Process measuring point trend diagnosis method and system based on subspace smooth incremental spectrum analysis

Abstract

The invention relates to a process measuring point trend diagnosis method and system based on subspace smooth increment spectrum analysis, and the method comprises the steps: carrying out the increment processing of a process measuring point historical data sequence, mapping the process measuring point historical data sequence into an increment subspace matrix, and carrying out the smoothing processing of each subspace vector in the increment subspace matrix; performing frequency domain projection analysis on all subspace vectors in the increment subspace matrix to obtain long-time increment power spectrum statistical characteristics of the process measurement points; performing short-segment data spectrum estimation based on the constructed simulation online subspace to obtain an error allowable maximum multiple and a corresponding short-segment data window length; and performing online diagnosis on the real-time process measuring point data according to the configuration condition of the process measuring point in combination with the determined long-time increment power spectrum statistical characteristics of the process measuring point, the maximum error allowable multiple and the corresponding short-section data window length to obtain a process measuring point trend diagnosis result. The method can be widely applied to the field of fault diagnosis of the process industry.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis in the process industry, and more specifically to a method and system for diagnosing process measurement point trends based on subspace smoothing incremental spectrum analysis. Background Technology

[0002] Process industries are pillar industries of the national economy and important engines driving the operation of modern industrial systems. In large-scale refining and fine chemical production processes, the continuity, scale, and automation of equipment are increasing, and complex physicochemical reactions accompany the production process. As core indicators reflecting the operating status of the equipment, process monitoring points include traditional process parameters (such as temperature, pressure, liquid level, and flow rate), as well as equipment and environmental parameters. The stability of these key monitoring points directly affects production efficiency, product quality, and energy consumption. Once these monitoring points show abnormal trends, it is often an early sign of equipment failure, control loop failure, or deterioration of process conditions. If these issues are not detected and addressed in time, they can lead to product defects and increased energy consumption, or even trigger a chain reaction, causing unplanned shutdowns of the equipment and serious safety and environmental accidents, resulting in incalculable losses to personnel, equipment, and the environment.

[0003] Trend diagnosis of process monitoring points is a crucial step in ensuring the stable operation of chemical processes. In actual production, processes exhibit complex, non-stationary and non-linear characteristics due to multiple factors such as fluctuations in raw material properties, equipment aging, and environmental interference. Currently, on-site management still relies primarily on fixed threshold alarms from distributed control systems (DCS) and operator experience-based inspections. This passive management model struggles to capture subtle early anomalies in process monitoring points and is easily affected by adjustments to operating conditions, failing to meet the urgent needs of modern process industries for digital and intelligent operation and maintenance.

[0004] Currently, scholars both domestically and internationally have conducted extensive research on process measurement point trend diagnosis methods, which can be broadly categorized into three types: analytical model-based, qualitative empirical knowledge-based, and data-driven methods. Due to the complexity of chemical process mechanisms, it is difficult to establish accurate mathematical models, and expert experience is hard to acquire. Therefore, data-driven methods are widely favored because they only require equipment operating data. Within the data-driven framework, methods are mainly divided into those based on multivariate statistical analysis, artificial intelligence, and signal processing.

[0005] Multivariate statistical analysis-based methods mine historical data of process variables and use projection techniques to extract statistical features for diagnostic purposes. Key algorithms include Principal Component Analysis (PCA), Partial Least Squares (PLS), and Independent Principal Component Analysis (ICA). These methods utilize statistical indicators to diagnose process states and effectively handle correlations between variables. However, these methods are primarily based on the assumption of static data distribution, often neglecting the dynamic time-series information and frequency domain energy distribution characteristics inherent in the data sequence. For common non-stationary oscillations, DC component drift, or disturbances at specific frequencies in chemical processes, simple statistical projection methods are insufficient to reveal their physical nature and have poor adaptability to nonlinear conditions.

[0006] Artificial intelligence-based methods primarily utilize machine learning algorithms to establish nonlinear mappings between fault characteristics and categories. These mainly include Artificial Neural Networks (ANNs), Support Vector Machines (SVMs), and fuzzy logic. However, these methods have significant limitations in practical engineering applications: firstly, they are typically "black box" models with poor physical interpretability, making it difficult for maintenance personnel to understand the underlying mechanisms of alarms; secondly, these methods heavily rely on massive amounts of well-labeled fault sample data for training. In actual chemical production, fault samples are often scarce, limiting the generalization ability and practical application of these methods.

[0007] Signal processing-based methods extract time-domain or frequency-domain features from measured signals using various transformation techniques, primarily including spectral analysis and wavelet transform. These methods, by analyzing the power spectral density (PSD) or time-frequency distribution of the signal, can keenly capture anomalies from an energy perspective, possessing clear physical meaning. While wavelet transform has excellent time-frequency localization characteristics, enabling simultaneous analysis of signal time and frequency details, it faces significant engineering application bottlenecks in industrial online diagnostics: firstly, wavelet transform has high computational complexity, resulting in excessive online computational load; secondly, selecting suitable wavelet bases for different parameters is extremely difficult, increasing on-site deployment and debugging costs. In contrast, PSD offers advantages such as low computational complexity, mature and efficient algorithms, and clear physical meaning, making it more suitable for real-time analysis of large-scale industrial data. To obtain high-precision spectral estimation, long data lengths are often required during offline training; however, in online diagnostics, short data windows must be used to achieve real-time alarm response. Shortening the data length leads to a decrease in spectral resolution and an increase in spectral estimation errors, causing the fluctuation characteristics of short-term data to be falsely amplified. Summary of the Invention

[0008] To address the aforementioned problems, the purpose of this invention is to provide a method and system for diagnosing process measurement point trends based on subspace smoothing incremental spectral analysis. This method constructs an integrated framework for offline training and online diagnosis: in the offline training phase, long-segment subspaces are constructed using long-segment incremental data to mine the statistical characteristics of the incremental power spectrum of measurement points; in the online diagnosis phase, for short-segment incremental data with real-time requirements, a "maximum allowable error multiple" mechanism is introduced to quantify the spectral estimation error caused by the short-segment incremental data, thereby achieving accurate real-time diagnosis of abnormal trends in process measurement points.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for diagnosing the trend of process measurement points based on subspace smoothing incremental spectral analysis, comprising the following steps: The historical data sequence of process measurement points is incrementally processed and mapped into an incremental subspace matrix, and the vectors of each subspace within the incremental subspace matrix are smoothed. Frequency domain projection analysis is performed on all subspace vectors within the incremental subspace matrix to obtain the statistical characteristics of the long-time incremental power spectrum at the process measurement points. Short-segment data spectrum estimation is performed based on the constructed simulated online subspace to obtain the maximum allowable error factor and the corresponding short-segment data window length; Based on the configuration of process measurement points, combined with the determined statistical characteristics of incremental power spectrum of measurement points, the maximum allowable error multiple, and the corresponding short data window length, online diagnosis is performed on real-time process measurement point data to obtain process measurement point trend diagnosis results.

[0010] Furthermore, the incremental processing of the historical data sequence of process measurement points and mapping it into an incremental subspace matrix, followed by smoothing of each subspace vector within the incremental subspace matrix, includes: Incremental processing is performed on the historical data sequence of process measurement points to obtain an incremental data sequence; Based on incremental data sequences, an incremental subspace matrix is ​​constructed using segmented mapping techniques. An iterative statistical smoothing technique is used to smooth all subspace vectors in the incremental subspace matrix.

[0011] Furthermore, the iterative statistical smoothing technique is used to smooth all subspace vectors in the incremental subspace matrix, including: ① For the incremental subspace matrix The first in Subspace vectors The mean and standard deviation are calculated; ②For the first Subspace vectors Each data point is traversed, and each data point is judged as an outlier using a pre-constructed outlier criterion based on statistical distribution. ③ Data points identified as outliers Perform median filtering for smooth replacement; ④ Use the replaced new data points to analyze the first... Subspace vectors Update to obtain the first smooth subspace vectors ; ⑤ Repeat steps ① to ④ above until the incremental subspace matrix is ​​obtained. We obtain the smooth incremental subspace matrix by continuing until no element satisfies the outlier criterion. .

[0012] Furthermore, the frequency domain projection analysis of all subspace vectors within the incremental subspace matrix to obtain the long-time incremental power spectrum statistical characteristics of the process measurement points includes: Regarding the first smooth subspace vectors Spectral estimation was performed using the Wiener-Khinchin theorem to obtain the smoothed incremental subspace matrix. power spectral density matrix ; Establish a frequency domain "statistical benchmark" for process measurement points, and analyze the power spectral density matrix. Statistical characteristic calculations are performed to obtain the long-term incremental power spectrum statistical characteristics of the process measurement points, including the baseline mean and baseline standard deviation.

[0013] Furthermore, the short-segment data spectrum estimation based on the constructed simulated online subspace, to obtain the maximum allowable error factor and the corresponding short-segment data window length, includes: Based on the short data window set in the actual online diagnosis, a simulated online incremental subspace matrix is ​​constructed and smoothed to obtain a smoothed simulated online incremental subspace matrix. Spectral estimation is performed on each simulated online incremental subspace vector in the smoothed simulated online incremental subspace matrix to obtain the short-time standard deviation; The obtained short-time standard deviation is compared point by point with the benchmark standard deviation to obtain the error amplification sequence at different frequency points; The maximum value in the error amplification sequence among all frequency points is selected as the global maximum allowable error factor. The obtained maximum permissible multiple of error is verified until a valid maximum permissible multiple and the corresponding short data window length are obtained.

[0014] Furthermore, based on the configuration of process measurement points, and combined with the long-term incremental power spectrum statistical characteristics, maximum allowable error multiple, and corresponding short-segment data window length of the determined process measurement points, online diagnosis is performed on the real-time process measurement point data to obtain process measurement point trend diagnosis results, including: Based on the configuration of process measurement points, anomalies in long-term straight-line data of real-time process measurement points are determined. Based on the statistical characteristics of the long-term incremental power spectrum of the determined process measurement points, the maximum allowable error multiple, and the corresponding short-segment data window length, trend anomalies are determined in the real-time process measurement point data. When an abnormal trend occurs, the local trend of the real-time process measurement data is extracted to determine the short-term sudden change anomaly. The diagnostic status code is output based on the anomaly determination result, and the corresponding alarm signal is triggered.

[0015] Furthermore, the step of determining long-term straight-line anomalies in real-time process measurement point data based on the process measurement point configuration includes: Real-time process measurement point data sequence Perform moving average filtering to obtain a smoothed sequence. ; Traversing the smooth sequence Count the non-decreasing values. Non-incrementing count value ; count value and The system is compared with a preset threshold, and the comparison result is used to determine whether there is an abnormality of walking in a straight line for a long time.

[0016] Furthermore, the step of determining trend anomalies in real-time process measurement point data based on the statistical characteristics of the long-term incremental power spectrum of the determined process measurement points, the maximum allowable error multiple, and the corresponding short-segment data window length includes: Based on real-time process measurement data Construct an online real-time incremental subspace matrix, and perform spectral estimation on each subspace vector of the online real-time incremental subspace matrix to obtain the power spectral density sequence; Using the maximum allowable error factor as the initial reference benchmark for the online diagnostic threshold, the power spectral density sequence of each subspace vector is judged, and the trend anomaly is determined based on the judgment result.

[0017] Furthermore, when an anomaly occurs, the local trend of the real-time process measurement data is extracted, and a short-term abrupt change anomaly is determined, including: Real-time process measurement data The end window is truncated, and the median of the beginning segment of the end window is extracted. and the median of the end segment ; Based on the median of the initial segment and the median of the end segment Calculate the normalized rate of change The short-term mutation anomalies are determined based on the calculation results.

[0018] Secondly, the present invention provides a process measurement point trend diagnosis system based on subspace smoothing incremental spectrum analysis, comprising: The incremental subspace construction and vector sequence smoothing module is used to perform incremental processing on the historical data sequence of process measurement points and map it into an incremental subspace matrix, and to smooth the vectors of each subspace within the incremental subspace matrix. The subspace increment spectrum statistical feature mining module is used to perform frequency domain projection analysis on all subspace vectors within the increment subspace matrix to obtain the long-time increment power spectrum statistical features of the process measurement points. The maximum allowable error factor calculation module is used to perform short-segment data spectrum estimation based on the constructed simulated online subspace, and obtain the maximum allowable error factor and the corresponding short-segment data window length; The online trend diagnosis module is used to perform online diagnosis on real-time process measurement point data based on the process measurement point configuration, combined with the determined statistical characteristics of the incremental power spectrum of the measurement points, the maximum allowable error multiple, and the corresponding short data window length, to obtain the process measurement point trend diagnosis results.

[0019] The present invention has the following advantages due to the adoption of the above technical solutions: 1. This invention is based on the subspace smoothing incremental spectrum analysis - maximum allowable error multiple - online threshold determination process, and performs trend diagnosis on non-stationary fluctuations of key measurement points in chemical processes, which solves the contradiction between spectrum estimation accuracy and data window length.

[0020] 2. This invention ensures the purity of the data through iterative statistical filtering of the incremental subspace vector, and effectively mines the frequency domain features of the data through a subspace spectral statistical integration strategy. 3. This invention quantifies and compensates for the physical errors caused by short data windows by using the maximum allowable error multiple.

[0021] This invention boasts advantages such as strong noise resistance, reduced manual costs due to threshold adjustment, no need for extensive fault sample training, and clear physical meaning. It is suitable for real-time monitoring of measurement points in large-scale industrial processes and possesses excellent commercial value and application prospects. Therefore, this invention can be widely applied in the field of fault diagnosis in process industries. Attached Figure Description

[0022] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. In the drawings: Figure 1 This is an overall flowchart of the process measurement point trend diagnosis method based on subspace smoothing incremental spectrum analysis provided in the embodiments of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] In process industries, the stability of critical process measurement points is crucial for ensuring the safe and efficient operation of equipment. Abnormal trends can lead to significant economic losses and safety accidents. Therefore, accurately diagnosing abnormal trends at process measurement points is essential. However, real-world industrial data often exhibits non-stationary trends and impulse noise, and traditional frequency domain methods suffer from an inherent contradiction between "real-time performance" and "accuracy": online diagnostics requires short data windows, but this significantly increases spectral estimation errors, rendering traditional methods based on fixed thresholds ineffective in identifying anomalies. Furthermore, the dynamic and variable nature of chemical production processes, the scarcity of fault samples, and the limitations imposed by data length on spectral estimation accuracy make it difficult for any single existing technology to meet practical engineering needs.

[0026] Therefore, in some embodiments of the present invention, a process measurement point trend diagnosis method based on subspace smoothing incremental spectral analysis is provided, which has unique advantages in resolving the contradiction between real-time performance in online diagnosis and accuracy in spectral estimation. Unlike traditional subspace analysis methods that use orthogonal decomposition (SVD) to separate signal and noise, this invention uses differential calculation to eliminate trend term noise, it uses iterative statistical smoothing techniques to remove impulse noise, and it mines the frequency domain features of long-segment data through subspace spectral statistical integration. Simultaneously, a complete system including offline training and online diagnosis is constructed. Addressing the spectral estimation error problem in online short-segment frequency domain diagnosis, the concept of a "maximum allowable error multiple" is proposed to provide an initial benchmark for threshold adjustment, thereby ensuring the timeliness and effectiveness of process measurement point trend diagnosis.

[0027] In summary, this invention integrates offline training and online diagnostic functions to construct a diagnostic architecture based on incremental subspace mapping, vector sequence smoothing, and a maximum allowable error multiple. First, the process measurement data is incrementally processed and mapped into an incremental subspace matrix. An iterative statistical filtering algorithm is used to smooth the vector sequence within the subspace, eliminating outlier interference in the time domain and ensuring the statistical stationarity of elements within the subspace. Second, during the offline training phase, a subspace spectrum estimation strategy is employed to calculate the spectrum of the denoised subspace vectors and to mine the baseline frequency domain features of long data segments through statistical ensemble analysis. Furthermore, addressing the Fast Fourier Transform variance inflation phenomenon caused by shortened data windows in online diagnostics, this invention automatically calculates the "maximum allowable error multiple" during the offline training phase to quantify the physical error boundary brought about by short data. Finally, in the online diagnostic phase, the maximum allowable error multiple is used as the initial reference benchmark for the diagnostic threshold, and combined with time-domain trend analysis, anomaly diagnosis is performed on the real-time acquired short data stream.

[0028] Correspondingly, in other embodiments of the present invention, a process measurement point trend diagnosis system based on subspace smoothing incremental spectrum analysis is provided.

[0029] Example 1 like Figure 1 As shown, this embodiment provides a process measurement point trend diagnosis method based on subspace smoothing incremental spectrum analysis, and constructs an integrated framework for offline training and online diagnosis, mainly including the following steps: 1) Constructing incremental subspaces and smoothing vector sequences: Incremental processing is performed on the historical data sequence of process measurement points and mapped to an incremental subspace matrix, and the vectors of each subspace within the incremental subspace matrix are smoothed.

[0030] In order to obtain high-confidence frequency domain statistical features, this embodiment first maps the one-dimensional time series of process measurement points into a high-dimensional incremental subspace that can characterize the dynamic changes of the process, and then smooths the row vector sequence in the incremental subspace to remove noise interference.

[0031] Specifically, it includes the following steps: 1.1) Incremental processing of data: Incremental processing is performed on the historical data sequence of process measurement points to obtain incremental data sequence.

[0032] Trend diagnosis of key measurement points is essentially a diagnosis of the rate of change of process variables (i.e., the first derivative). Given that historical data in actual production processes typically uses a uniform sampling frequency, diagnosing the first derivative can be equivalently transformed into diagnosing data increments. Therefore, this embodiment calculates the data increment between adjacent time points based on the historical data sequence of process measurement points. Let the historical data sequence be denoted as... ,in Given the total number of sampling points, the data increment between adjacent time points is... The calculation formula is: (1) In the formula, Indicates time Data increment; and Representing time respectively and time Sampling point data; In this embodiment, incremental processing can eliminate the implicit DC component and non-stationary linear trend in the original data, resulting in a data with a length of... Incremental data sequence , represented as: (2) 1.2) Mapping and Construction of Incremental Subspace: Based on the incremental data sequence, the incremental subspace matrix is ​​constructed using piecewise mapping techniques. .

[0033] To uncover the local statistical characteristics of the process measurement data and perform subsequent subspace spectral analysis, it is necessary to transform the one-dimensional incremental data sequence... Mapped to a higher-dimensional observation space. The subspace vector length is set to... (In this embodiment, we take) (In offline training, statistical significance can only be guaranteed if the data segment is long enough.) This is achieved using a segmented mapping technique, with incremental data sequences... Co-formation _n original subspace vectors (where Thus, the incremental subspace matrix is ​​constructed. , represented as: (3) Incremental subspace matrix Each subspace vector in can be denoted as , representing process measurement point data at An independent sample vector in a high-dimensional observation space. Through the spatial mapping operation described above, the original long-time-series incremental signal is transformed into a set of vectors in a high-dimensional space, thus forming the incremental subspace for subsequent spectral statistical analysis.

[0034] 1.3) Iterative smoothing of subspace vectors: Iterative statistical smoothing techniques are used to smooth the incremental subspace matrix. All subspace vectors are smoothed.

[0035] Due to the complex industrial environment, the incremental subspace matrix Each subspace vector in It is often mixed with sporadic impulse noise (outliers). If the spectrum is directly estimated from the incremental subspace matrix containing noise, it will lead to noise increase across the entire frequency band and destroy the statistical stationarity of the incremental subspace.

[0036] Therefore, this embodiment employs an iterative statistical smoothing technique on the incremental subspace matrix. vectors of each subspace Performing smoothing includes the following steps: 1.3.1) Calculation of statistical characteristics of subspace vectors: For the incremental subspace matrix The first in Subspace vectors The mean and standard deviation are calculated.

[0037] Let the first The subspace vectors are Then the mean of its internal data and standard deviation The calculation formula is: (4) 1.3.2) Outlier Identification and Location: For the first... Subspace vectors Each data point is iterated through, and each data point is judged as an outlier using a pre-constructed outlier criterion based on statistical distribution.

[0038] Specifically, traversing the first Each data point within the subspace vector If data points Satisfy the following formula: (5) In the formula, The preset outlier determination coefficient (in this embodiment) If the data point is determined to be... These are outliers that disrupt the stability of the subspace.

[0039] 1.3.3) Adaptive Median Filtering Smoothing Replacement: For data points identified as outliers... An adaptive window median filter is used for smooth replacement to eliminate impulse interference.

[0040] In this embodiment, the backtracking window length is first initialized. (Initially set to 1), and extract the data including the outlier itself and the time preceding it. A reference dataset is constructed using historical data points. During the extraction process, if the time index of a historical data point is non-negative, its actual value is used; if the time index of a historical data point is less than 0 (i.e., it exceeds the starting position of the current subspace vector), the value 0 is used to fill in the blanks.

[0041] Subsequently, the median of the reference dataset was calculated. The calculation formula is shown below, and this median is used to analyze the current outlier. Perform a replacement update.

[0042] (6) After the replacement is complete, the updated data point is validated again using the outlier criterion. If the data point is still identified as an outlier, the backtracking window length is increased. (For example, let) Repeat the above process of constructing the reference dataset and replacing the median until the data point no longer meets the outlier criterion.

[0043] 1.3.4) Iterative Convergence and Reconstruction: Using the replaced new data points to reconstruct the... Subspace vectors To ensure absolute stationarity of the subspace, updates are performed, and steps 1.3.1) to 1.3.3) are repeated until the incremental subspace matrix is ​​reached. The search continues until no element exists that satisfies the outlier criterion.

[0044] When the increment subspace matrix All of them After all row vectors have undergone the above iterative smoothing process, the resulting new matrix set is the smoothed incremental subspace, denoted as . The subspace vectors within it are denoted as .

[0045] 2) Based on subspace increment spectrum statistical feature mining: Frequency domain projection analysis is performed on all subspace vectors within the increment subspace matrix to obtain the long-time increment power spectrum statistical features of the process measurement points.

[0046] In this embodiment, each subspace vector obtained in step 1) The process measurement point is regarded as an independent observation sample under different time windows. By comprehensively analyzing the distribution pattern of all observation samples in the frequency domain projection, the long-term incremental power spectrum statistical characteristics of the process measurement point are explored.

[0047] Specifically, it includes the following steps: 2.1) Subspace Spectrum Estimation: For the first... smooth subspace vectors Spectral estimation was performed using the Wiener-Khinchin Theorem.

[0048] Specific methods include: 2.1.1) Calculation of autocorrelation function: Calculate the first... smooth subspace vectors Discrete autocorrelation function , used to characterize its correlation in the time domain.

[0049] The formula for calculating the autocorrelation function is as follows: (7) In the formula, It is a time-delay variable.

[0050] 2.1.2) Power spectral density calculation: Set the number of points in the fast Fourier transform to be... For autocorrelation function Perform a Fast Fourier Transform (FFT) to obtain the power spectral density sequence. .

[0051] Because the power spectrum of a real signal has conjugate symmetry, this embodiment only retains the amplitude of the positive frequency portion, thus obtaining the first... Power spectral density sequence of subspace vectors : (8) In the formula, This is a discrete frequency index, corresponding to the physical frequency. .

[0052] 2.1.3) Repeat steps 2.1.1) to 2.1.2) for... The autocorrelation function and power spectral density of each subspace vector are calculated to obtain the power spectral density matrix of the smoothed incremental subspace matrix. , represented as: (9) 2.2) Mining of Statistical Features of Incremental Power Spectrum at Measurement Points: Establishing a frequency domain "statistical benchmark" for process measurement points, and analyzing the power spectral density matrix. Statistical characteristic calculations were performed to obtain the long-term incremental power spectrum statistical characteristics of the process measurement points.

[0053] In this embodiment, for each frequency index Calculate the mean and standard deviation of each frequency point, and denote them as the baseline mean. and the benchmark standard deviation : (10) The aforementioned baseline mean and baseline standard deviation, together, constitute the "long-term incremental power spectrum statistical characteristics of the process measurement points," which are used to subsequently measure whether there are abnormal deviations in the online real-time diagnostic data.

[0054] 3) Calculate the maximum allowable error factor: Based on the constructed simulated online subspace, perform short-segment data spectrum estimation to obtain the maximum allowable error factor and the corresponding short-segment data window length.

[0055] According to signal processing theory, the variance of power spectrum estimation is inversely proportional to the data length. Online diagnostics requires real-time performance, necessitating spectral estimation within a short data window, which leads to decreased spectral resolution and estimation errors. To prevent normal fluctuations caused by shorter data windows from being misjudged as abnormal, it is necessary to simulate the online environment offline, quantify this error boundary, and calculate the "maximum allowable error factor" as an initial reference value for the online diagnostic threshold.

[0056] Specifically, it includes the following steps: 3.1) Simulated online subspace construction: Based on the short data window set in the actual online diagnosis, construct the simulated online incremental subspace matrix using the same method as in step 1), and perform smoothing processing.

[0057] Here, the construction method of the online incremental subspace matrix is ​​the same as in step 1), but when using the piecewise mapping technique, the length of the subspace vector is based on the short data window set in the actual online diagnosis. (In this embodiment, we take) If the division is carried out again, it can form A simulated online incremental subspace is constructed, and a simulated online incremental subspace matrix is ​​generated. : (11) In the formula, each simulated online incremental subspace vector This represents a simulated "short online data segment".

[0058] 3.2) Short-segment data spectrum estimation: Using the same method as in step 2.1), the simulated online incremental subspace matrix is ​​estimated. For each simulated online incremental subspace vector, spectral estimation is performed to obtain the short-time standard deviation.

[0059] For the simulation of online incremental subspace matrix For each simulated online incremental subspace vector, repeat the spectral estimation in step 2) to obtain the short-time power spectral density sequence. Subsequently, the statistical analysis of all simulated online incremental subspace vectors at each frequency index was performed. The standard deviation at a given point is denoted as the short-time standard deviation. : (12) This indicator reflects the current short data window length. Below, the inherent "observation noise" level of the process measurement data at each frequency point.

[0060] 3.3) Error amplification factor calculation: The short-time standard deviation obtained in step 3.2) is... Compared with the long-term reference standard deviation obtained in step 2.2), By performing point-by-point comparison, the error amplification sequence at the frequency index is obtained. .

[0061] Among them, frequency index Error amplification sequence at the location Defined as: (13) This sequence objectively quantifies the frequency at any given frequency point. Above, because the short data window length is from shortened to The degree of amplification of the resulting fluctuations.

[0062] 3.4) Threshold coefficient determination and verification: Select error amplification sequences from all frequency points. The maximum value in the range is taken as the maximum allowable multiple of the global error, denoted as . : (14) 3.5) Spectral estimation validity verification mechanism: The maximum allowable multiple of the error obtained in step 3.4). Perform verification until a valid maximum allowed multiple is obtained.

[0063] In this embodiment, the spectral estimation validity verification mechanism includes: like This indicates the current short data window length. If the timeframe is too short, the main low-frequency components in the data cannot form a complete cycle, resulting in a serious "spectral estimation error," meaning that short-term observations lose the long-term fluctuation characteristics. In this case, the timeframe must be increased. And repeat steps 3.1) to 3.4 above. like ,determination If effective, this can be used as the initial threshold benchmark for subsequent online diagnostics.

[0064] 4) Online trend diagnosis: Based on the configuration of process measurement points, combined with the determined statistical characteristics of incremental power spectrum of measurement points, the maximum allowable error multiple and the corresponding short data window length, online diagnosis is performed on the real-time process measurement point data to obtain the process measurement point trend diagnosis results.

[0065] After completing steps 1) to 3) above, offline training of the key measurement points is complete, and the baseline mean is obtained. Benchmark Standard Deviation Online diagnostic data length and the allowable error multiple The online diagnostic phase first involves accessing real-time process data. The short data window length is Perform the following parallel anomaly diagnosis: 4.1) Anomaly detection for prolonged straight-line travel: Based on the configuration of process measurement points, analyze the real-time measurement point process data. Perform abnormal judgment for long-term straight-line movement.

[0066] In chemical production sites, sensor drift, communication failures, control loop saturation, or AD conversion failures often cause process variables (PV) to exhibit spurious invariance, manifesting as long-term constant values ​​or monotonically slow drift (i.e., "straight-line" phenomenon). To address this anomaly, if the measurement point configuration requires long-term straight-line determination, the specific operating steps are as follows: 4.1.1) Moving average filtering: for real-time process measurement point data sequences Perform moving average filtering to obtain a smoothed sequence. This is to eliminate the interference of high-frequency measurement noise on the judgment of subtle trends.

[0067] 4.1.2) Monotonic cumulative counting: Traversing a smooth sequence Count the non-decreasing values. Non-incrementing count value .

[0068] 4.1.3) Anomaly detection: The count value... and The system is compared with a preset threshold, and the comparison result is used to determine whether there is an abnormality of walking in a straight line for a long time.

[0069] If the count value is not decreasing Or a non-incrementing count value If any item in the total data points exceeds a preset threshold (90% in this embodiment), it is determined to be an abnormal long-term straight-line movement.

[0070] 4.2) Trend Anomaly Judgment: Based on the determined statistical characteristics of the incremental power spectrum of the measurement points, the maximum allowable error multiple, and the corresponding short data window length, the real-time measurement point process data is analyzed. Perform trend anomaly detection.

[0071] 4.2.1) Constructing online incremental subspace vector and spectrum estimation: based on real-time process measurement data An online real-time incremental subspace matrix is ​​constructed, and the power spectral density sequence is obtained by spectral estimation of each subspace vector of the online real-time incremental subspace matrix.

[0072] Incremental processing is performed in the same manner as in step 1.1) to construct an online incremental subspace vector. Subsequently, the power spectral density sequence of this online real-time subspace vector is calculated using the same spectral estimation method as in step 2.1). .

[0073] 4.2.2) Threshold Decision: Utilize the maximum allowable multiple of error obtained from offline training. As an initial reference benchmark for online diagnostic thresholds, the power spectral density sequence of each subspace vector is used to make a decision, and the decision result is used to determine whether an abnormal trend has occurred.

[0074] The specific judgment formula is as follows: (15) like At any effective frequency point If the above inequality is satisfied, it indicates that the frequency domain energy fluctuation of the current data significantly exceeds the safety boundary defined by both "long-term statistical regularity" and "short-term error amplification mechanism," and is therefore judged as an abnormal fluctuation. Given the complexity of industrial environments, a single theoretical calculation value is often not applicable to all operating conditions. Therefore, it is necessary to fine-tune the threshold based on field experience to construct the final diagnostic threshold.

[0075] 4.3) Short-term sudden change anomaly judgment: When an abnormal trend occurs, the local trend of the real-time measurement point data is extracted to judge the short-term sudden change anomaly.

[0076] This step is performed only when a "trend anomaly" is triggered, and is designed to further respond to substantial and abrupt changes (such as sudden valve opening, pump shutdown, etc.). Short-term abrupt changes are more severe than trend anomalies.

[0077] 4.3.1) Local trend extraction: Extraction of local trends from real-time measurement point process data. The end window (taken in this embodiment) The last 30 points (for example only) are truncated, and the median of the first 10 points (for example only) of the last window is extracted. The median of the last 10 points (for example only) and the end segment. .

[0078] 4.3.2) Calculation of rate of change and anomaly judgment: based on the median of the initial segment. and the median of the end segment Calculate the normalized rate of change The short-term mutation anomalies are determined based on the calculation results.

[0079] Among them, the normalized rate of change The calculation formula is: (16) In the formula, To prevent extremely small positive numbers with a denominator of zero. If (That is, a change exceeding 9%) is judged as an abnormal, rapid increase in a short period of time; if If the rate of change does not exceed the threshold, the judgment of abnormal fluctuation is maintained.

[0080] 4.4) Error code output: Output diagnostic status code based on the error determination result and trigger the corresponding alarm signal.

[0081] The system outputs the final diagnostic status code based on the above logic (0 for normal, 1 for abnormal fluctuation, 2 for abnormal straight line movement over a long period of time, and 3 and 4 for abnormal rapid rise / fall over a short period of time, respectively), and triggers the corresponding alarm signal for on-site maintenance personnel to handle in a timely manner.

[0082] Example 2 The above-described embodiment 1 provides a process measurement point trend diagnosis method based on subspace smoothing incremental spectrum analysis. Correspondingly, this embodiment provides a process measurement point trend diagnosis system based on subspace smoothing incremental spectrum analysis. The system provided in this embodiment can implement the process measurement point trend diagnosis method based on subspace smoothing incremental spectrum analysis of embodiment 1. The system can be implemented by software, hardware, or a combination of both. For example, the system may include integrated or separate functional modules or functional units to execute the corresponding steps in the methods of embodiment 1. Since the system in this embodiment is basically similar to the method embodiment, the description process in this embodiment is relatively simple. For relevant details, please refer to the description of embodiment 1. The system embodiment provided in this embodiment is merely illustrative.

[0083] The process measurement point trend diagnosis system based on subspace smoothing incremental spectral analysis provided in this embodiment includes: The incremental subspace construction and vector sequence smoothing module is used to perform incremental processing on the historical data sequence of process measurement points and map it into an incremental subspace matrix, and to smooth the vectors of each subspace within the incremental subspace matrix. The subspace increment spectrum statistical feature mining module is used to perform frequency domain projection analysis on all subspace vectors within the increment subspace matrix to obtain the long-time increment power spectrum statistical features of the process measurement points. The maximum allowable error factor calculation module is used to perform short-segment data spectrum estimation based on the constructed simulated online subspace, and obtain the maximum allowable error factor and the corresponding short-segment data window length; The online trend diagnosis module is used to perform online diagnosis on real-time process measurement point data based on the process measurement point configuration, combined with the determined statistical characteristics of the incremental power spectrum of the measurement points, the maximum allowable error multiple, and the corresponding short data window length, to obtain the process measurement point trend diagnosis results.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for diagnosing process measurement point trends based on subspace smoothing incremental spectral analysis, characterized in that, Includes the following steps: The historical data sequence of process measurement points is incrementally processed and mapped into an incremental subspace matrix, and the vectors of each subspace within the incremental subspace matrix are smoothed. Frequency domain projection analysis is performed on all subspace vectors within the incremental subspace matrix to obtain the statistical characteristics of the long-time incremental power spectrum at the process measurement points. Short-segment data spectrum estimation is performed based on the constructed simulated online subspace to obtain the maximum allowable error factor and the corresponding short-segment data window length; Based on the configuration of process measurement points, combined with the determined statistical characteristics of incremental power spectrum of measurement points, the maximum allowable error multiple, and the corresponding short data window length, online diagnosis is performed on real-time process measurement point data to obtain process measurement point trend diagnosis results.

2. The process measurement point trend diagnosis method based on subspace smoothing incremental spectral analysis as described in claim 1, characterized in that, The incremental processing of the historical data sequence of process measurement points and mapping it into an incremental subspace matrix, followed by smoothing of each subspace vector within the incremental subspace matrix, includes: Incremental processing is performed on the historical data sequence of process measurement points to obtain an incremental data sequence; Based on incremental data sequences, an incremental subspace matrix is ​​constructed using segmented mapping techniques. An iterative statistical smoothing technique is used to smooth all subspace vectors in the incremental subspace matrix.

3. The process measurement point trend diagnosis method based on subspace smoothing incremental spectrum analysis as described in claim 2, characterized in that, The iterative statistical smoothing technique is used to smooth all subspace vectors in the incremental subspace matrix, including: ① For the incremental subspace matrix The first in Subspace vectors The mean and standard deviation are calculated; ②For the first Subspace vectors Each data point is traversed, and each data point is judged as an outlier using a pre-constructed outlier criterion based on statistical distribution. ③ Data points identified as outliers Perform median filtering for smooth replacement; ④ Use the replaced new data points to analyze the first... Subspace vectors Update to obtain the first smooth subspace vectors ; ⑤ Repeat steps ① to ④ above until the incremental subspace matrix is ​​obtained. We obtain the smooth incremental subspace matrix by continuing until no element satisfies the outlier criterion. .

4. The process measurement point trend diagnosis method based on subspace smoothing incremental spectral analysis as described in claim 3, characterized in that, The frequency domain projection analysis of all subspace vectors within the incremental subspace matrix yields the long-time incremental power spectrum statistical characteristics of the process measurement points, including: Regarding the first smooth subspace vectors Spectral estimation was performed using the Wiener-Khinchin theorem to obtain the smoothed incremental subspace matrix. power spectral density matrix ; Establish a frequency domain "statistical benchmark" for process measurement points, and analyze the power spectral density matrix. Statistical characteristic calculations are performed to obtain the long-term incremental power spectrum statistical characteristics of the process measurement points, including the baseline mean and baseline standard deviation.

5. The process measurement point trend diagnosis method based on subspace smoothing incremental spectral analysis as described in claim 4, characterized in that, The method of estimating the short-segment data spectrum based on the constructed simulated online subspace, obtaining the maximum allowable error factor and the corresponding short-segment data window length, includes: Based on the short data window set in the actual online diagnosis, a simulated online incremental subspace matrix is ​​constructed and smoothed to obtain a smoothed simulated online incremental subspace matrix. Spectral estimation is performed on each simulated online incremental subspace vector in the smoothed simulated online incremental subspace matrix to obtain the short-time standard deviation; The obtained short-time standard deviation is compared point by point with the benchmark standard deviation to obtain the error amplification sequence at different frequency points; The maximum value in the error amplification sequence among all frequency points is selected as the global maximum allowable error factor. The obtained maximum permissible multiple of error is verified until a valid maximum permissible multiple and the corresponding short data window length are obtained.

6. The process measurement point trend diagnosis method based on subspace smoothing incremental spectral analysis as described in claim 1, characterized in that, Based on the configuration of process measurement points, and combined with the long-term incremental power spectrum statistical characteristics, maximum allowable error multiple, and corresponding short-segment data window length of the determined process measurement points, online diagnosis of real-time process measurement point data is performed to obtain process measurement point trend diagnosis results, including: Based on the configuration of process measurement points, anomalies in long-term straight-line data of real-time process measurement points are determined. Based on the statistical characteristics of the long-term incremental power spectrum of the determined process measurement points, the maximum allowable error multiple, and the corresponding short-segment data window length, trend anomalies are determined in the real-time process measurement point data. When an abnormal trend occurs, the local trend of the real-time process measurement data is extracted to determine the short-term sudden change anomaly. The diagnostic status code is output based on the anomaly determination result, and the corresponding alarm signal is triggered.

7. The process measurement point trend diagnosis method based on subspace smoothing incremental spectral analysis as described in claim 6, characterized in that, The step of determining long-term straight-line anomalies in real-time process measurement point data based on the process measurement point configuration includes: Real-time process measurement point data sequence Perform moving average filtering to obtain a smoothed sequence. ; Traversing the smooth sequence Count the non-decreasing values. Non-incrementing count value ; count value and The system is compared with a preset threshold, and the comparison result is used to determine whether there is an abnormality of walking in a straight line for a long time.

8. The process measurement point trend diagnosis method based on subspace smoothing incremental spectral analysis as described in claim 6, characterized in that, The step of determining trend anomalies in real-time process measurement point data based on the statistical characteristics of the long-term incremental power spectrum of the determined process measurement points, the maximum allowable error multiple, and the corresponding short-segment data window length includes: Based on real-time process measurement data Construct an online real-time incremental subspace matrix, and perform spectral estimation on each subspace vector of the online real-time incremental subspace matrix to obtain the power spectral density sequence; Using the maximum allowable error factor as the initial reference benchmark for the online diagnostic threshold, the power spectral density sequence of each subspace vector is judged, and the trend anomaly is determined based on the judgment result.

9. The process measurement point trend diagnosis method based on subspace smoothing incremental spectral analysis as described in claim 6, characterized in that, When an anomaly occurs, the local trend of the real-time process measurement data is extracted, and a short-term abrupt change anomaly is determined, including: Real-time process measurement data The end window is truncated, and the median of the beginning segment of the end window is extracted. and the median of the end segment ; Based on the median of the initial segment and the median of the end segment Calculate the normalized rate of change The short-term mutation anomalies are determined based on the calculation results.

10. A process measurement point trend diagnosis system based on subspace smoothing incremental spectral analysis, characterized in that, include: The incremental subspace construction and vector sequence smoothing module is used to perform incremental processing on the historical data sequence of process measurement points and map it into an incremental subspace matrix, and to smooth the vectors of each subspace within the incremental subspace matrix. The subspace increment spectrum statistical feature mining module is used to perform frequency domain projection analysis on all subspace vectors within the increment subspace matrix to obtain the long-time increment power spectrum statistical features of the process measurement points. The maximum allowable error multiple calculation module is used to perform short-segment data spectrum estimation based on the constructed simulated online subspace, and obtain the maximum allowable error multiple and the corresponding short-segment data window length; The online trend diagnosis module is used to perform online diagnosis on real-time process measurement point data based on the process measurement point configuration, combined with the determined statistical characteristics of the incremental power spectrum of the measurement points, the maximum allowable error multiple, and the corresponding short data window length, to obtain the process measurement point trend diagnosis results.

Images