An Industrial Internet Device Data Acquisition and Online Monitoring System

By automatically identifying the state reorganization based on the binary message structure boundary and time identifier, the problem of determinism in device access and diagnostic conclusions in industrial internet data acquisition systems is solved, realizing adaptive device access and the uniqueness and reproducibility of diagnostic conclusions.

CN123123205APending Publication Date: 2026-11-17CHENGDU YIYUN TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611265078.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing industrial internet data acquisition systems cannot automatically identify the private binary protocols of heterogeneous existing devices, resulting in long access cycles and high maintenance costs for new devices; edge-cloud collaborative diagnosis cannot guarantee the certainty and reproducibility of diagnostic conclusions under network latency jitter and data out-of-order conditions.

Method used

The protocol boundary identification module automatically identifies the binary message structure boundary, and the semantic alignment module generates decoding rules. The edge gateway encapsulates the status token and uploads it to the cloud processing platform. The global spatiotemporal state machine module performs state recombination and rollback based on time identification to ensure the uniqueness and reproducibility of the diagnostic conclusions.

Benefits of technology

It enables adaptive access to heterogeneous devices and unified semantic organization, reduces reliance on manual configuration, reduces false alarms and missed alarms, and ensures the uniqueness and reproducibility of diagnostic conclusions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123205A_ABST
    Figure CN123123205A_ABST
Patent Text Reader

Abstract

This invention discloses an industrial internet equipment data acquisition and online monitoring system, belonging to the field of industrial internet equipment operation and maintenance monitoring technology. The system includes an edge gateway and a cloud processing platform. The edge gateway identifies the frame header, checksum, and load segment boundary of unknown protocol messages through a sequence feature extraction model, and uploads the load segment features to the cloud. The cloud combines the equipment mechanism knowledge graph to generate decoding rules and distributes them to achieve adaptive access of dumb devices. The edge gateway encapsulates the anomaly detection results and time stamp information into a status token and uploads it to the cloud. The cloud global spatiotemporal state machine reorganizes the state based on the time stamp information. When the time deviation exceeds the threshold, it performs state rollback and replay to eliminate diagnostic drift and improve the certainty and reproducibility of fault diagnosis conclusions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of industrial internet equipment operation and maintenance monitoring technology, and in particular to an industrial internet equipment data acquisition and online monitoring system. Background Technology

[0002] Against the backdrop of the rapid development of the Industrial Internet and intelligent manufacturing, data acquisition and online monitoring of industrial equipment are the core links for realizing fault prediction, health management and equipment operation and maintenance. There are a large number of sensors, programmable logic controllers, instruments and other equipment in the industrial field. The operating data generated by these devices needs to be continuously collected, aggregated and analyzed in order to detect abnormal equipment status in a timely manner and avoid production interruption or safety accidents caused by equipment failure.

[0003] Existing industrial internet data acquisition and monitoring systems typically employ an infrastructure that combines edge acquisition gateways with cloud monitoring platforms. The edge gateways are deployed in industrial sites to communicate with various field devices, while the cloud monitoring platform aggregates data uploaded from each edge node, runs diagnostic algorithms, and provides a visual monitoring interface to maintenance personnel.

[0004] A similar prior art solution to this invention operates as follows: The edge gateway polls and reads device data through a pre-configured communication protocol library, converts the read data into message queue telemetry transmission format or hypertext transfer protocol format, and uploads it to the cloud. The cloud runs threshold alarm rules or a trained diagnostic model based on the received time-series data to achieve online monitoring of device status. Specifically, this solution first requires engineers to manually establish a mapping relationship between register addresses and physical quantity meanings in the edge gateway's configuration tool according to the device's communication manual, forming a point table. Subsequently, the edge gateway periodically polls the device according to this point table, parses the collected raw messages into structured values ​​according to preset protocol rules, and then encapsulates the parsed structured data into messages and uploads them to the cloud platform. After receiving the data, the cloud platform stores it in the time-series database according to the device number and triggers preset threshold rules or a trained model for status evaluation.

[0005] The aforementioned existing technical solutions have two main shortcomings. First, the protocol parsing process relies heavily on manually pre-configured point tables. If existing devices exist with missing protocol documentation, undisclosed proprietary protocols from manufacturers, or whose message structures have changed due to firmware upgrades—these are known as "dumb devices"—the edge gateway cannot determine the boundary positions of the frame header, checksum, and payload segments in the message, thus failing to complete data parsing. Consequently, these devices cannot be integrated into a unified monitoring system, resulting in long access cycles and high maintenance costs for new devices, hindering plug-and-play interoperability. Second, the solution's design assumes that network transmission and... Point computing is ideally synchronous. Local features obtained from edge computing are directly fused with cloud data without being calibrated with a unified time base. Once millisecond-level delays, jitter, or packet loss and retransmission occur in the field network, the order in which multiple data streams arrive at the cloud will be inconsistent with their actual generation order. The cloud diagnostic model makes inferences based on the out-of-order data, which will lead to contradictory intermediate diagnostic conclusions for the same device at different times. That is, the diagnostic results have time drift, which will cause false alarms or missed alarms. It is difficult to guarantee the determinism and reproducibility of fault detection and recovery judgment in the time dimension under the harsh working conditions of industrial sites.

[0006] It is evident that existing industrial internet equipment data acquisition and online monitoring systems have room for improvement in both adaptive access to heterogeneous existing equipment and the certainty of edge-cloud collaborative diagnostic conclusions. There is an urgent need for a technical solution that can achieve adaptive protocol parsing without relying on manually predefined protocol mapping relationships, while ensuring that the diagnostic conclusions are unique and reproducible even under network latency jitter or data out-of-order conditions. Summary of the Invention

[0007] The technical problems to be solved by this invention include the following two aspects: First, when facing the private binary protocols of massive heterogeneous existing dumb devices, existing industrial internet data acquisition systems must rely on manual configuration of protocol mapping tables point by point, which cannot achieve automatic protocol identification and semantic-level parsing, resulting in long access cycles for new devices and high maintenance costs. Second, under the conditions of network latency jitter, data disorder, or fluctuations in edge node computing power, the local features output by the edge side are directly fused and calculated with the cloud without being calibrated with a unified time reference, which may lead to contradictory diagnostic conclusions for the same device at different times, and cannot guarantee the determinism and reproducibility of fault detection and recovery judgment in the time dimension.

[0008] To address the aforementioned technical problems, this invention provides an industrial internet device data acquisition and online monitoring system, comprising an edge gateway and a cloud processing platform, wherein the edge gateway is connected to the target device via a communication bus.

[0009] The edge gateway includes a protocol boundary identification module and a token encapsulation module. The protocol boundary identification module is used to continuously collect the raw binary message stream on the communication bus, input the raw binary message stream into a pre-trained sequence feature extraction model, classify the byte positions in the message through the sequence feature extraction model, identify the boundaries of the frame header, check bit and payload segment in the message, obtain the payload segment feature vector, and parse the message according to the identified boundaries to obtain structured data. The token encapsulation module is used to perform local anomaly detection on the structured data. When an anomaly is detected, it obtains time stamp information, encapsulates the observation data on which the detection is based, the anomaly, and the time stamp information into a status token, and sends it to the cloud processing platform in the form of an event stream.

[0010] The cloud processing platform includes a semantic alignment module and a global spatiotemporal state machine module. The semantic alignment module receives the load segment feature vector, calculates the similarity between the load segment feature vector and multiple known device feature vectors in a pre-built device mechanism knowledge graph, identifies knowledge graph nodes with similarity exceeding a preset threshold as matching nodes, generates decoding rules based on the physical quantity meaning associated with the matching nodes, and sends the decoding rules to the edge gateway for parsing the packet's load segment. The global spatiotemporal state machine module receives the state token sent by the edge gateway and processes it according to the time identifier carried by the state token. The information maps the state token to the corresponding time window for state reorganization. It calculates the time deviation between the time when the state token arrives at the global spatiotemporal state machine and the time identifier information. When the time deviation exceeds a preset time window threshold, the intermediate diagnostic state calculated within the time window is rolled back to the state before the corresponding time of the state token. The deterministic replay calculation is performed again strictly according to the time sequence of the time identifier information carried by the multiple state tokens within the time window. The replay calculation result is independent of the actual order in which the state tokens arrive at the global spatiotemporal state machine module, and the diagnostic conclusion of the time window is obtained.

[0011] The technical effects achieved by this invention are as follows:

[0012] 1. Since the protocol boundary identification module can automatically identify the structural boundaries of unknown binary protocol messages, and the semantic alignment module can automatically generate decoding rules by combining the device mechanism knowledge graph, there is no need for manual pre-configuration of point tables. Therefore, the access cycle of new devices is significantly shortened, the reliance on protocol documents and manual configuration experience is reduced, and adaptive access and unified semantic organization of heterogeneous dumb device data are realized.

[0013] 2. Because the edge gateway does not directly output deterministic local diagnostic conclusions, but instead encapsulates the observation data, anomaly characteristics, and time stamp information into a status token and uploads it to the cloud, the global spatiotemporal state machine restructures the status token based on the time stamp information rather than the arrival order. When a time deviation exceeds a preset threshold, state rollback and replay are performed. Therefore, regardless of changes in the on-site network conditions, the diagnostic conclusions obtained for the same device and the same time window can remain unique and reproducible. The state restructuring and deterministic replay calculation based on the time stamp rather than the arrival order ensures that the diagnostic conclusions are not affected by the data arrival order, remain unique and reproducible, and reduce the occurrence of false alarms and missed alarms. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall system architecture of the present invention;

[0015] Figure 2 This is a flowchart of the protocol boundary identification and semantic decoding rule generation process of the present invention;

[0016] Figure 3 This is a flowchart of the status token encapsulation and cloud-based deterministic diagnostics process of the present invention. Detailed Implementation

[0017] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0018] System Overall Architecture

[0019] The industrial internet device data acquisition and online monitoring system provided in this embodiment includes an edge gateway and a cloud processing platform. The edge gateway is deployed in the industrial field and connects to target devices via a bus bypass method. The target devices include standard devices with publicly available communication protocols and existing "dumb" devices without publicly available protocol documentation. The edge gateway internally includes a protocol boundary identification module and a token encapsulation module. The cloud processing platform internally includes a semantic alignment module and a global spatiotemporal state machine module. The cloud processing platform is also connected to an operation and maintenance monitoring visualization platform to present deterministic diagnostic conclusions to operation and maintenance personnel. The edge gateway and the cloud processing platform interact bidirectionally through an event flow channel. On one hand, the edge gateway continuously uploads load segment feature vectors and status tokens to the cloud processing platform; on the other hand, the cloud processing platform distributes the generated decoding rules to the edge gateway.

[0020] Specific implementation methods for protocol boundary identification and semantic decoding rule generation

[0021] After the edge gateway bypasses the communication bus of the target device, the protocol boundary identification module continuously samples the raw binary byte stream transmitted on the bus with a fixed-length sampling window to form the raw message sequence to be processed and treat it as time series data.

[0022] The original message sequence is first fed into a one-dimensional convolutional layer to extract pattern features between local bytes of the message, resulting in a local feature sequence. This convolution operation is represented as:

[0023] ;

[0024] in, Represents the first byte in the original byte sequence. The value at each byte position. This indicates the position of the one-dimensional convolution kernel. The weighting coefficient at the location, Indicates the kernel length. Indicates the bias term. Represents a non-linear activation function. The first convolution output represents the... Local eigenvalues.

[0025] The local feature sequence output from the one-dimensional convolutional layer is further input into a gated recurrent unit (ROU) to model long-range dependencies between byte sequences, thus providing a basis for identifying the boundaries between the frame header, checksum, and payload. The update process of the gated ROU is represented as follows:

[0026] ;

[0027] ;

[0028] ;

[0029] ;

[0030] in, This represents the convolutional features input at the current time step. This indicates the hidden state at the previous moment. This indicates the update gate output, used to control the proportion of historical states retained. This indicates the reset gate output, used to control the degree of influence of historical states on candidate states. Indicates the candidate hidden state. This indicates the hidden state at the current moment. , , These represent the weight matrices corresponding to the update gate, reset gate, and candidate states, respectively. , , These represent the corresponding bias terms. This represents a vector concatenation operation. This indicates element-wise multiplication.

[0031] The timing feature representation output by the gated loop unit is fed into a preset boundary discrimination output layer. This output layer classifies each byte position, determining whether it belongs to the frame header, checksum, payload segment, or frame trailer, thereby determining the structural boundary division of the current message. The classified payload segment bytes are extracted as payload segment feature vectors and uploaded to the cloud processing platform via the network interface. The protocol boundary identification module parses the message based on the boundary discrimination results, obtaining structured data containing timestamps, byte positions, and their categories, which are then used by the token encapsulation module for subsequent processing.

[0032] The semantic alignment module of the cloud processing platform pre-constructs a knowledge graph containing the communication mechanisms of various typical industrial devices. Each node in the graph corresponds to a feature vector of a known device, along with the meaning, unit, and data type of the associated physical quantity. After receiving the load segment feature vector, the semantic alignment module calculates the inner product of this feature vector and the feature vector of each known device in the knowledge graph. This inner product is then divided by the product of the magnitudes of the two vectors to obtain their similarity. This similarity is calculated as follows:

[0033] ;

[0034] in, This represents the feature vector of the load segment uploaded by the edge gateway. Represents the first in a knowledge graph A known device feature vector, This represents the dot product of two vectors. and Let these represent the magnitudes of the two vectors, respectively. This indicates the similarity between the two; the closer the value is to 1, the higher the degree of semantic matching.

[0035] The semantic alignment module identifies knowledge graph nodes with similarity exceeding a preset threshold as matching nodes. Combining the physical quantity meaning, unit, and data type information associated with these matching nodes, it generates decoding rules under a unified namespace. These decoding rules are then distributed to the edge gateway for the protocol boundary identification module to perform real-time semantic parsing of the payload segments of subsequent similar messages. The parsed structured data is uploaded to the cloud for storage and display in a unified namespace format. If multiple calculated similarities do not exceed the preset threshold, the semantic alignment module marks the payload segment feature vector as a sample to be confirmed. It receives manually labeled information for this sample and incorporates it into the device mechanism knowledge graph based on this labeling information for continuous optimization of subsequent similarity matching.

[0036] Specific Implementation Methods of State Token Encapsulation and Cloud-based Deterministic Diagnostics

[0037] The token encapsulation module performs local anomaly detection on the structured data parsed by the protocol boundary identification module to determine whether there are potential anomalies. When a potential anomaly is detected, the token encapsulation module reads the local clock of the edge gateway to obtain a nanosecond-level timestamp, and encapsulates the observation data, local anomalies, and the nanosecond-level timestamp together into a status token. This token is then continuously pushed to the event stream receiving queue of the cloud processing platform in the form of an event stream without waiting or buffering, ensuring that the status token is uploaded in real time according to the method of being generated and sent immediately.

[0038] The global spatiotemporal state machine module sequentially retrieves state tokens from the event stream receiving queue, reads the timestamps carried within each token, and calculates the absolute value of the difference between the actual time the state token arrives at the global spatiotemporal state machine module and its own timestamp. This time deviation is expressed as:

[0039] ;

[0040] in, Indicates the first The actual time when a state token arrives at the global spacetime state machine module. Indicates the first Each state token carries its own nanosecond-level timestamp. This indicates the time deviation between the two.

[0041] The global spatiotemporal state machine module determines whether the time deviation exceeds a preset time window threshold. If it does not, the status token is merged into the state calculation queue of its time window according to its timestamp order, and participates in the diagnostic state fusion calculation of that time window. If it exceeds the threshold, it is determined that network latency or out-of-order delivery has occurred. The global spatiotemporal state machine module locates the affected time window to which the timestamp of the status token belongs, rolls back the intermediate diagnostic state currently calculated for that time window to the historical state before the corresponding time of the status token, and then, based on the complete timestamp order of all received status tokens within the time window, performs a deterministic replay calculation of the diagnostic state of that time window to obtain the diagnostic intermediate state of that time window at that moment. When the duration of a time window exceeds its window termination boundary and no new status tokens belonging to that time window are received, the global spatiotemporal state machine module outputs a diagnostic conclusion for the corresponding device within that time window based on the replay of the state within that time window. The diagnostic conclusion includes a fault determination result and a recovery determination result, and pushes the diagnostic conclusion to the operation and maintenance monitoring visualization platform.

[0042] Through the aforementioned state rollback and replay mechanism, regardless of whether the order in which the state tokens arrive at the cloud processing platform is consistent with their actual generation order, the global spatiotemporal state machine module will ultimately complete the state calculation according to the timestamp order carried by the state tokens themselves. This ensures that the diagnostic conclusions obtained for the same device and the same time window are unique, and that the same results can be obtained by replaying the event stream when needed, thus possessing reproducibility.

[0043] Alternative implementation methods

[0044] In an alternative implementation, the sequence feature extraction model in the protocol boundary identification module can use a temporal convolutional network instead of the cascaded structure of one-dimensional convolutional layers and gated recurrent units, or it can use a network structure containing a self-attention mechanism. By modeling the long-range dependencies between byte sequences, it can also achieve automatic partitioning of frame headers, check bits, and payload segments, achieving a protocol adaptive parsing effect comparable to the aforementioned embodiments.

[0045] In another alternative implementation, the time identification information obtained by the edge gateway is not limited to the local nanosecond-level physical timestamp. A vector clock mechanism can also be used to mark the causal order of the state tokens generated by each edge gateway. The global spatiotemporal state machine module sorts the state tokens according to the causal order identifiers carried by multiple state tokens and performs state reorganization accordingly. In field environments where it is difficult to achieve high-precision physical time synchronization, the same technical effect of eliminating the disordered drift of diagnostic results and ensuring deterministic diagnosis can be achieved.

[0046] Application Examples

[0047] Taking an existing frequency converter in an industrial site that lacks publicly available protocol documentation as an example, this device is connected to an edge gateway via an RS485 bus. After bypassing the bus, the edge gateway continuously collects its communication messages. The protocol boundary identification module processes these messages to obtain a load segment feature vector, which is then uploaded to a cloud processing platform. The semantic alignment module calculates the similarity between this load segment feature vector and various frequency converter feature vectors already included in the knowledge graph. After determining the matching nodes, it generates decoding rules containing physical quantities such as speed and current, and sends them to the edge gateway. Based on this, the edge gateway completes real-time parsing of subsequent messages from the frequency converter, eliminating the need for engineers to pre-configure point tables. During the operation of the frequency converter, the edge gateway performs local anomaly detection on the parsed structured data. When an abnormal current fluctuation is detected, a status token is generated and uploaded to the cloud. The global spatiotemporal state machine module reassembles the status tokens reported by multiple devices in a unified time window based on the timestamp carried by the status token. When a brief delay occurs in the field network, causing the arrival order of the status tokens to be inconsistent with the actual generation order, state rollback and replay are performed. Finally, a deterministic diagnostic conclusion about the frequency converter within the corresponding time window is obtained and pushed to the operation and maintenance monitoring visualization platform for operation and maintenance personnel to view.

[0048] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. An industrial internet device data acquisition and online monitoring system, comprising an edge gateway and a cloud processing platform, wherein the edge gateway is connected to the target device via a communication bus, characterized in that: The edge gateway includes a protocol boundary identification module and a token encapsulation module; The protocol boundary identification module is used to continuously collect the original binary message stream on the communication bus, input the original binary message stream into a pre-trained sequence feature extraction model, classify the byte positions in the message through the sequence feature extraction model, identify the boundaries of the frame header, check bit and payload segment in the message, obtain the payload segment feature vector, and parse the message according to the identified boundaries to obtain structured data; The token encapsulation module is used to perform local anomaly detection on the structured data. When an anomaly is detected, it obtains time stamp information, encapsulates the observation data on which the detection is based, the anomaly, and the time stamp information into a status token, and sends it to the cloud processing platform in the form of an event stream. The cloud processing platform includes a semantic alignment module and a global spatiotemporal state machine module; The semantic alignment module is used to receive the load segment feature vector, calculate the similarity between the load segment feature vector and multiple known device feature vectors in the pre-built device mechanism knowledge graph, determine the knowledge graph nodes with similarity exceeding a preset threshold as matching nodes, generate decoding rules according to the physical quantity meaning associated with the matching nodes, and send the decoding rules to the edge gateway for the edge gateway to parse the load segment of the message; The global spatiotemporal state machine module is used to receive the state token sent by the edge gateway, map the state token to the corresponding time window according to the time identifier information carried by the state token for state reorganization, calculate the time deviation between the time when the state token arrives at the global spatiotemporal state machine and the time identifier information, and when the time deviation exceeds a preset time window threshold, roll back the intermediate diagnostic state calculated within the time window to the state before the corresponding time of the state token, and re-perform deterministic replay calculation strictly according to the time sequence of the time identifier information carried by the multiple state tokens within the time window. The replay calculation result is independent of the actual order in which the state tokens arrive at the global spatiotemporal state machine module, and the diagnostic conclusion of the time window is obtained.

2. The system as described in claim 1, characterized in that, The sequence feature extraction model includes a one-dimensional convolutional layer, a gated recurrent unit, and a boundary discrimination output layer connected in sequence. The one-dimensional convolutional layer is used to perform convolution operations on the original binary message stream to extract pattern features between local bytes of the message and obtain a local feature sequence. The gated recurrent unit is used to perform temporal modeling on the local feature sequence to obtain a temporal feature representation that incorporates long-range dependencies. The boundary discrimination output layer is used to classify each byte position according to the temporal feature representation, dividing the byte position into four categories: frame header, check bit, payload segment, and frame tail, thereby determining the structural boundary division of the message.

3. The system as described in claim 1, characterized in that, The semantic alignment module calculates the inner product of the load segment feature vector and the known device feature vector, and divides the inner product by the product of the magnitude of the load segment feature vector and the magnitude of the known device feature vector to obtain the similarity. The decoding rules include the physical quantity meaning, unit and data type information associated with the matching node.

4. The system as described in claim 3, characterized in that, The semantic alignment module is also used to mark the load segment feature vector as a sample to be confirmed when the calculated similarity does not exceed a preset threshold, receive manual annotation information for the sample to be confirmed, and incorporate the sample to be confirmed into the device mechanism knowledge graph according to the manual annotation information.

5. The system as described in claim 1, characterized in that, The time identifier information is a nanosecond-level timestamp collected by the edge gateway through the local clock. The global spatiotemporal state machine module obtains the time deviation by calculating the absolute value of the difference between the time when the state token arrives at the global spatiotemporal state machine and the nanosecond-level timestamp.

6. The system as described in claim 5, characterized in that, The global spatiotemporal state machine module is also used to output a diagnostic conclusion for the corresponding device within the time window when the current system time exceeds the termination boundary of the time window and no longer receives new state tokens belonging to the time window, based on the state calculated by replay within the time window. The diagnostic conclusion includes fault determination results and recovery determination results, and the diagnostic conclusion is sent to the operation and maintenance monitoring visualization platform.

7. The system as described in claim 1, characterized in that, The time identifier information is a causal order identifier generated by the edge gateway according to the vector clock mechanism. The global spatiotemporal state machine module sorts the multiple state tokens according to the causal order identifier carried by the multiple state tokens, and performs state reorganization according to the sorting result.

8. A method for data acquisition and online monitoring of industrial internet devices, applied to a system including an edge gateway and a cloud processing platform, characterized in that, include: The edge gateway continuously collects the raw binary message stream on the communication bus connected to the target device, inputs the raw binary message stream into a pre-trained sequence feature extraction model, classifies the byte positions in the message through the sequence feature extraction model, identifies the boundaries of the frame header, check bit and payload segment in the message, obtains the payload segment feature vector, and parses the message according to the identified boundaries to obtain structured data; The edge gateway performs local anomaly detection on the structured data. When an anomaly is detected, it obtains time stamp information, encapsulates the observation data on which the detection is based, the anomaly, and the time stamp information into a status token, and sends it to the cloud processing platform in the form of an event stream. The cloud processing platform receives the load segment feature vector, calculates the similarity between the load segment feature vector and multiple known device feature vectors in the pre-built device mechanism knowledge graph, determines knowledge graph nodes with similarity exceeding a preset threshold as matching nodes, generates decoding rules based on the physical quantity meaning associated with the matching nodes, and sends the decoding rules to the edge gateway for the edge gateway to parse the load segment of the message; The cloud processing platform receives the status token sent by the edge gateway, maps the status token to the corresponding time window according to the time identifier information carried by the status token, performs state reorganization, calculates the time deviation between the arrival time of the status token and the time identifier information, and when the time deviation exceeds a preset time window threshold, rolls back the intermediate diagnostic state calculated within the time window to the state before the corresponding time of the status token, and performs deterministic replay calculation strictly according to the time sequence of the time identifier information carried by the multiple status tokens within the time window. The replay calculation result is independent of the actual order in which the status tokens arrive at the global spatiotemporal state machine module, and obtains the diagnostic conclusion of the time window.

9. The method as described in claim 8, characterized in that, The sequence feature extraction model includes a one-dimensional convolutional layer, a gated recurrent unit, and a boundary discrimination output layer connected in sequence. The one-dimensional convolutional layer extracts pattern features between local bytes of the message to obtain a local feature sequence. The gated recurrent unit performs temporal modeling on the local feature sequence to obtain a temporal feature representation. The boundary discrimination output layer divides the byte positions into four categories: frame header, check bit, payload segment, and frame tail according to the temporal feature representation. The cloud processing platform calculates the inner product of the payload segment feature vector and the known device feature vector, and divides the inner product by the product of the magnitudes of the two vectors to obtain the similarity.

10. The method as described in claim 8, characterized in that, The time identifier information is a nanosecond-level timestamp collected by the edge gateway through the local clock. When the current system time exceeds the termination boundary of the time window and no longer receives new status tokens belonging to the time window, the cloud processing platform outputs a diagnostic conclusion including fault determination result and recovery determination result based on the status calculated by replay within the time window, and sends the diagnostic conclusion to the operation and maintenance monitoring visualization platform.