing power resources based on real-time data access throughput. The local lightweight inference module uses a quantized compressed lightweight neural network model to complete data anomaly identification locally. The data transmission layer only uploads the pre-processed and anomaly-identified valid feature data and anomaly alarm data to the upper-level control system.

2. The lightweight intelligent terminal for real-time data processing according to claim 1, characterized in that, The hardware layer adopts a RISC-V reduced instruction set architecture main control chip, cuts out unnecessary general floating point arithmetic units and multimedia processing peripherals, and integrates a dedicated time-series data acceleration processing core and a low-power solid-state storage unit. The power consumption of the whole machine is much lower than that of general edge terminals with the same computing power.

3. A lightweight intelligent terminal for real-time data processing according to claim 1, characterized in that, The real-time data preprocessing module adopts a sliding time window verification mechanism to directly remove duplicate and noisy data that exceed the preset threshold range within a unit time window, and retain only the original data that meets the data validity rules for subsequent feature pre-extraction operations.

4. A lightweight intelligent terminal for real-time data processing according to claim 1, characterized in that, The dynamic computing power scheduling module has a built-in computing power mapping table. It matches the corresponding computing power allocation strategy according to the real-time monitored data access volume per second, and automatically shuts down idle processing cores and peripheral interfaces during idle periods.

5. A lightweight intelligent terminal for real-time data processing according to claim 1, characterized in that, The lightweight neural network model used in the local lightweight inference module is a lightweight classification or time series prediction derivative model after INT8 quantization and pruning. The model size is much smaller than that of the general inference model, and the local single inference time meets the millisecond-level real-time processing requirements.

6. A lightweight intelligent terminal for real-time data processing according to claim 1, characterized in that, The data transmission layer has a built-in breakpoint resume and fragmented caching unit. When the network transmission quality is detected to be lower than a preset threshold, the valid data to be uploaded is cached in the local storage unit. After the network is restored, the cached data is automatically re-uploaded according to the time priority.

7. A lightweight intelligent terminal control system for real-time data processing, used to manage the lightweight intelligent terminal according to any one of claims 1 to 6, characterized in that, This includes a cloud-based management and control platform, a terminal configuration management module, a multi-source data aggregation module, and a global visualization module; The terminal configuration management module is used to send data processing rules, inference model parameters, and computing power scheduling strategies to each connected lightweight smart terminal. The multi-source data aggregation module is used to receive and uniformly store valid feature data and abnormal alarm data uploaded by various lightweight smart terminals. The global visualization module is used to visualize the operating status, data processing results, and abnormal alarm information of the access terminal.

8. A lightweight intelligent terminal control system for real-time data processing according to claim 7, characterized in that, It also includes a global computing power scheduling module, which is used to collect computing power load and network status data of all access terminals in real time, and dynamically adjust the task allocation strategy according to the real-time data processing needs of each region to balance the computing power load of each terminal.

9. A lightweight intelligent terminal control system for real-time data processing according to claim 7, characterized in that, It also includes a model iteration and update module, which iteratively optimizes the lightweight inference model based on the aggregated effective feature data and sends model update packages to the terminal using differential updates. The size of the differential update package is much smaller than that of the full model package.

10. A lightweight intelligent terminal control system for real-time data processing according to claim 7, characterized in that, It also includes a security management module, which performs two-way authentication on all access terminals, encrypts all data transmitted between the terminal and the cloud end-to-end, and automatically isolates abnormal terminals when an abnormal operating state is detected.

Images