Method and system for failure early warning of tobacco production equipment

Abstract

The application discloses a fault early warning method and system of a tobacco production equipment, and relates to the technical field of data processing.The method adopts a technical means combining graph attention heterogeneous fusion and causal convolution time series modeling, realizes automatic determination from multi-source operation data to hierarchical fault early warning, and overcomes the technical problems that a traditional monitoring system only alarms without diagnosis, lacks deep feature mining and intelligent early warning grading.The method specifically comprises the following steps: collecting and regularizing and denoising multi-source time series operation data flow to construct a standardized data base; completing cross-dimension feature association and deep fusion of multi-source data through a graph attention heterogeneous fusion network; capturing time series dependence of the equipment operation state through a causal convolution and an expansion convolution structure; enhancing feature expression capability through layer-by-layer deepening feature mapping by layer residual deduction; condensing core features of the equipment state through global pooling and dimension compression; and outputting real-time fault early warning grade data through comparison of a multilayer perception machine classifier and a hierarchical threshold.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a fault early warning method and system for tobacco production equipment. Background Technology

[0002] In modern industrial production settings such as tobacco processing workshops, the level of equipment automation is extremely high. Currently, workshops typically use SCADA systems (such as Siemens WinCC) to monitor the production line. When equipment malfunctions, the system displays alarm signals on the HMI interface (usually an indicator light turning red or an alarm entry popping up). Maintenance personnel need to locate the cause of the malfunction and find solutions based on the alarm code or symptoms, relying on their personal experience or consulting paper / electronic maintenance manuals.

[0003] However, with the increasing complexity of equipment processes, a single production line often involves hundreds or even thousands of critical I / O points (sensors, actuator status). When equipment stops or malfunctions, traditional HMI interfaces can only output alarm codes (Fault Codes) or simple indicator light statuses based on the PLC's Boolean logic. Maintenance personnel often face the following dilemmas when dealing with complex electromechanical-hydraulic coupling faults: Information silos: Alarm messages only indicate the "result" (such as motor overload) without revealing the "cause" (such as bearing wear or voltage fluctuation).

[0004] Experience-dependent: Troubleshooting relies heavily on the engineer's personal experience or consulting thick maintenance manuals, resulting in long average repair times and seriously affecting overall equipment efficiency.

[0005] Traditional monitoring methods suffer from the drawback of "only alarming, not diagnosing," heavily relying on human experience, resulting in long troubleshooting times and impacting production efficiency. Summary of the Invention

[0006] The main purpose of this application is to provide a fault early warning method and system for tobacco production equipment, so as to solve the problem that the traditional monitoring mode in the prior art has the pain point of "only alarming, not diagnosing", which relies heavily on human experience, resulting in long fault diagnosis time and affecting production efficiency.

[0007] To achieve the above objectives, this application provides the following technical solution: A fault early warning method for tobacco production equipment, the fault early warning method comprising: Step S1: Continuously collect multi-source time-series operation data streams during the operation of tobacco production equipment, and perform normalization, noise reduction, and dimension unification processing on the multi-source time-series operation data streams to obtain a standard time-series data stream; Step S2: Input the standard time-series data stream into the graph attention heterogeneous fusion network, and perform cross-dimensional feature association and deep fusion on the standard time-series data stream through a multi-head attention mechanism to obtain the device operation fusion feature set; Step S3: Input the device running fused feature set into the temporal causal convolutional network, and capture the temporal dependency relationship through the causal convolution and dilated convolution structure to obtain the temporal associated feature sequence; Step S4: Perform hierarchical residual deduction on the time-series associated feature sequence, and complete the feature deepening mapping layer by layer through multi-layer residual connections to obtain the deepened deduction feature sequence; Step S5: Perform global pooling and feature dimension compression on the in-depth inference feature sequence, remove redundant feature components and extract effective features to obtain the standard feature vector of device status; Step S6: Input the standard feature vector of the equipment state into the multilayer perceptron classifier, output the probability distribution of the equipment state through fully connected mapping and Softmax normalization operation, and output the real-time fault warning level data of the tobacco production equipment based on the comparison result of the probability distribution of the equipment state and the preset grading threshold.

[0008] Beneficial effects of steps S1 to S6: This method addresses the technical problems of isolated alarm information and heavy reliance on manual experience in fault diagnosis during tobacco production equipment fault early warning. It employs a combination of graph attention heterogeneous fusion and causal convolutional temporal modeling to achieve automated judgment from multi-source operational data to hierarchical fault early warning. This effectively overcomes the technical problems of traditional monitoring systems in the background, which only alarm but do not diagnose, lack deep feature mining and intelligent early warning hierarchical. Specifically, step S1 involves collecting and regularizing multi-source time-series operational data streams to construct a standardized data foundation; step S2 involves using a graph attention heterogeneous fusion network to complete cross-dimensional feature association and deep fusion of multi-source data; step S3 involves capturing the temporal dependencies of equipment operating status through causal convolution and dilated convolution structures; step S4 involves deepening feature mapping layer by layer through hierarchical residual deduction to enhance feature expression capabilities; step S5 involves refining the core features of equipment status through global pooling and dimensional compression; and step S6 involves outputting real-time fault warning level data through multilayer perceptron classifier and hierarchical threshold comparison.

[0009] As a further improvement to this application, step S1 involves continuously collecting multi-source time-series operational data streams during the operation of the tobacco production equipment, and performing normalization, noise reduction, and dimensionality unification processing on the multi-source time-series operational data streams to obtain a standard time-series data stream, including: Step S1.1: The sensor group deployed at the key components of the tobacco production equipment continuously collects operating parameters such as temperature, vibration, current, pressure, and rotation speed to obtain a multi-source raw operating data stream; Step S1.2: Resample and align the multi-source original running data stream at a uniform sampling frequency, and map each source data to a global time-series coordinate based on the timestamp identifier to obtain a time-aligned data stream; Step S1.3: Perform missing value detection and forward filling on the time-aligned data stream, and identify and remove abnormal outliers to obtain a cleaned data stream; Step S1.4: Smooth the purified data stream using a moving average filter to obtain a noise-reduced data stream; Step S1.5: Perform Min-Max normalization operation on the noise-reduced data stream according to each dimension, and map the values ​​of each dimension to a preset interval to obtain a dimensionally unified data stream; Step S1.6: The unified data stream is spliced ​​by tensors according to the device node dimension and the time sequence dimension to obtain the standard time sequence data stream.

[0010] Beneficial effects of steps S1.1 to S1.6: This series of steps, through multi-source operational parameter acquisition and step-by-step data processing, transforms heterogeneous operational data from different sensors and sampling frequencies into a standardized time-series data stream. This effectively solves the problem of feature modeling difficulties caused by temporal misalignment, noise interference, and dimensional differences in multi-source data. It provides a reliable and dimensionally unified input foundation for cross-dimensional feature association of graph attention heterogeneous fusion networks and temporal dependency modeling of temporal causal convolutional networks. Specifically, step S1.1 utilizes a multi-type sensor group to cover key dimensions such as temperature, vibration, current, pressure, and rotational speed, achieving comprehensive perception of the equipment's operating status; step S1.2 eliminates temporal misalignment of multi-source data through unified sampling frequency resampling and timestamp identification mapping to ensure temporal alignment; step S1.3 ensures the integrity and reliability of the data stream during long-term continuous acquisition through forward filling of missing values ​​and anomaly removal using the 3σ criterion; step S1.4 suppresses high-frequency noise interference through moving average filtering, improving the data signal-to-noise ratio to reduce feature extraction errors in subsequent networks; step S1.5 maps data from each dimension to a unified measurement range through Min-Max standardization, eliminating interference from dimensional differences on feature fusion; and step S1.6 processes multi-dimensional data into structured temporal tensors for direct reception by downstream networks through tensor concatenation of device node dimensions and temporal dimensions.

[0011] As a further improvement to this application, in step S2, the standard time-series data stream is input into a graph attention heterogeneous fusion network, and cross-dimensional feature association and deep fusion are performed on the standard time-series data stream through a multi-head attention mechanism to obtain a device operation fusion feature set, including: Step S2.1: Divide the standard time-series data stream into several nodes according to the device component dimension, and use the data segment corresponding to each node as the node feature vector to construct the device topology graph structure; Step S2.2: Based on the physical connection relationship and signal transmission path between the device components corresponding to each node, define the heterogeneous edge type and edge weight between nodes in the device topology graph structure to obtain a weighted heterogeneous graph; Step S2.3: Perform linear mapping on the node feature vector of each node in the weighted heterogeneous graph to generate query vector, key vector, and value vector; Step S2.4: Perform a dot product operation on the query vector and the key vector and normalize it using Softmax to obtain the attention weight matrix between each node; Step S2.5: The attention weight matrix and the value vector are weighted and summed, and the multi-head attention outputs are spliced ​​and linearly projected to obtain node-level fusion features; Step S2.6: Perform a global pooling operation on the node-level fusion features to aggregate the feature information of all nodes and obtain the device operation fusion feature set.

[0012] Beneficial effects of steps S2.1 to S2.6: This series of steps transforms standard temporal data streams into a fusion feature set of device operation through a graph attention heterogeneous fusion network. It models the physical relationships and signal transmission paths between device components using a graph structure, and achieves autonomous association and deep fusion of cross-dimensional features through a multi-head attention mechanism. This effectively solves the problem that traditional feature splicing methods cannot characterize the topological relationships and coupling features between device components, resulting in the loss of deep association information in multi-source operation data. It provides high-quality fusion features containing device topological semantics for subsequent temporal dependency modeling of temporal causal convolutional networks. Specifically, step S2.1 divides the discrete time-series data into a physically meaningful graph representation by dividing nodes according to device component dimensions and constructing a topological graph structure; step S2.2 encodes the physical connection relationships and signal transmission paths between device components into the graph structure to represent the coupling relationships between nodes by defining heterogeneous edge types and edge weights; step S2.3 generates query vectors, key vectors, and value vectors by linearly mapping node feature vectors, providing a learnable feature space projection for attention calculation; step S2.4 achieves autonomous measurement and weight allocation of the association strength between nodes by performing dot product operation on query vectors and key vectors and Softmax normalization; step S2.5 fuses feature information from different subspaces to generate node-level fused features by weighted summation of attention weight matrices and value vectors and multi-head splicing projection; and step S2.6 aggregates the feature information of all nodes through global pooling, transforming the graph-level representation into a device-level fused feature set for downstream network reception.

[0013] As a further improvement to this application, step S3 involves inputting the device's fused feature set into a temporal causal convolutional network, capturing temporal dependencies through causal convolution and dilated convolution structures, and obtaining a temporal associated feature sequence, including: Step S3.1: The device operation fusion feature set is serialized and expanded according to time steps to obtain the fusion feature time sequence; Step S3.2: Apply causal convolution operation to the fused feature time series sequence, wherein each convolution kernel performs convolution operation only on the data of the current time step and the historical time steps to obtain the causal feature sequence; Step S3.3: Apply dilated convolution operation to the causal feature sequence, and expand the sampling interval of the convolution kernel according to a preset dilation factor to expand the temporal receptive field without increasing the number of parameters, so as to obtain the dilated feature sequence. Step S3.4: Perform batch normalization and nonlinear activation operations on the dilated feature sequence sequentially to obtain the activated feature sequence; Step S3.5: The activation feature sequence is used as the input of the next layer of causal dilated convolution for multi-layer stacking iteration to expand the temporal receptive field and deepen the temporal dependency modeling layer by layer, so as to obtain the temporal correlation feature sequence.

[0014] Beneficial effects of steps S3.1 to S3.5: This series of steps uses a temporal causal convolutional network to model the temporal dependencies of the fused feature set of device operation. By utilizing the temporal direction constraints of causal convolution and the receptive field expansion mechanism of dilated convolution, it captures the long-range temporal correlation of device operation status while ensuring temporal causality. This effectively solves the problems of future information leakage and insufficient long-range dependency capture capability in traditional convolutional networks in temporal modeling, and provides a sequence of associated features containing rich temporal dynamic information for subsequent hierarchical residual inference. Specifically, step S3.1 transforms the fused feature set into a temporal sequence form by unfolding it sequentially by time step, providing a standard temporal input for causal convolution operations; step S3.2 constrains the convolution kernel to only perceive current and historical time step data through causal convolution, preventing future information leakage and ensuring the causal rationality of temporal modeling; step S3.3 expands the sampling interval by dilating convolution according to a preset dilation factor, increasing the temporal receptive field to capture longer-range temporal dependencies without increasing the number of parameters; step S3.4 stabilizes the feature distribution of each layer and enhances the nonlinear expressive power of the features through batch normalization and nonlinear activation operations; step S3.5 expands the receptive field layer by layer and deepens temporal dependency modeling through multi-layer stacking iteration, accumulating the output of multi-layer causal dilated convolution into a sequence of associated features containing long-range temporal dynamics.

[0015] As a further improvement of this application, step S4 involves performing hierarchical residual deduction on the temporal correlation feature sequence, completing the feature deepening mapping layer by layer through multi-layer residual connections to obtain the deepened deduction feature sequence, including: Step S4.1: Input the time-series correlation feature sequence into the first-layer inference unit, perform linear mapping through a fully connected layer and nonlinear activation operation to obtain the first-layer inference features; Step S4.2: Add the residuals of the first layer of inference features and the time-series associated feature sequence to obtain the first layer of residual inference features; Step S4.3: Input the first layer residual inference features into the second layer inference unit, perform linear mapping through a fully connected layer and nonlinear activation operation to obtain the second layer inference features; Step S4.4: Add the residuals of the second-layer derivation features and the first-layer residual derivation features to obtain the second-layer residual derivation features; Step S4.5: Input the second-layer residual extrapolation features into the subsequent extrapolation unit, and perform linear mapping, nonlinear activation and residual addition operations layer by layer according to the extrapolation logic of the first-layer extrapolation unit and the second-layer extrapolation unit to obtain the in-depth extrapolation feature sequence.

[0016] Beneficial effects of steps S4.1 to S4.5: This series of steps performs multi-layer feature deepening mapping on temporal correlation feature sequences through hierarchical residual deduction. By utilizing the residual connection mechanism, the original feature information is preserved and high-order abstract features are superimposed in the layer-by-layer transformation. This effectively solves the problem of information loss and feature degradation caused by gradient degradation in deep networks during feature deepening. It provides a deep deduction feature sequence with strong representation ability and high information fidelity for subsequent global pooling and feature dimension compression. Specifically, step S4.1 projects the temporal correlation feature sequence to a high-dimensional feature space through linear mapping and nonlinear activation operations in a fully connected layer to introduce nonlinear expressive power; step S4.2 adds the residuals of the first-layer inference features and the temporal correlation feature sequence to deepen the features while preserving the original temporal correlation information to alleviate gradient degradation; step S4.3 further extracts higher-order abstract features through fully connected mapping and nonlinear activation of the second-layer inference unit to deepen the feature expression level; step S4.4 adds the residuals of the second-layer inference features and the first-layer residual inference features to establish a direct information pathway between layers to ensure the integrity of deep information transmission; and step S4.5 accumulates the multi-level feature transformation results into a richly represented inference feature sequence by iteratively executing linear mapping, nonlinear activation, and residual addition operations layer by layer according to the preceding inference logic.

[0017] As a further improvement to this application, step S5 involves performing global pooling and feature dimension compression on the deepened inference feature sequence, removing redundant feature components and extracting effective features to obtain a standard feature vector of device status, including: Step S5.1: Perform global average pooling operation on the in-depth inference feature sequence along the time dimension to aggregate each feature channel into a single feature vector in the time dimension, and obtain the time-series aggregated feature vector; Step S5.2: Calculate the feature variance of each dimension in the time-series aggregated feature vector, and sort each dimension according to the size of the variance to obtain the feature importance ranking result; Step S5.3: Based on the feature importance ranking result, feature dimensions with variances below a preset threshold are identified as redundant feature components and removed to obtain an effective feature subset. Step S5.4: The effective feature subset is linearly reduced and compressed to a preset target dimension through a fully connected layer to obtain the standard feature vector of the device state.

[0018] Beneficial effects of steps S5.1 to S5.4: This series of steps reduces and condenses the feature sequence of the in-depth inference through global pooling and feature dimension compression. It aggregates the dynamic features in the time dimension into static feature vectors and removes redundant components based on feature importance. This effectively solves the problems of increased computational overhead and interference from classifier input noise caused by the dimensional redundancy of the feature sequence of the in-depth inference. It provides a standard feature vector of device state with simplified dimensions and high information density for the subsequent multilayer perceptron classifier. Specifically, step S5.1 aggregates dynamic temporal features into static feature vectors by performing global average pooling along the temporal dimension to eliminate temporal dimension redundancy; step S5.2 quantifies the information contribution of each feature component by calculating and sorting the variance of each feature dimension to construct a basis for evaluating feature importance; step S5.3 removes low-information features by identifying feature dimensions with variances below a preset threshold as redundant components; and step S5.4 compresses effective features to a preset target dimension through linear dimensionality reduction mapping using a fully connected layer, achieving dimensionality simplification of the feature vector while retaining core state information to adapt to the input requirements of the classifier.

[0019] As a further improvement of this application, step S6 involves inputting the standard feature vector of the equipment state into a multilayer perceptron classifier, outputting the equipment state probability distribution through fully connected mapping and Softmax normalization, and outputting real-time fault warning level data of the tobacco production equipment based on the comparison result between the equipment state probability distribution and a preset grading threshold, including: Step S6.1: Input the standard feature vector of the device state into the multilayer perceptron classifier, and perform feature transformation layer by layer through several fully connected layers and nonlinear activation functions to obtain the category space mapping vector; Step S6.2: Apply the Softmax function to the category space mapping vector to perform normalization operation, convert each component into a probability value, and obtain the device state probability distribution; Step S6.3: Extract the maximum probability value and its corresponding state category from the device state probability distribution to obtain the current device state determination result; Step S6.4: Compare the maximum probability value with a preset first grading threshold and a second grading threshold, and determine the fault warning level based on the comparison result, wherein the first grading threshold is lower than the second grading threshold. Step S6.5: Generate warning instruction data based on the fault warning level and the current equipment status determination result, and output the warning instruction data as the real-time fault warning level data of the tobacco production equipment.

[0020] Beneficial effects of steps S6.1 to S6.5: This series of steps transforms the standard feature vector of equipment status into real-time fault warning level data through a multilayer perceptron classifier and a hierarchical threshold comparison mechanism. It uses fully connected mapping and Softmax normalization to achieve probabilistic output of status and determines the warning level based on dual threshold hierarchies. This effectively solves the problem that traditional monitoring systems only output Boolean alarm signals and lack fault probability measurement and warning classification, and realizes a complete closed loop from feature vector to executable warning instructions. Specifically, step S6.1 maps the state standard feature vector to the category space through layer-by-layer feature transformation using several fully connected layers and nonlinear activation functions to establish a discriminative relationship between features and device state; step S6.2 converts the category space mapping vector into a probability distribution through normalization operation using the Softmax function to realize the probabilistic expression of the device state discrimination result; step S6.3 extracts the maximum probability value and its corresponding state category to obtain the most likely state judgment result of the current device from the probability distribution; step S6.4 compares it with the first and second level thresholds to determine the fault warning level according to the interval of the probability value to realize differentiated classification of warnings; step S6.5 generates warning instruction data based on the warning level and state judgment result, converting the classifier output into real-time fault warning level data that can be directly used for equipment operation and maintenance.

[0021] To achieve the above objectives, this application also provides the following technical solutions: A fault early warning system for tobacco production equipment, wherein the fault early warning system is applied to the fault early warning method described above, and the fault early warning system includes: The standard time-series data stream acquisition module is used to continuously collect multi-source time-series operation data streams during the operation of tobacco production equipment, and to perform normalization, noise reduction and dimension unification processing on the multi-source time-series operation data streams to obtain a standard time-series data stream. The device operation fusion feature set acquisition module is used to input the standard time series data stream into the graph attention heterogeneous fusion network, and perform cross-dimensional feature association and deep fusion on the standard time series data stream through a multi-head attention mechanism to obtain the device operation fusion feature set. The temporal correlation feature sequence acquisition module is used to input the device running fusion feature set into the temporal causal convolutional network, and capture temporal dependencies through causal convolution and dilated convolution structures to obtain the temporal correlation feature sequence; The in-depth inference feature sequence acquisition module is used to perform hierarchical residual inference on the time-series associated feature sequence, and complete the feature in-depth mapping layer by layer through multi-layer residual connection to obtain the in-depth inference feature sequence; The device status standard feature vector acquisition module is used to perform global pooling and feature dimension compression on the in-depth inference feature sequence, remove redundant feature components and extract effective features to obtain the device status standard feature vector. The real-time fault warning level data acquisition module is used to input the standard feature vector of the equipment status into the multilayer perceptron classifier, output the probability distribution of the equipment status through fully connected mapping and Softmax normalization operation, and output the real-time fault warning level data of the tobacco production equipment based on the comparison result of the probability distribution of the equipment status and the preset grading threshold.

[0022] To achieve the above objectives, this application also provides the following technical solutions: An electronic device includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the fault warning method for tobacco production equipment as described above.

[0023] To achieve the above objectives, this application also provides the following technical solutions: A computer-readable storage medium storing program instructions, which, when executed by a processor, enable the implementation of the fault early warning method for tobacco production equipment as described above. Attached Figure Description

[0024] Figure 1 This is a schematic flowchart illustrating the steps of an embodiment of a fault early warning method for tobacco production equipment according to this application; Figure 2 This is a schematic diagram of the functional modules of a fault early warning system for tobacco production equipment according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an embodiment of the electronic device of this application; Figure 4 This is a schematic diagram of the structure of one embodiment of the storage medium of this application. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0026] The terms "first," "second," and "third" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (e.g., as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive having. For example, a process, method, system, product, or device having a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0027] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be present in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0028] It should be noted that, due to the limited types and number of symbols or letters that can represent specific meanings, for embodiments with many formulas or codes, there may be situations where symbols or letters cannot meet the usage requirements. Therefore, the interpretation of formula symbols in the steps or sub-steps of the embodiments is only valid for the current step or sub-step.

[0029] If the same symbol has different interpretations in different steps or sub-steps, the interpretation in the current step or sub-step shall prevail; if the same symbol appears in different steps or sub-steps, but no interpretation is given in subsequent steps or sub-steps after its first appearance, the interpretation in the first step or sub-step shall be used.

[0030] For example, "i=1,2,...,n" is a writing convention and a well-known meaning. If "i=1,2,...,n" has already appeared once in an embodiment and needs to be used again in formulas with different scenarios and meanings in the future, then the meaning of "i=1,2,...,n" should be interpreted according to the corresponding formulas. If "i=1,2,...,n" is not used for the first time and is changed to a symbol or letter with a non-well-known meaning, such as "p=1,2,...,q", to avoid repetition, it is more likely to cause confusion, ambiguity, and unclear problems.

[0031] like Figure 1 As shown, this embodiment provides a fault early warning method for tobacco production equipment. In this embodiment, the fault early warning method includes the following steps: Step S1: Continuously collect multi-source time-series operation data streams during the operation of tobacco production equipment, and perform regularization, noise reduction, and dimension unification processing on the multi-source time-series operation data streams to obtain standard time-series data streams.

[0032] Furthermore, step S1 specifically includes the following steps: Step S1.1: The sensor group deployed at the key components of the tobacco production equipment continuously collects operating parameters such as temperature, vibration, current, pressure, and rotation speed to obtain a multi-source raw operating data stream.

[0033] Preferably, the core of this step is to collect multi-dimensional parameters of the equipment's operating status. This is achieved by deploying sensor groups on key components such as the main motor, bearings, and cutter rollers of the shredder, the cylinder and hot air system of the shredder, and the atomization system of the feeder, thereby enabling comprehensive perception of the equipment's operating parameters.

[0034] For the temperature parameter, a PT100 platinum resistance temperature sensor is deployed on the motor windings, bearing housings, and cylinder surface. The measurement range is -50 to 250 degrees Celsius, the accuracy is ±0.15 degrees Celsius, and the sampling interval is 1 second.

[0035] For vibration parameters, IEPE piezoelectric accelerometers are deployed on the bearing housing and cutter roller support housing, with a measurement range of 0 to 50g, a frequency range of 0.5 to 10000 Hz, and a sampling interval of 0.01 seconds.

[0036] For the current parameter, Hall current sensors are deployed in the power supply circuits of the main motor and auxiliary motor, with a measurement range of 0 to 200 amperes, an accuracy of ±0.5%, and a sampling interval of 0.1 seconds.

[0037] For the pressure parameter, a piezoresistive pressure sensor is deployed in the hydraulic system and pneumatic circuit, with a measurement range of 0 to 10 MPa, an accuracy of ±0.25%, and a sampling interval of 0.1 seconds.

[0038] For the rotational speed parameter, a magnetoelectric rotational speed sensor is deployed on the motor output shaft and the cutter roller spindle. The measurement range is 0 to 6000 revolutions per minute, the accuracy is ±1 revolution per minute, and the sampling interval is 0.05 seconds.

[0039] Preferably, the temperature sampling interval is 1 second; the vibration sampling interval is 0.01 seconds; the current sampling interval is 0.1 seconds; the pressure sampling interval is 0.1 seconds; and the rotational speed sampling interval is 0.05 seconds.

[0040] For example, taking the main motor of the shredder as an example, at time YYYY-MM-DD-HH:MM:SS-001, the sensor group collected the motor winding temperature of 78.2 degrees Celsius, the bearing vibration acceleration of 2.3g, the main motor current of 45.6 amperes, the hydraulic pressure of 4.2 MPa, and the spindle speed of 1480 revolutions per minute. The above data constitutes the multi-source raw operating data stream at that moment.

[0041] Step S1.2: Resample and align the original running data streams from multiple sources at a uniform sampling frequency, and map each source data to a global time-series coordinate based on the timestamp identifier to obtain a time-aligned data stream.

[0042] Preferably, the core of this step is to complete the time sequence alignment of multi-source data. By resampling at a unified sampling frequency and mapping with timestamp identifiers, the time sequence misalignment caused by the difference in sampling frequency among the sensors is eliminated.

[0043] For the resampling operation, the sampling interval of the vibration sensor with the highest frequency of 0.01 seconds is used as the reference frequency. Other parameters are upsampled using linear interpolation. Temperature, current, pressure and speed data are all resampled to a 0.01-second interval to align all data sources in the time dimension.

[0044] Specifically, for the timestamp identifier mapping, a globally incrementing timestamp identifier is assigned to each sampling point, with the identifier format being YYYY-MM-DD-HH:MM:SS-mmm, where mmm is a millisecond-level sequence number. Each source data is then mapped to the global time series coordinate system according to the timestamp identifier.

[0045] Preferably, the resampling reference frequency is 100 Hz (corresponding to a 0.01-second interval); the linear interpolation method uses first-order linear interpolation; and the timestamp accuracy is in the millisecond range.

[0046] For example, the 100 timestamps from YYYY-MM-DD-HH:MM:SS-000 to YYYY-MM-DD-HH:MM:SS-099 correspond to 100 sampling points within the current 100 milliseconds. Each sampling point contains five-dimensional parameter values: temperature, vibration, current, pressure, and rotational speed, forming a time-aligned data stream.

[0047] Step S1.3: Perform missing value detection and forward filling on the time-aligned data stream, and identify and remove abnormal outliers to obtain a cleaned data stream.

[0048] Preferably, this step focuses on data cleaning and outlier removal. It ensures data integrity by filling in missing values ​​forward and improves data reliability by identifying and removing outliers.

[0049] Specifically, for missing value detection and forward imputation, the parameters of each dimension of each sampling point in the time-aligned data stream are traversed. If a parameter value of a certain dimension is missing, the most recent valid value of that dimension is used for forward imputation. The imputation formula is as follows: .

[0050] in x is the fill value at time t, and x(t-1) is the previous valid sample value at time t.

[0051] Specifically, for outlier identification, the mean of the data within the sliding window is calculated for each dimension parameter. with standard deviation The window length is 1000 sampling points. If the parameter value of a certain sampling point is... satisfy If it is an outlier, it is identified as an abnormal outlier and replaced with... .

[0052] To ensure data continuity after outlier removal, the replaced outliers are calculated using the mean. Filling is used to maintain the temporal continuity of the data stream and avoid temporal breaks caused by culling.

[0053] Preferably, the maximum number of consecutive forward padding operations is 5; if more than 5 consecutive missing operations occur, a sensor fault alarm is triggered; the sliding window length is 1000 sampling points (corresponding to 10 seconds of data); outlier detection is in progress. The multiple of is 3.

[0054] For example, since the vibration acceleration data at time YYYY-MM-DD-HH:MM:SS-042 is missing, the vibration acceleration of 2.31g at time YYYY-MM-DD-HH:MM:SS-041 is used for forward filling; the current value at time YYYY-MM-DD-HH:MM:SS-078 is 187.3 amperes, while the mean current within the sliding window at that time is 45.8 amperes and the standard deviation is 2.1 amperes. It was determined to be an outlier and replaced with 45.8 amperes.

[0055] Step S1.4: Smooth the purified data stream by using a moving average filter to obtain a noise-reduced data stream.

[0056] Preferably, the core of this step is to smooth and reduce noise in the data. By using a moving average filter, high-frequency noise and random fluctuations in the sensor acquisition process are suppressed, and the signal-to-noise ratio of the data is improved to reduce the feature extraction error of the subsequent network.

[0057] For the moving average filtering, a sliding window is set for the parameter values ​​of each dimension in the cleaned data stream. The window length is N sampling points. The arithmetic mean of the data within the window is taken as the filtered output at the current time. The filtering formula is: .

[0058] Where y(t) is the filtered output value at time t. Let N be the i-th original sampled value within the window, and N be the window length.

[0059] Regarding the selection of the window length N, an excessively large N value will lead to increased signal delay, while an excessively small N value will result in limited noise reduction effect. In this embodiment, different window lengths are set for different dimension parameters: the vibration dimension window length is 51 (corresponding to 0.51 seconds), the temperature dimension window length is 101 (corresponding to 1.01 seconds), and the other dimensions window length is 21 (corresponding to 0.21 seconds).

[0060] In the boundary handling of the sliding window, when the window at the beginning of the data stream has less than N sampling points, the average of the number of available data points is used to ensure the availability of the data in the initial segment.

[0061] Preferably, the length of the vibration dimension filtering window is 51; the length of the temperature dimension filtering window is 101; and the length of the current, pressure, and rotational speed dimension filtering window is 21.

[0062] For example, the original vibration acceleration value at time YYYY-MM-DD-HH:MM:SS-050 is 2.35g, the window length is 51, and the average vibration acceleration value of the first 50 sampling points is 2.28g. Therefore, the filtered output value is 2.29g, which effectively suppresses instantaneous spike noise.

[0063] Step S1.5: Perform Min-Max normalization on the noise-reduced data stream according to each dimension, and map the values ​​of each dimension to a preset range to obtain a dimensionally unified data stream.

[0064] Preferably, the core of this step is to complete data normalization. By using Min-Max standardization, parameter values ​​of different dimensions and magnitudes are uniformly mapped to a preset range, eliminating the interference of dimension differences on attention weight calculation and convolution operation in the subsequent feature fusion network.

[0065] For Min-Max standardization, the standardization formulas for each dimension parameter are as follows: .

[0066] in The standardized value. The original value, The minimum value preset for this dimension. The maximum value preset for this dimension.

[0067] Among them, the preset minimum and maximum values ​​for each dimension are set according to the normal operating parameter range and extreme operating condition parameter range of the equipment, including the temperature dimension. 0 degrees Celsius 200 degrees Celsius; vibration dimension For 0g, 10g; Current dimension For 0 amperes, 200 amps; pressure dimension 0 MPa 10 MPa; rotational speed dimension 0 revolutions per minute It is 6000 revolutions per minute.

[0068] Preferably, the standardized mapping interval is [0,1]; the boundary values ​​of each dimension are set according to the rated parameters and extreme working conditions of the equipment, and are updated every 30 days based on historical data statistics.

[0069] For example, at time YYYY-MM-DD-HH:MM:SS-100, the motor winding temperature is 78.2 degrees Celsius. (Temperature dimension) 0, It is 200, after standardization Vibration acceleration 2.29g The main motor current is 45.8 amps. .

[0070] Step S1.6: Perform tensor concatenation on the dimension-unified data stream according to the device node dimension and the time sequence dimension to obtain the standard time sequence data stream.

[0071] Preferably, the core of this step is to complete the data structure reorganization, which organizes multidimensional normalized data into a time-series tensor of a unified dimension through tensor splicing, so that the graph attention heterogeneous fusion network and the temporal causal convolutional network can directly receive it.

[0072] For tensor splicing, the dimension-unified data stream is arranged according to the device node dimension. With time series dimension Arrange them to form a shape. The three-dimensional tensor, in which This represents the number of device nodes. The time window length, The feature dimensions for each node.

[0073] Among them, the number of device nodes Based on the number of key components in the workshop, in this embodiment, the shredder is set with 10 nodes (main motor, cutter roller, left bearing housing, right bearing housing, hydraulic pump, pneumatic valve, feeding roller, dust removal port, cooling fan, and control cabinet), the shredder is set with 8 nodes, and the feeder is set with 6 nodes, for a total of [number missing] nodes. It is 24.

[0074] Among them, for the timing window length Set to 500 sampling points (corresponding to 5 seconds of data), this window backs 500 sampling points from the current time, every... Each sampling point slides once.

[0075] Preferably, the tensor shape is (24, 500, 5), where 24 is the number of device nodes, 500 is the time window length, and 5 is the feature dimension (temperature, vibration, current, pressure, rotational speed); the time window length is 500 sampling points (5 seconds); and the window sliding step is 500 sampling points.

[0076] For example, the three-dimensional tensor constructed at time YYYY-MM-DD-HH:MM:SS-000 has a shape of (24,500,5). The 5-dimensional feature value of the 500th sampling point of the first node in the tensor is [0.391,0.229,0.229,0.420,0.247], which correspond to the standardized values ​​of temperature, vibration, current, pressure, and rotational speed, respectively. This tensor is the standard time-series data stream at the current time.

[0077] Beneficial effects of steps S1.1 to S1.6: This series of steps, through multi-source operational parameter acquisition and step-by-step data processing, transforms heterogeneous operational data from different sensors and sampling frequencies into a standardized time-series data stream. This effectively solves the problem of feature modeling difficulties caused by temporal misalignment, noise interference, and dimensional differences in multi-source data. It provides a reliable and dimensionally unified input foundation for cross-dimensional feature association of graph attention heterogeneous fusion networks and temporal dependency modeling of temporal causal convolutional networks.

[0078] Specifically, step S1.1 utilizes a multi-type sensor group to cover key dimensions such as temperature, vibration, current, pressure, and rotational speed, achieving comprehensive perception of the equipment's operating status; step S1.2 eliminates temporal misalignment of multi-source data through unified sampling frequency resampling and timestamp identification mapping to ensure temporal alignment; step S1.3 ensures the integrity and reliability of the data stream during long-term continuous acquisition through forward filling of missing values ​​and anomaly removal using the 3σ criterion; step S1.4 suppresses high-frequency noise interference through moving average filtering, improving the data signal-to-noise ratio to reduce feature extraction errors in subsequent networks; step S1.5 maps data from each dimension to a unified measurement range through Min-Max standardization, eliminating interference from dimensional differences on feature fusion; and step S1.6 processes multi-dimensional data into structured temporal tensors for direct reception by downstream networks through tensor concatenation of device node dimensions and temporal dimensions.

[0079] Step S2: Input the standard time-series data stream into the graph attention heterogeneous fusion network, and perform cross-dimensional feature association and deep fusion on the standard time-series data stream through a multi-head attention mechanism to obtain the device operation fusion feature set.

[0080] Furthermore, step S2 specifically includes the following steps: Step S2.1: Divide the standard time-series data stream into several nodes according to the device component dimension, and use the data segment corresponding to each node as the node feature vector to construct the device topology graph structure.

[0081] Preferably, the core of this step is to construct a graph structure, which transforms the nodes in the standard time-series data stream, divided by the dimensions of the equipment components, into a node representation of the graph structure, so that the operating data of each equipment component is organized in the form of a graph topology.

[0082] For node partitioning, the standard time-series data stream shape is (24, 500, 5). Each of the 24 device nodes uses a 500×5 data segment as its node feature vector, with the node feature vector having the following dimension: dimension.

[0083] Among them, the construction of the device topology diagram structure is based on Each node is a set of vertices in the graph. The connections between nodes are represented by the edge set E of the graph, and the graph G=(V,E) is constructed.

[0084] Specifically, the construction of the node feature matrix involves concatenating the feature vectors of all nodes row-wise to construct the node feature matrix X, with the following shape: .

[0085] Preferably, the number of nodes is 24; each node has a feature dimension of 24. The value is 2500; the shape of the node feature matrix X is (24, 2500).

[0086] For example, the main motor node of the shredder. The corresponding node feature vector is a 2500-dimensional vector obtained by flattening the 5-dimensional parameters of 500 sampling points. In this vector, the first to 500th dimensions are the temperature parameter sequence, the 501st to 1000th dimensions are the vibration parameter sequence, and so on.

[0087] Step S2.2: Based on the physical connection relationship and signal transmission path between the device components corresponding to each node, define the heterogeneous edge type and edge weight between nodes in the device topology graph structure to obtain a weighted heterogeneous graph.

[0088] Preferably, the core of this step is to perform graph structure weighting, which encodes the physical connection relationships and signal transmission paths between device components into the graph structure, so that the graph topology not only reflects the topological adjacency relationship but also reflects the coupling strength between components.

[0089] For heterogeneous edge types, three types of edges are defined based on the physical connection relationship between equipment components: mechanical connection edges (such as the mechanical transmission relationship between a motor and a bearing), electrical connection edges (such as the electrical drive relationship between a motor and a control cabinet), and thermal connection edges (such as the heat conduction relationship between the cylinder and the hot air system).

[0090] Among them, the definition of edge weights is based on the transmission efficiency. The value is set to a range of 0 to 1; the weight of the electrical connection edge is based on the electrical coupling coefficient. The value is set to a range of 0 to 1; the weight of the thermal connection edge is based on the heat conduction ratio coefficient. The value is set to a range of 0 to 1; the edge weight between nodes with no direct connection is 0.

[0091] Specifically, for the construction of the adjacency matrix, a weighted adjacency matrix A is constructed based on the edge type and edge weight, with the following shape: Where A[i][j] represents node With nodes The edge weights between nodes are defined as follows: A[i][j]=0 indicates that there is no connection between the two nodes.

[0092] Preferably, the mechanical connection edge weight Electrical connection edge weight Thermal connection edge weight The adjacency matrix A has the shape (24, 24).

[0093] For example, the main motor of the shredder. Left of the bearing housing There is a mechanical connection between them, and the transmission efficiency is 0.95, so A[1][3]=0.95; main motor With control cabinet There is an electrical connection between them, and the electrical coupling coefficient is 0.80, so A[1]

[10] =0.80; there is a thermal connection between the drying machine cylinder and the hot air system, and the heat conduction ratio coefficient is 0.70.

[0094] Step S2.3: Perform linear mapping on the node feature vector of each node in the weighted heterogeneous graph to generate query vector, key vector, and value vector.

[0095] Preferably, the core of this step is to complete the feature space projection, which projects the node feature vectors to the query space, key space and value space respectively through linear mapping, providing a learnable feature representation for multi-head attention computation.

[0096] For linear mappings, the node feature matrix X is multiplied by three learnable weight matrices. , , Generate the query matrix Q, key matrix K, and value matrix V, with the mapping formula as follows: .

[0097] .

[0098] .

[0099] in Shape , Shape , Shape , For the dimensions of query and key, The dimension of the value.

[0100] For the multi-head attention mechanism, h attention heads are set up, and each head learns a set of attention independently. , , The weight matrix, where the query, key, and value of the i-th head are respectively , , .

[0101] Preferably, the number of attention heads h is 4; the dimensions of query and key. 16; Dimensions of the value The weight matrix for each head is 16; the weight matrix shape for each head is (2500, 16); the query matrix... The shape is (24, 16).

[0102] For example, in the first attention head, the main motor node Node feature vectors (2500 dimensions) multiplied by The query vector is then obtained. (16 dimensions), multiplied by The key vector is then obtained. (16 dimensions), multiplied by The value vector is then obtained. (16 dimensions).

[0103] Step S2.4: Perform a dot product operation on the query vector and the key vector and normalize it using Softmax to obtain the attention weight matrix between each node.

[0104] Preferably, the core of this step is to calculate the attention weights, measure the association strength between nodes by performing a dot product operation between the query vector and the key vector, and convert the association strength into an attention weight in the form of a probability distribution through Softmax normalization.

[0105] For the dot product operation, the attention score matrix of the i-th attention head... The calculation formula is: .

[0106] in Shape , The transpose of the key matrix has the following shape: , Shape , The scaling factor is used to control the numerical range of the dot product result.

[0107] Among them, for Softmax normalization, the attention score matrix Perform a Softmax operation on each row to obtain the attention weight matrix. The calculation formula is: .

[0108] in For the node in the i-th attention head The attention weights for other nodes are distributed such that the sum of the weight values ​​is 1.

[0109] In the case of incorporating edge weights into a weighted heterogeneous graph, the adjacency matrix A is used as a mask and multiplied into the attention weight matrix, so that the attention weights between nodes with no physical connection are 0. The formula is as follows: .

[0110] in This represents element-wise multiplication. The attention weight matrix after incorporating physical topological constraints.

[0111] Preferably, scaling factor The attention weight matrix has a shape of (24,24); the attention weights for nodes with no connection after masking are 0.

[0112] For example, in the first attention head, the main motor For the left bearing housing The attention score was 12.5 for the control cabinet. The attention score was 8.3, and after Softmax normalization, the attention weights were 0.42 and 0.18, respectively. After incorporating the adjacency matrix mask, For unconnected feed rollers The attention weight was changed from 0.05 to 0.

[0113] Step S2.5: The attention weight matrix and value vector are weighted and summed, and the multi-head attention outputs are concatenated and linearly projected to obtain node-level fusion features.

[0114] Preferably, the core of this step is to complete multi-head attention aggregation and feature fusion. The value vector is weighted and summed by attention weights to aggregate the feature information of neighboring nodes, and the outputs of multiple attention heads are concatenated and linearly projected to generate node-level fused features.

[0115] Among them, for the weighted summation, the output of the i-th attention head The calculation formula is: .

[0116] in Shape , Shape , Shape , The j-th row represents the node Feature representation after incorporating neighbor information.

[0117] For multi-head concatenation, the outputs of h attention heads are concatenated column-wise to obtain the concatenated feature matrix. , shape .

[0118] For linear projection, the feature matrix will be concatenated. Multiply by the learnable output weight matrix The node-level fusion feature Z is obtained, and the formula is: .

[0119] in Shape , To output the feature dimension, the Z shape is... .

[0120] Preferably, the shape after multi-head splicing is (24, 64), that is, 4 heads × 16 dimensions = 64 dimensions; output weight matrix The shape is (64, 64); the shape of the node-level fusion feature Z is (24, 64).

[0121] For example, the first attention head output The shape is (24, 16), and after splicing 4 attention points... The shape is (24, 64), and the node-level fused feature Z shape after linear projection is (24, 64), where the main motor node... The fusion feature is a 64-dimensional vector, which has incorporated the feature information of physically related nodes such as bearing housings and control cabinets.

[0122] Step S2.6: Perform a global pooling operation on the node-level fusion features to aggregate the feature information of all nodes and obtain the device operation fusion feature set.

[0123] Preferably, the core of this step is to complete graph-level feature aggregation, which aggregates node-level fused features into a graph-level representation through global pooling, and integrates the local fused features of 24 device nodes into a device-level global fused feature set.

[0124] For global pooling, a dual-path parallel strategy of global average pooling and global max pooling is adopted. Average pooling extracts the overall distribution features of node features, while max pooling extracts the salient features of node features. The two outputs are concatenated according to the channel dimension and then dimensionality reduced by a fully connected layer to obtain the device operation fusion feature set.

[0125] For global average pooling, the mean value of the node-level fused feature Z is taken along the node dimension, and the formula is: .

[0126] in This is the output of average pooling.

[0127] For global max pooling, the maximum value of the node-level fused feature Z is taken along the node dimension, and the formula is: .

[0128] in This is for max-pooling output.

[0129] Among them, for dual-path splicing and dimensionality reduction, and spliced ​​as A 3D vector is mapped to a fully connected layer. Dimension, to obtain the device operation fusion feature set The formula is: .

[0130] in The weights of the fully connected layer are of shape [shape]. , For bias, This is a set of integrated features for equipment operation.

[0131] Preferably, global average pooling and global max pooling are executed in parallel; the dual-path concatenation dimension is 128 dimensions (64×2); the output dimension after dimensionality reduction by the fully connected layer is... 64-dimensional; Equipment operation fusion feature set .

[0132] For example, the node-level fusion feature Z(24,64) of 24 device nodes, after dual-path pooling, is averaged and output. The output is a 64-dimensional vector, max-pooled. The 64-dimensional vectors are concatenated to form a 128-dimensional vector. After dimensionality reduction using a fully connected layer, a 64-dimensional device operation fusion feature set is obtained. This feature set integrates the operating status information of all key components of the shredder, shredder, and feeder.

[0133] Beneficial effects of steps S2.1 to S2.6: This series of steps transforms standard temporal data streams into a fusion feature set of device operation through a graph attention heterogeneous fusion network. It models the physical relationships and signal transmission paths between device components using a graph structure, and achieves autonomous association and deep fusion of cross-dimensional features through a multi-head attention mechanism. This effectively solves the problem that traditional feature splicing methods cannot characterize the topological relationships and coupling features between device components, resulting in the loss of deep association information in multi-source operational data. It provides high-quality fusion features containing device topological semantics for subsequent temporal dependency modeling of temporal causal convolutional networks.

[0134] Specifically, step S2.1 divides the discrete time-series data into a physically meaningful graph representation by dividing nodes according to device component dimensions and constructing a topological graph structure; step S2.2 encodes the physical connection relationships and signal transmission paths between device components into the graph structure to represent the coupling relationships between nodes by defining heterogeneous edge types and edge weights; step S2.3 generates query vectors, key vectors, and value vectors by linearly mapping node feature vectors, providing a learnable feature space projection for attention calculation; step S2.4 achieves autonomous measurement and weight allocation of the association strength between nodes by performing dot product operations on query vectors and key vectors and Softmax normalization; step S2.5 fuses feature information from different subspaces to generate node-level fused features by weighted summation of attention weight matrices and value vectors and multi-head splicing projection; and step S2.6 aggregates the feature information of all nodes through global pooling, transforming the graph-level representation into a device-level fused feature set for downstream network reception.

[0135] Step S3: Input the device running fusion feature set into the temporal causal convolutional network, and capture the temporal dependency relationship through the causal convolution and dilated convolution structure to obtain the temporal associated feature sequence.

[0136] Furthermore, step S3 specifically includes the following steps: Step S3.1: The device operation fusion feature set is serialized and expanded according to time steps to obtain the fusion feature time sequence.

[0137] Preferably, the core of this step is to complete the serialization unrolling, which arranges the device operation fusion feature set of multiple consecutive time steps into a time sequence in chronological order, providing a standard time sequence input for causal convolution operations.

[0138] Specifically, for serialization unrolling, backtracking is performed with the current time as the endpoint. Each time step will fuse the device operation feature set at each time step. Arrange them in chronological order to construct a time-series sequence of fused features. The sequence length is The feature dimension of each time step is .

[0139] Regarding the selection of sequence length, It needs to cover the typical timescales of equipment failure evolution so that the causal convolutional network can capture a sufficiently long historical dependency to identify failure trends.

[0140] Preferably, the sequence length There are 100 time steps (each time step corresponds to one sampling window, i.e., 5 seconds of data, for a total duration of 500 seconds); the fused feature time series sequence The shape is (100, 64).

[0141] For example, if the current time is YYYY-MM-DD-HH:MM:SS-000, and we backtrack 100 time steps, we can arrange the device operation fusion feature set of the 100 time steps into a fusion feature time sequence of (100, 64).

[0142] Step S3.2: Apply causal convolution operation to the fused feature time series sequence, where each convolution kernel only performs convolution operation on the data of the current time step and the historical time steps to obtain the causal feature sequence.

[0143] Preferably, the core of this step is to complete the causal convolution operation. By constraining the convolution kernel to only perceive the current and historical time step data, the leakage of future information is prevented to ensure the causal rationality of the time series modeling.

[0144] For causal convolution, the kernel size is... The input data for time t is only related to time t. The data is convolved, and the output formula is: .

[0145] in For convolution kernel weights, Let be the input features at time ti. Let be the causal convolution output at time t.

[0146] For the zero-padding strategy of causal convolution, padding is performed at the beginning of the sequence. The zero values ​​are used to ensure that the output sequence has the same length as the input sequence, so that there is a corresponding causal convolution output at each time step.

[0147] Specifically, regarding the number of convolutional kernels, set... Multiple convolutional kernels are used to extract various feature patterns, and the output feature dimension is [value missing]. .

[0148] Preferably, the causal convolution kernel size The number of convolutional kernels is 3. The value is 64; the shape of the causal feature sequence is (100, 64); the number of zero-padding is 2.

[0149] For example, fusion feature time sequence The output at time t=50 is of shape (100, 64) after a causal convolution operation with a kernel size of 3. Using only the data from the current and the previous two time steps, the causal feature sequence output shape is (100, 64).

[0150] Step S3.3: Apply dilated convolution operation to the causal feature sequence, and expand the sampling interval of the convolution kernel according to the preset dilation factor to expand the temporal receptive field without increasing the number of parameters, so as to obtain the dilated feature sequence.

[0151] Preferably, the core of this step is to perform dilated convolution operation, which expands the receptive field by introducing an interval between the sampling points of the convolution kernel, thereby capturing longer-range temporal dependencies without increasing the number of parameters or the amount of computation.

[0152] For dilated convolution, the dilation factor is... The kernel size is The output formula at time t is: .

[0153] in For the weights of the dilated convolution kernel, It is the expansion factor. For interval The sampled values, the receptive field of dilated convolution .

[0154] Among them, for the accumulation of receptive fields in multi-layer dilated convolution, the dilation factor of the l-th layer dilated convolution... The overall sensory field after multiple layers are stacked , where L is the number of layers.

[0155] Preferably, the size of the dilated convolution kernel The expansion factor is 3; the first layer expansion factor is 3. =1, receptive field =3; second layer expansion factor The receptive field is 2, and the expansion factor of the third layer is 6. The receptive field is 4, and the receptive field is expanded to 12; the shape of the dilated feature sequence is (100, 64).

[0156] For example, in the second layer of dilated convolution, the dilation factor is 2, and the output at time t=50... The sampling interval is 2, and the receptive field is extended to 6 time steps.

[0157] Step S3.4: Perform batch normalization and nonlinear activation operations on the dilated feature sequence sequentially to obtain the activated feature sequence.

[0158] Preferably, the core of this step is to stabilize the feature distribution and introduce nonlinearity. Batch normalization is used to stabilize the feature distribution of each layer to accelerate training convergence, and nonlinear activation is used to enhance the expressive power of the features.

[0159] For batch normalization, the mean of the dilated feature sequence is calculated along the batch dimension. With variance Then, normalization and scaling / translation are performed, and the formula is: .

[0160] in and For learnable scaling and translation parameters, To prevent small constants from being divided by zero.

[0161] For nonlinear activation, the ReLU function is used to perform a nonlinear transformation on the batch-normalized output, as shown in the formula: .

[0162] in To activate the feature sequence, negative values ​​are truncated to 0 to introduce nonlinearity.

[0163] Preferably, batch normalization for Scaling parameters The initial value is 1, and the translation parameter is... The initial value is 0; the non-linear activation function is ReLU; the shape of the activation feature sequence is (100, 64).

[0164] For example, the value of a certain channel in a dilated feature sequence at a certain time is -0.5, which becomes -1.2 after batch normalization. After ReLU activation... Negative values ​​are truncated to preserve valid features.

[0165] Step S3.5: The activation feature sequence is used as the input of the next layer of causal dilated convolution for multi-layer stacking iteration to expand the temporal receptive field and deepen the temporal dependency modeling layer by layer, so as to obtain the temporal correlation feature sequence.

[0166] Preferably, the core of this step is to complete multi-layer stacking and temporal deepening. By connecting multi-layer causal dilated convolution, batch normalization and nonlinear activation layer by layer, the receptive field is expanded layer by layer and the temporal dependency modeling depth is accumulated, and finally the temporal correlation feature sequence is output.

[0167] For multi-layer stacked structures, a total of [number] are set. The layered causal dilated convolution module contains four sub-operations in each layer: causal convolution, dilated convolution, batch normalization, and nonlinear activation. The output of the l-th layer is used as the input of the l+1-th layer. Residual connections are set between layers to alleviate gradient degradation.

[0168] For interlayer residual connections, when the input and output dimensions are the same, the residual connection formula is: .

[0169] in This is the output after the l-th layer is activated. For the input of the l-th layer, The output after adding the residuals is used as the input of the (l+1)th layer.

[0170] Among them, for the final output, the first The residual output of the layer is the temporal correlation feature sequence. The features at each time step in this sequence have accumulated long-range temporal dependency information through multi-layer causal dilated convolution.

[0171] Preferably, the number of causal dilated convolutional layers It has 4 layers; the expansion factors of each layer are 1, 2, 4, and 8 respectively; the total receptive field There are 225 time steps (corresponding to 225 seconds); the shape of the temporal correlation feature sequence is (100, 64).

[0172] For example, after stacking 4 layers of causal dilated convolutions, the output features at time t=99 have accumulated temporal dependency information from t=55 to t=99, a total of 45 time steps. This feature sequence contains long-term temporal dynamic information about the evolution of the device state from normal operation to potential failure, which is the temporal correlation feature sequence.

[0173] Beneficial effects of steps S3.1 to S3.5: This series of steps uses a temporal causal convolutional network to model the temporal dependencies of the fused feature set of device operation. By utilizing the temporal direction constraints of causal convolution and the receptive field expansion mechanism of dilated convolution, it captures the long-range temporal correlation of device operation status while ensuring temporal causality. This effectively solves the problems of future information leakage and insufficient long-range dependency capture capability in traditional convolutional networks in temporal modeling, and provides a sequence of associated features containing rich temporal dynamic information for subsequent hierarchical residual inference.

[0174] Specifically, step S3.1 transforms the fused feature set into a temporal sequence form by unfolding it sequentially by time step, providing a standard temporal input for causal convolution operations; step S3.2 constrains the convolution kernel to only perceive current and historical time step data through causal convolution, preventing future information leakage and ensuring the causal rationality of temporal modeling; step S3.3 expands the sampling interval by dilating convolution according to a preset dilation factor, increasing the temporal receptive field to capture longer-range temporal dependencies without increasing the number of parameters; step S3.4 stabilizes the feature distribution of each layer and enhances the nonlinear expressive power of the features through batch normalization and nonlinear activation operations; step S3.5 expands the receptive field layer by layer and deepens temporal dependency modeling through multi-layer stacking iteration, accumulating the output of multi-layer causal dilated convolution into a sequence of associated features containing long-range temporal dynamics.

[0175] Step S4: Perform hierarchical residual deduction on the time-series associated feature sequence, and complete the feature deepening mapping layer by layer through multi-layer residual connections to obtain the deepened deduction feature sequence.

[0176] Furthermore, step S4 specifically includes the following steps: Step S4.1: Input the temporal correlation feature sequence into the first layer inference unit, perform linear mapping through a fully connected layer and nonlinear activation operation to obtain the first layer inference features.

[0177] Preferably, the core of this step is to complete the first layer of feature deepening mapping, which projects the temporally related feature sequence to a high-dimensional feature space through a fully connected layer and introduces nonlinear expressive power through nonlinear activation.

[0178] Specifically, for the linear mapping of the fully connected layer, the feature vector of each time step in the temporally associated feature sequence is... (64-dimensional) multiplied by the learnable weight matrix And add bias The linear mapping output is obtained. The formula is: .

[0179] in Shape , The input feature dimension is 64. For the hidden layer dimension, The output is a linear mapping at time t.

[0180] For nonlinear activation, the ReLU function is used to perform a nonlinear transformation on the linear mapping output, as shown in the formula: .

[0181] in The first layer of derivation features is truncated to 0 for negative values ​​to introduce nonlinearity.

[0182] Preferably, the input feature dimension The hidden layer dimension is 64. The weight matrix is ​​128; The shape is (64, 128); bias The shape is (128,*); the non-linear activation function is ReLU; the shape of the first layer of inferred features is (100,128).

[0183] For example, the feature vector at time t=99 in a time-series correlation feature sequence. (64-dimensional) After linear mapping, a 128-dimensional vector is obtained. The first layer of inferred features after ReLU activation It is a 128-dimensional vector, where the negative components have been truncated to 0.

[0184] Step S4.2: Add the residuals of the first-layer inferred features and the temporal correlation feature sequence to obtain the first-layer residual inferred features.

[0185] Preferably, the core of this step is to complete the first layer of residual connection. By adding the first layer of inferred features to the temporal correlation feature sequence element by element, the original temporal correlation information is preserved while the features are deepened, so as to alleviate gradient degradation.

[0186] Among them, for the summation of residuals, due to the dimension of the first layer of inference features Dimension of Time-Series Related Features Inconsistency requires the use of a projection matrix. The temporal correlation feature sequence is projected to the same dimension as the inferred feature, and then the residuals are summed. The formula is as follows: .

[0187] in Shape , The temporal correlation features at time t, These are the original features after projection. This represents the characteristics of the first layer of residual derivation.

[0188] Regarding the role of the projection matrix, The low-dimensional original features are mapped to a high-dimensional space to match the inferred feature dimension, so that the dimension of the residual addition operation is consistent.

[0189] Preferably, the projection matrix The shape is (64, 128); first-layer residual derivation characteristics The shape is (100, 128); the residuals are added element by element.

[0190] For example, the first-level derivation features at time t=99 A 128-dimensional vector, temporal correlation features through The projected vector is a 128-dimensional vector. The first layer of residual derivation features is obtained by adding the two vectors element by element. It is a 128-dimensional vector that contains both the enhanced higher-order features and the original temporal correlation information.

[0191] Step S4.3: Input the first layer residual inference features into the second layer inference unit, perform linear mapping through a fully connected layer and nonlinear activation operation to obtain the second layer inference features.

[0192] Preferably, the core of this step is to complete the second layer of feature deepening mapping, which projects the residual inference features of the first layer to a high-dimensional feature space through a fully connected layer and extracts higher-order abstract features through nonlinear activation.

[0193] For the linear mapping of the fully connected layer, the residual of the first layer is used to derive the features. (128-dimensional) multiplied by the learnable weight matrix And add bias The linear mapping output is obtained. The formula is: .

[0194] in Shape , The output of the second-level linear mapping at time t.

[0195] For nonlinear activation, the ReLU function is used to perform a nonlinear transformation on the linear mapping output, as shown in the formula: .

[0196] in This is the second layer of deductive features.

[0197] Preferably, the weight matrix Shape (128, 128); bias The shape is (128, *); the non-linear activation function is ReLU; the second layer of inferred features The shape is (100, 128).

[0198] For example, the first-level residual extrapolation characteristics at time t=99 (128 dimensions) After linear mapping, a 128-dimensional vector is obtained. The second layer of inferred features after ReLU activation It is a 128-dimensional vector, from which higher-order abstract features have been extracted.

[0199] Step S4.4: Add the residuals of the second-layer derivation features and the first-layer residual derivation features to obtain the second-layer residual derivation features.

[0200] Preferably, the core of this step is to complete the second layer of residual connection. By adding the second layer of deduced features to the first layer of residual deduced features element by element, a direct information pathway is established between each layer to ensure the integrity of deep information transmission.

[0201] For the addition of residuals, since the second-layer inference features and the first-layer residual inference features both have 128 dimensions, they are directly added element-wise, as shown in the formula: .

[0202] in This represents the characteristics of the second-level residual derivation. This is a second-level deduction feature. This represents the characteristics of the first layer of residual derivation.

[0203] Among them, the information preservation effect of residual connections is that the addition of residuals ensures that the output of deep networks always contains shallow feature information, thus preventing the original information from being submerged as the number of layers increases.

[0204] Preferably, the second-layer residual derivation features The shape is (100, 128); the residuals are added element-wise, without the need for a projection matrix.

[0205] For example, the second-level derivation features at time t=99 Characteristics derived from the first layer residual All are 128-dimensional vectors, and the second-layer residual inference features are obtained by adding them element by element. It is a 128-dimensional vector that accumulates the feature enhancement results of the first and second layers.

[0206] Step S4.5: Input the second-layer residual extrapolation features into the subsequent extrapolation units, and perform linear mapping, nonlinear activation and residual addition operations layer by layer according to the extrapolation logic of the first-layer extrapolation units and the second-layer extrapolation units to obtain the in-depth extrapolation feature sequence.

[0207] Preferably, the core of this step is to complete the iterative stacking of multi-level residual inference, and to connect the subsequent layers one by one according to the inference logic of the first two layers, so as to accumulate the multi-level feature transformation results into a rich in-depth inference feature sequence with rich representation levels.

[0208] Among them, for the iteration logic of subsequent layers, the l-th layer ( The derivation process is as follows: The residual characteristics of the (l-1)th layer are deduced. The l-th layer inference unit is input, and the l-th layer inference features are obtained through linear mapping and ReLU nonlinear activation in a fully connected layer. , and then with The residuals are summed to obtain the residual derivation characteristics of the l-th layer. The formula is: .

[0209] .

[0210] .

[0211] in The shape is (128, 128), and the dimensions of each layer are consistent, so no projection matrix is ​​needed.

[0212] Among them, for the final output, the first Residual extrapolation characteristics of the layer This refers to the in-depth extrapolation of the feature sequence, in which the features of each time step have accumulated rich hierarchical abstract information through multi-layer residual extrapolation.

[0213] Preferably, the total number of residual derivation layers It has 4 layers (including the first and second layers); each fully connected layer has a dimension of 128; weight matrix The shapes are all (128, 128); the shape of the in-depth deduction feature sequence is (100, 128).

[0214] For example, after stacking 4 layers of residual extrapolation, the output at time t=99 It is a 128-dimensional vector that accumulates the layer-by-layer deepening results from the temporal correlation feature sequence to the 4-layer nonlinear mapping, representing the richly layered deepening inference feature sequence, which is the final output.

[0215] Beneficial effects of steps S4.1 to S4.5: This series of steps performs multi-layer feature deepening mapping on temporal correlation feature sequences through hierarchical residual deduction. By utilizing the residual connection mechanism, the original feature information is preserved and high-order abstract features are superimposed during the layer-by-layer transformation. This effectively solves the problem of information loss and feature degradation caused by gradient degradation in deep networks during feature deepening. It provides a deep deduction feature sequence with strong representation ability and high information fidelity for subsequent global pooling and feature dimension compression.

[0216] Specifically, step S4.1 projects the temporal correlation feature sequence to a high-dimensional feature space through linear mapping and nonlinear activation operations in a fully connected layer to introduce nonlinear expressive power; step S4.2 adds the residuals of the first-layer inference features and the temporal correlation feature sequence to deepen the features while preserving the original temporal correlation information to alleviate gradient degradation; step S4.3 further extracts higher-order abstract features through fully connected mapping and nonlinear activation of the second-layer inference unit to deepen the feature expression level; step S4.4 adds the residuals of the second-layer inference features and the first-layer residual inference features to establish a direct information pathway between layers to ensure the integrity of deep information transmission; and step S4.5 accumulates the multi-level feature transformation results into a richly represented inference feature sequence by iteratively executing linear mapping, nonlinear activation, and residual addition operations layer by layer according to the preceding inference logic.

[0217] Step S5: Perform global pooling and feature dimension compression on the in-depth inference feature sequence, remove redundant feature components and extract effective features to obtain the standard feature vector of equipment status.

[0218] Furthermore, step S5 specifically includes the following steps: Step S5.1: Perform global average pooling operation on the in-depth inference feature sequence along the time dimension to aggregate each feature channel into a single feature vector in the time dimension, and obtain the time-series aggregated feature vector.

[0219] Preferably, the core of this step is to complete the temporal dimension aggregation, and compress the in-depth inference feature sequence of 100 time steps into a single feature vector through global average pooling, thereby eliminating the temporal dimension and realizing the conversion of dynamic features to static features.

[0220] Among them, for global average pooling, the feature sequence R (of shape ) is further derived. Along the time series dimension The average value is calculated using the following formula: .

[0221] in Let R[t][j] be the aggregated value of the j-th feature channel, and R[t][j] be the feature value of the j-th channel at time t. For sequence length, This is a time-series aggregated feature vector.

[0222] Among them, the information retention characteristic of global average pooling is that average pooling retains the overall distribution information of features within the entire time window, which has a smoothing effect on abnormal fluctuations, making the aggregated feature vector more reflective of the operating status characteristics of the device within the entire time window.

[0223] Preferably, the time-series dimension 100; Feature dimension 128; Temporal aggregated feature vector The shape is (128,*).

[0224] For example, if the shape of the in-depth inference feature sequence R is (100, 128), after averaging over 100 time steps, the aggregate value of the j=1th channel is... =0.347, the aggregated values ​​of 128 channels constitute the time-series aggregated feature vector. .

[0225] Step S5.2: Calculate the feature variance of each dimension in the time-series aggregated feature vector, and sort each dimension according to the size of the variance to obtain the feature importance ranking result.

[0226] Preferably, the core of this step is to complete the feature importance assessment by calculating the variance of each feature dimension to quantify the information contribution, and sorting them by variance to construct the feature importance ranking result.

[0227] Specifically, for variance calculation, the values ​​of the time-series aggregated feature vector across multiple historical time-series windows are statistically analyzed. For the j-th feature dimension, the most recent values ​​are collected. The aggregated values ​​of each time window are used to calculate the variance. .

[0228] in For the j-th dimension in The mean of each window, Let be the variance of the j-th dimension.

[0229] For variance sorting, the variance of each dimension is... Sort in descending order to obtain the feature importance ranking results. The larger the variance, the greater the fluctuation of the feature value of that dimension between different time windows, that is, the richer the information content and the higher the contribution to the device status judgment.

[0230] Preferably, the number of statistical windows The value is 50 (corresponding to the data from the most recent 50 time series windows); the variance is calculated as an unbiased estimate; and the sorting method is descending order.

[0231] For example, among the 128 feature dimensions, the dimension with the largest variance, j=37, is... =0.0821, the dimension with the smallest variance j=115. =0.0003, after sorting, j=37 is ranked 1st, and j=115 is ranked 128th.

[0232] Step S5.3: Based on the feature importance ranking results, feature dimensions with variances below a preset threshold are identified as redundant feature components and removed to obtain an effective feature subset.

[0233] Preferably, the core of this step is to remove redundant features. By using a preset variance threshold, feature dimensions with low information contribution are selected and removed, thereby reducing input noise interference and improving the discrimination accuracy of the subsequent classifier.

[0234] Specifically, for the variance threshold determination, the variance of each dimension is considered. Compared with the preset variance threshold If a comparison is made, If the j-th dimension is determined to be a redundant feature component, it will be removed, while the variance will be retained. The dimensions constitute an effective feature subset.

[0235] Among them, the selection of the variance threshold, It is determined by multiplying the median of the variances of all dimensions by a proportionality coefficient, as shown in the formula: .

[0236] in To perform the median calculation, This is the proportionality coefficient. This is the variance threshold.

[0237] Preferably, the proportionality coefficient 0.1; variance threshold It is 0.1 times the median; the number of effective feature dimensions retained after removal is dynamically determined according to the data distribution.

[0238] For example, the median variance across 128 feature dimensions is 0.0156, and the variance threshold is... Dimensions with a variance lower than 0.00156 are identified as redundant feature components and removed. Assuming a total of 32 dimensions are removed, the effective feature subset has 96 dimensions.

[0239] Step S5.4: The effective feature subset is linearly reduced and compressed to the preset target dimension through a fully connected layer to obtain the standard feature vector of the device state.

[0240] Preferably, the core of this step is to compress the feature dimension by mapping the effective feature subset to the preset target dimension through a fully connected layer, thereby reducing the dimension of the feature vector to adapt to the input requirements of the classifier while retaining the core state information.

[0241] For linear dimensionality reduction mapping, the effective feature subset is... (dimension is) Multiply by the learnable weight matrix And add bias Mapped to the target dimension The formula is: .

[0242] in Shape , For effective feature subset dimension, For the target dimension, This is the standard feature vector of the device status.

[0243] Among them, the selection of the target dimension, A balance needs to be struck between information retention and computational efficiency. Too many dimensions increase the computational overhead of the classifier, while too few dimensions lead to information loss.

[0244] Preferably, the effective feature subset dimension 96 (dynamically determined); target dimension The weight matrix is ​​64. The shape is (96, 64); bias Shape (64, *); Standard feature vector of device status The shape is (64,*).

[0245] For example, effective feature subset It is a 96-dimensional vector, after After linear dimensionality reduction mapping, it is compressed into a 64-dimensional standard feature vector of device state. This vector, while retaining the core device status information, achieves dimensionality reduction and can be directly used as input for subsequent multilayer perceptron classifiers.

[0246] Beneficial effects of steps S5.1 to S5.4: This series of steps reduces and condenses the feature sequence of the in-depth inference through global pooling and feature dimension compression. It aggregates dynamic features in the time dimension into static feature vectors and removes redundant components based on feature importance. This effectively solves the problems of increased computational overhead and interference from classifier input noise caused by dimensional redundancy in the feature sequence of the in-depth inference. It provides a standard feature vector of device state with simplified dimensions and high information density for the subsequent multilayer perceptron classifier.

[0247] Specifically, step S5.1 aggregates dynamic temporal features into static feature vectors by performing global average pooling along the temporal dimension to eliminate temporal dimension redundancy; step S5.2 quantifies the information contribution of each feature component by calculating and sorting the variance of each feature dimension to construct a basis for evaluating feature importance; step S5.3 removes low-information features by identifying feature dimensions with variances below a preset threshold as redundant components; and step S5.4 compresses effective features to a preset target dimension through linear dimensionality reduction mapping using a fully connected layer, achieving dimensionality simplification of the feature vector while retaining core state information to adapt to the input requirements of the classifier.

[0248] Step S6: Input the standard feature vector of the equipment status into the multilayer perceptron classifier, output the probability distribution of the equipment status through fully connected mapping and Softmax normalization operation, and output the real-time fault warning level data of the tobacco production equipment based on the comparison result of the equipment status probability distribution and the preset grading threshold.

[0249] Furthermore, step S6 specifically includes the following steps: Step S6.1: Input the standard feature vector of the device state into the multilayer perceptron classifier, and perform feature transformation layer by layer through several fully connected layers and nonlinear activation functions to obtain the category space mapping vector.

[0250] Preferably, the core of this step is to complete the mapping from features to the category space. Through the layer-by-layer fully connected mapping and nonlinear activation of the multilayer perceptron, the standard feature vector of the device state is transformed to the category space to establish the discriminative relationship between features and device state.

[0251] Specifically, for the structure of the multilayer perceptron classifier, the following settings are made: A fully connected layer, the first layer will store the standard feature vector of the device state. (64-dimensional) mapping to hidden dimensions The subsequent layers are The transformation is performed within the dimension, and the last layer maps the hidden features to the category space dimension. The formula is: .

[0252] .

[0253] .

[0254] .

[0255] .

[0256] in Shape , Shape , Shape , This is a category space mapping vector.

[0257] Specifically, for the category definition of the category space, the settings are as follows: The equipment status is categorized into five types: normal operation, minor abnormality, moderate abnormality, severe abnormality, and fault shutdown.

[0258] Preferably, the number of fully connected layers It has 3 layers (including 2 hidden layers and 1 output layer); hidden layer dimensions The number of categories is 32. The weight matrix is ​​5; The shape is (64, 32). The shape is (32,32). The shape is (32,5); the non-linear activation function is ReLU; the class space mapping vector is... The shape is (5,*).

[0259] For example, the standard feature vector of device status The vector is a 64-dimensional vector. It is mapped to a 32-dimensional hidden layer through a first fully connected layer and activated by ReLU. Then, it is transformed within the 32-dimensional hidden layer through a second fully connected layer and activated by ReLU. Finally, it is mapped to a 5-dimensional class space through the output layer to obtain the class space mapped vector. =[2.1,0.8,-0.3,-1.5,-2.0], which correspond to the original scores of the five categories: normal operation, minor abnormality, moderate abnormality, severe abnormality, and fault shutdown, respectively.

[0260] Step S6.2: Apply the Softmax function to the category space mapping vector to perform normalization operation, convert each component into a probability value, and obtain the device state probability distribution.

[0261] Preferably, the core of this step is to complete the probabilistic output, which uses the Softmax function to convert the category space mapping vector into a probability distribution, thereby realizing the probabilistic expression of the device status discrimination result.

[0262] Among them, for Softmax normalization, the class space is mapped to a vector. The exponential function of each component is calculated and normalized, and the formula is: .

[0263] Where p[c] is the probability value of the c-th category, Let be the original score of the c-th category, with the denominator being the exponential sum of the scores of all categories, and p being the probability distribution of the device state, with the sum of all components being 1.

[0264] For numerical stability, the exponential function is calculated before... Subtract the maximum value from each component To prevent exponential overflow, the formula is: .

[0265] Preferably, the number of categories The probability distribution p of the device state is 5; the shape of the probability distribution p is (5,*); the probability value range is [0,1]; the sum of all probability values ​​is 1.

[0266] For example, category space mapping vector =[2.1,0.8,-0.3,-1.5,-2.0], subtracting the maximum value of 2.1 gives [0,-1.3,-2.4,-3.6,-4.1], and taking the exponent gives [1.000,0.272,0.091,0.027,0.016]. The normalized probability distribution of the equipment state is p=[0.731,0.199,0.067,0.020,0.012], which correspond to 73.1% normal operation, 19.9% ​​minor abnormality, 6.7% moderate abnormality, 2.0% severe abnormality, and 1.2% failure shutdown, respectively.

[0267] Step S6.3: Extract the maximum probability value and its corresponding state category from the device state probability distribution to obtain the current device state determination result.

[0268] Preferably, the core of this step is to complete the state determination, extract the maximum probability value and its corresponding category from the probability distribution, and obtain the most likely state determination result of the current device.

[0269] For the extraction of the maximum probability value, the maximum value and its index are found by traversing each component of the device state probability distribution p. The formula is as follows: .

[0270] .

[0271] in The maximum probability value, The category index corresponding to the highest probability value indicates the current device status determination result is... .

[0272] Regarding the meaning of the state determination result, =1 indicates normal operation. =2 indicates a mild abnormality. =3 indicates moderate abnormality =4 indicates severe abnormality. =5 indicates a system failure and shutdown.

[0273] Preferably, the maximum probability value is extracted using the argmax operation; the state determination result includes both the probability value and the category index.

[0274] For example, the device state probability distribution p=[0.731,0.199,0.067,0.020,0.012], with the maximum probability value... =0.731, corresponding to the category index =1, the current device status is determined as (0.731, normal operation).

[0275] Step S6.4: Compare the maximum probability value with the preset first and second level thresholds, and determine the fault warning level based on the comparison result, wherein the first level threshold is lower than the second level threshold.

[0276] Preferably, the core of this step is to complete the early warning classification. By comparing the maximum probability value with two thresholds, the fault early warning level is determined, thereby realizing differentiated classification of early warnings to guide different levels of operation and maintenance response.

[0277] For the dual-threshold grading mechanism, a first grading threshold is preset. With the second grade threshold , { The grading rules are as follows: when When the condition is classified as normal operation or minor abnormality: like The system is classified as a Level 1 warning (normal monitoring). The equipment is operating well and routine monitoring should be maintained.

[0278] like The system is classified as a Level 2 warning (attention warning), indicating a potential risk trend in the equipment's operating status, prompting an increase in monitoring frequency.

[0279] like The system is classified as a Level 3 warning (fault warning), indicating a high degree of uncertainty in the equipment's operating status, requiring immediate investigation.

[0280] when When the abnormality is classified as moderate, severe, or fault-related shutdown: regardless The value is directly determined as a Level 4 warning (emergency warning), indicating that the equipment has a clear risk of failure and an immediate maintenance response is required.

[0281] Regarding the meaning of the threshold, The lower limit of confidence in the state determination is used to characterize the state determination. If the confidence level is lower than this value, it means that the classifier has insufficient confidence in the state determination result and there is an uncertainty risk in the state of the equipment. The upper limit of confidence in the state determination is represented; values ​​above this value indicate that the classifier has sufficient confidence in the state determination result.

[0282] Preferably, the first grading threshold The threshold value is 0.70; the second-level threshold is... The value is 0.85; there are four warning levels (normal monitoring, attention warning, fault warning, and emergency warning).

[0283] For example, the current device status determination result is (0.731, normal operation). For normal operation category, =0.731, The system is classified as a Level 2 warning (attention warning), indicating a potential risk trend in equipment operation, prompting an increase in monitoring frequency. If the next judgment result changes to (0.65, normal operation), If the condition is detected, it is classified as a Level 3 warning (fault warning) and immediate investigation is required.

[0284] Step S6.5: Generate early warning command data based on the fault early warning level and the current equipment status determination result, and output the early warning command data as the real-time fault early warning level data of the tobacco production equipment.

[0285] Preferably, the core of this step is to generate and output early warning instructions, integrate the fault early warning level and status judgment results into structured early warning instruction data, and output it to the equipment operation and maintenance system as real-time fault early warning level data.

[0286] The structure of the early warning instruction data includes the following fields: timestamp identifier, device number, device name, status category, maximum probability value, early warning level, and recommended measures. The recommended measures are automatically matched with a preset operation and maintenance strategy template based on the early warning level and status category.

[0287] Specifically, regarding the correspondence between warning levels and recommended measures, Level 1 warning (normal monitoring) corresponds to "maintain routine monitoring without intervention"; Level 2 warning (attention warning) corresponds to "increase the monitoring frequency to once every 30 seconds and pay attention to trend changes"; Level 3 warning (fault warning) corresponds to "immediately arrange personnel to check relevant equipment components and prepare spare parts"; and Level 4 warning (emergency warning) corresponds to "immediately stop the machine for inspection, activate the emergency plan, and notify the maintenance team".

[0288] Specifically, for the output interface of the early warning instruction data, the early warning instruction data is encapsulated in JSON format and pushed to the monitoring platform of the equipment operation and maintenance system through a message queue to achieve the delivery of real-time early warning information.

[0289] Preferably, the warning instruction data fields include seven fields: timestamp, device number, device name, status category, probability value, warning level, and recommended measures; the data format is JSON; and the transmission method is message queue push.

[0290] For example, the warning instruction data at time YYYY-MM-DD-HH:MM:SS-000 is: {"timestamp":"YYYY-MM-DD-HH:MM:SS-000","device_id":"KT-XXX-001","device_name":“a certain shredder”,"state_class":“normal operation”,"probability":0.731,"alert_level":“Level 2 warning”,"action":“Increase monitoring frequency to once every 30 seconds, pay attention to trend changes”}. This warning instruction data is the real-time fault warning level data of the tobacco production equipment, which has been pushed to the operation and maintenance monitoring platform.

[0291] Beneficial effects of steps S6.1 to S6.5: This series of steps transforms the standard feature vector of equipment status into real-time fault warning level data through a multilayer perceptron classifier and a hierarchical threshold comparison mechanism. It achieves probabilistic output of status using fully connected mapping and Softmax normalization, and determines the warning level based on dual threshold hierarchies. This effectively solves the problem that traditional monitoring systems only output Boolean alarm signals and lack fault probability measurement and warning level, realizing a complete closed loop from feature vector to executable warning command.

[0292] Specifically, step S6.1 maps the state standard feature vector to the category space through layer-by-layer feature transformation using several fully connected layers and nonlinear activation functions to establish a discriminative relationship between features and device state; step S6.2 converts the category space mapping vector into a probability distribution through normalization operation using the Softmax function to realize the probabilistic expression of the device state discrimination result; step S6.3 extracts the maximum probability value and its corresponding state category to obtain the most likely state judgment result of the current device from the probability distribution; step S6.4 compares it with the first and second level thresholds to determine the fault warning level according to the interval of the probability value to realize differentiated classification of warnings; step S6.5 generates warning instruction data based on the warning level and state judgment result, converting the classifier output into real-time fault warning level data that can be directly used for equipment operation and maintenance.

[0293] In summary, the overall beneficial effects of steps S1 to S6 in this embodiment are as follows: This method addresses the technical problems of isolated alarm information and heavy reliance on manual experience in fault diagnosis during tobacco production equipment fault early warning. It employs a combination of graph attention heterogeneous fusion and causal convolutional temporal modeling to achieve automated judgment from multi-source operational data to hierarchical fault early warning. This effectively overcomes the technical problems of traditional monitoring systems that only alarm without diagnosis and lack deep feature mining and intelligent early warning hierarchical classification.

[0294] Specifically, step S1 involves collecting and regularizing multi-source time-series operational data streams to construct a standardized data foundation; step S2 involves using a graph attention heterogeneous fusion network to complete cross-dimensional feature association and deep fusion of multi-source data; step S3 involves capturing the temporal dependencies of equipment operating status through causal convolution and dilated convolution structures; step S4 involves deepening feature mapping layer by layer through hierarchical residual deduction to enhance feature expression capabilities; step S5 involves refining the core features of equipment status through global pooling and dimensional compression; and step S6 involves outputting real-time fault warning level data through multilayer perceptron classifier and hierarchical threshold comparison.

[0295] like Figure 2 As shown, this embodiment provides an example of a fault early warning system for tobacco production equipment. In this embodiment, the fault early warning system is applied to the fault early warning method as described in the above embodiment.

[0296] Specifically, the fault early warning system includes a standard time-series data stream acquisition module 1, which is electrically or communicatively connected in sequence; a device operation fusion feature set acquisition module 2; a time-series correlation feature sequence acquisition module 3; a deep inference feature sequence acquisition module 4; a device status standard feature vector acquisition module 5; and a real-time fault early warning level data acquisition module 6.

[0297] The standard time-series data stream acquisition module 1 continuously collects multi-source time-series operation data streams during the operation of tobacco production equipment, and performs normalization, noise reduction, and dimensional unification processing on the multi-source time-series operation data streams to obtain a standard time-series data stream. The equipment operation fusion feature set acquisition module 2 inputs the standard time-series data stream into a graph attention heterogeneous fusion network, and performs cross-dimensional feature association and deep fusion on the standard time-series data stream through a multi-head attention mechanism to obtain an equipment operation fusion feature set. The time-series correlation feature sequence acquisition module 3 inputs the equipment operation fusion feature set into a time-series causal convolutional network, and captures time-series dependencies through causal convolution and dilated convolution structures to obtain a time-series correlation feature sequence. The module further deduces the feature sequence. The acquisition module 4 is used to perform hierarchical residual deduction on the time-series correlated feature sequence. Through multi-layer residual connections, the feature deepening mapping is completed layer by layer to obtain the deepened deduction feature sequence. The equipment state standard feature vector acquisition module 5 is used to perform global pooling and feature dimension compression on the deepened deduction feature sequence, remove redundant feature components and extract effective features to obtain the equipment state standard feature vector. The real-time fault warning level data acquisition module 6 is used to input the equipment state standard feature vector into the multi-layer perceptron classifier, output the equipment state probability distribution through fully connected mapping and Softmax normalization operation, and output the real-time fault warning level data of tobacco production equipment based on the comparison result of the equipment state probability distribution and the preset grading threshold.

[0298] like Figure 3 As shown, the electronic device 7 includes a processor 71 and a memory 72 coupled to the processor 71.

[0299] The memory 72 stores program instructions for implementing the fault warning method for tobacco production equipment in any of the above embodiments.

[0300] The processor 71 is used to execute program instructions stored in the memory 72 to perform fault warnings for tobacco production equipment.

[0301] The processor 71 can also be referred to as a CPU (Central Processing Unit). The processor 71 may be an integrated circuit chip with signal processing capabilities. The processor 71 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0302] Furthermore, Figure 4 This is a schematic diagram of the structure of a storage medium according to an embodiment of this application. See also: Figure 4The storage medium 8 in this embodiment stores program instructions 81 capable of implementing all the above methods. These program instructions 81 can be stored in the storage medium as a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods in each embodiment of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0303] In the several embodiments provided in this application, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the fault warning of a unit is only a logical function fault warning; in actual implementation, there may be other fault warning methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of the system or units may be electrical, mechanical, signal, or other forms.

[0304] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A fault early warning method for tobacco production equipment, characterized in that, The fault early warning method includes: Step S1: Continuously collect multi-source time-series operation data streams during the operation of tobacco production equipment, and perform normalization, noise reduction, and dimension unification processing on the multi-source time-series operation data streams to obtain a standard time-series data stream; Step S2: Input the standard time-series data stream into the graph attention heterogeneous fusion network, and perform cross-dimensional feature association and deep fusion on the standard time-series data stream through a multi-head attention mechanism to obtain the device operation fusion feature set; Step S3: Input the device running fused feature set into the temporal causal convolutional network, and capture the temporal dependency relationship through the causal convolution and dilated convolution structure to obtain the temporal associated feature sequence; Step S4: Perform hierarchical residual deduction on the time-series associated feature sequence, and complete the feature deepening mapping layer by layer through multi-layer residual connections to obtain the deepened deduction feature sequence; Step S5: Perform global pooling and feature dimension compression on the in-depth inference feature sequence, remove redundant feature components and extract effective features to obtain the standard feature vector of device status; Step S6: Input the standard feature vector of the equipment state into the multilayer perceptron classifier, output the probability distribution of the equipment state through fully connected mapping and Softmax normalization operation, and output the real-time fault warning level data of the tobacco production equipment based on the comparison result of the probability distribution of the equipment state and the preset grading threshold.

2. The fault early warning method according to claim 1, characterized in that, Step S1: Continuously collect multi-source time-series operational data streams during the operation of tobacco production equipment, and perform normalization, noise reduction, and dimensionality unification processing on the multi-source time-series operational data streams to obtain a standard time-series data stream, including: Step S1.1: The sensor group deployed at the key components of the tobacco production equipment continuously collects operating parameters such as temperature, vibration, current, pressure, and rotation speed to obtain a multi-source raw operating data stream; Step S1.2: Resample and align the multi-source original running data stream at a uniform sampling frequency, and map each source data to a global time-series coordinate based on the timestamp identifier to obtain a time-aligned data stream; Step S1.3: Perform missing value detection and forward filling on the time-aligned data stream, and identify and remove abnormal outliers to obtain a cleaned data stream; Step S1.4: Smooth the purified data stream using a moving average filter to obtain a noise-reduced data stream; Step S1.5: Perform Min-Max normalization operation on the noise-reduced data stream according to each dimension, and map the values ​​of each dimension to a preset interval to obtain a dimensionally unified data stream; Step S1.6: The unified data stream is spliced ​​by tensors according to the device node dimension and the time sequence dimension to obtain the standard time sequence data stream.

3. The fault early warning method according to claim 1, characterized in that, Step S2: Input the standard time-series data stream into a graph attention heterogeneous fusion network. Perform cross-dimensional feature association and deep fusion on the standard time-series data stream using a multi-head attention mechanism to obtain a device operation fusion feature set, including: Step S2.1: Divide the standard time-series data stream into several nodes according to the device component dimension, and use the data segment corresponding to each node as the node feature vector to construct the device topology graph structure; Step S2.2: Based on the physical connection relationship and signal transmission path between the device components corresponding to each node, define the heterogeneous edge type and edge weight between nodes in the device topology graph structure to obtain a weighted heterogeneous graph; Step S2.3: Perform linear mapping on the node feature vector of each node in the weighted heterogeneous graph to generate query vector, key vector, and value vector; Step S2.4: Perform a dot product operation on the query vector and the key vector and normalize it using Softmax to obtain the attention weight matrix between each node; Step S2.5: The attention weight matrix and the value vector are weighted and summed, and the multi-head attention outputs are spliced ​​and linearly projected to obtain node-level fusion features; Step S2.6: Perform a global pooling operation on the node-level fusion features to aggregate the feature information of all nodes and obtain the device operation fusion feature set.

4. The fault early warning method according to claim 1, characterized in that, Step S3: Input the device's fused feature set into a temporal causal convolutional network. Capture temporal dependencies through causal convolution and dilated convolution structures to obtain a temporal associated feature sequence, including: Step S3.1: The device operation fusion feature set is serialized and expanded according to time steps to obtain the fusion feature time sequence; Step S3.2: Apply causal convolution operation to the fused feature time series sequence, wherein each convolution kernel performs convolution operation only on the data of the current time step and the historical time steps to obtain the causal feature sequence; Step S3.3: Apply dilated convolution operation to the causal feature sequence, and expand the sampling interval of the convolution kernel according to a preset dilation factor to expand the temporal receptive field without increasing the number of parameters, so as to obtain the dilated feature sequence. Step S3.4: Perform batch normalization and nonlinear activation operations on the dilated feature sequence sequentially to obtain the activated feature sequence; Step S3.5: The activation feature sequence is used as the input of the next layer of causal dilated convolution for multi-layer stacking iteration to expand the temporal receptive field and deepen the temporal dependency modeling layer by layer, so as to obtain the temporal correlation feature sequence.

5. The fault early warning method according to claim 1, characterized in that, Step S4: Perform hierarchical residual deduction on the temporal correlation feature sequence, and complete the feature deepening mapping layer by layer through multi-layer residual connections to obtain the deepened deduction feature sequence, including: Step S4.1: Input the time-series correlation feature sequence into the first-layer inference unit, perform linear mapping through a fully connected layer and nonlinear activation operation to obtain the first-layer inference features; Step S4.2: Add the residuals of the first layer of inference features and the time-series associated feature sequence to obtain the first layer of residual inference features; Step S4.3: Input the first layer residual inference features into the second layer inference unit, perform linear mapping through a fully connected layer and nonlinear activation operation to obtain the second layer inference features; Step S4.4: Add the residuals of the second-layer derivation features and the first-layer residual derivation features to obtain the second-layer residual derivation features; Step S4.5: Input the second-layer residual extrapolation features into the subsequent extrapolation unit, and perform linear mapping, nonlinear activation and residual addition operations layer by layer according to the extrapolation logic of the first-layer extrapolation unit and the second-layer extrapolation unit to obtain the in-depth extrapolation feature sequence.

6. The fault early warning method according to claim 1, characterized in that, Step S5: Perform global pooling and feature dimension compression on the deepened inference feature sequence, remove redundant feature components and extract effective features to obtain the device state standard feature vector, including: Step S5.1: Perform global average pooling operation on the in-depth inference feature sequence along the time dimension to aggregate each feature channel into a single feature vector in the time dimension, and obtain the time-series aggregated feature vector; Step S5.2: Calculate the feature variance of each dimension in the time-series aggregated feature vector, and sort each dimension according to the size of the variance to obtain the feature importance ranking result; Step S5.3: Based on the feature importance ranking result, feature dimensions with variances below a preset threshold are identified as redundant feature components and removed to obtain an effective feature subset. Step S5.4: The effective feature subset is linearly reduced and compressed to a preset target dimension through a fully connected layer to obtain the standard feature vector of the device state.

7. The fault early warning method according to claim 1, characterized in that, Step S6: Input the standard feature vector of the equipment state into a multilayer perceptron classifier, output the equipment state probability distribution through fully connected mapping and Softmax normalization, and output real-time fault warning level data of the tobacco production equipment based on the comparison result of the equipment state probability distribution and the preset grading threshold, including: Step S6.1: Input the standard feature vector of the device state into the multilayer perceptron classifier, and perform feature transformation layer by layer through several fully connected layers and nonlinear activation functions to obtain the category space mapping vector; Step S6.2: Apply the Softmax function to the category space mapping vector to perform normalization operation, convert each component into a probability value, and obtain the device state probability distribution; Step S6.3: Extract the maximum probability value and its corresponding state category from the device state probability distribution to obtain the current device state determination result; Step S6.4: Compare the maximum probability value with a preset first grading threshold and a second grading threshold, and determine the fault warning level based on the comparison result, wherein the first grading threshold is lower than the second grading threshold. Step S6.5: Generate early warning instruction data based on the fault early warning level and the current equipment status determination result, and output the early warning instruction data as the real-time fault early warning level data of the tobacco production equipment.

8. A fault early warning system for tobacco production equipment, wherein the fault early warning system is applied to the fault early warning method as described in any one of claims 1 to 7, characterized in that, The fault early warning system includes: The standard time-series data stream acquisition module is used to continuously collect multi-source time-series operation data streams during the operation of tobacco production equipment, and to perform normalization, noise reduction and dimension unification processing on the multi-source time-series operation data streams to obtain a standard time-series data stream. The device operation fusion feature set acquisition module is used to input the standard time series data stream into the graph attention heterogeneous fusion network, and perform cross-dimensional feature association and deep fusion on the standard time series data stream through a multi-head attention mechanism to obtain the device operation fusion feature set. The temporal correlation feature sequence acquisition module is used to input the device running fusion feature set into the temporal causal convolutional network, and capture temporal dependencies through causal convolution and dilated convolution structures to obtain the temporal correlation feature sequence; The in-depth inference feature sequence acquisition module is used to perform hierarchical residual inference on the time-series associated feature sequence, and complete the feature in-depth mapping layer by layer through multi-layer residual connection to obtain the in-depth inference feature sequence; The device status standard feature vector acquisition module is used to perform global pooling and feature dimension compression on the in-depth inference feature sequence, remove redundant feature components and extract effective features to obtain the device status standard feature vector. The real-time fault warning level data acquisition module is used to input the standard feature vector of the equipment status into the multilayer perceptron classifier, output the probability distribution of the equipment status through fully connected mapping and Softmax normalization operation, and output the real-time fault warning level data of the tobacco production equipment based on the comparison result of the probability distribution of the equipment status and the preset grading threshold.

9. An electronic device, characterized in that, The method includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the fault warning method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that, when executed by a processor, enable the fault warning method as described in any one of claims 1 to 7.

Images