A pta polyester integrated equipment fault prediction method and system
By combining time series decomposition and Kalman filtering in modeling, the problem of distinguishing between normal process changes and equipment malfunctions in polyester production was solved. This achieved highly sensitive and accurate fault detection, reduced false alarm rates, and improved production stability and product quality consistency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 江苏嘉通能源有限公司
- Filing Date
- 2026-04-23
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies struggle to accurately distinguish between reasonable changes caused by normal process transmission and abnormal deviations caused by equipment malfunctions in polyester production, leading to untimely early warnings or a high false alarm rate.
A time series decomposition and Kalman filtering fusion model is adopted. By collecting real-time operation data of upstream sections and downstream observations, the basic trend components and periodic components are separated. The residual sequence is processed by the Kalman filtering algorithm, the time series feature vector is calculated and the dynamic monitoring threshold is generated. Fault prediction is carried out by combining cross-section dynamic correlation diagram and multi-level state space model.
It improves the sensitivity and accuracy of fault signal detection, reduces the risk of production interruption caused by false alarms or missed alarms, enhances the reliability and interpretability of early warning, and realizes the visualization of fault propagation paths and the adaptive generation of dynamic monitoring thresholds.
Smart Images
Figure CN122087428B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial automation monitoring and fault technology, and in particular to a fault prediction method and system for PTA polyester integrated equipment. Background Technology
[0002] Currently, polyester production, as a core link in the modern chemical industry, directly supports the development of many downstream industries such as textiles, packaging, and electronics. Its stable and efficient operation is of key significance to ensuring the security of the industrial chain.
[0003] In existing technologies, most monitoring methods based on industrial control systems rely primarily on single-point threshold alarms or statistical analysis based on historical data. These methods typically focus on data changes at local monitoring points when detecting equipment anomalies.
[0004] However, when upstream process conditions are frequently adjusted or affected by external raw material fluctuations, downstream measurements fluctuate accordingly. Traditional monitoring methods can easily misjudge normal coupling responses as faults or drown out early fault signals amidst normal transmission fluctuations, leading to untimely warnings or a high false alarm rate. Therefore, existing technologies suffer from the problem of accurately distinguishing between reasonable changes caused by normal process transmission and abnormal deviations caused by equipment malfunctions. Summary of the Invention
[0005] This invention provides a method and system for predicting faults in PTA polyester integrated equipment, in order to solve the problem of difficulty in accurately distinguishing between reasonable changes caused by normal process transmission and abnormal deviations caused by equipment malfunctions.
[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for predicting the failure of an integrated PTA polyester equipment, comprising: Real-time operation data from upstream sections and observations from downstream sections are collected and time-series aligned. The basic trend component and periodic component are separated from the time-aligned sequence to obtain the residual sequence. The residual sequence is processed using the Kalman filter algorithm to obtain the comprehensive mapping value; If the deviation between the comprehensive mapping value and the downstream observation value exceeds a preset allowable limit, the corresponding sampling point is determined to be a potential abnormal deviation signal, and the time-series feature vector of the abnormal deviation signal is obtained. Calculate the Mahalanobis distance between the time-series feature vector and the historical normal operating condition statistical distribution, and output the classification result label based on the Mahalanobis distance; Extract the abnormal deviation subset marked as equipment area fault from the classification result labels, and calculate the duration and cumulative amplitude of the abnormal deviation subset. If the duration exceeds a preset time limit and the cumulative amplitude exceeds a preset intensity benchmark, generate an early warning trigger command. If the early warning trigger command is triggered, a cross-section dynamic association diagram is generated based on the abnormal deviation subset, the path weights between the affected node sequences in the cross-section dynamic association diagram are calculated, and the path with the highest path weight is locked as the fault propagation path. A multi-level state space model is generated through the fault propagation path. Based on the prediction results of the multi-level state space model, a dynamic monitoring threshold is generated after Kalman filtering. The dynamic monitoring threshold is combined with the signal lag time of the located fault source node to establish a normal transmission interval. If the difference between the downstream observation value and the normal transmission interval exceeds the preset safety constraint boundary, a final warning signal is generated.
[0007] Secondly, the present invention provides a fault prediction system for integrated PTA polyester equipment, comprising: The data preprocessing and sequence decomposition module collects real-time operation data from upstream sections and downstream observations and performs time-series alignment. It then separates the basic trend components and periodic components from the time-aligned sequence to obtain the residual sequence. The upstream impact modeling and mapping module uses the Kalman filter algorithm to process the residual sequence and obtain the comprehensive mapping value. The abnormal signal identification and feature extraction module determines that if the deviation between the comprehensive mapping value and the downstream observation value exceeds a preset allowable limit, the corresponding sampling point is a potential abnormal deviation signal, and the time-series feature vector of the abnormal deviation signal is obtained. The fault mode classification module calculates the Mahalanobis distance between the time-series feature vector and the historical normal operating condition statistical distribution, and outputs the classification result label based on the Mahalanobis distance. The early warning triggering module extracts the abnormal deviation subset marked as equipment area fault from the classification result labels, and calculates the duration and cumulative amplitude of the abnormal deviation subset. If the duration exceeds a preset time limit and the cumulative amplitude exceeds a preset intensity benchmark, an early warning triggering command is generated. If the early warning trigger command is triggered, the fault propagation path analysis module generates a cross-section dynamic association diagram based on the abnormal deviation subset, calculates the path weight between the affected node sequences in the cross-section dynamic association diagram, and locks the path with the highest path weight as the fault propagation path. The dynamic threshold generation module generates a multi-level state space model through the fault propagation path, and generates a dynamic monitoring threshold based on the prediction results of the multi-level state space model after Kalman filtering. The final early warning module combines the dynamic monitoring threshold with the signal lag time of the located fault source node to establish a normal transmission interval. If the difference between the downstream observation value and the normal transmission interval exceeds the preset safety constraint boundary, a final early warning signal is generated.
[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention achieves accurate estimation of upstream process transmission effects by combining time series decomposition and Kalman filtering modeling. In a strongly coupled, long-delay cross-segment environment, it effectively removes the influence of normal fluctuations on downstream observations, thereby significantly improving the sensitivity and accuracy of fault signal detection and reducing the risk of production interruption caused by false alarms or missed alarms in the industrial control system.
[0009] (2) By adopting a composite triggering mechanism, the present invention can intelligently distinguish between actual equipment faults and deviations caused by normal process adjustments, and make dual judgments by combining duration and amplitude, avoiding the singularity and limitations of traditional threshold alarms, enhancing the reliability and interpretability of early warnings, and providing clearer decision support for operation and maintenance personnel.
[0010] (3) By constructing a cross-section dynamic association diagram and a multi-level state space model, this invention realizes the visualization of fault propagation paths and the adaptive generation of dynamic monitoring thresholds. It can not only trace the source of the fault and the scope of its impact, but also update the monitoring boundary in real time, enabling the industrial control system to have continuous learning and optimization capabilities, thereby improving the overall operational stability of the device and the consistency of product quality. Attached Figure Description
[0011] Figure 1 This is a schematic diagram of the fault prediction method for PTA polyester integrated equipment provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the fault prediction system for PTA polyester integrated equipment provided in the second embodiment of the present invention. Detailed Implementation
[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0013] Reference Figure 1 The first embodiment of the present invention provides a method for predicting faults in an integrated PTA polyester equipment, comprising the following steps: S11: Collect real-time operation data of upstream sections and observations of downstream sections and perform time-series alignment. Separate the basic trend component and periodic component from the time-aligned sequence to obtain the residual sequence. S12, The residual sequence is processed using the Kalman filter algorithm to obtain the comprehensive mapping value; S13, if the deviation between the comprehensive mapping value and the downstream observation value exceeds the preset allowable limit value, then the corresponding sampling point is determined to be a potential abnormal deviation signal, and the time-series feature vector of the abnormal deviation signal is obtained. S14, calculate the Mahalanobis distance between the time-series feature vector and the historical normal operating condition statistical distribution, and output the classification result label based on the Mahalanobis distance; S15, extract the abnormal deviation subset marked as equipment area fault from the classification result labels, and calculate the duration and cumulative amplitude of the abnormal deviation subset. If the duration is greater than the preset time limit and the cumulative amplitude exceeds the preset intensity benchmark, generate an early warning trigger command. S16, if the early warning trigger command is triggered, a cross-section dynamic association diagram is generated based on the abnormal deviation subset, the path weights between the affected node sequences in the cross-section dynamic association diagram are calculated, and the path with the highest path weight is locked as the fault propagation path. S17, A multi-level state space model is generated through the fault propagation path, and a dynamic monitoring threshold is generated based on the prediction results of the multi-level state space model after Kalman filtering. S18, the dynamic monitoring threshold is combined with the signal lag time of the located fault source node to establish a normal transmission interval. If the difference between the downstream observation value and the normal transmission interval exceeds the preset safety constraint boundary, a final warning signal is generated.
[0014] It should be noted that deviation refers to the difference between the downstream comprehensive mapping value calculated by the upstream impact model and the actual downstream observation value collected; abnormal deviation refers to the situation where the magnitude of the deviation exceeds the preset allowable limit value set according to historical statistics and process requirements; potential abnormal deviation signal refers to the abnormal deviation identified by the system that may be caused by equipment failure, which triggers further feature extraction and classification processes; residual sequence refers to the sequence remaining after removing the basic trend component and periodic component from the original time-aligned sequence, mainly containing random noise and unmodeled sudden disturbances; comprehensive mapping value refers to the expected impact value of upstream section operation fluctuations on downstream observation indicators estimated by algorithms such as Kalman filtering.
[0015] In step S11, real-time operation data from the upstream section and downstream observations are collected and time-series aligned. The basic trend component and periodic component are separated from the time-aligned sequence to obtain the residual sequence, including: Real-time operation data of the upstream section and downstream observation values of the PTA polyester integrated equipment are obtained. Linear interpolation is used to fill the downstream observation values and align them with the real-time operation data of the upstream section to obtain a time-aligned sequence. The low-frequency change trajectory in the time-aligned sequence is processed using a weighted moving average to determine the underlying trend components; The underlying trend component is removed from the time-aligned sequence to obtain a detrended sequence. Autocorrelation analysis is then performed on the detrended sequence to extract the periodic component. The basic trend component and the periodic component are removed from the time-aligned sequence to obtain the residual sequence.
[0016] It should be noted that real-time operational data from the upstream section of the PTA polyester integrated equipment and downstream observations were acquired. The upstream operational data included high-frequency data such as esterification reaction temperature, acetic acid concentration, and catalyst dosage, while the downstream observations included low-frequency laboratory indicators such as chip viscosity, hue b-value, and carboxyl content. Due to the difference in sampling frequencies between the upstream and downstream data, a linear interpolation method was used to fill in the downstream low-frequency data, aligning it with the timestamps of the upstream high-frequency data to obtain a time-aligned sequence.
[0017] In one embodiment, upstream temperature data is recorded every minute, and downstream viscosity data is sampled every 4 hours. By performing linear interpolation of the time ratio between adjacent test points, a viscosity sequence aligned with the upstream minute-level data can be generated, providing a unified time reference for subsequent analysis.
[0018] It should be noted that the exponentially weighted moving average (EWMA) algorithm is used to smooth the time-aligned series, with a smoothing factor selected in the range of 0.02 to 0.10. By stripping away low-frequency variation trajectories and identifying the underlying trend components, the value range needs to be optimized within this range based on the historical fluctuation cycles and noise levels of key state variables to filter out daily cyclical fluctuations and retain slow process trends. The processing sequence length covers 48 to 96 hours of a typical production batch, and the specific values are determined through a grid search of historical normal operating condition sequences, with the objective function being to minimize the decay time of the autocorrelation coefficient of the detrended sequence. In practical applications, for different key state variables (such as temperature and concentration), the most suitable value needs to be selected within this range by analyzing the fluctuation cycles and noise levels of their respective historical sequences. This effectively filters out short-cycle fluctuations such as daily fluctuations while retaining slow process trends.
[0019] In one embodiment, for a continuous 72-hour esterification reaction temperature sequence, the following was selected: EWMA smoothing was performed. After smoothing, the temperature rise trend caused by catalyst activity decay was clearly displayed.
[0020] It should be noted that the detrended sequence is obtained by removing the basic trend component from the time-aligned sequence. Autocorrelation analysis is performed on the detrended sequence to calculate the autocorrelation coefficients from order 0 to the maximum lag order. The maximum lag order is set to one-quarter of the sequence length to cover the potential periodic range. Then, based on the sampling frequency and prior knowledge of the process, the lag order is converted into a time unit. The position where the autocorrelation coefficient exceeds the 95% confidence interval threshold and shows a local maximum value is identified as a significant periodic peak. The lag time corresponding to the peak is extracted as a periodic component.
[0021] In one embodiment, the autocorrelation plot shows a significant peak at 24 hours, corresponding to the daily cycle caused by nighttime load adjustments of the device or diurnal fluctuations in the environment, and a weaker peak at 168 hours, corresponding to the weekly cycle caused by weekend production plan fine-tuning.
[0022] It should be noted that the basic trend component and the periodic component are removed from the time-aligned sequence to obtain the residual sequence. This residual sequence mainly characterizes random disturbances and sudden abnormal events that are not explained by the trend and periodicity during device operation.
[0023] In one embodiment, when the upstream acetic acid concentration undergoes an abnormal abrupt change, after removal, the residual sequence will exhibit a significant spike at the corresponding time, the value of which may exceed the normal fluctuation range, while the residual values under other normal operating conditions are distributed within this range, thus highlighting the abnormal event.
[0024] In step S12, the residual sequence is processed using the Kalman filter algorithm to obtain the comprehensive mapping value, including: Analyze the residual sequence, establish a state vector containing the real-time operation data of the upstream section, and generate prior state estimates by combining the preset state transition matrix; The Kalman gain coefficient is calculated based on the prior state estimate and the statistical characteristics of the observation noise. The prior state estimate is then corrected using the Kalman gain coefficient to obtain the posterior state estimate. Finally, a comprehensive mapping value is extracted from the posterior state estimate to characterize the impact of the real-time operation of the upstream section on the downstream observation.
[0025] It should be noted that, by analyzing the residual sequence under historical normal operating conditions, key upstream variables that have a significant transmission effect on downstream observations are identified. Their deviations or rates of change in the residual sequence are extracted as the dimension of the state vector to establish the state vector. The state transition matrix is a square matrix with the same dimension as the state vector. Its diagonal elements are determined by fitting a first-order autoregressive model of the historical sequence of each key variable to reflect its own time correlation. The off-diagonal elements are set according to the process mechanism or the historical cross-correlation relationship between variables and their values are much smaller than those of the diagonal elements. When generating the prior state estimate, the preset state transition matrix is multiplied by the state vector estimate of the previous time step.
[0026] In one embodiment, the constructed state vector represents the acetic acid concentration deviation and the esterification temperature fluctuation range. The preset state transition matrix is as follows: The state estimate from the previous moment. Multiplying this matrix yields the prior state estimate for the current time step.
[0027] It should be noted that the dimensionless value of the observed noise variance is determined based on the variance of the standardized residual sequence under historical normal operating conditions, and the dimensionless value of the process noise variance is determined based on the variance of the standardized historical data of upstream key variables within the normal fluctuation range. The Kalman gain coefficient is calculated using the prediction error covariance matrix and the observed noise variance matrix according to the standard Kalman filter equation. The prior state estimate is weighted and corrected using this gain coefficient to obtain the posterior state estimate. The dimensionless comprehensive mapping value used to quantify the expected impact of upstream key variables on downstream is extracted from the posterior state estimate.
[0028] In one embodiment, based on standardized data, the observation noise variance is set to 0.04, and the process noise variance is set to 0.09. In the prior estimate, the acetic acid concentration bias component is 0.6, while the current observation residuals indicate that its influence is stronger. The Kalman filter algorithm corrects this prior value with a larger gain coefficient. After correction, the resulting composite mapping value rises to 0.9. This value represents the composite mapping value of the upstream variable's influence on the downstream at the current moment, indicating an increase in the intensity of the anomalous influence.
[0029] In step S13, if the deviation between the integrated mapping value and the downstream observation value exceeds a preset allowable limit, the corresponding sampling point is determined to be a potential abnormal deviation signal, and the time-series feature vector of the abnormal deviation signal is obtained, including: An instantaneous deviation sequence is generated based on the difference between the comprehensive mapping value and the downstream observation value. If the deviation value of the sampling point in the instantaneous deviation sequence exceeds the preset allowable limit value, the corresponding sampling point is marked as a potential abnormal deviation signal. The deviation signal segments are extracted from consecutive time windows before and after the potential abnormal deviation signal, and the deviation signal segments are analyzed in the time domain and frequency domain to obtain the time domain waveform geometric properties and frequency domain energy density distribution. The time-domain waveform geometric properties and the frequency-domain energy density distribution are normalized and vector-concatenated to generate the time-series feature vector of the potential abnormal deviation signal.
[0030] It should be noted that the downstream observations are first standardized using the Z-score to eliminate the influence of dimensions, and this is denoted as the standardized observation. The instantaneous deviation is the difference between the comprehensive mapped value and the standardized observation; this deviation is dimensionless. The preset allowable limit value is a dimensionless threshold used to determine whether the instantaneous deviation is abnormal. It is set by statistically analyzing deviation data under historical stable operating conditions. First, the 95% confidence interval of the historical deviation data is calculated, for example, ±0.3, and then multiplied by a process safety margin coefficient (usually between 1.2 and 1.5) to obtain the preset allowable limit value. For example, if the margin coefficient is 1.33, the limit value is set to ±0.4. When N consecutive data points in the instantaneous deviation sequence exceed the allowable limit, a potential abnormal deviation signal is triggered. The number of consecutive exceedances N is set based on simulation analysis of historical false alarms and missed alarms to achieve a balance between detection sensitivity and anti-interference capability. The specific value of N needs to be determined by combining the inherent sampling period of the data acquisition system and the typical establishment time of the process anomaly at the monitoring point. For example, at a minute-level sampling frequency, the total duration corresponding to N=3 is usually set in the range of 5 to 15 minutes to ensure that continuous real anomalies can be effectively captured while avoiding short-term interference.
[0031] In one embodiment, if the model mapping predicts a downstream composite mapping value of 4.8 and the actual observed value is 5.3 at a certain time period, the instantaneous deviation is calculated to be +0.5. If the preset tolerance limit is ±0.4, and the deviation at two consecutive subsequent time points both exceed +0.4, the system determines it to be a potential anomaly.
[0032] It should be noted that, taking the occurrence time of the potential abnormal deviation signal as the center, a fixed-length continuous time window is extracted before and after it, thereby generating a deviation signal segment containing the complete evolution process of the signal. The length of the time window is determined according to the dynamic response time of the process, and is usually 2-3 times the typical propagation time of the fault in that section. For example, for the midstream section, the window length can be set to 30 to 90 minutes. The segment is analyzed in both the time domain and the frequency domain. In the time domain, its waveform geometric properties are extracted, including peak height, rise time, duration, and waveform symmetry. In the frequency domain, its energy density distribution is calculated using Fast Fourier Transform, and the energy concentration characteristics in the low-frequency and mid-frequency bands are analyzed to identify potential periodic or oscillation modes.
[0033] It should be noted that the time-domain waveform geometric attributes are extracted from the deviation signal segment. The maximum deviation amplitude within the segment is divided by the normalized peak height of the historical maximum observation value. The time for the signal to rise from the initial amplitude to the peak is divided by the total window length to obtain the normalized rise time. The total duration of the signal exceeding a preset threshold is divided by the total window length to obtain the normalized duration. Taking the moment when the signal first exceeds the threshold as the starting point and the moment when it last falls back below the threshold as the ending point, the integral area of the rising and falling segments is calculated, and the ratio of the difference between the two to their sum is taken as the waveform symmetry index. For the same signal... The frequency domain energy density distribution is extracted by performing a Fast Fourier Transform on the segment. The proportion of energy in the 0 to 0.1 period / hour frequency band is calculated to obtain the low-frequency energy proportion. The frequency corresponding to the energy peak in the 0.1 to 2 period / hour frequency band is identified as the mid-frequency dominant frequency. The frequency domain energy entropy is calculated based on the uniformity of energy distribution in the frequency domain. The above seven feature values are concatenated in a predetermined order: normalized peak height, normalized rise time, normalized duration, waveform symmetry index, low-frequency energy proportion, mid-frequency dominant frequency, and frequency domain energy entropy to generate a seven-dimensional time-series feature vector. That is, the time-series feature vector is a multi-dimensional vector combined in a predetermined order.
[0034] In one embodiment, for a deviation signal segment lasting approximately 45 minutes, the normalized peak height is calculated to be 0.85, the normalized rise time is 0.22, the normalized duration is 0.72, the waveform symmetry index is 0.15, the low-frequency energy proportion is 0.68, the mid-frequency dominant frequency is 0.35 cycles / hour, and the frequency domain energy entropy is 2.1. Combining these seven values in sequence constitutes a time-series feature vector characterizing the multi-scale dynamic properties of the anomalous signal.
[0035] In step S14, the Mahalanobis distance between the time-series feature vector and the historical normal operating condition statistical distribution is calculated, and the classification result label is output based on the Mahalanobis distance, including: Calculate the mean vector and covariance matrix of the time-series feature vectors under historical normal operating conditions; Calculate the Mahalanobis distance between the current time-series feature vector, the mean vector, and the covariance matrix. Compare the Mahalanobis distance with a preset confidence threshold and output the classification result label.
[0036] It should be noted that, firstly, data from the equipment's continuous operation under normal historical conditions for no less than 30 days is collected. The time-series feature vectors of each extracted sampling point are used to form a normal sample set. Then, the mean vector and covariance matrix of all historical time-series feature vectors in this sample set are calculated and stored as classification parameters.
[0037] It should be noted that the Mahalanobis distance is obtained by taking the square root of the weighted sum of the differences between the current feature vector and the mean vector using the inverse of the covariance matrix. This distance value is used to quantify the degree to which the current vector deviates from the historical normal distribution. The preset confidence threshold is determined by the statistical distribution of all Mahalanobis distances under historical normal operating conditions, specifically by taking its 95th quantile. For example, if the 95th quantile of the historical Mahalanobis distance is 3.2, then the threshold is set to 3.2. The calculated Mahalanobis distance is compared with this threshold. If it is greater than the threshold, the classification result label is output as equipment failure; otherwise, normal fluctuation is output.
[0038] In one embodiment, the Mahalanobis distance is calculated from 2000 samples collected under normal operating conditions. Its 95th percentile is 2.8. The Mahalanobis distance of the current time series feature vector to be judged is 4.1. Since 4.1 is greater than 2.8, the device fault label is output.
[0039] In step S15, an abnormal deviation subset marked as equipment area fault is extracted from the classification result labels, and the duration and cumulative amplitude of the abnormal deviation subset are calculated. If the duration exceeds a preset time limit and the cumulative amplitude exceeds a preset intensity benchmark, an early warning trigger command is generated, including: Extract the abnormal deviation subset marked as equipment area fault from the classification result labels. If the time interval between consecutive sampling points in the abnormal deviation subset is less than a preset time interval threshold, it is divided into continuous fault segments. Calculate the duration and cumulative amplitude of the continuous fault segment. If the duration exceeds a preset time limit and the cumulative amplitude exceeds a preset intensity benchmark, then generate an early warning trigger command.
[0040] It should be noted that all sampling points marked as equipment fault areas are selected from the classification results labels, and the timestamp and instantaneous deviation value of each sampling point are extracted to form an abnormal deviation subset. The preset time interval threshold is determined by statistical distribution analysis of the time intervals of consecutive fault sampling points in historical fault event samples. The continuous marked fault periods corresponding to confirmed fault events in the past 12 months are collected, and the time interval between adjacent sampling points within the period is calculated. The 90th percentile of these interval values is used as the threshold setting benchmark and corrected by combining the inherent sampling period of 1 minute of the data acquisition system. Statistical analysis shows that the time interval threshold at the minute-level sampling frequency is 2 minutes. The sampling points in the abnormal deviation subset are traversed in ascending order of timestamp, and the time interval between two adjacent points is calculated. If the interval is less than or equal to 2 minutes, they are classified into the same continuous fault segment; if the interval is greater than 2 minutes, they are divided into different independent fault segments.
[0041] It should be further explained that the historical fault events are obtained by collecting production operation data for 12 consecutive months, filtering out all abnormal events that trigger DCS alarms or are recorded by operators, and combining equipment maintenance records, maintenance work orders, process parameter trend charts, and post-event fault analysis reports to independently judge each event. If the event causes equipment component damage, performance degradation, or product indicators exceeding the standard, and is restored after maintenance, it is marked as a real fault event. If it is only a process fluctuation or instrument drift and does not cause actual loss, it is marked as a non-fault event. When the sample size of historical fault events is insufficient, for example, less than 30, typical fault characteristics, such as step, ramp, oscillation, etc., are injected into the normal operating condition data to generate simulated fault events.
[0042] In one embodiment, the time intervals between adjacent time points T1, T2, and T3 are all 1 minute, which is less than the threshold of 2 minutes, so they are aggregated into the same continuous fault segment, while the interval between T2 and T5 is 3 minutes, which is greater than the threshold, so they are divided into different fault segments.
[0043] It should be noted that when calculating the duration of a continuous fault segment, the difference between the timestamps of the first and last sampling points within the segment is used. When calculating the cumulative amplitude, the absolute values of all instantaneous deviations within the segment are summed using a trapezoidal integral. The preset time limit and preset intensity benchmark are set based on the statistical distribution analysis of valid fault event samples that have been confirmed to have an actual impact on production over the past 12 months. The duration and cumulative amplitude of each fault event are collected, and cumulative probability distribution curves are plotted. The time tolerance limit is set as the 10th percentile of the duration distribution to filter out more than 95% of brief noise disturbances. Statistically, this value is 20 minutes. The intensity benchmark limit is set as the 15th percentile of the cumulative amplitude distribution to ensure coverage of more than 85% of real fault events. Statistically, this value is 15.0 deviation units. When the duration of a continuous fault segment is greater than 20 minutes and the cumulative amplitude exceeds 15.0, the composite triggering condition is met, and an early warning trigger command is generated.
[0044] In one embodiment, a fault segment lasts from 08:15 to 08:42, a duration of 27 minutes (greater than 20 minutes), and the cumulative amplitude is 21.3 (greater than 15.0), triggering a warning command from the system.
[0045] In step S16, if the early warning trigger command is triggered, a cross-section dynamic correlation diagram is generated based on the abnormal deviation subset. The path weights between the affected node sequences in the cross-section dynamic correlation diagram are calculated, and the path with the highest weight is identified as the fault propagation path, including: If the early warning trigger command is triggered, the abnormal deviation subset is mapped to the preset section directed graph to locate the fault source node and downstream adjacent node. Calculate the maximum cross-correlation strength and signal lag time between the fault source node and the downstream adjacent node. If the maximum cross-correlation strength exceeds a preset correlation threshold, construct a cross-section dynamic correlation diagram based on the signal lag time. Identify the affected node sequence in the cross-section dynamic association diagram, and multiply the maximum cross-correlation strength of each edge through which the affected node sequence passes from the fault source node to the end node to generate path weights, and lock the path with the highest path weight as the fault propagation path.
[0046] It should be noted that if an early warning trigger command is triggered, the subset of abnormal deviations will be mapped to a pre-constructed directed graph of the process segment based on the production process flow diagram and equipment connection relationships. The method for constructing the directed graph of the process segment includes parsing the piping and instrumentation flow diagram of the PTA polyester integrated process, extracting each equipment node and the material and energy transfer direction, generating an initial directed graph with equipment as nodes and transfer relationships as directed edges, and then combining the point table and equipment labels in the distributed control system to uniquely identify and map the attributes of nodes and edges to form a computable topology. To ensure the integrity and accuracy of the graph, the relationships between each node pair are calculated using 30 consecutive days of historical normal operation data. The normalized cross-correlation strength is used to determine the edges with a cross-correlation strength below 0.3, which are then considered to have no significant transitive relationship and are deleted. This threshold is set based on the lower limit of the 95% confidence interval of the cross-correlation strength distribution under historical normal operating conditions. At the same time, process expert rules are introduced to manually review and correct isolated nodes and abnormal connections. When a process change or equipment is cut off, the directed graph is automatically updated by reading the equipment status bits and process configuration table in the distributed control system. For equipment cut-off, the corresponding node is marked as offline and its outgoing edges are temporarily hidden. For process changes, the parsing and verification steps are re-executed based on the changed piping and instrumentation diagram to generate a new version of the directed graph and replace the old version.
[0047] When locating the fault source node, each node in the directed graph of the section is traversed within the time interval covered by the abnormal deviation subset. The instantaneous deviation of each node's monitored value relative to its historical normal operating condition baseline value is calculated. The node that first shows a deviation, whose deviation duration exceeds a preset duration threshold, and whose deviation amplitude exceeds a preset amplitude threshold is marked as the fault source node. The historical normal operating condition baseline value is the moving average of the node at the same time in the previous 24 hours. The preset duration threshold is set to 3 minutes by taking the 10th percentile of the statistical distribution of the duration of the first abnormal node in the confirmed fault events in the past 12 months. The preset amplitude threshold is 3 times the standard deviation of the fluctuation of the node's monitored value under historical normal operating conditions. For example, in a topology containing a reactor, heat exchanger, and distillation column, if the reactor outlet temperature monitoring point first shows a deviation exceeding 3 times the standard deviation at t=0 minutes and the deviation lasts for more than 3 minutes, while the heat exchanger shows a similar deviation at t=4 minutes, then the reactor is located as the fault source node, and the heat exchanger directly connected downstream is the downstream adjacent node.
[0048] It should be noted that the normalized cross-correlation function of the deviation sequence between the fault source node and the downstream adjacent node within a preset time window is calculated, and the maximum value of the function is taken as the maximum cross-correlation strength. The time shift corresponding to the maximum value is taken as the signal lag time. The preset correlation threshold is determined by statistical distribution analysis of the maximum cross-correlation strength of all process-related node pairs under historical normal operating conditions. The maximum cross-correlation value of each node pair during normal production periods for 12 consecutive months is collected, and the value of 0.75, the 95th percentile of the dataset, is taken as the threshold to control the false correlation probability to within 5%. If the maximum cross-correlation strength of the node pair exceeds 0.75, it is considered that there is a significant propagation relationship. A cross-section dynamic correlation graph is constructed with the fault source node as the starting point, the downstream adjacent node as the ending point, and the signal lag time as the directed edge weight.
[0049] In one embodiment, when the maximum cross-correlation strength of the deviation sequence between the reactor (source node) and the heat exchanger (adjacent node) is 0.87 (greater than the threshold of 0.75) and the lag time is 4.2 minutes, a directed edge from the reactor to the heat exchanger is drawn in the correlation graph and labeled with the lag time of 4.2 minutes.
[0050] It should be noted that, starting from the fault source node, the dynamic correlation graph across the work section is traversed. Based on the direction of the directed edge and the signal lag time, the downstream nodes are visited in ascending order of time. Each complete access path from the source node to the end node is recorded as the sequence of affected nodes. The path weight matrix is arranged with the propagation start node as the row and each node on the path as the column. The matrix element is the product of the maximum mutual correlation strength of each edge traversed from the start node to the node along the path to quantify the cumulative propagation effect. The path weight value of the end node corresponding to each affected node sequence is calculated. The propagation chain corresponding to the maximum path weight is selected as the main fault propagation path.
[0051] In one embodiment, in a chain that includes a reactor, a heat exchanger, a feed point to a separation tower, and a top condenser, the chain from the reactor to the top condenser is determined to be the main fault propagation path because the calculated product weight is higher than other paths.
[0052] In step S17, a multi-level state-space model is generated through the fault propagation path. Based on the prediction results of the multi-level state-space model, a dynamic monitoring threshold is generated after Kalman filtering, including: A multi-level state-space model is constructed based on the transmission gain and lag time between nodes in the fault propagation path. The abnormal signal of the fault source node is input into the multi-level state space model to calculate the prediction result of the downstream adjacent node. Historical time-series fluctuation characteristics are obtained as initial noise statistical parameters for the Kalman filter. The predicted results are used as the observation benchmark for the Kalman filter. The Kalman filter is then used to predict the state covariance of the multi-level state-space model in real time, generating dynamic monitoring thresholds. It should be noted that, based on adjacent node pairs along the fault propagation path, monitoring signals from two nodes under normal operating conditions for more than 30 consecutive days are collected. A step excitation is applied to the upstream node, and the downstream response curve is recorded. The least squares method is used to fit a first-order inertial plus pure time delay transfer function model, identifying the transfer gain and lag time constant. A 2×2 transfer feature matrix is constructed using the transfer gain and lag time constant as elements. Key state variables such as temperature or pressure of each equipment node along the path are used as state variables. State equations are established based on the gain and time constant in the transfer feature matrices of adjacent nodes to describe the transfer and delay of state variables from the preceding node to the following node. All nodes are connected in series along the path to form a multi-level state-space model.
[0053] In one embodiment, in the path from the reactor to the heat exchanger to the separation tower, by fitting and averaging the historical step test data of 16 times, the transfer gain from the reactor to the heat exchanger is 0.85 and the lag time is 4.5 minutes, and the transfer gain from the heat exchanger to the separation tower is 0.92 and the lag time is 5.2 minutes. Based on this, a third-order state-space model is constructed.
[0054] It should be noted that the abnormal signal deviation is obtained by subtracting the recent historical moving average value of the actual monitored value of the fault source node at the current moment. The moving average window length is set to 4 hours based on the analysis of the fluctuation cycle of historical normal operating conditions to balance the tracking speed and noise suppression. This deviation is used as the excitation input to the multi-level state-space model. Based on the adjacent node transmission gain and lag time constant included in the previous step through step response test and least squares fitting, the state equation is solved by numerical integration using the fourth-order Runge-Kutta method or discrete recursion using the first-order backward Euler method. The dynamic response process of the state variables of the downstream nodes is calculated level by level, and the prediction results of each downstream node under the influence of the current fault propagation are output.
[0055] In one embodiment, the abnormal temperature rise of the reactor by 5°C is used as input. The model calculates that the temperature of the heat exchanger will rise by 4.25°C after 4.5 minutes based on the transfer gain of 0.85 from the reactor to the heat exchanger and the lag time of 4.5 minutes. Then, based on the transfer gain of 0.92 from the heat exchanger to the separation tower and the lag time of 5.2 minutes, the model calculates that the temperature of the separation tower will rise by 3.91°C after 8.0 minutes.
[0056] It should be noted that the standard deviation of the system's key state variables under historical normal operating conditions for 12 consecutive months is used as the time-series fluctuation characteristic. For example, the standard deviation of the reactor temperature is 1.2℃. This standard deviation is used as the initial process noise covariance matrix parameter of the Kalman filter. The multi-level state space model is used as the state transition equation of the Kalman filter, and the prediction result is used as the observation benchmark for Kalman filter observation update. In real-time operation, the Kalman filter algorithm is used to recursively update the state estimate and its error covariance matrix. The prediction interval of each state variable is calculated by multiplying the standard deviation of the error covariance matrix of the state estimate predicted at the current time by 1.96 times (corresponding to a 95% confidence level, which is determined based on the normal distribution assumption). The upper and lower boundary values of the prediction interval are generated as dynamic monitoring thresholds.
[0057] In one embodiment, if the standard deviation of the predicted reactor temperature state covariance is 1.5°C, the dynamic monitoring threshold is set to the predicted value plus or minus 3.0°C.
[0058] In step S18, the dynamic monitoring threshold is combined with the signal lag time of the located fault source node to establish a normal transmission interval. If the difference between the downstream observation value and the normal transmission interval exceeds the preset safety constraint boundary, a final warning signal is generated, including: The dynamic monitoring threshold is combined with the signal lag time of the fault source node to perform time phase compensation and establish a normal transmission interval. The absolute difference between the downstream observation and the normal transmission interval is calculated to obtain a transmission deviation measure; If the transmission deviation metric exceeds the preset safety constraint boundary, a final warning signal is generated.
[0059] It should be noted that, in order to dynamically set the evaluation time window length based on the total estimated delay from the current node to the fault source node on the fault propagation path, the window length is taken as 1.5 to 2.5 times the delay to ensure the time from the arrival of the coverage anomaly signal to the complete presentation of the dynamic response. For example, when the total estimated delay is 20 minutes, the window length is taken as 30 to 50 minutes. Within this window, if the downstream observation value falls within the normal transmission interval at each sampling moment, the deviation distance is zero. If it falls outside the interval, the absolute difference between the observation value and the nearest interval boundary is calculated as the deviation distance at that moment. The root mean square of the deviation distance for all sampling moments within the window is used to obtain the transmission deviation metric.
[0060] In one embodiment, if the predicted temperature dynamic monitoring threshold of the reactor at 10:00 is ±3.0℃ and the time delay from the reactor to the heat exchanger is 4.5 minutes, then the normal transfer range of the heat exchanger at 10:04:30 is set to ±3.0℃.
[0061] It should be noted that the subsequent observation sequence of the real-time sensor is collected, and the evaluation time window length is dynamically determined based on the total estimated delay from the current node to the fault source node along the fault propagation path. This time window length is 1.5 to 2.5 times the total estimated delay. For example, if the total estimated delay is 20 minutes, the window length is 30 to 50 minutes to ensure the time from the arrival of the coverage abnormal signal to the complete presentation of the dynamic response. Within this window, the corresponding normal transmission interval is obtained for each sampling moment. If the observation value falls within the interval, the deviation distance at that moment is zero. If it falls outside the interval, the absolute difference between the observation value and the nearest interval boundary is calculated as the deviation distance at that moment. The root mean square of the deviation distance for all sampling moments within the window is calculated to obtain the transmission deviation metric.
[0062] In one embodiment, the dynamic monitoring threshold of the reactor at 10:00 is the predicted value ± 3.0℃, and the signal lag time from the reactor to the heat exchanger is 4.5 minutes. Therefore, the normal transmission range of the heat exchanger at 10:04:30 is the predicted value ± 3.0℃.
[0063] It should be noted that the safety constraint boundary is determined through statistical analysis of a 12-month continuous sample of transmission deviation measurements during historical normal operation. The 95th percentile value of this sample set is calculated, and combined with a comprehensive consideration of the maximum permissible deviation tolerance for this monitoring point in the process safety regulations, the 95th percentile of the transmission deviation measurement under historical normal operating conditions for the heat exchanger node is 0.5℃, and the maximum permissible deviation tolerance is 0.6℃. Therefore, the safety constraint boundary is set at 0.5℃. Comparing the current transmission deviation measurement with this boundary, if it exceeds 0.5℃, an anomaly is determined, and a warning level is assigned based on the degree of exceedance. A medium-level warning is generated for exceedances between 0.5℃ and 1.0℃, and a high-level warning is generated for exceedances above 1.0℃. The system automatically pushes a warning signal containing the warning level, the exceeding value, the occurrence time, and related node information to the monitoring terminal.
[0064] In one embodiment, if the heat exchanger transmission deviation reaches 0.7°C, exceeding the 0.5°C safety constraint boundary, a medium-level warning signal is generated to prompt close attention to the heat exchanger's operating status.
[0065] It should also be noted that, simultaneously with or after the generation of the final warning signal, the system feeds back the updated dynamic monitoring threshold to the early warning triggering module in real time. Based on the updated threshold, the early warning triggering module reassesses whether the current and subsequent abnormal deviation subsets meet the triggering conditions. If the threshold update causes the duration or cumulative amplitude of the original abnormal deviation subset to no longer exceed the new limit, the system can automatically suppress or cancel the corresponding early warning; conversely, if historical data identifies new anomalies under the new, more sensitive threshold, the system can re-trigger the early warning. Through this feedback mechanism, a closed-loop adaptive adjustment from model optimization to warning decision-making is achieved, enabling the warning system to continuously track changes in process status and improve the accuracy and timeliness of warnings. After the final warning is generated, the system feeds back the updated dynamic monitoring threshold to the early warning triggering module in real time, updating the threshold every 30 minutes. If the threshold change is less than 5% within three consecutive cycles, it is considered that the model has converged and entered a stable monitoring state.
[0066] In summary, this invention discloses a fault prediction method for integrated PTA polyester equipment. The method includes: separating the trend and periodic components of upstream and downstream sequences through data preprocessing to obtain residual sequences; constructing a mapping model of upstream influence on downstream based on Kalman filtering to identify potential abnormal deviation signals and extract their temporal characteristics; using Mahalanobis distance to distinguish between equipment faults and normal fluctuations; analyzing the duration and amplitude of fault subsets to trigger early warnings; constructing a cross-section dynamic correlation graph to lock the fault propagation path; and optimizing the response model and updating dynamic monitoring thresholds through path simulation to achieve the final early warning. This invention deeply integrates time series analysis, state estimation, and machine learning techniques, making it particularly suitable for production scenarios with strong coupling and long time delays across multiple sections in industrial control systems. It achieves accurate identification of early equipment anomalies, significantly reduces false alarms and missed alarms, and provides clear guidance for fault location and process adjustment.
[0067] Reference Figure 2 The second embodiment of the present invention provides a fault prediction system for integrated PTA polyester equipment, comprising: The data preprocessing and sequence decomposition module collects real-time operation data from upstream sections and downstream observations and performs time-series alignment. It then separates the basic trend components and periodic components from the time-aligned sequence to obtain the residual sequence. The upstream impact modeling and mapping module uses the Kalman filter algorithm to process the residual sequence and obtain the comprehensive mapping value. The abnormal signal identification and feature extraction module determines that if the deviation between the comprehensive mapping value and the downstream observation value exceeds a preset allowable limit, the corresponding sampling point is a potential abnormal deviation signal, and the time-series feature vector of the abnormal deviation signal is obtained. The fault mode classification module calculates the Mahalanobis distance between the time-series feature vector and the historical normal operating condition statistical distribution, and outputs the classification result label based on the Mahalanobis distance. The early warning triggering module extracts the abnormal deviation subset marked as equipment area fault from the classification result labels, and calculates the duration and cumulative amplitude of the abnormal deviation subset. If the duration exceeds a preset time limit and the cumulative amplitude exceeds a preset intensity benchmark, an early warning triggering command is generated. If the early warning trigger command is triggered, the fault propagation path analysis module generates a cross-section dynamic association diagram based on the abnormal deviation subset, calculates the path weight between the affected node sequences in the cross-section dynamic association diagram, and locks the path with the highest path weight as the fault propagation path. The dynamic threshold generation module generates a multi-level state space model through the fault propagation path, and generates a dynamic monitoring threshold based on the prediction results of the multi-level state space model after Kalman filtering. The final early warning module combines the dynamic monitoring threshold with the signal lag time of the located fault source node to establish a normal transmission interval. If the difference between the downstream observation value and the normal transmission interval exceeds the preset safety constraint boundary, a final early warning signal is generated.
[0068] It should be noted that the PTA polyester integrated equipment fault prediction system provided in this embodiment of the invention is used to execute all the process steps of the PTA polyester integrated equipment fault prediction method in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.
[0069] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.
[0070] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.
Claims
1. A method for predicting faults in an integrated PTA polyester equipment, characterized in that, include: Real-time operation data from upstream sections and observations from downstream sections are collected and time-series aligned. The basic trend component and periodic component are separated from the time-aligned sequence to obtain the residual sequence. The residual sequence is processed using the Kalman filter algorithm to obtain the comprehensive mapping value; If the deviation between the comprehensive mapping value and the downstream observation value exceeds a preset allowable limit, the corresponding sampling point is determined to be a potential abnormal deviation signal, and the time-series feature vector of the abnormal deviation signal is obtained. Calculate the Mahalanobis distance between the time-series feature vector and the historical normal operating condition statistical distribution, and output the classification result label based on the Mahalanobis distance; Extract the abnormal deviation subset marked as equipment area fault from the classification result labels, and calculate the duration and cumulative amplitude of the abnormal deviation subset. If the duration exceeds a preset time limit and the cumulative amplitude exceeds a preset intensity benchmark, generate an early warning trigger command. If the early warning trigger command is triggered, a cross-section dynamic association diagram is generated based on the abnormal deviation subset, the path weights between the affected node sequences in the cross-section dynamic association diagram are calculated, and the path with the highest path weight is locked as the fault propagation path. A multi-level state space model is generated through the fault propagation path. Based on the prediction results of the multi-level state space model, a dynamic monitoring threshold is generated after Kalman filtering. The dynamic monitoring threshold is combined with the signal lag time of the located fault source node to establish a normal transmission interval. If the difference between the downstream observation value and the normal transmission interval exceeds the preset safety constraint boundary, a final warning signal is generated. The process of processing the residual sequence using a Kalman filter algorithm to obtain a comprehensive mapping value includes: analyzing the residual sequence, establishing a state vector containing real-time operation data of the upstream section, and generating a priori state estimate by combining it with a preset state transition matrix; calculating the Kalman gain coefficient based on the priori state estimate and the statistical characteristics of the observation noise, using the Kalman gain coefficient to correct the priori state estimate to obtain a posterior state estimate, and extracting a comprehensive mapping value from the posterior state estimate to characterize the impact of the real-time operation of the upstream section on the downstream observation value; In this process, based on the pairs of adjacent nodes along the fault propagation path, the monitoring signals of the two nodes under historical normal operating conditions are collected. A step excitation is applied to the upstream node and the downstream response curve is recorded. The least squares method is used to fit the first-order inertial plus pure time delay transfer function model, and the transfer gain and time delay constant are identified. The transfer gain and time delay constant are used as elements to construct a 2×2 transfer feature matrix. The temperature or pressure key state variables of each equipment node on the path are used as state variables. The state equation is established based on the gain and time constant in the transfer feature matrix of adjacent nodes to describe the transfer and delay of state variables from the previous node to the next node. All nodes are connected in series along the path to form a multi-level state space model.
2. The fault prediction method for PTA polyester integrated equipment according to claim 1, characterized in that, Real-time operational data from upstream sections and downstream observations are collected and time-series aligned. The underlying trend component and periodic component are separated from the time-aligned sequence to obtain the residual sequence, including: Real-time operation data of the upstream section and downstream observation values of the PTA polyester integrated equipment are obtained. Linear interpolation is used to fill the downstream observation values and align them with the real-time operation data of the upstream section to obtain a time-aligned sequence. The low-frequency change trajectory in the time-aligned sequence is processed using a weighted moving average to determine the underlying trend components; The underlying trend component is removed from the time-aligned sequence to obtain a detrended sequence. Autocorrelation analysis is then performed on the detrended sequence to extract the periodic component. The basic trend component and the periodic component are removed from the time-aligned sequence to obtain the residual sequence.
3. The PTA polyester integrated equipment fault prediction method according to claim 1, characterized in that, If the deviation between the comprehensive mapping value and the downstream observation value exceeds a preset allowable limit, the corresponding sampling point is determined to be a potential abnormal deviation signal, and the time-series feature vector of the abnormal deviation signal is obtained, including: An instantaneous deviation sequence is generated based on the difference between the comprehensive mapping value and the downstream observation value. If the deviation value of the sampling point in the instantaneous deviation sequence exceeds the preset allowable limit value, the corresponding sampling point is marked as a potential abnormal deviation signal. The deviation signal segments are extracted from consecutive time windows before and after the potential abnormal deviation signal, and the deviation signal segments are analyzed in the time domain and frequency domain to obtain the time domain waveform geometric properties and frequency domain energy density distribution. The time-domain waveform geometric properties and the frequency-domain energy density distribution are normalized and vector-concatenated to generate the time-series feature vector of the potential abnormal deviation signal.
4. The PTA polyester integrated equipment fault prediction method according to claim 1, characterized in that, Calculate the Mahalanobis distance between the time-series feature vector and the historical normal operating condition statistical distribution, and output the classification result label based on the Mahalanobis distance, including: Calculate the mean vector and covariance matrix of the time-series feature vectors under historical normal operating conditions; Calculate the Mahalanobis distance between the current time-series feature vector, the mean vector, and the covariance matrix. Compare the Mahalanobis distance with a preset confidence threshold and output the classification result label.
5. The PTA polyester integrated equipment fault prediction method according to claim 1, characterized in that, Extract the abnormal deviation subset marked as equipment area fault from the classification result labels, and calculate the duration and cumulative amplitude of the abnormal deviation subset. If the duration exceeds a preset time limit and the cumulative amplitude exceeds a preset intensity benchmark, generate an early warning trigger command, including: Extract the abnormal deviation subset marked as equipment area fault from the classification result labels. If the time interval between consecutive sampling points in the abnormal deviation subset is less than a preset time interval threshold, it is divided into continuous fault segments. Calculate the duration and cumulative amplitude of the continuous fault segment. If the duration exceeds a preset time limit and the cumulative amplitude exceeds a preset intensity benchmark, then generate an early warning trigger command.
6. The PTA polyester integrated equipment fault prediction method according to claim 5, characterized in that, If the early warning trigger command is triggered, a cross-section dynamic correlation diagram is generated based on the abnormal deviation subset. The path weights between the affected node sequences in the cross-section dynamic correlation diagram are calculated, and the path with the highest weight is identified as the fault propagation path, including: If the early warning trigger command is triggered, the abnormal deviation subset is mapped to the preset section directed graph to locate the fault source node and downstream adjacent node. Calculate the maximum cross-correlation strength and signal lag time between the fault source node and the downstream adjacent node. If the maximum cross-correlation strength exceeds a preset correlation threshold, construct a cross-section dynamic correlation diagram based on the signal lag time. Identify the affected node sequence in the cross-section dynamic association diagram, and multiply the maximum cross-correlation strength of each edge through which the affected node sequence passes from the fault source node to the end node to generate path weights, and lock the path with the highest path weight as the fault propagation path.
7. The PTA polyester integrated equipment fault prediction method according to claim 6, characterized in that, A multi-level state-space model is generated through the fault propagation path. Based on the prediction results of the multi-level state-space model, a dynamic monitoring threshold is generated after Kalman filtering, including: A multi-level state-space model is constructed based on the transmission gain and lag time between nodes in the fault propagation path. The abnormal signal of the fault source node is input into the multi-level state space model to calculate the prediction result of the downstream adjacent node. Historical time-series fluctuation characteristics are obtained as initial noise statistics parameters for the Kalman filter. The prediction results are used as the observation benchmark for the Kalman filter. The state covariance of the multi-level state-space model is predicted in real time using the Kalman filter to generate dynamic monitoring thresholds.
8. The PTA polyester integrated equipment fault prediction method according to claim 7, characterized in that, The dynamic monitoring threshold is combined with the signal lag time of the located fault source node to establish a normal transmission interval. If the difference between the downstream observation value and the normal transmission interval exceeds the preset safety constraint boundary, a final warning signal is generated, including: The dynamic monitoring threshold is combined with the signal lag time of the fault source node to perform time phase compensation and establish a normal transmission interval. The absolute difference between the downstream observation and the normal transmission interval is calculated to obtain a transmission deviation measure; If the transmission deviation metric exceeds the preset safety constraint boundary, a final warning signal is generated.
9. A fault prediction system for integrated PTA polyester equipment, characterized in that, include: The data preprocessing and sequence decomposition module collects real-time operation data from upstream sections and downstream observations and performs time-series alignment. It then separates the basic trend components and periodic components from the time-aligned sequence to obtain the residual sequence. The upstream impact modeling and mapping module uses the Kalman filter algorithm to estimate the comprehensive mapping value of the upstream transmission effect of the residual sequence; The abnormal signal identification and feature extraction module determines that if the deviation between the comprehensive mapping value and the downstream observation value exceeds a preset allowable limit, the corresponding sampling point is a potential abnormal deviation signal, and the time-series feature vector of the abnormal deviation signal is obtained. The fault mode classification module calculates the Mahalanobis distance between the time-series feature vector and the historical normal operating condition statistical distribution, and outputs the classification result label based on the Mahalanobis distance. The early warning triggering module extracts the abnormal deviation subset marked as equipment area fault from the classification result labels, and calculates the duration and cumulative amplitude of the abnormal deviation subset. If the duration exceeds a preset time limit and the cumulative amplitude exceeds a preset intensity benchmark, an early warning triggering command is generated. If the early warning trigger command is triggered, the fault propagation path analysis module generates a cross-section dynamic association diagram based on the abnormal deviation subset, calculates the path weight between the affected node sequences in the cross-section dynamic association diagram, and locks the path with the highest path weight as the fault propagation path. The dynamic threshold generation module generates a multi-level state space model through the fault propagation path, and generates a dynamic monitoring threshold based on the prediction results of the multi-level state space model after Kalman filtering. The final early warning module combines the dynamic monitoring threshold with the signal lag time of the located fault source node to establish a normal transmission interval. If the difference between the downstream observation value and the normal transmission interval exceeds the preset safety constraint boundary, a final early warning signal is generated. The process of processing the residual sequence using a Kalman filter algorithm to obtain a comprehensive mapping value includes: analyzing the residual sequence, establishing a state vector containing real-time operation data of the upstream section, and generating a priori state estimate by combining it with a preset state transition matrix; calculating the Kalman gain coefficient based on the priori state estimate and the statistical characteristics of the observation noise, using the Kalman gain coefficient to correct the priori state estimate to obtain a posterior state estimate, and extracting a comprehensive mapping value from the posterior state estimate to characterize the impact of the real-time operation of the upstream section on the downstream observation value; In this process, based on the pairs of adjacent nodes along the fault propagation path, the monitoring signals of the two nodes under historical normal operating conditions are collected. A step excitation is applied to the upstream node and the downstream response curve is recorded. The least squares method is used to fit the first-order inertial plus pure time delay transfer function model, and the transfer gain and time delay constant are identified. The transfer gain and time delay constant are used as elements to construct a 2×2 transfer feature matrix. The temperature or pressure key state variables of each equipment node on the path are used as state variables. The state equation is established based on the gain and time constant in the transfer feature matrix of adjacent nodes to describe the transfer and delay of state variables from the previous node to the next node. All nodes are connected in series along the path to form a multi-level state space model.