Industrial equipment interconnection method based on industrial control platform

Abstract

The present invention discloses an industrial equipment interconnection and intercommunication method based on an industrial control platform, which relates to the technical field of industrial equipment interconnection and intercommunication, and specifically includes the following steps: for a reissued data frame with repeated content, calculating the behavioral entanglement between it and the reissued data frame; if the entanglement exceeds a preset threshold, extracting the sequence offset, time delay, and field fluctuation amplitude to generate an indicator vector for describing the timeliness of the data identification; inputting the generated indicator vector into a pre-trained graph embedding algorithm model to generate a first parameter and a second parameter, the first parameter being used to describe the consistency of the data frame with the historical evolution path, and the second parameter being used to characterize the timeliness of the data frame identification. The present invention solves the problem that reissued data of industrial equipment cannot be accurately identified, and based on a causal graph and a graph embedding model, realizes the timeliness judgment and validity identification of status data.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment interconnection and intercommunication, and in particular to an industrial equipment interconnection and intercommunication method based on an industrial control platform. Background Art

[0002] In modern industrial production systems, with the diversification and intelligent development of manufacturing equipment, industrial sites often deploy industrial equipment from different manufacturers and with varying communication protocols and interface standards. To achieve centralized management and efficient collaboration of these devices, interoperability among industrial equipment has become a key component of industrial automation and digitalization. Industrial equipment interoperability refers to the bidirectional transmission and compatible communication of information, such as data collection, command control, and status feedback, between devices. This enables various types of equipment to operate collaboratively within a unified system, supporting data sharing, unified monitoring, and centralized control. In this context, industrial control platforms, as information aggregation and command hubs, possess powerful computing, communication management, and business logic processing capabilities, capable of handling the demands of large-scale device access and high-frequency data flow. Building an industrial equipment interoperability system based on an industrial control platform not only enables unified access and integrated management of diverse devices, but also supports real-time monitoring, intelligent analysis, and collaborative control through the platform's remote service capabilities. This provides industrial enterprises with highly centralized and intelligent production operations and promotes digital and efficient production management.

[0003] Existing industrial equipment interconnection technologies based on industrial control platforms usually achieve unified access and centralized management of various industrial equipment by integrating functional units such as multi-protocol communication modules, data acquisition modules, and equipment management modules into the industrial control platform. Specifically, at the device access layer, the platform connects devices of different brands and with different communication interfaces by supporting mainstream industrial communication protocols such as Modbus, OPC UA, PROFINET, and EtherCAT, or by deploying protocol adapters and data gateways. Subsequently, during the data parsing and mapping phase, the platform uniformly formats and semantically maps the collected raw data to ensure compatible data structures across different devices. Next, at the data management and service scheduling layer, the industrial control platform monitors device status in real time, stores historical data, and analyzes abnormal alarms based on pre-set data models and control logic. The platform then provides the results to upper-level application systems or operations and maintenance personnel through a unified interface. Furthermore, during remote control and collaborative execution, the platform can issue control commands to each connected device, enabling cross-device linkage and automated control strategies. Finally, at the security layer, existing systems generally ensure the security and integrity of industrial data during transmission and processing through access control, identity authentication, communication encryption, and data isolation. This results in an industrial device interconnection system centered on the industrial control platform, encompassing multiple links including device access, data parsing, control scheduling, and security protection.

[0004] The existing technology has the following deficiencies:

[0005] When industrial control platforms interconnect industrial devices, some devices equipped with data retransmission mechanisms automatically retransmit status packets from the previous cycle when the communication network experiences momentary packet loss or delay. While the content of this data remains consistent with the original data, the timestamp or sequence number may have been slightly altered. Due to the opaque nature of the device's transmission behavior, the platform cannot directly identify whether this data is "retransmitted" rather than "real-time." Therefore, provided the field content and format are valid, the platform defaults to treating it as the latest status data and adopts it accordingly. Because the platform's existing data processing methods fail to incorporate mechanisms to identify and verify the timeliness of the data source and the context of the data behavior (such as retransmission behavior), it is unable to determine whether the status field should be considered as the current device's valid operating status based on the timeliness of the data identifier when the data retransmission occurs and the content is duplicated, resulting in misjudgment of the device's true status. This can lead to incorrect control decisions based on the platform's erroneous device status, such as repeatedly triggering status exception handling, incorrectly interrupting current tasks, or delaying the issuance of critical commands. This can lead to serious consequences such as disrupted production rhythms, distorted scheduling logic, and loss of synchronization in the device control chain.

[0006] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0007] The purpose of the present invention is to provide an industrial equipment interconnection method based on an industrial control platform to solve the problems in the above-mentioned background technology.

[0008] In order to achieve the above objectives, the present invention provides the following technical solution: an industrial equipment interconnection method based on an industrial control platform, specifically comprising the following steps:

[0009] Construct a causal graph between industrial equipment status fields and control instructions to record the causal relationship between field combinations and task evolution paths;

[0010] Receive data frames from industrial equipment and determine whether the data frame is consistent with the historical frame field content and the timestamp or sequence number is different. If so, mark it as a retransmitted data frame with duplicate content;

[0011] For resent data frames with duplicate content, the behavioral entanglement between them and the resent data frame is calculated. If the entanglement exceeds a preset threshold, the sequence offset, time delay, and field fluctuation amplitude are extracted to generate an indicator vector describing the timeliness of the data identification.

[0012] The generated indicator vector is input into the pre-trained graph embedding algorithm model to generate the first parameter and the second parameter. The first parameter is used to describe the consistency of the data frame with the historical evolution path, and the second parameter is used to characterize the timeliness of the data frame identification.

[0013] Based on a joint determination of the first parameter and the second parameter, determining whether the status field corresponding to the data frame should be used as a valid operating status of the current device;

[0014] The device status snapshot is updated based on the judgment results, and the judgment results are fed back to the causal graph path to dynamically adjust the path confidence weight and state identification strategy.

[0015] Preferably, a causal graph between the industrial equipment status field and the control instruction is constructed, specifically:

[0016] By collecting historical operating data of industrial equipment during task execution, the field groups and corresponding control instructions contained in each status data frame are extracted, and a triple set consisting of field groups, control instructions and their occurrence sequence is constructed;

[0017] The triple set is represented as a directed edge relationship in a graph structure, with the field group as the starting node and the control instruction as the ending node. At the same time, time weight and frequency weight are assigned to the edges according to the task evolution timeline to form a directed causal graph that can be used for state evolution path reasoning.

[0018] Preferably, a data frame from an industrial device is received, and it is determined whether the data frame is consistent with the historical frame field content and the timestamp or sequence number is different. If so, it is marked as a retransmitted data frame with duplicate content, specifically:

[0019] Receive data frames from industrial equipment, extract multiple key status fields from the data frames, arrange them in the order of field collection and the industrial equipment data structure template, and construct a field sequence feature vector to represent the complete equipment status corresponding to the current data frame;

[0020] Perform hash coding on the field sequence feature vector and generate a field content signature value by combining the weighted field value and the field position index. This signature value is used to uniquely identify a group of status field combinations within the same device dimension.

[0021] Retrieve historical data frames with the same field content signature value in the historical data buffer, and select the one with the smallest time interval with the current data frame as the comparison frame;

[0022] Determine whether the timestamps or sequence numbers of the current data frame and the comparison frame are non-increasing, reversed, or discontinuous. If so, mark the current data frame as a candidate for re-sending data.

[0023] Calculate the time offset between the reception time of the current data frame and the recording time of the comparison frame. If the offset is less than the preset task cycle tolerance threshold and the field content of the current data frame remains completely consistent with the comparison frame, then determine that the current data frame is a retransmitted data frame with duplicate content.

[0024] Preferably, for a reissued data frame with repeated content, the behavioral entanglement between it and the reissued data frame is calculated. If the entanglement exceeds a preset threshold, the sequence offset, time delay, and field fluctuation amplitude are extracted to generate an indicator vector for describing the timeliness of the data identification. Specifically, the following steps are included:

[0025] Collect a preset number of field sequence vectors of the reissued data frame with repeated content and the data frame being reissued within the time window before and after, arrange them in chronological order as a field value time series, and organize the field sequence vectors at each moment into a vector list in the sampling order to form a vector sequence trajectory representing the field evolution trend, where each trajectory is used to reflect the temporal change process of the same field combination in the state space before and after the reissue behavior;

[0026] The behavioral entanglement between two vector sequence trajectories is calculated based on the cosine similarity and dynamic time warping algorithms. Cosine similarity is used to measure the consistency of local field changes, and dynamic time warping is used to compare the matching degree of the overall temporal structure.

[0027] If the behavior entanglement exceeds the preset threshold, the time delay between the timestamp of the resent data frame and the timestamp of the resent data frame is calculated, the difference between the sequence numbers is used as the sequence offset, and the differential mean of the corresponding values ​​of the fields is used as the field fluctuation amplitude;

[0028] The time delay, sequence offset and field fluctuation amplitude are normalized respectively and combined into an indicator vector with a unified dimension according to a preset order, which is used to characterize the identification timeliness of the reissued data frame.

[0029] Preferably, the behavioral entanglement between two vector sequence trajectories is calculated based on the cosine similarity and the dynamic time warping algorithm. The specific calculation method is as follows:

[0030] The field sequence vectors extracted from the retransmitted data frame with repeated content and the retransmitted data frame within the preset time window before and after each are arranged into two vector sequence tracks in sampling order. The field vectors in each track are normalized according to the maximum and minimum values ​​of the field, and the field values ​​are mapped to a unified interval.

[0031] Based on the two normalized vector sequence trajectories, a dynamic time warping algorithm is used to calculate the Euclidean distance between each pair of field vectors in the two trajectories. A one-to-one matching path is constructed based on the minimum cumulative cost, and the sampling time position of each pair of vectors in the path in their respective trajectories is recorded.

[0032] For all paired field vectors in the matching path, calculate the cosine similarity, form a similarity value sequence in the order of the paths, and calculate the arithmetic mean of the value sequence as the first value to measure the consistency of the two trajectories in the direction of local field change;

[0033] The sampling time position differences of each field vector pair in the matching path are averaged to obtain the degree of offset of the two trajectories in the overall time structure as the second value. The first and second values ​​are normalized respectively and then linearly combined according to the preset weighting ratio to obtain the behavioral entanglement degree used to characterize the degree of consistency of the overall changes of the two vector trajectories.

[0034] Preferably, the generated indicator vector is input into a pre-trained graph embedding algorithm model to generate the first parameter and the second parameter, specifically:

[0035] The generated indicator vector is used as the feature vector of the current data frame. The graph node corresponding to the field combination matching the indicator vector is located in the causal graph between the field combination and the control instruction. The adjacent nodes and edge weight information of the node in the graph structure are collected to construct the structural adjacency matrix and the node feature matrix.

[0036] The structural adjacency matrix and the node feature matrix containing the indicator vector are input into a pre-trained graph embedding algorithm model. The model outputs an embedding vector representing the temporal semantics of the graph node based on the integration of the field combination context structure and the indicator dynamic features.

[0037] Extracting several dimensional components for measuring the consistency between the current data frame and the historical field evolution trajectory from the embedding vector, calculating the mean of the dimensional components, and obtaining a first numerical parameter for characterizing the field evolution consistency;

[0038] Several dimensional components reflecting the field disturbance amplitude and time offset performance in the current data frame and the historical control path are extracted from the embedded vector, and the weighted average of the components is calculated to obtain a second numerical parameter for characterizing the state delay and disturbance amplitude.

[0039] Preferably, the pre-trained graph embedding algorithm model is specifically:

[0040] Based on the graph structure consisting of the causal paths between field combinations and control instructions and its historical state data, training samples are constructed. The historical indicator vector corresponding to each field combination node is used as input, and the state judgment results of this node in historical tasks are used as training targets. Through training, the graph embedding algorithm model can generate embedding vectors based on the graph structure and indicator characteristics.

[0041] After training, the model takes as input the structural adjacency matrix and the indicator vector and outputs an embedding vector, which contains the numerical dimensions used to generate the first and second parameters, respectively used to reflect the consistency of the data frame with the historical state path and the timeliness of its identification state.

[0042] Preferably, based on the joint judgment of the first parameter and the second parameter, it is judged whether the status field corresponding to the data frame should be used as the valid operating status of the current device, specifically:

[0043] Construct a two-dimensional parameter judgment space, using the first parameter as the horizontal coordinate and the second parameter as the vertical coordinate. Scatter distribution annotation is performed on the known valid and invalid state data frames during the historical task execution process, and the valid state judgment boundary is preset based on the distribution results.

[0044] receiving a first parameter and a second parameter corresponding to a current data frame, and calculating a coordinate point corresponding to the data frame in a two-dimensional parameter determination space;

[0045] Determine whether the coordinate point is within the preset valid state determination boundary. If it is within the boundary, the state field corresponding to the data frame is determined to be the valid operating state of the current device. If it is outside the boundary, it is determined to be an invalid state.

[0046] The preset effective state judgment boundary is constructed based on the distribution characteristics of historical data using a fixed threshold fitting method, and the threshold used is derived from the statistical analysis of historical state data.

[0047] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0048] 1. This invention constructs a causal graph between industrial equipment status fields and control instructions, and combines it with historical task evolution paths to achieve structured modeling of the causal relationship between status field combinations and control logic. This overcomes the technical bottleneck of traditional data processing methods, which cannot identify the behavioral context of reissued data. In particular, the state tracking mechanism established by combining field content signatures with time information enables the platform to actively identify reissued data frames with "consistent content but delayed time." This effectively solves the problem of device control mistriggering caused by the platform misjudging reissued data as real-time, improving the accuracy of information exchange between devices and the stability of system control decisions.

[0049] 2. The present invention introduces a new indicator called "behavioral entanglement". It uses the joint calculation method of vector trajectory, dynamic time warping and cosine similarity to accurately model the differences between the reissued data and the original data in the field evolution trend and time series structure, and then quantify the "timeliness" of the status data. The indicator vector generated on this basis is used as the input of the embedding model to achieve high-precision characterization of the data frame in the two dimensions of "field consistency" and "identification freshness". The trained graph embedding algorithm model is further mapped into semantic parameters, so that when facing complex data streams, the platform can jointly judge the reliability of the status based on the historical evolution context and indicator characteristics, which greatly improves the intelligent level of data validity identification.

[0050] 3. The present invention constructs a dynamic graph structure control mechanism with self-evolution capabilities by feeding back the judgment results to the causal graph path in real time and adjusting the path confidence weight and state recognition strategy accordingly. This mechanism not only continuously optimizes the platform's strategy for distinguishing device states, but also gradually learns the behavior patterns and communication rules between different devices during task iteration, enabling the platform to be highly adaptable to complex situations such as sudden retransmissions, communication disturbances, and data retransmissions during actual deployment. Therefore, the present invention has significant engineering applicability and system robustness, and can improve the security, real-time performance, and data credibility of industrial control systems in a multi-source heterogeneous device environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction to the drawings required for use in the embodiments will be given below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0052] Figure 1 The figure is a flow chart of the industrial equipment interconnection method based on the industrial control platform of the present invention. DETAILED DESCRIPTION

[0053] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.

[0054] The present invention provides Figure 1 The industrial equipment interconnection method based on the industrial control platform shown includes the following steps:

[0055] Construct a causal graph between industrial equipment status fields and control instructions to record the causal relationship between field combinations and task evolution paths;

[0056] In this embodiment, a causal graph between the industrial equipment status field and the control instructions is constructed, specifically:

[0057] By collecting historical operating data of industrial equipment during task execution, the field groups and corresponding control instructions contained in each status data frame are extracted, and a triple set consisting of field groups, control instructions and their occurrence sequence is constructed;

[0058] Field groups and control instructions can be extracted by parsing the raw data frames in industrial equipment operation logs and communication records. First, through the task scheduling record and data acquisition service interface in the industrial control platform, a continuous sequence of data frames from the equipment within each task cycle is obtained. Each data frame is field-decomposed, extracting key status fields such as temperature, pressure, speed, and alarm codes. These fields are then aggregated into field groups according to time windows. Simultaneously, the control instructions corresponding to these field groups are synchronously retrieved from the control layer records on the same timeline, and the instruction trigger points are identified by the state differences before and after the field changes. Next, the order of the field groups and control instructions is sorted by timestamp, constructing a set of field group-control instruction-sequence triples. This process can be modeled in software using a rule engine, a field difference mapping function, and a task timeline mapper. Ultimately, the data is stored as a graph data structure, which serves as the input for constructing nodes and edges in the causal graph.

[0059] The reason for structuring the correspondence between status field groups and control instructions into triples and expressing them in the form of a causal graph is that in complex industrial scenarios, device behavior is not triggered by a single field change, but rather by the combination of multiple fields in a specific state, triggering a certain control operation. Relying solely on traditional static thresholds to determine this relationship would make it difficult to accurately reflect the true causal chain between state and control. However, by extracting the evolutionary path between field groups and control instructions along the temporal evolution dimension during task execution, not only can the control behavior context be accurately modeled, but it also provides a contextual basis for subsequent anomaly identification and status judgment. In particular, in complex scenarios such as data retransmission and status rollback, the causal graph can effectively determine whether the current state represents a reasonable progression along the path, significantly enhancing the accuracy of judgments on state validity. Therefore, triple extraction is not only a key step in data preprocessing but also a prerequisite for subsequent graph embedding modeling and behavioral partial order analysis.

[0060] The triple set is represented as a directed edge relationship in a graph structure, with the field group as the starting node and the control instruction as the ending node. At the same time, time weight and frequency weight are assigned to the edges according to the task evolution timeline to form a directed causal graph that can be used for state evolution path reasoning.

[0061] By establishing a graph structure modeling module, the triples extracted in the previous step can be mapped into nodes and directed edges in a graph data structure. In the specific implementation, each field group is first used as a starting node and uniquely encoded, for example, by generating a field combination identifier using a hash algorithm. The control instruction, serving as the ending node, is also structured and uniquely identified. Next, directed edges are created in the direction of "field group → control instruction" to express the directional causal logic of state-induced behavior. Subsequently, based on the historical timeline of task execution, two attributes are assigned to each pair of edges between the starting and ending nodes: a time weight, obtained by counting the average time interval between the appearance of the triple within the task cycle, which expresses the evolutionary speed between state and instruction; and a frequency weight, which measures the strength of the causal relationship by calculating the proportion of times the field group triggers the same control instruction within a task. The entire graph construction process can be modeled, stored, and queried using a graph database (such as Neo4j) or a graph computing framework (such as NetworkX), enabling visualization, structuring, and reasoning about the evolving state and behavior of devices.

[0062] Mapping triple sets into directed causal graphs and assigning weights to them systematically captures the underlying logic of how state combinations evolve into control decisions in industrial equipment. Compared to traditional data tables or rule lists, graph structures offer stronger structural representation and path computation capabilities, making them particularly suitable for describing causal networks. In industrial control scenarios, device behavior is often sequential, overlapping, and driven by multiple sources. Simply judging by field values ​​can easily lose context. By modeling state field combinations and control instructions as nodes in a graph and representing their evolutionary logic with directional edges, this not only clearly expresses the control behavior that certain state field combinations tend to trigger, but also leverages path analysis algorithms to match and predict subsequent state input data. In particular, in exceptional scenarios such as data retransmission and behavior rollback, the directed graph can determine whether the current state is at a reasonable point in the evolutionary path, providing a causal basis for subsequent decisions about whether the state should be adopted as a valid state. This graph structure eliminates the reliance on a single data frame for device state judgment, relying instead on logical reasoning based on historical experience. This significantly enhances the platform's adaptive judgment capabilities in complex data dynamics.

[0063] Receive data frames from industrial equipment and determine whether the data frame is consistent with the historical frame field content and the timestamp or sequence number is different. If so, mark it as a retransmitted data frame with duplicate content;

[0064] In this embodiment, a data frame is received from an industrial device, and it is determined whether the data frame is consistent with the historical frame field content and the timestamp or sequence number is different. If the conditions are met, it is marked as a retransmitted data frame with duplicate content. Specifically:

[0065] Receive data frames from industrial equipment, extract multiple key status fields from the data frames, arrange them in the order of field collection and the industrial equipment data structure template, and construct a field sequence feature vector to represent the complete equipment status corresponding to the current data frame;

[0066] After receiving data frames from industrial equipment, the industrial control platform can use predefined field parsing templates to perform field-level structural analysis on each frame, extracting multiple representative key status fields. These fields typically include indicator variables that directly impact the control logic during equipment operation, such as operating temperature, spindle speed, system pressure, feed current, displacement position, alarm flags, operating status codes, and control mode codes. These fields are referred to as key status fields because they not only reflect the current operating conditions of the equipment but also directly determine whether the control platform triggers a response action. To ensure uniform structure and comparability in subsequent processing, each frame of data must be arranged according to the field collection sequence specified in the equipment manufacturer's agreement or unified communication standard. Fields must be aligned, completed, and type-checked using the equipment data structure template to ensure semantic integrity and consistency. When constructing the field sequence feature vector, the current values ​​of each field are combined into a multidimensional vector in a predetermined order. Numeric fields are normalized as necessary, and enumerated fields are discretized to ensure a standard numerical representation in subsequent processing. The field sequence feature vector constructed in this way not only retains the full picture of the device status but also possesses consistency, discriminability, and machine processability, facilitating subsequent operations such as data consistency judgment, behavior recognition, and graph structure mapping. Therefore, this step converts the raw data frame into a core intermediate representation that can be used for semantic comparison and temporal reasoning, and is the prerequisite for achieving intelligent judgment and status understanding.

[0067] Perform hash coding on the field sequence feature vector and generate a field content signature value by combining the weighted field value and the field position index. This signature value is used to uniquely identify a group of status field combinations within the same device dimension.

[0068] After constructing the field sequence feature vector, it can be converted into a set of comparable digital fingerprints through a hash encoding process, allowing for quick identification of field content consistency. The core of this process is to bind the value of each key state field to its position index in the sequence. This not only records the numerical characteristics of the field itself, but also encodes its semantic position within the entire state structure. Specifically, each field in the field sequence is first assigned a unique position index—for example, the first field is 0, the second is 1, and so on. Each field value is then multiplied by its corresponding index weight coefficient, which can be a linearly increasing coefficient or a preset factor based on the field's importance, to form a set of "weighted field values." All weighted field values ​​are then concatenated sequentially into a long integer or fixed-length string, which serves as the hash input. Finally, a digest operation is performed on this weighted string using a hash algorithm with strong structural stability and low collision rate (such as SHA-256 or MurmurHash3) to generate the field content signature value. The signature value is unique within the same industrial equipment dimension, that is, as long as the value and order of the status field combination remain consistent, the same signature value will be generated even if the timestamp is different. Therefore, it can be used as a quick basis for determining whether the data content is repeated. Among them, "hash coding processing" is a summary method that maps input of arbitrary length to a fixed-length digital output to ensure structural invariance; "weighted field value" introduces the fusion consideration of field content and its semantic status in the overall structure, enhancing the feature expression capability; "field position index" ensures the uniqueness of the field arrangement order. Different arrangements can generate different signatures even if the field values ​​are the same, avoiding misjudgment. This mechanism can be executed immediately by software after data is received. It has the characteristics of high efficiency, strong discrimination, and can be embedded in subsequent graph structure modeling. It is the core support means for realizing rapid comparison and discrimination of equipment status content.

[0069] Retrieve historical data frames with the same field content signature value in the historical data buffer, and select the one with the smallest time interval with the current data frame as the comparison frame;

[0070] In an industrial control platform, to identify whether a received data frame is a reissued data frame, its field content must be quickly compared with historical data frames to verify duplicates. A key step in achieving this is to search the historical data buffer for historical data frames with the same field content signature value and select the frame with the smallest time interval with the current data frame as the comparison frame. This process can be implemented by constructing a hash-indexed data structure. Specifically, a hash table structure is maintained in the platform's data buffer, with the field content signature value as the key and the corresponding historical data frame list as the value. Whenever a new data frame arrives and its field sequence vector is hash-encoded, the signature value is used as a search criterion to quickly search the hash table for historical records with the same signature value. If so, all historical data frames with the signature value are sorted in ascending timestamp order. The absolute time interval between each historical data frame and the current data frame's reception time is calculated, and the frame with the smallest time interval is selected as the comparison frame. This comparison frame is the historical "equivalent state" that best matches the current data behavior context and is used to further analyze whether the current data represents a normal state flow or a reissued behavior. This approach is based on the high-frequency sampling, asynchronous reporting, and even unstable communication networks of device data, which can cause multiple status frames to recur in a short period of time. Only by accurately identifying the most recent "equal content history frame" can we determine whether reissues have occurred based on temporal logic, avoiding misjudging normal updates as redundant retransmissions. This provides a reliable reference point for subsequent timing offset judgment and validity identification. This method not only offers efficient retrieval performance and good scalability, but is also highly adaptable to the characteristic differences and real-time requirements of industrial data streams at the device level, and can be efficiently implemented entirely through platform software logic.

[0071] Determine whether the timestamps or sequence numbers of the current data frame and the comparison frame are non-increasing, reversed, or discontinuous. If so, mark the current data frame as a candidate for re-sending data.

[0072] To determine whether the current data frame is a candidate for retransmission, the system compares the logical relationship between the timestamp and sequence number of the current data frame and its corresponding comparison frame, focusing on identifying any anomalies such as non-incremental, reversed, or numbering gaps. Specifically, the system extracts the timestamp and sequence number values ​​of the current and comparison frames. Next, it establishes a judgment logic: If the timestamp of the current data frame is less than that of the comparison frame, it indicates a chronological regression and obvious non-incremental behavior. If the timestamps of the two frames are the same but the sequence number is less than or greater than the expected consecutive number (for example, the current number should be the comparison frame number plus one, but there is a jump or rollback), it is determined to be a sequence numbering gap or reversal. To avoid misjudgment, a reasonable numbering window tolerance can be set to ensure that normal fluctuations are not mistaken for abnormal behavior. Furthermore, the platform can combine historical device behavior models to determine whether the frame exhibits retransmission logic characteristics, such as whether it periodically retransmits its most recent status or whether it performs batch retransmissions after communication anomalies, thereby improving judgment robustness. The fundamental reason for this is that during the data communication process of industrial equipment, data with the same content is often sent multiple times due to factors such as network fluctuations, node cache retransmissions, or delayed processing of edge devices. If the platform determines the status of the device based solely on the consistency of field content, it will not be able to distinguish between real-time data and delayed retransmissions, which may lead to serious consequences such as status snapshot errors and false triggering of control logic. Therefore, through logical judgment of timestamps and sequence numbers, candidate data for retransmission behavior can be captured with minimal overhead and maximum real-time performance, providing the necessary data basis for whether it will be adopted as a valid state in the future. This judgment process can be completely embedded in the data receiving module through software processes and executed in real time, with good engineering feasibility and versatility.

[0073] Calculate the time offset between the reception time of the current data frame and the recording time of the comparison frame. If the offset is less than the preset task cycle tolerance threshold and the field content of the current data frame remains completely consistent with the comparison frame, then determine that the current data frame is a retransmitted data frame with duplicate content.

[0074] To achieve final confirmation of a reissued data frame, the time offset between the reception time of the current data frame and the recording time of its corresponding comparison frame can be calculated, combined with field content consistency to determine whether the data frame is a reissued data frame with duplicate content. Specifically, the timestamp of the current data frame's reception by the industrial control platform is extracted and subtracted from the timestamp of the comparison frame recorded in the historical data buffer to calculate the absolute time offset between the two. This time offset value reflects the difference in time between the two identical data frames being received by the platform. If this difference is very small (typically less than the standard sampling interval of a complete task cycle), it may indicate a short reissue. The field contents of the two data frames are then compared for complete consistency. If the field values ​​are completely consistent (this can be achieved by comparing field sequence feature vectors or signature values), the current data frame is essentially confirmed to be a redundant reissued data frame.

[0075] The "preset task cycle tolerance threshold" refers to the upper limit of the data duplication time allowed for each type of task or status refresh process during normal device operation. This threshold can be set based on parameters such as the device sampling cycle, control execution cycle, or status refresh cycle. For example, if a device has a 100ms control cycle, the threshold can be set to within 200ms. This design accounts for small-scale network jitter and prevents incorrectly misclassifying legitimate sampling frames as retransmission frames. The platform can automatically generate this threshold by learning from historical device operating data, or engineers can manually configure it based on task logic. This approach ensures that the platform incorporates temporal semantics when determining data validity, rather than relying solely on field content consistency. This prevents mistaking legitimately reported data from devices for the latest status, especially in the presence of interference factors such as communication delays, buffer overflows, and retransmission mechanisms. The entire process can be embedded in a software logic module, making the real-time calculation and threshold comparison process computationally lightweight and responsive, meeting the dual requirements of high-frequency data stability and accuracy for industrial control platforms.

[0076] For resent data frames with duplicate content, the behavioral entanglement between them and the resent data frame is calculated. If the entanglement exceeds a preset threshold, the sequence offset, time delay, and field fluctuation amplitude are extracted to generate an indicator vector describing the timeliness of the data identification.

[0077] In this embodiment, for a retransmitted data frame with duplicate content, the behavioral entanglement between the retransmitted data frame and the retransmitted data frame is calculated. If the entanglement exceeds a preset threshold, the sequence offset, time delay, and field fluctuation amplitude are extracted to generate an indicator vector describing the timeliness of the data identification. The specific steps include:

[0078] Collect a preset number of field sequence vectors of the reissued data frame with repeated content and the data frame being reissued within the time window before and after, arrange them in chronological order as a field value time series, and organize the field sequence vectors at each moment into a vector list in the sampling order to form a vector sequence trajectory representing the field evolution trend, where each trajectory is used to reflect the temporal change process of the same field combination in the state space before and after the reissue behavior;

[0079] This process is implemented through the data processing logic module in the software system. First, the platform continuously maintains a historical data buffer based on a sliding time window to store data frames reported by all devices. After identifying a reissued data frame with duplicate content and its corresponding reissued data frame, the system extracts a preset number of consecutive data frames from the time axis preceding and following these two frames. For each frame, a predefined set of key fields (such as temperature, speed, load, and current) is extracted to form a field sequence vector. The system then sorts these field sequence vectors in chronological order to form a field value time series. Finally, all field sequence vectors are combined into a two-dimensional vector list in sampling order. This vector list is a vector sequence trajectory, reflecting the device's state changes before and after the reissue. This trajectory construction process is accomplished through standardized array operations, timestamp sorting functions, and a field group parsing module, ensuring efficient execution within the system and scalable field templates for different device types.

[0080] The fundamental purpose of constructing vector sequence trajectories of field evolution trends is to provide a more context-aware representation of state changes than single-point field content comparison. In scenarios where data reissue occurs, field content consistency alone cannot determine whether the data is currently representative. However, by extracting the field change patterns over a continuous period before and after a data frame, it is possible to analyze whether the field combination has been continuously evolving or stagnant before and after the reissue. For example, if the field fluctuates significantly during the time period of the reissued data frame, while the field values ​​remain static before and after the reissued data frame, it is likely a historical state retransmission. Conversely, if the field trajectory is coherent and the changes are consistent, it is likely a real-time retransmission. This trajectory constructed based on the temporal context provides the necessary foundation for subsequent entanglement calculations, enabling the system to perform dynamic state trend comparisons, thereby improving the accuracy and stability of reissued frame judgments and avoiding distorted state judgments caused by the incorrect adoption of old data. This strategy meets the requirements of high-frequency and high-precision state recognition in industrial control and is a significant improvement over traditional static field comparison methods.

[0081] The behavioral entanglement between two vector sequence trajectories is calculated based on the cosine similarity and dynamic time warping algorithms. Cosine similarity is used to measure the consistency of local field changes, and dynamic time warping is used to compare the matching degree of the overall temporal structure.

[0082] If the behavior entanglement exceeds the preset threshold, the time delay between the timestamp of the resent data frame and the timestamp of the resent data frame is calculated, the difference between the sequence numbers is used as the sequence offset, and the differential mean of the corresponding values ​​of the fields is used as the field fluctuation amplitude;

[0083] This process can be implemented in software using the data processing and threshold judgment module embedded in the industrial control platform. The system first calculates behavioral entanglement based on the constructed field sequence trajectory. Entanglement is a quantitative score that indicates the degree of overlap in the evolutionary trends of the reissued data frame and the reissued data frame. The platform sets a "preset threshold" as a benchmark for determining behavioral entanglement. This threshold is obtained through unsupervised cluster analysis of a large amount of historical device operation data. Data samples with clear reissue behaviors are selected, their trajectory entanglement is statistically analyzed, and the quantile of their empirical distribution (e.g., 95th percentile) is extracted as the initial threshold. This threshold is then dynamically adjusted based on actual system operation experience. If the entanglement of a frame exceeds this threshold, the platform deems the reissue behavior highly suspicious and requires further analysis of the timing and content differences. At this point, the platform extracts the timestamps and sequence numbers of the current reissued data frame and the corresponding reissued data frame, calculates the time delay and sequence offset, and further extracts the corresponding field items in the field vectors of the two frames and calculates the mean difference between their values ​​as the field fluctuation amplitude. These steps can be efficiently implemented within the program through time comparison functions, position index difference operations, and field-level vector difference and average operations.

[0084] Specifically, the time delay is calculated by subtracting the recording time of the resent frame from the reception time of the current resent frame. For example, if the resent frame is received at 10:00:00 and the resent frame is received at 10:00:03, the delay is 3 seconds. The sequence offset is the difference in sequence numbers. For example, if numbers 105 and 110 are different, the offset is 5. The field fluctuation amplitude is calculated by taking the average of the absolute differences between the corresponding field values. For example, if the fields include temperature, current, and pressure, and the original values ​​are [100, 5.1, 0.8] and the resent frame values ​​are [100.5, 5.0, 0.85], the field differences are [0.5, 0.1, 0.05], with an average of approximately 0.216. Using these three metrics, the system can characterize the "identification timeliness" of the frame data from the three dimensions of time, sequence, and field variation, thereby constructing the feature inputs required for the subsequent judgment model. This approach takes into account both the temporal continuity and field consistency of device behavior, and is suitable for accurately identifying historical retransmitted frames in complex and dynamic industrial data environments.

[0085] The time delay, sequence offset and field fluctuation amplitude are normalized respectively and combined into an indicator vector with a unified dimension according to a preset order, which is used to characterize the identification timeliness of the reissued data frame.

[0086] In software implementation, this process can be accomplished through the feature engineering module built into the industrial control platform. The platform first normalizes the time delay, sequence offset, and field fluctuation amplitude to eliminate dimensional interference caused by different physical meanings and numerical ranges. Normalization can be performed using a sliding window-based min-max linear transformation, mapping the current value to the [0, 1] interval between the minimum and maximum values ​​statistically determined from a recent period of historical data. This process dynamically adapts to state fluctuations in the industrial process. The platform then combines the three normalized values ​​into a three-dimensional vector according to a preset indicator order (e.g., time delay as the first dimension, sequence offset as the second dimension, and field fluctuation amplitude as the third dimension). This creates a uniformly dimensional indicator vector. This preset order is determined by the platform during the model design phase and remains consistent throughout the system to ensure that the vectors generated for all devices and tasks have stable dimensional semantics and processing order. The fundamental purpose of this step is to provide subsequent state recognition models (such as embedded neural networks and time-sensitive classifiers) with input vectors that are consistent in format, comparable in value, and clearly defined. This improves the accuracy and robustness of the system's time-sensitive identification of re-sent data frames and lays the foundation for intelligent device state assessment. This vectorization and normalization approach not only improves the model's generalization capabilities but also supports its reusability across multiple scenarios and devices.

[0087] In this embodiment, the behavioral entanglement between two vector sequence trajectories is calculated based on the cosine similarity and dynamic time warping algorithm. The specific calculation method is as follows:

[0088] The field sequence vectors extracted from the retransmitted data frame with repeated content and the retransmitted data frame within the preset time window before and after each are arranged into two vector sequence tracks in sampling order. The field vectors in each track are normalized according to the maximum and minimum values ​​of the field, and the field values ​​are mapped to a unified interval.

[0089] This processing step can be implemented in software during the data preprocessing phase. Its primary purpose is to uniformly map the field sequence vectors collected at different times into a standardized numerical space to facilitate subsequent trajectory comparison and behavioral entanglement calculation. Specifically, after receiving a retransmitted data frame with duplicate content and its corresponding retransmitted data frame, multiple historical data frames within a preset time window are first collected both forward and backward, from which corresponding field sequence vectors are extracted, such as key fields such as temperature, pressure, and load. Next, each collected vector trajectory is sorted by sampling time, forming two ordered field vector sequences. To eliminate the influence of different fields in terms of dimension and value range, the values ​​of each field in the vector are normalized. Specifically, the maximum and minimum values ​​of each field within the time window are used as upper and lower bounds, and a standard normalization formula is used to linearly map them to the interval [0, 1]. For example, if the minimum value of a field within the window is 40 and the maximum value is 140, the original value of 60 will be normalized to (60-40) / (140-40)=0.2. Through this process, each field value is converted to a dimensionless value and standardized on a standard scale, preventing the dominance of certain fields due to large values ​​during subsequent comparisons. This approach not only enhances the comparability between trajectories but also helps subsequent algorithms (such as cosine similarity and dynamic time warping) more accurately identify field change trends and structural differences. For example, consider two trajectories representing the state sequence within 5 seconds before and after the resent data frame and the state sequence within 5 seconds before and after the resent data frame. After the above normalization process, two standardized field value trajectories are obtained for behavioral similarity assessment.

[0090] Based on the two normalized vector sequence trajectories, a dynamic time warping algorithm is used to calculate the Euclidean distance between each pair of field vectors in the two trajectories. A one-to-one matching path is constructed based on the minimum cumulative cost, and the sampling time position of each pair of vectors in the path in their respective trajectories is recorded.

[0091] This processing step can be implemented using time series analysis algorithms integrated into the software, such as Python's dtaidistance, tslearn, or MATLAB's DTW module. The implementation process involves first obtaining two normalized field vector trajectories (e.g., representing the state evolution sequence of the retransmitted and retransmitted data frames over the preceding and subsequent time periods, respectively) and performing dynamic time warping (DTW) matching on them as input time series. During execution, the software calculates the Euclidean distance between each field vector in each trajectory and all vectors in the other trajectory, forming a two-dimensional cost matrix. Next, the software searches for a path with the lowest cost (minimum cumulative distance) within this cost matrix as the matching path, thereby establishing a one-to-one correspondence between two vector trajectories of different lengths or incomplete alignment. During this process, the software also records the sampling time position of each successfully matched pair of field vectors to support subsequent similarity calculation and structural analysis. This approach effectively addresses trajectory asynchrony caused by time-varying device data, such as rate fluctuations and sampling offsets, allowing comparisons to focus on the field's changing trends rather than the interference of temporal misalignment.

[0092] Dynamic Time Warping (DTW) is a widely used algorithm for time series alignment analysis. Its purpose is to nonlinearly stretch time series to achieve optimal pairing. It allows one point in one time series to be aligned with multiple points in the other, thus addressing uneven sampling or time delay. Euclidean distance measures the difference between two field vectors and is defined as the square root of the sum of the squared differences between the two vectors across all field dimensions. For example, if the two field vectors are [0.2, 0.3, 0.6] and [0.3, 0.4, 0.5], the Euclidean distance is √((0.2-0.3)²+(0.3-0.4)²+(0.6-0.5)²)=√(0.01+0.01+0.01)=√0.03≈0.173. The DTW algorithm searches the entire cost matrix for the "minimum cumulative cost path," meaning the path with the smallest sum of all Euclidean distances from the starting point to the end point. This path is then used as the optimal matching path between two trajectories, ensuring the most reasonable overall pairing and the lowest global error. This path serves as the basis for subsequent similarity and behavior alignment calculations.

[0093] For all paired field vectors in the matching path, calculate the cosine similarity, form a similarity value sequence in the order of the paths, and calculate the arithmetic mean of the value sequence as the first value to measure the consistency of the two trajectories in the direction of local field change;

[0094] This step can be accomplished using vector similarity analysis tools integrated into data analysis platforms, such as the NumPy or Scikit-learn libraries in Python. First, all paired field vector pairs are extracted from the matching paths obtained by dynamic time warping (DTW). For each pair of vectors, the cosine similarity function is applied. Cosine similarity measures the directional consistency between two vectors and is defined as the dot product of the two vectors divided by the product of their moduli, with a value range of [-1, 1]. The software performs this calculation on each pair of matching field vectors, and the results are organized into a sequence of similarity values, arranged in the order of the paths. Finally, all similarity values ​​in this sequence are arithmetic averaged to obtain a value reflecting the consistency of the local field evolution direction between the two field sequence trajectories. This value is referred to as the "first value" in this step. Cosine similarity is used as a basis for determining the direction of local field change because even if the field values ​​differ slightly, as long as the change trend is consistent, the cosine similarity remains high. This provides strong robustness to sudden jitter and sampling accuracy variations, improving the accuracy of behavioral pattern recognition.

[0095] For example, suppose two field vectors are A = [0.6, 0.8, 0.1] and B = [0.5, 0.75, 0.2]. First, calculate their dot product: 0.6 × 0.5 + 0.8 × 0.75 + 0.1 × 0.2 = 0.3 + 0.6 + 0.02 = 0.92. Then, calculate the moduli of the two vectors: ||A|| = √(0.36 + 0.64 + 0.01) = √1.01 ≈ 1.005, and ||B|| = √(0.25 + 0.5625 + 0.04) = √0.8525 ≈ 0.923. The final cosine similarity is 0.92 ÷ (1.005 × 0.923) ≈ 0.92 ÷ 0.927 ≈ 0.993, indicating that the two vectors are very consistent in direction. By processing all paired vectors in this way and ultimately taking the average of all cosine similarities (e.g., 0.976 after multiple pairs), we can obtain a consistency index for the direction of local field change. The closer this value is to 1, the more consistent the field change trends between the two trajectories.

[0096] The sampling time position differences of each field vector pair in the matching path are averaged to obtain the degree of offset of the two trajectories in the overall time structure as the second value. The first and second values ​​are normalized respectively and then linearly combined according to the preset weighting ratio to obtain the behavioral entanglement degree used to characterize the degree of consistency of the overall changes of the two vector trajectories.

[0097] This process can be implemented using the time series analysis module in software, such as Python's Pandas and NumPy libraries. First, for the matching paths generated by the dynamic time warping algorithm, the sampling time index (i.e., their respective temporal positions) of each pair of paired vectors in the original trajectory are extracted. The difference between the sampling time positions of each pair of vectors is then calculated, recorded, and averaged to obtain the average offset between the two trajectories in terms of their overall temporal structure, which serves as the second value. This value reflects the degree of temporal structural consistency between the reissued data and the original data, specifically whether they overlap significantly in the temporal dimension. Subsequently, the first value (i.e., the mean cosine similarity) and the second value are normalized to align their dimensions. The normalization method can be the minimum-maximum method, which maps each value to the interval [0, 1] for subsequent combined analysis. This step combines the two dimensions of "local field change trend consistency" and "overall temporal structure matching" to produce a more comprehensive quantitative indicator of behavioral consistency, helping to accurately determine the authenticity and validity of the data frame status.

[0098] After normalization, the system linearly combines the first and second values ​​according to a preset weighting ratio to generate the final behavioral entanglement. This linear combination is typically expressed as: Behavioral entanglement = α × first value + β × second value, where α and β are the weighting coefficients of the normalized first and second values, respectively, such that α + β = 1. This weighting ratio can be empirically tuned based on the reliability requirements of the industrial scenario, data characteristics, and statistical analysis of historical false positive rates. For example, in scenarios where state change trends are more critical, α can be set to 0.7 and β to 0.3. If the device is more sensitive to time delays, the weightings can be reversed. The weightings are determined through offline training or a cross-validation-based model evaluation strategy, and continuously optimized through a feedback mechanism to ensure that the behavioral entanglement accurately reflects the overall matching between vector trajectories. This approach prevents a single dimension from dominating the judgment results, making the recognition of reissued data more robust and accurate.

[0099] The generated indicator vector is input into the pre-trained graph embedding algorithm model to generate the first parameter and the second parameter. The first parameter is used to describe the consistency of the data frame with the historical evolution path, and the second parameter is used to characterize the timeliness of the data frame identification.

[0100] In this embodiment, the generated indicator vector is input into a pre-trained graph embedding algorithm model to generate the first parameter and the second parameter, specifically:

[0101] The generated indicator vector is used as the feature vector of the current data frame. The graph node corresponding to the field combination matching the indicator vector is located in the causal graph between the field combination and the control instruction. The adjacent nodes and edge weight information of the node in the graph structure are collected to construct the structural adjacency matrix and the node feature matrix.

[0102] To effectively map the generated indicator vectors to the corresponding field combination nodes in the causal graph, the causal graph must first be stored in a graph structure that supports retrieval and indexing. Each field combination is modeled as a graph node, with each edge representing the causal relationship between the field combination and a control instruction. Edge weights can be used to express instruction frequency or path confidence. The system traverses all field combination nodes in the graph, encoding their static features (such as field type, order, and Boolean flags). Similarity matching (such as Euclidean distance or cosine similarity) is performed with the indicator vector to identify the closest graph node. After identification, the system extracts the target node's adjacent nodes and their edge information, constructing a structural adjacency matrix (storing node connectivity in the form of an adjacency list or sparse matrix) and a node feature matrix (recording the features of the current node and its adjacent nodes, such as indicator values, task frequencies, and upstream and downstream state patterns). These provide complete input for the subsequent graph embedding model. This process achieves structural alignment between the indicator vector and the graph's semantic context, ensuring the accuracy of subsequent semantic computations.

[0103] Among them, the "structural adjacency matrix" is used to represent the topological structure between the nodes in the graph. It is a two-dimensional matrix. Each element in the matrix indicates whether there is an edge connection between the two nodes (such as a value of 1 indicates an edge, and 0 indicates no edge). It can be expanded to store edge weight information to represent the strength of the control logic; and the "node feature matrix" is used to store the feature information of each node in the graph. Each row corresponds to a node, and each column represents a certain feature dimension, such as the stability label of the device status, the historical execution frequency, the typical response instruction set, etc. The purpose of constructing these two matrices is to convert the original graph data structure into a numerical input that can be directly processed by the neural network model, so that the graph embedding model can simultaneously capture the internal attributes of the field combination node (through the feature matrix) and the topological relationship in the graph structure (through the adjacency matrix), and then accurately produce a high-dimensional vector representation that reflects the behavioral background and similarity of the state in the control chain, providing basic support for the subsequent state recognition and judgment mechanism.

[0104] The structural adjacency matrix and the node feature matrix containing the indicator vector are input into a pre-trained graph embedding algorithm model. The model outputs an embedding vector representing the temporal semantics of the graph node based on the integration of the field combination context structure and the indicator dynamic features.

[0105] This step is achieved by inputting both the structural adjacency matrix and the node feature matrix containing indicator vectors into a pre-trained graph embedding algorithm model to obtain an embedded semantic representation of the current field combination node. Specifically, this approach involves first employing mainstream graph neural network models such as GCN (graph convolutional network), GAT (graph attention network), or GraphSAGE (graph sampling aggregation). These models are capable of simultaneously processing structural adjacency relationships and node attribute features. The model receives the adjacency matrix, understands the connection structure between the current field combination node and its adjacent nodes, and establishes the semantic context of the field state combination within the device's operation path. It also receives the node feature matrix and performs nonlinear transformations and feature fusion on the state indicator information carried by the node itself (i.e., the previously generated indicator vector). The model has been trained on a large number of historical task graphs and is able to capture the co-occurrence logic and dynamic evolution patterns between field combinations and control paths. Therefore, when inputting this structure during inference, it outputs a fixed-dimensional embedding vector representing the current data frame in terms of both graph structure and dynamic semantics. For example, a field combination like "feed = 1, speed = 1500, temperature = 120" can be converted into a 128-dimensional vector through graph embedding. Some dimensions express its degree of alignment with the "heating instruction" path, while others represent its historical stability and sensitivity to disturbances, providing strong support for subsequent parameter generation and state judgment. This approach is important because it unifies the static structural position and dynamic operational performance of the data frame in the same semantic space, significantly improving the contextual understanding capabilities of temporal data discrimination.

[0106] Extracting several dimensional components for measuring the consistency between the current data frame and the historical field evolution trajectory from the embedding vector, calculating the mean of the dimensional components, and obtaining a first numerical parameter for characterizing the field evolution consistency;

[0107] The method for extracting and calculating the first numerical parameter used to characterize field evolution consistency is primarily based on the semantic separation mechanism of graph embedding vector dimensions. During the training phase of the graph embedding model, supervised or self-supervised learning is used to guide some embedding dimensions to specifically learn feature representations of "field evolution path similarity." For example, these dimensions are trained using consistent labels of evolution paths between historical data frames, so that they focus on semantics such as trajectory direction, field trend, and task stage in the embedding space. After model training is complete, during the actual inference phase, the components corresponding to these trained dimensions are extracted from the embedding vector of the current data frame. These are typically continuous or discontinuous dimensions (such as the 5th, 7th, 9th, and 12th dimensions). The arithmetic mean of these values ​​is taken as the first numerical parameter. For example, if the extracted field evolution dimension components are [0.78, 0.82, 0.75, 0.81], their mean is 0.79, representing the average consistency of the current data frame's field change trend with the historical path. A higher value indicates a closer fit between the evolution trajectory and historical patterns, and a higher credibility. The physical meaning of this parameter is to provide a quantitative indicator to measure whether the re-sent data frame continues the evolution logic of the known task path, and to assist in determining whether it can be trusted as the current valid state of the device.

[0108] Extracting several dimensional components used to measure the consistency between the current data frame and historical field evolution trajectories can be achieved through dimension semantic calibration during model training. Specifically, when constructing training data for the graph embedding model, a corresponding field combination evolution label or trajectory consistency score is constructed for each data frame. A multi-head graph embedding output mechanism is then implemented, pre-defining several dimensions in the embedding vector as "trajectory consistency-specific dimensions." Using a multi-task loss function, these dimensions are focused on learning trajectory trends rather than other semantics (such as temporal perturbations). After training, these dimensions have clear semantic assignments in the vector output, becoming dedicated channels for consistency judgment. During inference, the corresponding dimensions can be extracted according to this definition. For example, if dimensions 10-15 in the model are trajectory consistency dimensions, these sub-vectors are directly extracted and averaged. The advantages of this approach include stable and standardized extraction, eliminating the need for subsequent secondary clustering or classification, and ensuring interpretable and universal dimension semantics, thus supporting cross-scenario reuse and consistent model inference.

[0109] Several dimensional components reflecting the field disturbance amplitude and time offset performance in the current data frame and the historical control path are extracted from the embedded vector, and the weighted average of the components is calculated to obtain a second numerical parameter for characterizing the state delay and disturbance amplitude.

[0110] To generate a second numerical parameter to characterize state delay and disturbance amplitude, several dimensional components related to field disturbance amplitude and time drift can be extracted from the embedding vector output by the graph embedding model and a weighted average of these components can be performed. This is achieved as follows: During model training, labels for "time delay" and "field disturbance amplitude" corresponding to the retransmitted data frames are introduced. A set of dedicated dimensions is trained through a regression task to reflect the degree of temporal drift and field stability of the data frames. During inference, these pre-set dimensional components are extracted from the embedding vector, multiplied by the corresponding disturbance or delay weights (e.g., based on the impact of field fluctuations on system control logic), and the weighted sum is then divided by the total weight to obtain the second numerical parameter. For example, if the extracted dimension values ​​are [0.45, 0.62, 0.57] and the corresponding weights are [0.3, 0.4, 0.3], the weighted average is 0.45 × 0.3 + 0.62 × 0.4 + 0.57 × 0.3 = 0.555, which is the second numerical parameter. The physical meaning of this parameter is to measure whether the current data frame has significant delays or abnormal field fluctuations, indicating whether the data frame reliably reflects the current state. Higher values ​​indicate greater disturbances, more severe lags, and lower credibility.

[0111] Extracting several dimensional components reflecting the magnitude of field perturbations and time offsets between the current data frame and historical control paths can be achieved by implementing a task-layered supervision strategy during the graph embedding model training phase. Specifically, two types of labels are assigned to each data frame in the training data: one for the magnitude of field perturbations (e.g., the standard deviation of the field value fluctuations) and the other for the time offset (e.g., the ratio of the reissue delay to the standard task period). The model architecture then specifies several dimensions for regressing the magnitude of field perturbations and others for fitting the degree of time offset. These dimensions are jointly optimized using different loss functions to achieve semantically separate training. During inference, these dimensions have fixed index positions in the embedding vector and can be directly extracted by developers, such as dimensions 3 and 5 for field perturbations and dimensions 6 and 7 for time offsets. After extraction, weights are assigned to each dimension based on its predictive power during training, and a weighted average is performed to generate a comprehensive parameter for state assessment. This approach ensures that the extracted dimensional components are clearly defined, semantically consistent, and reusable, avoiding subsequent human intervention or semantic ambiguity.

[0112] In this embodiment, the pre-trained graph embedding algorithm model is specifically:

[0113] Based on the graph structure consisting of the causal paths between field combinations and control instructions and its historical state data, training samples are constructed. The historical indicator vector corresponding to each field combination node is used as input, and the state judgment results of this node in historical tasks are used as training targets. Through training, the graph embedding algorithm model can generate embedding vectors based on the graph structure and indicator characteristics.

[0114] After training, the model takes as input the structural adjacency matrix and the indicator vector and outputs an embedding vector, which contains the numerical dimensions used to generate the first and second parameters, respectively used to reflect the consistency of the data frame with the historical state path and the timeliness of its identification state.

[0115] The process of constructing training samples based on a graph structure consisting of causal paths between field combinations and control instructions and its historical status data, using the historical indicator vector corresponding to each field combination node as input and the status determination results of that node in historical tasks as training targets, and enabling the graph embedding algorithm model to generate embedding vectors based on the graph structure and indicator features through training, can be implemented in software as follows: First, a graph structure is constructed based on the sequential relationship between device status fields and control instructions issued by the platform in historical industrial data, where nodes represent field combinations and edges represent sequential causal relationships in control logic. Next, for each field combination node, label information indicating whether it was determined to be "valid" or "invalid" in multiple tasks is collected as training targets. Simultaneously, the indicator vector associated with each field combination before each determination (including indicators such as time delay, sequence offset, and field perturbation) is extracted as training input. When constructing the graph embedding model, the indicator vector of the field combination node is embedded in the node feature matrix. The adjacency relationships in the graph structure are used to guide the model in learning the semantic dependencies between nodes. Forward propagation is performed through graph neural network algorithms (such as GCN, GAT, etc.), and the model output is compared with the true state label. The model parameters are optimized through error backpropagation until the model can stably output node embedding vectors with distinguishing capabilities.

[0116] This is necessary because traditional state judgment algorithms often rely on static rules or threshold judgments, lacking a deep understanding of the context of device behavior changes. By constructing a graph structure of field combinations and instructions, not only is the causal chain between state and behavior preserved, but the graph embedding model can also effectively integrate the node's own characteristics (indicator vector) with its graph neighbor information, enabling the model to mine temporal patterns from the graph and identify subtle state differences. Ultimately, this training process enables the model to output a set of vector components containing semantic features when faced with new reissued data frames. The parameters extracted from these can be used to accurately distinguish whether the state is real and valid, thereby preventing the platform from making erroneous control decisions.

[0117] Based on a joint determination of the first parameter and the second parameter, determining whether the status field corresponding to the data frame should be used as a valid operating status of the current device;

[0118] In this embodiment, based on the combined judgment of the first parameter and the second parameter, it is determined whether the status field corresponding to the data frame should be used as the valid operating status of the current device, specifically:

[0119] Construct a two-dimensional parameter judgment space, using the first parameter as the horizontal coordinate and the second parameter as the vertical coordinate. Scatter distribution annotation is performed on the known valid and invalid state data frames during the historical task execution process, and the valid state judgment boundary is preset based on the distribution results.

[0120] receiving a first parameter and a second parameter corresponding to a current data frame, and calculating a coordinate point corresponding to the data frame in a two-dimensional parameter determination space;

[0121] Determine whether the coordinate point is within the preset valid state determination boundary. If it is within the boundary, the state field corresponding to the data frame is determined to be the valid operating state of the current device. If it is outside the boundary, it is determined to be an invalid state.

[0122] The judgment process can be implemented in software. The main steps include: First, based on historical device status data, the numerical distribution of the first and second parameters corresponding to all data frames identified as valid and invalid operating states is statistically analyzed. These historically labeled samples are used to plot a scatter plot in a two-dimensional coordinate system. Interpretable decision boundary construction methods, such as density clustering or support vector machines (SVM), are then used to generate a boundary curve that demarcates the valid and invalid state regions. Second, upon receiving each data frame to be judged, the system first calculates the coordinates of its point in the two-dimensional decision space based on the first and second parameters generated for that data frame. The system then invokes a built-in boundary judgment algorithm to determine whether the point lies within the valid state region. If it does, the current device status snapshot is updated to contain the data frame's field contents. If not, the data frame is discarded as invalid input, and its identification information is recorded for subsequent model self-verification updates. The main purpose of introducing the decision space in this method is to transform the judgment of continuous numerical parameters into a visual spatial decision process, improving the system's memory of historical experience and the stability of the judgment rules, avoiding the misjudgment problems associated with simple threshold judgments, and providing an intuitive basis for subsequent model adjustments and manual intervention.

[0123] The preset effective state judgment boundary is constructed based on the distribution characteristics of historical data using a fixed threshold fitting method. The threshold used is derived from the statistical analysis of historical state data to ensure the consistency and stability of the judgment criteria.

[0124] The preset valid state determination boundary can be implemented in the software through statistical analysis of historical state data. The specific approach is as follows: First, the system collects data frame samples that have been manually or rule-basedly labeled as "valid state" and "invalid state" in historical tasks, extracts the first and second parameter values ​​corresponding to these samples, and visualizes their scatter distribution on a two-dimensional plane. Subsequently, the value ranges of the first and second parameters of all valid state samples are counted, and their means and standard deviations are calculated respectively. A threshold range is set in each parameter dimension, with the deviation from the mean not exceeding a set multiple (e.g., 1.5 times the standard deviation) to construct a regular boundary. Alternatively, equal frequency binning and quantile analysis can be used to extract the 80%–90% main dense distribution area. These intervals are then used to jointly construct a rectangular or elliptical area as the valid state determination boundary, and the boundary parameters are written into the system model configuration to implement the regular application of the boundary. This method ensures that boundary construction is based on real data distribution characteristics rather than manual experience judgment, improving the stability and generalization ability of state judgment. At the same time, because it is based on fixed threshold fitting rather than model training, it has strong interpretability and debuggability, and is suitable for deployment and iterative optimization in industrial environments.

[0125] The device status snapshot is updated based on the judgment results, and the judgment results are fed back to the causal graph path to dynamically adjust the path confidence weight and state identification strategy.

[0126] This step can be implemented in software in the following way: After the system completes the valid status judgment for a data frame, it immediately writes the judgment result and the field combination of the data frame into the device status snapshot table. The status snapshot structure uses the field combination as the primary key, and is accompanied by the latest judgment label (valid / invalid), timestamp, and device ID and other identification information. Subsequently, the system locates the node position of the field combination in the causal graph, traces back along its path to the field combination node and control instruction edge of the previous hop, and updates the confidence weight on the path. The update method includes: if the current judgment is a valid state, increase the confidence score of the current path edge (for example, the cumulative number of hits +1, and calculate its relative frequency in all paths); if the judgment is an invalid state, reduce the confidence score of the path edge, and record the context information of the failed status judgment.

[0127] This mechanism is designed to achieve dynamic evolution of the credibility of paths in the graph, enabling the platform to gradually learn the rationality of paths from actual control experience, thereby improving the accuracy of subsequent state recognition and the precision of decision-making. In addition, by continuously feeding back new state judgment results to the graph, it can also trigger adaptive adjustments to the state recognition strategy. For example, when a path is judged invalid for a long time, the system can automatically weaken its weight or screening conditions for participating in subsequent embedded reasoning, thereby preventing outdated or abnormal paths from misleading real-time control. This feedback mechanism not only improves the real-time and robustness of the graph model, but also enhances the intelligent evolution capability of the entire control platform.

[0128] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0129] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. A computer program product comprises one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, the processes or functions according to the embodiments of the present application are fully or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). A computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0130] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0131] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0132] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative. For example, the division of units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0133] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0134] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0135] The above are only specific embodiments of the present application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. The industrial equipment interconnection method based on the industrial control platform is characterized by: The specific steps include: Construct a causal graph between industrial equipment status fields and control instructions to record the causal relationship between field combinations and task evolution paths; Receive data frames from industrial equipment and determine whether the data frame is consistent with the historical frame field content and the timestamp or sequence number is different. If so, mark it as a retransmitted data frame with duplicate content; For reissued data frames with duplicate content, the behavioral entanglement between them and the reissued data frame is calculated. If the entanglement exceeds a preset threshold, the sequence offset, time delay, and field fluctuation amplitude are extracted to generate an indicator vector used to describe the timeliness of the data identification. The behavioral entanglement is based on the normalized field sequence vectors of the reissued data frame with duplicate content and the reissued data frame, obtained by linearly combining the results of the dynamic time warping algorithm and the cosine similarity calculation according to a preset weighted ratio. The generated indicator vector is input into the pre-trained graph embedding algorithm model to generate the first parameter and the second parameter. The first parameter is used to describe the consistency of the data frame with the historical evolution path, and the second parameter is used to characterize the timeliness of the data frame identification. Based on a joint determination of the first parameter and the second parameter, determining whether the status field corresponding to the data frame should be used as a valid operating status of the current device; The device status snapshot is updated based on the judgment results, and the judgment results are fed back to the causal graph path to dynamically adjust the path confidence weight and state identification strategy.

2. The industrial equipment interconnection method based on the industrial control platform according to claim 1 is characterized in that: Construct a causal graph between industrial equipment status fields and control instructions, specifically: By collecting historical operating data of industrial equipment during task execution, the field groups and corresponding control instructions contained in each status data frame are extracted, and a triple set consisting of field groups, control instructions and their occurrence sequence is constructed; The triple set is represented as a directed edge relationship in a graph structure, with the field group as the starting node and the control instruction as the ending node. At the same time, time weight and frequency weight are assigned to the edges according to the task evolution timeline to form a directed causal graph that can be used for state evolution path reasoning.

3. The industrial equipment interconnection method based on the industrial control platform according to claim 2 is characterized in that: Receive data frames from industrial equipment and determine whether the data frame is consistent with the historical frame field content and the timestamp or sequence number is different. If so, mark it as a retransmitted data frame with duplicate content. Specifically: Receive data frames from industrial equipment, extract multiple key status fields from the data frames, arrange them in the order of field collection and the industrial equipment data structure template, and construct a field sequence feature vector to represent the complete equipment status corresponding to the current data frame; Perform hash coding on the field sequence feature vector and generate a field content signature value by combining the weighted field value and the field position index. This signature value is used to uniquely identify a group of status field combinations within the same device dimension. Retrieve historical data frames with the same field content signature value in the historical data buffer, and select the one with the smallest time interval with the current data frame as the comparison frame; Determine whether the timestamps or sequence numbers of the current data frame and the comparison frame are non-increasing, reversed, or discontinuous. If so, mark the current data frame as a candidate for re-sending data. Calculate the time offset between the reception time of the current data frame and the recording time of the comparison frame. If the offset is less than the preset task cycle tolerance threshold and the field content of the current data frame remains completely consistent with the comparison frame, then determine that the current data frame is a retransmitted data frame with duplicate content.

4. The industrial equipment interconnection method based on the industrial control platform according to claim 3 is characterized in that: For resent data frames with duplicate content, the behavioral entanglement between them and the resent data frame is calculated. If the entanglement exceeds a preset threshold, the sequence offset, time delay, and field fluctuation amplitude are extracted to generate an indicator vector describing the timeliness of the data identification. The specific steps include: Collect a preset number of field sequence vectors of the reissued data frame with repeated content and the data frame being reissued within the time window before and after, arrange them in chronological order as a field value time series, and organize the field sequence vectors at each moment into a vector list in the sampling order to form a vector sequence trajectory representing the field evolution trend, where each trajectory is used to reflect the temporal change process of the same field combination in the state space before and after the reissue behavior; The behavioral entanglement between two vector sequence trajectories is calculated based on the cosine similarity and dynamic time warping algorithms. Cosine similarity is used to measure the consistency of local field changes, and dynamic time warping is used to compare the matching degree of the overall temporal structure. If the behavior entanglement exceeds the preset threshold, the time delay between the timestamp of the resent data frame and the timestamp of the resent data frame is calculated, the difference between the sequence numbers is used as the sequence offset, and the differential mean of the corresponding values ​​of the fields is used as the field fluctuation amplitude; The time delay, sequence offset and field fluctuation amplitude are normalized respectively and combined into an indicator vector with a unified dimension according to a preset order, which is used to characterize the identification timeliness of the reissued data frame.

5. The industrial equipment interconnection method based on the industrial control platform according to claim 4 is characterized in that: The behavioral entanglement between two vector sequence trajectories is calculated based on the cosine similarity and dynamic time warping algorithm. The specific calculation method is as follows: The field sequence vectors extracted from the retransmitted data frame with repeated content and the retransmitted data frame within the preset time window before and after each are arranged into two vector sequence tracks in sampling order. The field vectors in each track are normalized according to the maximum and minimum values ​​of the field, and the field values ​​are mapped to a unified interval. Based on the two normalized vector sequence trajectories, a dynamic time warping algorithm is used to calculate the Euclidean distance between each pair of field vectors in the two trajectories. A one-to-one matching path is constructed based on the minimum cumulative cost, and the sampling time position of each pair of vectors in the path in their respective trajectories is recorded. For all paired field vectors in the matching path, calculate the cosine similarity, form a similarity value sequence in the order of the paths, and calculate the arithmetic mean of the value sequence as the first value to measure the consistency of the two trajectories in the direction of local field change; The sampling time position differences of each field vector pair in the matching path are averaged to obtain the degree of offset of the two trajectories in the overall time structure as the second value. The first and second values ​​are normalized respectively and then linearly combined according to the preset weighting ratio to obtain the behavioral entanglement degree used to characterize the degree of consistency of the overall changes of the two vector trajectories.

6. The industrial equipment interconnection method based on the industrial control platform according to claim 5 is characterized in that: The generated indicator vector is input into the pre-trained graph embedding algorithm model to generate the first and second parameters, specifically: The generated indicator vector is used as the feature vector of the current data frame. The graph node corresponding to the field combination matching the indicator vector is located in the causal graph between the field combination and the control instruction. The adjacent nodes and edge weight information of the node in the graph structure are collected to construct the structural adjacency matrix and the node feature matrix. The structural adjacency matrix and the node feature matrix containing the indicator vector are input into a pre-trained graph embedding algorithm model. The model outputs an embedding vector representing the temporal semantics of the graph node based on the integration of the field combination context structure and the indicator dynamic features. Extracting several dimensional components for measuring the consistency between the current data frame and the historical field evolution trajectory from the embedding vector, calculating the mean of the dimensional components, and obtaining a first numerical parameter for characterizing the field evolution consistency; Several dimensional components reflecting the field disturbance amplitude and time offset performance in the current data frame and the historical control path are extracted from the embedded vector, and the weighted average of the components is calculated to obtain a second numerical parameter for characterizing the state delay and disturbance amplitude.

7. The industrial equipment interconnection method based on the industrial control platform according to claim 6 is characterized in that: The pre-trained graph embedding algorithm model is as follows: Based on the graph structure consisting of the causal paths between field combinations and control instructions and its historical state data, training samples are constructed. The historical indicator vector corresponding to each field combination node is used as input, and the state judgment results of this node in historical tasks are used as training targets. Through training, the graph embedding algorithm model can generate embedding vectors based on the graph structure and indicator characteristics. After training, the model takes as input the structural adjacency matrix and the indicator vector and outputs an embedding vector, which contains the numerical dimensions used to generate the first and second parameters, respectively used to reflect the consistency of the data frame with the historical state path and the timeliness of its identification state.

8. The industrial equipment interconnection method based on the industrial control platform according to claim 7 is characterized in that: Based on the combined judgment of the first parameter and the second parameter, it is determined whether the status field corresponding to the data frame should be used as the valid operating status of the current device, specifically: Construct a two-dimensional parameter judgment space, using the first parameter as the horizontal coordinate and the second parameter as the vertical coordinate. Scatter distribution annotation is performed on the known valid and invalid state data frames during the historical task execution process, and the valid state judgment boundary is preset based on the distribution results. receiving a first parameter and a second parameter corresponding to a current data frame, and calculating a coordinate point corresponding to the data frame in a two-dimensional parameter determination space; Determine whether the coordinate point is within the preset valid state determination boundary. If it is within the boundary, the state field corresponding to the data frame is determined to be the valid operating state of the current device. If it is outside the boundary, it is determined to be an invalid state. The preset effective state judgment boundary is constructed based on the distribution characteristics of historical data using a fixed threshold fitting method, and the threshold used is derived from the statistical analysis of historical state data.

Images