Chemical process fault diagnosis method and system

Abstract

The invention discloses a chemical process fault diagnosis method and system, and relates to the technical field of chemical process fault diagnos.The scheme aims at the diagnosis bottleneck of catalyst inactivation type progressive faults, fine parameter offset is captured in real time through a dynamic reference model, weak signals are accumulated and amplified in combination with an attenuation weighting mechanism, and the fault diagnosis accuracy is improved. The problem that a traditional method is not sensitive to slow drifting, and consequently report omission is caused is solved, and a fault recognition window is remarkably advanced. The dynamic threshold value is updated in real time based on mobile statistics, and raw material fluctuation and sensor noise can be self-adapted; during working condition switching, the model is automatically reset and the threshold value is relaxed, so that false alarm triggered by parameter mutation is avoided, and the stability of production scheduling is guaranteed; the design that fault half-life period weight and moving window length are associated with an inactivation period is introduced, so that the model autonomously adapts to different catalyst characteristics.

Description

Technical Field

[0001] This invention relates to the field of chemical process fault diagnosis technology, and in particular to a chemical process fault diagnosis method and system. Background Technology

[0002] In high-temperature catalytic reactions such as ethylene cracking furnaces, progressive failures caused by catalyst deactivation are a typical challenge. These failures manifest as a daily decrease in catalyst activity of 0.1%-0.5%, leading to a slow shift in parameters such as reaction temperature and product yield, with initial changes even lower than the sensor noise level. Current mainstream solutions employ deep learning-based dynamic monitoring models, such as DA-CycleGAN and online LSTM, to predict equipment status through real-time data. These methods perform well in diagnosing sudden failures, but lack the ability to capture slowly evolving hidden risks.

[0003] Although recent research has introduced semi-supervised learning, such as the TE-DS model and transfer learning, to alleviate the problem of data scarcity, the models rely on significant feature shifts to trigger alarms. However, the initial signals of catalyst deactivation are masked by fluctuations in normal operating conditions, resulting in a lag in fault identification, often delayed by 3-7 days. In addition, false signals generated by fluctuations in raw material composition or sensor drift are easily misjudged as early deactivation, frequently triggering false alarms. For example, after a PTA plant applied the attention mechanism model, the number of false alarms was high each month, seriously interfering with production scheduling.

[0004] Some solutions employ multimodal fusion (such as combining infrared spectroscopy and temperature data) or adaptive threshold optimization. However, fixed threshold strategies struggle to distinguish between slow failures and noise. Transfer learning requires pre-storing a large amount of deactivation case data. In practice, the failure modes of similar catalysts differ significantly, leading to a sharp drop in cross-scenario generalization performance. Therefore, a chemical process fault diagnosis solution is urgently needed to address these issues. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a method and system for diagnosing faults in chemical processes, which solves the problems of high false negative rates for slow faults, frequent false alarms during operating condition switching, and reliance on a complete fault database in existing diagnostic methods.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a method for diagnosing faults in a chemical process, comprising, Step S1: Real-time acquisition of sensor timing data from the reaction equipment; Step S2: Dynamically generate a benchmark model based on an online incremental learning algorithm, and calculate the real-time residual between the sensor data and the benchmark model; Step S3: Input the residual into the circular memory pool module and generate a cumulative anomaly index by weighting it according to time decay. Step S4: Dynamically set alarm thresholds based on historical normal data distribution. When the cumulative abnormal index continuously exceeds the threshold, a progressive fault alarm is triggered. The real-time residual is calculated by the Euclidean distance between the sensor data and the baseline model prediction within the sliding window. The determination of continuous exceedance of the threshold requires that the cumulative abnormal index exceeds the threshold for a period of time that is greater than a preset duration.

[0008] As a preferred embodiment of the chemical process fault diagnosis method described in this invention, the baseline model employs incremental principal component analysis (PCA) or incremental partial least squares (PLS) algorithm. The incremental PCA / PLS algorithm retains the principal component weights of the preceding model during each update and adjusts the covariance matrix based on the new data window; When a raw material switching signal is detected, the initial value of the covariance matrix is ​​reset.

[0009] In a preferred embodiment of the chemical process fault diagnosis method of the present invention, the operation of the annular memory pool module includes: Store the most recent residual vector, apply weighting coefficients to historical residuals, and calculate the weighted norm sum as the cumulative outlier index; The storage depth is set according to the fault evolution cycle, and the weighting coefficient is negatively correlated with the fault decay rate.

[0010] As a preferred embodiment of the chemical process fault diagnosis method described in this invention, wherein: in the ring memory pool module, by analyzing the most recent... The residual vectors at each time step are exponentially weighted and accumulated, and the weighted norm sum is calculated to form the cumulative anomaly index at the current time step, including: Calculate the Euclidean norm of the residual vector in each sensor dimension, for time... The residual vector is defined as Then its Euclidean norm is: , in, Indicates time The residual vector, by the first The difference between the sensor measurements and the baseline model predictions constitutes the total value. Indicates time In the Residuals on each sensor This represents the number of sensors and the corresponding dimension of the residual vector. The Euclidean norm operator is used to measure the overall deviation of a vector. The time index represents the residual vector. Represents the sensor dimension index, from arrive Add them one by one; Introducing fault half-life Define the attenuation constant: , Therefore, in the sampling interval Within, the residual weight is defined as: , in, This represents the fault attenuation constant, used to characterize the exponential decay rate of the fault signal strength over time. This indicates that the natural logarithm function acts on the constant 2. This represents the fault half-life, which is the time required for the fault signal strength to decay to half of its original value. This represents the time interval between two consecutive residual samples. Represents the exponential decay coefficient, satisfying It is inversely proportional to the fault decay rate. This represents the natural exponential function, used to implement exponential time decay; Let the current time be The circular buffer has a storage depth of 1. Then only the interval is retained. The residual vector within the time frame is weighted according to the chronological order of the historical residuals, and then subjected to norm-weighted summation to obtain the cumulative outlier index: , in, Indicates the current time The cumulative anomaly index is a comprehensive measure used to characterize the degree of system anomalies over a recent period. This indicates the storage depth of the circular buffer, which is the number of historical residual vectors involved in the calculation. This represents the sequential index of the current accumulated item in the buffer, with a value range of 1. , Corresponding to the earliest time The residual, Corresponding to the current moment The residual, Indicates the first The time decay weights corresponding to each historical residual are exponentially given by: Reflecting the distance from the current moment, Indicates the first in the buffer At that moment The residual vector, This represents the Euclidean norm of the residual vector at that moment, used to quantify the magnitude of the deviation from the baseline model. This represents the index at the current time, and the exponentiation operation is based on the definition in the previous step. ; After completing the above calculations, the ring memory pool module will As the cumulative anomaly index at that moment, it is used for subsequent threshold determination. The circular buffer maintains the most recent anomaly index through a cyclic overlay method. The residual at each moment.

[0011] In a preferred embodiment of the chemical process fault diagnosis method of the present invention, the alarm threshold is determined by means of: The moving mean and standard deviation of the cumulative abnormal index under normal historical operating conditions were statistically analyzed. Set the threshold to the moving average plus a preset multiple of the standard deviation; The length of the moving average window is proportional to 1 / 10 of the typical deactivation cycle of the catalyst.

[0012] In a preferred embodiment of the chemical process fault diagnosis method of the present invention, before determining the dynamic alarm threshold, a cumulative abnormality index sequence under historical normal operating conditions is used. Calculate the moving mean and standard deviation, including: Introducing parameters This reflects the impact of the typical catalyst deactivation cycle on the statistical window. This represents the typical deactivation period of the catalyst, expressed in time, with the sampling interval denoted as . The window length for the moving mean and standard deviation is denoted as: , in, This represents the typical deactivation cycle of the catalyst and is directly related to the catalyst deactivation rate. Indicates two consecutive cumulative abnormality indices The time interval between samples, This indicates a floor operation to ensure the window length is an integer. The cumulative number of outlier index samples included in the moving statistical window is proportional to the catalyst deactivation cycle. Indicates time The cumulative abnormal index; Assuming up to the current time Calculate time The moving average of the cumulative anomaly index sequence collected during the historical normal operating conditions is: , of which all The corresponding times all meet normal operating conditions and do not include abnormal data that has triggered alarms; For the Take the nearest sampling point. Within a time interval, i.e. The cumulative anomaly index is used to construct a moving statistical set, and the moving mean is defined as: , in, Indicates time The moving average, To move the window length, This is the relative offset index within the window, with a value range of... ,when Time corresponds to the current moment , Indicates time The cumulative abnormality index, denominator The total number of samples within the window ensures that the mean is correctly normalized. The moving variance is defined based on the same batch of samples within the moving window. for: , The moving standard deviation is then: , in, Indicates time The moving standard deviation, For moving variance, The denominator of the unbiased sample variance is... and This measure indicates the degree of deviation of each sample within a window from the moving mean; the summation actually covers the most recent... The cumulative anomaly index samples at each time point are used, and the square and square root operations are used to obtain the standard deviation, which reflects the dispersion trend of the anomaly index within that window.

[0013] As a preferred embodiment of the chemical process fault diagnosis method of the present invention, it further includes: When a raw material component switching signal is detected, the initial parameters of the baseline model are automatically reset and the alarm threshold is relaxed. The degree of relaxation of the alarm threshold is positively correlated with the difference in operating parameters before and after the switch; The reset of the baseline model initial parameters includes clearing the historical data of the covariance matrix.

[0014] Secondly, the present invention provides a chemical process fault diagnosis system, comprising, include: The data acquisition unit acquires real-time data from the temperature, pressure, and flow sensors of the reaction equipment. The processing module is configured to execute chemical process fault diagnosis methods; The alarm unit triggers a tiered alarm based on the comparison between the cumulative anomaly index output by the processing module and the threshold. The processing module is deployed on an edge computing device and has a built-in real-time data stream processing engine. The tiered alarm system includes early warning and emergency alarms.

[0015] As a preferred embodiment of the chemical process fault diagnosis system of the present invention, the processing module includes: The baseline model generation submodule employs an online incremental learning architecture. The ring-shaped memory pool submodule has a configurable attenuation coefficient and storage depth; The dynamic threshold calculation submodule connects to the historical database to update statistical parameters. The circular memory pool submodule is implemented using a circular buffer, and the storage depth can be dynamically adjusted. The dynamic threshold calculation submodule includes a sliding window statistician and a standard deviation amplifier.

[0016] In a preferred embodiment of the chemical process fault diagnosis system described in this invention, the alarm unit synchronously outputs a fault evolution trend graph when an alarm is triggered. The trend chart overlays the cumulative abnormality index curve, dynamic threshold line, and historical normal range.

[0017] The beneficial effects of this invention are: This solution addresses the diagnostic bottleneck of progressive catalyst deactivation faults. It uses a dynamic benchmark model to capture subtle parameter shifts in real time, and combines this with a decay-weighted mechanism to amplify weak signals, solving the problem of missed detections caused by the insensitivity of traditional methods to slow drifts and significantly advancing the fault identification window. The dynamic threshold is updated in real time based on moving statistics, adapting to raw material fluctuations and sensor noise. During operating condition changes, the model is automatically reset and the threshold is relaxed to avoid false alarms triggered by parameter mutations, ensuring the stability of production scheduling. The design incorporates fault half-decay weighting and moving window length to correlate the deactivation cycle, enabling the model to autonomously adapt to different catalyst characteristics. Edge computing deployment and a tiered alarm mechanism reduce reliance on expert experience and improve on-site operability. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the chemical process fault diagnosis method in Example 1.

[0020] Figure 2 This is a schematic diagram of the framework of the chemical process fault diagnosis system in Example 1. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Example 1, referring to Figure 1 and Figure 2 This embodiment provides a method for diagnosing faults in a chemical process, including the following steps: Step S1: Real-time acquisition of sensor timing data from the reaction equipment; Step S2: Dynamically generate a benchmark model based on an online incremental learning algorithm, and calculate the real-time residual between the sensor data and the benchmark model; Step S3: Input the residual into the circular memory pool module and generate a cumulative anomaly index by weighting it according to time decay. Step S4: Dynamically set alarm thresholds based on historical normal data distribution. When the cumulative abnormal index continuously exceeds the threshold, a progressive fault alarm is triggered. Real-time residuals are calculated by the Euclidean distance between sensor data and baseline model predictions within a sliding window. The determination of continuous exceedance of the threshold requires that the cumulative abnormal index exceeds the threshold for a period of time that is greater than a preset duration. The baseline model employs either incremental principal component analysis (PCA) or incremental partial least squares (PLS) algorithm. The incremental PCA / PLS algorithm retains the principal component weights of the preceding model during each update and adjusts the covariance matrix based on the new data window; When a raw material switching signal is detected, reset the initial value of the covariance matrix; The operation of the ring memory pool module includes: Store the most recent residual vector, apply weighting coefficients to historical residuals, and calculate the weighted norm sum as the cumulative outlier index; The storage depth is set according to the fault evolution cycle, and the weighting coefficient is negatively correlated with the fault decay rate. In the circular memory pool module, by checking the most recent The residual vectors at each time step are exponentially weighted and accumulated, and the weighted norm sum is calculated to form the cumulative anomaly index at the current time step, including: Calculate the Euclidean norm of the residual vector in each sensor dimension, for time... The residual vector is defined as Then its Euclidean norm is: , in, Indicates time The residual vector, by the first The difference between the sensor measurements and the baseline model predictions constitutes the total value. Indicates time In the Residuals on each sensor This represents the number of sensors and the corresponding dimension of the residual vector. The Euclidean norm operator is used to measure the overall deviation of a vector. The time index represents the residual vector. Represents the sensor dimension index, from arrive Add them one by one; Introducing fault half-life Define the attenuation constant: , Therefore, in the sampling interval Within, the residual weight is defined as: , in, This represents the fault attenuation constant, used to characterize the exponential decay rate of the fault signal strength over time. This indicates that the natural logarithm function acts on the constant 2. This represents the fault half-life, which is the time required for the fault signal strength to decay to half of its original value. This represents the time interval between two consecutive residual samples. Represents the exponential decay coefficient, satisfying It is inversely proportional to the fault decay rate. This represents the natural exponential function, used to implement exponential time decay; Let the current time be The circular buffer has a storage depth of 1. Then only the interval is retained. The residual vector within the time frame is weighted according to the chronological order of the historical residuals, and then subjected to norm-weighted summation to obtain the cumulative outlier index: , in, Indicates the current time The cumulative anomaly index is a comprehensive measure used to characterize the degree of system anomalies over a recent period. This indicates the storage depth of the circular buffer, which is the number of historical residual vectors involved in the calculation. This represents the sequential index of the current accumulated item in the buffer, with a value range of 1. , Corresponding to the earliest time The residual, Corresponding to the current moment The residual, Indicates the first The time decay weights corresponding to each historical residual are exponentially given by: Reflecting the distance from the current moment, Indicates the first in the buffer At that moment The residual vector, This represents the Euclidean norm of the residual vector at that moment, used to quantify the magnitude of the deviation from the baseline model. This represents the index at the current time, and the exponentiation operation is based on the definition in the previous step. ; After completing the above calculations, the ring memory pool module will As the cumulative anomaly index at that moment, it is used for subsequent threshold determination. The circular buffer maintains the most recent anomaly index through a cyclic overlay method. The residuals at each time point are used to avoid storing irrelevant early data while ensuring that new data is incorporated into the calculation in a timely manner. Specifically, this calculation method combines the residual norm with time weights, taking into account both the direct reflection of the residual magnitude on the anomaly and the time decay characteristics, to accommodate different fault evolution rates, and utilizes the fault half-life. With sampling interval Constructing the exponential decay coefficient This allows recent residuals to have a higher weight in the cumulative anomaly index, while earlier residuals gradually decay, avoiding excessive influence of historical interference on current anomaly judgment. The circular buffer mechanism ensures a constant storage size, retaining only valuable latest information, and forms a closed loop in the form of weighted norm sum, mapping the dynamic evolution process to continuous and comparable values ​​in real time. It has the ability to respond quickly to sudden anomalies, and can gradually amplify the signal in the early stage of the fault, improving the fault detection sensitivity. At the same time, it has a certain tolerance for short-term fluctuations in the system, reducing the risk of false alarms. The methods for determining dynamic alarm thresholds include: The moving mean and standard deviation of the cumulative abnormal index under normal historical operating conditions were statistically analyzed. Set the threshold to the moving average plus M times the standard deviation, where M>0; The length of the moving average window is proportional to 1 / 10 of the typical deactivation cycle of the catalyst; Before determining the dynamic alarm threshold, a cumulative anomaly index sequence under historical normal operating conditions is used. Calculate the moving mean and standard deviation, including: Introducing parameters This reflects the impact of the typical catalyst deactivation cycle on the statistical window. This represents the typical deactivation period of the catalyst, expressed in time, with the sampling interval denoted as . The window length for the moving mean and standard deviation is denoted as: , in, This represents the typical deactivation cycle of the catalyst and is directly related to the catalyst deactivation rate. Indicates two consecutive cumulative abnormality indices The time interval between samples, This indicates a floor operation to ensure the window length is an integer. The cumulative number of outlier index samples included in the moving statistical window is proportional to the catalyst deactivation cycle. Indicates time The cumulative abnormal index, same Just use subscripts It refers to the cumulative anomaly index at any given time. Assuming up to the current time Calculate time The moving average of the cumulative anomaly index sequence collected during the historical normal operating conditions is: , of which all The corresponding times all meet normal operating conditions and do not include abnormal data that has triggered alarms; For the Take the nearest sampling point. Within a time interval, i.e. The cumulative anomaly index is used to construct a moving statistical set, and the moving mean is defined as: , in, Indicates time The moving average, To move the window length, This is the relative offset index within the window, with a value range of... ,when Time corresponds to the current moment , Indicates time The cumulative abnormality index, denominator The total number of samples within the window ensures that the mean is correctly normalized. The moving variance is defined based on the same batch of samples within the moving window. for: , The moving standard deviation is then: , in, Indicates time The moving standard deviation, For moving variance, The denominator of the unbiased sample variance is... and This measure indicates the degree of deviation of each sample within a window from the moving mean; the summation actually covers the most recent... The cumulative anomaly index samples at each time point are used to obtain the standard deviation through square and square root operations, which reflects the dispersion trend of the anomaly index within that window. After completing the above steps, you will receive real-time updates. and Used to build dynamic alarm thresholds (For example ,in It is a multiple of the standard deviation. It is a positive scalar with a standard deviation multiple, used to adjust the sensitivity of the alarm threshold based on the normal distribution statistics. Its value needs to be calibrated according to the actual experience and fault tolerance requirements of the chemical plant. Subsequently, this real-time statistic can be used to determine whether the cumulative abnormal index exceeds the limit in the threshold determination. Specifically, this step utilizes a moving window mechanism to integrate catalyst deactivation cycle information while maintaining statistical real-time performance, enabling dynamic tracking of the cumulative anomaly index data distribution. The window length... With typical catalyst deactivation cycle Proportional to this, the window selection avoids prematurely introducing noise outside the deactivation cycle while also aligning with the actual catalyst performance decay process, using a moving average. It depicts the central trend of the recent cumulative outlier index, while the moving standard deviation... This reflects the fluctuation range. Combining these two factors allows for dynamic adjustment of the threshold, moving synchronously with changes in normal operating conditions. This reduces the risk of false alarms caused by slow system drift or periodic fluctuations. Within the window, the unbiased variance denominator is used. Calculate the standard deviation to ensure the accuracy of statistical estimation, and also consider the sample size. By rounding up to avoid loss of precision, this method combines robustness and flexibility, and is applicable to chemical processes with different deactivation cycles. Chemical process fault diagnosis methods also include: When a raw material component switching signal is detected, the initial parameters of the baseline model are automatically reset, and the alarm threshold for the previous T minutes is relaxed. The degree of relaxation of alarm thresholds is positively correlated with the difference in operating parameters before and after the switch. Resetting the initial parameters of the baseline model includes clearing historical data for the covariance matrix.

[0025] This embodiment also provides a chemical process fault diagnosis system, including: The data acquisition unit acquires real-time data from the temperature, pressure, and flow sensors of the reaction equipment. The processing module is configured to execute chemical process fault diagnosis methods; The alarm unit triggers a tiered alarm based on the comparison between the cumulative anomaly index output by the processing module and the threshold. The processing module is deployed on edge computing devices and has a built-in real-time data stream processing engine. The tiered alarm system includes early warning (threshold continuously exceeded for <5 minutes) and emergency alarm (threshold continuously exceeded for >30 minutes). The processing module includes: The baseline model generation submodule employs an online incremental learning architecture. The ring-shaped memory pool submodule has a configurable attenuation coefficient and storage depth; The dynamic threshold calculation submodule connects to the historical database to update statistical parameters. The circular memory pool submodule is implemented using a circular buffer, and the storage depth can be dynamically adjusted. The dynamic threshold calculation submodule includes a sliding window statistician and a standard deviation amplifier; When an alarm is triggered, the alarm unit simultaneously outputs a fault evolution trend graph: The trend chart overlay displays the cumulative abnormality index curve, dynamic threshold line, and historical normal range. The historical normal range is indicated by a gray semi-transparent area, representing the range of the moving mean ± 2 standard deviations.

[0026] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for diagnosing faults in a chemical process, characterized in that, include, Step S1: Real-time acquisition of sensor timing data from the reaction equipment; Step S2: Dynamically generate a benchmark model based on an online incremental learning algorithm, and calculate the real-time residual between the sensor data and the benchmark model; Step S3: Input the residual into the circular memory pool module and generate a cumulative anomaly index by weighting it according to time decay. Step S4: Dynamically set alarm thresholds based on historical normal data distribution. When the cumulative abnormal index continuously exceeds the threshold, a progressive fault alarm is triggered. The real-time residual is calculated by the Euclidean distance between the sensor data and the baseline model prediction within the sliding window. The determination of continuous exceedance of the threshold requires that the cumulative abnormal index exceeds the threshold for a period of time that is greater than a preset duration.

2. The chemical process fault diagnosis method as described in claim 1, characterized in that, The baseline model employs incremental principal component analysis (PCA) or incremental partial least squares (PLS) algorithm. The incremental PCA / PLS algorithm retains the principal component weights of the preceding model during each update and adjusts the covariance matrix based on the new data window; When a raw material switching signal is detected, the initial value of the covariance matrix is ​​reset.

3. The chemical process fault diagnosis method as described in claim 1, characterized in that, The operation of the ring memory pool module includes: Store the most recent residual vector, apply weighting coefficients to historical residuals, and calculate the weighted norm sum as the cumulative outlier index; The storage depth is set according to the fault evolution cycle, and the weighting coefficient is negatively correlated with the fault decay rate.

4. The chemical process fault diagnosis method as described in claim 3, characterized in that, In the ring memory pool module, by analyzing the most recent... The residual vectors at each time step are exponentially weighted and accumulated, and the weighted norm sum is calculated to form the cumulative anomaly index at the current time step, including: Calculate the Euclidean norm of the residual vector in each sensor dimension, for time... The residual vector is defined as Then its Euclidean norm is: ， in, Indicates time The residual vector, by the first The difference between the sensor measurements and the baseline model predictions constitutes the total value. Indicates time In the The residuals on each sensor This represents the number of sensors and the corresponding dimension of the residual vector. The Euclidean norm operator is used to measure the overall deviation of a vector. The time index represents the residual vector. Represents the sensor dimension index, from arrive Add them one by one; Introducing a fault half-life Define the attenuation constant: ， Therefore, in the sampling interval Within, the residual weight is defined as: ， in, This represents the fault attenuation constant, used to characterize the exponential decay rate of the fault signal strength over time. This indicates that the natural logarithm function acts on the constant 2. This represents the fault half-life, which is the time required for the fault signal strength to decay to half of its original value. This represents the time interval between two consecutive residual samples. Represents the exponential decay coefficient, satisfying It is inversely proportional to the fault decay rate. This represents the natural exponential function, used to implement exponential time decay; Let the current time be The circular buffer has a storage depth of 1. Then only the interval is retained. The residual vector within the time frame is weighted according to the chronological order of the historical residuals, and then subjected to norm-weighted summation to obtain the cumulative outlier index: ， in, Indicates the current time The cumulative anomaly index is a comprehensive measure used to characterize the degree of system anomalies over a recent period. This indicates the storage depth of the circular buffer, which is the number of historical residual vectors involved in the calculation. This represents the sequential index of the current accumulated item in the buffer, with a value range of 1. , Corresponding to the earliest time The residual, Corresponding to the current moment The residual, Indicates the first The time decay weights corresponding to each historical residual are exponentially given by: Reflecting the distance from the current moment, Indicates the first in the buffer At that moment The residual vector, This represents the Euclidean norm of the residual vector at that moment, used to quantify the magnitude of the deviation from the baseline model. This represents the index at the current time, and the exponentiation operation is based on the definition in the previous step. ; After completing the above calculations, the ring memory pool module will As the cumulative anomaly index at that moment, it is used for subsequent threshold determination. The circular buffer maintains the most recent anomaly index through a cyclic overlay method. The residual at each moment.

5. The chemical process fault diagnosis method as described in claim 1, characterized in that, The alarm threshold is determined in the following ways: The moving mean and standard deviation of the cumulative abnormal index under normal historical operating conditions were statistically analyzed. Set the threshold to the moving average plus a preset multiple of the standard deviation; The length of the moving average window is proportional to 1 / 10 of the typical deactivation cycle of the catalyst.

6. The chemical process fault diagnosis method as described in claim 5, characterized in that, Before determining the dynamic alarm threshold, it is based on the cumulative abnormal index sequence under historical normal operating conditions. Calculate the moving mean and standard deviation, including: Introducing parameters This reflects the impact of the typical catalyst deactivation cycle on the statistical window. This represents the typical deactivation period of the catalyst, expressed in time, with the sampling interval denoted as . The window length for the moving mean and standard deviation is denoted as: ， in, This represents the typical deactivation cycle of the catalyst and is directly related to the catalyst deactivation rate. Indicates two consecutive cumulative abnormality indices The time interval between samples, This indicates a floor operation to ensure the window length is an integer. The cumulative number of outlier index samples included in the moving statistical window is proportional to the catalyst deactivation cycle. Indicates time The cumulative abnormal index; Assuming up to the current time Calculate time The moving average of the cumulative anomaly index sequence collected during the historical normal operating conditions is: , of which all The corresponding times all meet normal operating conditions and do not include abnormal data that has triggered alarms; For the Take the nearest sampling point. Within a time interval, i.e. The cumulative anomaly index is used to construct a moving statistical set, and the moving mean is defined as: ， in, Indicates time The moving average, To move the window length, This is the relative offset index within the window, with a value range of... ,when Time corresponds to the current moment , Indicates time The cumulative abnormality index, denominator The total number of samples within the window ensures that the mean is correctly normalized. The moving variance is defined based on the same batch of samples within the moving window. for: ， The moving standard deviation is then: ， in, Indicates time The moving standard deviation, For moving variance, The denominator of the unbiased sample variance is... and This measure indicates the degree of deviation of each sample within a window from the moving mean; the summation actually covers the most recent... The cumulative anomaly index samples at each time point are used, and the square and square root operations are used to obtain the standard deviation, which reflects the dispersion trend of the anomaly index within that window.

7. The chemical process fault diagnosis method as described in claim 1, characterized in that, Also includes: When a raw material component switching signal is detected, the initial parameters of the baseline model are automatically reset and the alarm threshold is relaxed. The degree of relaxation of the alarm threshold is positively correlated with the difference in operating parameters before and after the switch; The reset of the baseline model initial parameters includes clearing the historical data of the covariance matrix.

8. A chemical process fault diagnosis system, based on the chemical process fault diagnosis method according to any one of claims 1 to 7, characterized in that, include: The data acquisition unit acquires real-time data from the temperature, pressure, and flow sensors of the reaction equipment. The processing module is configured to execute chemical process fault diagnosis methods; The alarm unit triggers a tiered alarm based on the comparison between the cumulative anomaly index output by the processing module and the threshold. The processing module is deployed on an edge computing device and has a built-in real-time data stream processing engine. The tiered alarm system includes early warning and emergency alarms.

9. A chemical process fault diagnosis system as described in claim 8, characterized in that, The processing module includes: The baseline model generation submodule employs an online incremental learning architecture. The ring-shaped memory pool submodule has a configurable attenuation coefficient and storage depth; The dynamic threshold calculation submodule connects to the historical database to update statistical parameters. The circular memory pool submodule is implemented using a circular buffer, and the storage depth can be dynamically adjusted. The dynamic threshold calculation submodule includes a sliding window statistician and a standard deviation amplifier.

10. A chemical process fault diagnosis system as described in claim 8, characterized in that, When the alarm is triggered, the alarm unit simultaneously outputs a fault evolution trend graph: The trend chart overlays the cumulative abnormality index curve, dynamic threshold line, and historical normal range.

Images