Digital visual control method and system for industrial automatic equipment

Abstract

The invention relates to the technical field of equipment control, in particular to a digital visual control method and system for industrial automatic equipment. The method comprises the following steps: acquiring multi-source heterogeneous sensing data, and performing dynamic label analysis and signal decoupling to obtain a multi-source preprocessing data set; performing multi-level semantic association enhancement according to the multi-source preprocessing data set to obtain a dynamic semantic association network; performing eigenvector weighted regression based on the dynamic semantic association network to obtain a multi-modal unified eigenvector; performing equipment multi-dimensional dynamic behavior mapping by using the multi-modal unified feature vector to obtain an industrial process multi-dimensional dynamic model; acquiring real-time multi-source heterogeneous sensing data, and identifying a multi-dimensional abnormal outlier index of the equipment; and carrying out adaptive parameter reconstruction on the multi-dimensional abnormal outlier index of the equipment to obtain an equipment control instruction sequence, and issuing the equipment control instruction sequence to a corresponding equipment control unit. According to the invention, the response speed and the control efficiency of the industrial automatic equipment can be optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of device control, and in particular to an industrial automatic device digital visual control method and system. BACKGROUND

[0002] The digital visual control method directly presents device running state, process parameters, production data and other information in a graphical and dynamic form, greatly improving the understanding and control efficiency of the operating personnel on the device running state, and promoting the development of intelligent industrial devices. Traditional industrial automatic device control mostly uses control methods based on physical keys or simple human-machine interfaces. Limited by interface display capability and data interaction mode, it is impossible to achieve comprehensive and real-time monitoring and interaction of complex device states. With the popularization of the concept of Industry 4.0 and the development of device interconnection and big data technology, digital and visual control has gradually become an important trend in the development of industrial automation. However, the digital visual control system of traditional industrial automatic devices generally lacks highly integrated data fusion capability. There are a large number of heterogeneous sensors and devices in the industrial field environment, and the multi-source heterogeneous data generated by these devices is difficult to integrate efficiently, resulting in lagging and incomplete visual information of device states, which affects the accuracy of control decisions. Secondly, the existing visual control interface is mainly two-dimensional static display, lacking three-dimensional modeling and real-time interaction capability for complex device structure and dynamic running state, and it is difficult to meet the needs of operating personnel for in-depth understanding and real-time regulation and control of device internal running mechanism. SUMMARY

[0003] Therefore, it is necessary to provide an industrial automatic device digital visual control method and system to solve at least one of the above technical problems.

[0004] To achieve the above-mentioned purpose, an industrial automatic device digital visual control method comprises the following steps:

[0005] Step S1: acquiring multi-source heterogeneous sensor data, and performing dynamic label analysis and signal decoupling according to the multi-source heterogeneous sensor data to obtain a multi-source preprocessed data set;

[0006] Step S2: constructing an industrial process structure model according to the multi-source preprocessed data set, and performing multi-level semantic association enhancement between the industrial process structure model and semantic entities in the multi-source preprocessed data set to obtain a dynamic semantic association network;

[0007] Step S3: based on the dynamic semantic association network, and combining the motion physical quantity characteristics of the end effector of the mechanical arm and the conveyor belt in the multi-source preprocessed data set to perform multi-task feature vector fusion, obtaining a preliminary fused multi-task feature vector; performing feature vector weighted regression on the preliminary fused multi-task feature vector to obtain a multi-modal unified feature vector;

[0008] Step S4: mapping the multi-dimensional dynamic behavior of the equipment using the multi-modal unified feature vector, thereby obtaining a multi-dimensional dynamic model of the industrial process;

[0009] Step S5: obtaining real-time multi-source heterogeneous sensing data; predicting the running state of the equipment by the multi-dimensional dynamic model of the industrial process, and identifying multi-dimensional abnormal outlier indexes of the equipment according to the running state prediction result;

[0010] Step S6: performing adaptive parameter reconstruction on the multi-dimensional abnormal outlier indexes of the equipment, obtaining a device control instruction sequence, and issuing it to the corresponding device control unit.

[0011] Optionally, step S1 comprises:

[0012] Step S11: collecting multi-source heterogeneous sensing data, and performing time zone correction and data time sequence alignment on the multi-source heterogeneous sensing data, thereby obtaining a time-standardized multi-source data set;

[0013] Step S12: performing initial dynamic label analysis on each type of physical quantity in the time-standardized multi-source data set, thereby obtaining initial dynamic label analysis data;

[0014] Step S13: performing hierarchical signal decoupling processing on the initial dynamic label analysis data, extracting a principal component dynamic signal subset, and thereby obtaining hierarchical decoupling signal data;

[0015] Step S14: performing multi-dimensional dynamic signal correlation reconstruction using the hierarchical decoupling signal data, thereby obtaining multi-dimensional signal period correlation data;

[0016] Step S15: performing dynamic label fine-tuning on the multi-dimensional signal period correlation data, and integrating the dynamic label fine-tuning result, thereby obtaining a multi-source pre-processing data set.

[0017] Optionally, step S12 comprises:

[0018] Step S121: obtaining industrial automatic equipment running structure data, including equipment running period, equipment structure data, and process flow data;

[0019] Step S122: dividing the time-standardized multi-source data set into a physical feature space within the equipment running period, and labeling the division result in a feature space, thereby obtaining a multi-physical quantity initial label vector data;

[0020] Step S123: identifying equipment running events from the time-standardized multi-source data set, thereby obtaining equipment running event division data;

[0021] Step S124: Work condition event label screening is performed on the combined multi-physical quantity initial label vector data and the equipment operation event division data, to obtain multi-physical quantity event screening label data;

[0022] Step S125: Time sequence reconstruction is performed on the multi-physical quantity event screening label data, and label semantics in the time sequence reconstruction result are enhanced in combination with the equipment structure data and the process flow data, to obtain dynamic semantic label vector data;

[0023] Step S126: Label integration and standardization processing are performed based on the dynamic semantic label vector data, to obtain dynamic label initial analysis data.

[0024] Optionally, the step S14 comprises:

[0025] Step S141: Multi-dimensional signal grouping processing is performed based on the decoupling levels of different physical quantities and signals in the hierarchical decoupling signal data, and high-dimensional feature space mapping of each physical quantity is established, to obtain multi-dimensional signal feature mapping data;

[0026] Step S142: Dynamic alignment is performed on the time relationship between each physical quantity in the multi-dimensional signal feature mapping data, to identify the dynamic coupling relationship and synchronicity between different physical quantities, to obtain multi-dimensional signal dynamic time alignment data;

[0027] Step S143: Multi-modal relationship modeling is performed on the multi-dimensional signals in the multi-dimensional signal dynamic time alignment data, to obtain multi-modal relationship data of the multi-dimensional signals;

[0028] Step S144: Multi-dimensional signal periodic pattern detection is performed based on the multi-modal relationship data of the multi-dimensional signals, to obtain multi-dimensional signal periodic pattern data;

[0029] Step S145: According to the similarity and difference between the signal periodicities in the multi-dimensional signal periodic pattern data, the periodic signal features of each periodic pattern are clustered and weighted, to obtain multi-dimensional signal periodic correlation data.

[0030] Optionally, the step S2 comprises:

[0031] Step S21: Hierarchical analysis is performed on the equipment structure data in the multi-source pre-processing data set, to extract the topological features of the equipment, to obtain equipment structure topological feature data;

[0032] Step S22: Process flow modeling and relationship reasoning are performed based on the process flow data in the multi-source pre-processing data set, to obtain process flow feature data;

[0033] Step S23: Multi-level graph embedding fusion is performed on the equipment structure topological feature data and the process flow feature data, to obtain an industrial process structure model;

[0034] Step S24: semantic entity division is performed on the equipment running events and process labels in the multi-source preprocessed data set, and equipment running semantic entity data is generated;

[0035] Step S25: the industrial process structure model is associated with the equipment running semantic entity data in multiple levels to enhance the dynamic semantic association network data.

[0036] Optionally, the multi-task feature vector fusion in step S3 includes:

[0037] According to the dynamic semantic association network, high-correlation semantic entities of the mechanical arm end effector and the conveyor belt are extracted, and mechanical arm-related semantic entities and conveyor belt-related semantic entities are obtained;

[0038] The physical quantity dynamic features of the mechanical arm are extracted from the multi-source preprocessed data set by using the mechanical arm-related semantic entities, and mechanical arm physical quantity dynamic data are obtained;

[0039] The motion sensing features of the conveyor belt are extracted from the multi-source preprocessed data set by using the conveyor belt-related semantic entities, and conveyor belt physical quantity dynamic data are obtained;

[0040] The mechanical arm physical quantity dynamic data and the conveyor belt physical quantity dynamic data are time-aligned, and the time-aligned results are effectively dynamically screened to obtain an effective multi-dimensional physical quantity feature vector;

[0041] A multi-task index set of the industrial automatic equipment is obtained, and the effective multi-dimensional physical quantity feature vector is divided and aggregated in layers according to the characteristics of the multi-task index set, to obtain a multi-task layered feature vector set;

[0042] The multi-task layered feature vector set is multi-dimensionally embedded and fused to integrate the feature representations of different tasks, and a preliminary fused multi-task feature vector is generated.

[0043] Optionally, step S4 includes:

[0044] Step S41: the multi-modal unified feature vector is classified and disassembled according to the physical quantity categories, a multi-dimensional dynamic physical quantity vector set is constructed, and each node in the industrial process structure model is dynamically behavior vector mapped by using the multi-dimensional dynamic physical quantity vector set, to obtain an industrial process dynamic behavior graph;

[0045] Step S42: the node connection relationship in the industrial process dynamic behavior graph is modeled in each edge dynamic behavior relationship, to obtain an industrial process dynamic relationship matrix;

[0046] Step S43: the industrial process dynamic relationship matrix is used to perform multi-dimensional dynamic topology optimization on the industrial process dynamic behavior graph, to obtain an optimized industrial process dynamic behavior graph;

[0047] Step S44: Based on the optimized industrial process dynamic behavior graph, a multi-dimensional relationship modeling of structured-unstructured features is performed to obtain a multi-dimensional dynamic relationship network;

[0048] Step S45: According to the multi-dimensional dynamic relationship network, a graph structure learning is performed to obtain a multi-dimensional dynamic model of the industrial process.

[0049] Optionally, the multi-dimensional abnormal outlier index of the equipment in step S5 comprises:

[0050] A running state deviation degree of the running state prediction result is quantified to obtain a deviation degree tensor;

[0051] In each time step of 3 minutes, residual features of the physical quantity of the equipment in the deviation degree tensor are extracted to generate a state residual matrix;

[0052] Mean, variance and extreme value fluctuation analysis are respectively performed on each row and each column in the state residual matrix, and the fluctuation analysis results are organized into a multi-dimensional outlier detection index vector;

[0053] In combination with the fluctuation interval and threshold range of each physical quantity of each equipment in the multi-dimensional dynamic model of the industrial process, a dynamic adaptive discrimination condition is set, and the dynamic adaptive discrimination condition is used for abnormal outlier behavior detection on the multi-dimensional outlier detection index vector to obtain an abnormal outlier index point set;

[0054] The abnormal outlier index point set in each time step is summarized to construct a multi-dimensional abnormal outlier index set of the equipment.

[0055] Optionally, the adaptive parameter reconstruction in step S6 comprises:

[0056] Abnormal frequent parameter points in the multi-dimensional abnormal outlier index set of the equipment are extracted, and an abnormal feature distribution vector is constructed;

[0057] The abnormal feature distribution vector is compared with the state parameters of the dynamic model of the industrial process through difference analysis, and a reconstruction gain coefficient set is calculated;

[0058] A device control log of the device control unit is called, and a control parameter group is screened in combination with the multi-source pre-processing data set;

[0059] The control parameter group is weighted and offset by using the reconstruction gain coefficient set to generate a reconstruction parameter vector set;

[0060] The reconstruction parameter vector set is converted into an instruction sequence, and is aligned with the device control channel protocol specification of the device control unit to generate a device control instruction sequence.

[0061] The application realizes comprehensive and dynamic capture of equipment operation state and process parameters through efficient fusion of multi-source heterogeneous data, significantly improves the completeness and real-time performance of data, and avoids the control blind area caused by data lag or loss in traditional systems. Through hierarchical analysis of equipment structure and process flow, a multi-dimensional industrial process structure model is constructed, and the accurate expression of the relationship between equipment components is realized, laying a solid foundation for dynamic mapping and relationship modeling of equipment behavior. The vectorization embedding of high-dimensional semantic entities and the fusion of multi-level semantics and atlas relationship effectively enhance the expression ability of the semantic association network, improve the depth of understanding of the equipment operation semantics, and enable the dynamic semantic association network to accurately reflect the equipment state changes and their internal correlations. Based on the time alignment and screening of dynamic physical quantity characteristics, the effectiveness and representativeness of the physical quantity characteristic vector are ensured, and the interference of redundant data on the analysis results is avoided, and then through hierarchical aggregation and multi-dimensional embedding fusion of multi-task index set, the feature information of different task dimensions is comprehensively integrated, and the fusion degree and accuracy of feature representation are improved. The construction of industrial process dynamic behavior graph and its dynamic relationship matrix, combined with multi-dimensional dynamic topology optimization, effectively improves the dynamic performance of nodes and edges in the industrial process, making the equipment behavior mapping more detailed and accurate, thereby providing a solid data foundation for subsequent multi-dimensional dynamic relationship network construction. The multi-dimensional dynamic relationship network structure realizes the dynamic evolution description and relationship prediction of equipment behavior through multi-dimensional relationship modeling of structured and unstructured features, enhancing the adaptability and prediction ability of the model. The tensorization of the running state deviation degree and the residual feature extraction in the short time window enable the equipment state anomaly to be quickly identified and located, combined with multi-dimensional outlier detection indicators and dynamic adaptive discrimination conditions, effectively improving the detection accuracy and real-time response ability of abnormal outlier behavior. The extraction of abnormal frequent parameter points and the construction of abnormal feature distribution vector, through the difference comparison with the state parameters of the industrial process dynamic model, the calculated reconstruction gain coefficient set provides a scientific basis for the dynamic adjustment of control parameters, which helps to accurately identify the optimization direction of equipment control. The joint screening of equipment control log and multi-source preprocessed data ensures the diversity and effectiveness of the control parameter set, and the weighted offset of the reconstruction gain coefficient to the control parameter realizes targeted regulation, enhancing the adaptability and execution effect of the control instruction. The alignment of the generated equipment control instruction sequence with the control channel protocol specification ensures the correct transmission and execution of the instruction, improving the digital control precision and real-time performance of the industrial automatic equipment. By reasonably setting the parameters in each step, such as time window length, threshold range and vector dimension, not only the stability and robustness of data processing and feature extraction are guaranteed, but also the system response speed and control efficiency are optimized, thereby improving the performance and intelligent level of the digital visual control system of the industrial automatic equipment as a whole.

[0062] Optionally, the present specification also provides an industrial automatic equipment digital visualization control system for executing the industrial automatic equipment digital visualization control method as described above, the industrial automatic equipment digital visualization control system comprising:

[0063] A data preprocessing module is configured to acquire multi-source heterogeneous sensing data, and perform dynamic label analysis and signal decoupling according to the multi-source heterogeneous sensing data, to obtain a multi-source preprocessed data set.

[0064] A semantic association module is configured to construct an industrial process structure model according to the multi-source preprocessed data set, and perform multi-level semantic association enhancement between the industrial process structure model and semantic entities in the multi-source preprocessed data set, to obtain a dynamic semantic association network.

[0065] A mechanical arm-conveyor belt association module is configured to perform multi-task feature vector fusion based on the dynamic semantic association network and in combination with motion physical quantity features of an end effector of a mechanical arm and a conveyor belt in the multi-source preprocessed data set, to obtain a preliminary fused multi-task feature vector, and perform feature vector weighted regression on the preliminary fused multi-task feature vector, to obtain a multi-modal unified feature vector.

[0066] A device behavior mapping module is configured to perform device multi-dimensional dynamic behavior mapping using the multi-modal unified feature vector, to obtain an industrial process multi-dimensional dynamic model.

[0067] A device anomaly outlier analysis module is configured to acquire real-time multi-source heterogeneous sensing data, perform device running state prediction on the real-time multi-source heterogeneous sensing data through the industrial process multi-dimensional dynamic model, and identify device multi-dimensional anomaly outlier indexes according to the running state prediction result.

[0068] A control parameter reconstruction module is configured to perform adaptive parameter reconstruction on the device multi-dimensional anomaly outlier indexes, to obtain a device control instruction sequence, and deliver the device control instruction sequence to a corresponding device control unit.

[0069] The industrial automatic equipment digital visualization control system of the present application can implement any one of the industrial automatic equipment digital visualization control methods of the present application, and is configured to jointly operate and transmit signals between the modules, to complete the industrial automatic equipment digital visualization control method, and the modules in the system cooperate with each other, thereby optimizing the system response speed and control efficiency, and improving the performance and intelligent level of the industrial automatic equipment digital visualization control system as a whole. BRIEF DESCRIPTION OF DRAWINGS

[0070] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-restrictive embodiments made with reference to the accompanying drawings:

[0071] Fig. 1The step flow diagram of the industrial automatic equipment digital visualization control method of the present application is shown in the figure;

[0072] Fig. 2 The detailed step flow diagram of step S1 in the present application is shown in the figure;

[0073] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0074] The technical method of the present application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0075] In addition, the accompanying drawings are only schematic diagrams of the present application, and are not necessarily drawn to scale. The same reference signs in the drawings represent the same or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. The functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0076] It should be understood that although the terms "first", "second" and the like can be used herein to describe various elements, these elements should not be limited by these terms. These terms are only used to distinguish one element from another. For example, without departing from the scope of the exemplary embodiments, a first element can be called a second element, and similarly a second element can be called a first element. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0077] To achieve the above-mentioned purpose, please refer to Figs. 1-2 The present application provides an industrial automatic equipment digital visualization control method, which comprises the following steps:

[0078] Step S1: acquiring multi-source heterogeneous sensing data, and performing dynamic label analysis and signal decoupling according to the multi-source heterogeneous sensing data to obtain a multi-source preprocessed data set;

[0079] In this embodiment, multiple sensor devices are deployed on the intelligent automated assembly line, including infrared temperature sensors, motor current sensors, six-axis accelerometers, industrial cameras, and laser range finders, to collect heterogeneous data such as temperature, current, vibration images, and spatial position information covering the entire production process. To unify the data structure, the sampling frequency of all device signals is reset to 100 Hz, and synchronization is achieved using timestamp alignment. After collection, the identified assembly behaviors in the images are matched one by one with the acceleration, displacement, and current waveform segments using a pre-set label dictionary. For the interference components in the acceleration and current signals, different signals are decoupled and processed to remove signal aliasing caused by motion coupling using frequency domain filtering and envelope detection methods. After processing, all signals are converted into standard data tables based on time steps, with each row corresponding to a time point and each column corresponding to an analyzed signal type, and the current behavior label is attached. Finally, a unified structure and multi-source synchronized labeled dataset is generated as the basis for subsequent modeling.

[0080] Step S2: Construct an industrial process structure model based on the multi-source preprocessed dataset, and enhance the multi-level semantic association between the industrial process structure model and the semantic entities in the multi-source preprocessed dataset to obtain a dynamic semantic association network;

[0081] In this embodiment, based on the obtained preprocessed dataset, the physical layout and control process documents of the assembly line are combined to construct an industrial process structure model. The model is described in the form of a directed node graph, with each node representing an industrial unit or action step, such as "robot picking", "conveyor transmission", and "vision recognition", and the edges representing logical or data flow connection relationships. After construction, the node graph is associated with semantic entities such as device categories and action labels in the preprocessed dataset. For example, by checking the sensor number and device type during the "robot picking" time period in the label, the node "robot picking" is bound with three-axis acceleration, end position, and grasping image frames. Further, to enhance the semantic association level, a multi-level semantic enhancement mechanism is introduced to add a device attribute layer (such as device model and motion range) and a task semantic layer (such as action purpose and operation category) to the structure graph, forming a three-layer semantic graph structure. The final generated dynamic semantic association network contains both production process order and device attributes and operation semantics, which can support subsequent feature fusion and state modeling.

[0082] Step S3: Based on the dynamic semantic association network, and combined with the motion physical quantity features of the robot end effector and the conveyor belt in the multi-source preprocessed dataset, perform feature vector fusion to obtain a preliminary fused multi-task feature vector; perform feature vector weighted regression on the preliminary fused multi-task feature vector to obtain a multi-modal unified feature vector;

[0083] In this embodiment, on the basis of the constructed dynamic semantic association network, the key physical characteristics of the end of the mechanical arm in each operation period, such as speed, position change, attitude angle, and current load, are extracted, and the material type and color region distribution in the image frame in the associated node are extracted to form a preliminary multi-task feature vector. After time alignment of the information from different data sources and different dimensions, the information is stored in the form of a fixed-length vector, and each record represents an operation period. On this basis, in order to improve the data expression ability, different weighting coefficients are given to different feature components according to the weights of different nodes in the semantic network. The weighting process refers to the statistical distribution of the device load change trend and the image recognition confidence in the historical operation, and the key influencing factors are retained with a higher weight. The weighted feature vector is then uniformly represented through a set of linear regression mapping relationships, and the output is a fixed-length multi-modal fusion feature vector, each vector carrying the device physical state and semantic information in the current period, which is used for behavior mapping and state recognition.

[0084] Step S4: using the multi-modal unified feature vector to perform device multi-dimensional dynamic behavior mapping, thereby obtaining an industrial process multi-dimensional dynamic model;

[0085] In this embodiment, the obtained multi-modal unified feature vector is used in combination with the original industrial process structure model to dynamically map the actual running track of each node action in different production batches. In the mapping process, the dynamic behavior change of the corresponding node is analyzed according to the semantic action and time period carried by each feature vector. For example, if the "mechanical arm grabbing" shows different displacement distances and clamping times in different batches of materials, the system will automatically map this node as a multi-dimensional action description with "material type" and "time characteristics". On the whole process structure diagram, a multi-dimensional attribute set is established for each node, including average execution time, standard deviation, end displacement range, environmental temperature influence factor, etc. At the same time, a statistical chart is used to represent the change trend with time or batch, and finally a dynamically evolving industrial process behavior model is formed. The model can be used to reverse the operation change source, support device state simulation and process optimization.

[0086] Step S5: acquiring real-time multi-source heterogeneous sensing data; predicting the device running state of the real-time multi-source heterogeneous sensing data through the industrial process multi-dimensional dynamic model, and identifying the device multi-dimensional abnormal outlier index according to the running state prediction result;

[0087] In this embodiment, during real-time operation, data from all deployed sensors is continuously collected, and structured input records are generated for each time point. These real-time data inputs are matched in the generated multi-dimensional dynamic model. The model internally predicts the device operating state based on pre-learned feature distribution and behavior mapping for the current data. For example, when an abnormal large vibration amplitude is detected during the "conveyor transmission" process, and the temperature rise speed exceeds the historical 90th percentile, the model will mark this action as "deviating from the normal range", and record the deviation dimension. During the prediction process, multiple indicators are output for each operation node, including operation execution time prediction value, end displacement prediction range, and environmental influence tolerance threshold. The system compares the current observation value with the prediction result, and if the deviation exceeds the set tolerance, it is recorded as an abnormal outlier. Abnormal indicators include but are not limited to "high temperature", "action delay", and "end displacement anomaly". Each abnormality records the corresponding time period, affected device number, and risk level, which are used for subsequent control adjustment.

[0088] Step S6: Self-adaptive parameter reconstruction of multi-dimensional abnormal outlier indicators of the device is performed to obtain a device control instruction sequence, which is then sent to a corresponding device control unit.

[0089] In this embodiment, for the identified multi-dimensional abnormal outlier indicators, the system performs self-adaptive parameter adjustment based on predefined control strategy templates. Taking the "abnormal mechanical arm clamping force" indicator as an example, the system recalculates the acceleration, closing angle, and deceleration parameters in the control instruction by referring to the current current overload degree, clamping displacement range, and feedback position deviation. The specific control parameters are dynamically generated by consulting the device performance table and safety threshold range, and compared with the parameter difference in the last normal action to ensure that the adjustment range is within the controllable interval. The generated control instruction sequence is in a standard industrial control format and is sent to the corresponding device execution unit through a PLC communication interface. Each control instruction includes target position, motion speed, maximum tolerance, emergency stop threshold, etc. Before execution, the system also simulates the action path to verify safety.

[0090] Optionally, step S1 includes:

[0091] Step S11: Collect multi-source heterogeneous sensor data, and perform time zone correction and data time sequence alignment on the multi-source heterogeneous sensor data to obtain a time-standardized multi-source data set;

[0092] In this embodiment, on a smart manufacturing production line with complex equipment integration, a multi-source heterogeneous collection including thermocouple temperature sensors, torque sensors, three-dimensional image acquisition devices, environmental humidity detectors, and multiple sets of industrial voltage and current monitoring probes is deployed. Due to different manufacturers, there are multiple time zone deviations in the timestamp of data recording, and even some devices use local clocks, which makes the data unable to be directly aligned. To unify the data time standard, first, the clocks of various devices are calibrated through the NTP network time protocol to coordinate the GMT+8 as the reference time zone. Then, according to the synchronization frame interval setting of each type of sensing device, such as 10 frames per second for temperature sensors and 50 frames per second for current monitoring, all data are mapped to a unified 100Hz time axis using interpolation resampling. For time drift caused by network delay or device failure, a time offset tolerance of ±20ms is set, and the part exceeding the tolerance will be automatically marked as missing data. After the above operations, a standardized time series data table with millisecond time steps is generated, which contains the sensor source number, physical quantity type, numerical value, and corresponding timestamp fields, forming a unified time standardized multi-source data set.

[0093] Step S12: Dynamic label initial analysis of each type of physical quantity in the time standardized multi-source data set is performed to obtain dynamic label initial analysis data;

[0094] In this embodiment, the time standardized multi-source data set is further processed, and each type of physical quantity data is preliminarily dynamically labeled and analyzed according to its sensor source and industrial behavior background. In the specific implementation process, a dynamic label library is set, which contains initial behavior description labels such as "moving action start", "temperature rise trigger", "assembly contact occurrence", etc., and each type of sensor is configured with its typical numerical value interval in different operation stages. Taking the data of the end effector of the mechanical arm as an example, if the acceleration change is above 0.3g and lasts for more than 0.5 seconds, accompanied by a change in clamping torque of more than 10Nm, it is preliminarily marked as "object taking action occurs"; similarly, if the temperature sensor data continuously rises by more than 5 degrees Celsius within 5 seconds, it corresponds to "heat source intervention process starts". The above label analysis logic is embedded in the data processing flow by setting conditions, and each data record is labeled with an initial event label field. At the same time, in order to facilitate subsequent tracking, each type of label is bound to a unique event number and trigger condition record. The initial analysis data finally output adds a label column to the original time series data structure, so that each time step has corresponding preliminary semantic recognition information.

[0095] Step S13: Hierarchical signal decoupling processing is performed on the dynamic label initial analysis data to extract a principal component dynamic signal subset, thereby obtaining hierarchical decoupled signal data;

[0096] In this embodiment, the hierarchical signal decoupling process is further performed on the dataset after the preliminary label resolution is completed. First, the data is logically layered according to the sensor type and the corresponding device function, for example, acceleration, current, position displacement related to mechanical action are grouped into a physical motion layer, temperature, humidity are grouped into an environmental state layer, and image recognition confidence, color distribution, etc. are grouped into a sensory knowledge recognition layer. In each layer, identify the signal channels with dramatic numerical fluctuations and obvious noise superposition, perform variance analysis on each data channel in the sliding window based on the time-varying characteristics of the signal, and use threshold screening to retain the main components, for example, only retain the first three signal channels that account for more than 85% of the change rate. Taking the acceleration sensor as an example, only two direction data with the largest contribution to change are retained from the original X, Y, Z three-dimensional data, and further key features such as extreme points, change rates, peak spacing are extracted to construct a principal component set. The principal component set extracted in each layer is consistent with the original time axis, and the output is a hierarchical decoupled signal data with unified structure. The data in each layer is identified as the main signal, the secondary signal and the extraction time window in the form of multiple columns, constructing a high-reliability input for correlation modeling.

[0097] Step S14: Perform multi-dimensional dynamic signal correlation reconstruction using hierarchical decoupled signal data to obtain multi-dimensional signal periodic correlation data.

[0098] In this embodiment, the output hierarchical decoupled signal data is input into the multi-dimensional periodic modeling process, and the periodic correlation between different signals is reconstructed according to the device operation period and historical behavior label. First, local period detection is performed on each main signal sequence, and within a set observation window length (for example, every 60 seconds), key event time points are extracted as reference anchor points, such as clamping start, current peak, temperature inflection point, etc., and an event time vector is constructed for these anchor points. Then, combined with the time alignment positions between multiple signals, a periodic signal mapping table is constructed, where the rows represent the reference period, the columns represent the signal types, and the cells record the key event occurrence positions and standard deviations of the signals in the current period. In this way, the start time point in the acceleration signal, the peak point in the current signal, and the rising slope of the temperature signal are corresponded to form a cross-signal dimension periodic behavior alignment data structure. The multi-dimensional signal periodic correlation data finally reconstructed is represented in a three-dimensional matrix, with dimensions of period number, signal type, and key behavior index, serving as a support data structure for subsequent label fine-tuning.

[0099] Step S15: Perform dynamic label fine-tuning on the multi-dimensional signal periodic correlation data, and integrate the dynamic label fine-tuning results to obtain a multi-source preprocessed dataset.

[0100] In this embodiment, based on the constructed period correlation data structure, the dynamic labels obtained by early analysis are fine-tuned. First, the time difference of each key behavior event in the period and the offset of the initial label are analyzed. If the offset is within the set tolerance range (such as ±2 seconds), it is marked as “valid match”, otherwise the label trigger point is adjusted to the nearest key behavior anchor point. Second, for the case where there are multiple label conflicts, according to the period behavior consistency principle, only the label with the highest degree of agreement between signal change and label logic in the period is retained. For example, if “transportation is completed” and “clamping fails” are marked in a period, but the clamping torque curve shows that the clamping is not up to standard, then “clamping fails” is retained and “transportation is completed” is removed. After fine-tuning, all the modified labels are checked, and the discontinuous label intervals are removed and the missing label segments are filled. Finally, the adjusted labels and the original standardized time series data are re-integrated, and the complete multi-source preprocessed data set is output.

[0101] Optionally, step S12 comprises:

[0102] Step S121: Obtain industrial automation equipment operation structure data, including equipment operation period, equipment structure data and process flow data;

[0103] In this embodiment, in the automated hydraulic assembly unit, the operation structure data including equipment operation period (for example, 120 seconds for each working cycle), key component structure (such as the position information and connection relationship of the conveyor belt, the press head, the positioning base, and the loading and unloading mechanical arm), and corresponding process flow steps (such as clamping-pressing-releasing-resetting) are derived from the central industrial control. The equipment operation period is obtained through PLC control logic and stored in the form of period number and start and end time stamps, for example, period C001 starts at 13:00:00 and ends at 13:02:00. The equipment structure data is represented in a nested JSON structure, and each equipment element contains its number, function, belonging assembly level, and connection topology information. The process flow data defines the arrangement order of each step in time and the involved equipment elements, for example, the pressing step binds the equipment to the main press head, and the corresponding time period in the period is 30-55 seconds. Through data integration tools, the three types of information are unified to build a multi-dimensional structure description table, which serves as reference data input for subsequent data space division and event recognition.

[0104] Step S122: Physically divide the time-standardized multi-source data set within the equipment operation period, and label the division results in the feature space to obtain the multi-physical quantity initial labeling vector data;

[0105] In this embodiment, the multi-source sensing data that has completed time standardization processing is taken as the time window division basis, and the data subset corresponding to each cycle is extracted. Within each cycle, the data channels are divided into different physical feature spaces, such as temperature, pressure, current, displacement, etc., according to the element division in the device structure. Taking hydraulic pressure as an example, the pressure data is closely related to the action of the hydraulic cylinder, so the pressure sensor data is mapped with the component structure label; similarly, the temperature sensor is matched with the heat treatment process part. By setting the binding relationship table of device elements and sensor channels, the mapping from raw data to feature space is realized. In the labeling process, each piece of data is attached with a spatial label (such as "heat treatment area temperature signal", "positioning base displacement signal") according to the process sequence in the device running cycle, and the output is a multi-physical quantity initial labeling vector data, which is a two-dimensional vector matrix. Each row represents a time point, including timestamp, physical quantity value, spatial identifier and cycle number, which is convenient for subsequent working condition event recognition processing.

[0106] Step S123: device running event recognition is performed on the time-standardized multi-source data set to obtain device running event division data;

[0107] In this embodiment, the standardized multi-source data set after cycle division is used to identify the operation events in each cycle in combination with the process flow time period definition in the device running structure data. The identification operation judges the event boundary by analyzing the mutation characteristics of the key physical quantity on the time axis. For example, in the press fitting step, if the torque sensor data rapidly rises and stably maintains above a set pressure (such as 2.0 MPa) for more than 5 seconds, then this piece of data is marked as a "press fitting action segment". In addition, if the acceleration of the mechanical arm jumps from zero to more than 0.4g in a short time, and the corresponding displacement data changes more than 30mm, then it is identified as a "clamping or releasing action". After mapping these event determination rules to the data time axis, the device running event division data is output, which is structured as event number, start and end timestamp, event type and corresponding physical quantity channel set, and is bound with the cycle number, realizing the alignment of data and behavior events, which is used for subsequent multi-physical quantity event screening.

[0108] Step S124: working condition event label screening is performed in combination with the multi-physical quantity initial labeling vector data and the device running event division data to obtain multi-physical quantity event screening label data;

[0109] In this embodiment, the obtained multi-physical quantity initial labeling vector data is fused with the equipment operation event division data identified in step S123, and the data segments involved in the key working condition events are screened based on the overlapping of the time intervals and the space identifiers of the two. Specifically, if a certain segment of temperature data is labeled as "heat treatment zone", and its time period intersects with the event label "heating start", the data is retained as the heat treatment event related signal. Similarly, in the assembly process, if the torque data channel and the "clamping execution" event time period match more than 80%, the signal is retained and an event label is added. The data screened in this way retains the signal fragments that are strongly related to the actual event semantics, avoids redundant data interference, and forms multi-physical quantity event screening label data with the structure of

timestamp, channel type, space identifier, event number, event type

[0110] Step S125: Time sequence reconstruction is performed on the multi-physical quantity event screening label data, and the label semantics in the time sequence reconstruction result are enhanced in combination with the equipment structure data and the process flow data, to obtain dynamic semantic label vector data.

[0111] In this embodiment, the label time sequence reconstruction is performed across cycles for the multi-physical quantity event screening label data. In the operation, all cycle numbers are first uniformly expanded to a complete time axis, and similar events in different cycles are aligned through timestamp normalization operation, for example, "pressing action" is mapped to a standard position P2 in each cycle. Then, in combination with the trigger relationship between components in the equipment structure data and the execution order in the process flow definition, logical dependency relationships are added in the label sequence, for example, "clamping start" must be before "pressing start" and immediately after it, otherwise it is marked as a logical anomaly. When constructing the time sequence label vector, each event corresponds to a vector with a length of N, which includes event number, time position, related equipment element identifier and logical association relationship description field. Finally, a semantic enhanced label network is established through graph structure mapping method, and the nodes are label events and the edges are equipment structure connections or process dependency paths. The semantic enhanced label data structure can be derived as a set of three-dimensional tensors: [event number x feature dimension x time step], forming dynamic semantic label vector data.

[0112] Step S126: Based on the dynamic semantic label vector data, label integration and standardization processing are performed to obtain dynamic label initial analysis data.

[0113] In this embodiment, the constructed dynamic semantic label vector data is subjected to integration and standardization operation. Firstly, the semantic redundancy between labels is processed, for example, if the time overlap rate of "assembly action completion" and "clamping release completion" is greater than 90%, they are merged into a unified label "process stage switching completion"; secondly, the label expression format is uniformly coded, for example, all labels are converted into a fixed five-tuple form: [label code, event start time, event duration, involved device number, process stage number]. In order to further improve the universality of the label in different fields, the event type and process number coding system is defined according to the IEC62264 industrial semantic specification, and the original label code is converted into a standard number, such as "clamping action" is mapped to ACT-01, and "pressing completion" is mapped to PROC-04. After processing, the output of the dynamic label initial analysis data is a structured label data table, each record represents a standardized label event, has consistent expression ability across devices, cycles and semantic layers, and can be directly connected to process modeling and behavior prediction as input label data.

[0114] Optionally, step S14 comprises:

[0115] Step S141: based on the decoupling level of different physical quantities and signals in the hierarchical decoupling signal data, multi-dimensional signal grouping processing is performed, and high-dimensional feature space mapping of each physical quantity is established, to obtain multi-dimensional signal feature mapping data;

[0116] In this embodiment, for the obtained hierarchical decoupling signal data, firstly, according to the type of physical quantity, a first grouping is performed, including pressure, displacement, current, vibration and temperature signal channels; and then, according to the level number marked in the decoupling processing stage, a second grouping is performed, for example, the first layer of the pressure signal is the main cylinder data, and the second layer is the auxiliary circuit data. On this basis, a high-dimensional feature space mapping is constructed for each type of physical quantity group, for example, the original value, the change rate, the fluctuation amplitude and the peak-valley interval time of the pressure signal are taken as the feature dimensions to form a four-dimensional feature vector. Taking the main cylinder pressure channel P1 as an example, the signal sequence in the third to sixth second time period is constructed into a feature vector of [3.1 MPa, 0.4 MPa / s, 0.5 MPa, 1.2 s], which is further expanded to a local feature block of 10x4 in the time sequence window. All signals are stacked to construct a tensor data with a structure of [NxMxT], wherein N is the number of physical quantity categories, M is the number of feature dimensions of each category, and T is the number of time sliding windows. The multi-dimensional signal feature mapping data is output as the input basis for subsequent coupling analysis.

[0117] Step S142: dynamically aligning the time relationship between each physical quantity in the multi-dimensional signal feature mapping data, identifying the dynamic coupling relationship and synchronicity between different physical quantities, to obtain multi-dimensional signal dynamic time alignment data;

[0118] In this embodiment, the generated multi-dimensional signal feature mapping data is subjected to time relationship alignment operation. By unifying the global time stamp of sampling signals from different clock sources, and dynamically compensating for channels with sampling offset, for example, temperature channel delay is obvious, and the average value of the two time points before and after is used for interpolation correction. The time window length is selected as 5 seconds, and the sliding step is 0.5 seconds. In each window, the cooperative rising or falling section between each signal group is calculated. For example, whether the displacement signal rising period is synchronized with the current signal mutation, and whether the coincidence rate exceeds 70%. If the condition is met, it is determined as a dynamic coupling relationship. In this way, it can be identified that in the actual assembly process, the behavior of some physical quantities is strongly dependent on the triggering of another physical quantity, for example, the chain reaction relationship of “displacement→current→vibration”. The output is multi-dimensional signal dynamic time alignment data, including signal pair number, time synchronization segment start and end point, similarity index and other structure fields.

[0119] Step S143: Multi-modal relationship modeling of multi-dimensional signals in the multi-dimensional signal dynamic time alignment data is performed to obtain multi-dimensional signal multi-modal relationship data;

[0120] In this embodiment, based on the time-aligned signal data, the interaction relationship between different physical quantities is further structured modeled. Taking each type of signal as a node, the synchronized segment identified in the time alignment is used as a connection edge to establish a relationship graph. Node attributes such as physical quantity category and signal intensity statistical value are added to the graph, and edge attributes include average synchronization rate and maximum synchronization duration. For example, the current and pressure are synchronized to rise in multiple cycles, establishing a strong edge relationship, and the edge attribute is set to {sync_ratio: 0.82, duration_avg: 3.2s}. In the modeling process, the time structure is included in the weight calculation between nodes, and signals that maintain a synchronized relationship in consecutive cycles are given a higher coupling level. The generated multi-dimensional signal multi-modal relationship data is expressed in a graph structure file, with the structure G=(V, E), where V is a set of physical quantity nodes, and E is a set of relationship edges. It has rich time series relationship information and cross-physical quantity coupling structure, and is suitable for subsequent periodic detection and clustering identification.

[0121] Step S144: Multi-dimensional signal periodic pattern detection based on multi-dimensional signal multi-modal relationship data is performed to obtain multi-dimensional signal periodic pattern data;

[0122] In this embodiment, the established multi-modal relationship data is used to detect whether there is periodic resonance between signals. In each strongly connected signal pair, a plurality of continuous running period signal segments (for example, 30 seconds per period, and 5 consecutive period data) are intercepted, and the repetition rate of the key features in each period is calculated, such as whether the "rise-stable-fall" three-section structure in the current waveform is repeated in each period, and mode matching is performed with the coupled pressure signal. The window difference method is used, and the time deviation tolerance is set to ±1 second. The event time node offset in different periods is compared. If the error of most periods is within the tolerance, it is determined that the period is consistent. Finally, the identified periodic signals are combined and output as multi-dimensional signal period pattern data, each record containing signal number, period template number, matching period number, period length, average offset, and other information, facilitating analysis and aggregation of similarity in the next step.

[0123] Step S145: According to the similarity and difference between the periodicity of the multi-dimensional signal period pattern data, the periodic signal features of each periodicity pattern are clustered and weighted, and the multi-dimensional signal period correlation data is obtained.

[0124] In this embodiment, for the output periodic pattern data, the periodic signal combinations are clustered according to the periodic similarity index. The similarity judgment is mainly based on the proximity of the period length (error not more than 2 seconds), the consistency of the repeated segment time distribution (offset rate less than 10%), and the correspondence of the physical quantity waveform structure (such as whether it contains three peak segments). Taking period templates PM01 and PM02 as examples, if their period lengths are 28.5 seconds and 29 seconds respectively, the signal fluctuation structure difference is small, and they are clustered into one class. After clustering, weights are assigned to each class of periodic signal combinations. The weight evaluation is based on its occurrence frequency in the total sample, average signal strength, and coupling level in the multi-modal graph structure. For example, the occurrence frequency of period class CP01 is 12, accounting for 40% of the total number of periods, and the average coupling degree is 0.85, corresponding to a weight of 0.91. The final output multi-dimensional signal period correlation data structure includes period class number, period template set, signal channel list, clustering similarity score, and normalized weight field, which can directly serve the subsequent dynamic control modeling and prediction process.

[0125] Optionally, step S2 includes:

[0126] Step S21: Hierarchical analysis is performed on the device structure data in the multi-source preprocessed data set to extract the topological features of the device, and device structure topological feature data is obtained.

[0127] In this embodiment, for the device structure data in the multi-source preprocessed data set, first extract its device hierarchical structure list, including main device number, functional component division, interconnection interface type and spatial arrangement coordinates. Take the automatic folding unit in a packaging production line as an example, its structure data contains three nested structures: external rack, motor-driven arm and terminal suction disc. Hierarchical structure analysis is performed, the device structure depth is set to D=3, and is marked as level 0 (rack), level 1 (execution arm) and level 2 (suction disc). Extract the key connection points and transmission paths in each layer, such as the action relationship between "motor—connecting rod—slide rail", and establish the topological relationship matrix T(i,j)=1 indicating that i and j components are physically connected. Finally, the device structure topological feature data is obtained, the format is structure tree (tree root is main device, branch is sub-component) + adjacency matrix (indicates component connection) + coordinate table (indicates device spatial arrangement), the structure is like a triple {Tree, Adj, Coord}, which provides a physical basis for subsequent structure graph modeling.

[0128] Step S22: process modeling and relationship reasoning based on process flow data in the multi-source preprocessed data set, to obtain process flow feature data;

[0129] In this embodiment, process flow data is extracted from the multi-source preprocessed data set, including operation step sequence, process node name, step input and output, and trigger condition information. Take the automatic filling section as an example, the process steps are "bottle positioning→liquid injection→nozzle retraction→bottle conveying", and each step is labeled with trigger signal source and execution device number. Set the maximum parallel degree of the process to 4, and map different process paths to directed process chains, such as process chain P1 containing nodes {N1, N2, N3}, each node is attached with process parameters such as liquid injection volume and injection time. Build a process dependency graph by comparing the trigger dependency relationship between nodes, the graph structure Gf=(N,E), where N is the set of process nodes, E is the set of trigger connections, each edge is attached with delay time, upstream and downstream device binding relationship and other attributes. For example, edge e(i,j) indicates that node Ni is completed after 2 seconds to start node Nj, and the edge attribute is {delay:2s, device_link:DP100→DP103}. The output is process flow feature data, the structure includes node parameter table, dependency graph structure and node scheduling timing information, which has complete execution path restoration capability.

[0130] Step S23: multi-level graph embedding fusion of device structure topological feature data and process flow feature data, to obtain industrial process structure model;

[0131] In this embodiment, the topological features of the equipment structure and the process flow features are structurally fused. First, with the process flow node as the main trunk, the actual equipment number that each process node depends on is found, and its corresponding component path in the topological structure is mapped. For example, the filling node N2 corresponds to the motor DP103 and the adsorption head DP103-A2, and the corresponding path in the equipment topology tree is "host → filling module → end suction head". During the fusion process, each process node is mapped to the central node in the subgraph, and its attached equipment is used as the adjacent node in its graph to construct a nested graph structure Gi = (Vi, Ei), where Vi is the set of equipment subcomponents and process step nodes, and Ei is the process trigger relationship and physical connection edge set. A unified nested coding structure is adopted, and different embedded graphs are encoded using 5-dimensional vectors (representing equipment category, process type, operation time, trigger priority, state variables), and then uniformly embedded into a structural tensor. Where n is the number of fusion nodes and m is the embedding dimension. The final output is an industrial process structure model with complete topology, process logic, and equipment mapping information, supporting dynamic state injection.

[0132] Step S24: performing semantic entity division on the equipment operation events and process labels in the multi-source pre-processed data set to generate equipment operation semantic entity data;

[0133] In this example, semantic segmentation is performed on the equipment operation events and process tags recorded in the preprocessed dataset. For example, a heat sealing process log contains events such as "temperature control start," "heating element current change," and "material strip feed displacement response." The process tag is "preheating → sealing → holding pressure." Each event in the log is parsed into a four-tuple consisting of "action-device-time period-signal type," such as "current fluctuation → heater → 10.5-11.0s → analog quantity." These events are then categorized into corresponding stages based on their temporal position, forming preliminary semantic segments. Six semantic entity categories are set: heating, motion, position, feedback, temperature control, and synchronization, each with a corresponding color label for identification. The resulting semantic entity data structure is as follows: each event contains a tag ID, corresponding equipment number, occurrence time range, physical signal category, and semantic category code. For example: {"EV_A123," "Heater A," [10.5, 11.0], "current signal," "heating"}. The final output is the device operation semantic entity dataset, which has the structure of event sequence + semantic label table + device mapping table, and can serve as the basis for subsequent knowledge network nodes.

[0134] Step S25: Perform multi-level semantic association enhancement on the industrial process structure model and the equipment operation semantic entity data to obtain dynamic semantic association network data.

[0135] In this embodiment, based on the industrial process structure model, the generated semantic entities are mapped to the corresponding process nodes and device nodes in the structure model. The semantic enhancement range is set to all nodes with an adjacent depth of 2 in the structure diagram, i.e. each process node can be associated with its directly controlled devices and its upstream and downstream nodes. Taking the filling node N2 as an example, its control device DP103 can be attached with semantic entities such as "liquid addition", "pressure stabilization", "nozzle retraction", etc. The labels of these semantic entities are fused with event timestamps into three-dimensional vectors: semantic type number + trigger offset + signal feature intensity, which are embedded in the attribute table of each device node of the structure diagram. A semantic mapping matrix S(i,j,k) is established, where i is the structure diagram node number, j is the semantic type, and k is the semantic intensity level. The final dynamic semantic association network data is constructed, which has a multi-layer graph structure G=(V,E,S), V is the set of process and device nodes, E is the control and connection relationship, S is the semantic enhancement matrix, which has the ability of real-time state expression, semantic label tracking and behavior association support. This structure can be used for subsequent behavior mapping and anomaly detection analysis.

[0136] Especially important is that step S25 comprises:

[0137] Step S251: high-dimensional vector embedding representation is performed on each entity in the device operation semantic entity data, to obtain semantic entity embedding vector data;

[0138] In this embodiment, for the generated device operation semantic entity data, feature expansion is performed according to dimensions such as semantic category, trigger time, signal type, device type, and location label, and a high-dimensional semantic embedding vector is constructed. Taking "filling stage - pressurization action - current signal - DP103" as an example, the embedding vector dimension is set to 128, of which the first 20 dimensions are used for semantic category one-hot encoding, the middle 60 dimensions are used for signal change features (such as mean, maximum, and change rate), and the last 48 dimensions are used for mapping embedding of device type and structure position information. The five-tuple information of each semantic entity is processed through vector splicing and normalization to form a semantic entity embedding vector, and the vector matrix form is where n is the number of semantic entities. This vector not only retains the context relationship and physical source information of the entity, but also has adaptability to the semantic expression of nodes in the subsequent structure diagram.

[0139] Step S252: graph representation is performed on the industrial process structure model, each device component and process flow node is represented as a graph node, and the relationship between nodes is represented as an edge, to obtain process flow graph representation data;

[0140] In this embodiment, the industrial process structure model is converted into a graph, with equipment components and process step nodes as nodes in the graph, and the physical connection between equipment, the execution order between processes, and the control signal transmission as edges. Taking a typical filling and capping process line as an example, the node set V = {filling unit, spray head, motor A, capping unit, sensor T}, and the edge set E includes physical connection edges (motor-spray head), control edges (controller-motor), and sequential edges (filling->capping). Each node is attached with embedded attributes, including structural position, function type, belonging process stage, and operation frequency, forming a node attribute matrix where m is the number of nodes and d is the attribute dimension (e.g., 64). The final process graph representation data structure is constructed as G = (V, E, F), which has a complete process structure, component topology, and computable expression basis for process dependence.

[0141] Step S253: Multi-level semantic and graph relationship modeling of semantic entity embedding vector data and process graph representation data is performed to obtain semantic-graph dependence relationship data.

[0142] In this embodiment, the semantic entity embedding vector and the process graph representation data are modeled. First, according to the equipment or process behavior indicated by each semantic entity, its corresponding node in the graph is located, for example, the graph node corresponding to the "capping action" is "capping unit". The attachment relationship is determined by establishing an entity-node mapping table M(i,j), i is the entity number, and j is the node number. Further, according to the similar features in different dimensions of the entity embedding vector E and the graph node attribute matrix F, weighted mapping is performed, for example, the 25th dimension in E represents the signal peak value change rate, and the historical rate of the corresponding node in F is the comparison value, and the dependence edge is established by setting a threshold (e.g., ±5%). Finally, a set of triples R = {(ei, vj, w)} is established, where ei is the entity vector, vj is the graph node, and w is the similarity weight value, constituting the semantic-graph dependence relationship data structure. This structure supports semantic conduction and node semantic expression enhancement, and is an intermediate bridge for building a semantic fusion network.

[0143] Step S254: Multi-level semantic enhancement and fusion are performed using the semantic-graph dependence relationship data to enhance the semantic similarity and correlation between nodes, and generate preliminary dynamic semantic correlation network data.

[0144] In this embodiment, the semantic attributes of the semantic entities are transmitted to the corresponding nodes in the process map by using semantic-map dependency relationship data, and the semantic features are diffused to the upstream and downstream adjacent nodes. First, the direct injection of the entity vector value is performed on the attached node in the first 64-dimensional semantic description part, and the propagation depth is set to 2, that is, each node propagates semantic weight information to the second-order neighbor in the graph. In the process of semantic transmission, a normalized propagation weight scheme is adopted, for example, the part with an entity similarity greater than 0.7 is set as complete semantic fusion, and the part with an entity similarity less than 0.3 is not diffused, and the intermediate value determines the fusion strength according to linear interpolation. Finally, the enhanced representation matrix of the map nodes is formed In the attribute dimension of each node, the semantic feature projection values from multiple semantic entities are included, while the structural adjacency constraint is retained. The preliminary dynamic semantic association network data is composed of the enhanced node set V', the edge set E, and the updated attribute matrix F', which has the characteristics of rich semantic expression and close node relationship.

[0145] Step S255: Adjusting the node weight and smoothing the graph of the preliminary dynamic semantic association network data to obtain the dynamic semantic association network data.

[0146] In this embodiment, on the basis of the preliminary semantic association network, the graph smoothing and node weight adjustment are implemented to improve the overall semantic consistency and structural expression ability. First, the activation value of each node in the semantic dimension is normalized to ensure that the semantic similarity of adjacent nodes will not deviate due to the dimensional difference of the feature quantity. Secondly, through the graph convolution smoothing mechanism, the semantic attributes of the first-order adjacent nodes are input in the form of a weight matrix, for example, the set of adjacent nodes of node i is N(i), and the smoothed node representation is F”(i) = (1 / |N(i)|)∑F'(j), where F'(j) is the semantic feature vector of the adjacent node j. To prevent semantic dilution, a minimum activation value threshold of 0.1 is set for each node, and the semantic dimension below the threshold will be shielded to enhance the expression focus. The finally generated dynamic semantic association network data is network structure G” = (V', E, F”), which has the characteristics of continuous semantic expression, stable propagation ability, and structure-semantic co-expression ability, and is suitable for subsequent behavior modeling or anomaly identification application scenarios.

[0147] Optionally, the multi-task feature vector fusion in step S3 includes:

[0148] According to the dynamic semantic association network, the high-correlation semantic entities of the end effector of the mechanical arm and the conveyor belt are extracted to obtain the mechanical arm-related semantic entities and the conveyor belt-related semantic entities.

[0149] In this embodiment, in the dynamic semantic association network that has been constructed, by traversing the edge weight relationship between the semantic entity node and the "task key label", the semantic path representing the "assembly-handling" and "grabbing-placing" core tasks is focused on query, and the semantic nodes with a semantic similarity greater than 0.82 and a node category of "component" and "component state quantity" are extracted. Finally, 13 mechanical arm related semantic entities are selected, including joint torque, end speed, end displacement, joint current, load force, joint temperature and other physical quantities; 8 conveyor belt related semantic entities are selected, including belt speed, motor torque, conveyor position offset, belt vibration, running current and the like. These semantic entities provide direct indexing for subsequent feature extraction of dynamic physical quantities.

[0150] The physical quantity dynamic features of the mechanical arm are extracted from the multi-source preprocessed data set using the mechanical arm related semantic entities, and the mechanical arm physical quantity dynamic data is obtained.

[0151] In this embodiment, the 13 types of mechanical arm related semantic entities identified above are respectively mapped to the existing data fields in the multi-source preprocessed data set, and through the mapping relationship between the semantic labels and the data channel names, the corresponding fields of the mechanical arm are intercepted in the time standardized data matrix. The original data sampling frequency is set to 10 Hz, and the recording time period is 10 minutes, and finally a two-dimensional time series data matrix with a size of [6000x13] is formed, representing the dynamic change process of 13 types of mechanical arm physical quantities at each time frame. Each type of data is uniformly normalized before extraction, and is denoised by a sliding mean window (window width of 5 frames). The mechanical arm physical quantity dynamic data mainly includes the torque, angular velocity, angular position, current, voltage, temperature, load force, end effector speed and displacement of each joint of the mechanical arm, and these data reflect the motion state and working load of the mechanical arm.

[0152] The motion sensing features of the conveyor belt are extracted from the multi-source preprocessed data set using the conveyor belt related semantic entities, and the conveyor belt physical quantity dynamic data is obtained.

[0153] In this embodiment, according to the mapping relationship of semantic entities to sensor data fields, the corresponding 8 categories of dynamic physical quantity fields of the conveyor belt are extracted to form a time series data tensor with a dimension of [6000x8]. To ensure data consistency, the sampling frequency and time range are consistent with the robot. Due to high-frequency disturbances in some signal channels (such as belt vibration), a third-order wavelet filter is used for signal smoothing, and an acceleration derivative channel is added to the belt speed and position offset for subsequent correlation analysis. All data channels are uniformly mapped to the [-1, 1] interval range after extraction, facilitating subsequent time alignment processing. The conveyor belt physical quantity dynamic data includes belt speed, motor torque, transmission position offset, vibration amplitude, current and voltage, etc. These parameters describe the motion characteristics and operating status of the conveyor belt, helping to monitor the stability and operating efficiency of the conveyor belt.

[0154] The robot physical quantity dynamic data and the conveyor belt physical quantity dynamic data are time-aligned, and the time-aligned results are effectively screened to obtain an effective multi-dimensional physical quantity feature vector;

[0155] In this embodiment, the sliding window mechanism (window length set to 3 seconds, step 1 second) is used for window-by-window matching based on the time series of the robot physical quantity dynamic data and the conveyor belt physical quantity dynamic data, the maximum cross-correlation value is used to determine the optimal time offset, and frame-level time alignment is achieved. The aligned window data structure is a three-dimensional tensor of [180x(13+8)x30], and each sample contains 30 frames of physical quantity dynamic features. Further based on the channel screening index of change rate greater than the set threshold, invalid signals with weak amplitude or high stability are removed, and 9 key channels are retained to form an effective physical quantity feature tensor with a size of [180x9x30], which is the core expression of the physical quantity behavior.

[0156] A multi-task index set of the industrial automatic equipment is obtained, and the effective multi-dimensional physical quantity feature vector is hierarchically divided and aggregated according to the feature requirements of the multi-task index set to obtain a multi-task hierarchical feature vector set;

[0157] In this embodiment, according to the multi-task index set defined in the equipment operation strategy, the "position accuracy", "clamping stability", "motion coordination", "transmission smoothness" and "beat consistency" five types of tasks are matched with the feature requirements, and the effective physical quantity vector is attributed and divided according to the task correlation. Each type of task constructs a sub-tensor with a size of [180x3x30], where 3 represents the number of key channels selected for the corresponding task. Statistical feature vectors are generated by performing operations such as maximum value, mean value, standard deviation and main frequency on each sub-tensor, and are expanded to a task representation vector with a size of [180x15], and finally integrated to form a multi-task hierarchical feature vector set with a size of [180x5x15].

[0158] The multi-task hierarchical feature vector set is multi-dimensionally embedded and fused to integrate the feature representations of different tasks and generate a preliminary fused multi-task feature vector.

[0159] In this embodiment, the feature vector of each task dimension is input into a fully connected network with a two-layer residual connection structure. The first layer is an intra-task compression structure, which uses a ReLU activation unit with 32 hidden units for feature compression. The second layer sets a residual skip connection across the task channels to preserve task independence and enhance overall semantic fusion capability. The output of each sample forms a [1x24] embedding vector, representing the comprehensive expression ability of the physical quantities of the five types of tasks in this time segment. All vector combinations form a complete output matrix [180x24], which serves as a unified input structure for subsequent device state prediction or control strategy generation.

[0160] Especially important is that the effective dynamic physical quantity screening includes:

[0161] When the fluctuation amplitudes of the robotic arm physical quantity dynamic data and the conveyor belt physical quantity dynamic data in the same time window are greater than 3 units / s, and the correlation coefficient is greater than 0.70, it is determined that there is a strong dynamic coupling relationship between the physical quantity features in the time window, thereby obtaining a strong dynamic coupling relationship feature node.

[0162] In this embodiment, the length of the unified time window is set to 0.5 seconds, and the fluctuation amplitudes of the robotic arm physical quantity dynamic data and the conveyor belt physical quantity dynamic data in the time window are counted. If the physical quantity change amplitudes of the robotic arm and the conveyor belt both exceed 3 units / s, and the corresponding correlation coefficients are higher than 0.70, it is determined that the two physical quantity features exhibit a strong dynamic coupling relationship in this time period. Based on this determination, the system marks the coupling relationship feature node corresponding to the time window in the time series, forming a strong dynamic coupling feature node set.

[0163] Based on the strong dynamic coupling relationship feature node, the duration of the strong dynamic coupling relationship is extracted. When the duration exceeds 1.0 seconds, the strong dynamic coupling relationship feature node is determined to be valid and recorded as an effective strong dynamic feature node.

[0164] In this embodiment, in the strong dynamic coupling feature node set, the duration of each node is calculated, i.e., the length of time that continuously satisfies the strong coupling determination condition. If the duration of a certain feature node exceeds 1.0 seconds, the feature node is determined to be valid and is included in the set of effective strong dynamic feature nodes. This set reflects the strong coupling dynamic performance of the robotic arm and the conveyor belt in the continuous time interval.

[0165] When the fluctuation amplitudes of the mechanical arm physical quantity dynamic data and the conveyor physical quantity dynamic data in the same time window are less than 3.0 units / s, and the correlation coefficient is between 0.40 and 0.70, it is determined that there is a weak dynamic coupling relationship between the physical quantity characteristics in the time window, and a weak dynamic coupling relationship characteristic node is obtained;

[0166] In this embodiment, the time window with a fluctuation amplitude less than 3 units / s and a correlation coefficient between 0.40 and 0.70 is determined as a weak dynamic coupling relationship. The system also generates corresponding weak dynamic coupling relationship characteristic nodes for these time periods. This part of the node reflects the lower intensity but persistent dynamic association between the mechanical arm and the conveyor.

[0167] Based on the weak dynamic coupling relationship characteristic node, the duration of the weak dynamic coupling relationship is extracted. When the duration exceeds 2.0 seconds, the weak dynamic coupling relationship characteristic node is determined to be valid, and is recorded as an effective weak dynamic characteristic node;

[0168] In this embodiment, the duration of the weak dynamic coupling characteristic node is calculated. If the duration exceeds 2.0 seconds, the node is determined to be an effective weak dynamic characteristic node and is recorded into the set of effective weak dynamic characteristic nodes. This step ensures that the weak coupling characteristics selected have sufficient time stability and reflect more reliable dynamic relationships.

[0169] The effective strong dynamic characteristic nodes and the effective weak dynamic characteristic nodes are time-sequentially integrated to obtain an effective multi-dimensional physical quantity characteristic vector.

[0170] In this embodiment, the effective strong dynamic characteristic nodes and the effective weak dynamic characteristic nodes are time-sequentially integrated to generate a unified multi-dimensional physical quantity characteristic vector. The characteristic vector is stored in a matrix form, where the rows represent different coupling characteristic nodes, the columns represent the time sequence, and the matrix elements represent the coupling strength and duration characteristics, providing refined input data for subsequent multi-task index analysis.

[0171] Optionally, step S4 comprises:

[0172] Step S41: The multi-modal unified characteristic vector is classified and disassembled according to the physical quantity category, a multi-dimensional dynamic physical quantity vector set is constructed, and the multi-dimensional dynamic physical quantity vector set is used for dynamic behavior vector mapping of each node in the industrial process structure model, to obtain an industrial process dynamic behavior graph;

[0173] In this embodiment, the multi-modal unified feature vector obtained through multi-source fusion and preprocessing is split according to the physical quantity category (such as force sensing signal, temperature sensing signal, vibration signal, etc.). It is assumed that the feature vector is represented in matrix form, where the rows represent time steps (such as 1000 time points), and the columns represent multiple feature dimensions of each physical quantity (such as 5-dimensional features). After splitting, multiple matrix sets are formed, referred to as multi-dimensional dynamic physical quantity vector sets. Then, each node (node number N_i, i = 1 ~ M, where M is the total number of nodes) in the industrial process structure model is mapped to the corresponding physical quantity category, and the physical quantity vector is assigned to each node through the node mapping matrix to form a node dynamic behavior vector matrix, with each row representing the dynamic behavior characteristics of a node in the time dimension. The finally constructed industrial process dynamic behavior graph is a graph structure G = (V, E), where V is the node set, E is the connection edge between nodes, the node is attached with a dynamic behavior vector attribute, and the edge represents the physical coupling relationship between devices.

[0174] Step S42: Perform edge-by-edge dynamic behavior relationship modeling on the node connection relationship in the industrial process dynamic behavior graph to obtain an industrial process dynamic relationship matrix;

[0175] In this embodiment, based on the obtained dynamic behavior graph, for each node pair (N_i, N_j) connected by an edge in the graph, the dynamic behavior relationship between the nodes is obtained by calculating the dynamic correlation of the corresponding node dynamic behavior vectors. This calculation uses a sliding time window mechanism (window length T = 50 time points), and in each time window, the similarity (such as correlation coefficient) of the node dynamic characteristics is calculated to form a dynamic relationship value r_ij(t). The relationship values of all edges are integrated into a dynamic relationship matrix R with a dimension of M x M x T, where each element R(i, j, t) represents the dynamic interaction intensity between node i and node j at time t. This matrix is used to dynamically reflect the time evolution characteristics of the behavior relationship of the nodes in the industrial process.

[0176] Step S43: Perform multi-dimensional dynamic topology optimization on the industrial process dynamic behavior graph using the industrial process dynamic relationship matrix to obtain an optimized industrial process dynamic behavior graph;

[0177] In this embodiment, for the connection strength in the dynamic relationship matrix R, weak connection edges are removed through threshold screening (threshold set to 0.3), and the edge weight distribution is adjusted using a weighted adjacency matrix smoothing technique to reduce noise interference. Matrix operations based on the graph smoothing operator are used to regularize the adjacency matrix, enhancing the weight of the main dynamic connections, thereby generating an optimized adjacency matrix A_opt. The optimized industrial process dynamic behavior graph G_opt = (V, E_opt) is constructed from this, where E_opt is the edge set after screening and weight optimization, the sparsity of the graph is controlled, the representativeness and computational efficiency of the graph structure are maintained, and the effectiveness of the industrial process dynamic interaction network structure is ensured.

[0178] Step S44: Based on the optimized industrial process dynamic behavior graph, a multi-dimensional relationship modeling of structured-unstructured features is performed to obtain a multi-dimensional dynamic relationship network;

[0179] In this embodiment, the structured information (device connection relationship, process node topology) and the unstructured dynamic behavior vector are combined, and a three-dimensional tensor T (dimension M x feature dimension F x time step T) is constructed to realize the coupling relationship expression of multi-dimensional features. In the tensor T, the first dimension corresponds to each node of the industrial process, the second dimension corresponds to the physical quantity feature dimension, and the third dimension corresponds to the time sequence. The tensor simultaneously carries the structural topology information and the time dynamic information, and realizes the establishment of the multi-dimensional dynamic relationship network. Through multi-dimensional operation of the tensor, the complex relationship between nodes is mined to support subsequent dynamic analysis.

[0180] Step S45: According to the multi-dimensional dynamic relationship network, a graph structure learning is performed to obtain an industrial process multi-dimensional dynamic model.

[0181] In this embodiment, based on the multi-dimensional dynamic relationship network, a hierarchical spatio-temporal embedding structure is designed for dynamic learning of the graph structure. The structure includes three main levels: node feature encoding layer, time dependence capturing layer and topology relationship fusion layer. The node feature encoding layer is responsible for processing the multi-dimensional physical quantity features of each node changing over time. This layer adopts a multi-channel convolution method, which can extract time sequence features from the time sequence data of the node, capture the dynamic change law of the node feature, and generate a feature representation with low time resolution. The time dependence capturing layer models the time sequence of the encoded node features through recursive structures such as gated recurrent units (gate threshold range), further mines long-term dependence relationships, and improves the model's ability to capture the dynamic evolution of nodes. The topology relationship fusion layer is based on the dynamically changing adjacency relationship matrix, and combines the weighted graph convolution operation to realize the interactive fusion of node features in space. This layer integrates the spatial dependence and time dynamics (fluctuation interval) between nodes, and outputs an embedding representation that comprehensively reflects the multi-dimensional dynamic behavior of the device node and the topology change. Through iterative training of multi-level spatio-temporal features, the model can accurately capture the dynamic changes of the device running state and the network evolution trend in the industrial process, thereby constructing a multi-dimensional dynamic model that conforms to the actual complex industrial environment, and supporting subsequent behavior prediction and anomaly detection applications.

[0182] Optionally, the multi-dimensional abnormal outlier index of the device identified in step S5 includes:

[0183] The running state deviation degree of the quantitative running state prediction result is obtained to obtain a deviation degree tensor;

[0184] In this embodiment, for the prediction result of the running state of the industrial equipment, firstly, the deviation degree between the actual observation value and the prediction value collected in reality in real time is quantitatively calculated to form a deviation tensor. The tensor contains deviation data of multiple physical quantities in the dimension of equipment, and the time dimension is in units of minutes, reflecting the deviation of different equipment physical quantities in each time step. Specifically, the data structure of the deviation tensor is a three-dimensional tensor with the size of equipment number x physical quantity number x time step, facilitating subsequent multi-dimensional analysis.

[0185] In each time step of 3 minutes, the residual features of the equipment physical quantities in the deviation tensor are extracted to generate a state residual matrix.

[0186] In this embodiment, in each time step of three minutes, the residual data of the physical quantities of all equipment in the time period is extracted from the deviation tensor to form a state residual matrix. The rows of the matrix represent different equipment, and the columns correspond to the residual values of various physical quantities. Through such structured residual data, the running deviation of each equipment in the time window can be clearly reflected. The size of the residual matrix is usually equipment number x physical quantity number, such as 50 x 10, supporting fine-grained analysis.

[0187] The mean, variance and extreme value fluctuation analysis are respectively performed on each row and each column in the state residual matrix, and the fluctuation analysis results are organized into a multi-dimensional outlier detection index vector.

[0188] In this embodiment, the statistical features of each row (i.e., the residual distribution of each equipment) and each column (i.e., the residual distribution of each physical quantity on all equipment) of the state residual matrix are calculated, including mean, variance, and maximum and minimum value extreme value fluctuation indicators. These statistical results are integrated into a multi-dimensional outlier detection index vector, and the vector dimension is determined by the number of equipment, the number of physical quantities, and the number of statistical features. For example, a feature vector with a length of equipment number x physical quantity number x 3 (number of statistical indicators) can be generated to reflect the abnormal fluctuation of data distribution.

[0189] In combination with the fluctuation interval and threshold range of each equipment and each physical quantity in the multi-dimensional dynamic model of the industrial process, a dynamic adaptive discrimination condition is set, and the dynamic adaptive discrimination condition is used for abnormal outlier behavior detection on the multi-dimensional outlier detection index vector to obtain an abnormal outlier index point set.

[0190] In this embodiment, in combination with the pre-defined normal fluctuation interval and threshold of each equipment physical quantity in the multi-dimensional dynamic model of the industrial process, a dynamic adaptive discrimination condition is designed. The discrimination condition adjusts the threshold sensitivity in real time according to the multi-dimensional outlier index vector of the current state residual matrix, realizes personalized abnormal detection for different equipment and different physical quantities. Through this mechanism, the abnormal behavior existing in the outlier index vector can be dynamically determined, and finally the abnormal outlier index point set is identified.

[0191] The abnormal outlier indicator point sets in each time step are summarized to construct a device multi-dimensional abnormal outlier indicator set.

[0192] In this embodiment, the abnormal outlier indicator points detected in each time step are summarized to form a device multi-dimensional abnormal outlier indicator set. The indicator set is stored in the form of a time series, supporting the analysis of abnormal trends and mutation points of the device in a continuous operation period. The indicator set structure is time step x device number x abnormal indicator number, facilitating long-term monitoring and early warning, and ensuring comprehensive control and timely response of the industrial equipment operation state.

[0193] Optionally, the adaptive parameter reconstruction in step S6 includes:

[0194] Extract abnormal frequent parameter points in the device multi-dimensional abnormal outlier indicator set, and construct an abnormal feature distribution vector;

[0195] In this embodiment, based on the abnormal point information recorded in the device multi-dimensional abnormal outlier indicator set, first, the abnormal frequent parameter points are screened. The specific operation is to traverse all abnormal indicator points in the time series, count the frequency and duration of each parameter, and select parameters with a frequency exceeding a set threshold (such as more than 5 times per hour). Subsequently, a multi-dimensional array structure is used to construct an abnormal feature distribution vector, each dimension of which corresponds to an abnormal parameter, containing the frequency weight, average deviation amplitude and duration information of the parameter abnormality, thereby forming an abnormal feature expression that can reflect the spatial distribution and time evolution of the device abnormal state.

[0196] Differential analysis and comparison of the abnormal feature distribution vector and the state parameters of the industrial process dynamic model are performed to calculate a reconstruction gain coefficient set;

[0197] In this embodiment, the above abnormal feature distribution vector is compared with the state parameter data in the multi-dimensional dynamic model of the industrial process. The state parameters are stored in the form of a multi-dimensional matrix, reflecting the fluctuation interval and typical state range of each physical quantity of the device during normal operation. Through dimension matching, each parameter of the abnormal feature distribution vector is compared with the corresponding state parameter, and then a numerical set reflecting the abnormal degree is generated, i.e. the reconstruction gain coefficient set. The calculation formula can be: where G i is the reconstruction gain coefficient, A i is the abnormal value of the i-th parameter in the abnormal feature distribution vector, and S i is the normal state value of the corresponding i-th parameter in the industrial process dynamic model. The set serves as a weight vector, indicating the amplitude and direction of adjustment for each control parameter, supporting the subsequent optimization of control parameters to adapt to the abnormal state.

[0198] Call the device control log of the device control unit, and screen the control parameter groups in combination with the multi-source preprocessing data set;

[0199] In this embodiment, the device control log saved in the device control unit is called, which contains multiple control parameter configurations of the device at different times and corresponding operation responses. In combination with the current multi-source preprocessing data set, first, each control parameter group in the control log is screened. The screening rule is based on the correlation between the device abnormal parameters and the corresponding control parameters, and selects multiple best control parameter combinations matching the current abnormal state. Each control parameter group includes multiple dimensions such as joint speed of the mechanical arm, clamping force, and conveyor belt speed, and each parameter group corresponds to the excellent response effect exhibited by the device in the past actual operation. By associating historical control behavior with the current multi-source data, the screened parameter groups effectively cover the diversified control requirements in the abnormal scenario.

[0200] Weighting offset the control parameter groups by using the reconstruction gain coefficient set to generate a reconstruction parameter vector set;

[0201] In this embodiment, the multiple best control parameter groups screened are weighted and offset by using the reconstruction gain coefficient set. Specifically, the original value of each control parameter group is multiplied by the corresponding gain coefficient, and is added or subtracted to correct the abnormal deviation direction, to generate a new reconstruction parameter vector set. The set is organized in the form of a multi-dimensional array, contains multiple adjusted control parameter combinations, ensures the diversity and flexibility of the control strategy, and facilitates the dynamic selection of the most suitable parameter combination in the subsequent device operation process.

[0202] Convert the reconstruction parameter vector set into an instruction sequence, and align it with the device control channel protocol specification of the device control unit to generate a device control instruction sequence.

[0203] In this embodiment, the reconstruction parameter vector set is converted into a device control instruction sequence that meets the communication protocol specification of the device control unit. The instruction sequence is encoded according to the predetermined format, including control parameter number, adjusted value, timestamp, and command identification information. The instruction sequence is organized in time sequence to ensure that the device can receive and execute in the predetermined order. Finally, the control instruction is issued to the mechanical arm and the conveyor belt through the device control channel, realizing precise control adjustment in the abnormal state and ensuring the stability and safety of the device operation.

[0204] Optionally, the present specification also provides an industrial automatic device digital visualization control system for executing the industrial automatic device digital visualization control method as described above, which comprises:

[0205] The data preprocessing module is configured to acquire multi-source heterogeneous sensing data, and perform dynamic label analysis and signal decoupling according to the multi-source heterogeneous sensing data, to obtain a multi-source preprocessed data set.

[0206] The semantic association module is configured to construct an industrial process structure model according to the multi-source preprocessed data set, and perform multi-level semantic association enhancement between the industrial process structure model and semantic entities in the multi-source preprocessed data set, to obtain a dynamic semantic association network.

[0207] The mechanical arm-conveyor belt association module is configured to perform multi-task feature vector fusion based on the dynamic semantic association network and in combination with motion physical quantity features of an end effector of a mechanical arm and a conveyor belt in the multi-source preprocessed data set, to obtain a preliminary fused multi-task feature vector; and perform feature vector weighted regression on the preliminary fused multi-task feature vector, to obtain a multi-modal unified feature vector.

[0208] The device behavior mapping module is configured to perform device multi-dimensional dynamic behavior mapping by using the multi-modal unified feature vector, to obtain an industrial process multi-dimensional dynamic model.

[0209] The device anomaly outlier analysis module is configured to acquire real-time multi-source heterogeneous sensing data; perform device running state prediction on the real-time multi-source heterogeneous sensing data by using the industrial process multi-dimensional dynamic model, and identify device multi-dimensional anomaly outlier indexes according to a result of the running state prediction.

[0210] The control parameter reconstruction module is configured to perform adaptive parameter reconstruction on the device multi-dimensional anomaly outlier indexes, to obtain a device control instruction sequence, and deliver the device control instruction sequence to a corresponding device control unit.

[0211] Therefore, from any viewpoint, the embodiments should be considered as being exemplary and non-limiting, the scope of the application being defined by the appended claims and not by the above description, and all the changes falling within the meaning and the scope of the equivalent elements of the patent file are therefore intended to be comprised in the present application.

[0212] The above description is merely one specific implementation of the application, which enables a person skilled in the art to understand or implement the application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the application. Therefore, the application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A digital visualization control method for industrial automatic equipment, characterized in that: The following steps are involved: Step S1: Acquire multi-source heterogeneous sensor data, and perform dynamic label resolution and signal decoupling based on the multi-source heterogeneous sensor data to obtain a multi-source preprocessed data set; Step S2: constructing an industrial process structure model based on the multi-source preprocessed dataset, and performing multi-level semantic association enhancement between the industrial process structure model and the semantic entities in the multi-source preprocessed dataset to obtain a dynamic semantic association network; Step S3: Based on the dynamic semantic association network, multi-task feature vector fusion is performed in combination with the motion physical quantity features of the robot end effector and the conveyor belt in the multi-source preprocessed data set to obtain a preliminary fused multi-task feature vector; feature vector weighted regression is performed on the preliminary fused multi-task feature vector to obtain a multimodal unified feature vector; Step S4: Use the multimodal unified feature vector to map the multidimensional dynamic behavior of the equipment, thereby obtaining a multidimensional dynamic model of the industrial process; Step S5: Acquire real-time multi-source heterogeneous sensor data; predict the equipment operation status of the real-time multi-source heterogeneous sensor data using a multi-dimensional dynamic model of the industrial process, and identify multi-dimensional abnormal outlier indicators of the equipment based on the operation status prediction results; Step S6: Adaptively reconstruct the parameters of the multi-dimensional abnormal outlier indicators of the equipment to obtain the equipment control instruction sequence, and send it to the corresponding equipment control unit.

2. The digital visualization control method for industrial automatic equipment according to claim 1, characterized in that: Step S1 includes: Step S11: collecting multi-source heterogeneous sensor data, and performing time zone correction and data time series alignment on the multi-source heterogeneous sensor data to obtain a time-standardized multi-source data set; Step S12: performing dynamic label initial parsing on various physical quantities in the time-normalized multi-source data set to obtain dynamic label initial parsed data; Step S13: performing hierarchical signal decoupling processing on the dynamic tag initial parsed data, extracting the principal component dynamic signal subset, thereby obtaining hierarchical decoupled signal data; Step S14: performing multi-dimensional dynamic signal correlation reconstruction using the layered decoupled signal data to obtain multi-dimensional signal periodic correlation data; Step S15: Dynamically fine-tune the labels of the multi-dimensional signal period correlation data, and integrate the results of the dynamic label fine-tuning to obtain a multi-source pre-processed data set.

3. The digital visualization control method for industrial automatic equipment according to claim 2, characterized in that: Step S12 includes: Step S121: Acquire the operation structure data of the industrial automatic equipment, including the equipment operation cycle, equipment structure data and process flow data; Step S122: performing physical feature space division on the time-normalized multi-source data set during the equipment operation cycle, and performing feature space annotation on the division results to obtain initial annotated vector data of multiple physical quantities; Step S123: performing device operation event identification on the time-normalized multi-source data set to obtain device operation event classification data; Step S124: combining the multi-physical quantity initial labeled vector data and the equipment operation event classification data to perform operating condition event label screening to obtain multi-physical quantity event screening label data; Step S125: reconstructing the time series of the multi-physical quantity event screening label data, and enhancing the label semantics in the time series reconstruction results by combining the equipment structure data and process flow data to obtain dynamic semantic label vector data; Step S126: performing label integration and standardization processing based on the dynamic semantic label vector data to obtain dynamic label initial parsing data.

4. The digital visualization control method for industrial automatic equipment according to claim 2, characterized in that: Step S14 includes: Step S141: performing multidimensional signal grouping processing based on the decoupling levels of different physical quantities and signals in the hierarchical decoupling signal data, and establishing a high-dimensional feature space mapping of each physical quantity to obtain multidimensional signal feature mapping data; Step S142: dynamically aligning the time relationship between the physical quantities in the multi-dimensional signal feature mapping data, identifying the dynamic coupling relationship and synchronization between different physical quantities, and obtaining multi-dimensional signal dynamic time alignment data; Step S143: performing multimodal relationship modeling on the multidimensional signals in the multidimensional signal dynamic time alignment data to obtain multidimensional signal multimodal relationship data; Step S144: performing multidimensional signal periodic pattern detection based on the multidimensional signal multimodal relationship data to obtain multidimensional signal periodic pattern data; Step S145: clustering and weighting the periodic signal features of each periodic pattern according to the similarities and differences between the signal periodicities in the multidimensional signal periodic pattern data to obtain multidimensional signal periodic correlation data.

5. The digital visualization control method for industrial automatic equipment according to claim 1, characterized in that: Step S2 includes: Step S21: performing hierarchical analysis on the device structure data in the multi-source pre-processed data set, extracting the topological features of the devices, and obtaining device structure topological feature data; Step S22: performing process modeling and relationship reasoning based on the process flow data in the multi-source pre-processed data set to obtain process flow feature data; Step S23: Perform multi-level graph embedding and fusion of the equipment structure topology feature data and the process flow feature data to obtain an industrial process structure model; Step S24: performing semantic entity division on the equipment operation events and process labels in the multi-source pre-processed data set to generate equipment operation semantic entity data; Step S25: Perform multi-level semantic association enhancement on the industrial process structure model and the equipment operation semantic entity data to obtain dynamic semantic association network data.

6. The digital visualization control method for industrial automatic equipment according to claim 1, characterized in that: The multi-task feature vector fusion in step S3 includes: According to the dynamic semantic association network, the highly relevant semantic entities of the robot end effector and the conveyor belt are extracted to obtain the robot arm related semantic entities and the conveyor belt related semantic entities. The physical quantity dynamic features of the manipulator are extracted from the multi-source preprocessed data set using the manipulator-related semantic entities to obtain the manipulator's physical quantity dynamic data; The conveyor belt's motion sensing features are extracted from the multi-source preprocessed data set using conveyor belt related semantic entities to obtain the conveyor belt's physical quantity dynamic data. Perform time series alignment on the dynamic data of the physical quantities of the robot arm and the conveyor belt, and perform effective dynamic physical quantity screening on the time series alignment results to obtain effective multi-dimensional physical quantity feature vectors; Obtain a multi-task indicator set for industrial automatic equipment, and perform hierarchical division and aggregation of effective multi-dimensional physical quantity feature vectors according to the characteristic requirements of the multi-task indicator set to obtain a multi-task hierarchical feature vector set; The multi-task hierarchical feature vector set is multi-dimensionally embedded and fused to integrate the feature representations of different tasks and generate a preliminary fused multi-task feature vector.

7. The digital visualization control method for industrial automatic equipment according to claim 1, characterized in that: Step S4 includes: Step S41: Classify and decompose the multimodal unified feature vector according to the physical quantity category, construct a multidimensional dynamic physical quantity vector set, and use the multidimensional dynamic physical quantity vector set to perform dynamic behavior vector mapping on each node in the industrial process structure model to obtain an industrial process dynamic behavior graph; Step S42: performing edge-by-edge dynamic behavior relationship modeling on the node connection relationships in the industrial process dynamic behavior graph to obtain an industrial process dynamic relationship matrix; Step S43: performing multi-dimensional dynamic topology optimization on the industrial process dynamic behavior graph using the industrial process dynamic relationship matrix to obtain an optimized industrial process dynamic behavior graph; Step S44: performing structured-unstructured feature multidimensional relationship modeling based on the optimized industrial process dynamic behavior diagram to obtain a multidimensional dynamic relationship network; Step S45: Perform graph structure learning based on the multidimensional dynamic relationship network to obtain a multidimensional dynamic model of the industrial process.

8. The digital visualization control method for industrial automatic equipment according to claim 1, characterized in that: In step S5, the identification of multi-dimensional abnormal outlier indicators of the device includes: Quantify the running state deviation of the running state prediction result to obtain the deviation tensor; Within 3 minutes of each time step, the residual features of the device physical quantities in the deviation tensor are extracted to generate the state residual matrix; Perform mean, variance and extreme value fluctuation analysis on each row and column in the state residual matrix, and organize the fluctuation analysis results into a multidimensional outlier detection indicator vector; Combined with the fluctuation range and threshold range of each physical quantity of each device in the multidimensional dynamic model of the industrial process, dynamic adaptive discrimination conditions are set. The dynamic adaptive discrimination conditions are used to detect abnormal outlier behavior of the multidimensional outlier detection indicator vector to obtain the abnormal outlier indicator point set. The abnormal outlier indicator point set in each time step is summarized to construct a multi-dimensional abnormal outlier indicator set of the equipment.

9. The digital visualization control method for industrial automatic equipment according to claim 1, characterized in that: The adaptive parameter reconstruction in step S6 includes: Extract abnormal frequent parameter points from the multi-dimensional abnormal outlier indicator set of the equipment and construct the abnormal feature distribution vector; Perform difference analysis and comparison between the abnormal feature distribution vector and the state parameters of the industrial process dynamic model, and calculate the reconstruction gain coefficient set; Call the device control log of the device control unit and filter the control parameter group based on the multi-source pre-processed data set; Performing a weighted offset on the control parameter group using the reconstruction gain coefficient set to generate a reconstruction parameter vector set; The reconstructed parameter vector set is converted into an instruction sequence and aligned with the device control channel protocol specification of the device control unit to generate a device control instruction sequence.

10. A digital visualization control system for industrial automatic equipment, characterized in that: For executing the industrial automatic equipment digital visualization control method according to claim 1, the industrial automatic equipment digital visualization control system comprises: The data preprocessing module is used to obtain multi-source heterogeneous sensor data and perform dynamic label parsing and signal decoupling based on the multi-source heterogeneous sensor data to obtain a multi-source preprocessed data set; The semantic association module is used to build an industrial process structure model based on the multi-source pre-processed dataset, and perform multi-level semantic association enhancement between the industrial process structure model and the semantic entities in the multi-source pre-processed dataset to obtain a dynamic semantic association network; The robot-conveyor belt association module is used to fuse multi-task feature vectors based on a dynamic semantic association network and the motion physical quantity characteristics of the robot arm end effector and conveyor belt in the multi-source preprocessed dataset to obtain a preliminary fused multi-task feature vector; and perform feature vector weighted regression on the preliminary fused multi-task feature vector to obtain a multimodal unified feature vector. The equipment behavior mapping module is used to map the multi-dimensional dynamic behavior of equipment using a multimodal unified feature vector, thereby obtaining a multi-dimensional dynamic model of the industrial process; The equipment anomaly and outlier analysis module is used to obtain real-time multi-source heterogeneous sensor data; it uses a multi-dimensional dynamic model of the industrial process to predict the equipment operating status of the real-time multi-source heterogeneous sensor data, and identifies the equipment multi-dimensional anomaly and outlier indicators based on the operating status prediction results; The control parameter reconstruction module is used to adaptively reconstruct the parameters of the multi-dimensional abnormal outlier indicators of the equipment, obtain the equipment control instruction sequence, and send it to the corresponding equipment control unit.

Images