A paper bag raw material spiral conveying internet of things metering and fault pre-diagnosis method

Abstract

This invention relates to an IoT-based metering and fault prediction method for screw conveyors used in paper bag production. Through spatial planning, multiple low-power sensors, including vibration, optical, and current sensors, are strategically deployed at key mechanical nodes to collect raw multi-channel signals and align them based on a unified time reference, achieving synchronous integration of multi-dimensional data. Signal normalization and high-dimensional feature compression are used to construct a graph structure state attribute with both temporal and spatial constraints. Operating condition identification is performed by combining physical parameters and a rule base, dynamically injecting node attributes, calculating physical correlations and confidence weights between nodes, and generating a dynamic graph with physical meaning. By comparing sliding window subgraph sampling with primitive trajectories, local topological change features are extracted, enabling accurate identification of the trend, location, and rate of early mechanical micro-degradation. When the abnormal evolution rate exceeds a threshold, the system issues an early warning and optimizes the model based on feedback data. This method significantly improves the intelligence level of screw conveyor status monitoring.

Description

Technical Field

[0001] This invention relates to the fields of industrial equipment status sensing, heterogeneous data fusion and fault trend prediction technology, and in particular to an IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials. Background Technology

[0002] In the current field of early fault trend prediction and IoT metering for paper bag raw material screw conveyor systems, the mainstream approaches mainly rely on signal time-series feature engineering, shallow / deep machine learning models, and traditional criteria based on single physical quantities. Common technical paths in the industry are as follows: First, by deploying multi-source sensors such as vibration sensors, current sensors, and temperature modules at key mechanical nodes to acquire multi-dimensional data on equipment operation and achieve health status monitoring of the screw conveyor. Second, by using feature engineering methods such as signal spectrum analysis and time-domain statistical acquisition to screen for more sensitive fault discrimination parameters, and by setting empirical thresholds or conducting traditional machine learning (such as decision trees, support vector machines, and Naive Bayes) for status determination. Third, with the development of industrial digitalization, some systems are attempting to introduce deep learning models (such as convolutional neural networks and recurrent neural networks LSTM) or end-to-end multi-channel data fusion to achieve fault prediction and maintenance early warning by extracting features and identifying trends from large-scale sensor signals. In addition, a small number of documents and patents employ fusion paradigms such as multimodal data stitching, attention mechanisms, or evidence theory to achieve equipment status determination under multiple operating conditions.

[0003] The existing technology has the following major shortcomings: First, the generalization ability of heterogeneous data fusion models is limited. Most methods rely heavily on the consistency of data distribution between the model and the training conditions, and are poorly adaptable to changes in equipment state transitions and physical scenarios under new conditions. The models are prone to bias or degradation, and cannot meet the high robustness requirements of complex fields such as multiple operating conditions, multiple raw material ratios, and equipment aging.

[0004] Second, the ability to embed physical semantics is insufficient. The mainstream approach uses an end-to-end method with weak physical constraints for feature extraction and fusion, which makes it difficult to explicitly describe the material flow path, energy transfer law and key node action mechanism of the screw conveyor. This is not conducive to the identification of early minor faults (such as bearing micro-stripping and local wear of blades), and the output results lack interpretability and credibility support for field engineering applications.

[0005] Third, the model maintenance and upgrade costs are high. Traditional deep fusion models often require full retraining when migrating across different operating conditions. When encountering new processes / materials, a large number of labeled samples must be re-accumulated, execution domain alignment or model fine-tuning must be performed, resulting in long downtime for field equipment and making it difficult to achieve adaptive optimization closed loop. Summary of the Invention

[0006] This application provides an IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials, which aims to solve one of the problems or issues of the existing technology mentioned in the background.

[0007] This application provides a method for IoT metering and fault pre-diagnosis of spiral conveying of paper bag raw materials, specifically including: S1: Obtain multi-source raw signal streams of key mechanical points of the screw conveyor. The multi-source raw signal streams include waveform data collected by a low-power vibration accelerometer, current time-series data recorded by a motor current sampling module, and spectral reflectance data detected by an optical thickness sensor at the feed inlet. Based on a unified timestamp, the above multi-source raw signal streams are synchronously aligned to generate a time-synchronized multidimensional data sequence.

[0008] S2: Based on the time-synchronized multidimensional data sequence, perform fixed-length slicing processing to truncate the waveform data of each channel in each sampling period into standardized data segments, and map the standardized data segments into the initial attribute vectors of graph nodes to generate node attribute sets.

[0009] S3: Based on the physical property parameters of the paper bag raw material and the equipment operating status parameters, the current operating condition is symbolically analyzed to generate operating condition sensing anchor tag. The operating condition sensing anchor tag represents a specific operating mode such as high humidity and low speed viscous state or dry powder and high speed segregation state.

[0010] S4: Based on the condition-sensing anchor tag, match the nearest graph structure template, and dynamically inject the node attribute set into the corresponding mechanical functional unit node in the graph structure template. At the same time, calculate the edge weight based on the signal cross-correlation coefficient between adjacent nodes and the matching confidence of the condition-sensing anchor tag, thereby constructing a dynamic graph with physical semantic constraints.

[0011] S5: Perform a sliding window subgraph sampling operation on the dynamic graph to extract the local topological structure change vector between consecutive time steps. The local topological structure change vector represents the evolution trend of the energy transfer path between the feed section node and the middle support node.

[0012] S6: Input the local topological change vector into the micro-anomaly evolution primitive library for symbolic distance matching calculation. The micro-anomaly evolution primitive library stores local subgraph evolution trajectory benchmarks constructed based on typical early degradation modes. By comparison, the early fault feature triplet composed of fault trend type, starting position and evolution rate is output.

[0013] S7: Determine whether the evolution rate in the early fault feature triplet exceeds a preset critical threshold. If the determination result is that it exceeds the threshold, generate an early fault warning instruction containing the fault type and prediction time; otherwise, generate a system normal operation status identifier.

[0014] S8: Update the device health record of the remote management terminal based on the early fault warning command or the system normal operation status indicator, and perform incremental correction on the local subgraph evolution trajectory benchmark in the micro-anomaly evolution primitive library according to the newly generated actual fault feedback data to complete the model adaptive optimization.

[0015] The IoT-based metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials provided in this application has the following beneficial effects: (1) To address the problems of traditional industrial equipment fault prediction methods relying on deep neural network black-box modeling, requiring a large amount of labeled data, and being difficult to adapt to dynamic operating conditions, this solution constructs a lightweight dynamic graph based on physical semantic constraints. The original waveforms of multi-source sensors are directly used as the initial attributes of graph nodes. Combined with symbolic chemical condition anchor points generated by raw material characteristics and operating parameters, a preset graph structure template is matched, thus realizing an explicit and structured representation of the equipment state evolution process. This mechanism abandons the traditional frequency domain analysis or statistical feature extraction path, avoids information loss caused by the subjectivity of feature engineering, and overcomes the performance degradation defects caused by data distribution shifts when existing methods migrate across operating conditions. Since the graph edge relationship is jointly driven by the cross-correlation of signals and the confidence of operating condition matching, it can adaptively reflect the changing trend of the dynamic coupling between energy transfer path and material flow, significantly improving the sensitivity of early weak anomaly capture and the robustness of pattern recognition. High-precision trend positioning can still be achieved without the need for a large number of historical fault labels.

[0016] (2) Unlike mainstream models that rely on end-to-end learning or multimodal fusion strategies for fault classification, this solution innovatively transforms the fault identification problem into a graph structure evolution trajectory matching task constrained by physical rules. By constructing a "micro-anomaly evolution primitive library" offline to store the dynamic evolution features of local subgraphs corresponding to typical degradation modes, the online stage only requires sign distance comparison of local structural change vectors within the sliding window to complete the inference output. This design not only significantly reduces computational complexity, enabling the entire link to run in real time at the edge, with response latency controlled at the level of hundreds of milliseconds, meeting the stringent requirements of high real-time performance for continuous operation of paper bag production lines, but more importantly, it endows the system with strong interpretability—each type of early warning result can be traced back to the interaction degradation path between specific functional units and its corresponding physical mechanism, facilitating rapid decision-making and handling by on-site maintenance personnel. In addition, the construction of the primitive library only requires a small number of typical degradation case samples, effectively alleviating the problem of scarce fault data in industrial scenarios, and possessing good deployment feasibility and engineering practicality.

[0017] (3) This scheme achieves explicit modeling of equipment behavior differences under different raw material states (such as high humidity and viscosity, dry powder segregation) by guiding dynamic reconstruction of the graph structure through working condition perception anchor points. This enables the same system to operate stably in multi-variety, variable parameter production environments without retraining the model or adjusting hyperparameters, significantly enhancing the system's environmental adaptability and generalization performance. Compared with traditional methods that often suffer from soaring false alarm rates or missed detections when facing new working conditions, this architecture ensures the consistency and reliability of diagnostic logic by explicitly encoding the constitutive relationships of the conveying system (such as mechanical linkage logic and material-structure interaction). The overall technical path completely avoids data-driven learning paradigms such as attention mechanisms, domain alignment optimization, and multimodal splicing. Instead, it relies on symbolic rules and graph structure evolution to achieve knowledge embedding, forming a closed-loop adaptive lightweight monitoring system. It is particularly suitable for small and medium-sized manufacturing sites with limited computing power and scarce maintenance resources, providing a new paradigm of state perception that is highly reliable, low-dependency, and easily scalable for industrial edge intelligence.

[0018] The aforementioned technologies work together to construct an equipment health management system that combines high real-time performance, strong interpretability, low data dependency, and good adaptability to operating conditions. Its core advantage lies in deeply integrating physical prior knowledge into the entire process of graph structure modeling, realizing a paradigm shift from "data fitting" to "mechanism guidance," and effectively solving key bottleneck problems faced by existing intelligent diagnostic methods in practical implementation, such as unreliable models, high deployment costs, and high maintenance difficulties. Attached Figure Description

[0019] Figure 1 This is the main flowchart of an IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials.

[0020] Figure 2 This is a sub-flowchart of an IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials.

[0021] Figure 3 This is another sub-flowchart of an IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0024] like Figure 1 As shown, this application provides an IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials, specifically including: S1: Obtain multi-source raw signal streams of key mechanical points of the screw conveyor. The multi-source raw signal streams include waveform data collected by a low-power vibration accelerometer, current time-series data recorded by a motor current sampling module, and spectral reflectance data detected by an optical thickness sensor at the feed inlet. Based on a unified timestamp, the above multi-source raw signal streams are synchronously aligned to generate a time-synchronized multidimensional data sequence.

[0025] S2: Based on the time-synchronized multidimensional data sequence, perform fixed-length slicing processing to truncate the waveform data of each channel in each sampling period into standardized data segments, and map the standardized data segments into the initial attribute vectors of graph nodes to generate node attribute sets.

[0026] S3: Based on the physical property parameters of the paper bag raw material and the equipment operating status parameters, the current operating condition is symbolically analyzed to generate operating condition sensing anchor tag. The operating condition sensing anchor tag represents a specific operating mode such as high humidity and low speed viscous state or dry powder and high speed segregation state.

[0027] S4: Based on the condition-sensing anchor tag, match the nearest graph structure template, and dynamically inject the node attribute set into the corresponding mechanical functional unit node in the graph structure template. At the same time, calculate the edge weight based on the signal cross-correlation coefficient between adjacent nodes and the matching confidence of the condition-sensing anchor tag, thereby constructing a dynamic graph with physical semantic constraints.

[0028] S5: Perform a sliding window subgraph sampling operation on the dynamic graph to extract the local topological structure change vector between consecutive time steps. The local topological structure change vector represents the evolution trend of the energy transfer path between the feed section node and the middle support node.

[0029] S6: Input the local topological change vector into the micro-anomaly evolution primitive library for symbolic distance matching calculation. The micro-anomaly evolution primitive library stores local subgraph evolution trajectory benchmarks constructed based on typical early degradation modes. By comparison, the early fault feature triplet composed of fault trend type, starting position and evolution rate is output.

[0030] S7: Determine whether the evolution rate in the early fault feature triplet exceeds a preset critical threshold. If the determination result is that it exceeds the threshold, generate an early fault warning instruction containing the fault type and prediction time; otherwise, generate a system normal operation status identifier.

[0031] S8: Update the device health record of the remote management terminal based on the early fault warning command or the system normal operation status indicator, and perform incremental correction on the local subgraph evolution trajectory benchmark in the micro-anomaly evolution primitive library according to the newly generated actual fault feedback data to complete the model adaptive optimization.

[0032] Step S1: Obtain multi-source raw signal streams from key mechanical points of the screw conveyor. These multi-source raw signal streams include waveform data collected by a low-power vibration accelerometer, current timing data recorded by a motor current sampling module, and spectral reflectance data detected by an optical thickness sensor at the feed inlet. Based on a unified timestamp, the above multi-source raw signal streams are synchronized and aligned to generate a time-synchronized multidimensional data sequence. Specifically, this includes: S1.1: The layout of sensor nodes is planned for key mechanical points such as the screw of the feed section of the screw conveyor, the middle support bearing and the discharge valve body. Based on the coupling relationship of the physical process of paper bag raw material conveying, the installation coordinates of the low power vibration accelerometer, the motor current sampling module and the optical thickness sensor at the feed port are determined to generate a multi-source sensor array spatial topology.

[0033] Spatial topology planning was performed on the sensor node layout of key mechanical points such as the screw in the feed section of the screw conveyor, the central support bearing and the discharge valve body. A position priority index was established based on the degree of coupling between each mechanical point and the physical conveying process of the paper bag raw materials during operation. The sensor deployment order was determined according to the order of action of the mechanical functional units in the energy transfer chain.

[0034] A three-dimensional coordinate system for calculating the mechanical position is established by combining the centerline of the screw conveyor main shaft with the raw material flow path. Within this coordinate system, a spatial constraint model generated based on the equipment structure drawings is used to define the installable area, and the optimal installation position of the sensor is solved by a three-dimensional geometric matching method.

[0035] The installation posture and orientation of the low-power vibration accelerometer, motor current sampling module and feed port optical thickness sensor are optimized. The objective function of maximizing signal coverage is adopted to calculate the angle between the sensing direction of each sensor and the motion direction of the mechanical position, so as to ensure that the incident path of vibration and optical signals maintains maximum physical sensitivity.

[0036] Using a multi-source signal coverage simulation model, the vibration response amplitude, current signal stability, and spectral reflection interference under different layout schemes are comprehensively scored. Based on the scoring results, the installation coordinate group that achieves the highest comprehensive index value is selected.

[0037] By binding the aforementioned installation coordinate group with the mechanical position identifier, a spatial topology data object of the multi-source sensor array is constructed, and a physical coupling relationship matrix between nodes is attached to provide a spatial structure benchmark for subsequent parallel acquisition and unified time axis mapping.

[0038] By planning the spatial topology layout and optimizing the installation parameters, the target mechanical position information from the previous step is transformed into an executable sensor layout scheme, thereby achieving a high-coverage, high-sensitivity multi-source signal acquisition infrastructure.

[0039] S1.2: Based on the spatial topology of the multi-source sensor array, drive each sensing unit to collect raw physical quantities in parallel, use a low-power vibration accelerometer to obtain high-frequency mechanical vibration waveform data, record the stator current timing data of the drive motor through the motor current sampling module, and use the feed port optical thickness sensor to detect the spectral reflectance data of the raw material layer to form a multi-channel heterogeneous raw signal set.

[0040] Based on the determined sensor installation coordinate parameters in the spatial topology of the multi-source sensor array, the array control module is invoked to simultaneously issue acquisition start commands to the low-power vibration accelerometer, motor current sampling module, and feed port optical thickness sensor, and the acquisition channels of each sensing unit are opened under the trigger pulse synchronization mechanism.

[0041] A high-frequency sampling rate parameter is set in the acquisition channel of the vibration accelerometer. Based on the mechanical vibration characteristics of the screw conveyor, the sampling frequency is set to a value that can cover the target mechanical resonance frequency band. The instantaneous acceleration vector time series of the screw in the feeding section, the middle support bearing and the discharge valve body are recorded through the high-frequency waveform continuous acquisition module.

[0042] In the motor current sampling module acquisition channel, the power frequency multiple sampling rate parameter is set. Based on the stator current timing fluctuation characteristics of the drive motor, the sampling period is locked within a time span that includes at least one complete mechanical rotation cycle of the motor rotor. The instantaneous value sequence of the current of each phase winding is recorded through the current timing continuous acquisition module.

[0043] Narrowband spectral scanning parameters are set in the optical thickness sensor acquisition channel. Based on the optical reflection characteristics and fiber distribution density of the paper bag raw material, the incident light wavelength range and sampling integration time are adjusted. The multi-wavelength reflection intensity sequence of the raw material layer is recorded through the spectral reflection continuous acquisition module.

[0044] Vibration waveform data, current timing data, and spectral reflectance data generated by each acquisition channel are classified and stored according to the sensor node identifier and acquisition batch number. A multi-channel heterogeneous raw signal set is constructed using a unified acquisition batch index key, so as to realize the aggregation of raw physical quantities that can be directly entered into the clock source calibration process.

[0045] By employing the aforementioned acquisition triggering strategy and synchronization mechanism, the spatial distribution information of the array topology is effectively combined with the high-fidelity original signals of each channel, achieving consistency of multi-source original data in both spatial and temporal dimensions, and providing an accurate and complete source data foundation for subsequent time-series alignment and map construction.

[0046] For example, when the paper bag raw material screw conveyor is operating at medium humidity and medium speed, the array control module sets the sampling frequency of the low-power vibration accelerometer to 5120 Hz to cover the resonant frequency band of the support bearing; the motor current sampling module sets the sampling frequency to 256 Hz, with a sampling period covering two motor rotor rotation cycles to ensure the capture of stator current fluctuation details; the optical thickness sensor adjusts the incident light wavelength range to 450 nm. A 650 nm sampling and integration time of 50 ms was used for precise measurement of raw material layer thickness and fiber optical properties. In this configuration, the vibration accelerometer generated an acceleration vector sequence of 5120 points, the current sampling module generated a current sequence of 256 points, and the optical thickness sensor generated a reflection intensity matrix containing 10 wavelength channels. After batch indexing, a set of heterogeneous source data with three channels was formed. Field verification showed that this set could achieve consistency in signal amplitude and phase after subsequent timestamp alignment processing, significantly improving the reliability of early fault feature extraction.

[0047] S1.3: Perform a global clock source calibration operation on the multi-channel heterogeneous raw signal set, and use a network time protocol or hardware trigger pulse to map the local sampling time of each sensor to a unified system time axis to eliminate phase deviation between different sensing channels and generate raw data stream packets with time sequence labels.

[0048] A global clock source calibration operation is performed on the vibration waveform data, motor current timing data and spectral reflectance data in the multi-channel heterogeneous raw signal set to eliminate sampling phase differences with a unified time reference.

[0049] The local system clocks of each sensor acquisition terminal are synchronized using the network time protocol. The master clock source is a standard time signal provided by a high-stability crystal oscillator module synchronized with GPS.

[0050] The master clock signal is distributed to each acquisition terminal, and an offset vector set is established by recording the clock offset of each acquisition terminal. The offset is calculated as the difference between the master clock sampling time and the local sampling time.

[0051] A hardware trigger pulse method is used to synchronously trigger sampling events at the multi-source acquisition terminals. The rising edge of the trigger pulse serves as the sampling start point, ensuring that the sampling start time of each sensor is consistent.

[0052] Time mapping operations are performed based on the offset vector set to project the local sampling times of each sensor onto the unified system time axis.

[0053] The mapped sampling time is appended to the time identifier field of the corresponding data frame, and a raw data stream packet with a time sequence label is generated.

[0054] By using global clock source calibration and time mapping processing, the multi-channel heterogeneous raw signal set acquired in the previous step is transformed into a raw data stream packet with a unified time index, achieving the technical effect of phase consistency across sensing channels.

[0055] S1.4: Perform sliding window alignment processing based on the original data stream packets with time-series labels. Based on a unified timestamp, interpolate and resample the vibration waveform data, current time-series data, and spectral reflectance data within the same time slice and perform length matching to construct a multi-source original signal stream matrix with strict time-dimensional correspondence.

[0056] S1.5: Perform a data structure encapsulation operation on the multi-source original signal stream matrix to integrate the multi-dimensional signal vectors aligned by the time axis into a serialized data object with a unified index key value, so as to output a time-synchronized multi-dimensional data sequence that can be directly used for subsequent fixed-length slice processing.

[0057] Step S2: Based on the time-synchronized multidimensional data sequence, perform fixed-length slicing processing to truncate the waveform data of each channel within each sampling period into standardized data segments, and map the standardized data segments into initial attribute vectors for graph nodes to generate node attribute sets. Specifically, this includes: The graph structure template is a static graph structure skeleton pre-built and stored during the offline phase based on the physical topology of the screw conveyor, the spatial layout of mechanical functional units, and the energy transfer paths under typical operating conditions. Each graph structure template contains the following three core components: (1) Definition of mechanical functional unit nodes Node set V = {v1, v2, ..., v k Each node corresponds to a key mechanical functional unit of the screw conveyor, including but not limited to: feed section screw node: characterizing the screw state in the raw material input section; middle support bearing node: characterizing the mechanical vibration and current characteristics of the middle support bearing; discharge valve body node: characterizing the material flow and valve body state at the output end.

[0058] Each node is predefined in the template: node id : Unique identifier; node type Mechanical functional unit type (e.g., "feed section screw"); spatial location Spatial coordinates or topological location index; attribute schema : The dimension and data type of the attribute vector to be injected.

[0059] (2) Directed edge connection between nodes The set of directed edges E = {e ij |Node pairs that have energy transfer or physical coupling relationships (v i , v j Each directed edge e ij Indicates from node v i to node v j The direction of energy flow or signal transmission. For example: Feed section screw → middle support bearing: This indicates that the mechanical vibration and current fluctuations generated by the screw rotation are transmitted to the bearing; Central support bearing → discharge valve body: This indicates that the residual energy transmitted through the bearing affects the discharge end.

[0060] Directed edge connections are stored in the form of an adjacency matrix A or an edge list, where A ij = 1 indicates that there exists a value from v i to v j A directed edge ij = 0 indicates no direct connection.

[0061] (3) Edge physical semantic tags Each directed edge e ij Pre-associate a physical semantic label ij This tag describes the physical meaning of the energy transfer or coupling mechanism represented by the edge under specific operating conditions. Example tags include: "Screw-bearing vibration transmission": This refers to the mechanical vibration propagation path from the screw in the feed section to the central support bearing. "Bearing-valve body current coupling": This refers to the coupling effect of current fluctuations in the central support bearing on the control signal of the discharge valve body.

[0062] Template pre-storage and indexing All graph structure templates are stored in the local storage of the edge nodes as structured files (such as JSON, XML, or binary serialization format) and indexed by a "working condition-topology mapping dictionary". The keys of this dictionary are working condition-aware anchor point labels (such as "high humidity, low speed, viscous state"), and the values ​​are the corresponding template identifiers. When a matching working condition label is retrieved in step S4.1, the corresponding template instance can be loaded quickly.

[0063] Relationship between template and node attribute vector The "initial attribute vector of graph nodes" generated by S2 consists of the real-time state values ​​(such as vibration energy and current spectrum characteristics) corresponding to each node. The "dynamic injection of node attributes" operation in S4.2 precisely injects these real-time state values ​​according to the predefined nodes in the template. id and attributes schema The data is then filled into the corresponding node attribute fields of the template, thus forming an "intermediate state of the full attribute graph structure".

[0064] Example In the screw conveyor for paper bag raw materials, a pre-stored graph structure template (corresponding to the "high humidity, low speed, viscous state") contains three nodes: Node N1: Feed section screw (attribute) schema Vibration energy, current spectrum, spectral reflectance coefficient); Node N2: Central support bearing (attribute) schema (Peak vibration acceleration, current harmonic distortion rate); Node N3: Discharge valve body (attribute) schema (Feedback on spectral absorption peak position and valve opening).

[0065] The directed edges are: N1→N2 (labeled: "screw-bearing vibration transmission"), N2→N3 (labeled: "bearing-valve body current coupling"). The adjacency matrix is ​​[[0,1,0],[0,0,1],[0,0,0]].

[0066] When the node attribute vectors generated by S2 are: N1 = [0.82, 0.76, 0.91], N2 = [2.15, 2.10], N3 = [0.45, 0.43], S4.2 injects these values ​​into the template according to the schema to form a complete intermediate state of the full attribute graph structure, which is used for subsequent edge weight calculation and subgraph sampling.

[0067] S2.1: Perform sliding window truncation processing on the time-synchronized multidimensional data sequence, and divide the continuous vibration waveform data, current time series data and spectral reflectance data into independent standardized data segments according to the preset sampling period length, so as to generate a set of multi-channel discrete signal blocks with a unified time span.

[0068] When performing windowing segmentation on the vibration waveform, motor current timing, and spectral reflectance signals contained in the time-synchronized multidimensional data sequence, a continuous index sequence is established on a unified time axis based on a preset sampling period length to determine the start and end boundaries of each sampling period. This index sequence is mapped to each sensor data channel. High-frequency sampling points are truncated in the vibration waveform channel using the same start and end boundaries, the corresponding current value set is extracted in the motor current timing channel, and spectral data with the same time span is extracted in the spectral reflectance channel. Length consistency checks are performed on each data segment after windowing. Insufficient sampling points are adjusted or excessive sampling points are truncated using an interpolation algorithm to ensure that the three types of data segments are completely matched in terms of time span and sample quantity. The calibrated data segments are bound to the time window index and sensor channel identifier respectively, forming a set of multi-channel discrete signal blocks with a unified time span. Continuous traversal through a sliding window can cover the entire time-synchronized multidimensional data sequence, transforming the uniformly aligned data output from the previous step into multi-channel discrete signal blocks with standardized time spans, thus providing the standard input required for subsequent amplitude normalization and high-dimensional feature tensor construction.

[0069] For example, in the scenario of monitoring the operation of a screw conveyor, the preset sampling period length is 0.5 seconds, the sampling frequency of the vibration acceleration sensor is 2000Hz, the sampling frequency of the motor current sensor is 500Hz, and the sampling frequency of the optical thickness sensor is 100Hz. Start and end index boundaries are generated every 0.5 seconds on a unified time axis. Each window in the vibration waveform channel extracts 1000 data points, each window in the current time series channel extracts 250 current values, and each window in the spectral reflectance channel extracts 50 spectral sampling values. For the current time series channel, the 250 points are extended to 1000 points using linear interpolation to match the vibration waveform length; for the spectral reflectance channel, the 50 points are extended to 1000 points. The interpolation method uses the following formula: in, For the target interpolation time point, and Given the sampling time points, The extended sample value is calculated using this formula to correspond to the signal value. After execution, a multi-channel discrete signal block with 1000 points is obtained for each segment. Under this operating condition, the output multi-channel discrete signal block set can be directly used for amplitude normalization processing, and in actual verification, it significantly improves the temporal consistency and feature stability when fusing data from different sensors.

[0070] S2.2: Perform amplitude normalization operation based on the multi-channel discrete signal block set, use the maximum and minimum value scaling method to eliminate the difference in dimensions of different sensors and suppress background noise interference, so as to output a standard normalized signal matrix with consistent numerical distribution.

[0071] S2.3: Perform a dimension reorganization operation on the standard normalized signal matrix, and splice the data segments of each channel into a high-dimensional tensor structure according to the spatial topological order of the key mechanical positions of the screw conveyor, so as to form the original feature tensor representing the instantaneous state of a single mechanical functional unit.

[0072] A spatial topology sequence index mapping operation is performed on each channel data segment in the input standard normalized signal matrix. Based on the preset spatial topology table of the key mechanical points of the screw conveyor, the vibration, current and spectral channels are respectively mapped to the node index positions of the feed section screw, the middle support bearing and the discharge valve body.

[0073] The mapped multi-channel data segments are processed by vector concatenation in the order of node index. The tensor construction method is used to stack the three types of channel data segments axially at the index position of each mechanical functional unit to form local tensor units, while maintaining strict separation between the time dimension and the channel dimension.

[0074] Perform a dimension expansion operation on each local tensor unit and use zero-padding or mirror expansion to unify the length of data segments in each channel to ensure the consistency of data dimensions in each channel during the subsequent construction of high-dimensional tensor structures.

[0075] A cross-channel permutation and combination operation is performed on the extended local tensor unit. The vibration, current and spectrum channels are arranged in a fixed channel order in the channel dimension position of the high-dimensional tensor structure through a preset permutation matrix, thereby generating the original feature tensor unit that represents the instantaneous state of the mechanical functional unit.

[0076] All the original feature tensor units of the mechanical functional units are stacked in parallel along the batch dimension according to the node order of the spatial topology table to form a complete high-dimensional tensor structure, which is used as the output original feature tensor of this sub-step.

[0077] Through the aforementioned dimensional recombination operation, the standard normalized signal matrix is ​​transformed into a structured high-dimensional tensor structure with consistent spatial topological relationships, enabling accurate representation of the instantaneous state of a single mechanical functional unit and providing high-quality input for the subsequent linear projection transformation to generate compact state encoding vectors.

[0078] For example, in one sampling cycle, the feed section vibration waveform channel segment length is 1024, the current timing channel segment length is 800, and the spectral reflection channel segment length is 512, all with normalized values ​​ranging from -1 to 1. After performing spatial topological sequence index mapping, the three channels are assigned to node index 0 of the feed section screw, node index 1 of the central support bearing, and node index 2 of the discharge valve body, respectively. When extending the local tensor unit dimension, the mirror expansion method is used to extend the current channel length from 800 to 1024 and the spectral channel length from 512 to 1024. In the permutation and combination operation, the preset permutation matrix is ​​[1,2,3], indicating that the vibration channel is arranged first, the current channel is arranged second, and the spectral channel is arranged third. The final generated original feature tensor unit dimension is (3,1024), and stacking three node units along the batch dimension yields the complete high-dimensional tensor structure dimension (3,3,1024). In model validation, after subsequent linear projection dimensionality reduction, the state encoding vector of this high-dimensional tensor can significantly improve the stability and sensitivity of micro-anomaly identification in fault trend identification primitive library matching, meeting the technical requirements of low-computing-power devices for accurate characterization of high-channel states.

[0079] S2.4: Perform a linear projection transformation based on the original feature tensor, and compress the high-dimensional time-domain signal into a low-dimensional latent space through a predefined weight mapping matrix to generate a compact state encoding vector containing mechanical vibration energy and material flow characteristics.

[0080] The goal of performing low-dimensional latent space mapping on the original feature tensor representing the instantaneous state of a single mechanical functional unit is to reduce data dimensionality while preserving physical features highly correlated with fault trend prediction. The input is a multi-channel, high-dimensional time-domain feature tensor formed by amplitude normalization and splicing in spatial topological order, containing a sequence of sampling points for vibration waveforms, current time series, and spectral reflectance. The original feature tensor is centered, and the mean of each dimension is subtracted to eliminate the influence of offset on the projection result.

[0081] Calculate the characteristic covariance matrix for the centered tensor and evaluate the correlation between signals from each channel.

[0082] Based on the covariance matrix, eigenvalue decomposition is performed to extract eigenvectors and construct a weight mapping matrix, which is used for linear transformation.

[0083] Matrix multiplication is used to multiply the high-dimensional input with the weight mapping matrix to realize the projection operation of the feature vector in the latent space.

[0084] The projection results are normalized to ensure that the encoding vectors of different mechanical functional units are comparable in numerical distribution.

[0085] By eliminating low-contribution projected components through feature selection, only the core components reflecting mechanical vibration energy and material flow characteristics are retained.

[0086] By using linear projection matrix compression, the original feature tensor constructed in the previous step is transformed into a compact state encoding vector containing fault-sensitive features, thereby reducing computational load and improving the efficiency of subsequent graph structure template injection.

[0087] For example, in a screw conveyor for paper bag raw materials, the sampling period is 50ms, and the original feature tensor size is (3 channels × 1024 sampling points). After amplitude normalization and centering adjustment, a covariance matrix is ​​obtained. The covariance matrix is ​​3×3 with diagonal elements of 46.2, 39.8, and 51.5. Eigenvalue decomposition yields an eigenvector matrix W, where the first row has elements of (0.58, 0.41, 0.71), the second row has elements of (0.63, 0.59, 0.46), and the third row has elements of (0.52, 0.69, 0.49). A linear projection operation is performed on the original feature tensor X and W, resulting in a compact state encoding vector. The dimension is reduced from the original 1024 to 256, preserving both the high amplitude characteristics of the vibration energy peak in the feed section within the latent space and the low-frequency energy distribution index of the discharge valve's spectral signal. After normalization, the numerical distribution of the encoding vector tends to be consistent among different mechanical functional units, which can significantly improve the matching stability under both high humidity and low speed viscous state and dry powder high speed segregation state conditions when injecting the graph structure template.

[0088] S2.5: Perform structured encapsulation processing on the compact state coding vector, bind it to the corresponding mechanical functional unit identifier and add metadata tags, so as to finally output the graph node initial attribute vector set that can be directly injected into the graph structure template.

[0089] like Figure 2 As shown, step S3 involves: based on the physical property parameters of the paper bag raw material and the equipment operating status parameters, performing symbolic analysis on the current operating condition to generate operating condition sensing anchor tags. These anchor tags represent specific operating modes, such as a high-humidity, low-speed viscous state or a dry powder, high-speed segregation state. Specifically, this includes: S3.1: Perform multi-source parameter extraction processing on the spectral reflectance data detected by the optical thickness sensor at the feed inlet of the screw conveyor and the current time series data recorded by the motor current sampling module to obtain a set of original working condition parameters, including the nominal range of raw material moisture content, fiber length distribution characteristics, instantaneous screw speed, fixed conveying angle, and instantaneous load fluctuation.

[0090] Spectral feature decomposition is performed on the spectral reflectance data detected by the optical thickness sensor at the feed inlet. The absorption peak positions and peak areas related to the moisture content of the raw material are identified by a combination of multi-peak fitting and band integration.

[0091] The identified spectral peaks are converted into physical quantities to calculate the nominal range of moisture content in the raw materials.

[0092] Multi-scale statistical analysis was performed on the spectral reflectance curve of the raw material to extract the peak width to peak height ratio in the preset fiber sensitive band, which was used as the input index for fiber length distribution characteristic value.

[0093] Instantaneous spectrum transformation is performed on the motor current timing data. Fast Fourier transform is used to extract the instantaneous value of the screw speed corresponding to the main current frequency, and the fixed value of the conveying tilt angle is directly assigned in combination with the equipment structural constant.

[0094] An envelope curve is established for the current time series data. A sliding window is used to calculate the local extreme value difference and quantify the instantaneous load fluctuation value, forming a physical quantity characterization of the material input fluctuation.

[0095] Through the above chain processing method, the spectral reflectance data and current time series data are transformed into a set of original working condition parameters, including the nominal range of raw material moisture content, fiber length distribution characteristics, instantaneous screw speed, fixed conveying angle, and instantaneous load fluctuation. This enables the generation of the full physical state input required for subsequent working condition sensing anchor points.

[0096] For example, in the actual operation of a paper bag raw material conveying equipment, the spectral reflectance data collected by the optical thickness sensor at the feed inlet is decomposed to obtain an absorption peak with a center wavelength of 1450nm and a peak area of ​​12.5 units. Using the formula... The nominal moisture content range calculated using H = 0.8 × 12.5 + 3.2 is 13.2%. Spectroscopic analysis yielded a peak-to-width ratio of 2.1 in the 1300-1600 nm band, from which the fiber length distribution characteristic value was estimated to be 0.65. FFT analysis of the motor current time series data showed a main frequency of 50 Hz, corresponding to an instantaneous screw speed of 3000 rpm, with the conveying angle fixed at 15 degrees based on structural parameters. Load fluctuation analysis yielded an instantaneous load fluctuation value of 0.08 kN. This initial set of operating parameters was accurately parsed as a "medium-moisture, medium-speed stable state" in the subsequent rule engine for matching operating condition anchor points. Verification results showed that this anchor point label maintained high consistency when repeatedly generated under the same batch of raw materials, significantly improving the stability of early failure trend prediction.

[0097] S3.2: Based on the original set of operating condition parameters, perform multidimensional threshold boundary mapping operation, and use a predefined operating condition division rule library to divide the continuously changing nominal range of raw material moisture content and instantaneous screw speed into multiple discretized state intervals such as high moisture, medium moisture, dry powder, low speed, medium speed, and high speed, so as to generate a single-dimensional state classification vector with clear physical boundaries.

[0098] The operating condition classification rule base is a knowledge base pre-built in the offline stage based on the physical properties of the paper bag raw materials (moisture content, fiber length distribution) and the operating mechanism of the screw conveyor (speed, load fluctuation, conveying angle). Its core function is to convert real-time sensor data into standardized operating condition labels (such as "high humidity, low speed, viscous state" or "dry powder, high speed, segregation state"), providing semantic anchors for subsequent graph structure template matching. The rule base consists of the following two parts: I. One-dimensional state discretization rules This section defines the discretization interval boundaries for each continuous physical parameter. Each parameter is divided into several discrete state intervals based on its physical meaning, specifically including: Raw material moisture content: Based on the moisture content values ​​obtained from the spectral reflectance data of the optical thickness sensor, it is divided into three discrete states: high moisture (moisture content not less than 15%), medium moisture (moisture content between 8% and 15%), and dry powder (moisture content less than 8%).

[0099] Screw speed: It is obtained by extracting the main frequency from the motor current timing data through fast Fourier transform and is divided into three discrete states: high speed (speed above 60 rpm), medium speed (speed between 30 and 60 rpm), and low speed (speed below 30 rpm).

[0100] Fiber length distribution characteristic value: calculated from the ratio of peak width to peak height of spectral reflectance data in the preset fiber sensitive band, and divided into three discrete states: long fiber (characteristic value greater than 0.7), medium fiber (characteristic value between 0.4 and 0.7), and short fiber (characteristic value less than 0.4).

[0101] Instantaneous load fluctuation value: obtained by quantifying the local extreme value difference of the motor current envelope curve, and divided into three discrete states: high fluctuation (fluctuation value greater than 0.12 kN), medium fluctuation (fluctuation value between 0.05 and 0.12 kN), and low fluctuation (fluctuation value less than 0.05 kN).

[0102] The above-mentioned boundary values ​​can be adjusted offline based on the equipment model, raw material batch, and on-site calibration results, but do not rely on online learning.

[0103] II. Logic Combination Rules for Complex Operating Conditions This section combines single-dimensional discrete states using Boolean logic, mapping them to composite operating condition labels with clear physical meanings. Each rule uses an IF statement. The THEN form, as shown in the example below: High humidity, low speed, viscous state: If the moisture content is high, the screw speed is low, the fiber length characteristic value is long fiber or medium fiber, and the load fluctuation is not lower than medium fluctuation, then output the working condition label "HS_LS_Stick".

[0104] High-speed segregation of dry powder: If the moisture content is dry powder, the screw speed is high, the fiber length characteristic value is short fiber, and the load fluctuation is low fluctuation, then the operating condition label "DP_HS_Sep" will be output.

[0105] Medium Moisture Medium Speed ​​Steady State: If the moisture content is medium, the screw speed is medium, and the load fluctuation is medium or low, then the operating condition label "MM_MS_Stable" will be output.

[0106] Other combined operating conditions can be expanded based on actual operational experience. The rule base is stored in the non-volatile memory of the edge nodes in the form of lookup tables or decision trees. When S3 performs symbolic parsing, it first maps the real-time parameters to one-dimensional discrete states, then traverses the logical combination rules, matches the first rule that meets the conditions, and outputs the corresponding operating condition perception anchor label. If no rule matches, it outputs the "unknown operating condition" label and triggers the default conservative control strategy.

[0107] Example: A sampling yielded a moisture content of 17% (high humidity), a rotation speed of 28 rpm (low speed), a fiber length characteristic value of 0.72 (long fiber), and a load fluctuation of 0.10 kN (medium fluctuation). The rule base matched the "high humidity, low speed, viscous state" rule, outputting the label "HS_LS_Stick". Simultaneously, the confidence level (0.95 in this example) can be calculated based on the sensor signal-to-noise ratio and boundary distance.

[0108] Through the pre-built working condition classification rule library mentioned above, this solution can quickly and interpretably transform the original physical parameters into working condition semantic labels without relying on complex machine learning models, providing reliable and stable anchor points for graph structure template matching.

[0109] When performing multidimensional threshold boundary mapping operation on the original set of operating parameters, the input object is the original set of operating parameters containing continuous measured values ​​such as the nominal range of raw material moisture content and the instantaneous value of screw speed.

[0110] Using the moisture content boundary set in the working condition division rule library, the nominal range of raw material moisture content is calculated to perform interval positioning. Based on the preset upper and lower humidity thresholds, the measured value is mapped to the discrete humidity level of high humidity, medium humidity or dry powder.

[0111] The rotational speed boundary set in the same rule base is called to perform interval positioning calculation on the instantaneous value of the screw rotational speed. The low-speed, medium-speed, and high-speed segment boundaries within the design rotational speed range of the screw conveyor are used to map the measured value to the corresponding discretized speed level.

[0112] Humidity level and speed level are converted into one-dimensional state classification vector elements, and the parameter type is corresponding to the index position in the internal structure to form a one-dimensional state classification vector with clear physical boundaries.

[0113] By using this discretization method based on physical quantity thresholds, the continuous operating condition measurement values ​​from the previous step are transformed into well-defined state vectors, thus achieving input standardization for the subsequent logical combination coding process.

[0114] S3.3: Perform logical combination encoding operation on the single-dimensional state classification vector, use Boolean algebra and state machine transformation processing method to couple and associate the high humidity state interval and the low speed state interval, and combine the interaction relationship between fiber length distribution characteristic value and instantaneous load fluctuation value to construct a composite working condition logical expression that characterizes complex coupling effects such as high humidity low speed viscous state or dry powder high speed segregation state.

[0115] The initial conditions for performing logical combination encoding on the single-dimensional state classification vector are: moisture state labels such as raw material moisture content state range (high moisture, medium moisture, dry powder, etc.) obtained through multi-dimensional threshold boundary mapping, and speed labels such as low speed, medium speed, high speed, etc. for screw speed state range, all possessing numerical descriptions of fiber length distribution characteristics and instantaneous load fluctuation values. The moisture and speed states are input into the Boolean algebra combination processing module, and the coupling association of state labels is achieved using logical operators and integrated circuit-level state logic units. For the moisture state labels... With speed status label Establish Boolean functions The calculation result is the state truth value that meets the target combination condition. The Boolean truth value is then compared with the fiber length distribution characteristic value. Instantaneous load fluctuation value The input set is composed to construct the multi-input condition matrix for the state machine transition processing method. The output state of the composite operating condition under different combinations of conditions is defined through a state transition table, for example, when... For high humidity and low speed true value, at the same time Located in the long fiber range and When the system is in a high-fluctuation range, the state machine outputs "high humidity, low speed, viscous state". A matrix indexing method is used to map the combined conditions to a set of complex operating condition logical expressions, ensuring that each logical expression is based on a clear physical parameter coupling relationship. Through this processing method, the single-dimensional state classification vector from the previous step is transformed into a complex operating condition logical expression containing the interaction relationships of humidity, speed, fiber characteristics, and load fluctuations, achieving precise symbolization of complex operating modes.

[0116] S3.4: Based on the composite working condition logic expression, perform symbolic tag generation processing to convert each determined composite working condition logic expression into a unique string identifier, so as to output a standardized working condition perception anchor point tag sequence.

[0117] When performing symbolic label generation processing on the logical expression of the composite working condition, a composite logical vector containing the raw material moisture content state, screw speed state, fiber length distribution characteristics and instantaneous load fluctuation characteristics is used as input.

[0118] The composite logic vector is parsed one by one, and a mapping index relationship between state combinations and symbolic labels is established. Each composite logic vector is then converted into a corresponding string identifier.

[0119] During the mapping process, hash encoding is used to check the uniqueness of string identifiers, ensuring that there are no conflicts in the symbol space for different working condition logic combinations.

[0120] The mapping results are standardized by formatting, unifying string naming rules, adding working condition category prefixes and attribute separators, and ensuring that the tag structure can be directly parsed by the subsequent topology template retrieval module.

[0121] The standardized string identifiers are sorted and indexed according to time series to form a sequence of condition-aware anchor tags with sequential information.

[0122] Through the above mapping and formatting processes, the composite working condition logical expression constructed in the previous step is transformed into unique and standardized symbolic label data, thereby achieving the interpretability and structured retrieval of the labels.

[0123] S3.5: Perform confidence verification and formatting encapsulation operations on the condition sensing anchor tag sequence, evaluate the reliability of tag generation based on the signal-to-noise ratio of each sensor data at the current time, and attach timestamp metadata to finally generate an interpretable condition sensing anchor tag object that can be directly used for graph structure template matching.

[0124] like Figure 3 As shown, step S4 involves: matching the nearest graph structure template based on the condition-sensing anchor point label, dynamically injecting the node attribute set into the corresponding mechanical functional unit node in the graph structure template, and simultaneously calculating edge weights based on the signal cross-correlation coefficient between adjacent nodes and the matching confidence of the condition-sensing anchor point label, thereby constructing a dynamic graph with physical semantic constraints. Specifically, this includes: S4.1: Perform template index retrieval processing on the working condition sensing anchor tag object, and use the preset working condition-topology mapping dictionary to match the string identifiers such as high humidity low speed viscous state or dry powder high speed segregation state to the nearest basic topology frame in the pre-stored graph structure template library, so as to generate a candidate graph structure template instance containing the mechanical functional unit node definition and initial connection relationship.

[0125] The input conditions include an interpretable condition-aware anchor tag object output by step S3, which contains a unique string identifier, timestamp metadata, and confidence information; and a pre-stored graph structure template library, in which each template contains the definition of mechanical functional unit nodes and initial connection relationships, and the index relationship is maintained by a condition-topology mapping dictionary.

[0126] Perform string identifier decoding on the condition-aware anchor tag object to extract the physical operation mode code corresponding to the complex composite condition logic, ensuring that the index retrieval matches the standardized status tag.

[0127] Based on the decoded physical operation mode encoding access condition—topology mapping dictionary entry, retrieve the mapping key that is consistent with or closest to the encoding, and obtain the corresponding template library index value.

[0128] The corresponding entry in the graph structure template library is called using the template library index value to read its basic topology framework information, including the list of mechanical functional unit nodes, node spatial coordinates, and the initial directed edge connection matrix between nodes.

[0129] A similarity calculation method is used to verify the matching between the current working condition mode and the template framework, and the Euclidean distance of the physical coupling feature vector is calculated to quantify the matching degree.

[0130] By using template index retrieval and matching calculation, the working condition perception anchor point label object is mapped to a candidate graph structure template instance containing complete mechanical functional unit node definitions and initial connection relationships, realizing the unambiguous transformation from working condition to topology structure, and providing an accurate structural carrier for subsequent dynamic injection of node attributes.

[0131] For example, when the paper bag raw material screw conveyor system is operating in a high-humidity, low-speed, viscous state, the string identifier of the condition sensing anchor point tag object is "HS_LS_Stick", the timestamp is 20240601093000, and the confidence level is 0.92. After decoding, the physical operation mode encoding vector is obtained [humidity = high, speed = low, material state = viscous]. In the condition-topology mapping dictionary, this encoding matches the index value Tmpl. 03 The corresponding template includes three nodes: the feed section screw, the central support bearing, and the discharge valve body, as well as the initial connection matrix [[0,1,0],[1,0,1],[0,1,0]]. The Euclidean distance between the physical coupling feature vector of the current mode and the template feature vector is calculated. The humidity difference is 0, the velocity difference is 0, and the material state difference is 0.1. The calculated result is 0.1, which is lower than the preset matching threshold of 0.15, thus it is determined to be the optimal match. A candidate graph structure template instance is generated for subsequent node attribute injection processing in S4.2. In this scenario, template index retrieval achieves fast and accurate condition-topology mapping under low computing power conditions, significantly improving the structural matching efficiency for fault trend prediction.

[0132] S4.2: Based on the candidate graph structure template instance, perform a dynamic node attribute injection operation. Fill the initial attribute vector set of graph nodes generated in the previous steps into the corresponding feed section screw, middle support bearing and discharge valve body node positions in the candidate graph structure template instance according to the mechanical functional unit identifier binding relationship, so as to generate a full attribute graph structure intermediate state with real-time state characteristics.

[0133] The input conditions include the condition-aware anchor point label object output by S3.5, the candidate graph structure template instance output by S4.1, and the initial attribute vector set of graph nodes generated in the previous steps. Based on the node definition and identifier mapping of the candidate graph structure template instance, node attribute binding index construction processing is performed to establish a bidirectional lookup table between mechanical functional unit identifiers and node initial attribute vectors. Attribute vector matching and verification operations are performed on this lookup table to determine whether each mechanical functional unit node has a corresponding valid initial attribute vector. If a missing vector is found, the missing data filling module is called to generate a complete vector. The complete initial attribute vector set after matching and verification is injected into the corresponding feed section screw, middle support bearing, and discharge valve body node attribute fields of the candidate graph structure template instance according to the lookup table index order, ensuring the consistency of node attributes with the spatial topology of the mechanical functional unit. Consistency standardization processing is performed on the injected node attribute fields, using a linear normalization strategy to map each node attribute value to a unified numerical range to eliminate differences in different feature dimensions. A vector encapsulation mechanism is employed to bind standardized attribute values, node identifiers, timestamp metadata, and operating condition anchor tags into a unified set of node attribute data objects, updating candidate graph structure template instances to form an intermediate state of a full-attribute graph structure. Through dynamic node attribute injection processing, the results of the previous step are transformed into graph structure data containing real-time state characteristics, achieving the expected technical effect of embedding a set of node attributes highly consistent with the current operating condition and sensor acquisition status into the template instance.

[0134] For example, in the operating scenario of a paper bag raw material screw conveyor with a conveying angle of 15° and a screw speed of 60 rpm, the working condition perception anchor point label object is "high humidity, low speed, viscous state". The candidate graph structure template instance contains three mechanical functional unit node identifiers: N1 (feed section screw), N2 (middle support bearing), and N3 (discharge valve body). In the initial attribute vector set of the preceding nodes, the vibration feature vector corresponding to N1 is [0.82, 0.76, 0.91], the current feature vector corresponding to N2 is [2.15, 2.10, 2.20], and the spectral reflectance feature vector corresponding to N3 is [0.45, 0.43, 0.47]. After performing a matching check, it was confirmed that each node possessed a valid attribute vector. Subsequently, each vector was normalized; for example, the normalized feature vectors for N1 were 0.9, 0.84, 1.0; for N2, 0.98, 0.96, 1.0; and for N3, 0.96, 0.92, 1.0. The normalized feature vectors were then bound to the node identifier and the current timestamp to generate a set of node attribute data objects, which were then injected into the corresponding node position of the template instance. The intermediate state of the full attribute graph structure after this injection process contains the standardized real-time state features of the three nodes, which can be directly used in the cross-correlation coefficient calculation stage. Verification results show that this state set significantly improves the stability of energy transfer efficiency calculation in sliding window cross-correlation analysis.

[0135] S4.3: Perform cross-correlation analysis on the standardized data segments between adjacent mechanical functional unit nodes in the intermediate state of the full attribute graph structure, and use the sliding window cross-correlation processing method to calculate the phase lag and amplitude correlation between vibration waveform data and current time series data in the time-synchronized multidimensional data sequence, so as to generate the original cross-correlation matrix characterizing energy transfer efficiency.

[0136] Cross-correlation analysis is performed on standardized data fragments of adjacent mechanical functional unit nodes in the intermediate state of the full-attribute graph structure. The analysis object is defined as the node attribute vector dynamically injected by S4.2, derived from vibration waveform data and current time-series data in a time-synchronized multidimensional data sequence. Based on the binding relationship between nodes, standardized data fragments corresponding to each node are extracted. Under a unified time axis, a sliding window cross-correlation processing method is used to construct signal pairs of continuous time segments. For each signal pair, mean removal is performed to eliminate DC components. By calculating the cross-correlation function values ​​under different time delays within the sliding window, the correlation curves of the signal pairs under various delay parameters are obtained. The peak position of the correlation curve is determined and the peak amplitude is quantified. The peak position is mapped to the phase lag, and the peak amplitude is mapped to the amplitude correlation degree. Combining the phase lag and amplitude correlation degree of all node pairs, a two-dimensional matrix representing the original cross-correlation coefficient matrix of energy transfer efficiency is constructed. The cross-correlation formula is then used. in The average value of the signal. For lazy indexing, and These are standardized data segments between nodes. The cross-correlation coefficients within the window are calculated using this formula, and the peak positions and amplitudes are determined according to different delays, thus quantifying the correlation between phase lag and amplitude. Through cross-correlation analysis, the intermediate state of the full-attribute graph structure from the previous step is transformed into the original cross-correlation coefficient matrix characterizing energy transfer efficiency, achieving a precise characterization of the physical signal coupling strength and delay characteristics between mechanical functional units.

[0137] S4.4: Perform weighted fusion calculation based on the confidence verification results of the original cross-correlation matrix and the working condition sensing anchor tag object. Use a linear weighted fusion strategy to couple the values ​​in the original cross-correlation matrix with the signal-to-noise ratio evaluation values ​​attached to the working condition sensing anchor tag object to generate a set of composite edge weight values ​​that have both signal physical correlation and working condition matching reliability.

[0138] Based on the confidence verification results of the original cross-correlation matrix and the condition-sensing anchor point label objects, the cross-correlation vector between the corresponding mechanical functional unit nodes in the matrix is ​​selected as the initial signal physical correlation input. This vector is normalized and matched according to the signal-to-noise ratio (SNR) evaluation value sequence attached to the condition-sensing anchor point objects to ensure that the SNR values ​​from different sensor sources are distributed within a uniform scale. A linear weighted fusion strategy is adopted to construct fusion weight coefficients to couple the cross-correlation coefficients and SNR calculations. The values ​​of the weight coefficients are jointly set based on the stability index of the condition label confidence and the cross-correlation coefficients. After the fusion coefficients are generated, the fusion values ​​of all edges are matrix-encapsulated, mapping the fusion results to a composite edge weight value set according to the node connection relationship, and attaching a condition matching confidence metadata tag for subsequent edge weight update operations. Through the above processing method, the cross-correlation matrix and condition label SNR evaluation values ​​from the previous step are transformed into a composite edge weight value set that combines signal physical correlation and condition matching reliability, achieving highly adaptive construction of lightweight dynamic graph edge attributes.

[0139] S4.5: Perform edge weight attribute update and structured encapsulation processing on the intermediate state of the full attribute graph structure, assign the set of composite edge weight values ​​to the directed edge attribute fields connecting the nodes of each mechanical functional unit in the intermediate state of the full attribute graph structure, and add physical semantic constraint metadata tags to finally output a dynamic graph with physical semantic constraints that can be directly used for subgraph sampling analysis.

[0140] The directed edge attribute field assignment operation is performed on the set of composite edge weight values ​​that have been calculated in the intermediate state of the full attribute graph structure. The connection relationship between each set of edge weights and the corresponding mechanical functional unit nodes is strictly indexed and matched to ensure that the node index in the assignment process is completely consistent with the template definition.

[0141] Metadata injection processing is performed on the assigned directed edges, and the physical semantic identifier corresponding to the working condition perception anchor point label is written as an additional attribute into the edge attribute field to preserve the energy transfer semantics of the edge under different working conditions.

[0142] A structured encapsulation operation is performed on the set of edge attributes containing composite edge weights and physical semantics. The graph data serialization protocol is used to transform it into a unified format object that conforms to the subsequent subgraph sampling and analysis interface, ensuring that nodes, edges and attributes remain consistent in the data structure.

[0143] The structured graph structure is subjected to consistency verification. The topology integrity verification method is used to detect the legality of edge direction, node index and weight value domain, so as to eliminate graph structure anomalies caused by assignment or encapsulation process.

[0144] The output is a lightweight dynamic graph object that passes the consistency check. It serves as the direct input for subsequent sliding window subgraph sampling. It has physical semantic constraints and can accurately characterize the energy interaction state between mechanical functional units of the screw conveyor under specific working conditions.

[0145] By assigning directed edge attributes and injecting physical semantic metadata, the composite edge weight set from the previous step is transformed into structured and interpretable graph edge attribute data, thereby achieving the usability and robustness of lightweight dynamic graphs in cross-condition fault trend prediction.

[0146] For example, under the high-humidity, low-speed viscous state of paper bag raw materials, the composite edge weight value set of the feed section screw node and the central support bearing node has been calculated, where the composite weight value of this connecting edge is 0.78 and the signal-to-noise ratio evaluation value is 0.92. During the assignment operation, 0.78 is bound to the directed edge attribute field in the template, and the physical semantic tag "high-humidity, low-speed viscous state" is injected. The structured encapsulation adopts a node-edge-attribute JSON format, where the node indices are the feed section screw (ID=1) and the central support bearing (ID=2), the directed edge is defined as ID=1→ID=2, and the attribute field includes the weight value 0.78, the signal-to-noise ratio 0.92, and the physical semantic tag. The topology integrity verification method checks that the directed edge exists in the template structure and has the correct direction, and that the weight value is within the allowable range 0≤w≤1. The final output lightweight dynamic graph object is used for sliding window subgraph sampling in step S5. The verification results show that the energy transfer path evolution trend description under this condition is accurate and the interpretability is enhanced.

[0147] Step S5: Perform a sliding window subgraph sampling operation on the dynamic graph to extract local topological change vectors between consecutive time steps. These local topological change vectors characterize the evolution trend of the energy transfer path between the feed section nodes and the central support nodes. Specifically, this includes: S5.1: Perform time series indexing on the dynamic graph with physical semantic constraints, and arrange the graph instances in the order of collection according to the unified timestamp to generate a time series graph queue, so as to establish a continuous time reference axis for sliding window sampling.

[0148] When performing time-series indexing on lightweight dynamic graph instances with physical semantic constraints, the graph instances output from step S4.5 are used as input. Each instance contains structured data consisting of node attributes and edge weights, as well as unified timestamp metadata. For all graph instances, the timestamp parsing module is first called to read the timestamp values ​​and verify the format validity, ensuring that the time information uses a unified millisecond precision encoding. Then, the timestamp values ​​are sent to the sorting engine, and all graph instances are globally sorted according to the timestamp in ascending order, forming a sequence set that increases in sequence according to the acquisition order. Subsequently, a continuity check is performed on the sorted sequence set, using the difference between adjacent timestamps to compare with the preset acquisition period to determine whether there are time jumps or missing values. When a discontinuity is detected, an empty graph placeholder is inserted to maintain the integrity of the time reference axis. Finally, the graph sequences that have undergone continuity correction are bound to a unified sequence index number, and a one-to-one correspondence between the time axis and the acquired sequences is achieved through index number mapping. Finally, the time series graph queue with established continuous time reference axes is encapsulated into a data structure object that can be called by sliding window sampling, ensuring the temporal continuity and reference stability of subsequent S5.2 truncation operations. Through index sorting, continuity correction, and index mapping, the graph instance from the previous step is transformed into a time series graph queue with continuous time reference axes, achieving temporal consistency and data integrity in the sliding window sampling process.

[0149] For example, in the scenario of fault diagnosis for the spiral conveyor of paper bag raw materials, the lightweight dynamic graph instance generated in step S4.5 includes unified timestamp metadata (in milliseconds), node attributes, and edge weight fields. The timestamp parsing module is configured to support millisecond-level precision, and the sorting engine is set to ascending order based on integer timestamp values. The acquisition cycle is set to 500ms. The continuity detection module calculates the absolute error of the difference between adjacent instance timestamps compared to 500ms. If the error is greater than 501000 seconds, an empty graph placeholder is inserted. When there are two instance timestamps of 160000 and 161050, with a difference of 1050ms, the detection module inserts an empty graph instance at the 160500ms position to maintain timeline continuity. The index binding module assigns index value 1 to the first instance, index value 2 to the second instance, and so on, forming a sequence index mapping table. After executing the index construction step, the time-series graph queue can be windowed in the subsequent S5.2 with a time span of 3 acquisition cycles. Under different operating conditions, the sampling window is guaranteed to cover the graph instances of the current and previous two cycles. The verification results show that sliding window subgraph sampling can significantly improve the temporal coherence of local topological structure change vectors and the accuracy of evolution trend determination.

[0150] S5.2: Based on the time series graph queue, perform a fixed span sliding window truncation operation, and use a preset time step parameter to select a set of subgraph sequences containing the current time and several previous times in the time series graph queue, so as to generate a local subgraph observation window with temporal continuity.

[0151] For the input time series graph queue, a selection logic for the truncation range is constructed based on a preset time step parameter to ensure the continuity of the selected subgraph sequence set on the time axis and the coherence of the physical state.

[0152] Using the time step parameter, index positioning operations are performed in the time series map queue to generate a time index set covering the current time and several previous acquisition times. This set is then constrained to a fixed span window to ensure that the time base for subsequent differential operations is consistent.

[0153] Perform a graph instance extraction operation on the time index set, call the queue pointer to locate lightweight dynamic graph instances at different time steps, and read the corresponding node and edge weight data one by one according to the window span.

[0154] During the extraction process, boundary integrity checks are performed on the start and end boundaries of the window to ensure that all time step map instances within the window fully contain the condition-aware anchor point labels and the actual measured node attribute data, in order to avoid null values ​​or missing values ​​in subsequent local topology analysis.

[0155] The subgraph sequence after boundary verification is encapsulated in a structured manner, combining the graph instances at each time step into local subgraph observation window objects in chronological order, and attaching time span metadata tags to indicate the physical duration of the window.

[0156] By using a fixed-span sliding window truncation method, the time series graph queue generated in the previous step is transformed into a local subgraph observation window with temporal coherence, thereby enabling controllable sampling of local topological changes between continuous time steps.

[0157] For example, during the operation of a screw conveyor for paper bag raw materials, the time-series graph queue contains 100 collection points arranged with a uniform timestamp, and the preset time step parameter is 5 steps. When performing index positioning, an index set covering the current time and the previous 4 times is generated, for example, the index sequence is [96,97,98,99,100]. The lightweight dynamic graph instance corresponding to this index is extracted. Each instance contains three nodes: the feed section screw, the middle support bearing, and the discharge valve body, as well as the edge weight data between each node. Boundary integrity verification confirms that the above five graph instances all have condition-aware anchor point labels and full attribute data. When packaging, an additional time span label value of 5 × 0.2 seconds, i.e., 1 second of physical duration, is added. This local subgraph observation window object can be used in subsequent difference operations to calculate the change in edge weight between the feed section node and the middle support node within this second, obtain the evolution characteristics of the instantaneous energy transfer path, and thus significantly improve the accuracy and stability of fault trend prediction.

[0158] S5.3: Perform edge weight difference operation between nodes on the adjacent time step subgraphs within the observation window of the local subgraph, and use the point-by-point subtraction method to calculate the change in weight value of the connecting edge between the feed section node and the middle support node in continuous time steps, so as to generate an edge weight difference matrix that characterizes the instantaneous fluctuation of energy transfer efficiency.

[0159] For adjacent time-step subgraphs within the local subgraph observation window, a lightweight dynamic graph instance with physical semantic constraints generated in the preceding steps, along with its edge weight attribute field, is loaded as input. An index extraction operation is performed on the connection edge weight sequence between the feed section node and the central support node, constructing an ordered vector set containing weight values ​​from multiple time steps. The weight values ​​for each time step are normalized according to a unified unit system to eliminate dimensional differences between different acquisition batches and ensure the accuracy of the difference calculation. A point-by-point subtraction operation is performed based on the normalized weight vector set. The differences are stored sequentially in the corresponding row and column positions of the difference matrix according to the time step index, forming a two-dimensional numerical array structure representing the instantaneous changes in continuous time steps. An outlier detection operation is performed on the difference matrix, using a preset threshold to identify physically unreasonable mutations for subsequent filtering, improving the reliability of evolution trend analysis. Through differential processing, the continuous time-step weight value vector is transformed into an edge weight difference matrix representing the instantaneous fluctuations in energy transfer efficiency, achieving the quantification and availability of weak changes in fault signals and subsequent feature extraction.

[0160] For example, under the high-humidity, low-speed, viscous operating condition of the paper bag raw material screw conveyor, the edge weight values ​​between the feeding section node and the middle support node are collected. The time span is set to 5 seconds, the sampling interval is 0.5 seconds, and a total of 10 time step weight values ​​are obtained. After normalization to the 0~1 interval, the sequence [0.82, 0.80, 0.77, 0.75, 0.74, 0.72, 0.71, 0.69, 0.68, 0.66] is obtained. The change in each time step is calculated using difference, and the first column of the difference matrix is ​​[-0.02, -0.03, -0.02, -0.01, -0.02, -0.01, -0.02, -0.01, -0.02], and the remaining columns are empty or zero values ​​used for matrix alignment and filling. After outlier detection, no physically unreasonable abrupt changes are found, and the matrix directly enters the subsequent S5.4 phase lag feature extraction process. The difference matrix in this scenario showed a gradual decay and stable amplitude trend in subsequent analysis, which significantly improved the system's sensitivity to the slow decline in energy transfer efficiency caused by early bearing wear. In experimental verification, it achieved the effect of issuing an early warning several minutes in advance.

[0161] S5.4: Based on the edge weight difference matrix, perform topological phase lag feature extraction processing, and use the cross-correlation peak offset detection method to analyze the change trajectory of signal propagation delay between corresponding nodes of vibration waveform data and current time series data, so as to generate a phase lag evolution vector that reflects the mechanical coupling state drift.

[0162] Signal synchronization evaluation is performed on the continuous time step weight change data in the edge weight difference matrix. A time domain pairing relationship is established for the vibration waveform data and current time series data corresponding to each pair of nodes to ensure that the two types of signals are cross-correlation calculated on a unified sampling reference.

[0163] The paired signal sequence is input into the cross-correlation peak detection unit, and the correlation coefficient curve under each time delay is calculated using the sliding cross-correlation function. The signal propagation delay is obtained by locating the global peak position of the correlation coefficient curve.

[0164] Perform a difference operation on the delay sequence at different time steps, and obtain the delay change vector by subtracting point by point.

[0165] The delay change vector is smoothed by filtering, and the moving average method is used to eliminate instantaneous noise fluctuations. The length of the moving average window is determined according to the sampling rate of the node signal and the characteristics of the target frequency band.

[0166] The smoothed delay change curve is mapped to the phase lag evolution quantity and converted into phase units.

[0167] The obtained continuous time-step phase lag evolution quantities are integrated into a phase lag evolution vector, and time indexes and node identifiers are added to achieve temporal feature extraction of mechanical coupling state drift. Through a chain-like processing method based on cross-correlation peak offset detection and phase transformation, the edge weight difference matrix of the previous step is transformed into a phase lag evolution vector characterizing mechanical coupling state drift, thereby achieving accurate quantification of the system's energy transfer delay change.

[0168] S5.5: Perform a multi-dimensional weight fusion encapsulation operation on the edge weight difference matrix and the phase lag evolution vector, and use tensor splicing technology to integrate the numerical change features and temporal phase features into a high-dimensional state descriptor to output the local topological change vector that finally represents the evolution trend of the energy transfer path between the feed section node and the middle support node.

[0169] Multidimensional feature alignment is performed on the edge weight difference matrix and phase lag evolution vector generated by the previous steps. A dimension normalization strategy is adopted to ensure the consistency of different types of features in terms of numerical scale, so as to eliminate the influence of the difference in dimensions on the fusion accuracy.

[0170] Tensor shape mapping is performed on the two types of features after normalization. The index of the predefined node sequence is used to match the index positions of each row and column of the difference matrix with the index positions of the phase lag vector, forming a matching dimension that can be directly spliced.

[0171] Multidimensional weight calculation is performed based on the matching dimension. Physically meaningful weighting coefficients are used to adjust the proportions of the two types of features. The weighting coefficients are derived from the importance assessment value of the signal type corresponding to the current working condition perception anchor point label.

[0172] The adjusted difference matrix and evolution vector are structurally merged using tensor splicing technology. During the splicing process, the index metadata tags of each dimension are retained to ensure the traceability of node positions in subsequent matching operations.

[0173] The high-dimensional state descriptor after tensor splicing is subjected to structured encapsulation processing. The descriptor is bound to the energy transfer path identifier between the corresponding feed section node and the central support node, and time window attribute information is attached to serve as the final output of the local topology change vector.

[0174] Through the above-mentioned multi-dimensional weight fusion and structured encapsulation processing, the difference and phase lag features of the previous step are transformed into local topological change vectors that are physically semantically complete, temporally continuous, and can be used for micro-anomaly evolution matching, thereby achieving accurate characterization of the energy transfer evolution trend of the feed section and the central support node.

[0175] For example, when the paper bag raw material screw conveyor operates in a high-humidity, low-speed, viscous state, the two-dimensional size of the edge weight difference matrix is ​​3×3, with the numerical range normalized to 0~1. The phase lag evolution vector length is 3, and its normalized range is also 0~1. The operational condition sensing anchor point label assesses the importance coefficient of the current vibration waveform to the current time series as 0.6, and the importance coefficient of the current time series as 0.4. This is applied to weight adjustment to obtain a new difference matrix and evolution vector. The difference matrix and evolution vector are matched according to the node sequence index, for example, the feed section node, the middle support node, and the discharge valve body node are arranged in sequence. After tensor splicing, a high-dimensional state descriptor with a size of 3×6 is formed, where the first 3 columns are the weighted difference matrix values, and the last 3 columns are the weighted evolution vector values. The descriptor binds the energy transfer path identifier "feed section → middle support" and attaches a time window attribute such as "t0~t10". In the symbol distance matching of the micro-anomaly evolution primitive library, it can significantly improve the matching stability of bearing micro-stripping degradation mode. The verification results show that the local topology change vector generated under the high humidity and low speed conditions can stably characterize the trend of simultaneous occurrence of transmission efficiency decline and phase drift.

[0176] Step S6: The local topology change vector is input into the micro-anomaly evolution primitive library for symbolic distance matching calculation. The micro-anomaly evolution primitive library stores local subgraph evolution trajectory benchmarks constructed based on typical early degradation modes. Early fault feature triplets, composed of fault trend type, starting position, and evolution rate, are output through comparison. Specifically, this includes: The micro-anomaly evolution primitive library is a standardized fault mode library generated offline using historical fault data, accelerated life tests, or high-fidelity simulations. Its core function is to provide "evolutionary trajectory templates" for various typical early degradation processes, allowing for similarity matching of real-time subgraph change vectors. The primitive library is constructed and stored as follows: (1) Data source The data for building the primitive library comes from the following sources: Historical failure case studies: Early failure data recorded during the actual operation of the screw conveyor, including a complete subgraph evolution sequence from normal state to failure occurrence; Accelerated life testing: Simulates the degradation process of different failure modes (such as bearing wear, screw bending, material adhesion, etc.) under controlled environment and collects local subgraph structure change vectors; High-fidelity simulation: Based on the finite element or discrete element method, the degradation simulation of key mechanical functional units is performed to generate virtual evolution trajectories.

[0177] (2) The structure of the primitive library Each primitive (i.e., an "evolutionary trajectory baseline") contains the following information: Fault type labels: such as "feed section screw wear", "pitting corrosion of the middle support bearing", "discharge valve body adhesion and blockage"; Starting position identifier: The mechanical functional unit node where the fault initially occurred (such as the screw in the N1 feed section); Evolutionary trajectory sequence: an ordered sequence of vectors representing structural changes in a local subgraph {ΔG1, ΔG2, ..., ΔG}. m}, each ΔG t This represents the offset of the local topology (such as the change in edge weights between feed section nodes and middle nodes, node attribute drift, etc.) relative to the normal state at time step t. Evolution rate range: typical evolution rate range (e.g., edge weight decay rate 0.01~0.05 / hour). Matching threshold: The lower limit of similarity used to determine whether a real-time changing vector belongs to the primitive.

[0178] (3) Symbolic distance matching mechanism In step S6, the local topological change vector ΔG obtained by real-time sliding window subgraph sampling is... real Symbolic distance is calculated with each evolutionary trajectory benchmark in the primitive library.

[0179] Matching result output: Fault Trend Type: The fault type corresponding to the primitive with the highest matching degree; Starting position: The position of the starting node defined in the matching primitive; Evolution rate: via ΔGreal The amplitude change rate is obtained by weighting the typical rate range in the primitive.

[0180] (4) Incremental correction mechanism As described in step S8, once the system obtains actual fault feedback data, the primitive library supports incremental correction: Cluster the new evolutionary trajectories with the existing primitives, and if the similarity exceeds the threshold, merge them and update the evolution rate range; If it is a new fault mode, a new primitive entry is created and added to the library.

[0181] Example In the screw conveyor for paper bag raw materials, the micro-anomaly evolution primitive library stores a primitive called "feed section screw wear," whose evolution trajectory sequence is: {ΔG1: edge weight -0.02, ΔG2: -0.05, ΔG3: -0.09, ΔG4: -0.14}, with an evolution rate of 0.03 to 0.05 / hour. When the subgraph change vector detected in real time is [-0.03, -0.06, -0.10], the symbol distance calculation shows that the distance with this primitive in the primitive library is 0.12 (below the matching threshold of 0.2), indicating a successful match. The fault type "feed section screw wear" is output, the starting position is "N1," and the evolution rate is 0.04 / hour.

[0182] By using the pre-built micro-anomaly evolution primitive library, this scheme can quickly identify early fault trends without relying on a large amount of online training data, and has the ability to adapt based on actual feedback, which significantly improves the accuracy of fault pre-diagnosis and its engineering practicality.

[0183] S6.1: Perform dimension alignment and standardization preprocessing on the local topological change vector, and use the vector space mapping method to project the high-dimensional state descriptor to the same feature subspace as the baseline trajectory of the micro-anomaly evolution primitive library, so as to generate a set of standardized evolution feature vectors with comparability.

[0184] The local topological change vector is subjected to input analysis processing to identify its numerical change feature components and temporal phase feature components, and a feature naming mapping relationship is established with the baseline trajectory of the micro-anomaly evolution primitive library.

[0185] The feature scale normalization operation is performed on the numerical change feature components, and the maximum and minimum value scaling method is used to linearly transform each numerical change component to the interval [0,1], so as to eliminate the differences in different feature dimensions and ensure the consistency of vector space mapping.

[0186] The temporal phase characteristic components are subjected to centralization and variance standardization, transforming the original phase lag values ​​into standardized values ​​with a mean of zero and a variance of one. For the original value of the feature, This is the mean of the feature within the observation window. This represents the corresponding standard deviation.

[0187] The normalized numerical variation components and the standardized temporal phase components are concatenated into a high-dimensional state description matrix according to the feature naming mapping order, and a predefined vector space mapping matrix is ​​used. Perform matrix multiplication to project the high-dimensional state description matrix onto a feature subspace consistent with the baseline trajectory of the primitive library.

[0188] The projected vector set is subjected to feature position alignment processing. The positions of the vector components are rearranged according to the node order of the primitive library trajectory and the feature index table to ensure that the relative positions of the corresponding features in the vector space are completely consistent and comparable.

[0189] Through the above-mentioned normalization, standardization, spatial projection and position alignment processing methods, the local topological structure change vector of the previous step is transformed into a standardized evolutionary feature vector set consistent with the trajectory of the primitive library, so as to achieve scale consistency and structural alignment of feature matching across working conditions.

[0190] S6.2: Based on the standardized evolutionary feature vector set, perform a multi-mode traversal retrieval operation, and use a dynamic time warping strategy to compare the real-time evolutionary trajectory with the typical local subgraph evolutionary trajectory benchmarks such as bearing micro-stripping or helical blade wear stored in the micro-anomaly evolutionary primitive library in order to screen out the candidate degradation mode set with the highest topological similarity.

[0191] When performing a multi-mode traversal retrieval operation on the standardized evolutionary feature vector set, this feature vector set is used as the input object for the retrieval query, and an index relationship is established with each benchmark trajectory in the micro-anomaly evolutionary primitive library. Based on this index relationship, a traversal scheduler is configured to control the loading process of candidate benchmark trajectories one by one, ensuring that the local subgraph evolution trajectories corresponding to different fault mechanisms are all included in the comparison range. For each loaded benchmark trajectory, dynamic time warping (DTW) mapping preprocessing is performed to align the standardized evolutionary feature vector set with the time step sequence of the benchmark trajectory on the time axis, eliminating morphological misalignment caused by time scale differences. During the alignment process, a cumulative distance matrix is ​​constructed, and an optimal path search is performed on this matrix to obtain the path with the minimum cumulative distance between the two sequences. The elements of the cumulative distance matrix are calculated using the following formula: in, Indicates the first The real-time feature vector and the first Euclidean distance between the feature vectors of the baseline trajectory These are the element values ​​of the cumulative distance matrix. After path alignment based on the minimum cumulative distance, the temporal morphological differences in local topological changes are extracted, and these differences are compared with a preset similarity threshold. The baseline trajectory that meets the threshold condition and has the minimum cumulative distance is marked as a candidate degradation mode, ultimately forming a set of candidate degradation modes with the highest topological similarity. By combining a dynamic temporal warping strategy with traversal retrieval, the standardized evolutionary feature vector set from the previous step is transformed into a multi-mode morphological comparison result, achieving the effect of candidate mode selection technology based on temporal structural similarity.

[0192] For example, in a paper bag raw material screw conveyor system operating under high humidity, low speed, and viscous conditions, the standardized evolutionary feature vector set is set to 80 time steps in length. The micro-anomaly evolutionary primitive library contains 15 typical baseline trajectories, each with a length of 100 time steps. This feature vector set is input into the traversal scheduler, and the baseline trajectories are loaded one by one. In the DTW comparison, the Euclidean distance calculation vector dimension is set to 12. When initializing the cumulative distance matrix, the first element... equal The matrix elements are updated sequentially according to the formula. The optimal path search employs a three-way adjacency strategy to ensure a smooth path covering all time steps. After comparison, among the 15 baseline trajectories, 3 trajectories have a cumulative distance less than the set threshold of 30, and one of these trajectories has a cumulative distance of 18, the minimum. This trajectory corresponds to the bearing micro-stripping pattern, is marked as a candidate degradation pattern, and added to the output set. Verification shows that this pattern significantly improves the agreement with the actual equipment operating state in terms of topological change trends, and the output set can be used for subsequent symbolic distance measurements to quickly pinpoint the fault trend type and starting location.

[0193] S6.3: Perform symbolic distance metric calculation on the candidate degradation mode set to quantify the degree of dispersion of the standardized evolution feature vector and the baseline trajectory of each candidate degradation mode on the node attribute drift and edge weight evolution path, so as to generate a numerical symbolic distance scalar sequence representing the abnormal deviation.

[0194] For each local subgraph evolution trajectory baseline in the candidate degradation mode set, the standardized evolutionary feature vector set generated in the previous step is loaded as the comparison object.

[0195] A dimension-wise symbolic mapping is performed on the node attribute drift paths of the standardized evolutionary feature vector and the corresponding candidate degradation mode trajectory benchmark, converting the continuous numerical sequence into a finite letter sequence based on a symbol dictionary, so as to carry out edit distance calculation.

[0196] An index mapping matrix is ​​established for the symbolized node attribute drift sequence and edge weight evolution sequence to ensure the consistency of character position with time step and eliminate non-aligned interference terms.

[0197] The edit distance operation is performed on the two sequences above. The cost of character replacement, insertion, and deletion operations is uniformly set as a constant, and the shortest operation path length is calculated using recursive dynamic programming.

[0198] Edit distance values ​​are calculated for the node attribute drift sequence and the edge weight evolution sequence respectively, and the two types of distances are fused by weighted synthesis.

[0199] The weighted edit distance value corresponding to each candidate degradation mode is recorded as a numerical symbolic distance scalar, forming a sequence of symbolic distance scalars for subsequent optimal matching retrieval.

[0200] By using the weighted editing distance metric processing method described above, the standardized evolution feature vector from the previous step is transformed into a quantifiable anomaly deviation index, thereby improving the accuracy of degradation pattern recognition under cross-operating conditions.

[0201] For example, under the operating condition of a screw conveyor, the standardized evolutionary feature vector is set to 50 dimensions, the symbol dictionary includes four state labels: A, B, C, and D, the node attribute drift sequence is a symbol sequence of length 50, and the edge weight evolution sequence is also of length 50. The replacement cost is set to 1, and the insertion and deletion costs are both 1. The calculated edit distance of the node attribute sequence is 12, and the edit distance of the edge weight sequence is 8. The weight coefficients are set to 0.6 and 0.4, and the weighted fusion yields an overall symbol distance value of 10.4. The above indicators are calculated for the three trajectory benchmarks in the candidate pattern set, resulting in a symbol distance scalar sequence of {10.4, 15.2, 9.6}. This sequence will then be used as the input for the subsequent minimum value selection step, enabling the identification of candidate patterns that closely approximate the actual degenerate trajectory, thereby significantly improving the accuracy and robustness of fault trend matching.

[0202] S6.4: Based on the numerical symbol distance scalar sequence, perform minimum value optimization and parameter inversion processing, lock the local subgraph evolution trajectory benchmark of the optimal matching according to the principle of minimum symbol distance, and extract the fault mechanism label, spatial node index and temporal gradient slope corresponding to the benchmark to construct an early fault feature triplet containing fault trend type, starting position and evolution rate.

[0203] Minimum value optimization is performed based on the calculated numerical symbolic distance scalar sequence. All symbolic distance scalars in the sequence are arranged in ascending order to generate an ordered index table, establishing a mapping relationship between symbolic distance and candidate degradation mode baseline trajectories. An optimal matching locking operation is performed on the first element of the ordered index table to extract the unique storage index code and structural definition information of its corresponding local subgraph evolution trajectory baseline. Fault mechanism label parsing is performed on this optimally matched baseline trajectory. The label mapping submodule in the primitive library is called to read the fault mechanism identifier bound to the trajectory baseline to obtain interpretable fault type data fields. Based on the trajectory baseline structural definition, the spatial topology is parsed to locate the spatial node index of the mechanical functional unit node corresponding to the fault mode in the global graph, and the index information is appended to the current feature extraction result. Gradient calculation is performed on the time dimension features of the baseline trajectory, and the evolution rate value is calculated on the edge weight evolution path of the trajectory using the time step slope estimation method. The extracted fault mechanism labels, spatial node indices, and evolution rate values ​​are sequentially integrated to construct an early fault feature triplet containing fault trend type, starting position, and evolution rate. Through the above processing method, the local topological structure change vector of the previous step is transformed into a standardized triple with clear fault type, spatial location and evolution rate indicators, so as to realize the early fault feature extraction based on the optimal matching trajectory.

[0204] S6.5: Perform confidence encapsulation and structured output operations on the early fault feature triplet, evaluate the reliability of the matching results by combining the distribution variance of the symbolic distance scalar sequence, and attach metadata identifiers to generate a standardized early fault feature data object that can be directly used for threshold judgment logic.

[0205] Step S7: Determine whether the evolution rate in the early fault feature triplet exceeds a preset critical threshold. If the determination result is yes, generate an early fault warning instruction containing the fault type and prediction time; otherwise, generate a system normal operation status indicator. Specifically, this includes: S7.1: Perform structured parsing on the early fault feature triplet, and use the data unpacking method to extract the evolution rate value that represents the speed of mechanical degradation, so as to obtain the real-time evolution rate scalar for threshold determination.

[0206] S7.2: Perform a dynamic benchmark loading operation based on the real-time evolution rate scalar, and retrieve the corresponding preset critical threshold from the pre-stored working condition-threshold mapping table according to the currently matched working condition perception anchor point label, so as to generate a dynamic judgment benchmark value that adapts to the current operating mode.

[0207] When performing dynamic benchmark loading based on the real-time evolution rate scalar, the real-time evolution rate scalar output in step S7.1 is used as the input object, and the condition-aware anchor tag object generated in the preceding step S3 is retrieved as the index basis. The condition-aware anchor tag object undergoes string parsing processing, translating its unique identifier into a queryable key-value field for index retrieval from the pre-stored condition-threshold mapping table. Key-value matching calculations are performed on the condition-threshold mapping table to locate the mapping record that is completely consistent with the key-value field of the current condition-aware anchor tag, and the original value of the preset critical threshold stored in that record is extracted. The original value of the preset critical threshold undergoes unit standardization processing, converting it to the same unit of measurement and numerical resolution as the real-time evolution rate scalar to eliminate dimensional differences during cross-domain data comparison. Dynamic adaptation calculations are performed on the unit-standardized critical threshold, utilizing the instantaneous load fluctuation coefficient in the condition operating parameters. and the coefficient of variation of screw speed The linear correction calculation is performed using the following formula: in To dynamically determine the benchmark value, The original value of the critical threshold is used. The dynamic judgment benchmark value is encapsulated into a standardized numerical object and bound to the current timestamp and operating condition tag metadata to achieve a perfect match between the scale and time sequence of the real-time evolution rate scalar.

[0208] By parsing the anchor tags of the working condition perception, retrieving the working condition-threshold mapping table, and performing dimensional unification and coefficient correction operations, the comparison benchmark corresponding to the real-time evolution rate scalar of the previous step is dynamically generated, realizing the adaptive adjustment of the evolution rate judgment standard under the variable operating mode.

[0209] S7.3: Perform a size comparison operation on the real-time evolution rate scalar and the dynamic judgment benchmark value, and use numerical comparison logic to determine whether the real-time evolution rate scalar is strictly greater than the dynamic judgment benchmark value, so as to generate a Boolean judgment signal containing a true or false result.

[0210] S7.4: Based on the Boolean decision signal, perform branch logic routing processing. If the Boolean decision signal is true, call the early warning generation module to encapsulate the fault trend type and prediction time into an early fault warning instruction. If the Boolean decision signal is false, generate a system normal operation status identifier to output the final status decision object.

[0211] Based on the Boolean judgment signal input, it is loaded into the initial state register unit of the logic routing processor as the trigger condition for the early warning decision branch control. For the true value judgment signal, the execution module calls the instruction generation function. In the decision processing chain, it sequentially accesses the fault type index library and prediction time calculator of the early warning generation module. It uses the fault trend type identifier and evolution rate value output from the previous steps to generate a prediction time scalar, and then performs data structure binding to form the core field set of the early fault early warning instruction. For the false value judgment signal, the execution module generates a state identifier construction instruction. In the decision processing chain, it accesses the running state identifier template library and calls the standard state generator, binding the current device running state to normal mode, and simultaneously attaching a default health state code and a zero-warning time scalar to generate a system running normal state identifier. The above two types of instructions or identifiers are instantiated into data objects. Through a structured encapsulator, the field set or state identifier is inserted into the unified output interface of the state decision object, ensuring that the object meets the data input compatibility specifications of the subsequent format encapsulation module. By using branch logic routing, the judgment result of the previous step is converted into a data object with fault warning or normal state characteristics, so as to realize accurate differentiation of the running state based on the evolution rate threshold and preparation for downstream module calls.

[0212] S7.5: Perform standardized format encapsulation operation on the early fault warning command or system normal operation status identifier, add unified timestamp metadata and device identity identifier to generate a standard status report data packet that can be directly transmitted to the remote management terminal.

[0213] Step S8: Update the device health record of the remote management terminal based on the early fault warning command or the system normal operation status indicator, and incrementally correct the local subgraph evolution trajectory benchmark in the micro-anomaly evolution primitive library according to the newly generated actual fault feedback data to complete the model adaptive optimization. Specifically, this includes: S8.1: Perform protocol encapsulation and transmission processing on early fault warning instructions or system normal operation status indicators, and use the Internet of Things communication protocol to send standard status report data packets containing fault type, prediction time and device identification identifier to the remote management terminal to generate a device status update request sequence to be written.

[0214] S8.2: Based on the device status update request sequence, perform historical data association retrieval operation, use the time axis alignment method to locate the corresponding device's past operation record in the device health file stored in the remote management terminal, and append the fault trend type and evolution rate value in the current standard status report data packet to the end of the file to generate an updated version of the device health file containing the full life cycle status trajectory.

[0215] S8.3: Perform structured parsing processing on the newly generated actual fault feedback data confirmed by on-site maintenance personnel, use data cleaning rules to extract the actual fault occurrence time, specific fault mechanism label and measured degradation curve data, and perform spatiotemporal matching between the above measured degradation curve data and the predicted data in the previously generated updated equipment health record to construct an actual fault feature verification sample set for model correction.

[0216] When performing structured parsing on newly generated actual fault feedback data confirmed by on-site maintenance personnel, the updated equipment health record is used as the reference data source, and the original fault record file submitted by the maintenance end is used as the parsing object. For the various raw fields contained in the fault record file, the timestamp parsing rules in the data cleaning rule set are invoked to uniformly convert the fault occurrence time represented in different formats into a standard timestamp under the system time base, and records with missing values ​​or abnormal formats are removed. Based on the fault mechanism label dictionary defined in the equipment health record, the natural language description or coded fields in the feedback data are matched to unique fault mechanism label identifiers through label mapping rules, achieving standardized binding of mechanism labels. For the original sensor measurement sequences attached to the feedback data, after processing with noise filtering and amplitude normalization methods, the feature extraction module obtains a degradation state sequence that can be used for curve modeling, and time series sorting is performed to form continuous measured degradation curve data. The measured degradation curve data and the predicted degradation curve data in the updated equipment health record are bidirectionally matched on the time axis and spatial node index. By calculating the synchronization offset and node correspondence between the two curves, a spatiotemporal matching result set with an index chain is generated. This spatiotemporal matching result set is encapsulated into a set of actual fault feature verification samples containing the fault occurrence time, specific fault mechanism labels, and curve correspondences, which serves as input data for subsequent deviation quantification calculation steps. Through structured parsing and bidirectional spatiotemporal matching processing, the results of the previous step are transformed into standardized verification samples that can be directly used for model correction, realizing a complete technical closed loop from fault feedback data to model optimization input.

[0217] For example, during the operation of a screw conveyor for paper bag raw materials, maintenance personnel confirmed that a certain piece of equipment experienced an abnormality in its central support bearing within 48 hours of a system warning. The feedback data log file was in CSV format, containing the following fields: fault time (format: "2024 / 05 / 18 14:26:05"), fault description (text: "abnormal vibration in the central support"), sensor ID, and the corresponding original vibration waveform sequence (sampling frequency 500Hz, number of sampling points 6000). During the parsing process, the timestamp parsing rules converted the fault time into a millisecond-level timestamp of 1631931965000 under a unified system time base, and the mechanism label mapping rules matched the fault description to the label "bearing micro-peeling". Noise filtering (5th-order low-pass Butterworth filtering, cutoff frequency 150Hz) and amplitude normalization (maximum value scaled to 1.0) were performed on the original vibration waveform sequence, and then a time point index was generated according to the sampling frequency, resulting in a measured degradation curve data vector with a length of 6000. Time axis alignment was performed on the predicted degradation curve data (length 6000, consistent sampling frequency). Dynamic time warping was used to establish the correspondence matrix between the two curves, where the time synchronization offset was... At 12 o'clock, the spatial node index alignment error was 0.02. The fault occurrence time, the label "bearing micro-stripping", and the curve correspondence matrix were encapsulated into a validation sample set and stored in the validation data buffer for subsequent trajectory deviation vector generation, thereby significantly improving the model's adaptive adjustment capability for central support bearing-type faults.

[0218] S8.4: Based on the actual fault characteristics verification sample set, perform deviation quantization calculation operation, and use dynamic time warping strategy to compare the difference between the measured degradation curve data and the local subgraph evolution trajectory benchmark stored in the micro-anomaly evolution primitive library on the node attribute drift path, so as to generate a trajectory deviation vector sequence that characterizes the model prediction error.

[0219] S8.5: Perform weighted fusion correction processing on the trajectory deviation vector sequence, use gradient descent to optimize and adjust the shape parameters and evolution slope of the local subgraph evolution trajectory benchmark corresponding to the fault mechanism label in the micro-anomaly evolution primitive library, and re-serialize and store the corrected benchmark in the primitive library to complete the model adaptive optimization closed loop with working condition adaptability.

[0220] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0221] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0222] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for IoT-based metering and fault pre-diagnosis of paper bag raw material spiral conveying, specifically comprising: S1: Acquire multi-source raw signal streams from key mechanical points of the screw conveyor and generate time-synchronized multi-dimensional data sequences; S2: Based on time-synchronized multidimensional data sequences, the waveform data of each channel in each sampling period is truncated into standardized data segments and mapped to the initial attribute vectors of graph nodes to generate node attribute sets; S3: Based on the physical property parameters of the paper bag raw materials and the equipment operating status parameters, the current working condition is symbolically analyzed to generate working condition perception anchor point tags; S4: Match the nearest graph structure template based on the anchor label of the working condition perception, and dynamically inject the node attribute set into the corresponding mechanical functional unit node in the graph structure template. At the same time, calculate the edge weight based on the cross-correlation coefficient between adjacent nodes and the matching confidence of the anchor label of the working condition perception, and construct a dynamic graph with physical semantic constraints. S5: Perform a sliding window subgraph sampling operation on the dynamic graph to extract the local topological change vector between consecutive time steps. The local topological change vector represents the evolution trend of the energy transfer path between the feed section node and the middle support node. S6: Input the local topological structure change vector into the micro-anomaly evolution primitive library, and output the early fault feature triplet composed of fault trend type, starting position and evolution rate by comparison; S7: Determine whether the evolution rate in the early fault feature triplet exceeds the critical threshold. If the determination result is that it exceeds the threshold, generate an early fault warning instruction containing the fault type and prediction time; otherwise, generate a system normal operation status indicator.

2. The IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials according to claim 1, characterized in that, The multi-source raw signal stream includes waveform data acquired by a low-power vibration accelerometer, current timing data recorded by a motor current sampling module, and spectral reflectance data detected by an optical thickness sensor at the feed inlet.

3. The IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials according to claim 1, characterized in that, The operating condition sensing anchor tag represents a specific operating mode, such as a high-humidity, low-speed viscous state or a dry powder, high-speed segregation state.

4. The IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials according to claim 1, characterized in that, The graph structure template includes mechanical functional unit node definitions, directed edge connection relationships between nodes, and edge physical semantic tags, which are used to match the working condition perception anchor point tags and serve as the structural skeleton for node attribute injection.

5. The IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials according to claim 1, characterized in that, The micro-anomaly evolution primitive library is a standardized failure mode library generated from historical failure data, accelerated life tests, or high-fidelity simulations. The micro-anomaly evolution primitive library stores local subgraph evolution trajectory benchmarks constructed based on typical early degradation modes.

6. The IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials according to claim 1, characterized in that, Following S7, the following also includes: S8: Update the device health record of the remote management terminal based on the early fault warning command or the normal operation status indicator of the system, and perform incremental correction on the local subgraph evolution trajectory benchmark in the micro-anomaly evolution primitive library according to the newly generated actual fault feedback data to complete the model adaptive optimization.

7. The IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials according to claim 1, characterized in that, Step S3 specifically includes: Multi-source parameter extraction processing is performed on the spectral reflectance data detected by the optical thickness sensor at the feed inlet of the screw conveyor and the current time series data recorded by the motor current sampling module to obtain a set of original working condition parameters, including the nominal range of raw material moisture content, fiber length distribution characteristics, instantaneous screw speed, fixed conveying angle, and instantaneous load fluctuation. Based on the original set of operating parameters, a multidimensional threshold boundary mapping operation is performed. The operating condition classification rule base is used to divide the continuously changing nominal range of raw material moisture content and instantaneous screw speed into multiple discretized state ranges such as high moisture, medium moisture, dry powder, low speed, medium speed, and high speed, so as to generate a single-dimensional state classification vector with clear physical boundaries. Logical combination encoding is performed on the single-dimensional state classification vector. Boolean algebra and state machine transformation processing methods are used to couple and associate the high humidity state interval and the low speed state interval. The interaction relationship between fiber length distribution characteristic value and instantaneous load fluctuation value is combined to construct a composite working condition logical expression that characterizes complex coupling effects such as high humidity low speed viscous state or dry powder high speed segregation state. Based on the compound working condition logic expression, symbolic label generation processing is performed to convert each specific compound working condition logic expression into a unique string identifier and output a standardized working condition perception anchor point label sequence. The confidence verification and formatting encapsulation operations are performed on the condition sensing anchor point tag sequence. The reliability of tag generation is evaluated based on the signal-to-noise ratio of each sensor data at the current time, and timestamp metadata is added. Finally, an interpretable condition sensing anchor point tag object that can be directly used for graph structure template matching is generated.

8. The IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials according to claim 1, characterized in that, Step S4 specifically includes: The template index retrieval process is performed on the working condition sensing anchor point label object. The string identifiers such as high humidity low speed viscous state or dry powder high speed segregation state are matched to the nearest basic topology frame in the pre-stored graph structure template library using the preset working condition-topology mapping dictionary, so as to generate candidate graph structure template instances containing mechanical functional unit node definitions and initial connection relationships. Based on the candidate graph structure template instance, a node attribute dynamic injection operation is performed. The initial attribute vector set of graph nodes generated in the previous steps is filled into the corresponding feed section screw, middle support bearing and discharge valve body node positions in the candidate graph structure template instance according to the mechanical functional unit identifier binding relationship, so as to generate a full attribute graph structure intermediate state with real-time state characteristics. Generate the original cross-correlation matrix based on the intermediate states of the full-attribute graph structure; Based on the confidence verification results of the original cross-correlation matrix and the working condition sensing anchor tag object, a weighted fusion calculation is performed. The linear weighted fusion strategy is used to couple the values ​​in the original cross-correlation matrix with the signal-to-noise ratio evaluation values ​​attached to the working condition sensing anchor tag object to generate a set of composite edge weight values ​​that have both signal physical correlation and working condition matching reliability. The intermediate state of the full-attribute graph structure is processed by edge weight attribute update and structured encapsulation. The set of composite edge weight values ​​is assigned to the directed edge attribute field connecting each mechanical functional unit node in the intermediate state of the full-attribute graph structure, and physical semantic constraint metadata tags are added to finally output a dynamic graph with physical semantic constraints that can be directly used for subgraph sampling analysis.

9. The IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials according to claim 1, characterized in that, The process of generating the original cross-correlation matrix based on the intermediate state of the full-attribute graph structure specifically involves performing cross-correlation analysis on the standardized data segments between adjacent mechanical functional unit nodes in the intermediate state of the full-attribute graph structure, and using the sliding window cross-correlation processing method to calculate the phase lag and amplitude correlation between vibration waveform data and current time series data in the time-synchronized multidimensional data sequence, so as to generate the original cross-correlation matrix characterizing energy transfer efficiency.

10. The IoT metering and fault pre-diagnosis method for spiral conveying of paper bag raw materials according to claim 1, characterized in that, Step S5 specifically includes: Time series indexing is constructed for dynamic graphs with physical semantic constraints. Graph instances are arranged in the order of collection based on a unified timestamp to generate a time series graph queue, thereby establishing a continuous time reference axis for sliding window sampling. Based on the time series graph queue, a fixed-span sliding window truncation operation is performed. Using a preset time step parameter, a set of subgraph sequences containing the current time and several previous times is selected in the time series graph queue to generate a local subgraph observation window with temporal continuity. Perform edge weight difference operation between nodes on adjacent time step subgraphs within the local subgraph observation window. Use the point-by-point subtraction method to calculate the change in weight value of the connecting edge between the feed section node and the middle support node in continuous time steps, so as to generate an edge weight difference matrix that characterizes the instantaneous fluctuation of energy transfer efficiency. Based on the edge weight difference matrix, topological phase lag feature extraction is performed. The cross-correlation peak offset detection method is used to analyze the change trajectory of signal propagation delay between corresponding nodes of vibration waveform data and current time series data, so as to generate a phase lag evolution vector that reflects the mechanical coupling state drift. The edge weight difference matrix and the phase lag evolution vector are subjected to multi-dimensional weight fusion and encapsulation operation. Tensor splicing technology is used to integrate the numerical change features and temporal phase features into a high-dimensional state descriptor, and the final output is a local topological change vector that represents the evolution trend of the energy transfer path between the feed section node and the middle support node.

Images