A method of monitoring an alkylation cycle system

Abstract

The application discloses a kind of monitoring methods of alkylation circulation system, it is related to alkylation circulation system monitoring technical field, method includes receiving alkylation circulation pipeline setting node place multiple sensors collection and the operation parameter data of gathering upload by PLC controller;To each road operation parameter data executes wavelet packet transform, and original time domain signal is decomposed into multiple band components;Combination all band components generate multidimensional observation matrix, complete blind source separation by independent component analysis model, after distinguishing signal component and noise component, reconstruct to go out noise after operation parameter signal by inverse transform;Build time-space two-dimensional space-time grid to complete multi-parameter space-time registration;Parameter value in each grid cell is compared with corresponding normal working condition baseline range, and the working condition that deviates amplitude exceeds threshold generates monitoring alarm information.This method can improve the accuracy of working condition deviation monitoring, reduce the false alarm and the probability of missing report.

Description

Technical Field

[0001] This invention relates to the field of alkylation cycle system monitoring technology, and in particular to a monitoring method for alkylation cycle systems. Background Technology

[0002] The alkylation circulation system is the core pipeline system for transporting the acid-hydrocarbon mixture in the sulfuric acid alkylation process. Real-time monitoring of its operating parameters is fundamental to ensuring stable process operation and identifying abnormal operating conditions. Currently, the conventional monitoring method in this field involves deploying multiple sensors for flow rate, pressure, temperature, and medium concentration at various locations along the circulation pipeline. After a PLC controller collects and aggregates the data from each source, a fixed threshold comparison method is used to determine whether the operating conditions deviate from the normal range. Some solutions add low-pass or median filtering before the threshold comparison to suppress random fluctuations in the data.

[0003] Among the existing monitoring technologies mentioned above, the accuracy of monitoring deviations from operating conditions is very low. On the one hand, the sensor signals of the alkylation cycle system are simultaneously affected by multiple sources of noise, such as pump blade pulsation, pipeline turbulence, and water hammer impact. The noise and the actual operating condition signal overlap in the frequency domain. Conventional filtering methods, while suppressing noise, smooth out the transient characteristics of operating condition changes, causing the actual operating condition anomalies to be masked. On the other hand, different sensors are spatially dispersed and have asynchronous sampling times. The data of various operating parameters lack a unified spatiotemporal reference, making it impossible to correlate and compare them in the same spatiotemporal dimension.

[0004] The above factors together result in a high false alarm rate and false negative rate for existing monitoring schemes, making it difficult to accurately and reliably identify real deviations from operating conditions. Summary of the Invention

[0005] To address the technical problems of high false alarm and false negative rates in existing alkylation cycle system monitoring schemes, which make it difficult to accurately and reliably identify real operational deviation events, this invention provides a monitoring method for alkylation cycle systems.

[0006] The technical solution adopted in this invention is as follows: This application provides a monitoring method for an alkylation cycle system, comprising the following steps:

[0007] Step 1: Receive operating parameter data collected by various sensors deployed at the set nodes of the alkylation circulation pipeline and aggregated and uploaded by the PLC controller. The operating parameter data includes at least flow parameters, pressure parameters, temperature parameters and medium concentration parameters, and the operating parameter data carries a collection timestamp.

[0008] Step 2: Perform wavelet packet transform processing on each operating parameter data to decompose the original time-domain signal of each operating parameter into multiple frequency band components;

[0009] Step 3: Combine all frequency band components corresponding to each operating parameter to generate a multidimensional observation matrix; input the multidimensional observation matrix into an independent component analysis model to perform blind source separation processing to obtain multiple independent components; classify each independent component into signal components and noise components, retain the signal components and remove the noise components; perform inverse wavelet packet transform processing on the retained signal components corresponding to each operating parameter to reconstruct the denoised operating parameter signals corresponding to each path;

[0010] Step 4: Construct a two-dimensional spatiotemporal grid, where the horizontal axis of the spatiotemporal grid is the time dimension and the vertical axis is the spatial position dimension along the circulation pipeline; based on the acquisition timestamps corresponding to each denoised operating parameter signal and the spatial position of the sensor deployment, map each denoised operating parameter signal to the corresponding grid cell of the spatiotemporal grid to complete the spatiotemporal registration process.

[0011] Step 5: Obtain the normal operating condition baseline range of each operating parameter corresponding to each spatial location; compare the value of each operating parameter in each grid cell of the spatiotemporal grid with the normal operating condition baseline range of the corresponding operating parameter at the corresponding spatial location; when any operating parameter value deviates from the corresponding normal operating condition baseline range and the deviation exceeds the preset deviation threshold, generate monitoring alarm information, the monitoring alarm information including at least the name of the deviated operating parameter, the deviation spatial location information, and the deviation magnitude value.

[0012] The beneficial effects of the present invention are at least one of the following:

[0013] This invention performs multi-scale frequency band decomposition of the operating parameter signal through wavelet packet transform, and performs blind source separation by combining independent component analysis to distinguish signal components from noise components. Then, the denoised operating parameter signal is reconstructed by inverse transform. Compared with conventional filtering methods, it can reduce the interference of multi-source noise while reducing the smoothing loss of the transient change characteristics of the real operating condition, and provide a more accurate data basis for subsequent judgment of operating condition deviation.

[0014] This invention constructs a two-dimensional spatiotemporal grid and completes the spatiotemporal registration of multi-source operating parameters, unifying parameter data from different deployment locations and sampling times under the same spatiotemporal coordinate system, ensuring the spatiotemporal consistency of multi-parameter correlation comparison, and avoiding judgment deviations caused by parameter spatiotemporal misalignment.

[0015] This invention, based on noise-reduced and spatiotemporally registered operating parameters, combines deviation comparison and alarm output with the normal operating condition baseline range, which helps to reduce false alarms and missed alarms in operating condition deviation monitoring and improves the stability and reliability of real-time monitoring of alkylation cycle systems. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Example 1

[0019] A monitoring method for an alkylation cycle system, such as Figure 1 As shown, it includes the following steps:

[0020] Step 1: Receive operating parameter data collected by various sensors deployed at the set nodes of the alkylation circulation pipeline and aggregated and uploaded by the PLC controller. The operating parameter data includes at least flow parameters, pressure parameters, temperature parameters, and medium concentration parameters, and the operating parameter data carries a collection timestamp.

[0021] In the specific implementation process, multiple types of sensors are deployed at the set nodes of the alkylation circulation pipeline, including electromagnetic flowmeters installed on the main pipeline outlet of the circulation pump, diffused silicon pressure transmitters installed at the pump inlet and outlet and the inlet and outlet of the acid sedimentation separator, armored K-type thermocouples installed at multiple locations in the reactor and at the pump bearing housing, and online acid concentration analyzers installed on the sampling branches of the acid circulation pipeline.

[0022] The output signals of each sensor are connected to the nearest PLC remote I / O station via shielded cables. The PLC synchronously samples each channel at a sampling frequency of no less than 100Hz, and after attaching an absolute timestamp and channel identifier to each sampling point, it uploads the data to the data processing server via industrial Ethernet.

[0023] The operating parameter data received by the data processing server includes at least the flow rate parameter Q, pressure parameter P, temperature parameter T, and medium concentration parameter C, and each parameter data carries a collection timestamp.

[0024] Step 2: Perform wavelet packet transform processing on each operating parameter data to decompose the original time-domain signal of each operating parameter into multiple frequency band components.

[0025] In the specific implementation process, wavelet packet transform is performed on each channel's operating parameter data. The wavelet packet transform also recursively decomposes the low-frequency and high-frequency subbands of each layer, forming a complete frequency band binary tree, thus transforming the original time-domain signal... Decomposed into Layer Each frequency band component. Wavelet packet transform in the th... Layer The recursive relationship of the nodes is as follows:

[0026]

[0027]

[0028] In the formula For the first Layer The low-frequency subband coefficients of each node, For the first Layer The high-frequency subband coefficients of each node These are the coefficients of a wavelet low-pass filter. These are the coefficients of the wavelet high-pass filter. For decomposition layer index, For node indexing, For translation parameters, The variables are for summation. The wavelet basis functions are Daubechies wavelets, with vanishing moments of order 4 to 6.

[0029] To further improve denoising efficiency and parameter adaptability, different decomposition levels can be configured for various sensors based on the differences in the frequency domain distribution of noise in different sensor signals. In one possible implementation, the wavelet packet transform processing performed on each channel of operating parameter data includes the following:

[0030] Configure the corresponding number of decomposition layers according to the sensor type to which the operating parameters belong; the number of decomposition layers corresponding to flow parameters is greater than that corresponding to pressure parameters, and the number of decomposition layers corresponding to temperature parameters is less than that corresponding to pressure parameters.

[0031] For example, the flow sensor takes The pump unit operates at speeds typically between 1500 rpm and 3000 rpm, with 4 to 6 blades and a blade passing frequency between 100 Hz and 300 Hz. The signal sampling frequency is 1000 Hz, and the Nyquist frequency is 500 Hz. The 5-layer decomposition generates 32 sub-bands within the 0 to 500 Hz range, each approximately 15.6 Hz wide, effectively separating the blade passing frequency and its second and third harmonics into different sub-band nodes.

[0032] pressure sensor The layer generates 16 sub-bands in the 0 to 500 Hz range, each with a width of approximately 31.3 Hz, which is sufficient to separate water hammer impact pulse noise from normal pressure fluctuations.

[0033] Temperature sensor The layer generates eight sub-bands in the 0-500Hz range, taking advantage of the inherent low-pass characteristics of temperature signals. The concentration sensor also utilizes... Since the drift noise in the detection cell is a slow-drifting low-frequency noise, it can be effectively separated by decomposition into three layers.

[0034] The physical basis for the differentiated configuration of decomposition layers lies in the systematic differences in the frequency domain distribution characteristics of noise in different sensor signals. Pump blades, through frequency pulsation, represent a deterministic periodic interference generated by rotating machinery. Its energy is concentrated in a few discrete harmonic frequencies and their surrounding narrow bands, requiring high-resolution frequency band division for isolation. Noise in temperature signals mainly originates from ambient temperature fluctuations and electromagnetic interference coupling. This noise is significantly attenuated in the high-frequency band due to the thermal inertia of thermocouples, and the signal's spectral energy is concentrated in the low-frequency range. The differentiated decomposition layer configuration, adapted to the spectral characteristics of each sensor signal, minimizes the overall computational load while ensuring effective denoising.

[0035] Step 3: Combine all frequency band components corresponding to each operating parameter to generate a multidimensional observation matrix; input the multidimensional observation matrix into the independent component analysis model to perform blind source separation processing to obtain multiple independent components; classify each independent component into signal components and noise components, retain the signal components and remove the noise components; perform inverse wavelet packet transform processing on the retained signal components corresponding to each operating parameter to reconstruct the denoised operating parameter signals corresponding to each path.

[0036] In one possible implementation, the multidimensional observation matrix is ​​input into an independent component analysis model to perform blind source separation processing, resulting in multiple independent components. These independent components are then classified into signal components and noise components, including the following:

[0037] Eigenvalue decomposition is performed on the covariance matrix of the multidimensional observation matrix to obtain multiple eigenvalues ​​arranged in descending order; the cumulative eigenvalue contribution rate of the first K eigenvalues ​​is calculated, and the K value corresponding to the first time the cumulative eigenvalue contribution rate exceeds the preset contribution rate threshold is determined as the number of independent components output;

[0038] When classifying each independent component, the correlation between each independent component and the main frequency band of the original signal of the corresponding operating parameter is calculated. Independent components with a correlation exceeding the preset signal component classification threshold are classified as signal components, and the rest are classified as noise components.

[0039] In the specific implementation process, all frequency band components obtained from the decomposition of each operating parameter in step 2 are combined to form a multidimensional observation matrix X. X is an M-row × N_t-column matrix, where M is the total number of all frequency band components from all sensors, and N_t is the number of sampling points on the time axis. Each row of X represents the sampling sequence of a frequency band component on the time axis. The mathematical model for independent component analysis is as follows:

[0040]

[0041] In the formula, A is an M×K dimensional mixing matrix, K is the number of independent components, and S is a K×N_t dimensional source signal matrix, whose row vectors are statistically independent.

[0042] Independent component analysis (ICA) estimates the separation matrix W = A^{-1} from the observation matrix X using an optimization algorithm, so that each component in Y = W·X approximates the true independent source signal S.

[0043] The optimization objective is to maximize the non-Gaussianity measure among the output components. In this embodiment, negative entropy is used as the non-Gaussianity measure.

[0044] The number of independent components K is determined by: the covariance matrix of the multidimensional observation matrix X. Perform eigenvalue decomposition to obtain eigenvalues ​​arranged in descending order. Cumulative contribution rate of the first K eigenvalues Defined as:

[0045]

[0046] Take the satisfaction The minimum K value is taken as the number of independent components in the output, where The preset contribution rate threshold is set to 0.95.

[0047] After separating into K independent components, each component is further divided into signal and noise components. The division is based on the following criteria: for each independent component... Calculate its spectrum and determine the interval where the dominant frequency is located. Then, calculate the Pearson correlation coefficient between the spectral energy in this interval and the spectral energy in the same frequency band interval of the original sensor signal.

[0048] The correlation coefficient exceeds the preset signal component classification threshold. The independent components are divided into signal components, and the rest are divided into noise components. The signal components are retained, and the row vectors corresponding to the noise components are set to zero. An inverse wavelet packet transform is performed on the retained signal components of each sensor to reconstruct the signal components from the sub-band space back to the original time domain space, obtaining the denoised operating parameter signal corresponding to that sensor, denoted as . , , , .

[0049] Step 4: Construct a two-dimensional spatiotemporal grid, where the horizontal axis of the spatiotemporal grid is the time dimension and the vertical axis is the spatial position dimension along the circulation pipeline; based on the acquisition timestamps corresponding to each denoised operating parameter signal and the spatial position of the sensor deployment, map each denoised operating parameter signal to the corresponding grid cell of the spatiotemporal grid to complete the spatiotemporal registration process.

[0050] In the specific implementation process, a two-dimensional spatiotemporal grid is constructed. The horizontal axis For the time dimension, resolution Take the greatest common divisor of the common sampling periods of all sensors; vertical axis The spatial dimension along the circulation pipeline is defined as follows: with the circulation pump outlet as the spatial zero point, a complete circulation loop extends along the pipeline from the forward direction to the return circulation pump inlet. This is based on the total length of the pipeline. of As a uniform division, a total of Each spatial grid cell.

[0051] The denoised operating parameter signals are based on their acquisition timestamps. and sensor deployment space location Mapped to the corresponding mesh cell In this context, the mapping rule is:

[0052]

[0053]

[0054] The flow sensor and pressure sensor are point sensors, with each sampling point mapped to a single grid cell in the corresponding spatial location. The temperature sensor is deployed at multiple locations on the reactor, and its signal is mapped to multiple consecutive adjacent grid cells within the spatial range of the reactor.

[0055] Step 5: Obtain the normal operating condition baseline range of each operating parameter corresponding to each spatial location; compare the value of each operating parameter in each grid cell of the spatiotemporal grid with the normal operating condition baseline range of the corresponding operating parameter at the corresponding spatial location; when any operating parameter value deviates from the corresponding normal operating condition baseline range and the deviation exceeds the preset deviation threshold, generate monitoring alarm information, the monitoring alarm information including at least the name of the deviated operating parameter, the deviation spatial location information, and the deviation magnitude value.

[0056] First, establish the baseline range for normal operating conditions. The continuous operating period after system commissioning, where no faults are confirmed and process parameters are within the design range, is taken as the baseline modeling period. Then, for each spatial location within this period... The historical values ​​of each operating parameter at the location are statistically analyzed, and the average value is calculated. and standard deviation .

[0057] The baseline range for normal operating conditions is defined as follows:

[0058]

[0059] in The calculation formula is:

[0060]

[0061] In the formula For this parameter at this spatial location within this historical period, it is the first... Each sample value This represents the total number of samples. In each monitoring cycle, the current parameter values ​​for each operating parameter within each grid cell of the spatiotemporal grid are recorded. and corresponding baseline range Compare them.

[0062] when When the deviation exceeds the baseline range, calculate the deviation magnitude. :

[0063]

[0064] in For the baseline range boundary and The closest boundary value - when hour ,when hour .

[0065] when When the deviation exceeds the preset threshold, a monitoring alarm message is generated, which includes the name of the deviation operation parameter, the spatial location of the deviation, and the magnitude of the deviation. .

[0066] This embodiment utilizes the difference in statistical independence between signal and noise to achieve joint frequency-spatial domain denoising. The noise sources of the pump blades, such as frequency pulsation, pipeline turbulence, and electromagnetic interference, are physically different from the actual changes in process parameters (acid concentration fluctuations, temperature drift, and flow regulation), and statistically approximately satisfy the independence assumption.

[0067] Wavelet packet transform maps the signals from various sensors to the frequency band space, separating source signals with different frequency characteristics in the frequency domain. Independent component analysis further utilizes statistical independence to extract the individual sources from the mixed signal. The two are cascaded to form a two-stage separation architecture that first pre-separates in the frequency domain and then decouples by statistical independence. Compared with using wavelet thresholding for denoising or using independent component analysis alone, this architecture can more accurately preserve the true transient characteristics of process parameters while suppressing multiple types of noise.

[0068] Spatiotemporal grid registration unifies the denoised multi-sensor data into the same spatiotemporal coordinate system, enabling subsequent baseline comparison to be performed at the same time and spatial location, thus eliminating false alarms caused by spatiotemporal misalignment. Baseline comparison employs... The statistical boundary is considered as the range of normal operating conditions, based on the assumption of a normal distribution of normal operating condition data. The interval accounts for a certain percentage of the theoretically normal data. Parameter values ​​outside this range are statistically significant.

[0069] This embodiment effectively separates real process signals from noise, and performs baseline comparison on the denoised signal, thereby improving alarm accuracy and reducing false alarm rate. Multi-parameter correlation comparison after spatiotemporal grid registration significantly suppresses isolated false alarms caused by instantaneous changes in a single sensor.

[0070] Example 2

[0071] Considering that step 4 of Example 1 maps the denoised signal to the spatiotemporal grid by collecting timestamps and deployment spatial locations, but in the actual alkylation circulation pipeline, the medium is transported at a limited flow rate, and there is a certain transport delay between the upstream and downstream sensors. In order to further improve the physical alignment accuracy of upstream and downstream data in the spatiotemporal grid, transport delay compensation can be introduced during the mapping process so that the upstream and downstream data in the same grid cell correspond to the same fluid cell.

[0072] In one possible implementation, step 4 involves mapping each denoised operating parameter signal to the corresponding grid cell of the spatiotemporal grid to complete the spatiotemporal registration process, which includes the following:

[0073] For upstream and downstream sensors arranged sequentially along the flow direction of the circulation pipeline, the transport time of the medium from the upstream sensor position to the downstream sensor position is calculated based on the current circulation flow rate.

[0074] The denoised operating parameter signal corresponding to the upstream sensor is shifted backward along the time axis by the transport duration, and then the shifted upstream denoised operating parameter signal and downstream denoised operating parameter signal are mapped to the same spatial coordinates of the spatiotemporal grid.

[0075] In the specific implementation process, for any pair of upstream sensors arranged sequentially along the flow direction of the circulation pipeline and downstream sensors Let the installation locations of the two be respectively and ( The length of the pipeline between the two is The cross-sectional area of ​​the pipeline is .

[0076] The denoised flow signal reconstructed in step 3 Obtain the current flow rate value from the medium. Location transmission to Location transport time The calculation formula is:

[0077]

[0078] Will The corresponding denoised operating parameter signal is shifted backward along the time axis. That is, the grid cells mapped to the spatiotemporal grid. place, and The denoised signal is mapped to the same spatial coordinates. It should be noted that the transport time varies in real time with the circulating flow rate. When the circulating flow rate fluctuates due to operating condition adjustments, It then updates adaptively.

[0079] In this embodiment, the physical essence of transport delay compensation is to convert the sensor's fixed-point measurement data from the Euler observation perspective into tracking data of the same fluid unit from the Lagrange tracking perspective. Without compensation, the upstream sensor... The parameter values ​​measured at each time point reflect what The state of the fluid unit at the upstream location is constantly monitored; the downstream sensor is in the same location. The parameter values ​​measured at each time point reflect what Constantly through upstream location, experience After time transmission The state of the fluid unit that reaches the downstream position at any given time; the two correspond to different fluid units. After... After the time shift, the upstream sensor is Real-time data and downstream sensors The data at any given moment physically tracks the same fluid unit, eliminating causal mismatches caused by transport delays.

[0080] In this embodiment, the physical consistency of upstream and downstream data in the spatiotemporal grid is significantly improved after transport delay compensation. The systematic spatial registration deviation caused by transport delay is eliminated after compensation, and the spatial position of each parameter in the baseline comparison is improved. The values ​​at each location have a correct physical correspondence, avoiding false alarms caused by causal timing mismatch.

[0081] Example 3

[0082] Considering that step 5 of Example 1 uses a uniform deviation threshold for baseline comparison of all grid cells, in actual deployment, the number of sensors from which the parameter values ​​of different grid cells originate may be different. The reliability of parameter values ​​in areas with multiple sensors overlapping is higher, followed by single-sensor areas, and the reliability of interpolation areas is relatively lower.

[0083] To address the aforementioned technical problems, in one possible implementation, the process after completing the spatiotemporal registration further includes the following:

[0084] Count the number of source sensors corresponding to the denoised operating parameter signals within each grid cell;

[0085] When the number of source sensors is greater than 1, a weighted average calculation is performed on the signal values ​​corresponding to each sensor, the calculation result is used as the parameter value of the grid cell, and the grid cell is marked with the first confidence level.

[0086] When the number of source sensors is equal to 1, the signal value corresponding to that sensor is used as the parameter value of that grid cell, and the grid cell is marked with a second confidence level;

[0087] When the number of source sensors is 0 and the grid cell is located between two data grid cells, the parameter value of the grid cell is calculated by interpolation, and the grid cell is labeled with the third confidence level.

[0088] In step 5, when performing baseline comparison processing, the corresponding preset deviation threshold is adjusted according to the confidence level of the grid cell. The preset deviation threshold corresponding to the first confidence level is less than that of the second confidence level, and the preset deviation threshold corresponding to the second confidence level is less than that of the third confidence level.

[0089] In the specific implementation process, for each grid cell The number of source sensors that map the denoised operating parameter signals to this unit is statistically analyzed. ,according to The value of is handled in three cases:

[0090] Scenario 1: This unit is simultaneously covered by multiple sensors. A weighted average calculation is performed on the signal values ​​corresponding to these multiple sensors, with the weights... Assignment based on sensor accuracy class—accuracy classes for flow sensors and pressure sensors are selected as follows: Temperature sensor and concentration sensor Weighted average As the parameter value of this unit, the unit is labeled as the first confidence level.

[0091] Scenario 2: This unit is covered by only a single sensor, and the signal value of that sensor is directly taken as the parameter value of this unit. This unit is marked as the second confidence level.

[0092] Scenario 3: Furthermore, this unit is located between two data-carrying units, within the sensor coverage gap, and uses linear interpolation: ,in and These are the parameter values ​​of the nearest data units to the left and right sides of this unit, respectively. and For the corresponding spatial location index, this cell is marked as the third confidence level.

[0093] For example, during baseline comparison in step 5, the deviation threshold is adjusted based on the confidence level of the grid cell. The preset deviation threshold corresponding to the first confidence level. The preset deviation threshold corresponding to the second confidence level The preset deviation threshold corresponding to the third confidence level .

[0094] In this embodiment, the inverse relationship between confidence level and deviation threshold reflects the basic principle that sample size affects the width of confidence interval in statistical inference. The data of the first confidence level has undergone multi-sensor cross-validation, which is equivalent to multiple independent samples measuring the same physical quantity. The variance of the measured value is small, so a narrower deviation threshold can be used for strict judgment. The second confidence level relies only on single-sensor data without cross-validation and uses the standard deviation threshold. The third confidence level is a pure interpolation estimate without actual measurement support. A wide deviation threshold is used to tolerate interpolation error, and an alarm is triggered only when a large deviation occurs.

[0095] In this embodiment, the confidence-driven differential deviation threshold makes the sensitivity of the alarm positively matched with the reliability of the data. In the core pipe section with multiple sensors overlapping, the narrow threshold of 15% can sensitively capture early slight deviations. In the branch pipe section with sparse sensor coverage, the wide threshold of 30% effectively suppresses false alarms caused by interpolation errors. This mechanism reduces false alarms caused by sparse data areas by about 50% while maintaining high sensitivity in the core monitoring area.

[0096] Example 4

[0097] In Example 1, the wavelet packet transform decomposition level and signal component classification threshold are set based on offline test data before system commissioning. Considering that the noise characteristics of the sensor may change slowly during long-term operation, in order to further extend the effective working range of the denoising pipeline throughout its entire life cycle, one possible implementation also includes recalibrating the decomposition level at a preset first cycle interval, specifically including the following:

[0098] The signal components output by independent component analysis within the current period are obtained and reconstructed into the operating parameter reconstructed signal through inverse wavelet packet transform. The reconstruction error between the reconstructed operating parameter signal and the original operating parameter data is calculated. When the reconstruction error exceeds the preset reconstruction error threshold, the number of decomposition layers is adjusted layer by layer, and wavelet packet transform processing, multidimensional observation matrix construction and blind source separation processing are re-executed to generate new signal components and reconstruct the operating parameter reconstructed signal until the reconstruction error falls back to below the preset reconstruction error threshold.

[0099] In the specific implementation process, recalibration is performed to preset the first cycle interval. Triggered by the day. In each calibration cycle, all signal components retained by the independent component analysis (ICA) in step 3 of that cycle are acquired. These components are then reconstructed using two inverse transform steps: inverse independent component analysis and inverse wavelet packet transform, to reconstruct the estimated values ​​of the original signals from each sensor. For a given sensor, its original signal is calculated. With reconstructed signal During this evaluation period Root mean square error at each sampling point:

[0100]

[0101] In the formula For sampling point index, This represents the total number of sampling points within the evaluation period.

[0102] For recalibrating the number of decomposition layers: when Exceeding the preset reconstruction error threshold (in When the standard deviation of the original signal from the sensor within that period is given, it is determined that the current decomposition level is insufficient to adequately separate the noise from the effective components of the signal. The wavelet packet transform decomposition level of the sensor is then adjusted. Add one layer, re-execute wavelet packet transform processing, multidimensional observation matrix construction, and blind source separation processing to generate new signal components and reconstruct the reconstructed signal with running parameters, and calculate the adjusted signal. .like Falling back to The following will retain the adjusted decomposition levels; if Still more Then continue to increase the number of decomposition layers layer by layer until... Falling back to The following may result in the number of decomposition layers reaching the preset limit. .

[0103] In one possible implementation, the signal component classification threshold is recalibrated at a preset second period interval, specifically including:

[0104] Obtain all independent components from the independent component analysis output within the current period, and calculate the correlation between each independent component and the main frequency band of the original signal corresponding to the operating parameters; count the number of independent components whose correlation falls between the current signal component classification threshold and the lower limit of the preset classification threshold, and use this as the number of boundary neighbor components; when the number of boundary neighbor components exceeds the preset neighbor component number threshold, lower the signal component classification threshold by a preset step size, recount the number of boundary neighbor components, and judge again; successively lower the signal component classification threshold until the number of boundary neighbor components does not exceed the preset neighbor component number threshold, or the signal component classification threshold drops to the lower limit of the preset classification threshold; reclassify all independent components into signal components and noise components, and perform inverse wavelet packet transform reconstruction processing on the retained signal components.

[0105] In the specific implementation process, a second cycle interval is preset. Using days as the cycle, all independent components output from step 3 (independent component analysis) are acquired in each cycle. Calculate the correlation between each independent component and the main frequency band of the original signal corresponding to the operating parameters. For each independent component Determine the interval where its dominant frequency is located, and calculate the frequency within that interval. The Pearson correlation coefficient between the spectral energy and the spectral energy of the corresponding original sensor signal within the same frequency band is used as... .

[0106] The statistical correlation falls within the current signal component classification threshold. Compared with the preset classification threshold lower limit The number of independent components between them is denoted as the number of boundary neighborhood components.

[0107]

[0108] Boundary neighborhood components are independent components with similar correlation but not reaching the classification threshold. They are located in the boundary zone between signal and noise and are the component group with the highest risk of confusion under the current classification boundary.

[0109] when Exceeding the preset threshold for the number of neighboring components If the current classification threshold is deemed too strict, edge members of the signal component risk being misclassified as noise. Lower by a preset step size Recount the number of boundary neighborhood components. And again with In comparison, if Still more and Not yet dropped Then continue to lower Iterate until or .

[0110] After the above iterations converge, all independent components are reclassified and divided into signal components and noise components based on the final determined signal component classification threshold. The retained signal components are then subjected to inverse wavelet packet transform reconstruction.

[0111] In this embodiment, the Reconstruction Error (RMSE) is a comprehensive indicator measuring the information loss during the denoising process. An increase in RMSE indicates that some real process signals are misidentified as noise and discarded during the separation process. The root cause of this phenomenon may be sensor degradation, which alters the signal characteristics in specific frequency bands—bands where signal energy was originally concentrated experience energy attenuation due to sensor degradation, leading to misidentification as noise by independent component analysis (ICA). Increasing the number of decomposition layers allows for more precise separation of noise and signal through frequency domain subdivision. When sensor degradation causes noise spectrum broadening, and the existing decomposition layers are insufficient to concentrate the noise into independent subbands, adding another layer further narrows the subband width, helping to re-confine the broadened noise energy to a few sub-nodes, forming a clearer boundary with the signal subband.

[0112] Example 5

[0113] In step 5 of Example 1, baseline comparison is performed on each grid cell in each monitoring cycle to generate monitoring alarm information one by one. However, in actual process parameter deviation events, the deviation often spans multiple consecutive grid cells in space and lasts for multiple monitoring cycles in time, which may generate multiple alarm information pointing to the same deviation event. In order to facilitate the operation and maintenance personnel to quickly grasp the global spatiotemporal range of the deviation, multi-scale alarm aggregation processing can be performed after generating alarm information to merge related alarms into a unified alarm event.

[0114] In one possible implementation, after generating the monitoring alarm information, the following content is also included:

[0115] Multiple monitoring alarm messages that are spatially adjacent and temporally continuous are aggregated into a single alarm event. The alarm event includes the start and end times of aggregation, the spatial range of aggregation, the list of operating parameters involved, and the maximum deviation of each operating parameter. The alarm event is then output to replace the original individual monitoring alarm messages.

[0116] In the specific implementation process, the alarm aggregation processing flow is as follows: taking the generated original monitoring alarm information set as input, and processing each pair of alarm information in the set... and Perform temporal and spatial proximity determinations. Temporal proximity is determined when the timestamp interval between two alarms does not exceed 1.5 times the length of a monitoring cycle. Spatial proximity is determined by the spatial location grid cell index of the two alarms. and satisfy .

[0117] Using a graph connectivity component search approach, all alarm messages that meet the criteria of temporal and spatial proximity are clustered into a single alarm event cluster. Each alarm event cluster is aggregated into one alarm event, containing the following four aggregation information: aggregation start and end times (the earliest and latest alarm timestamps within the cluster); aggregation spatial range (the smallest spatial grid cell index within the cluster); and aggregation spatial range. and the index of the largest spatial grid cell The corresponding pipeline location segment start and end coordinates; the list of operating parameters, taking the union of all different operating parameter names appearing within the cluster; the maximum deviation of each operating parameter, taking the deviation of all alarms for that parameter within the cluster for each operating parameter. The maximum value.

[0118] Output alarm events instead of the original alarm messages one by one; the original alarm messages will no longer be output separately.

[0119] In this embodiment, the rationale for alarm aggregation lies in the fact that parameter deviations caused by the same physical root cause, such as a failure of an acid concentration regulating valve, have continuity in the spatiotemporal dimension. The medium between adjacent spatial grid cells is connected by pipelines, and a deviation at one location will inevitably affect its upstream and downstream adjacent locations; once a deviation occurs at the same spatial location, the deviation state will persist until process adjustments are intervened. This spatiotemporal continuity makes the originally discrete grid alarms physically belong to the spatial expansion and temporal continuation of the same deviation event, and the aggregated output is closer to the physical reality of the deviation event than the output of each individual alarm.

[0120] This example uses a case where a malfunction in the acid concentration regulating valve causes a continuous deviation of the acid concentration in three consecutive grid cells between the reactor outlet and the acid sedimentation separator inlet for three minutes. Assuming a monitoring period of one minute, the original result would be... After aggregation, the original alarm information is processed so that the operation and maintenance personnel only receive one alarm event. This event clearly indicates the time and space range and the maximum deviation. In large-scale deployment scenarios, alarm aggregation can reduce the number of information items that operators need to process by more than an order of magnitude.

[0121] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A monitoring method for an alkylation cycle system, characterized in that, Includes the following steps: Step 1: Receive operating parameter data collected by various sensors deployed at the set nodes of the alkylation circulation pipeline and aggregated and uploaded by the PLC controller. The operating parameter data includes at least flow parameters, pressure parameters, temperature parameters and medium concentration parameters, and the operating parameter data carries a collection timestamp. Step 2: Perform wavelet packet transform processing on each operating parameter data to decompose the original time-domain signal of each operating parameter into multiple frequency band components; Step 3: Combine all frequency band components corresponding to each operating parameter to generate a multi-dimensional observation matrix; The multidimensional observation matrix is ​​input into the independent component analysis model to perform blind source separation processing, resulting in multiple independent components; Each independent component is classified into signal components and noise components. The signal components are retained and the noise components are removed. Inverse wavelet packet transform is performed on the retained signal components corresponding to each operating parameter to reconstruct the denoised operating parameter signals for each channel. Step 4: Construct a two-dimensional spatiotemporal grid, where the horizontal axis of the spatiotemporal grid is the time dimension and the vertical axis is the spatial position dimension along the circulation pipeline; based on the acquisition timestamps corresponding to each denoised operating parameter signal and the spatial position of the sensor deployment, map each denoised operating parameter signal to the corresponding grid cell of the spatiotemporal grid to complete the spatiotemporal registration process. Step 5: Obtain the normal operating condition baseline range of each operating parameter corresponding to each spatial location; compare the value of each operating parameter in each grid cell of the spatiotemporal grid with the normal operating condition baseline range of the corresponding operating parameter at the corresponding spatial location. When any operating parameter value deviates from the corresponding normal operating condition baseline range and the deviation exceeds the preset deviation threshold, a monitoring alarm message is generated. The monitoring alarm message includes at least the name of the deviated operating parameter, the spatial location information of the deviation, and the value of the deviation magnitude.

2. The monitoring method for the alkylation cycle system according to claim 1, characterized in that, The process of performing wavelet packet transform on each operating parameter data includes the following: Configure the corresponding number of decomposition layers according to the sensor type to which the operating parameters belong; The number of decomposition layers corresponding to the flow rate parameter is greater than that corresponding to the pressure parameter, while the number of decomposition layers corresponding to the temperature parameter is less than that corresponding to the pressure parameter.

3. The monitoring method for the alkylation cycle system according to claim 1, characterized in that, The multidimensional observation matrix is ​​input into the independent component analysis model to perform blind source separation processing, resulting in multiple independent components; The independent components are classified into signal components and noise components, including the following: Eigenvalue decomposition is performed on the covariance matrix of the multidimensional observation matrix to obtain multiple eigenvalues ​​arranged in descending order; the cumulative eigenvalue contribution rate of the first K eigenvalues ​​is calculated, and the K value corresponding to the first time the cumulative eigenvalue contribution rate exceeds the preset contribution rate threshold is determined as the number of independent components output; When classifying each independent component, the correlation between each independent component and the main frequency band of the original signal of the corresponding operating parameter is calculated. Independent components with a correlation exceeding the preset signal component classification threshold are classified as signal components, and the rest are classified as noise components.

4. The monitoring method for the alkylation cycle system according to claim 1, characterized in that, In step 4, the denoised operating parameter signals are mapped to the corresponding grid cells of the spatiotemporal grid to complete the spatiotemporal registration process, which includes the following: For upstream and downstream sensors arranged sequentially along the flow direction of the circulation pipeline, the transport time of the medium from the upstream sensor position to the downstream sensor position is calculated based on the current circulation flow rate. The denoised operating parameter signal corresponding to the upstream sensor is shifted backward along the time axis by the transport duration, and then the shifted upstream denoised operating parameter signal and downstream denoised operating parameter signal are mapped to the same spatial coordinates of the spatiotemporal grid.

5. The monitoring method for the alkylation cycle system according to claim 1, characterized in that, After completing the spatiotemporal registration process, the following content is also included: Count the number of source sensors corresponding to the denoised operating parameter signals within each grid cell; When the number of source sensors is greater than 1, a weighted average calculation is performed on the signal values ​​corresponding to each sensor, the calculation result is used as the parameter value of the grid cell, and the grid cell is marked with the first confidence level. When the number of source sensors is equal to 1, the signal value corresponding to that sensor is used as the parameter value of that grid cell, and the grid cell is marked with a second confidence level; When the number of source sensors is 0 and the grid cell is located between two data grid cells, the parameter value of the grid cell is calculated by interpolation, and the grid cell is labeled with the third confidence level. In step 5, when performing baseline comparison processing, the corresponding preset deviation threshold is adjusted according to the confidence level of the grid cell. The preset deviation threshold corresponding to the first confidence level is less than that of the second confidence level, and the preset deviation threshold corresponding to the second confidence level is less than that of the third confidence level.

6. The monitoring method for the alkylation cycle system according to claim 2, characterized in that, It also includes recalibrating the number of decomposition layers at a preset first period interval, specifically including the following: The signal components output by independent component analysis within the current period are obtained and reconstructed using inverse wavelet packet transform to obtain the reconstructed signal of the running parameters. Calculate the reconstruction error between the reconstructed operating parameter signal and the original operating parameter data; When the reconstruction error exceeds the preset reconstruction error threshold, the number of decomposition layers is adjusted layer by layer, and wavelet packet transform processing, multidimensional observation matrix construction and blind source separation processing are re-executed to generate new signal components and reconstruct the signal with running parameters until the reconstruction error falls back below the preset reconstruction error threshold.

7. The monitoring method for the alkylation cycle system according to claim 3, characterized in that, The signal component classification threshold is recalibrated at a preset second period interval, specifically including: Obtain all independent components of the independent component analysis output within the current period, and calculate the correlation between each independent component and the main frequency band of the original signal of the corresponding operating parameters; The number of independent components whose statistical correlation falls between the current signal component classification threshold and the preset classification threshold lower limit is used as the number of boundary neighborhood components. When the number of boundary neighborhood components exceeds the preset threshold for the number of neighborhood components, the threshold for classifying the signal components is lowered by a preset step size, the number of boundary neighborhood components is recounted, and the judgment is made again. The signal component classification threshold is gradually lowered until the number of boundary neighborhood components does not exceed the preset neighborhood component number threshold, or the signal component classification threshold is reduced to the preset classification threshold lower limit. All independent components are reclassified into signal components and noise components. The retained signal components are then reconstructed using inverse wavelet packet transform.

8. The monitoring method for the alkylation cycle system according to claim 1, characterized in that, After generating the monitoring alarm information, the following content is also included: Multiple monitoring alarm messages that are spatially adjacent and temporally continuous are aggregated into a single alarm event. The alarm event includes the start and end times of aggregation, the spatial range of aggregation, the list of operating parameters involved, and the maximum deviation of each operating parameter. The alarm event is then output to replace the original individual monitoring alarm messages.

Images