Centrifugal pump early cavitation online monitoring method and system based on reinforcement learning

Abstract

The application discloses a centrifugal pump early cavitation online monitoring method and system based on reinforcement learning, and relates to the technical field of intelligent monitoring, comprising the following steps: parameterized modeling is carried out on the pump body flow passage, impeller inlet, blade suction surface, volute tongue, ring gap and complex curved surface machining surface of the centrifugal pump for ships; adjustable parameter combinations are screened by using centrifugal pump design equation constraints and distance criteria to obtain a parameterized structure sample set; non-steady-state multiphase flow and structure response joint simulation is carried out on the parameterized structure sample set; the evolution relationship between the bubble volume distribution, pressure pulsation, low pressure area, trailing edge vorticity, entropy production concentration area, head deviation and efficiency change is established to obtain a cavitation evolution causal atlas. The application improves the interpretability and reliability of online cavitation confirmation records by using counterfactual inference and ablation analysis to screen key evidence domains and effective actions.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology, and in particular to an online monitoring method and system for early cavitation of centrifugal pumps based on reinforcement learning. Background Technology

[0002] Centrifugal pumps for ships are widely used in cooling, ballasting, fire fighting, water supply, and liquid transportation, and their operating status directly affects the ship's power supply and fluid transportation safety. With the development of sensor technology, computational fluid dynamics, time-frequency analysis, and intelligent diagnostics, centrifugal pump cavitation monitoring has evolved from manual inspection and judgment based on single pressure or vibration thresholds to online monitoring combining multiple sources of signals such as inlet pressure pulsation, outlet pressure pulsation, acoustic emission, casing vibration, motor current, and shaft power. Simultaneously, methods such as unsteady multiphase flow simulation, fluid-structure interaction analysis, blade passing frequency identification, synchronous compression time-frequency analysis, and deep feature extraction are increasingly being used to describe the relationships between bubble formation, pressure pulsation propagation, hydraulic performance degradation, and structural response.

[0003] Existing early cavitation monitoring technologies for centrifugal pumps typically focus on operational signal identification or simulation result analysis. However, there is a lack of clear correlation between signal characteristics and structurally sensitive components such as the impeller inlet, blade suction surface, volute tongue, and mouth ring gap. This leads to easy confusion between weak cavitation, primary cavitation, and non-cavitation disturbances. Especially under conditions of variable flow rate, variable speed, self-priming, and fluctuating inlet pressure in marine centrifugal pumps, a single evidence domain is insufficient to reliably characterize the location of cavitation initiation, the source of flow loss, and the path of hydraulic performance degradation. Monitoring results lack interpretable causal evidence and are difficult to further map to correctable structural parameters. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an online monitoring method for early cavitation of centrifugal pumps based on reinforcement learning to solve the problem of difficulty in coordinating the discrimination of multi-source signals and structural mechanisms in online monitoring of early cavitation of centrifugal pumps.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an online monitoring method for early cavitation of centrifugal pumps based on reinforcement learning, comprising,

[0008] Parametric modeling was performed on the pump body flow channel, impeller inlet, blade suction surface, volute tongue, mouth ring clearance and complex curved surface of the centrifugal pump for ships. Adjustable parameter combinations were screened using centrifugal pump design equation constraints and distance criteria to obtain a parametric structure sample set.

[0009] Unsteady multiphase flow and structural response were jointly simulated on a parameterized structural sample set to establish the evolution relationship between bubble volume distribution, pressure pulsation, low-pressure zone, trailing edge vorticity, entropy production concentration zone, head deviation and efficiency change, and to obtain a causal map of cavitation evolution.

[0010] Real operating signals are collected and converted into synchronous compressed time-frequency maps. The synchronous compressed time-frequency maps, parameterized structured sample sets, and cavitation evolution causal maps are combined to form a reinforcement learning state. Features are extracted by a residual network, actions are evaluated by a critic network, and the state is distinguished by feedback from the class center distance, thus generating a cavitation precursor state record.

[0011] The system drives multi-agent reinforcement learning to make monitoring action decisions on cavitation precursor state records, and uses counterfactual inference and ablation analysis to screen effective actions and obtain online cavitation confirmation records.

[0012] The online cavitation confirmation records were processed by correctable parameter mapping, non-dominated sorting, and hypervolume improvement to obtain online monitoring results.

[0013] As a preferred embodiment of the reinforcement learning-based online monitoring method for early cavitation of centrifugal pumps described in this invention, the generation of the parameterized structured sample set specifically includes:

[0014] Structural variables of the pump body flow channel, impeller inlet, blade suction surface, volute tongue, mouth ring clearance, and complex curved surface machining are extracted to form a structural parameter table;

[0015] The blade inlet angle, blade outlet angle, impeller outlet width, volute tongue clearance, blade thickness distribution, flow channel cross-sectional area variation and surface roughness are incorporated into the structural parameter table to form an adjustable parameter combination.

[0016] Candidate parameter combinations are formed by limiting the adjustable parameter combinations based on head, efficiency, net positive suction head, self-priming time, and manufacturing accessibility.

[0017] The candidate parameter combinations are verified, and those that do not meet the head calculation results, net positive suction head (NPSH) calculation results, and structural manufacturing boundaries are eliminated to form qualified parameter combinations.

[0018] The structural difference degree between qualified parameter combinations is calculated based on the normalized difference of each structural variable. Qualified parameter combinations with a structural difference degree not less than the preset structural difference threshold and where each structural variable is within the manufacturing achievable range are selected to obtain the parameterized structural sample set.

[0019] As a preferred embodiment of the reinforcement learning-based online monitoring method for early cavitation of centrifugal pumps described in this invention, the generation of the cavitation evolution causal map specifically includes:

[0020] Each set of adjustable parameters in the parameterized structure sample set is filled into the three-dimensional geometric template of the pump body flow channel, impeller, volute and mouth ring gap to generate a structural geometric model;

[0021] The structural geometry model is divided into a rotating domain, a stationary domain, and a fluid-structure interaction interface. The mesh boundaries corresponding to the impeller inlet, blade suction surface, volute tongue, and mouth ring gap are generated to obtain the simulation structural model.

[0022] A working condition tuple is constructed based on the flow rate point, speed point, self-priming status indicator and inlet pressure point. Each simulation structural model is paired with each working condition tuple to form a simulation working condition sequence.

[0023] Under the simulated operating condition sequence, the bubble volume distribution, pressure pulsation, low-pressure area of ​​blade suction surface, trailing edge vorticity, entropy production concentration area, head deviation and efficiency change are calculated to form a simulation response record.

[0024] The location of cavitation initiation is determined by the bubble volume distribution and the low-pressure area of ​​the blade suction surface; online monitoring evidence is determined by pressure pulsation; the source of flow loss is determined by the trailing edge vorticity and entropy production concentration area; and the hydraulic performance degradation result is determined by the head deviation and efficiency change, thus forming the evolution relationship between flow behavior, hydraulic performance and sensor response.

[0025] The evolutionary relationships, structural geometric model numbers, and operating condition tuple numbers are combined to form graph nodes and graph edges, resulting in a cavitation evolution causal graph.

[0026] As a preferred embodiment of the reinforcement learning-based online monitoring method for early cavitation of centrifugal pumps described in this invention, the formation of the reinforcement learning state specifically includes:

[0027] Collect inlet pressure pulsation, outlet pressure pulsation, acoustic emission, casing vibration, motor current and shaft power during the actual operation of the centrifugal pump to form a real operating signal;

[0028] The actual operating signal is converted into a synchronous compressed time-frequency map, and the blade passage frequency sideband, pressure pulsation texture and impact acoustic emission components are extracted to form a time-frequency evidence map.

[0029] The time-frequency evidence map is located to the corresponding structural parameters in the parameterized structural sample set according to the operating speed, blade passage frequency and pressure pulsation frequency band, forming a structural evidence record;

[0030] By embedding structural evidence records into the cavitation evolution causal map corresponding to the cavitation initiation location and flow loss source, a reinforcement learning state is obtained.

[0031] As a preferred embodiment of the reinforcement learning-based online monitoring method for early cavitation of centrifugal pumps described in this invention, the generation of cavitation precursor state records specifically includes:

[0032] The reinforcement learning state is fed into the residual network to extract the deep features corresponding to the blade passing frequency sideband, impact acoustic emission and pressure pulsation texture, and obtain the cavitation time-frequency features.

[0033] The time-frequency characteristics of cavitation are fed into a critic network to evaluate the degree of similarity between the monitoring actions and the early cavitation evolution paths in the cavitation evolution causal map, and to obtain the action evaluation results.

[0034] By using the category center distance feedback to separate the characteristic distances between weak cavitation, primary cavitation, and non-cavitation perturbations, the state distinction results are obtained;

[0035] By associating the action evaluation results with the state differentiation results, a cavitation precursor state record is generated.

[0036] As a preferred embodiment of the reinforcement learning-based online monitoring method for early cavitation of centrifugal pumps described in this invention, the step of executing monitoring action decisions specifically includes:

[0037] The cavitation precursor state records are respectively assigned to pressure pulsation agent, acoustic emission agent, vibration agent, current agent and structurally sensitive part agent to form a multi-agent state;

[0038] Each agent generates a time-frequency analysis window, blade passage frequency correlation band, short-term rotational speed maintenance confirmation, suspected cavitation location and cavitation precursor level based on the corresponding multi-agent state, forming an evidence domain monitoring action;

[0039] Based on the suspected cavitation location and the level of cavitation precursors, the monitoring actions of each evidence domain are merged to form a candidate monitoring action set.

[0040] As a preferred embodiment of the reinforcement learning-based online monitoring method for early cavitation of centrifugal pumps described in this invention, the step of screening effective actions through counterfactual inference and ablation analysis specifically includes:

[0041] The candidate monitoring action set and cavitation precursor state records are sent to the critic network to generate benchmark evaluation records;

[0042] Following the order of evidence domains—pressure pulsation, acoustic emission, vibration, current, and structurally sensitive parts—the state segments and corresponding monitoring actions of each evidence domain are shielded one by one, while the remaining evidence domains retain their original states and actions, thus generating an ablation state sequence.

[0043] The ablation state sequence is fed into a critic network to generate ablation evaluation records;

[0044] By comparing the ablation evaluation records with the baseline evaluation records, changes in cavitation precursor levels, suspected cavitation locations, and early cavitation evolution paths were obtained.

[0045] The shielded evidence domains that cause a decrease in the cavitation precursor level, migration of suspected cavitation sites, and interruption of the early cavitation evolution path are marked as key evidence domains, and the monitoring actions corresponding to the key evidence domains are retained as effective actions.

[0046] The shielded evidence domains that do not change the cavitation precursor level, suspected cavitation location, and early cavitation evolution path are marked as auxiliary evidence domains, and the monitoring actions corresponding to the auxiliary evidence domains are used as background evidence.

[0047] The consistency of effective actions, key evidence domains, auxiliary evidence domains, background evidence, and bubble volume distribution, low-pressure area on blade suction surface, and pressure pulsation path in the cavitation evolution causal map is checked to generate an online cavitation confirmation record.

[0048] As a preferred embodiment of the reinforcement learning-based online monitoring method for early cavitation of centrifugal pumps described in this invention, the modifiable parameter mapping specifically includes:

[0049] Extract effective actions, key evidence domains, auxiliary evidence domains, background evidence, suspected cavitation locations, and cavitation precursor levels from online cavitation confirmation records to form cavitation confirmation elements;

[0050] The suspected cavitation locations in the cavitation confirmation elements are mapped to the impeller inlet, blade suction surface, volute tongue, and mouth ring gap to form a location mapping record;

[0051] The location mapping records are mapped to the blade inlet angle, blade outlet angle, impeller outlet width, volute tongue clearance, blade thickness distribution, and surface roughness to form a set of correctable parameters;

[0052] The set of correctable parameters, cavitation precursor levels, key evidence domains, and background evidence are combined to form candidate records for structural correction.

[0053] As a preferred embodiment of the reinforcement learning-based online monitoring method for early cavitation of centrifugal pumps described in this invention, the online monitoring results specifically include:

[0054] The set of correctable parameters in the candidate records of structural correction is paired with the corresponding suspected cavitation location and cavitation precursor level. The correction amount of blade inlet angle, volute tongue clearance, impeller outlet width, local surface treatment range and operating condition suggestions are set to form candidate structural correction directions.

[0055] By evaluating the direction of candidate structure correction based on factors such as reduced bubble germination, weakened blade suction surface separation, improved self-priming stability, maintained efficiency, and manufacturing accessibility, a multi-objective evaluation record is formed.

[0056] Perform non-dominated sorting on the multi-objective evaluation records, retain the candidate structure correction directions that do not have a dominant relationship, and form a non-dominated candidate set;

[0057] By utilizing hypervolume improvement, the structural correction direction with the largest hypervolume improvement value is selected from the non-dominated candidate set, and the online monitoring results are obtained.

[0058] Secondly, this invention provides an online monitoring system for early cavitation of centrifugal pumps based on reinforcement learning, comprising:

[0059] The structural sampling module performs parametric modeling of gaps and complex curved surfaces, and uses centrifugal pump design equation constraints and distance criteria to screen adjustable parameter combinations to obtain a parametric structural sample set.

[0060] The map generation module performs joint simulation of unsteady multiphase flow and structural response on the parameterized structural sample set, establishes the evolution relationship between bubble volume distribution, pressure pulsation, low-pressure zone, trailing edge vorticity, entropy production concentration zone, head deviation and efficiency change, and obtains the cavitation evolution causal map.

[0061] The state generation module collects real operating signals and converts them into synchronous compressed time-frequency maps. It combines synchronous compressed time-frequency maps, parameterized structured sample sets, and cavitation evolution causal maps to form a reinforcement learning state. The residual network extracts features, the critic network evaluates actions, and the class center distance feedback distinguishes states, generating cavitation precursor state records.

[0062] The action screening module drives multi-agent reinforcement learning to make monitoring action decisions on cavitation precursor state records, and filters effective actions through counterfactual inference and ablation analysis to obtain online cavitation confirmation records;

[0063] The results generation module performs correctable parameter mapping, non-dominated sorting, and overvolume improvement processing on the online cavitation confirmation records to obtain online monitoring results.

[0064] The beneficial effects of this invention are as follows: By parametrically modeling the pump body flow channel, impeller inlet, blade suction surface, volute tongue, mouth ring clearance, and complex curved surface of the centrifugal pump for ships, structural variables can participate in subsequent analysis in a unified data form, avoiding reliance solely on empirical structures or single operating signals for cavitation judgment, and improving the structural targeting of early cavitation monitoring; by using centrifugal pump design equation constraints and distance criteria to screen the parametric structural sample set, representative structural differences can be retained while ensuring head, efficiency, net positive suction head (NPSH), self-priming time, and manufacturing feasibility, reducing invalid simulations and duplicate samples, and improving the utilization rate of computational resources; by forming a causal map of cavitation evolution through joint simulation of unsteady multiphase flow and structural response, the cavitation evolution can be analyzed, including bubble volume distribution, low-pressure area of ​​blade suction surface, etc. The relationships between pressure pulsation, trailing edge vorticity, entropy production concentration zone, head deviation, and efficiency changes are expressed through causal paths, providing traceable physical evidence for online monitoring results. By converting real operating signals into synchronous compressed time-frequency maps and combining them with parameterized structural sample sets and cavitation evolution causal maps to form an enhanced learning state, cavitation precursor features in inlet pressure pulsation, outlet pressure pulsation, acoustic emission, casing vibration, motor current, and shaft power can be mapped to specific structural parts, improving the ability to distinguish between weak cavitation, primary cavitation, and non-cavitation disturbances. By processing cavitation precursor states through residual networks, critic networks, and category center distance feedback, the consistency between monitoring actions and early cavitation evolution paths can be improved while preserving the blade passage frequency sideband, pressure pulsation texture, and impact acoustic emission components. Attached Figure Description

[0065] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Fig. 1 This is a flowchart of an online monitoring method for early cavitation in centrifugal pumps based on reinforcement learning.

[0067] Fig. 2 A flowchart for constructing a parameterized structure sample set and a causal map of cavitation evolution.

[0068] Fig. 3 Flowchart for enhancing learning status and identifying early signs of cavitation.

[0069] Fig. 4 A flowchart for generating multi-agent action screening and online monitoring results. Detailed Implementation

[0070] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0071] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0072] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0073] Reference Figs. 1-4 This is one embodiment of the present invention, which provides an online monitoring method for early cavitation of centrifugal pumps based on reinforcement learning, including the following steps:

[0074] S1. Parametric modeling is performed on the pump body flow channel, impeller inlet, blade suction surface, volute tongue, mouth ring clearance and complex curved surface of the centrifugal pump for ships. Adjustable parameter combinations are selected by using centrifugal pump design equation constraints and distance criteria to obtain a parametric structure sample set.

[0075] S1.1. Based on the pump body flow channel, impeller inlet, blade suction surface, volute tongue, mouth ring clearance, and complex curved surface machining of the marine centrifugal pump, a parametric model that can be quickly corrected is established. The parametric model uses rated flow rate, rotational speed, impeller inlet diameter, impeller outlet diameter, and original design structure as fixed inputs, and uses pump body flow channel section control points, impeller inlet profile points, blade suction surface profile points, volute tongue boundary points, mouth ring clearance dimensions, and complex curved surface roughness as structural variables, forming a structural parameter table. The structural parameter table must at least record the name, symbol, unit, original design value, manufacturing reach, and corresponding geometric location of the structural variables.

[0076] In the structural parameter table, the blade inlet angle, blade outlet angle, impeller outlet width, volute tongue clearance, blade thickness distribution, flow channel cross-sectional area variation, and surface roughness are defined as adjustable parameters, and adjustable parameter combinations are generated according to the manufacturing achievable range.

[0077] To further clarify, the manufacturing achievable range is as follows: blade inlet angle 15° to 30°, blade outlet angle 20° to 35°, impeller outlet width 0.85 to 1.15 times the original impeller outlet width, volute tongue clearance 0.04 to 0.08 times the impeller outlet diameter, blade thickness distribution 0.90 to 1.10 times the original blade thickness distribution, flow channel cross-sectional area change 0.85 to 1.15 times the original flow channel cross-sectional area, and surface roughness Ra 0.8μm to 6.3μm; the mouth ring clearance is calculated based on the minimum mouth ring diameter clearance in API 610. When the mouth ring diameter is not greater than 2 inches, the mouth ring diameter clearance is not less than 0.010 inches. For every 0.5 inch increase in the mouth ring diameter, the mouth ring diameter clearance increases by 0.001 inches. When selecting parameters, the mouth ring clearance is not less than the minimum value and not greater than twice the minimum value.

[0078] S1.2. Adjustable parameter combinations are defined by head, efficiency, net positive suction head (NPSH), self-priming time, and manufacturing feasibility. Specifically, the calculated head is controlled within 95% to 105% of the design head, the calculated efficiency is controlled to be no less than 95% of the original design efficiency, the calculated NPSH is controlled to be no less than the required NPSH plus 0.5m, the self-priming time is controlled to be no greater than the design self-priming time, and all structural variables are controlled within the manufacturing feasibility range. Adjustable parameter combinations that meet the above constraints form candidate parameter combinations.

[0079] When verifying candidate parameter combinations, a preliminary screening is performed using the fundamental design equations of the centrifugal pump within the physical embedded dynamic sampling algorithm. The head calculation result is derived from the Euler pump equation in the field of fluid machinery, which is derived from the conservation of angular momentum; the net positive suction head (NPSH) calculation result is derived from the energy head expression of Bernoulli's equation on the suction side of the centrifugal pump. Combining the impeller outlet velocity triangle and the suction side energy head calculation, the expression is:

[0080] ;

[0081] ;

[0082] ;

[0083] ;

[0084] in, This indicates the calculated head, in meters (m). This indicates the impeller inlet circumferential velocity, in m / s. This indicates the circumferential velocity at the impeller exit, expressed in m / s. This represents the circumferential component of the absolute velocity at the impeller inlet, in m / s. This represents the circumferential component of the absolute velocity at the impeller outlet, in m / s. This represents the acceleration due to gravity, and the unit is m / s². Indicates the first Impeller diameter, in meters; This indicates the impeller speed, in r / min. This indicates the rated flow rate, expressed in m³ / s. This indicates the impeller outlet radius, in meters (m). This indicates the impeller outlet width, in meters (m). Indicates the blade exit angle, in degrees; This indicates the calculated net positive suction head (NPSH), in meters (m). This represents the absolute pressure on the inhalation side, expressed in Pa. This indicates the vaporization pressure of the liquid at the operating temperature, expressed in Pa. This indicates the density of the liquid being transported, expressed in kg / m³. This represents the height difference between the liquid surface on the suction side and the center of the impeller inlet, expressed in meters (m). This indicates the friction loss along the suction side piping, in meters (m). This indicates the inhalation flow velocity, expressed in m / s.

[0085] S1.3. After completing the head and net positive suction head (NPSH) calculations, perform structural manufacturing boundary verification on the candidate parameter combinations. Candidate parameter combinations are eliminated if any structural variable exceeds the manufacturing achievable range, any adjacent curved surfaces self-intersect, the minimum thickness of the blade suction surface is less than 0.90 times the minimum thickness of the original blade, the mouth ring clearance is less than the minimum mouth ring clearance of API 610, or the volute tongue clearance is less than 0.04 times the impeller outlet diameter. The remaining candidate parameter combinations form qualified parameter combinations.

[0086] After the qualified parameter combinations are formed, the structural difference between qualified parameter combinations is calculated according to the normalized difference of each structural variable, and the qualified parameter combination with the highest difference and representativeness is selected by using the distance criterion.

[0087] The structural dissimilarity calculation originates from the Euclidean distance in mathematics, and its expression is:

[0088] ;

[0089] in, Indicates the first The first qualified parameter combination and the first The degree of structural difference between qualified parameter combinations; This indicates the number of structural variables involved in the calculation of structural dissimilarity. Indicates the first In the qualified parameter combination, the first The numerical values ​​of each structural variable; Indicates the first In the qualified parameter combination, the first The numerical values ​​of each structural variable; Indicates the first The upper limit of the range of structural variables within the manufacturing reach; Indicates the first The lower bound of a structural variable within the range achievable by manufacturing.

[0090] The upper limit of the achievable manufacturing range refers to the first The maximum permissible value for each structural variable while satisfying the basic design equations, structural manufacturing boundaries, assembly clearances, and machining accuracy requirements of the centrifugal pump; the lower limit of the manufacturing achievable range refers to the first... The minimum allowable values ​​for each structural variable while satisfying the basic design equations, structural manufacturing boundaries, assembly clearances, and machining accuracy requirements of the centrifugal pump. The manufacturing reach is determined jointly based on the original design structure, machining capabilities, assembly tolerances, minimum blade thickness, mouth ring clearance, volute tongue clearance, and surface roughness of complex curved surfaces.

[0091] If the upper and lower limits of a certain structural variable are the same, the structural variable will not participate in the structural difference calculation, but the original design value of the structural variable will still be retained in the qualified parameter combination.

[0092] When the structural difference between any two qualified parameter combinations is less than the structural difference threshold, retain the qualified parameter combination whose head calculation result, net positive suction head (NPSH) calculation result, and efficiency calculation result are closest to the design requirements, and discard the other qualified parameter combination.

[0093] The structural difference threshold is obtained by statistically analyzing the normalized Euclidean distance between qualified parameter combinations. That is, the structural difference degree between each pair of all qualified parameter combinations is first calculated, and then the distance value that can eliminate similar structural combinations and retain the main structural changes is used as the structural difference threshold.

[0094] After screening using the distance criterion, qualified parameter combinations with structural differences not less than the structural difference threshold and all structural variables within the manufacturing reach are selected to obtain the parameterized structural sample set.

[0095] S2. Perform unsteady multiphase flow and structural response joint simulation on the parameterized structural sample set, establish the evolution relationship between bubble volume distribution, pressure pulsation, low pressure zone, trailing edge vorticity, entropy production concentration zone, head deviation and efficiency change, and obtain the causal map of cavitation evolution.

[0096] S2.1. Sequentially read each set of adjustable parameters from the parameterized structure sample set, and fill each set of adjustable parameters into the three-dimensional geometric templates of the pump body flow channel, impeller, volute, and mouth ring clearance, respectively. This allows the impeller inlet, blade suction surface, volute tongue, mouth ring clearance, and complex curved surface machining to complete geometric updates according to the corresponding structural variables, generating a structural geometric model. The structural geometric model number is consistent with the parameterized structure sample set number.

[0097] The structural geometry model is used as the meshing object, and is divided into rotating domains, stationary domains, and fluid-structure interaction interfaces. The rotating domain corresponds to the impeller and its internal flow channels, the stationary domain corresponds to the pump body flow channels and volute, and the fluid-structure interaction interface corresponds to the contact boundaries where fluid pressure can be transmitted to the blade suction surface, volute tongue, and the solid wall adjacent to the mouth ring gap. Subsequently, the mesh boundaries corresponding to the impeller inlet, blade suction surface, volute tongue, and mouth ring gap are generated, and the blade suction surface, volute tongue, and mouth ring gap are locally refined to allow the low-pressure region, trailing edge vorticity, and bubble volume distribution to be identified at their corresponding geometric locations, thus obtaining the simulation structural model.

[0098] S2.2. Using the simulation structural model as the simulation object, construct operating condition tuples according to flow rate points, speed points, self-priming status indicators, and inlet pressure points. Flow rate points include partial flow rate points near the design flow rate, the design flow rate point, and the high flow rate point. Speed ​​points are speed points within the operating speed range of the marine centrifugal pump. Self-priming status indicators include self-priming start, self-priming transition, and self-priming completion. Inlet pressure points include normal inlet pressure, inlet pressure close to vaporization pressure, and inlet pressure below the safe net positive suction head (NPSH). Pair each simulation structural model with each operating condition tuple to form a simulation operating condition sequence.

[0099] Unsteady multiphase flow and structural response were jointly simulated under a series of simulation conditions. For unsteady multiphase flow, the mass and momentum conservation equations were solved using the finite volume method. The cavitation phase transition process was calculated using the cavitation calculation method based on the Rayleigh-Plesset bubble dynamics equations. Turbulence calculations employed the SST k-ω turbulence model commonly used in unsteady flow of centrifugal pumps. The fluid-structure interaction interface transferred the unsteady wall pressure to the blade suction surface, the volute tongue, and the adjacent structural surfaces of the mouth ring gap, which was used to obtain the structural response.

[0100] Under each simulation condition sequence, the bubble volume distribution, pressure pulsation, low-pressure region of the blade suction surface, trailing edge vorticity, entropy production concentration region, head deviation, and efficiency variation are extracted to form a simulation response record. The bubble volume distribution records the spatial distribution of vapor phase volume fraction in the pump body flow channel, impeller inlet, blade suction surface, volute tongue, and mouth ring gap; the low-pressure region of the blade suction surface records the continuous region where the static pressure is lower than the vaporization pressure at the operating temperature of the pumped liquid; the pressure pulsation records the pressure signal that changes with time at the impeller inlet, blade suction surface, volute tongue, and mouth ring gap boundary; the trailing edge vorticity records the velocity curl intensity near the blade trailing edge; the entropy production concentration region records the flow loss region where viscous dissipation and turbulent dissipation are concentrated; and the head deviation and efficiency variation record the hydraulic performance changes obtained from the joint simulation of unsteady multiphase flow and structural response.

[0101] To ensure consistent comparisons of pressure pulsation, head deviation, and efficiency variation with different operating condition sets, pressure pulsation is dimensionless using the pressure coefficient form from fluid mechanics, while head deviation and efficiency variation are calculated using relative changes from mathematical perspectives, expressed as follows:

[0102] ;

[0103] ;

[0104] ;

[0105] in, Indicates the first Pressure pulsation coefficient at each sampling time; Indicates the first The instantaneous pressure at each sampling time, in Pa; This represents the average pressure at the same measuring point during one impeller rotation cycle, expressed in Pa. This indicates the density of the liquid being transported, expressed in kg / m³. This indicates the circumferential velocity at the impeller exit, expressed in m / s. Indicates the first Head deviation at each sampling moment; Indicates the first The simulated head at each sampling time, in meters; Indicates the design head, in meters (m). Indicates the first Efficiency changes at each sampling time; Indicates the first Simulation efficiency at each sampling time; Indicates design efficiency.

[0106] S2.3. The evolutionary relationships are determined by simulation response records. The cavitation initiation location is determined based on the overlap of bubble volume distribution and the low-pressure zone of the blade suction surface. Online monitoring evidence is determined based on the time-domain peak value, periodic variation, and frequency-domain dominant frequency of the pressure pulsation coefficient. The sources of flow loss are determined based on the spatial overlap of the trailing edge vorticity and entropy production concentration areas. The hydraulic performance degradation results are determined based on head deviation and efficiency changes. The cavitation initiation location, online monitoring evidence, sources of flow loss, and hydraulic performance degradation results are correlated according to the same structural geometric model number and the same operating condition tuple number to form the evolutionary relationship between flow behavior, hydraulic performance, and sensor response.

[0107] The evolutionary relationships, structural geometric model numbers, and operating condition tuple numbers are used to form graph nodes and graph edges. Graph nodes include structural geometric model nodes, operating condition tuple nodes, cavitation initiation location nodes, online monitoring evidence nodes, flow loss source nodes, and hydraulic performance degradation result nodes. Graph edges represent the sequential and corresponding relationships between nodes in the same simulation operating condition sequence, specifically including structural geometric model nodes pointing to operating condition tuple nodes, operating condition tuple nodes pointing to cavitation initiation location nodes, cavitation initiation location nodes pointing to online monitoring evidence nodes, cavitation initiation location nodes pointing to flow loss source nodes, and flow loss source nodes pointing to hydraulic performance degradation result nodes. Graph nodes and graph edges form a causal graph of cavitation evolution.

[0108] S3. Collect real operating signals and convert them into synchronous compressed time-frequency maps. Combine the synchronous compressed time-frequency maps, parameterized structured sample sets, and cavitation evolution causal maps into a reinforcement learning state. Extract features through a residual network, evaluate actions through a critic network, and distinguish states through feedback from the class center distance to generate cavitation precursor state records.

[0109] S3.1. Collect inlet pressure pulsation, outlet pressure pulsation, acoustic emission, casing vibration, motor current, and shaft power during the actual operation of the centrifugal pump to form a real operating signal. Inlet and outlet pressure pulsation reflect unsteady pressure changes in the pump body flow channel and near the impeller inlet; acoustic emission reflects the high-frequency impact components generated by bubble collapse; casing vibration reflects the structural response after the flow excitation is transmitted to the pump body casing; and motor current and shaft power reflect the impact of hydraulic performance changes on the drive end load.

[0110] To ensure that the actual operating signal can cover the blade passage frequency band and the impact sound emission component, the sampling frequency is determined according to the Nyquist sampling theorem in the field of signal processing.

[0111] The sampling frequency is obtained from the highest frequency to be monitored, and the expression is:

[0112] ;

[0113] in, This indicates the sampling frequency, measured in Hz. This indicates the highest monitoring frequency that needs to be retained in the actual operating signal, expressed in Hz.

[0114] The highest frequency to be monitored is determined by the blade passage frequency band, pressure pulsation frequency band, and effective acoustic emission frequency band, so that the cavitation precursor components are not lost when the actual operating signal is converted into a synchronous compression time-frequency diagram.

[0115] S3.2. Convert the actual operating signal into a synchronous compressed time-frequency diagram. The synchronous compressed time-frequency diagram originates from the synchronous compression transform in the field of time-frequency analysis. After obtaining the initial time-frequency energy distribution through short-time Fourier transform or wavelet transform, the synchronous compression transform recompresses the diffused time-frequency energy back to the instantaneous frequency position, which is used to improve the frequency resolution of the blade passage frequency sideband, pressure pulsation texture, and impact acoustic emission components. During the conversion, the inlet pressure pulsation, outlet pressure pulsation, acoustic emission, casing vibration, motor current, and shaft power are first mean-reduced and amplitude-normalized, respectively, and then synchronous compressed time-frequency diagrams are generated according to the same time window.

[0116] In the synchronous compression time-frequency diagram, the blade passage frequency is determined based on the operating speed and the number of blades, and the blade passage frequency sideband is extracted.

[0117] The frequency of blade passage is derived from the kinematics of rotating machinery, that is, the number of times a fixed measuring point is swept by the blades per revolution of the impeller is equal to the number of blades, expressed as:

[0118] ;

[0119] in, This indicates the frequency at which the blade passes through, measured in Hz. Indicates the number of blades; This indicates the operating speed, expressed in r / min.

[0120] S3.3. Extract the blade passage frequency sideband, pressure pulsation texture, and impact acoustic emission component from the synchronous compression time-frequency diagram to form a time-frequency evidence map. The blade passage frequency sideband consists of the energy stripes within the blade passage frequency and adjacent frequency bands that change with the operating speed; the pressure pulsation texture is the periodic time-frequency texture formed by the inlet and outlet pressure pulsations within the impeller rotation cycle; the impact acoustic emission component is the short-duration, high-energy, high-frequency pulse component in the acoustic emission signal.

[0121] The time-frequency evidence map is located to the corresponding structural parameters in the parameterized structural sample set according to the operating speed, blade passage frequency, and pressure pulsation frequency band, forming a structural evidence record. During the location, the relevant structural parameters of the impeller inlet, blade suction surface, and blade trailing edge are first determined by the operating speed and blade passage frequency. Then, the relevant structural parameters of the volute tongue and mouth ring clearance are determined by the pressure pulsation frequency band. Finally, the low-pressure zone of the blade suction surface and the relevant location of bubble rupture are determined by the impact acoustic emission component.

[0122] Structural evidence records are embedded into the corresponding cavitation initiation locations and flow loss sources in the cavitation evolution causality map to obtain reinforcement learning states. During embedding, matching is performed according to the structural geometric model number and the operating condition tuple number. When the blade transmission frequency sideband, pressure pulsation texture, and impact acoustic emission components in the structural evidence records correspond to the cavitation initiation locations, online monitoring evidence, and flow loss sources in the cavitation evolution causality map, the time-frequency evidence map, corresponding structural parameters, cavitation initiation locations, flow loss sources, and hydraulic performance degradation results are combined into reinforcement learning states. These reinforcement learning states represent the physical location of the current real operating signal in the parameterized structural sample set and the cavitation evolution causality map.

[0123] S3.4. The reinforcement learning state is fed into the residual network to extract deep features corresponding to the blade's frequency sideband, impact acoustic emission, and pressure pulsation texture, thus obtaining the cavitation time-frequency features. The residual network belongs to the existing network structure in the field of deep learning. The residual network retains the shallow time-frequency texture through cross-layer connections and extracts deep time-frequency features through convolution operations. After the reinforcement learning state is input into the residual network, the blade's frequency sideband is used to characterize the impeller's periodic excitation changes, the impact acoustic emission is used to characterize the bubble bursting impact, and the pressure pulsation texture is used to characterize the unsteady pressure changes in the flow channel. These three types of features together form the cavitation time-frequency features.

[0124] The time-frequency characteristics of cavitation are fed into a critic network to evaluate the proximity of monitoring actions to early cavitation evolution paths in the cavitation evolution causal graph, thus obtaining action evaluation results. Monitoring actions include non-cavitation disturbance markers, weak cavitation markers, primary cavitation markers, and continued observation. The critic network uses the reinforcement learning state and monitoring actions as inputs, comparing the results corresponding to the monitoring actions with early cavitation evolution paths in the cavitation evolution causal graph. When monitoring actions bring the time-frequency characteristics of cavitation closer to the continuous path of cavitation initiation location, online monitoring evidence, and flow loss sources, the action evaluation result improves; when the results corresponding to the monitoring actions deviate from the early cavitation evolution path, the action evaluation result decreases.

[0125] S3.5. Utilize the category center distance feedback to differentiate the feature distances between weak cavitation, primary cavitation, and non-cavitation perturbations, thus obtaining the state distinction results. The category center distance feedback originates from Euclidean distance in mathematics and the concept of category centers in classification learning. First, calculate the category centers for weak cavitation, primary cavitation, and non-cavitation perturbations based on the confirmed cavitation time-frequency features. Then, calculate the distance between the current cavitation time-frequency feature and each category center. The expression is:

[0126] ;

[0127] ;

[0128] in, Indicates category The category center; This indicates weak cavitation, primary cavitation, or non-cavitation disturbance. Indicates category The number of cavitation time-frequency characteristics has been confirmed in the data; Indicates category The Middle One confirmed time-frequency characteristic of cavitation; Indicates the current time-frequency characteristics and category of cavitation. The distance between the category centers; The number of dimensions representing the time-frequency characteristics of cavitation; Indicating the current cavitation time-frequency characteristics, the first... 3D eigenvalues; Indicates category In the category center, the first Dimensional eigenvalues.

[0129] When the current cavitation time-frequency feature has the smallest distance from the category center of primary cavitation, and can be distinguished from the category center of weak cavitation and non-cavitation disturbance, the state classification result is recorded as primary cavitation; when the current cavitation time-frequency feature has the smallest distance from the category center of weak cavitation, and does not reach the corresponding distance relationship of primary cavitation, the state classification result is recorded as weak cavitation; when the current cavitation time-frequency feature has the smallest distance from the category center of non-cavitation disturbance, the state classification result is recorded as non-cavitation disturbance.

[0130] The action evaluation results and state differentiation results are correlated to generate a cavitation precursor state record. The cavitation precursor state record includes the actual operating signal number, the synchronous compressed time-frequency map number, the time-frequency evidence map number, the structural evidence record number, the reinforcement learning state number, the cavitation time-frequency characteristics, the action evaluation result, the state differentiation result, the corresponding cavitation initiation location, the corresponding flow loss source, and the corresponding early cavitation evolution path.

[0131] S4. Drive multi-agent reinforcement learning to make monitoring action decisions on cavitation precursor state records, and filter effective actions through counterfactual inference and ablation analysis to obtain online cavitation confirmation records.

[0132] S4.1. The cavitation precursor state records are assigned to pressure pulsation agents, acoustic emission agents, vibration agents, current agents, and structurally sensitive agents according to the source of evidence, forming a multi-agent state. The pressure pulsation agent corresponds to inlet pressure pulsation, outlet pressure pulsation, pressure pulsation texture, and pressure pulsation frequency band; the acoustic emission agent corresponds to acoustic emission and impact acoustic emission components; the vibration agent corresponds to casing vibration and blade passage frequency band; the current agent corresponds to motor current, shaft power, and hydraulic performance degradation results; and the structurally sensitive agent corresponds to the impeller inlet, blade suction surface, volute tongue, mouth ring gap, and cavitation initiation location and flow loss source in the cavitation evolution causal map in the parameterized structural sample set.

[0133] The pressure pulsation agent determines the corresponding time-frequency analysis window, blade passage frequency correlation band, suspected cavitation location, and cavitation precursor level based on the pressure pulsation frequency band, pressure pulsation texture, and pressure pulsation path. The acoustic emission agent determines the corresponding time-frequency analysis window, impact acoustic emission occurrence time, suspected cavitation location, and cavitation precursor level based on the impact acoustic emission component. The vibration agent determines the corresponding time-frequency analysis window, blade passage frequency correlation band, short-term speed maintenance confirmation, and cavitation precursor level based on casing vibration and blade passage frequency bands. The current agent determines the short-term speed maintenance confirmation, load change corresponding time period, and cavitation precursor level based on motor current, shaft power, and hydraulic performance degradation results. The structurally sensitive part agent determines the suspected cavitation location and cavitation precursor level based on structural evidence records, cavitation initiation location, and flow loss source. The pressure pulsation agent, acoustic emission agent, vibration agent, current agent, and structurally sensitive part agent each form evidence domain monitoring actions.

[0134] S4.2. Merge the monitoring actions of each evidence domain according to the suspected cavitation location and the cavitation precursor level to form a candidate monitoring action set. During merging, the monitoring actions of the evidence domain where the suspected cavitation location points to the impeller inlet, blade suction surface, or volute tongue are grouped into the same location record; the monitoring actions of the evidence domain with the same cavitation precursor level are grouped into the same level record; when different cavitation precursor levels exist under the same suspected cavitation location, the cavitation precursor level that can be continuously connected in the early cavitation evolution path in the cavitation evolution causal map is used as the candidate cavitation precursor level.

[0135] The candidate monitoring action set and cavitation precursor state records are used as evaluation inputs for the critic network. The critic network first reads the evidence domain name, time-frequency analysis window, blade passage frequency correlation band, short-term rotational speed maintenance confirmation, suspected cavitation location, and cavitation precursor level in the candidate monitoring action set. Then, it reads the cavitation time-frequency characteristics, state differentiation results, corresponding cavitation initiation location, corresponding flow loss source, and corresponding early cavitation evolution path in the cavitation precursor state record, and checks them item by item according to the node connection relationship in the cavitation evolution causal graph.

[0136] Item-by-item verification refers to verifying the consistency between suspected cavitation sites and the cavitation initiation locations in the cavitation evolution causality diagram; verifying the consistency between the time-frequency analysis window and the blade through frequency correlation bands and the online monitoring evidence in the cavitation evolution causality diagram; verifying the consistency between short-term speed maintenance confirmation and the speed points in the operating condition tuple; verifying the consistency between the cavitation precursor level and the cavitation development stage in the early cavitation evolution path; and verifying the consistency between structurally sensitive parts and the sources of flow loss in the cavitation evolution causality diagram.

[0137] When the suspected cavitation location, time-frequency analysis window, blade passage frequency correlation band, short-term rotational speed maintenance confirmation, cavitation precursor level, and structurally sensitive location in the candidate monitoring action set can all continuously correspond to the order of cavitation initiation location—online monitoring evidence—flow loss source—hydraulic performance degradation result in the cavitation evolution causal graph, the critic network determines that the candidate monitoring action set is consistent with the early cavitation evolution path; when any check item cannot find a corresponding node in the cavitation evolution causal graph, or there is no continuous graph edge between corresponding nodes, the critic network records the inconsistent check item.

[0138] The benchmark consistency level, consistency results of suspected cavitation sites, consistency results of online monitoring evidence, consistency results of rotational speed, consistency results of cavitation precursor level, consistency results of structurally sensitive sites, inconsistency check items, benchmark cavitation precursor level, benchmark suspected cavitation sites, benchmark early cavitation evolution path, and benchmark action evaluation results are written into the same record to form a benchmark evaluation record.

[0139] Following the order of evidence domains—pressure pulsation, acoustic emission, vibration, current, and structurally sensitive areas—state segments and corresponding monitoring actions of individual evidence domains are sequentially masked. Unmasked evidence domains retain their original states and actions, forming an ablation state sequence. When masking pressure pulsation, inlet pressure pulsation, outlet pressure pulsation, pressure pulsation texture, pressure pulsation frequency band, and corresponding monitoring actions are excluded from the critic network evaluation. When masking acoustic emission, acoustic emission, impact acoustic emission components, and corresponding monitoring actions are excluded. When masking vibration, casing vibration, blade passage frequency band, and corresponding monitoring actions are excluded. When masking current, motor current, shaft power, and corresponding monitoring actions are excluded. When masking structurally sensitive areas, structural evidence records, cavitation initiation locations, flow loss sources, and corresponding monitoring actions are excluded. Each item in the ablation state sequence masks only one evidence domain, facilitating the assessment of the impact of a single evidence domain on the cavitation confirmation result.

[0140] S4.3. The critic network evaluates each item in the ablation state sequence, forming an ablation evaluation record. Using the baseline evaluation record as the factual condition and the ablation evaluation record as the counterfactual condition, the changes in cavitation precursor level, suspected cavitation location, and early cavitation evolution path are obtained by comparing the cavitation precursor level, suspected cavitation location, and early cavitation evolution path under the factual and counterfactual conditions.

[0141] The ablation evaluation records are compared with the baseline evaluation records for consistency. If the ablation cavitation precursor level corresponding to the shielded evidence domain is lower than the baseline cavitation precursor level, or the suspected ablation cavitation site differs from the baseline suspected cavitation site, or the early ablation cavitation evolution path lacks any node among the following: cavitation initiation location, online monitoring evidence, flow loss source, and hydraulic performance degradation result, the shielded evidence domain is marked as a critical evidence domain, and the monitoring actions corresponding to the critical evidence domain are retained as valid actions. If the ablation cavitation precursor level, suspected ablation cavitation site, and early ablation cavitation evolution path are all consistent with the baseline evaluation records, the shielded evidence domain is marked as an auxiliary evidence domain, and the monitoring actions corresponding to the auxiliary evidence domain are used as background evidence.

[0142] S4.4. Verify the consistency of effective actions, key evidence domains, auxiliary evidence domains, and background evidence with the bubble volume distribution, low-pressure area on the blade suction surface, and pressure pulsation path in the cavitation evolution causality map, and generate an online cavitation confirmation record. During verification, the suspected cavitation location in the effective actions needs to be consistent with the cavitation initiation location in the cavitation evolution causality map; the pressure pulsation, acoustic emission, or vibration in the key evidence domain needs to be consistent with the pressure pulsation path in the cavitation evolution causality map; the structurally sensitive location needs to be consistent with the impeller inlet, blade suction surface, volute tongue, or mouth ring gap location in the parameterized structural sample set; the auxiliary evidence domain and background evidence are used to record the sources of evidence that do not change the cavitation confirmation results but can supplement the explanation of changes in the actual operating signals.

[0143] S5. Perform correctable parameter mapping, non-dominated sorting, and overvolume improvement processing on the online cavitation confirmation records to obtain online monitoring results.

[0144] S5.1. Extract effective actions, key evidence domains, auxiliary evidence domains, background evidence, suspected cavitation locations, and cavitation precursor levels from online cavitation confirmation records to form cavitation confirmation elements. Effective actions are used to characterize monitoring actions that still affect the cavitation confirmation results after counterfactual inference and ablation analysis; key evidence domains are used to characterize the sources of evidence that cause changes in cavitation precursor levels, suspected cavitation locations, or early cavitation evolution paths; auxiliary evidence domains and background evidence are used to characterize the sources of evidence that do not change the cavitation confirmation results but can supplement changes in actual operating signals; suspected cavitation locations are used to indicate the location of cavitation occurrence; cavitation precursor levels are used to indicate non-cavitation disturbances, weak cavitation, or initial cavitation states.

[0145] Using suspected cavitation locations from the cavitation confirmation elements as matching objects, the suspected cavitation locations are mapped to the impeller inlet, blade suction surface, volute tongue, and mouth ring clearance, forming a location mapping record. When the suspected cavitation location is located at the impeller inlet, the location mapping record includes the impeller inlet and the blade inlet angle associated with the impeller inlet; when the suspected cavitation location is located at the blade suction surface, the location mapping record includes the blade suction surface and the blade inlet angle, blade outlet angle, blade thickness distribution, and surface roughness associated with the blade suction surface; when the suspected cavitation location is located at the volute tongue, the location mapping record includes the volute tongue and the volute tongue clearance associated with the volute tongue; when the suspected cavitation location is located at the mouth ring clearance, the location mapping record includes the mouth ring clearance and the impeller outlet width and surface roughness adjacent to the mouth ring clearance. The location mapping record retains the online cavitation confirmation record number, suspected cavitation location, cavitation precursor level, and corresponding structural location.

[0146] S5.2. Map the structural positions in the part mapping record to the blade inlet angle, blade outlet angle, impeller outlet width, volute tongue clearance, blade thickness distribution, and surface roughness to form a set of correctable parameters. The impeller inlet corresponds to the blade inlet angle; the blade suction surface corresponds to the blade inlet angle, blade outlet angle, blade thickness distribution, and surface roughness; the volute tongue corresponds to the volute tongue clearance; and the mouth ring clearance corresponds to the impeller outlet width and surface roughness. Each correctable parameter in the set is read from the structural parameter table within its manufacturing reach; correctable parameters outside the manufacturing reach are not processed further.

[0147] The set of correctable parameters, cavitation precursor levels, key evidence domains, and background evidence are combined to form candidate records for structural correction.

[0148] The set of correctable parameters in the candidate structural correction records is paired with the corresponding suspected cavitation locations and cavitation precursor levels. Corrections are set for the blade inlet angle, volute tongue clearance, impeller outlet width, local surface treatment range, and operating condition recommendations to form candidate structural correction directions. The blade inlet angle correction is used to change the fluid injection conditions at the impeller inlet; the volute tongue clearance correction is used to change the pressure pulsation path near the volute tongue; the impeller outlet width correction is used to change the impeller outlet flow channel area and flow distribution; the local surface treatment range is used to determine the areas near the blade suction surface, impeller inlet, volute tongue, or mouth ring clearance where surface roughness needs improvement; and the operating condition recommendations are used to limit the short-term speed maintenance confirmation, inlet pressure change confirmation, and flow point adjustment range. Each candidate structural correction direction retains its corresponding manufacturing reachability; candidate structural correction directions that do not meet the manufacturing reachability are not included in the multi-objective evaluation record.

[0149] S5.3. A multi-objective evaluation record is formed by evaluating the direction of candidate structure modification based on factors such as reduced bubble initiation, weakened blade suction surface separation, improved self-priming stability, maintained efficiency, and manufacturing accessibility. Reduced bubble initiation is determined by the changing trends of bubble volume distribution and the low-pressure region of the blade suction surface in the cavitation evolution causal map; weakened blade suction surface separation is determined by the changing trends of trailing edge vorticity and entropy production concentration areas; improved self-priming stability is determined by the correspondence between self-priming state indicators, self-priming time, and inlet pressure point; maintained efficiency is determined by the correspondence between efficiency changes and head deviation; and manufacturing accessibility is jointly determined by the manufacturing reachability range, structural manufacturing boundaries, assembly clearances, and surface roughness of complex curved surfaces.

[0150] A non-dominated ranking is performed on the multi-objective evaluation records, retaining candidate structural correction directions that do not have a dominance relationship, forming a non-dominated candidate set. The non-dominated ranking originates from the Pareto dominance relationship in the multi-objective optimization domain. For any two candidate structural correction directions, if the first candidate structural correction direction is not lower than the second candidate structural correction direction in all five evaluation categories (reduction of bubble germination, weakening of blade suction surface separation, improvement of self-priming stability, efficiency maintenance, and manufacturing accessibility), and at least one evaluation category shows a higher result than the second candidate structural correction direction, then the first candidate structural correction direction dominates the second candidate structural correction direction. Candidate structural correction directions dominated by other candidate structural correction directions are eliminated, and candidate structural correction directions not dominated by other candidate structural correction directions enter the non-dominated candidate set.

[0151] S5.4. Using hypervolume improvement, the structural correction direction with the largest hypervolume improvement value is selected from the non-dominated candidate set to obtain the online monitoring results. Hypervolume improvement originates from the hypervolume index in the field of multi-objective optimization, used to measure the multi-objective improvement space covered by the candidate structural correction direction in the non-dominated candidate set relative to the reference point. The reference point consists of the lowest acceptable result among five evaluation results in the multi-objective evaluation record: reduced bubble germination, weakened blade suction surface separation, improved self-priming stability, maintained efficiency, and manufacturing accessibility. For each candidate structural correction direction in the non-dominated candidate set, the hypervolume difference before and after adding the candidate structural correction direction is calculated, and the difference is taken as the hypervolume improvement value; the candidate structural correction direction with the largest hypervolume improvement value is taken as the structural correction direction corresponding to the online monitoring result.

[0152] This embodiment also provides an online monitoring system for early cavitation of centrifugal pumps based on reinforcement learning, including: a structural sampling module, which performs parametric modeling of gaps and complex curved surfaces, and uses centrifugal pump design equation constraints and distance criteria to screen adjustable parameter combinations to obtain a parametric structural sample set;

[0153] The map generation module performs joint simulation of unsteady multiphase flow and structural response on the parameterized structural sample set, establishes the evolution relationship between bubble volume distribution, pressure pulsation, low-pressure zone, trailing edge vorticity, entropy production concentration zone, head deviation and efficiency change, and obtains the cavitation evolution causal map.

[0154] The state generation module collects real operating signals and converts them into synchronous compressed time-frequency maps. It combines synchronous compressed time-frequency maps, parameterized structured sample sets, and cavitation evolution causal maps to form a reinforcement learning state. The residual network extracts features, the critic network evaluates actions, and the class center distance feedback distinguishes states, generating cavitation precursor state records.

[0155] The action screening module drives multi-agent reinforcement learning to make monitoring action decisions on cavitation precursor state records, and filters effective actions through counterfactual inference and ablation analysis to obtain online cavitation confirmation records;

[0156] The results generation module performs correctable parameter mapping, non-dominated sorting, and overvolume improvement processing on the online cavitation confirmation records to obtain online monitoring results.

[0157] In summary, this invention parametrically models the pump body flow channel, impeller inlet, blade suction surface, volute tongue, mouth ring clearance, and complex curved machined surfaces of a shipborne centrifugal pump. This allows structural variables to participate in subsequent analysis in a unified data format, avoiding reliance on empirical structures or single operating signals for cavitation judgment and improving the structural specificity of early cavitation monitoring. By using centrifugal pump design equation constraints and distance criteria to screen the parametric structural sample set, representative structural differences can be retained while ensuring head, efficiency, net positive margin, self-priming time, and manufacturing feasibility. This reduces invalid simulations and duplicate samples, improving the utilization rate of computational resources. Furthermore, by jointly simulating unsteady multiphase flow and structural response to form a causal map of cavitation evolution, it can analyze bubble volume distribution, low-pressure areas on the blade suction surface, and pressure... The relationships between force pulsation, trailing edge vorticity, entropy production concentration zone, head deviation, and efficiency changes are expressed through causal paths, providing traceable physical evidence for online monitoring results. By converting real operating signals into synchronous compressed time-frequency maps and combining them with parameterized structural sample sets and cavitation evolution causal maps to form an enhanced learning state, cavitation precursor features in inlet pressure pulsation, outlet pressure pulsation, acoustic emission, casing vibration, motor current, and shaft power can be mapped to specific structural parts, improving the ability to distinguish between weak cavitation, primary cavitation, and non-cavitation disturbances. By processing cavitation precursor states through residual networks, critic networks, and category center distance feedback, the consistency between monitoring actions and early cavitation evolution paths can be improved while preserving the blade passage frequency sideband, pressure pulsation texture, and impact acoustic emission components.

[0158] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for online monitoring of early cavitation in centrifugal pumps based on reinforcement learning, characterized in that: include, Parametric modeling was performed on the pump body flow channel, impeller inlet, blade suction surface, volute tongue, mouth ring clearance and complex curved surface of the centrifugal pump for ships. Adjustable parameter combinations were screened using centrifugal pump design equation constraints and distance criteria to obtain a parametric structure sample set. Unsteady multiphase flow and structural response were jointly simulated on a parameterized structural sample set to establish the evolution relationship between bubble volume distribution, pressure pulsation, low-pressure zone, trailing edge vorticity, entropy production concentration zone, head deviation and efficiency change, and to obtain a causal map of cavitation evolution. Real operating signals are collected and converted into synchronous compressed time-frequency maps. The synchronous compressed time-frequency maps, parameterized structured sample sets, and cavitation evolution causal maps are combined to form a reinforcement learning state. Features are extracted by a residual network, actions are evaluated by a critic network, and the state is distinguished by feedback from the class center distance, thus generating a cavitation precursor state record. The system drives multi-agent reinforcement learning to make monitoring action decisions on cavitation precursor state records, and uses counterfactual inference and ablation analysis to screen effective actions and obtain online cavitation confirmation records. The online cavitation confirmation records were processed by correctable parameter mapping, non-dominated sorting, and hypervolume improvement to obtain online monitoring results.

2. The online monitoring method for early cavitation of centrifugal pumps based on reinforcement learning as described in claim 1, characterized in that: The generation of the parameterized structured sample set specifically includes, Structural variables of the pump body flow channel, impeller inlet, blade suction surface, volute tongue, mouth ring clearance, and complex curved surface machining are extracted to form a structural parameter table; The blade inlet angle, blade outlet angle, impeller outlet width, volute tongue clearance, blade thickness distribution, flow channel cross-sectional area variation and surface roughness are incorporated into the structural parameter table to form an adjustable parameter combination. Candidate parameter combinations are formed by limiting the adjustable parameter combinations based on head, efficiency, net positive suction head, self-priming time, and manufacturing accessibility. The candidate parameter combinations are verified, and those that do not meet the head calculation results, net positive suction head (NPSH) calculation results, and structural manufacturing boundaries are eliminated to form qualified parameter combinations. The structural difference degree between qualified parameter combinations is calculated based on the normalized difference of each structural variable. Qualified parameter combinations with a structural difference degree not less than the preset structural difference threshold and where each structural variable is within the manufacturing achievable range are selected to obtain the parameterized structural sample set.

3. The online monitoring method for early cavitation of centrifugal pumps based on reinforcement learning as described in claim 2, characterized in that: The generation of the cavitation evolution causal map specifically includes, Each set of adjustable parameters in the parameterized structure sample set is filled into the three-dimensional geometric template of the pump body flow channel, impeller, volute and mouth ring gap to generate a structural geometric model; The structural geometry model is divided into a rotating domain, a stationary domain, and a fluid-structure interaction interface. The mesh boundaries corresponding to the impeller inlet, blade suction surface, volute tongue, and mouth ring gap are generated to obtain the simulation structural model. A working condition tuple is constructed based on the flow rate point, speed point, self-priming status indicator and inlet pressure point. Each simulation structural model is paired with each working condition tuple to form a simulation working condition sequence. Under the simulated operating condition sequence, the bubble volume distribution, pressure pulsation, low-pressure area of ​​blade suction surface, trailing edge vorticity, entropy production concentration area, head deviation and efficiency change are calculated to form a simulation response record. The location of cavitation initiation is determined by the bubble volume distribution and the low-pressure area of ​​the blade suction surface; online monitoring evidence is determined by pressure pulsation; the source of flow loss is determined by the trailing edge vorticity and entropy production concentration area; and the hydraulic performance degradation result is determined by the head deviation and efficiency change, thus forming the evolution relationship between flow behavior, hydraulic performance and sensor response. The evolutionary relationships, structural geometric model numbers, and operating condition tuple numbers are combined to form graph nodes and graph edges, resulting in a cavitation evolution causal graph.

4. The online monitoring method for early cavitation of centrifugal pumps based on reinforcement learning as described in claim 3, characterized in that: The formation of the reinforcement learning state specifically includes, Collect inlet pressure pulsation, outlet pressure pulsation, acoustic emission, casing vibration, motor current and shaft power during the actual operation of the centrifugal pump to form a real operating signal; The actual operating signal is converted into a synchronous compressed time-frequency map, and the blade passage frequency sideband, pressure pulsation texture and impact acoustic emission components are extracted to form a time-frequency evidence map. The time-frequency evidence map is located to the corresponding structural parameters in the parameterized structural sample set according to the operating speed, blade passage frequency and pressure pulsation frequency band, forming a structural evidence record; By embedding structural evidence records into the cavitation evolution causal map corresponding to the cavitation initiation location and flow loss source, a reinforcement learning state is obtained.

5. The online monitoring method for early cavitation of centrifugal pumps based on reinforcement learning as described in claim 4, characterized in that: The generation of cavitation precursor state records specifically includes... The reinforcement learning state is fed into the residual network to extract the deep features corresponding to the blade passing frequency sideband, impact acoustic emission and pressure pulsation texture, and obtain the cavitation time-frequency features. The time-frequency characteristics of cavitation are fed into a critic network to evaluate the degree of similarity between the monitoring actions and the early cavitation evolution paths in the cavitation evolution causal map, and to obtain the action evaluation results. By using the category center distance feedback to separate the characteristic distances between weak cavitation, primary cavitation, and non-cavitation perturbations, the state distinction results are obtained; By associating the action evaluation results with the state differentiation results, a cavitation precursor state record is generated.

6. The online monitoring method for early cavitation of centrifugal pumps based on reinforcement learning as described in claim 5, characterized in that: The decision-making process for executing monitoring actions specifically includes, The cavitation precursor state records are respectively assigned to pressure pulsation agent, acoustic emission agent, vibration agent, current agent and structurally sensitive part agent to form a multi-agent state; Each agent generates a time-frequency analysis window, blade passage frequency correlation band, short-term rotational speed maintenance confirmation, suspected cavitation location and cavitation precursor level based on the corresponding multi-agent state, forming an evidence domain monitoring action; Based on the suspected cavitation location and the level of cavitation precursors, the monitoring actions of each evidence domain are merged to form a candidate monitoring action set.

7. The online monitoring method for early cavitation of centrifugal pumps based on reinforcement learning as described in claim 6, characterized in that: The method of screening effective actions through counterfactual inference and ablation analysis specifically includes: The candidate monitoring action set and cavitation precursor state records are sent to the critic network to generate benchmark evaluation records; Following the order of evidence domains—pressure pulsation, acoustic emission, vibration, current, and structurally sensitive parts—the state segments and corresponding monitoring actions of each evidence domain are shielded one by one, while the remaining evidence domains retain their original states and actions, thus generating an ablation state sequence. The ablation state sequence is fed into a critic network to generate ablation evaluation records; By comparing the ablation evaluation records with the baseline evaluation records, changes in cavitation precursor levels, suspected cavitation locations, and early cavitation evolution paths were obtained. The shielded evidence domains that cause a decrease in the cavitation precursor level, migration of suspected cavitation sites, and interruption of the early cavitation evolution path are marked as key evidence domains, and the monitoring actions corresponding to the key evidence domains are retained as effective actions. The shielded evidence domains that do not change the cavitation precursor level, suspected cavitation location, and early cavitation evolution path are marked as auxiliary evidence domains, and the monitoring actions corresponding to the auxiliary evidence domains are used as background evidence. The consistency of effective actions, key evidence domains, auxiliary evidence domains, background evidence, and bubble volume distribution, low-pressure area on blade suction surface, and pressure pulsation path in the cavitation evolution causal map is checked to generate an online cavitation confirmation record.

8. The online monitoring method for early cavitation of centrifugal pumps based on reinforcement learning as described in claim 7, characterized in that: The modifiable parameter mapping specifically includes, Extract effective actions, key evidence domains, auxiliary evidence domains, background evidence, suspected cavitation locations, and cavitation precursor levels from online cavitation confirmation records to form cavitation confirmation elements; The suspected cavitation locations in the cavitation confirmation elements are mapped to the impeller inlet, blade suction surface, volute tongue, and mouth ring gap to form a location mapping record; The location mapping records are mapped to the blade inlet angle, blade outlet angle, impeller outlet width, volute tongue clearance, blade thickness distribution, and surface roughness to form a set of correctable parameters; The set of correctable parameters, cavitation precursor levels, key evidence domains, and background evidence are combined to form candidate records for structural correction.

9. The online monitoring method for early cavitation of centrifugal pumps based on reinforcement learning as described in claim 8, characterized in that: The obtained online monitoring results specifically include, The set of correctable parameters in the candidate records of structural correction is paired with the corresponding suspected cavitation location and cavitation precursor level. The correction amount of blade inlet angle, volute tongue clearance, impeller outlet width, local surface treatment range and operating condition suggestions are set to form candidate structural correction directions. By evaluating the direction of candidate structure correction based on factors such as reduced bubble germination, weakened blade suction surface separation, improved self-priming stability, maintained efficiency, and manufacturing accessibility, a multi-objective evaluation record is formed. Perform non-dominated sorting on the multi-objective evaluation records, retain the candidate structure correction directions that do not have a dominant relationship, and form a non-dominated candidate set; By utilizing hypervolume improvement, the structural correction direction with the largest hypervolume improvement value is selected from the non-dominated candidate set, and the online monitoring results are obtained.

10. A reinforcement learning-based online monitoring system for early cavitation in centrifugal pumps, based on the reinforcement learning-based online monitoring method for early cavitation in centrifugal pumps according to any one of claims 1 to 9, characterized in that: include, The structural sampling module performs parametric modeling of gaps and complex curved surfaces, and uses centrifugal pump design equation constraints and distance criteria to screen adjustable parameter combinations to obtain a parametric structural sample set. The map generation module performs joint simulation of unsteady multiphase flow and structural response on the parameterized structural sample set, establishes the evolution relationship between bubble volume distribution, pressure pulsation, low-pressure zone, trailing edge vorticity, entropy production concentration zone, head deviation and efficiency change, and obtains the cavitation evolution causal map. The state generation module collects real operating signals and converts them into synchronous compressed time-frequency maps. It combines synchronous compressed time-frequency maps, parameterized structured sample sets, and cavitation evolution causal maps to form a reinforcement learning state. The residual network extracts features, the critic network evaluates actions, and the class center distance feedback distinguishes states, generating cavitation precursor state records. The action screening module drives multi-agent reinforcement learning to make monitoring action decisions on cavitation precursor state records, and filters effective actions through counterfactual inference and ablation analysis to obtain online cavitation confirmation records; The results generation module performs correctable parameter mapping, non-dominated sorting, and overvolume improvement processing on the online cavitation confirmation records to obtain online monitoring results.

Images