An edge-cloud collaborative real-time control system for the Industrial Internet of Things

By using time-frequency analysis and adaptive frequency domain processing in the edge-cloud collaborative real-time control system, the problem of difficult feature extraction of sensor data from edge nodes in time-varying noise environments is solved, and efficient and reliable closed-loop control is achieved.

CN122732147BActive Publication Date: 2026-11-17TIANJIN TOPTECH TECH CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610900261.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-06-22
Publication Date
2026-11-17
Estimated Expiration
2046-06-22

AI Technical Summary

Technical Problem

Existing industrial IoT edge nodes struggle to adaptively extract frequency domain features from sensor data when faced with time-varying noise environments, resulting in limited control response accuracy and resource utilization efficiency.

Method used

Employing a time-frequency analysis module, an envelope confidence module, a collaborative mapping module, a clustering module, and a backtracking calibration module, the system dynamically analyzes the sensor data stream, adaptively calculates the confidence interval, and performs collaborative mapping and clustering of the sparse spectrum in the frequency domain to generate closed-loop real-time control commands.

Benefits of technology

It improves the anti-interference capability and control accuracy of edge nodes in complex industrial sites, reduces computational complexity and resource consumption, and improves the system's adaptability and resource utilization efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122732147B_ABST
    Figure CN122732147B_ABST
Patent Text Reader

Abstract

This invention relates to the field of industrial Internet of Things (IIoT) control technology, specifically disclosing an edge-cloud collaborative real-time control system for IIoT, comprising a time-frequency analysis module, an envelope confidence module, a collaborative mapping module, a clustering module, a backtracking calibration module, and a queue encapsulation module. The system dynamically analyzes a time-series frame set to obtain a sparse frequency spectrum; performs residual error envelope analysis based on edge node states to obtain confidence intervals; obtains weight distributions based on collaborative mapping using a global policy library; clusters the weight distributions to obtain element clusters; performs backtracking calibration based on historical optimized trajectories to obtain correction values; and encapsulates the data in a queue to obtain closed-loop real-time control commands. This invention can improve the control efficiency of IIoT.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of industrial Internet of Things (IoT) control technology, and in particular to an edge-cloud collaborative real-time control system for industrial IoT. Background Technology

[0002] In Industrial Internet of Things (IIoT) scenarios, applications such as smart production lines, unmanned inspections, and remote equipment maintenance place stringent demands on the real-time performance and reliability of control systems. Existing technologies typically employ edge nodes to handle on-site data acquisition and simple control logic execution, while offloading complex computations and long-term data storage to the cloud. For example, in an automated assembly line, an edge gateway collects sensor data at millisecond intervals and executes local closed-loop control, while simultaneously reporting equipment status and process parameters to the cloud for global optimization analysis. However, edge nodes have limited computing and storage resources, making it difficult to support large-scale model inference or high-precision control algorithms; while the cloud possesses powerful computing capabilities, network transmission latency and uncertainty prevent direct control command issuance from meeting real-time requirements. To address this, the industry has proposed an edge-cloud collaborative architecture, aiming to retain real-time control tasks for execution at the edge, while offloading non-real-time analysis and training tasks to the cloud, and achieving task orchestration and model updates through bidirectional data channels.

[0003] Existing industrial IoT edge nodes typically employ fixed-parameter frequency domain analysis methods to process sensor data streams when performing real-time control tasks. However, industrial environments are complex, and the intensity and spectral distribution of noise, such as equipment vibration and electromagnetic interference, exhibit time-varying characteristics. Fixed-threshold frequency domain filtering methods struggle to adaptively adapt to changes in noise levels: setting the threshold too high may filter out weak, effective control signals; setting it too low fails to effectively suppress sudden noise, leading to decreased accuracy in subsequent control decisions. Furthermore, conventional methods lack differentiated treatment for the importance of different frequency components, treating all spectral components equally and failing to focus computational resources on the critical frequency regions most relevant to the control task. These issues limit the control response accuracy and resource utilization efficiency of edge nodes in dynamically changing industrial environments. Therefore, improving the ability of edge nodes to autonomously extract and structure the frequency domain features of sensor data in time-varying noise environments has become an urgent problem to be solved. Summary of the Invention

[0004] In view of the above problems, the purpose of this invention is to provide an edge-cloud collaborative real-time control system for the Industrial Internet of Things (IIoT) to solve the problems mentioned in the background.

[0005] This invention provides an edge-cloud collaborative real-time control system for the Industrial Internet of Things (IIoT), characterized in that the system includes a time-frequency analysis module, an envelope confidence module, a collaborative mapping module, a clustering and partitioning module, a backtracking calibration module, and a queue encapsulation module, wherein: The time-frequency analysis module is used to dynamically analyze the time-series frame set in the industrial Internet of Things sensor data stream to obtain the frequency domain sparse spectrum of the time-series frame set. The time envelope confidence module is used to perform envelope analysis on the residual error formed after threshold truncation of the frequency domain sparse spectrum based on the edge node status of the time frame set, so as to obtain the confidence interval of the frequency domain sparse spectrum. The collaborative mapping module is used to perform collaborative mapping on the confidence interval based on the edge-cloud global policy library of the Industrial Internet of Things to obtain the weight distribution of the frequency domain sparse spectrum. The clustering module is used to cluster the weight distribution to obtain the element clusters of the weight distribution; The backtracking calibration module is used to perform trajectory backtracking adaptive calibration on the element cluster based on the historical optimized trajectory of the edge and cloud of the Industrial Internet of Things, so as to obtain the correction amount of the element cluster. The queue encapsulation module is used to encapsulate the correction amount in a queue to obtain the closed-loop real-time control command of the industrial Internet of Things.

[0006] In a preferred embodiment, the step of dynamically parsing the time-series frame set in the industrial IoT sensor data stream to obtain the frequency domain sparse spectrum of the time-series frame set includes: The industrial IoT sensor data stream is subjected to time-series window sliding segmentation to obtain multiple overlapping frame segments of the industrial IoT sensor data stream. Based on the multiple overlapping frame segments, the temporal energy distribution of the time-series frame set in the industrial IoT sensor data stream is aggregated and mapped to obtain the energy concentration region of the time-series frame set. The frequency domain basis decomposition of the energy concentration region is performed to obtain the frequency domain component characteristics of the energy concentration region; Based on the frequency domain component characteristics, the time-series frame set is subjected to spectral filtering to obtain the frequency domain sparse spectrum of the time-series frame set.

[0007] In a preferred embodiment, the step of performing envelope analysis on the residual error formed after threshold truncation of the frequency domain sparse spectrum based on the edge node states in the time-series frame set to obtain the confidence interval of the frequency domain sparse spectrum includes: Feature separation is performed on the edge nodes in the time-series frame set to obtain the fluctuation amplitude parameter sequence of the edge nodes; Based on the fluctuation amplitude parameter sequence, the threshold truncation threshold of the frequency domain sparse spectrum is adaptively controlled and calculated to obtain the adaptive depth control value of the threshold truncation threshold. Based on the adaptive depth control value, the spectral line components of the frequency domain sparse spectrum are filtered out to obtain the residual error formed after the frequency domain sparse spectrum is truncated by a threshold. The remaining error is fitted with an envelope contour to obtain the upper envelope trajectory and the lower envelope trajectory of the remaining error. Based on the upper envelope trend trajectory and the lower envelope trend trajectory, the error amplitude variation in the frequency domain sparse spectrum is delimited by interval boundary to obtain the confidence interval of the frequency domain sparse spectrum.

[0008] In a preferred embodiment, the adaptive depth control value is calculated using the following formula: ; In the formula, The adaptive depth control value is... The threshold is the cutoff threshold. This is the cyclic index variable for the fluctuation amplitude parameter. For the first The aforementioned fluctuation amplitude parameter The total number of the fluctuation amplitude parameter sequences. It is a very small positive number. A natural constant An exponential function with base 0. It is a logarithmic function.

[0009] In a preferred embodiment, the step of fitting the envelope contour of the residual error based on the adaptive depth control value to obtain the upper envelope trajectory and lower envelope trajectory of the residual error includes: The residual error is subjected to extreme point identification to obtain the original extreme point distribution sequence of the residual error; Based on the adaptive depth control value, the positive and negative half axes of the original extreme point distribution sequence are subjected to extreme value amplitude layer discrimination to obtain the upper envelope anchor point pre-selection set and the lower envelope anchor point pre-selection set of the remaining error. Linear interpolation is performed on adjacent anchor points in the pre-selected set of upper envelope anchor points and the pre-selected set of lower envelope anchor points to obtain the upper envelope trend trajectory and lower envelope trend trajectory of the remaining error.

[0010] In a preferred embodiment, the edge-cloud global policy library based on the Industrial Internet of Things (IIoT) performs collaborative mapping on the confidence interval to obtain the weight distribution of the frequency domain sparse spectrum, including: Based on the working conditions of the policy entries in the global policy library of the Industrial Internet of Things (IIoT) at the edge and cloud, the confidence interval is retrievaled by tags to obtain the tag index sequence of the global policy library of the Industrial Internet of Things at the edge and cloud. Based on the label index sequence, anchor points are located in the confidence interval to obtain the mapped anchor point positions of the confidence interval; Based on the weight configuration rules in the global strategy library of the Industrial Internet of Things in the edge-cloud, weight factors are assigned to the mapping anchor positions to obtain the weight factor sequence of the mapping anchor positions. Based on the weight factor sequence, spectral line weights are configured for the spectral line components in the frequency domain sparse spectrum to obtain the weight distribution of the frequency domain sparse spectrum.

[0011] In a preferred embodiment, the step of clustering the weight distribution to obtain the element clusters of the weight distribution includes: The component spacing of the weight distribution is measured to obtain the set of component distances of the weight distribution; Based on the component distance set, the weight components are clustered to obtain the initial grouping of the weight distribution; The initial grouping is subjected to an inter-group separation threshold determination to obtain the merging decision identifier interval of the initial grouping; Based on the merging decision identifier interval, adjacent groups in the initial grouping are merged and grouped to obtain the element clusters of the weight distribution.

[0012] In a preferred embodiment, the edge-cloud historical optimization trajectory based on the Industrial Internet of Things (IIoT) performs trajectory backtracking adaptive calibration on the element cluster to obtain the correction amount of the element cluster, including: The historical optimization trajectories of the edge and cloud of the industrial Internet of Things are extracted sequentially to obtain the reference trajectory fragment sequence of the element cluster; Based on the reference trajectory segment sequence, the weight components in the element cluster are subjected to structured trajectory deviation backtracking to obtain the trajectory offset vector set of the element cluster; The trajectory offset vector set is subjected to offset quantization to obtain the significant offset components of the trajectory offset vector set; Based on the significant offset components, the weight distribution in the element cluster is offset-compensated and superimposed to obtain the correction amount of the element cluster.

[0013] In a preferred embodiment, the step of performing structured trajectory deviation backtracking on the weight components in the element cluster based on the reference trajectory segment sequence to obtain the trajectory offset vector set of the element cluster includes: Perform time-series matching on the reference trajectory segment sequence to obtain the time correspondence between the reference trajectory segment sequence and the weight components; Based on the time correspondence, the weight components in the element cluster are compared with the segments at the same position in the reference trajectory segment sequence to obtain the original deviation value sequence of the weight components. The original deviation value sequence is filtered to obtain a set of valid deviation values ​​from the original deviation value sequence. Based on the set of effective deviation values, deviation vectors are constructed for the weight components in the element cluster to obtain the set of trajectory offset vectors for the element cluster.

[0014] In a preferred embodiment, the step of queuing and encapsulating the correction amount to obtain the closed-loop real-time control command of the industrial Internet of Things includes: The correction amounts are sorted by time identifier to obtain an ordered list of the correction amounts; Based on the ordered list, each correction value is stored in the first-in-first-out buffer of the industrial Internet of Things, generating a filling sequence for the first-in-first-out buffer; Add batch markers to the correction amounts in the filling sequence to obtain the encapsulated data blocks of the correction amounts; The encapsulated data block of the correction amount is structured and encoded to obtain closed-loop real-time control instructions for industrial IoT edge nodes.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention obtains the sparse spectrum in the frequency domain through a time-frequency analysis module and adaptively calculates the confidence interval based on the edge node state using an envelope confidence module. Compared with fixed threshold truncation, this mechanism can dynamically adjust the filtering depth according to the severity of signal fluctuations, retain key frequency components in strong noise environments, and effectively suppress residual noise in stable environments, thereby improving the anti-interference capability of edge nodes in complex industrial environments.

[0016] 2. This invention uses a collaborative mapping module to configure weights differently for different frequency components based on a global policy library, and a clustering module to group the weight distribution into independent element clusters. This processing flow transforms the raw spectral data into frequency domain feature groups with clear physical meaning, providing structured input for the subsequent control decision module and reducing the computational complexity of the control algorithm. Attached Figure Description

[0017] Other objects and results of the invention will become more apparent and readily understood by referring to the following description taken in conjunction with the accompanying drawings, and with a more complete understanding of the invention. In the drawings: Figure 1 This is a system architecture diagram of an edge-cloud collaborative real-time control system for industrial Internet of Things (IoT) provided in an embodiment of the present invention. Detailed Implementation

[0018] Existing industrial IoT control systems typically employ either a centralized cloud architecture or a purely edge-based local control approach. While centralized cloud architectures enable global data analysis and optimization, they are susceptible to network latency when processing large numbers of real-time control commands. On the other hand, while purely edge-based local control offers fast response times, it struggles to obtain global production status information for collaborative optimization, thus limiting the overall operational efficiency of the system.

[0019] Some existing edge-cloud collaboration solutions predetermine the division of labor between the edge and the cloud during system deployment, failing to dynamically adjust based on equipment status, network load, or sudden task demands in industrial settings. This static division can easily lead to edge node overload or idle cloud resources, reducing the system's adaptability and resource utilization efficiency in complex industrial environments.

[0020] To address the aforementioned problems, this invention provides an edge-cloud collaborative real-time control system for the Industrial Internet of Things (IIoT). Specific embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0021] To illustrate the edge-cloud collaborative real-time control system for the Industrial Internet of Things provided by this invention, Figure 1 An exemplary illustration is provided for an edge-cloud collaborative real-time control system for the Industrial Internet of Things (IIoT) according to an embodiment of the present invention.

[0022] The following description of exemplary embodiments is merely illustrative and is in no way intended to limit the invention or its application or use. Techniques and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques and equipment should be considered part of the specification.

[0023] like Figure 1 The figure shown is a system architecture diagram of an edge-cloud collaborative real-time control system for industrial Internet of Things provided by an embodiment of the present invention.

[0024] The edge-cloud collaborative real-time control system 100 for industrial IoT described in this invention may include a video analysis module 101, an envelope confidence module 102, a collaborative mapping module 103, a clustering and partitioning module 104, a backtracking calibration module 105, and a queue encapsulation module 106, wherein: The time-frequency analysis module 101 is used to dynamically analyze the time-series frame set in the industrial IoT sensor data stream to obtain the frequency domain sparse spectrum of the time-series frame set, including: In this embodiment of the invention, the industrial IoT sensor data stream is subjected to time-series window sliding segmentation to obtain multiple overlapping frame segments of the industrial IoT sensor data stream. Based on the multiple overlapping frame segments, the temporal energy distribution of the time-series frame set in the industrial IoT sensor data stream is aggregated and mapped to obtain the energy concentration region of the time-series frame set. The energy concentration region is subjected to frequency domain basis splitting to obtain the frequency domain component characteristics of the energy concentration region; Based on the frequency domain component characteristics, the time-series frame set is subjected to spectral filtering to obtain the frequency domain sparse spectrum of the time-series frame set.

[0025] When performing a time-series window sliding segmentation operation on an Industrial Internet of Things (IIoT) sensor data stream, the edge node first reads its local configuration memory to obtain a pre-set window length and sliding step value. The window length is determined based on an integer multiple of the sampling period of the industrial field sensor data, and the sliding step value is set to half of the window length. Starting from the first data frame of the sensor data stream, the edge node extracts all consecutive data frames extending from that starting position within the window length range. These consecutive data frames together constitute the first overlapping frame segment. Subsequently, the edge node moves the extraction starting position backward by one sliding step and extracts consecutive data frames again within the forward window length range at the new starting position, obtaining the second overlapping frame segment. Since the sliding step is less than the window length, the first and second overlapping frame segments overlap in time, and the duration of the overlap is equal to the window length minus the sliding step. The edge node repeats the above moving and extraction operations until the current starting position plus the window length exceeds the position of the last frame of the sensor data stream, ultimately obtaining multiple overlapping frame segments of the IIoT sensor data stream.

[0026] After extracting the overlapping frame segments, the edge nodes begin performing aggregation mapping operations on the temporal energy distribution. For each overlapping frame segment, the edge node extracts the signal amplitude of all data frames within it, squares each amplitude, and then sums all the squared results to obtain a single value, which is the temporal energy value of that overlapping frame segment. The edge nodes arrange the calculated temporal energy values ​​sequentially on a unified time axis according to the center time of each overlapping frame segment from earliest to latest, forming a series of discrete numerical points that change over time. These discrete numerical points collectively describe the temporal energy distribution of the time-series frame set. Next, the edge nodes perform a merging judgment on adjacent discrete numerical points. If the difference between two adjacent numerical points is less than a preset merging threshold, the two values ​​are added together and replaced with a new numerical point, and their corresponding time intervals are merged. After multiple merging operations, several consecutive time intervals with values ​​significantly higher than the surrounding areas appear in the temporal energy distribution. Each such interval is identified and marked as an energy concentration region.

[0027] For each energy concentration region, the edge node performs a frequency domain basis splitting operation. The edge node completely copies the continuous data frames contained in the current energy concentration region from the original sensor data stream, forming an independent time-domain signal block. The edge node pre-stores a set of standard frequency waveforms, each corresponding to a fixed frequency value, and the duration of all waveforms is equal to the duration of the current time-domain signal block. These standard frequency waveforms are called the frequency domain basis. The edge node sequentially extracts each standard frequency waveform, multiplies it point-by-point with the amplitude of the data frame at the corresponding moment in the current time-domain signal block, and then sums all the products to obtain a projection coefficient. After traversing all standard frequency waveforms, the edge node obtains a series of projection coefficients, each coefficient arranged from low to high according to the frequency value of its corresponding waveform. This series of projection coefficients is collectively referred to as the frequency domain component characteristics of the energy concentration region. Each projection coefficient in the frequency domain component characteristics represents the contribution of the corresponding frequency component in the original time-domain signal block.

[0028] After obtaining the frequency domain component features, the edge nodes begin spectrum filtering to generate a sparse frequency domain spectrum. The edge nodes read a pre-set retention threshold from their local memory. The setting process for this retention threshold is as follows: During system deployment, the edge nodes continuously collect one hour of field sensor data. At this time, the industrial equipment is in standby mode, and the collected data only contains environmental noise and inherent sensor noise. The edge nodes calculate the average energy value of this data segment as the background noise level, and then multiply the background noise level by three times an empirical coefficient to obtain the retention threshold. The edge nodes iterate through each projection coefficient in the frequency domain component features, comparing the absolute value of the projection coefficient with the retention threshold. If the absolute value of the projection coefficient is greater than the retention threshold, the edge nodes retain the projection coefficient and its corresponding frequency value unchanged. If the absolute value of the projection coefficient is less than or equal to the retention threshold, the edge nodes directly assign the projection coefficient a value of zero. After all projection coefficients have been compared and assigned values, the projection coefficients that have not been set to zero and their corresponding frequency values ​​constitute a new spectrum. Since most of the projection coefficients are zero and only a few are non-zero, the edge node defines this new spectrum as the frequency domain sparse spectrum of the time frame set.

[0029] The beneficial effects are as follows: Multiple overlapping frame segments are obtained through time-series window sliding segmentation. The overlapping areas between adjacent segments completely preserve the continuous transition information of industrial equipment state changes, avoiding the loss of key dynamic features caused by simple segmentation. The aggregation mapping operation of time-domain energy distribution can accurately locate the period of strongest signal activity in the sensor data stream, i.e., the energy concentration area, allowing edge nodes to perform in-depth processing only on a small number of key periods, thus significantly reducing computational resource consumption. The frequency domain basis splitting operation converts the time-domain signal block of each energy concentration area into clear frequency-domain component features. These features can effectively distinguish the vibration frequency, current harmonics, or temperature fluctuation frequency generated by different industrial equipment, providing a reliable frequency-domain basis for subsequent refined control. The spectrum filtering operation uses a retention threshold value obtained in advance through on-site background noise measurement to eliminate frequency components corresponding to environmental noise and sensor inherent noise, generating a highly simplified frequency-domain sparse spectrum. This frequency-domain sparse spectrum retains all the main frequency information required for closed-loop control decisions while greatly compressing the data volume, thereby reducing the data transmission bandwidth occupied between edge nodes and the cloud and shortening network transmission latency. All the above operations are executed within the edge nodes in the form of hardware logic or embedded firmware, avoiding black-box processing that relies on external software and ensuring the traceability and determinism of each processing step. Overall, this specific implementation process provides the edge-cloud collaborative real-time control system with efficient and reliable data preprocessing capabilities, laying a clean and compressed data foundation for subsequent envelope analysis, collaborative mapping, clustering, backtracking calibration, and queue encapsulation, ultimately improving the generation speed and execution reliability of closed-loop real-time control commands.

[0030] The temporal envelope confidence module 102 is used to perform envelope analysis on the residual error formed after threshold truncation of the frequency domain sparse spectrum based on the edge node states in the temporal frame set, to obtain the confidence interval of the frequency domain sparse spectrum, including: In this embodiment of the invention, feature separation is performed on the edge nodes in the time-series frame set to obtain the fluctuation amplitude parameter sequence of the edge nodes; Based on the fluctuation amplitude parameter sequence, the threshold truncation threshold of the frequency domain sparse spectrum is adaptively controlled and calculated to obtain the adaptive depth control value of the threshold truncation threshold. Based on the adaptive depth control value, spectral line components are filtered out from the frequency domain sparse spectrum to obtain the residual error formed after threshold truncation of the frequency domain sparse spectrum. The remaining error is fitted with an envelope contour to obtain the upper envelope trajectory and the lower envelope trajectory of the remaining error. Based on the upper envelope trend trajectory and the lower envelope trend trajectory, the error amplitude variation in the frequency domain sparse spectrum is delimited by interval boundary to obtain the confidence interval of the frequency domain sparse spectrum.

[0031] The formula for calculating the adaptive depth control value is as follows: ; In the formula, The adaptive depth control value is... The threshold is the cutoff threshold. This is the cyclic index variable for the fluctuation amplitude parameter. For the first The aforementioned fluctuation amplitude parameter The total number of the fluctuation amplitude parameter sequences. It is a very small positive number. A natural constant An exponential function with base 0. It is a logarithmic function.

[0032] The step of fitting the envelope contour of the remaining error based on the adaptive depth control value to obtain the upper envelope trend trajectory and the lower envelope trend trajectory of the remaining error includes: The residual error is subjected to extreme point identification to obtain the original extreme point distribution sequence of the residual error; Based on the adaptive depth control value, the positive and negative half axes of the original extreme point distribution sequence are subjected to extreme value amplitude layer discrimination to obtain the upper envelope anchor point pre-selection set and the lower envelope anchor point pre-selection set of the remaining error. Linear interpolation is performed on adjacent anchor points in the pre-selected set of upper envelope anchor points and the pre-selected set of lower envelope anchor points to obtain the upper envelope trend trajectory and lower envelope trend trajectory of the remaining error.

[0033] Edge nodes read a pre-stored set of time-series frames from their local memory. This set contains the state records of all edge nodes in the current Industrial Internet of Things (IIoT) sensor data stream. For each edge node in the time-series frame set, the edge node performs feature separation. Feature separation extracts the numerical sequence of the signal amplitude changing over time from the state data of each edge node. Specifically, the edge node iterates through the state data of each edge node in chronological order, records the signal amplitude at each sampling moment, removes irrelevant information such as node identifiers and timestamps, and retains only the amplitude value itself. The amplitude values ​​of all edge nodes are then combined and arranged according to their respective nodes and sampling moments to form a one-dimensional numerical sequence. This sequence is defined as the fluctuation amplitude parameter sequence of the edge nodes. Each value in the fluctuation amplitude parameter sequence reflects the signal fluctuation intensity of the corresponding edge node at the corresponding moment.

[0034] After obtaining the fluctuation amplitude parameter sequence, when the edge node calculates the adaptive depth control value based on the fluctuation amplitude parameter sequence, it first reads the pre-set threshold cutoff value from the local memory. This threshold cutoff value is a fixed reference value, which is set by multiplying the average amplitude by an empirical coefficient after collecting a segment of field sensor data without external excitation during the system initialization phase. The edge node traverses all values ​​in the fluctuation amplitude parameter sequence except the last value, and performs a direction determination operation for each pair of adjacent values: it calculates the difference between the second and first values, adds the absolute value of the difference to a very small positive number, uses this as the denominator, and then divides the difference by the denominator to obtain a direction ratio. The direction ratio can only be positive or negative one, representing an increase or decrease in signal amplitude, respectively. The edge node also calculates the result of dividing the absolute value of this difference by the threshold cutoff value, adds one to this result, and takes the logarithm to the base of the natural constant to obtain a logarithmic result. Multiplying the aforementioned directional ratio by the logarithmic result yields the contribution term for that pair of adjacent values. The edge node sums up the contributions of all adjacent value pairs to obtain a total sum. The edge node then calculates the absolute value of each value in the fluctuation amplitude parameter sequence divided by the threshold cutoff, sums all the results, and adds them to the same minimum positive number to obtain a sum in the denominator. The edge node divides the aforementioned total sum by the sum in the denominator to obtain an exponential value. The edge node calculates an exponential function with the natural constant as the base and this exponential value as the exponent, and finally multiplies the threshold cutoff by the result of this exponential function to obtain the adaptive depth control value. The magnitude of the adaptive depth control value dynamically changes according to the intensity of numerical changes in the fluctuation amplitude parameter sequence: when adjacent values ​​frequently alternate between positive and negative or the absolute value of the difference is large, the exponent value is positive, the adaptive depth control value is greater than the threshold cutoff, and the filtering intensity is enhanced; when the fluctuation amplitude parameter sequence is generally stable and the absolute value is small, the exponent value is negative, the adaptive depth control value is less than the threshold cutoff, and the filtering intensity is weakened. Extremely small positive values ​​are fixed at the factory for edge nodes; their purpose is to prevent the denominator from being zero during direction determination operations and denominator summation calculations.

[0035] After obtaining the adaptive depth control value, the edge nodes perform spectral component filtering on the frequency-domain sparse spectrum. The frequency-domain sparse spectrum consists of multiple spectral components, each containing a frequency value and its corresponding amplitude. The edge nodes extract each spectral component one by one and compare the absolute value of its amplitude with the adaptive depth control value. If the absolute value of the amplitude of the spectral component is greater than the adaptive depth control value, the edge nodes retain the spectral component unchanged. If the absolute value of the amplitude of the spectral component is less than or equal to the adaptive depth control value, the edge nodes set the amplitude of the spectral component to zero, i.e., remove the frequency component from the frequency-domain sparse spectrum. After all spectral components have been compared and zeroed, the remaining non-zero spectral components in the frequency-domain sparse spectrum correspond to the main components of the signal, while the zeroed spectral components are considered as truncation errors. The edge nodes extract the original amplitudes of all zeroed spectral components in frequency order to form an independent numerical sequence, which is called the residual error formed after threshold truncation of the frequency-domain sparse spectrum. The residual error reflects the distribution of the weak energy frequency components that were filtered out.

[0036] Edge nodes perform envelope contour fitting on the residual error. First, the edge node locates all extreme points in the residual error numerical sequence. An extreme point is a position where the value is greater than or less than the values ​​of its immediate neighbors. All extreme points are arranged in order of their x-coordinates (frequency), resulting in the original extreme point distribution sequence. The edge node divides the extreme points in the original distribution sequence into two groups based on their numerical values: positive extreme points are assigned to the positive half-axis set, and negative extreme points are assigned to the negative half-axis set. The edge node then selects the extreme point with the largest local amplitude from the positive half-axis set. Specifically, it iterates through the positive half-axis set, checking for other positive extreme points within a predetermined range before and after each extreme point. If the amplitude of this extreme point is greater than all other positive extreme points in its neighborhood, it is selected into the upper envelope anchor point pre-selection set. Similarly, the edge node selects the extreme point with the smallest local amplitude (i.e., the point with the largest absolute negative value) from the negative half-axis set and selects it into the lower envelope anchor point pre-selection set. Edge nodes connect all anchor points in the pre-selected set of upper envelope anchor points in frequency order. During connection, linear interpolation is used to fill the values ​​of intermediate points between adjacent anchor points, resulting in a continuous trajectory curve, called the upper envelope trend trajectory of the residual error. Edge nodes perform the same connection operation on the pre-selected set of lower envelope anchor points to obtain the lower envelope trend trajectory of the residual error. The upper and lower envelope trend trajectories together delineate the upper and lower boundaries of the fluctuation of the residual error value.

[0037] Edge nodes define interval boundaries for error amplitude variations in the frequency-domain sparse spectrum based on the upper and lower envelope trend trajectories. Specifically, the edge node compares the original amplitude of each spectral line component in the frequency-domain sparse spectrum with the values ​​of the upper and lower envelope trend trajectories at the same frequency position. For each frequency point, the edge node records the interval between the values ​​of the upper and lower envelope trend trajectories where the original amplitude falls. These intervals for all frequency points are collected to form a set of upper and lower boundaries covering the entire frequency domain; this set is defined as the confidence interval of the frequency-domain sparse spectrum. The upper limit of each frequency point within the confidence interval is the value of the upper envelope trend trajectory at that point, and the lower limit is the value of the lower envelope trend trajectory at that point. The confidence interval characterizes the reasonable fluctuation range of the effective frequency components in the frequency-domain sparse spectrum; amplitudes outside the confidence interval are considered abnormal errors.

[0038] After reading the residual error sequence, the edge node scans each numerical point in the sequence one by one, comparing each numerical point with its two adjacent numerical points. If the current numerical point is greater than both its left and right adjacent numerical points, or less than both its left and right adjacent numerical points, the edge node marks this numerical point as an extreme point. The edge node arranges all the marked extreme points in ascending order of frequency, forming an ordered list containing the location and magnitude of the extreme points. This list is called the original extreme point distribution sequence of the residual error.

[0039] The edge node reads the previously calculated adaptive depth control value from local memory and uses this value as the benchmark for amplitude discrimination. The edge node divides all extreme points in the original extreme point distribution sequence into two independent subsets based on the sign of their amplitudes: positive amplitude extreme points are assigned to the positive half-axis set, and negative amplitude extreme points are assigned to the negative half-axis set. For the positive half-axis set, the edge node iterates through each extreme point, checking if its amplitude is greater than the amplitudes of its preset number of adjacent positive extreme points. The preset number is determined by the value of the adaptive depth control value; the larger the value, the wider the neighborhood range checked. If the amplitude of an extreme point is the maximum value within its neighborhood, the edge node selects it for a temporary storage area, named the upper envelope anchor point pre-selection set. For the negative half-axis set, the edge node iterates through each extreme point, checking if its amplitude is less than the amplitudes of its preset number of adjacent negative extreme points, i.e., whether the absolute value of the negative value is the maximum within its neighborhood. If the magnitude of an extreme point is the minimum within its neighborhood, the edge node selects it for another temporary storage area, which is named the lower envelope anchor point pre-selection set. The anchor points in the upper envelope anchor point pre-selection set are all critical inflection points on the upper edge of the residual error, and the anchor points in the lower envelope anchor point pre-selection set are all critical inflection points on the lower edge of the residual error.

[0040] The edge node extracts all anchor points from the pre-selected set of upper envelope anchor points and sorts them in ascending order of their corresponding frequency values. Starting with the smallest frequency anchor point, the edge node connects adjacent anchor points sequentially. During the connection, several intermediate points are inserted at equal intervals within the frequency range of the two anchor points. The amplitude of each intermediate point is calculated by linearly proportionally from the amplitudes of the two endpoint anchor points: the amplitude of the intermediate point equals the amplitude of the left anchor point plus the difference between the amplitudes of the right and left anchor points multiplied by the ratio of the difference between the frequency of the intermediate point and the frequency of the left anchor point divided by the difference between the frequencies of the two anchor points. After connecting all the original anchor points and the inserted intermediate points in frequency order, a continuous and smooth curve is formed. This curve is called the upper envelope trend trajectory of the residual error. The edge node performs the same linear interpolation connection operation on the pre-selected set of lower envelope anchor points, resulting in another continuous and smooth curve, which is called the lower envelope trend trajectory of the residual error. The upper envelope trend trajectory lies above the residual error sequence, and the lower envelope trend trajectory lies below the residual error sequence. Together, the two trajectories outline the outer envelope band of the residual error value fluctuation.

[0041] The beneficial effects are as follows: By performing feature separation operations on the edge nodes of the time-series frame set, a pure fluctuation amplitude parameter sequence is obtained. This sequence completely preserves the dynamic change information of the signal amplitude of each edge node and eliminates redundant data unrelated to the amplitude, providing a reliable input basis for subsequent adaptive threshold adjustment. Adaptive control calculation of the threshold truncation based on the fluctuation amplitude parameter sequence is performed. The truncation depth is dynamically adjusted by utilizing the difference direction and amplitude between adjacent values ​​in the fluctuation amplitude parameter. This allows the adaptive depth control value to change in real time with the intensity of signal fluctuations in the industrial field, avoiding the defects of fixed thresholds that excessively retain noise when the signal is stable or erroneously delete effective components when the signal is intense. Using the adaptive depth control value to filter spectral line components in the sparse spectrum of the frequency domain can accurately distinguish between strong signal components and weak noise components. The weak components are stored as residual errors, which both compresses the amount of main signal data and completely preserves error information for subsequent analysis. The remaining error is fitted with an envelope profile to obtain the upper and lower envelope trajectories. The upper and lower boundaries of error fluctuations are constructed through extreme point selection and linear interpolation. This method does not rely on any external model and extracts the envelope entirely based on the numerical distribution characteristics of the remaining error itself. Based on the upper and lower envelope trajectories, the error amplitude variation in the sparse frequency spectrum is bounded, and the resulting confidence interval provides reasonable upper and lower limits for the amplitude at each frequency point. This confidence interval can serve as input to the subsequent collaborative mapping module, guiding the generation of the weight distribution. The entire implementation process is executed sequentially in the hardware logic layer or embedded firmware of the edge node. The inputs and outputs of each step are clearly defined, avoiding the uncertainties caused by black-box processing. Ultimately, in industrial IoT scenarios, this system can adaptively adjust the frequency domain filtering intensity, effectively suppressing environmental noise interference while retaining key control frequency components, thus improving the anti-interference capability and control accuracy of the edge-cloud collaborative real-time control system in complex industrial environments.

[0042] The original extreme point distribution sequence of the residual error is obtained by identifying extreme points, completely capturing the position and amplitude of all local peaks and valleys in the residual error, providing a precise set of feature points for subsequent envelope construction. Based on adaptive depth control values, the positive and negative semi-axis of the original extreme point distribution sequence are used for hierarchical discrimination of extreme value amplitudes. The neighborhood comparison range is dynamically adjusted using adaptive depth control values, allowing for the selection of sparser anchor points to avoid an overly dense envelope when signal fluctuations are severe, and denser anchor points to ensure envelope accuracy when the signal is stable, thus achieving adaptive control of anchor point selection. Linear interpolation is performed on adjacent anchor points in the pre-selected sets of upper and lower envelope anchor points to generate continuous upper and lower envelope trend trajectories without introducing complex curve fitting, ensuring the smoothness of the envelope curve and computational efficiency. The generation of the two envelope trend trajectories depends entirely on the extreme value distribution of the residual error itself and the adaptive depth control values, requiring no external reference model, enabling the system to independently and reliably complete envelope extraction at industrial IoT edge nodes. The above operations provide clear upper and lower boundary criteria for the subsequent delineation of confidence intervals, thereby improving the accuracy of error amplitude assessment in the sparse spectrum of the frequency domain and enhancing the robustness of the edge-cloud collaborative real-time control system in noisy environments.

[0043] The collaborative mapping module 103 is used to perform collaborative mapping on the confidence interval based on the edge-cloud global policy library of the Industrial Internet of Things to obtain the weight distribution of the frequency domain sparse spectrum, including: In this embodiment of the invention, based on the working conditions of the policy entries in the global policy library of the industrial Internet of Things in the edge and cloud, the confidence interval is retrievaled with tags to obtain the tag index sequence of the global policy library of the industrial Internet of Things in the edge and cloud. Based on the label index sequence, anchor points are located in the confidence interval to obtain the mapped anchor point positions of the confidence interval; Based on the weight configuration rules in the global strategy library of the Industrial Internet of Things in the edge-cloud, weight factors are assigned to the mapping anchor positions to obtain the weight factor sequence of the mapping anchor positions. Based on the weight factor sequence, spectral line weights are configured for the spectral line components in the frequency domain sparse spectrum to obtain the weight distribution of the frequency domain sparse spectrum.

[0044] Edge nodes read from a pre-stored global policy library for the Industrial Internet of Things (IIoT) in their local memory. This library contains several policy entries, each corresponding to a type of industrial field condition and carrying a unique tag. Simultaneously, the edge node obtains a confidence interval calculated by the envelope confidence module, which includes the upper and lower limits of the amplitude at each frequency point in the sparse frequency spectrum. After determining the current policy entry condition based on the real-time operating status of the IIoT, the edge node searches the global policy library for policy entries matching this condition. The search method involves comparing the amplitude range of each frequency point in the confidence interval with the predefined condition characteristic interval in the policy entry. When the overall distribution of the confidence interval matches the condition characteristics of a policy entry, the edge node records the tag of that policy entry. All successfully matched policy entry tags are arranged in order of matching, forming an ordered tag list, which is defined as the tag index sequence of the global policy library for the IIoT.

[0045] After obtaining the label index sequence, the edge nodes extract each label from the sequence. Each label corresponds to a pre-stored anchor position parameter in the global policy library. This anchor position parameter indicates the specific frequency point that should be focused on on the frequency axis of the confidence interval. Based on the order of the labels in the label index sequence, the edge nodes sequentially retrieve the anchor position parameters pointed to by each label and map these anchor positions onto the frequency axis of the confidence interval. The mapping operation is as follows: for each anchor position parameter, find the frequency point in the frequency coordinates of the confidence interval that is closest to the parameter value, and mark this frequency point as a mapped anchor position. All marked mapped anchor positions are collected in ascending order of frequency, forming a position list, which is called the set of mapped anchor positions for the confidence interval.

[0046] Next, the edge node reads the weight configuration rules from the global policy library of the Industrial Internet of Things (IIoT) in the edge-cloud region. These rules are predefined mapping tables that record the correspondence between each possible mapping anchor location and a weight factor. The edge node iterates through each mapping anchor location in the set of confidence intervals, searching for the corresponding weight factor in the weight configuration rules. After the search is complete, all the found weight factors are arranged in order of frequency of their corresponding mapping anchor locations, forming a numerical sequence. This sequence is called the weight factor sequence for mapping anchor locations, where each weight factor is a value between zero and one.

[0047] Edge nodes acquire the frequency-domain sparse spectrum generated by the time-frequency analysis module. This spectrum contains multiple spectral line components, each corresponding to a frequency value and its amplitude. Each weight factor in the weight factor sequence of the mapping anchor positions is applied to the corresponding frequency spectral line component in the frequency-domain sparse spectrum by multiplying the original amplitude of the spectral line component by the weight factor. For frequency points in the frequency-domain sparse spectrum that do not appear in the set of mapping anchor positions, the edge node sets their weight factor to a preset default value, which is set during system initialization based on the background noise level of the industrial site. After traversing all spectral line components in the frequency-domain sparse spectrum and performing the above multiplication operation on each spectral line component, a new set of amplitudes is obtained. After weight configuration, all spectral line components form a new amplitude distribution, which is called the weight distribution of the frequency-domain sparse spectrum.

[0048] The beneficial effects are as follows: Based on the industrial IoT's global policy library at the edge and cloud, label retrieval of confidence intervals for operating conditions generates a label index sequence that accurately associates the real-time operating conditions of the current industrial site with predefined control modes in the policy library, avoiding operating condition mismatch problems caused by static configuration. The mapped anchor point positions obtained after anchoring the confidence intervals based on the label index sequence directly point to the key frequency points in the sparse frequency spectrum most relevant to the current operating condition, providing accurate spatial positioning basis for subsequent weight allocation. Weight factors are assigned to the mapped anchor point positions using the weight configuration rules in the global policy library at the edge and cloud, generating a weight factor sequence. The weight factors in this sequence are entirely determined by expert experience in the policy library, eliminating the need for online calculations and reducing the processing overhead of edge nodes. Based on the weight factor sequence, spectral line weights are configured for the spectral line components in the sparse frequency spectrum. The final weight distribution highlights frequency components closely related to the current control task and suppresses irrelevant or interfering frequency components. The entire collaborative mapping process is completed within the edge nodes using table lookups and multiplication operations, without relying on external models or online learning, ensuring deterministic response speed and simplicity of implementation. This weight distribution, as input to the clustering module, significantly improves the effectiveness of subsequent clustering and the relevance of control command generation. This enables the edge-cloud collaborative real-time control system to adaptively adjust the control focus according to different operating conditions, enhancing the system's adaptability in variable industrial environments.

[0049] The clustering module 104 is used to cluster the weight distribution to obtain the element clusters of the weight distribution, including: In this embodiment of the invention, the component spacing of the weight distribution is measured to obtain the component distance set of the weight distribution; Based on the component distance set, the weight components are clustered to obtain the initial grouping of the weight distribution; The initial grouping is subjected to an inter-group separation threshold determination to obtain the merging decision identifier interval of the initial grouping; Based on the merging decision identifier interval, adjacent groups in the initial grouping are merged and grouped to obtain the element clusters of the weight distribution.

[0050] Edge nodes acquire the weight distribution output by the collaborative mapping module. This distribution contains a series of weight components arranged in frequency order, each corresponding to a frequency point and having a numerical value. When measuring the component spacing of the weight distribution, the edge node sequentially traverses every two adjacent weight components, calculating their spacing on the frequency axis. The spacing is equal to the frequency value of the subsequent weight component minus the frequency value of the preceding weight component. All spacing distances between adjacent components are collected in frequency order to form a numerical set, which is called the component distance set of the weight distribution.

[0051] After obtaining the component distance set, the edge nodes perform clustering operations on the weight components. Starting from the first weight component of the weight distribution, the edge nodes sequentially determine whether the distance between the current component and the next component is less than a preset clustering threshold. This clustering threshold is set as follows: during system initialization, the edge nodes statistically analyze the frequency distribution density of typical control signals in the Industrial Internet of Things (IIoT), and multiply the reciprocal of this density by an empirical coefficient to obtain the clustering threshold. If the current distance is less than the clustering threshold, the current component and the next component are grouped into the same group; if the current distance is greater than or equal to the clustering threshold, a group boundary is set between the current component and the next component. The edge nodes repeat the above checks until all weight components have been traversed, forming an initial group from all weight components grouped into the same group. This results in several initial groups, the set of which is called the initial grouping of the weight distribution.

[0052] The edge node performs an inter-group separation threshold determination for each initial group. Inter-group separation refers to the degree of separation between two adjacent initial groups. Specifically, it is calculated by taking the largest weight component value in the left initial group and the smallest weight component value in the right initial group, calculating the difference between these two values, and then dividing by the center distance between the two initial groups on the frequency axis. The edge node reads a preset separation threshold from its local memory. This threshold is set during system deployment based on the allowable error range of the industrial IoT control system, corresponding to the specified error range. The inter-group separation of each adjacent group pair is compared with this threshold to determine whether each pair needs to be merged. All results are arranged in group order to form a list of interval identifiers, called the initial group merging decision identifier interval.

[0053] Based on the merge decision identifier interval, edge nodes perform a merge and agglomeration operation on adjacent groups in the initial grouping. Each identifier in the merge decision identifier interval is traversed. If the current identifier indicates that merging is required, the two adjacent initial groups corresponding to that identifier are merged into a new group. The merging method involves placing all weight components within the two groups into the same container and rearranging them according to frequency order. If the current identifier indicates that merging is not required, the two groups remain unchanged. After one traversal, for the newly generated group, the edge node recalculates the merge decision identifier between the new group and its adjacent groups and performs the merge operation again, until no merging occurs in two consecutive traversals. All the resulting groups are called element clusters of the weight distribution. Each element cluster corresponds to a frequency interval that is continuous in the frequency domain and has tightly clustered internal weight components. The number and boundary positions of the element clusters directly reflect the distribution of the main control feature components in the current industrial IoT sensor data.

[0054] The beneficial effects are as follows: by measuring the component spacing of the weight distribution, a component distance set is obtained. This set accurately records the interval information of adjacent weight components on the frequency axis, providing a direct numerical basis for subsequent clustering. Based on the component distance set, the weight components are clustered. Using a preset clustering threshold, closely adjacent weight components in frequency are grouped into the same initial group, achieving preliminary identification of clustering features in the frequency domain. After determining the inter-group separation threshold for the initial group, a merging decision identifier interval is obtained. This interval objectively determines whether adjacent groups belong to the same control feature by comparing the inter-group separation with the preset threshold, avoiding errors caused by subjectively setting group boundaries. Adjacent groups in the initial group are merged and grouped according to the merging decision identifier interval. Through multiple iterations until stable, the final weight distribution element clusters accurately reflect the frequency range occupied by each independent control feature in the industrial IoT sensor data. The weight components within each element cluster are highly correlated, and the inter-cluster separation is clear, enabling the subsequent backtracking calibration module to independently perform trajectory backtracking and correction calculations for each element cluster. This clustering process does not rely on any external models or learning algorithms; it is entirely based on the numerical characteristics of the weight distribution itself and preset industrial field experience thresholds, ensuring determinism and low latency in the processing. The resulting element clusters significantly improve the edge-cloud collaborative real-time control system's ability to separate multi-source control signals, laying a structured data foundation for the generation of accurate closed-loop control commands.

[0055] The backtracking calibration module 105 is used to perform trajectory backtracking adaptive calibration on the element cluster based on the edge-cloud historical optimization trajectory of the Industrial Internet of Things, to obtain the correction amount of the element cluster, including: In this embodiment of the invention, the historical optimization trajectory of the edge-cloud of the industrial Internet of Things is extracted to obtain the reference trajectory fragment sequence of the element cluster; Based on the reference trajectory segment sequence, the weight components in the element cluster are subjected to structured trajectory deviation backtracking to obtain the trajectory offset vector set of the element cluster; The trajectory offset vector set is subjected to offset quantization to obtain the significant offset components of the trajectory offset vector set; Based on the significant offset components, the weight distribution in the element cluster is offset-compensated and superimposed to obtain the correction amount of the element cluster.

[0056] Based on the reference trajectory segment sequence, the structured trajectory deviation backtracking of the weight components in the element cluster is performed to obtain the trajectory offset vector set of the element cluster, including: Perform time-series matching on the reference trajectory segment sequence to obtain the time correspondence between the reference trajectory segment sequence and the weight components; Based on the time correspondence, the weight components in the element cluster are compared with the segments at the same position in the reference trajectory segment sequence to obtain the original deviation value sequence of the weight components. The original deviation value sequence is filtered to obtain a set of valid deviation values ​​from the original deviation value sequence. Based on the set of effective deviation values, deviation vectors are constructed for the weight components in the element cluster to obtain the set of trajectory offset vectors for the element cluster.

[0057] Edge nodes retrieve historical optimization trajectories from cloud storage for the Industrial Internet of Things (IIoT). These historical optimization trajectories record the weight distribution values ​​and corresponding temporal positions used in each optimization adjustment during past control cycles. For each cluster output by the clustering module, the edge node extracts a trajectory segment corresponding to the frequency domain range of the cluster from the historical optimization trajectory. The extraction method is as follows: traverse each recorded point in the historical optimization trajectory; if the frequency coordinate of the recorded point falls within the boundary of the cluster, its value is extracted. All extracted recorded points are arranged in chronological order to form a sequence of segments corresponding to the current cluster; this sequence is called the reference trajectory segment sequence for the cluster.

[0058] After obtaining the reference trajectory segment sequence, the edge nodes begin performing structured trajectory deviation backtracking on the weight components in the element cluster. Structured trajectory deviation backtracking refers to comparing each current weight component in the element cluster with a reference value at the same temporal position in the reference trajectory segment sequence. Specifically, the edge node extracts a weight component from the element cluster in frequency order and simultaneously extracts a reference value from the reference trajectory segment sequence that has the same frequency and corresponding temporal position. The difference between the weight component value and the reference value is calculated; this difference is called the original deviation. After performing the above comparison on all weight components, a set of original deviation values ​​is obtained. The edge node then performs directional merging on this set of original deviation values, combining original deviation values ​​with the same sign and adjacent frequencies into a deviation vector. Each deviation vector contains three attributes: start frequency, end frequency, and deviation intensity. All deviation vectors are collected to form a set, which is called the trajectory offset vector set of the element cluster.

[0059] Next, the edge nodes perform offset quantization on the trajectory offset vector set to extract significant offset components of interest. The specific method for offset quantization is as follows: the edge node iterates through each deviation vector in the trajectory offset vector set, calculates the absolute value of the deviation intensity of that vector, and compares this absolute value with a preset significant offset threshold. This significant offset threshold is set as follows: during system initialization, a set of deviation vector samples is collected under normal operating conditions of the industrial IoT, the standard deviation of the deviation intensity of these samples is calculated, and three times this standard deviation is used as the significant offset threshold. If the absolute value of the deviation intensity of a deviation vector is greater than or equal to the significant offset threshold, the deviation vector is marked as a significant offset component; otherwise, it is ignored. All marked deviation vectors are extracted to form a new set, which is called the significant offset component of the trajectory offset vector set.

[0060] Edge nodes perform offset compensation overlay operations on the weight distribution within element clusters based on significant offset components to obtain the correction amount. Offset compensation overlay refers to applying each deviation vector in the significant offset component to the original weight distribution of the element cluster in the opposite direction of its deviation direction. Specifically, for each deviation vector in the significant offset component, the corresponding frequency interval in the element cluster is found, and the original value of each weight component within that interval is subtracted from the deviation intensity of the deviation vector. For frequency intervals not covered by any significant offset component, the original weight component values ​​remain unchanged. After completing the above operation for all frequency intervals, the weight component values ​​in the element cluster are updated to a new set of values, which is called the correction amount of the element cluster. The correction amount reflects the reverse calibration result of significant deviations in the historical optimization trajectory, making the current control weight distribution approach the ideal distribution implicit in the historical optimization trajectory.

[0061] Edge nodes retrieve historical optimization trajectories from cloud storage for the Industrial Internet of Things (IIoT). These historical optimization trajectories record the weight distribution values ​​and corresponding temporal positions used in each optimization adjustment during past control cycles. For each cluster output by the clustering module, the edge node extracts a trajectory segment corresponding to the frequency domain range of the cluster from the historical optimization trajectory. The extraction method is as follows: traverse each recorded point in the historical optimization trajectory; if the frequency coordinate of the recorded point falls within the boundary of the cluster, its value is extracted. All extracted recorded points are arranged in chronological order to form a sequence of segments corresponding to the current cluster; this sequence is called the reference trajectory segment sequence for the cluster.

[0062] After obtaining the reference trajectory segment sequence, the edge nodes begin performing structured trajectory deviation backtracking on the weight components in the element cluster. Structured trajectory deviation backtracking refers to comparing each current weight component in the element cluster with a reference value at the same temporal position in the reference trajectory segment sequence. Specifically, the edge node extracts a weight component from the element cluster in frequency order and simultaneously extracts a reference value from the reference trajectory segment sequence that has the same frequency and corresponding temporal position. The difference between the weight component value and the reference value is calculated; this difference is called the original deviation. After performing the above comparison on all weight components, a set of original deviation values ​​is obtained. The edge node then performs directional merging on this set of original deviation values, combining original deviation values ​​with the same sign and adjacent frequencies into a deviation vector. Each deviation vector contains three attributes: start frequency, end frequency, and deviation intensity. All deviation vectors are collected to form a set, which is called the trajectory offset vector set of the element cluster.

[0063] Next, the edge nodes perform offset quantization on the trajectory offset vector set to extract significant offset components of interest. The specific method for offset quantization is as follows: the edge node iterates through each deviation vector in the trajectory offset vector set, calculates the absolute value of the deviation intensity of that vector, and compares this absolute value with a preset significant offset threshold. This significant offset threshold is set as follows: during system initialization, a set of deviation vector samples is collected under normal operating conditions of the industrial IoT, the standard deviation of the deviation intensity of these samples is calculated, and three times this standard deviation is used as the significant offset threshold. If the absolute value of the deviation intensity of a deviation vector is greater than or equal to the significant offset threshold, the deviation vector is marked as a significant offset component; otherwise, it is ignored. All marked deviation vectors are extracted to form a new set, which is called the significant offset component of the trajectory offset vector set.

[0064] Edge nodes perform offset compensation overlay operations on the weight distribution within element clusters based on significant offset components to obtain the correction amount. Offset compensation overlay refers to applying each deviation vector in the significant offset component to the original weight distribution of the element cluster in the opposite direction of its deviation direction. Specifically, for each deviation vector in the significant offset component, the corresponding frequency interval in the element cluster is found, and the original value of each weight component within that interval is subtracted from the deviation intensity of the deviation vector. For frequency intervals not covered by any significant offset component, the original weight component values ​​remain unchanged. After completing the above operation for all frequency intervals, the weight component values ​​in the element cluster are updated to a new set of values, which is called the correction amount of the element cluster. The correction amount reflects the reverse calibration result of significant deviations in the historical optimization trajectory, making the current control weight distribution approach the ideal distribution implicit in the historical optimization trajectory.

[0065] The beneficial effects are as follows: By extracting the sequence of historical optimized trajectories from the edge and cloud of the Industrial Internet of Things (IIoT), a reference trajectory segment sequence for element clusters is obtained. This sequence provides a historical reference benchmark for each element cluster, corresponding one-to-one with the frequency position of the current weight component, thus providing a clear target reference for subsequent deviation calculations. Based on the reference trajectory segment sequence, structured trajectory deviation backtracking is performed on the weight components in the element clusters, generating a set of trajectory offset vectors for each element cluster. This set describes the direction and degree of systematic deviation between the current control state and the historical optimal trajectory in the form of deviation vectors. The significant offset components obtained after offset quantization of the trajectory offset vector set are effectively filtered out by a preset significant offset threshold, eliminating unimportant deviations caused by random noise and minor fluctuations, retaining only statistically significant deviation vectors, and avoiding over-correction. Based on the significant offset components, offset compensation is superimposed on the weight distribution in the element clusters. The resulting element cluster correction eliminates significant deviations that have appeared in the historical trajectory while maintaining the local continuity and smoothness of the weight distribution. The entire backtracking calibration process does not rely on online learning or external models; it is completed entirely based on historical optimized trajectories stored in the cloud and preset empirical thresholds, ensuring the repeatability of the operation and low computational latency. The final generated correction value provides the queue encapsulation module with precise control adjustment values, enabling the edge-cloud collaborative real-time control system to continuously optimize the current control effect by utilizing historical experience, thereby improving the convergence speed and steady-state accuracy of the system in repetitive industrial tasks.

[0066] By extracting the sequence of historical optimized trajectories from the edge and cloud of the Industrial Internet of Things (IIoT), a reference trajectory segment sequence for element clusters is obtained. This sequence provides a historical reference benchmark for each element cluster, corresponding one-to-one with the frequency position of the current weight component, thus providing a clear target reference for subsequent deviation calculations. Based on the reference trajectory segment sequence, structured trajectory deviation backtracking is performed on the weight components in the element clusters, generating a set of trajectory offset vectors for each element cluster. This set describes the direction and degree of systematic deviation between the current control state and the historical optimal trajectory in the form of deviation vectors. Significant offset components obtained after offset quantization are used, and a preset significant offset threshold effectively filters out unimportant deviations caused by random noise and minor fluctuations, retaining only statistically significant deviation vectors and avoiding over-correction. Offset compensation is then applied to the weight distribution in the element clusters based on the significant offset components. The resulting element cluster correction eliminates significant deviations already present in the historical trajectories while maintaining the local continuity and smoothness of the weight distribution. The entire backtracking calibration process does not rely on online learning or external models; it is entirely based on historical optimized trajectories stored in the cloud and preset empirical thresholds, ensuring repeatability and low computational latency. The final generated correction value provides the queue encapsulation module with precise control adjustment values, enabling the edge-cloud collaborative real-time control system to continuously optimize the current control effect by utilizing historical experience, thereby improving the convergence speed and steady-state accuracy of the system in repetitive industrial tasks.

[0067] The queue encapsulation module 106 is used to encapsulate the correction amount in a queue to obtain the closed-loop real-time control command of the industrial Internet of Things, including: In this embodiment of the invention, the correction amounts are sorted by time identifier to obtain an ordered list of the correction amounts; Based on the ordered list, each correction value is stored in the first-in-first-out buffer of the industrial Internet of Things, generating a filling sequence for the first-in-first-out buffer; Add batch markers to the correction amounts in the filling sequence to obtain the encapsulated data blocks of the correction amounts; The encapsulated data block of the correction amount is structured and encoded to obtain closed-loop real-time control instructions for industrial IoT edge nodes.

[0068] The edge node acquires all correction values ​​output by the backtracking calibration module. Each correction value carries a timestamp, which records the control cycle number at the time the correction value was generated. The edge node reads the timestamps of all correction values ​​and rearranges them in ascending order of the timestamps. After rearranging, a list organized in chronological order is formed, which is called the ordered list of correction values.

[0069] After obtaining the ordered list, the edge node locates the pre-configured FIFO (First-In, First-Out) buffer in the Industrial Internet of Things (IIoT). This buffer is a storage area with a fixed depth, equal to the maximum number of corrections it can hold within a control cycle. The edge node writes each correction in the ordered list sequentially into a free storage cell of the FIFO buffer, starting from the buffer's beginning address and moving the write pointer one cell forward after each correction is written. Once all corrections have been written, all storage cells in the FIFO buffer from the beginning address to the current write pointer position contain correction data. These corrections, arranged in the order they were written, are called the FIFO buffer's fill sequence.

[0070] Edge nodes perform an append batch tagging operation on the corrections in the fill sequence. Appending a batch tagging means appending a common batch number to all corrections in the current fill sequence. This batch number is generated by a counter inside the edge node, which automatically increments by one after each encapsulation operation. Specifically, the edge node reads the current counter value as the batch number and copies this batch number to the reserved field at the beginning of each correction in the fill sequence. After appending the batch number, the fill sequence, along with the batch numbers, forms a single, independent data unit, called the encapsulated data block of the corrections. The corrections in the encapsulated data block maintain their original order, and the batch number identifies that these corrections belong to the output of the same control cycle.

[0071] Edge nodes perform structured encoding on the encapsulated data blocks of correction quantities to obtain closed-loop real-time control commands for the industrial IoT edge nodes. Structured encoding refers to converting each correction quantity in the encapsulated data block into a binary code stream according to a pre-agreed communication protocol format between the edge node and the execution device. Specifically, the edge node reads the batch number from the encapsulated data block and converts it into a fixed-length binary code. Then, it sequentially reads each correction quantity, converting the start frequency, end frequency, and compensation value into binary code, and concatenates these into the batch number binary code according to the field order specified in the protocol. After all correction quantities have been converted, a preset end marker code is added to the end of the code stream. The resulting complete binary code stream is called the closed-loop real-time control command for the industrial IoT edge nodes. This command is sent to the corresponding edge node execution device via industrial fieldbus or real-time Ethernet to drive the actuator to complete the corresponding control actions.

[0072] The beneficial effects are as follows: by sorting the correction values ​​by time stamp to obtain an ordered list, it ensures that the correction values ​​enter the subsequent processing flow in the order of their generation time, avoiding control timing errors caused by out-of-order processing. The filling sequence generated after storing each correction value in a first-in-first-out buffer based on the ordered list creates a queue of correction values ​​in the buffer that is strictly arranged according to the writing order, providing a regular data organization form for batch encapsulation. Adding batch tags to the correction values ​​in the filling sequence yields encapsulated data blocks. All correction values ​​with the same batch number are associated with an atomic operation unit, allowing the receiving end to determine the integrity and timeliness of the control instructions based on the batch number. The closed-loop real-time control instructions obtained after structured encoding of the encapsulated data blocks follow the communication protocol between the edge node and the execution device. Each field in the instruction is clearly defined, and the receiving end can execute it directly without additional parsing. The entire queue encapsulation process is completed in the hardware queue and encoder of the edge node, without involving dynamic memory allocation or software protocol stacks, ensuring the determinism of control instruction generation and extremely low latency. The resulting closed-loop real-time control commands directly drive the execution devices in the Industrial Internet of Things, realizing a complete closed loop from sensor data acquisition to control command output, which significantly improves the system's real-time response capability and control reliability.

[0073] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A real-time edge-cloud collaborative control system for the Industrial Internet of Things, characterized in that, The system includes a time-frequency analysis module, an envelope confidence module, a collaborative mapping module, a clustering and partitioning module, a backtracking calibration module, and a queue encapsulation module, wherein: The time-frequency analysis module is used to dynamically analyze the time-series frame set in the industrial Internet of Things sensor data stream to obtain the frequency domain sparse spectrum of the time-series frame set. The envelope confidence module is used to perform envelope analysis on the residual error formed after threshold truncation of the frequency domain sparse spectrum based on the edge node status of the time-series frame set, so as to obtain the confidence interval of the frequency domain sparse spectrum. The collaborative mapping module is used to perform collaborative mapping on the confidence interval based on the edge-cloud global policy library of the Industrial Internet of Things to obtain the weight distribution of the frequency domain sparse spectrum. The clustering module is used to cluster the weight distribution to obtain the element clusters of the weight distribution; The backtracking calibration module is used to perform trajectory backtracking adaptive calibration on the element cluster based on the historical optimized trajectory of the edge and cloud of the Industrial Internet of Things, so as to obtain the correction amount of the element cluster. The queue encapsulation module is used to encapsulate the correction amount in a queue to obtain the closed-loop real-time control command of the industrial Internet of Things.

2. The edge-cloud collaborative real-time control system for industrial IoT as described in claim 1, characterized in that, The time-frequency analysis module performs dynamic analysis on the time-series frame set in the industrial IoT sensor data stream to obtain the frequency domain sparse spectrum of the time-series frame set, including: The industrial IoT sensor data stream is subjected to time-series window sliding segmentation to obtain multiple overlapping frame segments of the industrial IoT sensor data stream. Based on the multiple overlapping frame segments, the temporal energy distribution of the time-series frame set in the industrial IoT sensor data stream is aggregated and mapped to obtain the energy concentration region of the time-series frame set. The frequency domain basis decomposition of the energy concentration region is performed to obtain the frequency domain component characteristics of the energy concentration region; Based on the frequency domain component characteristics, the time-series frame set is subjected to spectral filtering to obtain the frequency domain sparse spectrum of the time-series frame set.

3. The edge-cloud collaborative real-time control system for industrial IoT as described in claim 1, characterized in that, The envelope confidence module performs envelope analysis on the residual error formed after threshold truncation of the frequency domain sparse spectrum based on the edge node states in the time-series frame set, to obtain the confidence interval of the frequency domain sparse spectrum, including: Feature separation is performed on the edge nodes in the time-series frame set to obtain the fluctuation amplitude parameter sequence of the edge nodes; Based on the fluctuation amplitude parameter sequence, the threshold truncation threshold of the frequency domain sparse spectrum is adaptively controlled and calculated to obtain the adaptive depth control value of the threshold truncation threshold. Based on the adaptive depth control value, the spectral line components of the frequency domain sparse spectrum are filtered out to obtain the residual error formed after the frequency domain sparse spectrum is truncated by a threshold. The remaining error is fitted with an envelope contour to obtain the upper envelope trajectory and the lower envelope trajectory of the remaining error. Based on the upper envelope trend trajectory and the lower envelope trend trajectory, the error amplitude variation in the frequency domain sparse spectrum is delimited by interval boundary to obtain the confidence interval of the frequency domain sparse spectrum.

4. The edge-cloud collaborative real-time control system for industrial IoT as described in claim 3, characterized in that, The formula for calculating the adaptive depth control value is as follows: ; In the formula, The adaptive depth control value is... The threshold is the cutoff threshold. This is the cyclic index variable for the fluctuation amplitude parameter. For the first The aforementioned fluctuation amplitude parameter The total number of the fluctuation amplitude parameter sequences. It is a very small positive number. A natural constant An exponential function with base 0. It is a logarithmic function.

5. A real-time edge-cloud collaborative control system for industrial IoT as described in claim 3, characterized in that, The envelope confidence module performs envelope contour fitting on the residual error based on the adaptive depth control value to obtain the upper and lower envelope trajectory of the residual error, including: The residual error is subjected to extreme point identification to obtain the original extreme point distribution sequence of the residual error; Based on the adaptive depth control value, the positive and negative half axes of the original extreme point distribution sequence are subjected to extreme value amplitude layer discrimination to obtain the upper envelope anchor point pre-selection set and the lower envelope anchor point pre-selection set of the remaining error. Linear interpolation is performed on adjacent anchor points in the pre-selected set of upper envelope anchor points and the pre-selected set of lower envelope anchor points to obtain the upper envelope trend trajectory and lower envelope trend trajectory of the remaining error.

6. The edge-cloud collaborative real-time control system for industrial IoT as described in claim 1, characterized in that, The collaborative mapping module executes a global policy library based on the industrial Internet of Things (IIoT) to collaboratively map the confidence interval, obtaining the weight distribution of the frequency domain sparse spectrum, including: Based on the working conditions of the policy entries in the global policy library of the Industrial Internet of Things (IIoT) at the edge and cloud, the confidence interval is retrievaled by tags to obtain the tag index sequence of the global policy library of the Industrial Internet of Things at the edge and cloud. Based on the label index sequence, anchor points are located in the confidence interval to obtain the mapped anchor point positions of the confidence interval; Based on the weight configuration rules in the global strategy library of the Industrial Internet of Things in the edge-cloud, weight factors are assigned to the mapping anchor positions to obtain the weight factor sequence of the mapping anchor positions. Based on the weight factor sequence, spectral line weights are configured for the spectral line components in the frequency domain sparse spectrum to obtain the weight distribution of the frequency domain sparse spectrum.

7. The edge-cloud collaborative real-time control system for industrial IoT as described in claim 1, characterized in that, The clustering module performs clustering on the weight distribution to obtain element clusters of the weight distribution, including: The component spacing of the weight distribution is measured to obtain the set of component distances of the weight distribution; Based on the component distance set, the weight components are clustered to obtain the initial grouping of the weight distribution; The initial grouping is subjected to an inter-group separation threshold determination to obtain the merging decision identifier interval of the initial grouping; Based on the merging decision identifier interval, adjacent groups in the initial grouping are merged and grouped to obtain the element clusters of the weight distribution.

8. The edge-cloud collaborative real-time control system for industrial IoT as described in claim 1, characterized in that, The backtracking calibration module executes historical optimization trajectories based on the industrial Internet of Things (IIoT) in the edge-cloud environment, performs trajectory backtracking adaptive calibration on the element cluster, and obtains the correction amount for the element cluster, including: The historical optimization trajectories of the edge and cloud of the industrial Internet of Things are extracted sequentially to obtain the reference trajectory fragment sequence of the element cluster; Based on the reference trajectory segment sequence, the weight components in the element cluster are subjected to structured trajectory deviation backtracking to obtain the trajectory offset vector set of the element cluster; The trajectory offset vector set is subjected to offset quantization to obtain the significant offset components of the trajectory offset vector set; Based on the significant offset components, the weight distribution in the element cluster is offset-compensated and superimposed to obtain the correction amount of the element cluster.

9. A real-time edge-cloud collaborative control system for industrial IoT as described in claim 8, characterized in that, The backtracking calibration module performs structured trajectory deviation backtracking on the weight components of the element cluster based on the reference trajectory segment sequence, obtaining a set of trajectory offset vectors for the element cluster, including: Perform time-series matching on the reference trajectory segment sequence to obtain the time correspondence between the reference trajectory segment sequence and the weight components; Based on the time correspondence, the weight components in the element cluster are compared with the segments at the same position in the reference trajectory segment sequence to obtain the original deviation value sequence of the weight components. The original deviation value sequence is filtered to obtain a set of valid deviation values ​​from the original deviation value sequence. Based on the set of effective deviation values, deviation vectors are constructed for the weight components in the element cluster to obtain the set of trajectory offset vectors for the element cluster.

10. A real-time edge-cloud collaborative control system for industrial IoT as described in claim 1, characterized in that, The queue encapsulation module performs queued encapsulation of the correction amount to obtain the closed-loop real-time control instructions for the industrial Internet of Things, including: The correction amounts are sorted by time identifier to obtain an ordered list of the correction amounts; Based on the ordered list, each correction value is stored in the first-in-first-out buffer of the industrial Internet of Things, generating a filling sequence for the first-in-first-out buffer; Add batch markers to the correction amounts in the filling sequence to obtain the encapsulated data blocks of the correction amounts; The encapsulated data block of the correction amount is structured and encoded to obtain closed-loop real-time control instructions for industrial IoT edge nodes.

Citation Information

Patent Citations

  • Edge-cloud collaborative user load prediction control and transaction system and implementation method thereof

    CN111754023A

  • Pumped storage unit technology water supply control system and method based on distributed Internet of Things

    CN120575626A