An online monitoring system for power equipment failure in polyester staple production

Abstract

The present application relates to the technical field of industrial equipment monitoring, in particular to a power equipment fault online monitoring system for polyester staple fiber production, comprising feature extraction, dynamic analysis, evolution prediction, anomaly identification, feature construction and diagnosis matching modules. The system collects power equipment operation data in real time, extracts time sequence features and calculates dynamic offset vectors, then predicts state evolution trajectory and identifies preliminary anomalies. The feature construction module will backtrack to extract feature segments associated with anomalies, obtain fault component signals with different physical meanings through multi-level signal deconstruction, and input into a fault feature atlas generator to construct a three-dimensional fault feature atlas reflecting fault evolution rules. The diagnosis matching module determines the specific fault type and its evolution stage by pattern matching the three-dimensional atlas with the preset fault knowledge base. The present application realizes early detection of fault germination, accurate identification of fault type and dynamic judgment of fault evolution stage.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment monitoring technology, and in particular to an online fault monitoring system for power equipment used in polyester staple fiber production. Background Technology

[0002] Current online monitoring of industrial power equipment generally employs single-point alarms based on fixed thresholds or anomaly detection methods based on statistical models. These technologies typically set static safety operating boundaries for single-dimensional or multi-dimensional time-series data such as vibration, temperature, and current; an alarm is triggered once the monitored data exceeds these boundaries. Furthermore, some solutions use machine learning models to classify extracted features to determine whether the equipment is in a "normal" or "faulty" state. The core drawback of these existing technologies is that their analysis remains at the level of identifying anomalies in the data, the alarm mechanism is relatively lagging, and the generated alarm information is coarse-grained. When an anomaly is detected, the equipment has often already entered a clear fault state, making it impossible to effectively identify early, subtle faults. At the same time, a single "fault" label cannot inform maintenance personnel of the specific type of fault, the root cause component, or how the fault is developing, resulting in insufficient timeliness of early warnings and inadequate targeted maintenance guidance.

[0003] Conventional techniques lack the ability to deeply analyze and characterize the physical processes and dynamic evolution of faults. Power equipment faults are dynamic processes involving the interweaving and changing of multiple physical quantities. Existing methods often extract features as statistical summaries or frequency domain energy of signals, failing to effectively separate and highlight the characteristic components corresponding to different fault mechanisms. This results in low distinguishability between different fault modes, easily leading to false alarms and missed alarms. More importantly, they cannot depict the complete evolution trajectory of a fault from its inception to its development and deterioration, leaving predictive maintenance without precise basis for judging the fault's development stage and remaining service life. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and propose an online monitoring system for power equipment faults in polyester staple fiber production.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an online monitoring system for power equipment faults in polyester staple fiber production, comprising:

[0006] The feature extraction module collects real-time operating status data of electrical equipment in the polyester staple fiber production line and performs multi-dimensional time series feature extraction to generate a time series feature sequence characterizing the operating status of the electrical equipment.

[0007] The dynamic analysis module calculates the dynamic offset vector between the operating state of the power equipment and the standard operating state based on the time-series feature sequence.

[0008] The evolution prediction module uses the dynamic offset vector to drive the fault mode evolution model and predict the future operating state evolution trajectory of the power equipment.

[0009] The anomaly identification module generates multidimensional anomaly signal thresholds covering different fault types based on the evolution trajectory of the operating state, compares the time-series feature sequence with the multidimensional anomaly signal thresholds in real time, and identifies and locates preliminary fault anomaly signals.

[0010] The feature construction module traces back to the time-series feature sequence based on the initial fault anomaly signal, extracts complete feature segments associated with the initial fault anomaly signal, performs multi-level signal deconstruction, decomposes the fault component signals under different physical meanings, and inputs the fault component signals to the fault feature map generator to construct a three-dimensional fault feature map that reflects the fault evolution law.

[0011] The diagnostic matching module performs pattern matching between the three-dimensional fault feature map and a preset fault knowledge base to determine the specific fault type and its evolution stage of the power equipment.

[0012] As a further aspect of the present invention, the real-time acquisition of operating status data of electrical equipment in the polyester staple fiber production line specifically includes:

[0013] Deploy sensor arrays at the power input terminals, critical load nodes, and control loops of power equipment;

[0014] The sensor array synchronously collects the instantaneous values ​​of three-phase voltage, three-phase current, power factor, and temperature of key nodes of the power equipment.

[0015] The instantaneous values ​​of three-phase voltage, three-phase current, power factor, and temperature are collected and then aligned with timestamps and their sampling rates are standardized.

[0016] The data, after being timestamped and sampled to a uniform rate, is integrated to form a raw running status data stream indexed by time.

[0017] As a further aspect of the present invention, the step of performing multi-dimensional time series feature extraction to generate a time series feature sequence characterizing the operating state of power equipment specifically includes:

[0018] Voltage data segments, current data segments, power factor data segments, and temperature data segments are extracted from the original operating status data stream within a fixed time window.

[0019] Spectral analysis is performed on the voltage data segment to extract the fundamental amplitude, main harmonic distortion rate, and voltage imbalance characteristics.

[0020] The current data segment is subjected to joint time-domain and frequency-domain analysis to extract the characteristics of the effective current value change rate, peak factor and total harmonic distortion rate of the current.

[0021] Statistical analysis was performed on the power factor data segment to extract the power factor mean, fluctuation variance, and number of abrupt changes.

[0022] Perform trend analysis on the temperature data segment to extract the temperature rise slope, local extreme points, and temperature rise rate characteristics;

[0023] All extracted features are arranged in order of time window and combined to form the time-series feature sequence.

[0024] As a further aspect of the present invention, the step of calculating the dynamic offset vector between the operating state and the standard operating state of the power equipment based on the time-series feature sequence specifically includes:

[0025] Retrieve the standard characteristic sequence of power equipment under standard healthy operating conditions, which is pre-stored in the database;

[0026] The current feature vector in the time-series feature sequence is compared element-by-element with the standard feature vector at the corresponding time point in the standard feature sequence;

[0027] Calculate the difference between the current feature vector and the standard feature vector in each feature dimension;

[0028] A multidimensional vector is constructed using the differences across all feature dimensions as components; this multidimensional vector is the dynamic offset vector.

[0029] The magnitude of the dynamic offset vector reflects the overall degree of offset, and the direction reflects which type or types of operational features have experienced abnormal offset.

[0030] As a further aspect of the present invention, the step of using the dynamic offset vector to drive the fault mode evolution model to predict the future operating state evolution trajectory of power equipment specifically includes:

[0031] Input the current dynamic offset vector into the pre-trained fault mode evolution model;

[0032] The fault mode evolution model learns the evolution rules from historical fault data and performs multi-step iterative deduction with the current dynamic offset vector as the initial state.

[0033] Each iteration predicts the dynamic offset vector after the next time step.

[0034] The dynamic offset vectors predicted by multiple consecutive time steps are connected in chronological order to form the trajectory of the future operating state evolution of the power equipment.

[0035] The operational status evolution trajectory describes the path and extent to which the operational status of power equipment deviates from the standard state over a future period of time.

[0036] As a further aspect of the present invention, the step of generating multidimensional abnormal signal thresholds covering different fault types based on the operational state evolution trajectory specifically includes:

[0037] Extract the maximum value of the predicted future dynamic offset vector in each feature dimension from the evolution trajectory of the operating state;

[0038] For each known fault type, the typical offset range caused by the fault type in each feature dimension is determined based on its historical data;

[0039] By combining the predicted maximum value of the future dynamic offset vector and the typical offset range of different fault types, a comprehensive warning upper limit threshold is calculated for each feature dimension;

[0040] Based on the differences in the sensitivity of different feature dimensions to various types of faults, different weighting coefficients are assigned to the warning upper limit threshold.

[0041] Finally, a multidimensional abnormal signal threshold set is generated, consisting of multiple weighted feature dimension thresholds.

[0042] As a further aspect of the present invention, the step of comparing the time-series feature sequence with the multidimensional abnormal signal threshold in real time to identify and locate the preliminary fault abnormal signal specifically includes:

[0043] Obtain the feature vector of the latest time window from the time-series feature sequence;

[0044] Each feature value in the feature vector of the latest time window is compared with the warning upper limit threshold of the corresponding feature dimension in the multidimensional abnormal signal threshold set;

[0045] If a certain feature value exceeds its corresponding warning upper limit threshold, the initial anomaly labeling of the feature dimension is triggered.

[0046] Record all feature dimensions that trigger the initial anomaly marker, along with the specific values ​​and time points when they exceed the limits;

[0047] The initial abnormality markers and their related information that are triggered simultaneously are packaged into a preliminary fault abnormality signal.

[0048] As a further aspect of the present invention, the step of tracing back from the preliminary fault anomaly signal to the time-series feature sequence and extracting a complete feature fragment associated with the preliminary fault anomaly signal specifically includes:

[0049] Based on the time point information contained in the preliminary fault anomaly signal, trace back a set length of historical time.

[0050] Extract all feature data from the time-series feature sequence from the backtracking start point to the current time point;

[0051] Using the feature dimension that triggers the initial anomaly marker as the core, extract the complete numerical sequence of all relevant dimensions from the historical feature data;

[0052] The extracted complete numerical sequences of each dimension are time-aligned and standardized.

[0053] The processed numerical sequences of each dimension are integrated to form the complete feature fragment that reflects the entire process of an anomaly from its latent state to its manifestation.

[0054] As a further aspect of the present invention, the step of performing multi-level signal decomposition to obtain fault component signals under different physical meanings specifically includes:

[0055] The complete feature segment is considered as a composite multidimensional time series signal;

[0056] By applying the empirical mode decomposition method, the time series signal of each dimension is adaptively decomposed into a series of intrinsic mode function components and a residual component;

[0057] The Hilbert transform is performed on the intrinsic mode function components obtained from the decomposition to obtain their time-frequency distribution characteristics;

[0058] Based on the similarity of time-frequency distribution characteristics, the intrinsic mode function components from different original feature dimensions are recombined to form the fault component signals that respectively characterize trends, periodic fluctuations, shock events, and random noise.

[0059] Each of the fault component signals carries information about a certain aspect of the physical process in the original composite signal.

[0060] As a further aspect of the present invention, the step of inputting the fault component signal to a fault feature map generator to construct a three-dimensional fault feature map reflecting the fault evolution law specifically includes:

[0061] Time is used as the first dimension, the type of fault component signal is used as the second dimension, and the amplitude, energy, or frequency characteristics of the fault component signal are used as the third dimension.

[0062] The characteristic value calculated for each fault component signal at its corresponding time point and signal type is taken as a data point in a three-dimensional space.

[0063] Arrange and connect the data points calculated from all fault component signals in three-dimensional space according to time sequence and signal type relationship;

[0064] By using a spatial interpolation algorithm, discrete data points are fitted into a continuous three-dimensional surface or voxel model, which is the three-dimensional fault feature map.

[0065] The three-dimensional fault feature map intuitively displays the interrelationships and structural features of different fault components over time, and is used to match it with the fault pattern map in the preset fault knowledge base, thereby determining the specific fault type and its evolution stage.

[0066] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0067] By performing multi-level signal deconstruction on complete feature segments associated with preliminary abnormal signals, fault component signals representing different physical meanings are obtained, and a three-dimensional fault feature map is constructed using a fault feature map generator. This technical solution deconstructs and reconstructs one-dimensional, mixed time-series signals into a dynamic map with clear visualization of three dimensions: time, fault component type, and signal intensity. This allows the feature patterns hidden in complex signals and related to specific physical failure processes to be separated and highlighted. This can intuitively reveal the intrinsic composition of the fault and the cooperative or competitive evolution relationship of each component over time, providing an information depth far exceeding traditional time-domain or frequency-domain analysis for understanding the essence of the fault, and realizing a three-dimensional and structured representation of fault characteristics.

[0068] By pre-setting a fault knowledge base containing 3D atlas templates of various typical fault modes at different evolution stages, and performing high-dimensional spatial pattern matching between the real-time generated 3D fault feature atlas and this knowledge base, this technical solution upgrades the diagnostic process from traditional feature vector classification to the comparison and identification of the "3D fingerprint" of fault evolution. This enables simultaneous and accurate classification of fault types and precise judgment of the current evolution stage. The matching process not only identifies the type of fault but also determines whether the fault is in an early, weak stage or a late, severe stage, thus directly supporting the leap from simple alarms to precise predictive maintenance decisions, achieving refinement of fault diagnosis in both the type and temporal dimensions. Attached Figure Description

[0069] Figure 1 This is a timing diagram of the online fault monitoring system for power equipment used in polyester staple fiber production according to the present invention;

[0070] Figure 2 Flowchart for calculating the dynamic offset vector;

[0071] Figure 3 A graph showing the monitoring and analysis of total harmonic distortion of current in polyester staple fiber power equipment;

[0072] Figure 4A multi-dimensional operational status monitoring diagram for polyester staple fiber power equipment;

[0073] Figure 5 This is a three-dimensional fault feature map. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0075] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0076] See Figure 1 An online monitoring system for power equipment faults in polyester staple fiber production is disclosed. The system includes a feature extraction module, a dynamic analysis module, an evolution prediction module, an anomaly identification module, a feature construction module, and a diagnostic matching module. The feature extraction module collects real-time operating status data of the power equipment in the polyester staple fiber production line and performs multi-dimensional time-series feature extraction to generate a time-series feature sequence characterizing the operating status of the power equipment. The dynamic analysis module calculates the dynamic offset vector between the operating status of the power equipment and the standard operating status based on the time-series feature sequence. The evolution prediction module uses the dynamic offset vector to drive a fault mode evolution model to predict the future operating status evolution trajectory of the power equipment. The anomaly identification module generates multi-dimensional anomaly signal thresholds covering different fault types based on the operating status evolution trajectory, and compares the time-series feature sequence with the multi-dimensional anomaly signal thresholds in real time to identify and locate preliminary fault anomaly signals. The feature construction module traces back to the time-series feature sequence based on the initial fault anomaly signal, extracts complete feature fragments associated with the initial fault anomaly signal, and performs multi-level signal deconstruction to obtain fault component signals under different physical meanings. These fault component signals are then input into the fault feature map generator to construct a three-dimensional fault feature map reflecting the fault evolution pattern. The diagnosis matching module performs pattern matching between the three-dimensional fault feature map and a preset fault knowledge base to determine the specific fault type and evolution stage of the power equipment.

[0077] In one embodiment of the present invention, in a specific implementation, a sensor array is deployed at the power input terminal, key load nodes, and control loop of the power equipment. The sensor array synchronously collects the instantaneous values ​​of the three-phase voltage, three-phase current, power factor, and temperature of the key nodes of the power equipment. The collected instantaneous values ​​of the three-phase voltage, three-phase current, power factor, and temperature are timestamped and their sampling rates are standardized. The data that has undergone timestamping and sampling rate standardization are integrated to form a raw operating status data stream indexed by time. In some embodiments, voltage, current, power factor, and temperature data segments within fixed time windows are extracted from the raw operating status data stream. Spectral analysis is performed on the voltage data segments to extract the fundamental amplitude, main harmonic distortion rate, and voltage imbalance characteristics. Joint time-domain and frequency-domain analysis is performed on the current data segments to extract the effective current value change rate, peak factor, and total harmonic distortion rate characteristics. In a specific implementation, the effective current value change rate is calculated using the formula:

[0078]

[0079] in: Indicates time window The rate of change of the effective value of the current, Indicates time window The effective value of the current, Indicates the previous time window The effective value of the current, This represents the time interval between consecutive time windows. Optionally, statistical analysis is performed on the power factor data segment to extract the power factor mean, fluctuation variance, and number of abrupt changes; trend analysis is performed on the temperature data segment to extract the temperature rise slope, local extrema, and temperature rise rate. It can be understood that the extracted fundamental amplitude, major harmonic distortion rate, voltage imbalance characteristics, current RMS change rate, peak factor, total harmonic distortion rate, power factor mean, fluctuation variance, number of abrupt changes, temperature rise slope, local extrema, and temperature rise rate characteristics are arranged and combined according to the time window sequence to form a time-series feature sequence. In some embodiments, the length of the fixed time window is set according to the operating cycle of the electrical equipment in the polyester staple fiber production line, and timestamp alignment uses a Global Positioning System clock signal to ensure the time synchronization of the sensor array data. Optionally, sampling rate unification processing adjusts the sampling frequency of all sensor data to be consistent through an interpolation algorithm, and spectrum analysis uses the Fast Fourier Transform method to calculate the fundamental amplitude, major harmonic distortion rate, and voltage imbalance from the voltage data segment. In practice, joint time-domain and frequency-domain analysis simultaneously calculates time-domain statistics such as the peak factor and frequency-domain transformations such as the total harmonic distortion rate of the current data segment. Trend analysis uses linear regression to fit the temperature rise slope and identify local extrema for the temperature data segment. It can be understood that the generation process of the time-series feature sequence is continuous, with feature extraction and combination operations automatically triggered at the end of each fixed time window.

[0080] In the implementation of the online monitoring system for power equipment faults in polyester staple fiber production, the timestamp alignment relies on a hierarchical precision time synchronization architecture. This architecture is designed for industrial environments where local network latency and equipment clock drift may exist in the production workshop. The system deploys a highly stable hardware clock server as the master clock source within the workshop's local area network. This master clock source obtains absolute standard time by receiving timing signals from the BeiDou Navigation Satellite System or the Global Positioning System and continuously sends synchronization clock messages to all data acquisition terminals within the network using a precision time protocol. Each sensor data acquisition terminal deployed at the power input end of the power equipment, key load nodes, and control loops integrates a network interface chip supporting the precision time protocol and a local clock circuit. While continuously acquiring data such as the instantaneous values ​​of three-phase voltage and three-phase current, the data acquisition terminal accurately records the local clock timestamp corresponding to the instant each data sample is converted by the analog-to-digital converter.

[0081] Before sending the packaged sensor data to the feature extraction module, the data acquisition terminal calculates the phase offset and transmission path delay between its local clock and the master clock source based on the latest received precision time protocol synchronization message. Based on this calculation, it dynamically compensates and corrects the original local timestamps recorded in the data packets, ultimately assigning each data point a unified time stamp synchronized with the master clock source. Upon receiving data packets from different sensor arrays, the feature extraction module arranges all instantaneous voltage, current, power factor, and temperature values ​​along the same absolute time axis based on these corrected high-precision time stamps. For individual data packet out-of-order or time stamp anomalies that may be caused by minute network jitter, the system uses a time-series-based sliding window verification algorithm for identification and correction. This ensures that the multi-dimensional data corresponding to each index moment in the final integrated original operating status data stream strictly originates from the same instant in the physical world.

[0082] Sampling rate unification is a crucial step in fusing multi-source heterogeneous sensor data after timestamp alignment. Due to differences in hardware performance and design purpose among sensors deployed at different monitoring locations, their inherent data output sampling rates may differ. For example, a current sensor used to capture rapid transient processes might have a sampling rate set to 10,000 times per second, while a temperature sensor used to monitor slow thermal processes might only have a sampling rate of once per second. During sampling rate unification, the system pre-sets a system-level target sampling frequency based on the specific requirements of the signal frequency band for fault diagnosis of electrical equipment in the polyester staple fiber production line. This target frequency is typically no less than twice the highest inherent sampling frequency among all sensors to satisfy the Nyquist sampling theorem and preserve necessary signal details. For sensor data with inherent sampling rates higher than the target sampling rate, the system uses an anti-aliasing finite-length unit impulse response digital filter for downsampling. The cutoff frequency of this filter is carefully designed to effectively suppress high-frequency noise and aliasing components while preserving key frequency band information of the equipment status.

[0083] For sensor data with an inherent sampling rate lower than the target sampling rate, the system employs an interpolation algorithm based on signal characteristics and contextual constraints for upsampling. For example, for relatively slow-changing signals like temperature, a spline interpolation algorithm that preserves the local waveform shape can be used; for current and voltage signals where phase continuity is important, a sinusoidal interpolation method based on their historical spectral characteristics may be used. After adjusting the sampling rate of all data channels to the target frequency, the system re-checks the alignment of the data in each channel on a unified time axis and compensates for any small group delays that may be introduced by interpolation or filtering. Finally, it generates a well-organized multivariate time series dataset with identical and equally spaced time indices across all data dimensions. This dataset serves as the rigorously preprocessed foundational data source upon which subsequent feature extraction depends.

[0084] In one embodiment of the present invention, see [reference] Figure 2 In practical implementation, the system calls upon the standard feature sequence of power equipment under standard healthy operating conditions, which is pre-stored in the database. It then compares the current feature vector in the time-series feature sequence with the standard feature vector at the corresponding time point in the standard feature sequence element-by-element, calculating the difference between the current and standard feature vectors in each feature dimension. Using the differences across all feature dimensions as components, a multi-dimensional vector is constructed; this multi-dimensional vector is the dynamic offset vector. In some embodiments, the magnitude of the dynamic offset vector reflects the overall degree of offset, while the direction of the dynamic offset vector reflects which specific type or categories of operating features have experienced abnormal offset. It can be understood that the magnitude of the dynamic offset vector... Calculation using the formula:

[0085]

[0086] in: This represents the magnitude of the dynamic offset vector. This represents the total number of feature dimensions. Indicates the first The difference across each feature dimension, the difference It is the current feature vector of the th The eigenvalue and the standard eigenvector The arithmetic difference of the eigenvalues.

[0087] Optionally, the current dynamic offset vector is input into a pre-trained fault mode evolution model (FMODE). The FMODE learns evolution patterns from historical fault data and performs multi-step iterative deduction with the current dynamic offset vector as the initial state. In specific implementations, each iteration predicts the dynamic offset vector one time step later. The predicted dynamic offset vectors from multiple consecutive time steps are connected chronologically to form the future operating state evolution trajectory of the power equipment. This trajectory describes the path and extent to which the operating state of the power equipment deviates from the standard state over a future period. In some embodiments, the FMODE uses a long short-term memory neural network structure. The training data for the FMODE comes from the time-series feature sequences recorded during the historical operation of the power equipment in the polyester staple fiber production line, along with the finally confirmed fault types and evolution stage labels. It can be understood that the multi-step iterative deduction process is recursively performed within the FMODE. The FMODE outputs the predicted dynamic offset vector state for the next time step based on the input current dynamic offset vector state, and uses this predicted output as the input state for the next recursive operation. In practice, the length of the time step is consistent with the length of the fixed time window used when generating the time series feature sequence, and the number of time steps for continuous prediction is preset according to actual monitoring needs. Optionally, the evolution trajectory of the running state is represented in the data structure as a sequence of dynamic offset vectors, where each dynamic offset vector corresponds to a future prediction time point.

[0088] The Fault Mode Evolution (FMODE) model is constructed using a Long Short-Term Memory (LSTM) neural network. This network structure is specifically designed for the temporal dependencies of sequential data such as dynamic offset vectors. The input layer dimension of the model is identical to the feature dimension of the dynamic offset vector, receiving individual dynamic offset vectors input sequentially over time. The LSM neural network contains multiple cascaded memory units. Each memory unit contains a cell state, a forget gate, an input gate, and an output gate. The cell state serves as the information transmission path throughout the time series, maintaining long-term fault evolution state memory. The forget gate, through a Sigmoid activation function layer, determines which old and irrelevant fault offset information to discard from the previous cell state. The input gate, through a combination of Sigmoid and tanh layers, determines which new dynamic offset vector information to store in the cell state. The output gate calculates and outputs the current hidden state based on the current cell state and the current input dynamic offset vector. This hidden state is passed to the memory unit at the next time step and mapped to the predicted dynamic offset vector for the next time step through a fully connected output layer.

[0089] The training process of the model uses a large number of time-series feature sequences recorded in historical fault data and their corresponding fault evolution stage labels labeled by experts. The continuous time-series feature sequences are divided into a large number of overlapping dynamic offset vector subsequences through a time sliding window. Each subsequence is used as a training sample, and the corresponding real dynamic offset vector of the next time step is used as a supervision label. The model minimizes the mean square error loss function between the predicted dynamic offset vector and the real dynamic offset vector through the backpropagation algorithm and optimizer, thereby learning the dynamic evolution law of the power equipment of the polyester staple fiber production line from the healthy state to different fault modes.

[0090] The multi-step iterative deduction process is executed recursively within the trained Fault Mode Evolution (FEMA) model. When the system inputs the current dynamic offset vector into the model, the model performs forward propagation calculations internally, outputting a single-step prediction of the dynamic offset vector for the next time step. For multi-step prediction, the system requires a recursive prediction controller to manage the iteration process. This controller takes the model's current single-step prediction output, passes it through a state transition interface, and reconstructs it into the next "current" dynamic offset vector that conforms to the model's input format. This reconstructed vector is then used as the input state for the next forward calculation. This process continues recursively; each recursive call to the model further predicts the dynamic offset vector for the next time step based on the prediction state output from the previous recursion. To ensure the long-term stability of multi-step predictions, the recursive prediction controller implements a strategy that dynamically switches between "teacherforcing" and free-running mode. In the initial few steps of the prediction, the controller can partially use historical real dynamic offset vector information to correct the recursive input, thus mitigating the accumulation of prediction errors. As the number of prediction steps increases, the system relies entirely on the model's own prediction output from the previous time step as the input for the next time step, i.e., it enters free-running mode. Finally, by arranging and connecting the dynamic offset vectors output from each recursive prediction according to their corresponding future time points, a complete trajectory describing the future deviation path of the equipment's operating state is formed. Each point on this trajectory represents the model's vectorized prediction of the degree of deviation of the equipment's state at a specific future time.

[0091] In one embodiment of the present invention, in a specific implementation, the maximum value of the predicted future dynamic offset vector in each feature dimension is extracted from the operational state evolution trajectory. For each known fault type, the typical offset range caused by that fault type in each feature dimension is determined based on its historical data. A comprehensive early warning upper limit threshold is calculated for each feature dimension by combining the maximum value of the predicted future dynamic offset vector and the typical offset range of different fault types. In some embodiments, the first... Warning upper limit threshold for each feature dimension Calculation using the formula:

[0092]

[0093] in: Indicates the first The upper limit threshold for early warning in each feature dimension. The first one extracted from the trajectory of the evolution of the running state The maximum value of the predicted future dynamic offset vector for each feature dimension. Indicates the first The first type of fault is in the The upper limit of the historical typical offset range in each feature dimension. This represents the total number of known fault types. and For adjustment coefficients, Indicates assignment to the first The weighting coefficients of the warning upper limit thresholds for each feature dimension are determined. Different weighting coefficients are assigned to the warning upper limit thresholds based on the differences in sensitivity of different feature dimensions to various types of faults, ultimately generating a multidimensional abnormal signal threshold set composed of multiple weighted feature dimension thresholds.

[0094] The latest time window's feature vector is obtained from the time-series feature sequence. Each feature value in the latest time window's feature vector is compared with the warning upper limit threshold of the corresponding feature dimension in the multi-dimensional anomaly signal threshold set. In specific implementations, if a feature value exceeds its corresponding warning upper limit threshold, a preliminary anomaly marker for that feature dimension is triggered. All feature dimensions that trigger preliminary anomaly markers, along with their specific values ​​and timestamps of exceeding the limit, are recorded. It can be understood that a set of simultaneously triggered preliminary anomaly markers and their related information are packaged into a preliminary fault anomaly signal. In some embodiments, the comparison operation is performed in real-time and cyclically; a new round of comparison is initiated immediately after the time-series feature sequence updates the feature vector of a latest time window. Optionally, the recorded information for the preliminary anomaly marker includes the feature dimension number, feature dimension name, specific value exceeding the threshold, precise timestamp of the exceeding limit, and the identifier of the latest time window to which it belongs. In specific implementations, the packaging operation is performed based on the time window identifier and fault correlation rules. The fault correlation rules define which feature dimensions' anomaly markers tend to appear simultaneously to characterize the same potential initial fault mode. It can be understood that the preliminary fault anomaly signal is passed as a structured data object to subsequent feature construction modules for processing.

[0095] Different weighting coefficients are assigned to the warning upper limit threshold based on the varying sensitivity of different feature dimensions across various fault types. This assignment process relies on a quantitative evaluation mechanism that performs offline calculations and learning based on historical fault data. The system needs to establish a comprehensive historical fault case library, where each case includes time-series data of relevant feature dimensions before and after the fault occurrence, as well as the final confirmed fault type label. For each known fault type, the system extracts the deviation of each feature dimension value relative to its health baseline state during the period when the fault characteristics are clearly manifested in all relevant cases. The system evaluates the sensitivity and stability of each feature dimension in identifying this type of fault by calculating the statistical distribution characteristics of the deviation amplitude under the same fault type, including its mean, standard deviation, and covariance relationship with other dimensions. A dimension that consistently produces a significant and stable deviation when a certain fault occurs has a high sensitivity score.

[0096] The specific calculation of weighting coefficients employs objective weighting methods. For example, the entropy method is used, where the system calculates the information entropy of each feature dimension under each fault type. The smaller the entropy value, the lower the disorder of the data in that dimension under that type of fault, the greater the amount of information it provides, and therefore its weight should be higher. Another method is the coefficient of variation method, where the system calculates the coefficient of variation of the offset of each feature dimension in different fault cases. The larger the coefficient of variation, the stronger the dimension's ability to distinguish faults. Finally, the system integrates the performance of each feature dimension under all known fault types and determines a global comprehensive weighting coefficient for each feature dimension through methods such as weighted averaging or principal component analysis. This coefficient is directly multiplied by the aforementioned calculated warning upper limit threshold, thus giving the feature dimensions that are more sensitive to faults a stricter, weighted over-limit judgment boundary.

[0097] A set of simultaneously triggered preliminary anomaly markers and their related information are packaged into a preliminary fault anomaly signal. This packaging operation is based on time window identifiers and fault correlation rules. The fault correlation rules define which feature dimensions of anomaly markers tend to appear simultaneously to characterize the same potential early fault mode. The establishment of fault correlation rules is derived from the analysis of the joint distribution of multi-dimensional features under historical normal and fault conditions. The system automatically discovers frequently co-occurring anomaly combinations of feature dimensions from a large amount of historical data through data mining techniques, such as association rule learning or cluster analysis. For example, the system may analyze historical data using the Apriori algorithm and find that the two anomaly markers "exceeding the limit for total harmonic distortion of current" and "exceeding the limit for temperature rise slope" have a high probability of being triggered simultaneously in the same or adjacent time windows before the occurrence of early bearing wear failure. This frequent itemset will then be extracted into a fault correlation rule. During real-time monitoring, when the anomaly identification module detects multiple feature dimensions triggering preliminary anomaly markers within a time window, it submits these markers and their detailed information to a rule matching engine. This engine compares the currently triggered anomaly marker combination with a pre-defined fault correlation rule base. If the current combination matches or is highly similar to a dimension combination defined in a rule, the rule matching engine categorizes and binds these discrete anomaly markers according to the potential fault mode indicated by that rule, attaching information such as the rule ID and matching confidence level, and encapsulating them into a structured preliminary fault anomaly signal data packet. This packaging mechanism ensures that scattered anomaly indicators can be aggregated into fault warning events with clear physical orientation, providing a clear starting point and context for subsequent feature backtracking and in-depth analysis.

[0098] See Figure 3 This is a monitoring and analysis chart of the total harmonic distortion (THD) of current in the electrical equipment of a polyester staple fiber production line. The chart shows the trend of THD over 60 time windows and compares it with a preset weighted warning threshold, visually reflecting the evolution of the equipment from normal operation to severe failure. In the 31st time window, the THD first exceeded the 5.0% threshold, triggering an initial anomaly marker. From the 50th window onwards, the distortion rate rose sharply, reaching a maximum of 17.0%, far exceeding the warning threshold, indicating that the failure had entered an irreversible and severe stage. During the development and severe stages of the failure, the curve exhibited dramatic fluctuations, reflecting the extreme instability of the equipment's operating state. By analyzing the slope and amplitude of the distortion rate, the speed of failure development can be predicted, allowing for advance maintenance planning.

[0099] In one embodiment of the present invention, in a specific implementation, based on the time point information contained in the initial fault anomaly signal, a historical time of a set length is traced back. All feature data from the tracing start point to the current time point is extracted from the time-series feature sequence. Using the feature dimension that triggered the initial anomaly marker as the core, complete numerical sequences of all relevant dimensions in the historical feature data are extracted. The extracted complete numerical sequences of each dimension are then time-aligned and standardized. The processed numerical sequences of each dimension are integrated to form a complete feature fragment reflecting the entire process of the anomaly from latency to manifestation. In some embodiments, the set length of the historical time is determined by system configuration parameters. These parameters define the window size for the tracing analysis. When extracting complete numerical sequences of all relevant dimensions in the historical feature data, other feature dimensions associated with the core anomaly dimension are determined based on a preset dimension correlation mapping table. It can be understood that the correlation mapping table defines the physical or electrical correlation between different feature dimensions; for example, the current effective value change rate dimension is correlated with the temperature rise slope dimension. The standardization process uses the Z-score standardization method, and its calculation formula is:

[0100]

[0101] in: Represents the standardized first The value of each data point Represents the first element in the original numerical sequence. The value of each data point Representing dimensions The numerical mean over the entire time span of the complete feature segment. Representing dimensions The numerical standard deviation over the entire time span of the complete feature segment. See Table 1 for an example of the mapping relationship between a core anomaly dimension and its related dimensions.

[0102] Table 1: Mapping Table of Core Anomaly Dimensions and Their Related Dimensions

[0103]

[0104] Optionally, time alignment ensures that all data points in relevant dimensions have uniform and equally spaced timestamps. Integration arranges the processed numerical sequences of each dimension in dimensional order to form a multidimensional matrix, which is the data structure of the complete feature fragment. The complete feature fragment is considered a composite multidimensional time series signal. Empirical Mode Decomposition (EMD) is applied to adaptively decompose the time series signal of each dimension into a series of intrinsic mode function (IMF) components and a residual component. In practice, Hilbert transform is performed on the decomposed IMF components to obtain their time-frequency distribution characteristics. Based on the similarity of the time-frequency distribution characteristics, IMF components from different original feature dimensions are recombined. It can be understood that the recombining operation, based on a clustering algorithm, groups IMF components with similar time-frequency distribution characteristics into the same group. Each group of recombined component signals represents one physical meaning among trend, periodic fluctuation, shock events, and random noise. Each fault component signal carries information about a certain aspect of the physical process in the original composite signal. In some embodiments, the empirical mode decomposition process is performed independently on each column of data in the multidimensional matrix of complete feature fragments, and the Hilbert transform is used to calculate the instantaneous frequency and instantaneous amplitude of each intrinsic mode function component. Optionally, the similarity of time-frequency distribution features is measured by calculating the correlation coefficient between the instantaneous frequency sequences of different intrinsic mode function components, and the clustering algorithm uses the K-means algorithm to group the intrinsic mode function components.

[0105] See Figure 4 This is a multi-dimensional operational status monitoring chart for polyester staple fiber power equipment, showing the dynamic changes in voltage, current, power factor, and temperature over 100 seconds, clearly reflecting the process from normal operation to abnormality. An extremely low and constant power factor is the most prominent anomaly, leading to increased reactive power losses in the power grid and is also a typical manifestation of internal equipment failure. The temperature continuously rises over time, while the current drops later, possibly due to decreased insulation performance caused by overheating, resulting in abnormal current fluctuations. The current fluctuation characteristics provide data for load balancing and production scheduling, allowing for optimization of equipment operating conditions and reduction of energy consumption. The abnormal power factor exposes shortcomings in energy utilization efficiency, providing a clear direction for improvement in reactive power compensation system upgrades.

[0106] In one embodiment of the present invention, time is used as the first dimension, the type of the fault component signal as the second dimension, and the amplitude, energy, or frequency characteristics of the fault component signal as the third dimension. The characteristic value calculated for each fault component signal at its corresponding time point and signal type is treated as a data point in a three-dimensional space. All the data points calculated for the fault component signals are arranged and connected in the three-dimensional space according to time order and signal type relationship. In some embodiments, the third dimension of the fault component signal representing a trend uses a smoothed amplitude characteristic; the third dimension of the fault component signal representing periodic fluctuations uses the energy characteristics of the main frequency components; and the third dimension of the fault component signal representing an impact event uses the peak value characteristic of the transient amplitude. It can be understood that the three-dimensional coordinates of each data point... Determined through calculation, where This indicates the specific time corresponding to the data point. This indicates the fault component signal type identifier corresponding to the data point. Indicates time Fault component signal type The corresponding third-dimensional feature value. Fault component signal type identifier. It is a discrete enumerated value used to distinguish different types of fault component signals, such as trends, periodic fluctuations, shock events, and random noise. Third-dimensional eigenvalues The specific calculation method depends on the type of fault component signal. To determine, for example, the amplitude characteristic can be determined using the following formula:

[0107]

[0108] in: Indicates a point in time Calculated fault component signal type amplitude characteristics, Indicates the type of fault component signal At discrete time points instantaneous amplitude, This indicates the length of the local time window used to calculate the amplitude characteristics.

[0109] In practical implementation, discrete data points are fitted into a continuous three-dimensional surface or voxel model using a spatial interpolation algorithm. This three-dimensional surface or voxel model is the three-dimensional fault feature map. The three-dimensional fault feature map intuitively displays the interrelationships and structural characteristics of different fault components over time, and is used to match it with the fault pattern map in a preset fault knowledge base to determine the specific fault type and its evolutionary stage. Optionally, the spatial interpolation algorithm uses the three-dimensional kriging interpolation method, which constructs a variogram model based on the spatial correlation of data points and performs optimal unbiased estimation of the feature values ​​at unknown locations. In some embodiments, the three-dimensional surface model is generated by triangulating discrete data points and applying a surface subdivision algorithm, and the voxel model is generated by dividing the three-dimensional space into a uniform voxel grid and assigning values ​​to each voxel. It can be understood that the preset fault knowledge base stores typical three-dimensional fault feature map samples corresponding to various known faults during their occurrence and development. The matching process calculates the similarity between the three-dimensional fault feature map to be diagnosed and each sample map in the knowledge base in terms of topological structure and feature value distribution. Optionally, the matching process uses a descriptor based on the three-dimensional shape context to measure the similarity between the maps, and the determination of the evolution stage is based on the analysis of the matching degree sequence between the map to be diagnosed and the sample maps of different stages of the same fault type in the knowledge base.

[0110] See Figure 5 This is a 3D fault feature map, a visual model constructed in a 3D space of time, component type, and eigenvalue after deconstructing the original composite signal into trend components, periodic fluctuation components, impact event components, and random noise components. Different fault types form unique "feature fingerprints" in 3D space. For example, bearing wear manifests as an increase in the amplitude of the periodic fluctuation component, while an electrical short circuit causes a sharp spike in the impact event component. By matching the real-time generated 3D map with a pre-set fault knowledge base, the fault type can be quickly identified, and its evolution stage can be accurately determined. Based on the predicted trajectory of fault evolution, maintenance plans can be rationally arranged to avoid unnecessary downtime and reduce production losses caused by sudden faults. By monitoring the slope of the trend component and the frequency of the impact event component, abnormal signals can be identified before a fault becomes apparent, providing sufficient time for preventative maintenance.

[0111] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An online fault monitoring system for power equipment used in polyester staple fiber production, characterized in that, The system includes: The feature extraction module collects real-time operating status data of electrical equipment in the polyester staple fiber production line and performs multi-dimensional time series feature extraction to generate a time series feature sequence characterizing the operating status of the electrical equipment. The dynamic analysis module calculates the dynamic offset vector between the operating state of the power equipment and the standard operating state based on the time-series feature sequence. The evolution prediction module uses the dynamic offset vector to drive the fault mode evolution model and predict the future operating state evolution trajectory of the power equipment. The anomaly identification module generates multidimensional anomaly signal thresholds covering different fault types based on the evolution trajectory of the operating state, compares the time-series feature sequence with the multidimensional anomaly signal thresholds in real time, and identifies and locates preliminary fault anomaly signals. The feature construction module traces back to the time-series feature sequence based on the initial fault anomaly signal, extracts complete feature segments associated with the initial fault anomaly signal, performs multi-level signal deconstruction, decomposes the fault component signals under different physical meanings, and inputs the fault component signals to the fault feature map generator to construct a three-dimensional fault feature map that reflects the fault evolution law. The diagnostic matching module performs pattern matching between the three-dimensional fault feature map and a preset fault knowledge base to determine the specific fault type and its evolution stage of the power equipment. The step of extracting multidimensional time series features to generate a time series feature sequence characterizing the operating state of power equipment specifically includes: Extract voltage, current, power factor, and temperature data segments within fixed time windows from the original operating status data stream. Spectral analysis is performed on the voltage data segment to extract the fundamental amplitude, main harmonic distortion rate, and voltage imbalance characteristics. The current data segment is subjected to joint time-domain and frequency-domain analysis to extract the characteristics of the effective current value change rate, peak factor and total harmonic distortion rate of the current. Statistical analysis was performed on the power factor data segment to extract the power factor mean, fluctuation variance, and number of abrupt changes. Perform trend analysis on the temperature data segment to extract the temperature rise slope, local extreme points, and temperature rise rate characteristics; All extracted features are arranged in order of time window and combined to form the time-series feature sequence; The method of using the dynamic offset vector to drive the fault mode evolution model to predict the future operating state evolution trajectory of power equipment specifically includes: Input the current dynamic offset vector into the pre-trained fault mode evolution model; The fault mode evolution model learns the evolution rules from historical fault data and performs multi-step iterative deduction with the current dynamic offset vector as the initial state. Each iteration predicts the dynamic offset vector after the next time step. The dynamic offset vectors predicted by multiple consecutive time steps are connected in chronological order to form the trajectory of the future operating state evolution of the power equipment. The operational status evolution trajectory describes the path and extent to which the operational status of power equipment deviates from the standard state over a future period of time; The step of tracing back from the initial fault anomaly signal to the time-series feature sequence and extracting complete feature segments associated with the initial fault anomaly signal specifically includes: Based on the time point information contained in the preliminary fault anomaly signal, trace back a set length of historical time. Extract all feature data from the time-series feature sequence from the backtracking start point to the current time point; Using the feature dimension that triggers the initial anomaly marker as the core, extract the complete numerical sequence of all relevant dimensions from the historical feature data; The extracted complete numerical sequences of each dimension are time-aligned and standardized. The processed numerical sequences of each dimension are integrated to form the complete feature fragment that reflects the entire process of an anomaly from its latent state to its manifestation.

2. The online fault monitoring system for power equipment used in polyester staple fiber production according to claim 1, characterized in that, The real-time acquisition of operating status data of electrical equipment in the polyester staple fiber production line specifically includes: Deploy sensor arrays at the power input terminals, critical load nodes, and control loops of power equipment; The sensor array synchronously collects the instantaneous values ​​of three-phase voltage, three-phase current, power factor, and temperature of key nodes of the power equipment. The instantaneous values ​​of three-phase voltage, three-phase current, power factor, and temperature are collected and then aligned with timestamps and their sampling rates are standardized. The data, after being timestamped and sampled to a uniform rate, is integrated to form a raw running status data stream indexed by time.

3. The online fault monitoring system for power equipment used in polyester staple fiber production according to claim 2, characterized in that, The dynamic offset vector between the operating state and the standard operating state of power equipment is calculated based on the time-series feature sequence, specifically including: Retrieve the standard characteristic sequence of power equipment under standard healthy operating conditions, which is pre-stored in the database; The current feature vector in the time-series feature sequence is compared element-by-element with the standard feature vector at the corresponding time point in the standard feature sequence; Calculate the difference between the current feature vector and the standard feature vector in each feature dimension; A multidimensional vector is constructed using the differences across all feature dimensions as components; this multidimensional vector is the dynamic offset vector. The magnitude of the dynamic offset vector reflects the overall degree of offset, and the direction reflects which type or types of operational features have experienced abnormal offset.

4. The online fault monitoring system for power equipment used in polyester staple fiber production according to claim 3, characterized in that, Based on the aforementioned operational state evolution trajectory, a multidimensional abnormal signal threshold covering different fault types is generated, specifically including: Extract the maximum value of the predicted future dynamic offset vector in each feature dimension from the evolution trajectory of the operating state; For each known fault type, the typical offset range caused by the fault type in each feature dimension is determined based on its historical data; By combining the predicted maximum value of the future dynamic offset vector and the typical offset range of different fault types, a comprehensive warning upper limit threshold is calculated for each feature dimension; Based on the differences in the sensitivity of different feature dimensions to various types of faults, different weighting coefficients are assigned to the warning upper limit threshold. Finally, a multidimensional abnormal signal threshold set is generated, consisting of multiple weighted feature dimension thresholds.

5. The online fault monitoring system for power equipment used in polyester staple fiber production according to claim 4, characterized in that, The time-series feature sequence is compared with the multidimensional abnormal signal threshold in real time to identify and locate preliminary fault abnormal signals, specifically including: Obtain the feature vector of the latest time window from the time-series feature sequence; Each feature value in the feature vector of the latest time window is compared with the warning upper limit threshold of the corresponding feature dimension in the multidimensional abnormal signal threshold set; If a certain feature value exceeds its corresponding warning upper limit threshold, the initial anomaly labeling of the feature dimension is triggered. Record all feature dimensions that trigger the initial anomaly marker, along with the specific values ​​and time points when they exceed the limits; The initial abnormality markers and their related information that are triggered simultaneously are packaged into a preliminary fault abnormality signal.

6. The online fault monitoring system for power equipment used in polyester staple fiber production according to claim 5, characterized in that, The process of performing multi-level signal deconstruction to obtain fault component signals under different physical meanings specifically includes: The complete feature segment is considered as a composite multidimensional time series signal; By applying the empirical mode decomposition method, the time series signal of each dimension is adaptively decomposed into a series of intrinsic mode function components and a residual component; The Hilbert transform is performed on the intrinsic mode function components obtained from the decomposition to obtain their time-frequency distribution characteristics; Based on the similarity of time-frequency distribution characteristics, the intrinsic mode function components from different original feature dimensions are recombined to form the fault component signals that respectively characterize trends, periodic fluctuations, shock events, and random noise. Each of the fault component signals carries information about a certain aspect of the physical process in the original composite signal.

7. The online fault monitoring system for power equipment used in polyester staple fiber production according to claim 6, characterized in that, The fault component signal is input to a fault feature map generator to construct a three-dimensional fault feature map reflecting the fault evolution law, specifically including: Time is used as the first dimension, the type of fault component signal is used as the second dimension, and the amplitude, energy, or frequency characteristics of the fault component signal are used as the third dimension. The characteristic value calculated for each fault component signal at its corresponding time point and signal type is taken as a data point in a three-dimensional space. Arrange and connect the data points calculated from all fault component signals in three-dimensional space according to time sequence and signal type relationship; By using a spatial interpolation algorithm, discrete data points are fitted into a continuous three-dimensional surface or voxel model, which is the three-dimensional fault feature map. The three-dimensional fault feature map intuitively displays the interrelationships and structural features of different fault components over time, and is used to match it with the fault pattern map in the preset fault knowledge base, thereby determining the specific fault type and its evolution stage.

Images