Multi-parameter sensing collaborative cleaning system for natural gas filter of gas turbine

Abstract

The invention discloses a multi-parameter sensing gas turbine natural gas filter collaborative cleaning system, and particularly relates to the field of industrial filter operation and maintenance, and the system comprises a multi-mode sensing module, a mode recognition module, a self-adaptive strategy generation module and an instruction execution module. According to the multi-parameter sensing collaborative cleaning system for the gas turbine natural gas filter, traditional centralized analysis is replaced with the mode recognition module, and the intelligent level of operation and maintenance management of the gas turbine natural gas filter and the overall production economic benefits are remarkably improved; through the self-adaptive strategy generation module, the cleaning effect is guaranteed, meanwhile, energy consumption and filter element loss are reduced, the service life of a filter element is prolonged, and the replacement cost is reduced; automatic issuing of a cleaning strategy and collaborative scheduling of operation and maintenance resources are achieved through the instruction execution module, the filter element replacement cost and non-planned shutdown economic loss are effectively reduced, intelligence and automation of operation and maintenance management are achieved, energy consumption and manual intervention requirements are reduced, and therefore the calculation cost of management personnel is reduced.

Description

Technical Field

[0001] This invention relates to the field of industrial filter operation and maintenance technology, and more specifically, to a multi-parameter sensing gas turbine natural gas filter collaborative cleaning system. Background Technology

[0002] With the deep integration of industrial intelligence and digital operation and maintenance management, optimizing the full life cycle business value of gas turbine natural gas filtration systems has become a core management element affecting unit operating efficiency, operation and maintenance cost control, and overall production economic benefits. Its essence is a business decision support system based on multi-dimensional data analysis, which maximizes the return on resource investment and minimizes the full life cycle cost. Traditional technologies lack business intelligence and multi-dimensional decision support, resulting in resource misallocation and passive response management. They lack business intelligence analysis and predictive decision-making capabilities, making it impossible to achieve dynamic business value optimization and dynamic matching of operation and maintenance strategies with business benefits.

[0003] To address the limitations of traditional technologies, existing technologies introduce multi-parameter sensors as supplementary monitoring units. Through time-series data calibration and preliminary fusion algorithms, parallel acquisition and simple collaborative judgment of multi-source data are achieved, providing basic data support for operation and maintenance decision optimization. This promotes the evolution of management models from single-variable experience-based management to multi-variable data-supported management, and initially forms the prototype of data-driven operation and maintenance.

[0004] However, in practical use, it still has some shortcomings. For example, the raw data collected by each monitoring unit mostly adopts a centralized processing paradigm, lacking a collaborative decision-making mechanism that is deeply adapted to the actual operation and maintenance scenario, and cannot meet the needs of scenario-based and dynamic decision-making in operation and maintenance. At the same time, it lacks an intelligent business prediction engine and investment decision optimization platform based on real-time multi-source data to minimize the cost of the entire life cycle. It is difficult to build differentiated and collaborative cleaning strategies that are adapted to different pollution modes, resulting in unreasonable allocation of operation and maintenance resources, increasing the cost of filter replacement and the economic losses caused by unplanned downtime. This restricts the intelligent level of operation and maintenance management of gas turbine natural gas filters and the improvement of overall production economic benefits, and fails to achieve the optimal balance between operation and maintenance costs and operating benefits. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, the present invention provides a multi-parameter sensing gas turbine natural gas filter collaborative cleaning system, which solves the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A multi-parameter sensing gas turbine natural gas filter collaborative cleaning system includes: Multimodal sensing module: used to collect first industrial data, including at least differential pressure, flow rate, particulate matter concentration and humidity, in real time through distributed acquisition units deployed in process units and associated pipeline networks, and to preprocess and extract features from the first industrial data to generate first industrial features corresponding to the first industrial data; Pattern recognition module: used to perform pattern recognition on the first industrial feature through a preset pattern classifier to obtain a second industrial feature representing the current state of the process unit; Adaptive strategy generation module: In response to the second industrial feature, it dynamically generates a third industrial feature containing executable instructions by establishing a physical mechanism optimization model to perform multi-parameter coupled simulation and effect prediction. Instruction execution module: used to convert the third working condition characteristics into an operation and maintenance report and send it to the management personnel terminal.

[0007] Preferably, the pattern recognition module, in the second industrial feature, includes at least the pollution pattern label characterizing the current pollution state of the target filter, which at least includes: Dust-dominated type, its defining characteristics are: pressure difference increases steadily over time, particulate matter concentration is consistently higher than the first threshold, and flow fluctuation variance is lower than the second threshold. The oil mist adhesion type is characterized by a stepwise increase in pressure difference, a continuous increase in ambient humidity reading above the third threshold, and particulate matter concentration between the fourth and fifth thresholds. The mixed gradual type is characterized by a slow change trend in parameters such as pressure difference, flow rate, particulate matter concentration, and humidity without significant abrupt changes. Sudden pollution type is characterized by particulate matter concentration changing at a rate exceeding the sixth threshold per unit time, accompanied by a rapid increase in pressure differential.

[0008] Preferably, the pattern recognition module uses a preset pattern classifier that is a heterogeneous hybrid model, integrating at least two classifiers built based on machine learning principles, and performs the following: Each classifier identifies the first industrial feature in parallel and outputs its respective identification result and corresponding confidence level for the pollution pattern label. The output contamination pattern label is determined by the identification result corresponding to the highest confidence level.

[0009] Preferably, when the confidence scores of all classifiers are lower than a preset confidence threshold, the pattern recognition module outputs an unknown pattern label and triggers a cleaning interruption process, the cleaning interruption process including: Send an interrupt signal to the pattern recognition module and the adaptive policy generation module; At the same time, it automatically switches to a preset safe cleaning strategy, which uses a preset, relatively low cleaning pressure and a short pulse duration; After the safe cleaning strategy is executed, the multimodal perception module and the pattern recognition module are triggered to re-collect data and perform pattern recognition until the confidence level of the identified pollution pattern label meets the confidence level threshold. The first industrial feature corresponding to the triggering of the unknown pattern label will be marked and stored.

[0010] Preferably, the adaptive strategy generation module's physical mechanism optimization model includes at least a first equation for simulating dust stripping effects, a second equation for simulating oil mist shear effects, and a third equation for evaluating compressed air pipeline pressure disturbances.

[0011] Preferably, in the adaptive strategy generation module, the first equation is used to simulate the dynamic process of dry dust being peeled off from the filter element surface under pulse backflushing, specifically expressed as: in, This is expressed as dust removal efficiency. , These are respectively expressed as empirical coefficients characterizing the stripping properties of specific filter media and dust types. Represented as pulse pressure, Expressed as pulse duration, Indicated as reference impulse, Represented as reference pulse pressure, This is represented as the reference pulse duration. This is expressed as real-time dust concentration. This is expressed as a reference concentration.

[0012] Preferably, in the adaptive strategy generation module, the second equation is used to simulate the rheological process of viscous oil mist contaminants being removed under shear force, specifically expressed as: in, Expressed as oil mist removal rate, , , These are respectively represented as model empirical coefficients. Represented as pulse pressure, This is expressed as the critical starting pressure. Expressed as pulse duration, The viscosity is expressed as a reference viscosity. Expressed as the equivalent dynamic viscosity of an oil-gas mixture.

[0013] Preferably, the adaptive strategy generation module, a third-party program, is used to quantify the pressure disturbance caused by a single or multiple cleaning actions on the compressed air pipeline system, specifically expressed as follows: in, This is expressed as a decrease in the total pressure of the pipeline network. Expressed as the gas state coefficient, Expressed as the equivalent volume of the pipeline network, This represents the number of filters that perform cleaning simultaneously or overlapping. Represented as the first Each backflush valve is under cleaning pressure. The air flow rate is expressed as the mass flow rate under standard gas conditions. This is expressed as the valve flow coefficient. Represented as the first The pulse duration of each filter.

[0014] The technical effects and advantages of this invention are as follows: 1. This invention replaces traditional centralized analysis with a pattern recognition module, achieving accurate classification and status assessment of pollution patterns, forming an intelligent prediction engine that is deeply adapted to operation and maintenance scenarios, and significantly improving the intelligence level and overall production and economic benefits of gas turbine natural gas filter operation and maintenance management. 2. This invention dynamically generates personalized cleaning strategies that adapt to different pollution modes through an adaptive strategy generation module, and constructs an investment decision optimization platform that minimizes the entire life cycle cost. While ensuring the cleaning effect, it reduces energy consumption and filter element wear, extends the filter element life, and reduces replacement costs. 3. This invention realizes the automated distribution of cleaning strategies and the collaborative scheduling of operation and maintenance resources through the instruction execution module, which effectively reduces the cost of filter replacement and economic losses from unplanned downtime, realizes intelligent and automated operation and maintenance management, reduces energy consumption and manual intervention requirements, thereby reducing the computing costs of management personnel. Attached Figure Description

[0015] Figure 1 This is a block diagram of a multi-parameter sensing gas turbine natural gas filter collaborative cleaning system provided according to an embodiment of this application. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0018] Hereinafter, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first," "second," and "third" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0019] As attached Figure 1 The multi-parameter sensing gas turbine natural gas filter collaborative cleaning system shown includes a multi-modal sensing module, a pattern recognition module, an adaptive strategy generation module, and an instruction execution module.

[0020] The multimodal sensing module is used to collect first industrial data, including at least differential pressure, flow rate, particulate matter concentration and humidity, in real time through distributed acquisition units deployed in process units and associated pipeline networks, and to preprocess and extract features from the first industrial data to generate first industrial features corresponding to the first industrial data.

[0021] Specifically, each data acquisition subunit in the distributed acquisition unit deployed in the filter and its pipeline network includes at least a differential pressure acquisition subunit, a flow acquisition subunit, a particulate matter concentration acquisition subunit, and a humidity acquisition subunit; each data acquisition subunit performs preprocessing operations according to its physical characteristics, and the preprocessing operations include at least noise reduction processing and signal correction.

[0022] In one embodiment, the differential pressure acquisition subunit uses an algorithm based on a first-order low-pass filter for noise reduction, and its cutoff frequency is set according to the dominant frequency of the fluid pulsation in the main pipeline; the flow rate acquisition subunit uses wavelet transform for noise reduction, selects the Db4 wavelet basis, and performs threshold processing on the high-frequency coefficients to eliminate random interference in the vortex signal; the humidity acquisition subunit eliminates drift by periodically reading the built-in reference resistance value and performing real-time compensation based on a pre-stored three-dimensional calibration table of temperature, humidity, and resistance; the particulate matter concentration acquisition subunit linearizes the signal by using a piecewise linear interpolation algorithm to map the original sensor voltage value to a standard mass concentration unit.

[0023] Furthermore, each data acquisition subunit in the distributed acquisition unit has plug-and-play functionality; when a new or replaced data acquisition subunit is added, its identity, measurement range, and available feature types are reported to the system through a unified hardware interface, and it is automatically connected to the data stream of the multimodal sensing module.

[0024] Furthermore, after each data acquisition subunit performs preprocessing operations based on its physical characteristics, a self-diagnostic operation is performed: if the continuous output value of a data acquisition subunit exceeds its measurement range or remains unchanged for a long time, it is diagnosed as a data anomaly and a coordination request is sent to its associated subunit; the data acquisition subunit receiving the request reconstructs the characteristics of the data acquisition subunit corresponding to the data anomaly based on a pre-stored linear regression model trained on historical data, and adds it to the first industrial characteristic; at the same time, a central timing coordinator is used to dynamically adjust the data sampling frequency of each data acquisition subunit according to the identified pollution pattern; and high-precision timing alignment is performed on asynchronous data streams from different data acquisition subunits based on hardware timestamps to generate a time-synchronized first industrial characteristic.

[0025] In this embodiment, feature extraction is performed on the first industrial data. All extracted feature values ​​are arranged in a predefined order of differential pressure, flow rate, particulate matter, and humidity, and combined into a fixed-dimensional numerical array, namely the first industrial feature. Its contents include, but are not limited to: the differential pressure acquisition subunit calculating the arithmetic mean of the preprocessed signal within a one-minute time window and the linear fitting slope within a five-minute window; the flow rate acquisition subunit calculating the standard deviation of the preprocessed signal within a ten-second window; the particulate matter concentration acquisition subunit calculating the numerical integral of the preprocessed signal within thirty seconds; and the humidity acquisition subunit directly outputting its instantaneous reading after real-time compensation.

[0026] The pattern recognition module is used to perform pattern recognition on the first industrial feature through a preset pattern classifier to obtain a second industrial feature that represents the current state of the process unit.

[0027] In one possible implementation, the second industrial feature includes a contamination pattern label characterizing the current contamination state of the target filter; wherein the contamination pattern label includes at least dust-dominated, oil mist-adhesive, mixed progressive, and sudden contamination types.

[0028] It should be noted that the dust-dominated type is characterized by a steady increase in pressure difference over time, a particulate matter concentration that remains above the first threshold, and a flow rate fluctuation variance that is below the second threshold; the oil mist adhesion type is characterized by a stepwise increase in pressure difference, a continuous increase in ambient humidity reading above the third threshold, and a particulate matter concentration between the fourth and fifth thresholds; the mixed gradual type is characterized by a slow change trend in pressure difference, flow rate, particulate matter concentration, and humidity parameters without significant abrupt changes; and the sudden pollution type is characterized by a change rate of particulate matter concentration exceeding the sixth threshold per unit time, accompanied by a rapid increase in pressure difference.

[0029] In this embodiment, each judgment threshold is automatically calibrated through cluster analysis of historical normal operation data, and can be fine-tuned through incremental learning during operation; the initial value of the first threshold is set to the 85th percentile of historical data; the initial value of the second threshold is set to the 30th percentile of historical flow variance data; those skilled in the art can adjust the above percentile values ​​according to the gas source quality and filter specifications of the specific power plant to adapt to the on-site working conditions.

[0030] In one possible implementation, the pattern classifier is deployed on the edge computing nodes of the distributed acquisition unit; the preset pattern classifier is a heterogeneous hybrid model that has been pruned and quantized, and it is preset to integrate at least two classifiers built based on different machine learning principles.

[0031] In this embodiment, the heterogeneous hybrid model is composed of a support vector machine and a random forest classifier in parallel. Its initial training data comes from thousands of data samples in the historical operation database, which were labeled by experts based on the filter element disassembly report. The labeling is based on the physical filter element disassembly report and laboratory component analysis report during the filter shutdown and maintenance. In this embodiment, when the dust deposition on the filter element surface exceeds 80%, the corresponding historical operation data is labeled as dust-dominated; when there is a large amount of sticky oil between the filter paper fibers, it is labeled as oil mist adhesion type. The classifier built based on support vector machine is used to find the optimal separating hyperplane in the high-dimensional feature space; the classifier built based on random forest constructs multiple decision trees and summarizes their results, focusing on mining statistical patterns from features such as particulate matter concentration and humidity, and is not sensitive to noise and missing values.

[0032] Furthermore, based on the pattern classifier, the recognition process includes: each classifier identifies the first industrial feature in parallel and outputs its respective recognition result and corresponding confidence level for the pollution pattern label; the output pollution pattern label is determined by the recognition result corresponding to the highest confidence level; when the confidence levels of all the classifiers are lower than a preset confidence threshold, an unknown pattern label is output and a cleaning interruption process is triggered.

[0033] It should be noted that when each classifier identifies the first industrial feature in parallel, its input includes the first industrial feature in the current period and the previous N periods. In this embodiment, the window size N is set to 10 to 30 sampling periods according to the dynamic characteristics of the system. The confidence score is a deterministic score calculated based on the entropy of the classifier's output probability distribution. At the same time, each classifier is configured with an incremental learning mechanism. In this embodiment, every Sunday morning, all data samples with confidence scores higher than the threshold and good feedback after execution in the past week are automatically used as a new training set, i.e., the first industrial feature and the finally verified pollution pattern label pair. The random forest classifier is updated using an incremental learning algorithm, while the support vector machine model is optimized by fine-tuning the model parameters.

[0034] In this embodiment, the support vector machine and random forest model are fused using a weighted voting method; the random forest classifier is assigned a weight of 0.6 because it is insensitive to noise; the support vector machine is assigned a weight of 0.4; the final classification result is the class with the highest sum of weighted confidence; the cleaning interruption process is triggered only when the difference between the highest weighted confidence scores of the two models is less than 0.1, i.e., the confidence threshold.

[0035] Furthermore, the cleaning interruption process includes: sending an interrupt signal to the pattern recognition module and the adaptive strategy generation module; simultaneously, automatically switching to a preset safe cleaning strategy, which employs a preset, relatively low cleaning pressure and a short pulse duration; after executing the safe cleaning strategy, triggering the multimodal perception module and the pattern recognition module to re-acquire data and identify contamination patterns until the confidence level of the identified contamination pattern label meets the confidence level threshold.

[0036] It should be noted that when an unknown pattern label is output, in addition to triggering a cleaning interruption process, the process also includes: marking and storing the first industrial feature, execution process data and recovery effect corresponding to the triggering of the unknown pattern label; automatically pushing it to the manual operation interface of the remote monitoring center to prompt engineers to intervene and make final annotations; once manual annotations are obtained, it will be given priority to be added to the next round of incremental learning training set as a high-quality sample.

[0037] The adaptive strategy generation module is used to respond to the second industrial feature by establishing a physical mechanism optimization model to perform multi-parameter coupled simulation and effect prediction, and dynamically generate a third industrial feature containing executable instructions. The third industrial feature includes at least: target filter number, instruction type, pulse pressure value, pulse duration and execution priority.

[0038] It should be noted that the process of generating cleaning instructions corresponding to the current pollution state by performing multi-parameter coupled simulation and effect prediction includes: responding to the second industrial feature, and according to the pollution mode label it carries, calling the corresponding physical equation model and its current set of empirical coefficients from the memory; constructing a multi-objective function with cleaning efficiency, energy consumption and system impact as optimization objectives, and receiving data in real time, including at least the system main pipe intake flow rate, the current pressure of the compressed air network and the ambient temperature, through the system data bus as real-time boundary conditions for the physical equation model; solving the combination of cleaning parameters using a non-dominated sorting genetic algorithm with an elitist strategy, and outputting one or more Pareto optimal solutions; from the Pareto optimal solution set, calculating the comprehensive score of each solution according to the weight coefficients using a weighted sum method, and selecting the solution with the highest score to encapsulate as the third industrial feature.

[0039] In this embodiment, a non-dominated sorting genetic algorithm with an elite strategy is used to solve the problem. The key operating parameters are pre-configured as follows: the population size is set to 50, the number of generations is set to 100, the crossover probability is set to 0.8, the mutation probability is set to 0.1, and the decision variables are cleaning pressure and pulse duration. The initial population is randomly generated within the variable constraints. According to the pollution mode label output by the second industrial feature, the weight coefficients of the three sub-objectives of cleaning efficiency, energy consumption, and system impact in the objective function are dynamically adjusted. Among them, for oil mist adhesion type pollution, cleaning efficiency is given a higher weight, and the weight vector (0.7, 0.15, 0.15) is used. For dust-dominated type pollution, energy consumption is given a higher weight, and the weight vector (0.2, 0.6, 0.2) is used.

[0040] In one possible implementation, the physical mechanism optimization model is a lightweight differential equation model, which includes at least a first equation for simulating the dust stripping effect, a second equation for simulating the oil mist shearing effect, and a third equation for evaluating the pressure disturbance of the compressed air pipeline network.

[0041] Furthermore, the first equation is used to simulate the kinetic process of dry dust being stripped from the filter surface under pulse backflushing, specifically expressed as follows: in, This is expressed as dust removal efficiency. , These are respectively expressed as empirical coefficients characterizing the stripping properties of specific filter media and dust types. Represented as pulse pressure, Expressed as pulse duration, Indicated as reference impulse, Represented as reference pulse pressure, This is represented as the reference pulse duration. This is expressed as real-time dust concentration. Indicated as reference concentration; It should be noted that the dust removal efficiency The value ranges from [0,1], representing the relative removal ratio of the dust layer in a single cleaning cycle, and is an empirical coefficient characterizing the peeling characteristics of a specific filter material and dust type. , The reference impulse is pre-obtained through systematic identification and curve fitting of historical operational and experimental data and stored in non-volatile memory; Used for normalizing impulse terms Reference pulse pressure Compared with reference pulse duration All are constants used for normalization; reference concentration A constant used to normalize the effect of concentration.

[0042] Furthermore, the second equation is used to simulate the rheological process of viscous oil mist contaminants being removed under shear force, specifically expressed as follows: in, Expressed as oil mist removal rate, , , These are respectively represented as model empirical coefficients. Represented as pulse pressure, This is expressed as the critical starting pressure. Expressed as pulse duration, The viscosity is expressed as a reference viscosity. Expressed as the equivalent dynamic viscosity of an oil-gas mixture.

[0043] It should be noted that the oil mist removal rate The value range is [0,1]; critical starting pressure The minimum pressure required to overcome oil mist adhesion; the equivalent dynamic viscosity of the oil-gas mixture. Affected by intake air humidity and temperature; model empirical coefficients , , The reference viscosity is obtained in advance through systematic identification and curve fitting of historical operating data and experimental data, and stored in non-volatile memory; The constant used for normalization.

[0044] Furthermore, the third-party process is used to quantify the pressure disturbance caused by a single or multiple cleaning actions to the compressed air pipeline system, specifically expressed as follows: in, This is expressed as a decrease in the total pressure of the pipeline network. Expressed as the gas state coefficient, Expressed as the equivalent volume of the pipeline network, This represents the number of filters that perform cleaning simultaneously or overlapping. Represented as the first Each backflush valve is under cleaning pressure. The air flow rate is expressed as the mass flow rate under standard gas conditions. This is expressed as the valve flow coefficient. Represented as the first The pulse duration of each filter.

[0045] It should be noted that the gas state coefficient Expressed as the gas state coefficient, Expressed as the specific gas constant of compressed air, Expressed as the absolute temperature of the pipeline network, it relates the mass of air flowing in or out to the resulting pressure change; its value reflects the relationship between the pressure and mass density of a specific gas at a specific temperature; the equivalent volume of the pipeline network. Characterizes the buffering capacity of the system; Each backflush valve is under cleaning pressure. airflow This refers to mass flow rate, not volumetric flow rate, indicating that at a specific valve opening, the mass of air flowing through is proportional to the square root of the pressure upstream of the valve; the disturbance is calculated using the third process. When, the summation term ∑ covers Each filter refers to all filters planned in the previous cycle that are to be cleaned within the same time window. The preparatory cleaning parameters of the filters are obtained by querying a global cleaning task queue, and then joint simulation prediction is performed.

[0046] The instruction execution module is used to convert the third working condition feature into an operation and maintenance report and send it to the management personnel terminal.

[0047] It should be noted that after the management personnel terminal receives the maintenance report, it coordinates and controls the action sequence of one or more actuators in the industrial production line. Specifically, this includes: using a time scheduler, which employs a priority-based dynamic weighted round-robin algorithm: calculating a dynamic priority score for each cleaning task to be executed, the priority score being determined by a weighted average of the task's static priority, waiting time factor, and urgency level output by the pattern recognition module; the time scheduler sorts tasks based on the priority scores and groups them according to the compressed air consumption of each task, ensuring that the cumulative air consumption of all tasks in the same group does not exceed 80% of the real-time supply capacity of the pipeline network, thereby fundamentally avoiding a sudden drop in compressed air pipeline network pressure through staggered execution.

[0048] Furthermore, to address system disturbances under dynamic operating conditions, online rescheduling is implemented: during the execution of preset coordinated control, the total pressure change of the compressed air pipeline network is monitored in real time; in this embodiment, the trigger condition for online rescheduling is set as follows: when the real-time pipeline network pressure is more than 10% lower than the lower limit of the predicted value output by the adaptive strategy generation module for 2 seconds; once the condition is met, all non-urgent tasks will be immediately suspended, and a more conservative new timing plan will be recalculated and issued based on the current actual pressure, the queue of tasks to be executed, and a preset list of uninterruptible tasks, to ensure that the system pressure remains stable within a safe window.

[0049] Furthermore, to ensure the execution accuracy of the cleaning command corresponding to each executable command in the maintenance report, each execution unit in the backflushing actuator cluster is equipped with an independent micro pressure sensor and an electronically controlled proportional valve. After receiving the target pressure command, the execution unit is driven by its embedded microcontroller to open the proportional valve and continuously reads the feedback value of the micro pressure sensor at a period of 10 milliseconds, thereby forming a fast PID control closed loop.

[0050] In this embodiment, the instruction execution module continuously records the historical number of operations and load of each backflush actuator, and dynamically balances the workload of each actuator based on its data during scheduling to extend the overall lifespan of the equipment. At the same time, it diagnoses the key performance indicators of the actuators: the response time of the actuator is defined as the time required from the issuance of the control signal to the return value of its pressure sensor reaching 90% of the target value. If it exceeds the rated value, it is considered abnormal. The tightness of the closure is judged by monitoring the rate of small pressure leakage downstream of the actuator during non-operation periods. If the leakage rate exceeds a preset threshold, it is determined that the closure is not tight. For actuators with abnormal performance, they are automatically marked and their task allocation priority is significantly reduced.

[0051] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-parameter sensing gas turbine natural gas filter collaborative cleaning system, characterized in that, include: Multimodal sensing module: used to collect first operating condition data, including at least differential pressure, flow rate, particulate matter concentration and humidity, in real time through distributed acquisition units deployed in process units and associated pipeline networks, and to preprocess and extract features from the first operating condition data to generate first operating condition features corresponding to the first operating condition data. Pattern recognition module: used to perform pattern recognition on the first working condition features through a preset pattern classifier to obtain the second working condition features that characterize the current state of the process unit; Adaptive strategy generation module: In response to the second working condition feature, it dynamically generates a third working condition feature containing executable instructions by establishing a physical mechanism optimization model to perform multi-parameter coupled simulation and effect prediction. Instruction execution module: used to convert the third working condition characteristics into an operation and maintenance report and send it to the management personnel terminal.

2. The multi-parameter sensing gas turbine natural gas filter collaborative cleaning system according to claim 1, characterized in that: The pattern recognition module, in the second industrial feature, includes at least the pollution pattern label characterizing the current pollution state of the target filter, which at least includes: Dust-dominated type, its defining characteristics are: pressure difference increases steadily over time, particulate matter concentration is consistently higher than the first threshold, and flow fluctuation variance is lower than the second threshold. The oil mist adhesion type is characterized by a stepwise increase in pressure difference, a continuous increase in ambient humidity reading above the third threshold, and particulate matter concentration between the fourth and fifth thresholds. The mixed gradual type is characterized by a slow change trend in parameters such as pressure difference, flow rate, particulate matter concentration, and humidity without significant abrupt changes. Sudden pollution type is characterized by particulate matter concentration changing at a rate exceeding the sixth threshold per unit time, accompanied by a rapid increase in pressure differential.

3. The multi-parameter sensing gas turbine natural gas filter collaborative cleaning system according to claim 2, characterized in that: The pattern recognition module has a preset pattern classifier that is a heterogeneous hybrid model, which integrates at least two classifiers built based on machine learning principles, and performs the following: Each classifier identifies the first industrial feature in parallel and outputs its respective identification result and corresponding confidence level for the pollution pattern label. The output contamination pattern label is determined by the identification result corresponding to the highest confidence level.

4. The multi-parameter sensing gas turbine natural gas filter collaborative cleaning system according to claim 3, characterized in that: The pattern recognition module outputs an unknown pattern label and triggers a cleaning interruption process when the confidence scores of all classifiers are lower than a preset confidence threshold. The cleaning interruption process includes: Send an interrupt signal to the pattern recognition module and the adaptive policy generation module; At the same time, it automatically switches to a preset safe cleaning strategy, which uses a preset, relatively low cleaning pressure and a short pulse duration; After the safe cleaning strategy is executed, the multimodal perception module and the pattern recognition module are triggered to re-collect data and perform pattern recognition until the confidence level of the identified pollution pattern label meets the confidence level threshold. The first industrial feature corresponding to the triggering of the unknown pattern label will be marked and stored.

5. The multi-parameter sensing gas turbine natural gas filter collaborative cleaning system according to claim 1, characterized in that: The adaptive strategy generation module includes at least a first equation for simulating dust stripping effects, a second equation for simulating oil mist shearing effects, and a third equation for evaluating compressed air pipeline pressure disturbances in its physical mechanism optimization model.

6. The multi-parameter sensing gas turbine natural gas filter collaborative cleaning system according to claim 5, characterized in that: The adaptive strategy generation module uses a first equation to simulate the dynamic process of dry dust being peeled off from the filter surface under pulse backflushing, specifically expressed as: in, This is expressed as dust removal efficiency. , These are respectively expressed as empirical coefficients characterizing the stripping properties of specific filter media and dust types. Represented as pulse pressure, Expressed as pulse duration, Indicated as reference impulse, Represented as reference pulse pressure, This is represented as the reference pulse duration. This is expressed as real-time dust concentration. This is expressed as a reference concentration.

7. The multi-parameter sensing gas turbine natural gas filter collaborative cleaning system according to claim 5, characterized in that: The adaptive strategy generation module uses a second equation to simulate the rheological process of viscous oil mist contaminants being removed under shear force, specifically expressed as follows: in, Expressed as oil mist removal rate, , , These are respectively represented as model empirical coefficients. Represented as pulse pressure, This is expressed as the critical starting pressure. Expressed as pulse duration, The viscosity is expressed as a reference viscosity. Expressed as the equivalent dynamic viscosity of an oil-gas mixture.

8. The multi-parameter sensing gas turbine natural gas filter collaborative cleaning system according to claim 5, characterized in that: The adaptive strategy generation module, a third-party program, is used to quantify the pressure disturbance caused by a single or multiple cleaning actions on the compressed air pipeline system, specifically as follows: in, This is expressed as a decrease in the total pressure of the pipeline network. Expressed as the gas state coefficient, Expressed as the equivalent volume of the pipeline network, This represents the number of filters that perform cleaning simultaneously or overlapping. Represented as the first Each backflush valve is under cleaning pressure. The air flow rate is expressed as the mass flow rate under standard gas conditions. This is expressed as the valve flow coefficient. Represented as the first The pulse duration of each filter.

Images