Edge cloud collaborative compression method for industrial data

Abstract

This invention discloses an edge-cloud collaborative compression method for industrial data, relating to the field of edge computing technology. The industrial gateway segments heterogeneous raw data streams into time-domain semantic data segments based on PLC step status words; it retrieves process correlation tables to determine the protection attribute markers for each channel, and combines this with sliding window incremental variance to determine signal status markers, forming a dual-marker combined index; based on this index, it queries a compression strategy mapping table to generate stage statistical summaries for stable state data and extract boundary key point data for fluctuating state data; upon alarm, it extracts abnormal time-domain raw data and causal correlation channel data to assemble an abnormal event packet; simultaneously, it calculates the original extreme values ​​and stage integral values ​​as original hard constraints, and asynchronously uploads compressed data packets according to priority; it receives new compression strategy versions from the cloud based on reverse synchronization of hard constraints and updates the compression strategy mapping table. This invention achieves hierarchical compression and adaptive adjustment of reconstruction accuracy for industrial data.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, specifically to an edge-cloud collaborative compression method for industrial data. Background Technology

[0002] Industrial data acquisition and edge-cloud collaborative processing are important development directions in the fields of intelligent manufacturing and the Industrial Internet. The aim is to achieve real-time monitoring of equipment status, process optimization, and fault diagnosis by efficiently transmitting equipment operation data from the industrial field to the cloud. In recent years, with advancements in edge computing and data compression algorithms, the transmission and storage of industrial data have gradually moved towards greater efficiency and adaptability.

[0003] Existing technologies often compress and transmit industrial data using fixed sampling rates or fluctuation judgment algorithms based on a single threshold to alleviate network bandwidth pressure between edge gateways and the cloud. These methods effectively reduce data transmission volume under most normal steady-state operating conditions and possess a certain degree of adaptive compression capability. However, in complex industrial production scenarios, equipment often involves frequent PLC step state switching and dynamic interweaving of multi-channel signals. This frequent state switching leads to non-stationary frequency variation and abrupt changes in industrial data, causing drastic fluctuations in the data's time-domain characteristic distribution and physical extrema. Existing technologies fail to effectively combine PLC step semantics with the process correlation between channels to dynamically adjust compression strategies. Furthermore, they lose original physical extrema and stage integrals during data compression, resulting in accumulated errors and physical quantity distortion when the cloud reconstructs the compressed data. This makes it difficult to accurately recreate the true operating conditions, increasing the risk of misjudgment of equipment status and inaccurate process control by the cloud. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an edge-cloud collaborative compression method for industrial data.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: This invention discloses an edge-cloud collaborative compression method for industrial data, comprising the following steps: Acquire heterogeneous raw data streams from multiple source devices and segment them into temporal semantic data segments; Retrieve the pre-configured process association table to determine the protection attribute markers of each channel within the time-domain semantic data segment. Simultaneously, calculate the incremental variance of the data within the time-domain semantic data segment using a sliding window to determine the signal status markers. Combine the protection attribute markers with the signal status markers to form a dual-marker combined index. According to the dual-label combined index, the preset compression strategy mapping table is queried, and the compression strategy data corresponding to each channel is read. The stage statistical summary is extracted from the stable state data, the boundary key point data is extracted from the fluctuating state data, and when the alarm is triggered, the abnormal correlation original data in the abnormal time domain and the causal correlation channel data locked by searching the process correlation table are extracted and assembled into an abnormal event package. At the same time, the original maximum value, original minimum value and corresponding sampling time and stage integral value of the time domain semantic data segment before compression are calculated and extracted to generate the original hard constraint quantity. The compressed data package containing the stage statistical summary, the boundary key point data, the abnormal event package, and the original hard constraint quantity will be asynchronously uploaded to the cloud server according to risk priority. Receive the new compression strategy version data from the cloud server via reverse synchronization, and update the compression strategy mapping table based on the new preset error threshold; The new compression strategy version data includes at least the adjusted new preset error threshold. The new compression strategy version data is obtained by reconstructing the curve of the compressed data packet and comparing the reconstruction result with the original hard constraint quantity. It is generated when the comparison deviation continues to exceed the preset error threshold within a preset judgment period.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention enables industrial gateways to simultaneously combine static process safety attributes and dynamic signal fluctuation characteristics to perform fine-grained classification of different data channels. This allows for the extraction of statistical summaries and boundary key points for stable and fluctuating states, respectively. While significantly reducing redundant transmission bandwidth, this invention avoids the damage to critical process and safety channel data caused by traditional unified compression algorithms, thus ensuring the feature integrity of core data.

[0007] 2. This invention enables the cloud to not only obtain complete abnormal data with physical context and time-series correlation when receiving compressed data packets, but also to obtain the global physical constraint benchmark before data compression. This eliminates the data isolation problem caused by sudden alarms and effectively prevents the accumulation of physical quantity deviations caused by lossy compression during cloud reconstruction.

[0008] 3. This invention enables the compression strategy at the edge to be dynamically adjusted based on the deviation assessment between the original hard constraint and the reconstruction result in the cloud. Therefore, it can adaptively match the step drift and complex working condition changes in the industrial production process, and maintain long-term stable compression accuracy and transmission efficiency. Attached Figure Description

[0009] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the present invention. In the accompanying drawings, the same reference numerals refer to the same components. Wherein: Figure 1 is a flowchart of the working principle of the present invention; Figure 2 is a timing flowchart of the present invention. Detailed Description of the Embodiments

[0010] It is easy to understand that, according to the technical solution of the present invention, without changing the essential spirit of the present invention, those of ordinary skill in the art can propose various structural modes and implementation modes that can be replaced with each other. Therefore, the following detailed embodiments and accompanying drawings are only exemplary descriptions of the technical solution of the present invention, and should not be regarded as the entirety of the present invention or as a definition or limitation of the technical solution of the present invention.

[0011] In the prior art, industrial field data acquisition and transmission mostly rely on fixed sampling rate upload or single threshold fluctuation compression, which makes it difficult to balance transmission bandwidth and reconstruction accuracy. When the working conditions of industrial equipment are switched frequently or multi-channel data is interleaved, the traditional methods are prone to data features being interfered by noise and having excessively high redundancy, resulting in a significant increase in cloud network transmission pressure. Existing edge gateway devices cannot synchronously combine process safety attributes and dynamic signal features to perform fine-grained distribution processing on data streams. Especially when a sudden fault causes multi-channel linkage anomalies, a single compression model will suffer from systematic information loss or loss of key anomaly context, making it difficult to meet the needs of accurate industrial on-site monitoring and accident traceability.

[0012] To solve the above problems, research has found that there is a coupling mapping rule between the PLC work step state of the equipment and the process correlation of the data channels, and a dual-label orthogonal decision model is established to achieve differentiated compression and distribution. It is further found that stationary data has high sensitivity to macroscopic statistical features, fluctuating data has high sensitivity to key time-domain boundary points, and abnormal data has extremely high requirements for original high-frequency sampling and causal channel correlation. Accordingly, the idea of dynamically matching compression strategies according to the orthogonal decision matrix and synchronously extracting original hard constraint quantities is proposed. Further experiments verify that hard constraint quantities such as stage integral values and the cloud reverse adjustment mechanism are introduced into the dynamic update process of the compression strategy to form a closed-loop feedback system.

[0013] After introducing the basic concept of the present invention, the embodiments of the present invention will be specifically described below with reference to the accompanying drawings.

[0014] Embodiment: As shown in Figure 1 and Figure 2As shown, the edge-cloud collaborative compression method for industrial data is implemented based on the hardware architecture of the industrial gateway. The industrial gateway includes a microprocessor (such as an embedded multi-core CPU or ARM processor), industrial communication interfaces (such as RS485, Ethernet bus, and CAN bus interfaces) physically connected to multi-source device buses, high-frequency random access memory (RAM), solid-state flash memory (Flash), and network communication modules (such as 5G / 4G or fiber optic Ethernet cards). The microprocessor calls pre-stored program instructions in the solid-state flash memory and allocates a data buffer in memory to drive the various hardware modules to collaboratively complete data reading, calculation, conversion, and transmission.

[0015] The microprocessor continuously acquires heterogeneous raw data streams from multiple sources through an industrial communication interface and writes the data into a memory buffer in real time. Based on the hardware step signals or system status instructions read from the microprocessor, the microprocessor divides the continuous heterogeneous raw data streams into time-domain segments to generate independent time-domain semantic data segments, thus completing the structured preprocessing of the data source.

[0016] The microprocessor retrieves the process association table pre-stored in the solid-state flash memory, obtains the attribute field corresponding to each data acquisition channel in the time-domain semantic data segment, assigns it a value, and determines it as a protection attribute marker. At the same time, the microprocessor allocates a sliding window memory area to perform continuous incremental variance calculation on the data points of each channel in the time-domain semantic data segment, logically compares the obtained incremental variance with the preset value, and determines the signal status marker reflecting the data fluctuation. The microprocessor overlaps and combines the protection attribute marker and the signal status marker to generate a dual-marker combination index containing multi-dimensional marker information.

[0017] The microprocessor uses a dual-label composite index as the query condition to retrieve the compression strategy mapping table pre-loaded in memory and read the compression strategy data corresponding to each channel. For stable state data, the microprocessor calculates its starting value, ending value, and mean, and assembles them to generate a stage statistical summary; for fluctuating state data, the microprocessor extracts its extreme points and inflection points, and assembles them to generate boundary key point data. When the microprocessor detects a hardware interrupt or software alarm trigger signal, it extracts the original abnormal correlation data in the corresponding abnormal time domain from the memory circular buffer, and again searches the process correlation table to lock the channel data with a causal mapping relationship, assembling them together into an abnormal event package.

[0018] Simultaneously, before compression, the microprocessor performs a full scan of the original time-domain semantic data segment, extracting the original maximum and minimum values ​​and their corresponding sampling time points. It then performs trapezoidal integration based on the sampling values ​​of adjacent sampling points and the sampling time interval to calculate the stage integral value. These parameters are packaged to generate the original hard constraint quantities, which serve as the posterior verification benchmark for cloud reconstruction. These original hard constraint quantities do not participate in the initial curve reconstruction in the cloud; they are only used for posterior verification deviation comparison with the corresponding statistics of the reconstructed signal after reconstruction.

[0019] The microprocessor packages the generated stage statistical summary, boundary key point data, anomaly event packets, and original hard constraint quantities into a compressed data packet with a unified structure. The microprocessor assigns different priority identifiers to each data packet according to preset data type security levels, and then calls the network communication module to asynchronously send the compressed data packet to the cloud server via the network in descending order of risk priority.

[0020] Finally, when the cloud server uses compressed data packets to reconstruct the curve and finds that the deviation between the reconstruction result and the original hard constraint quantity continues to exceed the preset error threshold within the preset judgment period, the cloud server generates and distributes a new compression strategy version data containing the adjusted new preset error threshold. The network communication module of the industrial gateway receives the new compression strategy version data and writes it into memory. The microprocessor parses out the new preset error threshold and performs online overwrite updates on the compression strategy mapping table in the solid-state flash memory and memory, completing the closed-loop adjustment of the compression parameters.

[0021] Preset judgment period The value ranges from 3 to 10 production cycles. The specific value is determined based on the process cycle time of the equipment. Equipment with shorter cycle times (such as injection molding machines and stamping machines) takes a larger value, while equipment with longer cycle times (such as reaction vessels and kilns) takes a smaller value.

[0022] This application achieves accurate classification, extraction, and compression of heterogeneous industrial data by combining process correlation and incremental variance to generate a dual-label combined index, significantly reducing data transmission bandwidth usage. By extracting the original hard constraint quantities and updating the compression strategy mapping table in conjunction with the cloud reverse synchronization mechanism, it identifies the accumulated deviation of physical quantities in lossy compression reconstruction, and suppresses the continuous expansion of deviation through subsequent compression strategy adjustments, thus ensuring the adaptability and accuracy stability of edge data processing.

[0023] This application further proposes monitoring the PLC step status word and segmenting the time-domain semantic data. Specifically, during the microprocessor's monitoring of the heterogeneous raw data stream, the microprocessor receives the status register data sent by the PLC in real time via an industrial communication interface through polling or interruption, and extracts the PLC step status word. The microprocessor continuously samples and compares the value of this status word. When the microprocessor detects that the PLC step status word changes from the current step value... Change to another step value To eliminate transient jitter in register values ​​caused by electromagnetic interference in industrial environments, the microprocessor starts an internal hardware timer to count. Only when the status word remains at a certain value... The duration reaches the preset confirmation period Only then does the microprocessor determine that a valid transition has occurred in the process step state. (Preset confirmation cycle) The value is determined based on the empirical value and experimental calibration of the PLC control cycle of the field equipment. A specific example range can be set to... (For example, take) or In addition, the microprocessor simultaneously parses high-priority control instructions on the bus. If it detects that the system has triggered a preset recipe change instruction, the microprocessor does not need to wait for the status word to change and immediately triggers the data segmentation logic.

[0024] When the conditions for a valid transition of the above-mentioned step status word or the triggering of a recipe change instruction are met, the microprocessor executes a segmentation operation. The microprocessor acquires the system's high-precision RTC clock, marks the timestamp of the current sampling point as the end time of the previous data segment, and disables data writing for that data segment. Simultaneously, the microprocessor allocates a new data structure in memory and establishes a new time-domain semantic data segment. The microprocessor reads the preset device hardware information and the current running recipe parameters from the solid-state flash memory, writes the read device identifier, recipe identifier, and preset stage identifier mapped according to the step transition value into the header attribute area of ​​the data structure; sets the current timestamp as the start time of the new data segment, and continuously appends subsequent real-time sampled data to the real-time data point set.

[0025] To further improve the accuracy of anti-shake determination in the process step, the determination process of the preset confirmation cycle can be calculated using the following continuous sliding window determination formula: ; In the formula, This represents the verification probability of the new process step state, with a value range normalized to [0,1]. N represents the preset confirmation period. The total number of times the internal microprocessor samples the PLC step status word (the value range is...). Second-rate); This represents the state word value obtained from the k-th sample. This is an indicator function that takes the value 1 when the condition is met, and 0 otherwise.

[0026] When calculated When the probability exceeds a preset probability threshold (e.g., 0.9 or 0.95), the microprocessor determines that the step switching is valid.

[0027] This application introduces a status word debouncing mechanism with a preset confirmation cycle and a dual triggering logic for formula change instructions. It combines multi-dimensional semantic identifiers such as equipment, stage, and formula to assemble time-domain semantic data segments, eliminating the erroneous segmentation phenomenon caused by industrial site noise. This ensures strict synchronization between data segmentation and physical production steps and process formulas, providing a data benchmark with a clear structure and explicit semantics for subsequent refined classification and compression.

[0028] This application further proposes that, after the microprocessor generates time-domain semantic data segments, it retrieves a pre-configured process association table and determines protection attribute markers. Specifically, the microprocessor calls the pre-configured process association table stored in solid-state flash memory (Flash). The process association table is a mapping database pre-built and written during the system initialization phase based on the hardware safety specifications, equipment process flow diagrams (P&ID), and physical channel topology of the industrial site.

[0029] The microprocessor extracts the unique hardware identifier (such as channel ID or bus address) of each data acquisition channel within the time-domain semantic data segment. Using the extracted hardware identifier and the preset stage identifier of the current data segment as a composite search key, the microprocessor queries the channel protection level mapping item in the process association table. The mapping item categorizes the physical security requirements of each channel at different production steps into different protection levels. Based on the query results, the microprocessor determines the stage protection level corresponding to each channel within the current time-domain semantic data segment and marks its protection attribute (denoted as ). Perform the assignment operation: When the mapping item indicates that the current channel is only used for routine environmental parameters or auxiliary status monitoring (such as cooling fan speed, routine chassis temperature, etc.) in the current process step, the microprocessor marks the protection attribute of that channel. Assigning a value to a normal channel flag (e.g., setting an enumeration value) When the mapping item shows that the data of the current channel is directly related to the product processing quality or the control of core process parameters (such as spindle speed, critical injection pressure, real-time temperature of the reactor, etc.), the microprocessor marks the protection attribute of that channel. Assigning values ​​to process protection channel markers (e.g., setting enumeration values) When the mapping item shows that the current channel is directly associated with hardware safety interlocks, emergency stop circuits, or overpressure / overtemperature safety valve status (such as safety light curtain trigger signal, emergency check valve pressure, explosion-proof area gas concentration, etc.), the microprocessor marks the protection attribute of that channel. Assigning values ​​to critical security channel markers (e.g., setting enumeration values) ).

[0030] To quantify the channel security weights under different protection levels, the mapping logic of protection attribute tags can be represented by the following standardized assignment function: ; In the formula, This represents the normalized protection weight of the i-th data channel, and its value range is defined as (0,1]. , and These are pre-calibrated empirical constants that satisfy the relation Specific example ranges can be set to Microprocessor according to The calculated values ​​are matched and written into the corresponding protection attribute tags.

[0031] This application achieves a refined classification of multi-channel heterogeneous data in terms of physical security and process quality by combining industrial physical topology with dynamic extraction of channel protection level mapping terms at each process step. This ensures that high-security and core process channels are given priority protection in subsequent data compression, providing an objective and reliable hardware and process semantic basis for the selection of differentiated compression strategies.

[0032] This application further proposes that the microprocessor allocates a fixed-length circular queue in memory as a sliding window for each data channel. With the continuous input of new sampled data points, the microprocessor uses an incremental update algorithm (such as the Welford algorithm) to update the statistics within the sliding window in real time, without needing to repeatedly traverse and calculate all historical data points within the window. Let the sample size of the current sliding window be W (its value range can be set according to the sampling frequency and dynamic response requirements). For example, if W=30), when a new data point is input... At that time, the microprocessor calculates the mean increment and the increment variance of the current window sequentially. .

[0033] When establishing the current temporal semantic data segment, the microprocessor reads the pre-configured fixed normalized lower limit from the process association table or sensor parameter table based on the device identifier, channel identifier, and stage identifier. and fixed normalization upper limit The fixed normalization upper and lower limits are determined based on the sensor's physical range or the historical calibration range of the corresponding production stage, and remain unchanged within the current time-domain semantic data segment. The k-th data point is normalized according to the following formula: ; in, To prevent the use of preset positive numbers with a denominator of zero, after the current data segment ends, the fixed normalization upper and lower limits used in the next production cycle can be updated based on the valid data range of that data segment, but the normalization benchmark of the current data segment will not be changed retroactively. ∈[0,1].

[0034] The formula for calculating incremental variance is as follows: ; ; In the formula, This represents the window mean after the k-th update; This indicates the data point that left the sliding window; This represents the real-time incremental variance after normalization.

[0035] Through the above recursive calculations, the microprocessor achieves low-complexity extraction of high-frequency data fluctuation features. This is followed by the completion of incremental variance... After the calculation, the microprocessor executes the signal status flag (denoted as...). The microprocessor employs a hierarchical judgment logic. First, it performs the highest-priority exception judgment: the microprocessor detects hardware interrupt status or software / hardware alarm flags. When it detects a hardware alarm event associated with the current channel, or a real-time data point of the current channel, it will... When the preset process upper or lower limit (i.e., the preset process boundary) is exceeded, the microprocessor ignores the variance calculation results and directly marks the signal status. It is identified as an abnormal state flag (e.g., by setting an enumeration value). ).

[0036] Without identifying it as an abnormal state, the microprocessor will calculate the incremental variance. Compared with the preset stability threshold stored in solid-state flash memory Perform numerical comparisons. Preset a stability threshold. The basis for this is the experimentally calibrated value of the background noise variance during the steady-state operation of the equipment multiplied by a safety margin factor (the specific example range can be set to...). When the incremental variance satisfies When the microprocessor determines that the data is in a stable state, it marks the signal state. Assigning a value to a stable state flag (e.g., setting an enumeration value) When the incremental variance satisfies At that time, the microprocessor determines that the data has dynamic fluctuations and marks the signal state. Assigning a value to a fluctuation state flag (e.g., setting an enumeration value) ).

[0037] For cases where the signal status is marked as an abnormal status, the microprocessor performs three levels of differentiated processing based on the protection attribute flag: When the abnormal state is detected and the protection attribute is marked as a normal channel, only the original data frames for the first preset duration (e.g. ±1s) before and after the alarm is triggered are captured, and the maximum number of associated channels is set to the first preset value (e.g. 3). When an abnormal state occurs and the protection attribute is marked as a process protection channel mark, the original data frame for the second preset duration (e.g., ±3s) before and after the alarm is triggered is captured. The upper limit of the number of associated channels is set to the second preset value (e.g., 5), and the last few steady-state sampling points before the alarm are retained as baseline references. When an abnormal state occurs and the protection attribute is marked as a critical safety channel, the original data frame is captured for a third preset time period (e.g., ±5s) before and after the alarm is triggered. All causal channels are retained (without any upper limit on the number of channels) and all high-frequency sampling data is completely retained. Any form of lossy compression is prohibited.

[0038] This application reduces edge computing power consumption through a sliding window incremental update algorithm, and achieves accurate three-level division of signal state by combining alarm / process boundary conditions with stable threshold comparison. This eliminates the time delay of traditional full calculation and provides a highly timely basis for selecting compression strategies in real time.

[0039] This application further proposes that, specifically, after the microprocessor generates the double-marked composite index, the microprocessor will mark it with a protection attribute. With signal status flags The concatenated double-label composite index serves as a composite key to retrieve the compression strategy mapping table pre-loaded into high-frequency random access memory (RAM), and performs data extraction and compression assembly according to the retrieved strategy instructions.

[0040] The signal state is marked as a stable state mark ( In the case of [missing information], the microprocessor performs three levels of differentiation refinement based on the protection attribute markers: First, when the protection attribute is marked as a normal channel marker ( When the microprocessor performs statistical calculations on all data points within the time-domain semantic data segment, it extracts the initial sampled value. End sampling value Time-domain arithmetic mean Overall variance , and assemble to generate a fixed-length stage statistical summary; Second, when the protection attribute is marked as a process protection channel mark ( When the microprocessor extracts the above-mentioned stage statistical summary, it additionally extracts and retains the original high-frequency sampling data points at the start and end of the stage for seamless boundary docking. Third, when the protection attribute is marked as a critical security passage marker ( When this occurs, the microprocessor calls the preset low-loss error threshold. (The values ​​are derived from the experimental calibration of the physical sensor's accuracy, and the range is...) For example, take The algorithm employs either the Rotating Door Algorithm (SDT) or the Douglas-Puk Algorithm (DP) for low-loss compression, prioritizing the retention of state transition points and fluctuation records exceeding the preset low-loss error threshold to ensure the traceability of key security channels.

[0041] The signal state is marked as a fluctuation state mark ( In the case of a certain error threshold, the microprocessor determines the error threshold based on the preset error threshold. (Its value range can be set to) For example, take Feature points are extracted from the data segment, and points are scanned point by point to identify maximum and minimum values ​​(extreme points), points where the value exceeds the preset upper or lower limit of the process (preset process boundary crossing points), and sampling points within the preset neighborhood before and after the PLC step change (adjacent points during PLC step switching, the neighborhood window can be used to select adjacent points). (Sampling points), assembling to generate boundary key point data. Preferably, when the protection attribute of the fluctuation channel is marked as a process protection channel mark or a critical safety channel mark, the microprocessor automatically switches the compression threshold to a preset protection error threshold. ,in Strictly meet: ; In the formula, The preset protection attenuation coefficient is defined in the range of (0,1) (specific examples can be taken as follows). or ), making By narrowing the allowable error range, higher density boundary key points are extracted.

[0042] This application establishes a dynamic balance between extremely high pressure reduction of steady-state ordinary data and high-precision preservation of safety / process key data through a differentiated compression strategy of precise matching of dual-marker combined index. This not only significantly reduces the transmission redundancy of non-critical steady-state data, but also ensures that the core channels and fluctuation edge features are not distorted, thereby improving the overall throughput efficiency and reconstruction fidelity of industrial data edge cloud transmission.

[0043] This application further proposes that, specifically, in the process of the microprocessor extracting boundary key point data from the fluctuation state data, for power frequency fluctuation data with repetitive cyclic characteristics, the microprocessor calls a hardware timer to obtain the stage start time of the current time-domain semantic data segment. With the end of the phase The microprocessor stores the original sampling timestamps for each sampling point within this data segment. The linear normalization mapping is calculated using the following formula: ; In the formula, This represents a dimensionless normalized time coordinate, with values ​​strictly normalized to [0,1]. Subsequently, the microprocessor discretely sets a preset number K of fixed time anchor points on this normalized time axis, either uniformly or non-uniformly. (in The value of the number of fixed time anchor points K is based on the highest harmonic frequency of the power frequency obtained from Fourier frequency domain analysis. Specific examples can be set to K=64, 128, or 256.

[0044] Next, the microprocessor uses the device identifier, recipe identifier, and preset stage identifier in the current time-domain semantic data segment as a composite index key to retrieve the local template library pre-stored in solid-state flash memory, and reads the cycle template curve that precisely matches the current operating conditions. and the corresponding template version identifier .

[0045] Subsequently, for each fixed time anchor point The microprocessor retrieves and locks the preceding and following raw data points immediately on either side of the normalized time axis. and ,satisfy The microprocessor performs piecewise linear interpolation calculations to obtain anchor point sample values ​​at fixed time anchor points. : ; The microprocessor will obtain the anchor point sample value The nominal template value of the periodic template curve at the corresponding anchor point Calculate the difference to generate the j-th residual term. The microprocessor arranges the K residual terms in chronological order to generate a stage residual sequence of fixed length K. When the template residual rate exceeds the template failure threshold for a preset number of consecutive production cycles, the corresponding template is marked as a template to be updated. From historical time-domain semantic data segments that have never experienced an alarm, have not crossed process boundaries, and whose template residual rate is below the template update sample threshold, the median of the sampled values ​​at each fixed time anchor point is extracted to generate the updated cycle template curve, and a new template version identifier is configured. The new template takes effect at the next production cycle boundary.

[0046] Finally, the microprocessor generates the stage residual sequence. Template version identifier The extracted boundary key point data is associated and mapped, packaged and encapsulated into a uniformly structured fluctuating compressed data, and written into a memory buffer and included in the compressed data package before being uploaded to the cloud server.

[0047] This application eliminates the time-domain stretching and deformation caused by production rhythm fluctuations through time-axis linear normalization and piecewise linear interpolation. By using local periodic templates for subtraction, it transmits only a fixed-length stage residual sequence and template version identifier, which greatly reduces the amount of data transmitted for periodic fluctuations and ensures high-precision alignment and low-loss reconstruction of data in the cloud.

[0048] This application further proposes that, specifically, during the process of the microprocessor detecting a system-triggered hardware or software alarm signal, the microprocessor initiates an exception capture interrupt service routine. First, the microprocessor locks the hardware or logic channel that triggered the alarm and records it as the primary exception channel. Then, using the channel identifier of the primary exception channel and the current PLC step status word as the search key, the microprocessor calls a pre-configured process association table in the solid-state flash memory (Flash) to query and obtain a list of causally related channels that have a causal timing relationship with the primary exception channel.

[0049] To prevent large-scale channel linkage caused by a single point of failure in industrial sites, which could lead to an increase in the size of abnormal event packets and block network transmission, the microprocessor reads the pre-configured association priority rules and the preset channel limit. (Its value range is usually set according to the gateway cache capacity and network bandwidth limit.) For example, take The microprocessor filters the list of causal association channels in descending order of association priority, selecting only those causal association channels that do not exceed the preset maximum number of channels for subsequent data extraction.

[0050] Next, the microprocessor determines the pre-buffer duration based on the preset pre-buffer duration stored in the solid-state flash memory. (For example, take) ) and time to return to stability (For example, take) The abnormal time domain interval was calculated. .in, This is the alarm trigger time. The microprocessor continuously monitors the data value of the main abnormal channel. When the data value falls back to the preset normal range and remains there for a certain period of time until it stabilizes, the alarm will be triggered. When, this moment is defined as That is, the effective recovery time of the abnormal signal. The termination time of the abnormal time domain is determined as follows: .

[0051] The microprocessor continuously extracts raw sampled data frames from the circular buffer of memory, which are high-frequency samples of the main exception channel and the selected causal correlation channel in the exception time domain, through DMA (direct memory access).

[0052] Furthermore, to effectively address the frequent occurrence of complex or cascading anomaly alarms in industrial settings, the microprocessor, while generating the current anomaly window, simultaneously searches the memory-maintained queue of historical, unuploaded anomaly windows. The microprocessor performs a time-domain overlap determination between the time interval of the current anomaly window and the time intervals of each historical anomaly window in the queue: if multiple anomaly windows overlap in time domain, or if the time interval between two adjacent anomaly windows... Not greater than the preset merging threshold (Its value range can be set according to the process recovery cycle.) If this occurs, the microprocessor will trigger the window merging mechanism.

[0053] Suppose there are M exception windows to be merged. The microprocessor first merges them according to their start times. Sort the windows in ascending order. After sorting, determine whether adjacent windows meet the merging conditions: based on the union operation and merging rules for temporal overlap anomaly windows, the merged anomaly temporal intervals are formed after the union operation. The formula is as follows: ; In the formula, M represents the total number of abnormal windows participating in the merger; and These represent the start and end times of the m-th abnormal window, respectively.

[0054] The microprocessor performs a union operation on the time intervals of multiple abnormal windows that meet the merging conditions, merging them into a unified extended abnormal time domain, only when the main abnormal channels of multiple abnormal windows are the same or confirmed to belong to the same abnormal causal chain by searching the process correlation table, and the estimated data volume after merging is not greater than the data volume of the independently packaged data. The microprocessor then extracts the common original cache data within this extended time domain once and associates it with multiple alarm codes. Simultaneously, it establishes a mapping index table between each alarm code and its original window time period and cache address, assembling it into a final abnormal event package containing multiple alarm codes and a single common original cache data. If the aforementioned preconditions are not met, merging is abandoned, and the data is independently packaged and uploaded.

[0055] The microprocessor starts from the current main anomaly channel and performs a directed depth-first traversal along the predefined causal timing edges in the process association table. If the node of another anomaly channel can be reached from the node of the current main anomaly channel within a preset maximum tracing depth (e.g., 3 hops), then the two are determined to belong to the same anomaly causal chain. During the traversal, signal propagation paths within the same equipment unit take priority, while cross-unit propagation requires confirmation of a direct material or energy coupling relationship through the process association table.

[0056] After merging, the microprocessor establishes an alarm code mapping index table, mapping each alarm code to the start and end times of its respective abnormal window, and pointing to the same common cache data area. The causal relationship channel is the union of the associated channels of each window to be merged, and is uniformly truncated according to the retention time corresponding to the highest protection level, ensuring that the causal context corresponding to any alarm is completely covered.

[0057] This application automatically locks the causal channel through the process association table and introduces a quantity limit control. Combined with the precise extraction of the pre- and post-buffer durations, as well as the temporal union operation and public cache deduplication mechanism for high-frequency overlapping abnormal windows, it not only fully preserves the causal time context and original high-frequency characteristics of complex industrial accidents, but also greatly reduces the amount of data redundancy during sudden alarms, ensuring the efficient and reliable uploading of abnormal event packets.

[0058] This application further proposes that, after the microprocessor assembles and generates a compressed data packet containing stage statistical summaries, boundary key point data, anomaly event packets, and original hard constraint quantities, specifically, the microprocessor calls a network communication module (such as a 5G / 4G module or Ethernet card) to execute asynchronous upload and dynamic memory control logic. First, the microprocessor allocates four independent circular first-in-first-out (FIFO) queues in random access memory (RAM), establishing a queue containing security anomalies (highest priority). Equipment alarm (second highest priority) Boundary key points (medium priority) ) and stage statistical summary (low priority) The microprocessor binds and associates the generated raw hard constraint quantities with their corresponding abnormal event packets (such as extended packets containing safety anomalies or device alarms) or boundary key point data, and assigns them the same priority attribute (i.e., assigns them different priority attributes). , or This ensures that hard constraint quantities and associated compressed feature data are transmitted synchronously.

[0059] Based on this, the microprocessor reads the hardware system status in real time and monitors the gateway's dynamic memory usage. (Defined as the ratio of used memory to total memory, with values ​​normalized to [0,1]), and compared with a pre-set security threshold. (For example, 60% or 70%) and preset critical thresholds (For example, using 85% or 90%) to perform step-by-step comparison and adjustment: First, when the microprocessor determines that the gateway memory usage rate meets the requirements... At that time, the system is in a memory-safe state, and the microprocessor schedules the network communication module according to... Data packets in each queue are sent asynchronously in order of priority.

[0060] Second, when the microprocessor determines that the gateway memory usage rate meets the requirements... At this time, the system is in a memory warning state, and the microprocessor triggers low-priority flow control, suspending the transmission of low-priority stage statistics summary queues. This allocates all network communication resources to high-priority security anomalies, device alarms, and boundary critical point data packets. This prioritizes ensuring the timely transmission of critical features.

[0061] Third, when the microprocessor determines that the gateway memory usage rate has further increased and meets the requirements... At this time, the system is in a critical memory congestion state. The microprocessor triggers the memory protection mechanism: the microprocessor locks the circular buffer area storing abnormal event packets, and before receiving the TCP / ACK confirmation instruction returned by the cloud server through the network communication module, it is strictly prohibited to use this buffer area as a priority area for overwriting or releasing, to prevent core fault data packet loss; at the same time, the microprocessor starts a background compression thread to perform multi-stage secondary merging and compression on the stage statistical summaries accumulated in the queue.

[0062] Specifically, the microprocessor scans the accumulated statistical summaries of consecutive stages in memory, retrieving and matching the device identifier, stage identifier, recipe identifier, and policy version identifier in their header attributes. When a consecutive stage is determined... When all four identifiers of the statistical summary for each stage are completely identical, the microprocessor extracts this... The corresponding fields in each data segment are merged and calculated to generate a merged statistical summary. The merge calculation formula is defined as follows: ; ; In the formula, and These represent the maximum and minimum values ​​of the merged population, respectively. and These represent the combined cumulative score and the cumulative duration, respectively. This represents the combined mean obtained through weighted calculation; and These represent the number of original sampling points and the stage mean corresponding to the statistical summary of the i-th stage, respectively. After merging to generate a single merged statistical summary, the microprocessor releases the memory space occupied by low-priority data that has been confirmed and uploaded to the cloud or replaced by the merge, bringing the memory usage rate back to a safe range.

[0063] This application establishes a risk priority queue and binds hard constraints, combined with a tiered control mechanism for gateway memory occupancy and a multi-segment secondary merging algorithm for continuous statistical summaries, to achieve lossless transmission of core security data and key features under network congestion or sudden high traffic conditions. At the same time, it completely avoids the risk of gateway memory overflow by dynamically reducing memory usage at the edge.

[0064] This application further proposes that, in the industrial data edge-cloud collaborative architecture, the cloud server primarily relies on reconstruction accuracy, and optionally considers network bandwidth utilization and industrial process quality indicators to generate updated compression strategies as needed. In this embodiment, reconstruction deviation is specifically used as the quantitative basis for threshold tightening. Specifically, the microprocessor of the edge gateway listens to the policy control channel issued by the cloud through a network communication module (such as an MQTT / HTTPS protocol interface), and receives new compression strategy version data packets from the cloud server in real time.

[0065] The cloud server calculates the average reconstruction deviation within a preset judgment period. The current preset error threshold is denoted as And generate the tightening ratio according to the following rules. : ; in, This is a preset adjustment coefficient (range 0.5~1.0). and These represent the lower and upper limits of the adjustment ratio for a single transaction (example value range: The cloud server further generates a new preset error threshold according to the following formula: ; in, The minimum permissible error threshold is determined based on the sensor's measurement accuracy and analog-to-digital conversion resolution. Therefore, the new preset error threshold never exceeds the current preset error threshold. When the reconstruction accuracy is insufficient, the edge end is forced to extract higher-density boundary key points in the next production cycle, thus actively suppressing compression loss.

[0066] The new compression policy version data packet contains a complete control metadata structure, which the microprocessor temporarily stores in the policy buffer of dynamic random access memory (DRAM). The parsed field structure of this data packet includes: target device identifier. Target channel identification Target stage identifier New strategy version number Previous strategy version number Effective production cycle identifier and the new preset error threshold (as updated) or ).

[0067] The microprocessor first reads and parses the effective production cycle identifier. To avoid inconsistencies in data compression benchmarks and disruption of temporal semantic integrity caused by dynamically modifying thresholds during the current production cycle (such as in-process batch processing, step cutting, or reaction phase), the microprocessor employs an "asynchronous reception, cycle boundary aligned update" control mechanism. 1. Version consistency check: The microprocessor reads the version number of the currently running policy and compares it with the version number of the previous policy in the data packet. If the two match, the verification passes and the update is allowed to enter the pending queue; if they do not match, the microprocessor sends a version conflict exception alarm to the cloud and rejects the invalid update.

[0068] 2. Cycle Boundary Trigger Determination: The microprocessor continuously monitors the PLC step status word and production batch register transmitted on the bus (such as Profinet or Modbus-TCP). When the microprocessor detects the current production cycle trigger end signal (e.g., PLC outputs a batch end flag or the step count reaches zero), and captures the production cycle identifier of the upcoming new production cycle and the effective production cycle identifier... When they are completely identical, the microprocessor triggers a policy update interrupt.

[0069] 3. Parameter Atom Overlay: During the idle time window between two batches, the microprocessor determines the target device identifier. Target channel identification and target phase identifiers As a composite index, the compression strategy mapping table in solid-state flash memory (Flash) and running memory (RAM) is retrieved, and the corresponding threshold parameter item is atomically replaced with the new preset error threshold. And update the strategy version number to .

[0070] Through the above mechanism, the updated compression strategy takes effect seamlessly the moment the next production cycle starts, and all subsequent feature extraction and segmentation of sampled data are based on the new version number. and threshold implement.

[0071] Through the above technical solutions, this application ensures the consistency of data semantics and statistical benchmarks: by parsing the effective production cycle identifier and performing threshold coverage at the cycle switching boundary, it completely avoids data gaps and statistical biases caused by sudden changes in compression accuracy during production, ensuring the integrity and closed-loop consistency of data analysis and reconstruction for the same batch at both the edge and cloud ends. It achieves highly reliable edge-cloud closed-loop adaptive optimization: combining version number verification and atomic update mechanisms, it ensures that the edge gateway can seamlessly respond to dynamic compression strategies issued by the cloud based on the global situation, preventing decompression failures caused by version mismatches and realizing the intelligent and dynamic evolution of industrial edge data compression accuracy.

[0072] This application further proposes that, in the microprocessor, based on a new preset error threshold... After completing the atomic overwrite and update of parameter items in the compression strategy mapping table, specifically, the system enters the online trial operation and closed-loop verification phase of the new strategy. The microprocessor then operates at the new preset error threshold. During each production cycle after the policy takes effect, the reconstructed verification data (including the previously extracted stage statistical summary, boundary key point data, and original hard constraint quantities as verification benchmarks, such as global extrema and stage integral values) will be extracted in real time from the cloud server based on the corresponding temporal semantic data segments of the compressed data package. This data will also include the new policy version number. The compressed data packet is uploaded asynchronously to the cloud server.

[0073] After receiving the compressed data packet, the cloud server uses a mirror reconstruction algorithm to restore the original continuous signal and compares its accuracy with the synchronously received original hard constraint quantity to calculate the actual reconstruction deviation within the current production cycle k. .set up and These represent the maximum and minimum values ​​of the cloud-based reconstructed data during the k-th production cycle. The stage integral values ​​for reconstructing the data; the corresponding original hard constraints are respectively , and The three types of deviations for the k-th production cycle are defined as follows: Maximum deviation: ; Minimum deviation: ; Integral deviation: ; Comprehensive Reconstruction Deviation Defined as the maximum value of the above three types of deviation: .when If the error exceeds the preset error threshold corresponding to the previous policy version number for a consecutive preset number of production cycles, a security downgrade decision is triggered.

[0074] Let the preset error threshold corresponding to the current new strategy version be . The preset error threshold corresponding to the previous strategy version was (And under normal circumstances, because this scheme adopts a unidirectional tightening strategy, it satisfies...) ).

[0075] Cloud servers for continuous Comprehensive reconfiguration deviation within a production cycle Perform statistical analysis and execute the following hierarchical judgment and response: Level 1 (Accuracy Qualified): When In continuous If the condition is met within each cycle, the cloud determines that the current compression strategy meets the reconstruction accuracy requirements and maintains the existing strategy unchanged.

[0076] Level 2 (Precision Fine-tuning Tightening): When In continuous All of them satisfy within each period At this point, the cloud server determines that although the current strategy has not severely failed, the compression loss has exceeded the newly set tolerance boundary and is within an acceptable degradation range. The cloud server then calculates the further reduction error threshold based on the tightening rules described above. and generate a file containing The next version of the compression strategy, after further tightening, is sent to the edge gateway to moderately improve compression accuracy in the next production cycle, in order to reduce deviations. Within.

[0077] Level 3 (Emergency Strategy Security Degradation): When... In continuous All of them satisfy within each period Furthermore, the deviation increases continuously with the increase of the production cycle number k (i.e., the slope condition is satisfied). When the cloud determines that the current new strategy has caused structural deviations due to model mismatch or sensor drift, further tightening the threshold will lead to a sharp drop in compression ratio or even algorithm failure. At this point, the cloud server immediately issues a strategy rollback command, forcing the edge gateway to revert to the parameter configuration corresponding to the previous strategy version number.

[0078] In this scheme, the compression strategy is adjusted in a unidirectional tightening direction, meaning the new preset error threshold is always satisfied. The aim is to continuously improve reconstruction accuracy, and there is no situation where the threshold is relaxed to increase the compression ratio. Based on this, It is defined as the minimum accuracy safety baseline for the reconstruction accuracy of this channel under the current operating conditions—that is, the minimum accuracy requirement that the previous version of the strategy has been verified through long-term operation and can guarantee the effectiveness of industrial monitoring.

[0079] Rollback triggering is not simply due to deviation exceeding the limit. That is, execution is not required, but rather the simultaneous fulfillment of two conditions: (1) the reconstruction deviation is in continuous Each production cycle is greater than (1) The safety bottom line is breached; (2) The deviation shows a continuous increasing trend (slope condition) The combined determination of the two conditions ensures that rollback is triggered only when the new strategy has caused structural failure, rather than simply due to insufficient fine-tuning caused by random fluctuations, thereby eliminating the interference of normal operating condition fluctuations on rollback decisions.

[0080] The above design ensures logical consistency between policy tightening and security degradation: in and Within the established tolerance range, a step-by-step fine-tuning process is implemented, with a rollback initiated only when the historical safety threshold is breached and the trend deteriorates, restoring the strategy to a verified and stable version.

[0081] The cloud server maintains a historical deviation monitoring queue in real time. When the cloud server detects a continuous deviation... Within a production cycle ( The value range can be set according to the production cycle time. For example, take The reconstruction deviations consistently exceed the historical error threshold corresponding to the previous strategy version number, i.e., they satisfy the following: Furthermore, the reconstruction deviation shows a continuous increasing trend over time (i.e., it satisfies the slope condition). When the cloud server determines that the new policy has caused cumulative drift or model mismatch risk in the actual industrial field, it sends a policy rollback command containing the target device identifier, target channel identifier, and target stage identifier to the edge gateway through the policy control channel.

[0082] After receiving the policy rollback instruction through the network communication module, the microprocessor of the edge gateway triggers the security degradation interrupt handler: 1. Version Locking and Extraction: The microprocessor locks the currently effective policy version and extracts the previous policy version number recorded in the new compressed policy version data packet stored in dynamic memory or the local metadata log. .

[0083] 2. Log Recovery and Overwriting: The microprocessor uses the target device identifier, target channel identifier, target stage identifier, and previous policy version number. As a composite index, it retrieves historically stored local policy change logs in solid-state flash memory (Flash), reading and... Precisely corresponding historical preset error threshold Subsequently, the microprocessor will switch at the end of the current production cycle boundary. The overlay of atoms is written into the corresponding parameter item in the compression strategy mapping table.

[0084] 3. Status Flag Reset: The microprocessor resets the version number of the currently effective policy from [previous version number]. Switch back and set At the same time, it sends a security downgrade completion confirmation signal (ACK) to the cloud server to end the security downgrade process.

[0085] Through the above technical solutions, this application constructs an industrial-grade edge-cloud collaborative self-healing and closed-loop fault-tolerant mechanism: by continuously monitoring and trend analysis of reconstruction deviations in the cloud, it can keenly detect the risk of new strategy mismatch caused by aging field sensors, environmental drift, or sudden changes in operating conditions, and automatically issue instructions to restore to a stable version without human intervention, completely eliminating the risk of continuous data distortion and system loss of control that may be caused by online strategy upgrades. It ensures high availability and traceability of industrial data throughout its entire lifecycle: combined with version retrieval and atomic parameter recovery from local logs at the edge, it ensures seamless connection during strategy switching and security degradation processes, keeping the data compression baseline within a traceable and accurately reconstructable safe range, significantly improving the long-term robustness of industrial edge gateways under complex and harsh operating conditions.

[0086] The following is a specific embodiment of an edge-cloud collaborative compression method for industrial data: A large chemical plant is equipped with an edge smart gateway that connects to the physical sensor array (sampling rate 100Hz) of the ammonia synthesis reaction tower via Modbus-RTU and Profinet bus to monitor reaction process data in real time. During the switching of reaction tower operating conditions, the edge gateway captures the top pressure of the reaction tower (peak 15.2MPa), catalyst bed temperature (currently maximum 480℃), and circulating gas flow rate (dynamic range 12000-15000m³) in real time.3 The gateway directly reads the step status word of the PLC controller (Siemens S7-1500), parses it to obtain that it is currently in the "pressurization and heating" step (step code 0x03), and triggers the temporal semantic segmentation of the data stream.

[0087] The gateway microprocessor extracts signal status markers including reaction tower top pressure (marked as critical safety channel 0x02 by protection attribute), bed temperature (process protection channel 0x01), and circulating gas flow rate (normal channel 0x00). It calculates the incremental variance for a continuous sliding window (60s in length, 50% overlap) to determine if the current bed temperature is in a fluctuating state. The circulating gas flow rate is in a stable state. The microprocessor retrieves the strategy mapping table and calculates the initial value of 13500 m³ / s for the recirculating gas flow rate. 3 / h, End value 13520m 3 / h, average 13510m 3 / h, variance 0.8, statistical summary of the assembly generation stage; for bed temperature, based on a preset error threshold. Extract the extreme point 480℃, the boundary crossing point 450℃, and adjacent points for process step switching, and automatically switch to the preset protection error threshold. High-density boundary key point data were extracted. Simultaneously, the global extreme value of the reaction tower top pressure (15.2 MPa) and the stage integral value within this time period were extracted and used as the original hard constraint quantities.

[0088] To address the periodic fluctuations in circulating airflow, the gateway obtains the start time of the phase in the time-domain semantic segment. and the end time Perform linear normalization on the timestamp Set K=64 fixed time anchor points on the normalized time axis. Read the periodic template curve and version identifier ID001 that match the current recipe from the local template library. Perform piecewise linear interpolation on the original data points on both sides of the anchor point to calculate the anchor point sample value, and calculate the difference with the template curve to generate a stage residual sequence with a fixed length of 64.

[0089] When the bed temperature exceeds the safety limit (485℃) and triggers an equipment alarm, the gateway locks the main anomaly channel (bed temperature), searches the process correlation table, and selects the causal correlation channel (reaction tower top pressure, cooling water flow rate) according to priority. This is based on the pre-buffer duration. Duration of recovery to stability Capture the original data frame. When a time interval between the detected and adjacent abnormal windows is... Less than the preset merging threshold At that time, the gateway further determines that the main abnormal channels (bed temperature) corresponding to the two abnormal windows belong to the same abnormal causal chain, and estimates that the amount of encoded data after merging is less than the amount of encoded data encapsulated separately; after the above conditions are met, a union operation is performed on the abnormal time interval. s, and establish a mapping index between the dual alarm codes and their respective original window periods and cache addresses, and assemble them into an abnormal event package containing the dual alarm codes and a single common original cache data.

[0090] Data packets are uploaded asynchronously according to priority. When the gateway memory usage reaches a preset security threshold ( When a low-priority phase statistical summary is sent, transmission of security anomaly and boundary critical point packets is prioritized. Three consecutive phase statistical summaries with the same identifier are merged, the overall extreme value and cumulative integral value are extracted, a weighted average is calculated, and memory space is released promptly.

[0091] After receiving data packets, the cloud server reconstructs the signal. When it detects a continuous increasing trend in reconstruction deviations over five consecutive production cycles, it sends a policy rollback command to the gateway. The gateway responds by parsing the previous policy version number, restoring the historical threshold overwrite mapping table from local logs, and seamlessly switching back to the stable version.

[0092] This embodiment solves the technical problems of network congestion and high storage costs caused by concurrent uploading of high-frequency industrial data by using edge-cloud collaborative semantic segmentation, differentiated strategy matching, and multi-dimensional data compression. While ensuring high fidelity of critical process and safety alarm data, it significantly reduces the transmission redundancy of non-critical steady-state data, and achieves high-precision closed-loop control and dynamic self-healing for data reconstruction at both the edge and cloud ends, ensuring the long-term stable operation of large-scale industrial sites.

[0093] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. An edge-cloud collaborative compression method for industrial data, characterized in that, The industrial gateway performs the following steps: Acquire heterogeneous raw data streams from multiple source devices and segment them into temporal semantic data segments; Retrieve the pre-configured process association table to determine the protection attribute markers of each channel within the time-domain semantic data segment. Simultaneously, calculate the incremental variance of the data within the time-domain semantic data segment using a sliding window to determine the signal status markers. Combine the protection attribute markers with the signal status markers to form a dual-marker combined index. The preset compression strategy mapping table is queried according to the dual-marker combined index, and the compression strategy data corresponding to each channel is read. A stage statistical summary is extracted from the stable state data, and boundary key point data is extracted from the fluctuating state data. When an alarm is triggered, the original abnormal correlation data in the abnormal time domain and the causal correlation channel data locked by searching the process correlation table are extracted and assembled into an abnormal event package. At the same time, the original maximum value, original minimum value and corresponding sampling time and stage integral value of the time domain semantic data segment before compression are calculated and extracted to generate the original hard constraint quantity. The compressed data package containing the stage statistical summary, the boundary key point data, the abnormal event package, and the original hard constraint quantity will be asynchronously uploaded to the cloud server according to risk priority. Receive the new compression strategy version data from the cloud server via reverse synchronization, and update the compression strategy mapping table based on the new preset error threshold; The new compression strategy version data includes at least the adjusted new preset error threshold. The new compression strategy version data is obtained by reconstructing the curve of the compressed data packet and comparing the reconstruction result with the original hard constraint quantity. It is generated when the comparison deviation continues to exceed the preset error threshold within a preset judgment period.

2. The edge-cloud collaborative compression method for industrial data according to claim 1, characterized in that, Acquire heterogeneous raw data streams from multiple source devices and segment them into temporal semantic data segments, specifically including: Monitor the PLC step status word in the heterogeneous raw data stream. When the PLC step status word changes from the current step value to another step value and continues to reach a preset confirmation period, or when a preset recipe change instruction is triggered, end the current time domain semantic data segment, and simultaneously establish the time domain semantic data segment containing the device identifier, preset stage identifier, recipe identifier, stage start time, stage end time, and real-time data point set.

3. The edge-cloud collaborative compression method for industrial data according to claim 1, characterized in that, Retrieve the pre-configured process association table to determine the protection attribute markers for each channel within the time-domain semantic data segment, specifically including: Extract the channel protection level mapping item from the process association table, determine the stage protection level corresponding to each channel in the time domain semantic data segment, and assign the protection attribute mark to the ordinary channel mark, process protection channel mark, or critical safety channel mark.

4. The edge-cloud collaborative compression method for industrial data according to claim 1, characterized in that, The incremental variance of the data within the time-domain semantic data segment is calculated using a sliding window increment to determine the signal state marker, including: Continuous incremental updates are performed on the data points within the time-domain semantic data segment, and the incremental variance within the current sliding window is calculated. When an alarm event associated with the current channel is detected, or when the data of the current channel crosses the corresponding preset process boundary, the signal status flag is determined as an abnormal status flag. In the absence of an abnormal state marker, if the incremental variance is not greater than a preset stability threshold, the signal state marker is determined to be a stable state marker. If the incremental variance is greater than the preset stability threshold when the abnormal state is not identified, the signal state will be identified as a fluctuating state.

5. The edge-cloud collaborative compression method for industrial data according to claim 3, characterized in that, Statistical summaries are extracted from steady-state data to generate phase-specific data, and boundary key point data are extracted from fluctuating-state data to generate boundary key point data, including: Query the compression strategy mapping table based on the double-labeled combined index: When the signal state is marked as a stable state and the protection attribute is marked as a normal channel, the start value, end value, mean, and variance within the time-domain semantic data segment are extracted and assembled to generate the stage statistical summary. When the signal status is marked as a stable state and the protection attribute is marked as a process protection channel, the stage statistical summary is extracted, and the original data points at the beginning and end of the stage are retained. When the signal state is marked as a stable state and the protection attribute is marked as a critical safety channel, compression is performed using a preset low-loss error threshold to retain the state change record. When the signal state is marked as a fluctuation state, the extreme points, preset process boundary crossing points, and adjacent points of PLC step switching within the time domain semantic data segment are extracted based on a preset error threshold, and the boundary key point data is assembled. When the protection attribute is marked as a process protection channel mark or a key safety channel mark, a preset protection error threshold smaller than the preset error threshold is used for extraction.

6. The edge-cloud collaborative compression method for industrial data according to claim 2, characterized in that, Extracting boundary key point data from fluctuation state data also includes: Based on the start and end times of the phase of the temporal semantic data segment, the sampling timestamps of the fluctuation state data are linearly normalized to form normalized time coordinates; a preset number of fixed time anchors are set on the normalized time coordinates. Retrieves the cycle template curve and template version identifier that match the equipment identifier, recipe identifier, and current stage identifier from the local template library; Piecewise linear interpolation is performed on adjacent original data points located on both sides of each fixed time anchor point to obtain the anchor point sample value corresponding to each fixed time anchor point, and the difference is calculated with the periodic template curve to generate a fixed-length stage residual sequence. The stage residual sequence, the template version identifier, and the boundary key point data are associated and encapsulated into fluctuation compressed data, and then uploaded in the compressed data package.

7. The edge-cloud collaborative compression method for industrial data according to claim 1, characterized in that, Extract the original abnormal correlation data within the abnormal time domain and the causal correlation channel data locked by retrieving the process correlation table, and assemble them into an abnormal event package, including: When an alarm is detected, the main abnormal channel that triggered the alarm is locked, and the process correlation table is retrieved to obtain a list of causally related channels that have a causal time relationship with the main abnormal channel. According to the pre-configured association priority, select no more than the preset maximum number of causal association channels from the list of causal association channels; Based on the protection attribute marker of the main abnormal channel, the corresponding pre-buffer duration, recovery stabilization duration, and associated channel retention rules are determined from the compression strategy mapping table; If multiple abnormal windows with overlapping time domains are determined, and if multiple abnormal windows have overlapping time domains, or the time interval between adjacent abnormal windows is not greater than a preset merging threshold, and the corresponding main abnormal channels are the same or belong to the same abnormal causal chain, and the expected data volume after merging is not greater than the data volume of each encapsulated window, then a union operation is performed on the time intervals of the multiple abnormal windows, and a mapping index is established between the alarm code and the main abnormal channel and the location of the common cache data, and the abnormal event package containing multiple alarm codes and a single common original cache data is assembled; otherwise, the corresponding abnormal event packages are assembled separately.

8. The edge-cloud collaborative compression method for industrial data according to claim 1, characterized in that, The compressed data package containing the stage statistical summary, the boundary key point data, the anomaly event package, and the original hard constraint quantities will be asynchronously uploaded to the cloud server according to risk priority, including: A priority queue containing security anomalies, equipment alarms, boundary key points, and stage statistical summaries is established based on preset data types, and the original hard constraint quantities are associated with the corresponding stage statistical summaries, boundary key point data, or anomaly event packages and set to the same priority. When the gateway memory usage rate is determined to be lower than the preset security threshold, each data packet is sent asynchronously in descending order of priority queue. When it is determined that the gateway memory usage rate has reached a preset security threshold but is lower than a preset critical threshold, the transmission of the stage statistical summary is suspended, and the security anomaly, the device alarm, and the boundary key point data are transmitted first. When the gateway memory usage rate is determined to reach a preset critical threshold, the cache area corresponding to the abnormal event packet will not be used as a priority coverage area before receiving the acceptance confirmation from the cloud server. For consecutive statistical summaries of multiple stages with the same device identifier, stage identifier, recipe identifier, and strategy version identifier, merge them, retrieve the original hard constraint quantities associated with the multiple stage statistical summaries, calculate the overall maximum value, overall minimum value, and cumulative integral value, and calculate the merged mean based on the number of sampling points in each stage to generate a merged statistical summary, and release the cache occupied by the low-priority data that has been uploaded and confirmed.

9. The edge-cloud collaborative compression method for industrial data according to claim 1, characterized in that, Receive the new compression strategy version data reverse-synchronized from the cloud server, including: Receive the new compression strategy version data, which includes the target device identifier, target channel identifier, target stage identifier, new strategy version number, previous strategy version number, effective production cycle identifier, and the new preset error threshold. Parse the effective production cycle identifier; When the current production cycle is detected to have ended and the production cycle identifier of the new production cycle is consistent with the effective production cycle identifier, the new preset error threshold is overwritten and the corresponding parameter item in the compression strategy mapping table is updated according to the target device identifier, target channel identifier, and target stage identifier.

10. The edge-cloud collaborative compression method for industrial data according to claim 9, characterized in that, Updating the compression strategy mapping table based on the new preset error threshold further includes: After the new preset error threshold takes effect, obtain the reconstruction verification data of the corresponding temporal semantic data segment generated and returned by the cloud server based on the compressed data packet; When the cloud server determines, based on the reconstruction verification data, that the reconstruction deviation is greater than the new preset error threshold but not greater than the preset error threshold corresponding to the previous strategy version number within a consecutive preset number of production cycles, it receives the next compression strategy version data generated by the cloud server after further tightening. When the reconstruction deviation is greater than the preset error threshold corresponding to the previous strategy version number for a consecutive preset number of production cycles and the deviation trend shows a continuous increase, a strategy security downgrade instruction issued by the cloud server is received, and the target channel is switched to a preset low-loss compression strategy; the preset low-loss compression strategy includes at least the following parameters: using a low-loss error threshold less than the preset error threshold in the current strategy for boundary key point extraction, retaining all original extreme points, and prohibiting any form of lossy compression on the key security channel data; Once the reconstruction deviation falls below the new preset error threshold within a preset number of consecutive production cycles, the new compression strategy version data will be resumed.

Images