A performance test system shared by a circulating pump and a submersible pump

Abstract

This invention relates to the field of fluid machinery testing technology, and in particular to a performance testing system shared by circulating pumps and submersible pumps. The system includes a data acquisition and control unit, a drive unit, a load adjustment unit, and an intelligent control and analysis server, configured to perform: startup and intelligent identification, digital twin model construction and test planning, adaptive basic performance sampling, boundary performance exploration, real-time safety monitoring and prediction, synchronous acquisition of multi-source heterogeneous information, and digital profile generation and model training. Before physical testing begins, this invention pre-calculates key feature points under all operating conditions through numerical simulation. Based on these pre-calculated key points, the system can automatically generate a dynamic test sequence that balances safety and efficiency, enabling subsequent real-world testing to be targeted and precise, capturing the core performance framework of the pump with the fewest possible tests, greatly improving the scientific rigor and efficiency of the testing.

Description

Technical Field

[0001] This invention relates to the field of fluid machinery testing technology, and in particular to a performance testing system for both circulating pumps and submersible pumps. Background Technology

[0002] As critical fluid transport equipment, the performance and reliability testing of circulating pumps and submersible pumps is a core component of product development, quality inspection, and condition assessment. Traditional performance testing involves simulating different operating conditions on a test bench, measuring parameters such as flow rate, head, power, and efficiency, and exploring their safe operating boundaries to plot performance curves and evaluate energy efficiency and reliability. This process is crucial for ensuring that pump products meet design specifications and achieve efficient and stable operation.

[0003] However, existing performance testing technologies have significant limitations. Mainstream methods rely on pre-set, fixed-step testing procedures, with operators manually or semi-automatically adjusting operating conditions and recording data based on experience. The entire process suffers from low intelligence, and testing efficiency and completeness heavily depend on human experience. Specifically, existing technologies typically cannot automatically identify the characteristics of unknown pump types before testing begins, resulting in a lack of targeted testing solutions. Test points are often evenly distributed or determined based on experience, making it difficult to accurately capture key features such as hump regions and cavitation thresholds, leading to wasted testing resources or missed features. Exploring performance and fault boundaries relies heavily on operator trial and error, posing safety risks and making standardization and systematization difficult. Safety monitoring is mostly passive, relying on over-threshold alarms, failing to anticipate potential risks. The final test results are often scattered data and curves, lacking in-depth integrated analysis and intuitive quantitative profiling of the overall equipment performance and health status. Furthermore, each testing stage is relatively independent, failing to form an intelligent closed loop capable of continuously accumulating knowledge and self-optimizing.

[0004] To address the aforementioned issues, this invention proposes a performance testing system shared by both circulating pumps and submersible pumps. This invention achieves intelligent pump identification by acquiring transient startup signals, and based on this, constructs a parameterized digital twin for simulation and test planning. This guides the physical testing system to perform adaptive sparse sampling and fine scanning. Simultaneously, a collaborative AI agent autonomously explores performance boundaries under safety constraints, and predictive safety protection is achieved through advanced simulation using the digital twin. Finally, the system integrates multi-source heterogeneous data and uses interpretable artificial intelligence to generate a comprehensive digital profile of the pump. Summary of the Invention

[0005] To overcome the problems mentioned in the background art, the present invention proposes a performance testing system that can be used for both circulating pumps and submersible pumps.

[0006] The technical solution of this invention is: a performance testing system shared by circulating pumps and submersible pumps, including a data acquisition and control unit, a drive unit, a load adjustment unit, and an intelligent control and analysis server. The data acquisition and control unit includes a test bench and a sensor array, and is configured to perform the following steps:

[0007] S11: Start-up and intelligent identification, start the pump under test at a preset extremely low frequency and power, and simultaneously collect multi-dimensional weak signals during the start-up transient.

[0008] S12: Digital twin model construction and test planning. Based on the identified type and parameters, combined with the input design parameters, a parameterized initial digital twin is constructed, and a sequence of key physical test target points, including boundary points and performance inflection points, is automatically planned through simulation analysis.

[0009] S13: Adaptive basic performance sampling. First, sparse scanning tests are performed on the key target points of the plan to obtain the skeleton of the performance curve. Then, the local features of the curve are calculated based on the measured data, and the density and distribution of subsequent sampling points are adaptively adjusted to complete the fine scanning.

[0010] S14: Boundary performance exploration. Based on the obtained basic performance data, the predictive AI agent is used to predict the performance boundary and drive the exploration AI agent to autonomously adjust the operating parameters under safety constraints, actively exploring the pump's fault boundary and performance boundary.

[0011] S15: Real-time security monitoring and prediction. During the test, the high-speed simulation of the digital twin is run in parallel to predict the system state in the future and issue an early warning and trigger active intervention before the system exceeds the safety threshold.

[0012] S16: Synchronous acquisition of multi-source heterogeneous information. During the test, time-series and image data from vibration, current, thermal imaging and noise sensors are acquired synchronously and aligned in time.

[0013] S17: Digital profile generation and model training. Integrate all test data and multi-source information, use interpretable machine learning models for training and analysis, generate a multi-dimensional, quantifiable comprehensive performance and health digital profile of the tested pump, and update the historical database.

[0014] As a preferred embodiment, the activation and intelligent identification steps specifically include:

[0015] S21: Control command sent. The intelligent control and analysis server sends a control command to the drive unit, which instructs the drive unit to start the pump under test at an initial frequency that is 10% to 20% lower than the rated frequency of the pump under test, and a corresponding starting torque that is lower than the rated torque.

[0016] S22: Multi-dimensional weak signal synchronous acquisition. During the startup process, electrical characteristic signals, mechanical characteristic signals and control response signals are synchronously acquired through a sensor array.

[0017] S23: Feature extraction and fusion judgment. The collected signal is processed to extract feature vectors representing the pump type and basic parameters, and input into the pre-trained classification judgment model to output the judgment result of the type and key parameters of the pump under test.

[0018] As a preferred option, the steps for building and testing a digital twin model specifically include:

[0019] S31: Model template matching and parameterized instantiation. Based on the identified pump type and parameters, the corresponding multi-domain physical model template is matched from the digital twin model library, and the identified parameters and user-input design parameters are used as initialization parameters to generate a parameterized initial digital twin.

[0020] S32: Simulation analysis and feature point pre-calculation. Within the initial digital twin, the performance curve cluster of the pump is calculated through numerical simulation over the entire operating range. Based on the mathematical characteristics of the curves and preset rules, the performance inflection point and safety boundary point are pre-calculated.

[0021] S33: Dynamic test sequence generation. All pre-calculated key target points are evaluated and sorted to generate a dynamic test sequence that guides the execution of physical tests.

[0022] As a preferred embodiment, the simulation analysis and feature point pre-calculation steps specifically include:

[0023] S321: Pre-location of the highest efficiency point. The predicted flow rate corresponding to the highest efficiency point is obtained by solving for the point where the first derivative of the efficiency-flow curve is zero.

[0024] S322: Head curve hump identification. By analyzing the head-flow curve, the point where the head change rate changes from negative to positive within a specific flow range is identified as the starting point of the hump.

[0025] S323: Cavitation Boundary Point Prediction. Based on the required net positive suction head (NPSH) model, the required NPSH of the pump at different flow rates is calculated and compared with the effective NPSH of the system to predict the critical flow rate at which cavitation occurs. The judgment condition is as follows:

[0026] ;

[0027] in, For safety margin, This represents the system's effective net positive suction head (NPSH). This is the required net positive suction head (NPSH).

[0028] S324: Minimum continuous thermally stable flow point prediction. Based on the pump power-flow curve and heat dissipation model, it calculates the pump temperature rise under different low flow conditions and predicts the critical flow rate at which the temperature rise exceeds the allowable limits of materials and seals.

[0029] S325: Maximum allowable working pressure point prediction. Based on the pump body structural strength model, the stress of the pump pressure-bearing components is calculated at different head near the shut-off point, and the critical head at which the allowable stress is reached is predicted.

[0030] As a preferred embodiment, the adaptive basic performance sampling step specifically includes:

[0031] S41: Sparse scanning of key target points. The control drive unit and load adjustment unit adjust the pump to each key target operating point in sequence according to the generated dynamic test sequence and wait for it to stabilize before performing the first measurement to obtain a sparse initial sampling point dataset that covers the key area of ​​the performance curve.

[0032] S42: Curve local feature analysis. Based on the obtained initial sampling point data, the target performance curve is piecewise fitted, and the local mathematical features of each curve segment are calculated.

[0033] S43: Adaptive sampling decision-making, which dynamically determines the location and density of new sampling points to be added based on the calculated local features;

[0034] S44: Incremental sampling and model update. Perform incremental testing and sampling at the new determined position, add the new data points to the dataset, and update the performance curve model.

[0035] S45: Iteration and Termination. Repeat steps S41 to S44 until the overall fitting accuracy of the target performance curve meets the preset threshold, thus completing the fine scan.

[0036] Preferably, in the adaptive sampling decision step, the adaptive sampling decision is based on curvature and a preset curvature threshold, with the following specific rules:

[0037] A11: For the smooth sections of a curve with curvature less than or equal to the minimum curvature threshold, set a low sampling density;

[0038] A12: For general variation segments of curves with curvature greater than the minimum curvature threshold and less than the maximum curvature threshold, set a medium sampling density;

[0039] A13: For high curvature segments of curves where the curvature is greater than or equal to the maximum curvature threshold, set a high sampling density;

[0040] A14: The specific location of the newly added sampling point is selected from the sub-interval with the largest prediction uncertainty in the corresponding interval.

[0041] As a preferred option, the boundary performance exploration steps specifically include:

[0042] S51: Agent initialization and modeling. Based on the obtained basic performance curves and multi-source historical data, the predictive AI agent and the exploratory AI agent are initialized, and a reinforcement learning exploratory environment model including state space, action space, safety constraints and reward function is constructed.

[0043] S52: Boundary exploration loop. By exploring the AI ​​agent, it selects and executes control actions based on the current state and its policy network, driving the test bench to change the working conditions. It predicts the AI ​​agent's assessment of state risks and predicts the boundary. The system monitors safety constraints and collects multi-source response data.

[0044] S53: Boundary confirmation and marking. When the exploration process triggers the boundary determination condition, the system confirms and records the precise working condition and multi-source feature data of the boundary point.

[0045] S54: Model update and policy optimization. Using the state-action-reward data sequence generated by exploration, update the policy network of the exploration AI agent, and use the newly discovered boundary data to optimize the prediction AI agent model.

[0046] S55: Exploration terminated. When the preset exploration termination conditions are met, the test in this phase will be terminated.

[0047] As a preferred option, the real-time security monitoring and prediction steps specifically include:

[0048] S61: Real-time data synchronization and twin state initialization, synchronously injecting the real-time control commands and sensor data streams of the physical test system into the digital twin, so that the virtual state of the twin is aligned with the current state of the physical system;

[0049] S62: Future state multi-step prediction. Using the current alignment state as the initial condition, the digital twin runs a closed-loop simulation at a speed faster than real-time to predict the evolution trajectory of key system state parameters within a preset time window in the future.

[0050] S63: Predictive assessment of security risks, analyzing predicted trajectories to determine whether there is a risk of violating preset security constraints at any future time;

[0051] S64: Tiered early warning and proactive intervention: Based on the urgency and severity of the risk, different levels of early warning are triggered and predetermined safety control strategies are automatically executed.

[0052] As a preferred embodiment, the digital profile generation and model training steps specifically include:

[0053] S71: Multi-source heterogeneous data fusion and feature engineering, which gathers all the obtained structured test data and unstructured multi-source sensing data, performs spatiotemporal alignment, cleaning and feature extraction, and constructs a feature matrix;

[0054] S72: Interpretable profile model training, using the feature matrix as input, trains an interpretable machine learning model, and the output of the machine learning model is a multi-dimensional quantitative indicator representing the overall state of the pump.

[0055] S73: Digital profile synthesis and report generation, which compares and combines the quantitative indicators output by the model with key features and thresholds to generate a structured digital profile report that includes multi-dimensional labels and scores.

[0056] S74: Knowledge base iterative update, archive the complete data chain, feature matrix, digital profile and corresponding interpretable model feature contribution of this test to the historical database.

[0057] Preferably, the interpretable machine learning model in the training step of the interpretable profile model is a tree ensemble model. The input of the tree ensemble model is a feature vector, and the output is a series of quantifiable scores, including:

[0058] The overall performance score characterizes the degree of conformity with design standards for similar pumps.

[0059] Mechanical health rating indicates the condition of bearings and shaft systems;

[0060] Electrical health level, characterizing the electrical condition of a motor;

[0061] Operational stability score, comprehensive evaluation of vibration and noise.

[0062] The beneficial effects of this invention are:

[0063] 1. Compared to existing technologies that typically rely on engineers' experience or fixed standards to select test conditions, this approach is highly subjective and inertial, potentially overlooking important performance inflection points and safety boundaries, making it difficult to scientifically guarantee test efficiency and completeness. This invention, before initiating physical testing, first constructs a parameterized initial digital twin based on the identification results, and pre-calculates the performance curve clusters under all operating conditions, as well as key feature points such as the hump region and cavitation boundaries, through numerical simulation. Based on these pre-calculated key points, the system can automatically generate a dynamic test sequence that balances safety and efficiency. This deep integration of virtual simulation and physical testing enables subsequent real-world testing to be targeted and precise, capturing the pump's core performance framework with minimal testing attempts, greatly improving the scientific rigor and efficiency of the testing.

[0064] 2. Compared to existing technologies that generally use equal-interval or fixed-pattern sampling methods for operating point sampling, which waste resources in flat areas of the performance curve and may lose details due to insufficient sampling in critical areas of drastic curve changes, this invention adopts a two-stage adaptive fine sampling strategy. First, it performs sparse scanning on the key points of the digital twin planning to outline the performance skeleton. Then, it calculates features such as the local curvature of the curve based on the measured data and drives the Bayesian optimization model to evaluate the prediction uncertainty. Based on this, it dynamically decides the location and density of new sampling points, so that the test resources can be intelligently concentrated in areas with complex curve shapes and high uncertainty. This achieves rapid convergence to a high-precision performance curve with minimal testing costs and is a successful application of optimal experimental design in automated testing.

[0065] 3. Compared to existing technologies that rely primarily on operator experience and cautious trial-and-error to explore pump performance and fault boundaries, this method is inefficient, lacks repeatability, and carries high safety risks, making it difficult to systematically discover all potential boundaries. This invention introduces a dual-agent collaborative mechanism of predictive and exploratory agents. The predictive agent acts as a seer, predicting performance mutation trends and boundary proximity risks based on multi-source data. The exploratory agent acts as a driver, autonomously learning through reinforcement learning how to adjust operating conditions within hard safety constraints to efficiently and safely approach and confirm various boundaries. This transforms the boundary exploration process into an intelligent optimization search problem within a safety cage, not only automating and intelligentizing the exploration process but also proactively discovering hidden boundaries that are difficult for humans to detect, greatly expanding the depth and breadth of testing.

[0066] 4. Compared to existing technologies that primarily employ passive safety monitoring and protection based on fixed thresholds—that is, triggering alarms or shutdowns only after parameters exceed limits—this approach is reactive and may cause unnecessary transient shocks to the equipment or unplanned test interruptions. This invention constructs a digital twin that runs synchronously and in parallel with the physical system, enabling it to perform closed-loop simulations faster than real-time. It continuously predicts the system's state evolution trajectory in the near future. By calculating the time margin for the predicted trajectory to first exceed the safety threshold, the system can achieve tiered early warning and early intervention. This shifts the safety defense line from the current threshold wall to a future time window, achieving a fundamental shift from passive response to predictive proactive defense. While maximizing equipment safety, it allows testing to approach real limits more closely.

[0067] 5. Compared to existing technologies that generate test reports that are mostly simple lists of performance curves and key data points, lacking a comprehensive and quantifiable evaluation of pump health status and operational quality, and whose judgment criteria and processes are often opaque; this invention deeply integrates multi-source heterogeneous information such as vibration, current, thermal imaging, and noise, and extracts a large number of deep features characterizing performance, mechanical, electrical, thermal, and acoustic states. It uses interpretable machine learning models to fuse and analyze these features, outputting multi-dimensional quantitative scores including comprehensive performance, mechanical health, electrical health, and operational stability. It also uses feature contribution analysis technology to clearly show the key basis behind each score; generating a comprehensive, quantitative, transparent, and traceable digital profile for each tested pump, making it an authoritative digital archive describing the pump's comprehensive status and characteristics. Attached Figure Description

[0068] Figure 1 The diagram shown is a schematic of the performance testing system for both circulating pumps and submersible pumps of the present invention.

[0069] Figure 2 The diagram shows the working process of the performance testing system for both circulating pumps and submersible pumps of the present invention. Detailed Implementation

[0070] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0071] Please see Figure 1 - Figure 2 This invention provides an embodiment: a performance testing system shared by circulating pumps and submersible pumps, including a data acquisition and control unit, a drive unit, a load adjustment unit, and an intelligent control and analysis server. The data acquisition and control unit includes a test bench and a sensor array, and is configured to perform the following steps:

[0072] Step 1: Startup and Intelligent Recognition

[0073] The pump under test is started at a preset extremely low frequency and power, and multi-dimensional weak signals during the start-up transient are collected simultaneously. The pump model and parameters are intelligently identified, including:

[0074] S21: Control command sent. The intelligent control and analysis server sends a control command to the drive unit, which instructs the drive unit to start the pump under test at an initial frequency that is 10% to 20% lower than the rated frequency of the pump under test, and a corresponding starting torque that is lower than the rated torque.

[0075] S22: Multi-dimensional weak signal synchronous acquisition. During the startup process, electrical characteristic signals, mechanical characteristic signals and control response signals are synchronously acquired through a sensor array.

[0076] S23: Feature extraction and fusion judgment. The collected signal is processed to extract feature vectors representing the pump type and basic parameters, and input into the pre-trained classification judgment model to output the judgment result of the type and key parameters of the pump under test.

[0077] In this embodiment, the electrical characteristic signals in the multi-dimensional weak signal synchronous acquisition step include:

[0078] Instantaneous values ​​of the three-phase starting current of the motor, collected by a power analyzer;

[0079] The corresponding instantaneous value of the starting voltage;

[0080] Based on current and voltage signals, the instantaneous active power and total harmonic distortion (THD) of the starting current during the startup process are calculated. The THD of the starting current is calculated using the following formula:

[0081] ;

[0082] in, To the total harmonic distortion of the starting current, This is the total effective value of the starting current. The fundamental effective value of the starting current is given, and the total effective value of the starting current is calculated using the following formula:

[0083] ;

[0084] The fundamental effective value of the starting current is calculated using the following formula:

[0085] ;

[0086] in, For the integration time period, The instantaneous value of the current varies with time. For Fast Fourier Transform, subscript This indicates that the component corresponding to the fundamental frequency in the FFT transform result is taken.

[0087] In this embodiment, the mechanical characteristic signals in the multi-dimensional weak signal synchronous acquisition step include:

[0088] The vibration time-domain signal during startup was acquired by a vibration acceleration sensor mounted on the pump bearing housing;

[0089] Start-up load torque estimated by a torque sensor or via electrical signals

[0090] The vibration spectrum at the initial moment of startup, as well as the main frequency and sideband characteristics of the vibration spectrum, are obtained by performing a fast Fourier transform on the vibration signal.

[0091] Based on the rotation equation, the resultant moment of inertia of the system is estimated using the difference between the driving torque and the load torque. The estimation formula is as follows:

[0092] ;

[0093] in, To synthesize the moment of inertia, For driving torque, For load torque, This represents the change in rotational speed during the startup process.

[0094] In this embodiment, the control response signal in the multi-dimensional weak signal synchronous acquisition step is:

[0095] The error curve formed by the drive frequency command output by the speed closed-loop controller of the drive unit and the feedback of the actual speed of the pump under test;

[0096] The calculated error curve represents the integral absolute value error during the startup process:

[0097] ;

[0098] in, The integral absolute value error of the difference curve during the startup process. The error curve is shown below. The time to reach steady state.

[0099] In this embodiment, the pre-trained classification and judgment model is a support vector machine, a random forest, or a lightweight deep neural network.

[0100] The classification model takes a feature vector representing the pump type and basic parameters as input and outputs the following:

[0101] Type labels, including circulating pumps, submersible pumps, and others;

[0102] Rated power ratings include low power, medium power, and high power;

[0103] Impeller type prediction, including centrifugal, mixed flow and axial flow tendencies;

[0104] The model's training data comes from the start-up characteristic data of pumps with known types and parameters accumulated from historical tests.

[0105] Step Two: Digital Twin Model Construction and Testing Planning

[0106] Based on the identified type and parameters, combined with the input design parameters, a parameterized initial digital twin is constructed. Simulation analysis is then used to automatically plan a sequence of key physical test target points, including boundary points and performance inflection points. Specifically, this includes:

[0107] S31: Model template matching and parameterized instantiation. Based on the identified pump type and parameters, the corresponding multi-domain physical model template is matched from the digital twin model library, and the identified parameters and user-input design parameters are used as initialization parameters to generate a parameterized initial digital twin.

[0108] S32: Simulation analysis and feature point pre-calculation. Within the initial digital twin, the performance curve cluster of the pump is calculated through numerical simulation over the entire operating range. Based on the mathematical characteristics of the curves and preset rules, the performance inflection point and safety boundary point are pre-calculated.

[0109] S33: Dynamic test sequence generation. All pre-calculated key target points are evaluated and sorted to generate a dynamic test sequence that guides the execution of physical tests.

[0110] Specifically, the model templates in the digital twin model library contain coupled equations from the fields of fluid dynamics, electrical machinery, rotor dynamics, and thermodynamics; the initialization parameters include rated flow rate, rated head, rated speed, specific speed, efficiency, moment of inertia, and material properties.

[0111] Specifically, the parameterized initial digital twin is defined by a set of core equations, and its head-flow performance curve is parameterized by the following characteristic polynomial:

[0112] ;

[0113] in, , and The coefficients are determined by interpolation from a family of typical curves obtained by mapping from the model library based on the identified pump type, specific speed, and design point. The head of the test pump is a function of the flow rate. To test the volumetric flow rate of the pump.

[0114] As a preferred embodiment, the simulation analysis and feature point pre-calculation steps specifically include:

[0115] S321: Pre-location of the highest efficiency point. The predicted flow rate corresponding to the highest efficiency point is obtained by solving the equation where the first derivative of the efficiency-flow curve is zero.

[0116] ;

[0117] in, Test the pump efficiency. This indicates that the efficiency of the test pump is based on a function of the flow rate. To test the volumetric flow rate of the pump;

[0118] S322: Head curve hump identification. By analyzing the head-flow curve, the point where the rate of change of head changes from negative to positive within a specific flow range is identified as the starting point of the hump. The mathematical expression of the judgment logic is as follows:

[0119] like and ,but ;

[0120] in, The flow rate at the current observation point. The flow rate is the value of the previous adjacent point. The flow rate is the value of the next adjacent point. In order to traffic The corresponding head value at that time In order to traffic The corresponding head value at that time In order to traffic The corresponding head value at that time This represents the flow rate value corresponding to the starting point of the hump area.

[0121] S323: Cavitation Boundary Point Prediction. Based on the required net positive suction head (NPSH) model, the required NPSH of the pump at different flow rates is calculated and compared with the effective NPSH of the system to predict the critical flow rate at which cavitation occurs. The judgment condition is as follows:

[0122] ;

[0123] in, For safety margin, This represents the system's effective net positive suction head (NPSH). This is the required net positive suction head (NPSH).

[0124] S324: Minimum continuous thermally stable flow point prediction. Based on the pump power-flow curve and heat dissipation model, it calculates the pump temperature rise under different low flow conditions and predicts the critical flow rate at which the temperature rise exceeds the allowable limits of materials and seals.

[0125] S325: Maximum allowable working pressure point prediction. Based on the pump body structural strength model, the stress of the pump pressure-bearing components is calculated at different head near the shut-off point, and the critical head at which the allowable stress is reached is predicted.

[0126] In this embodiment, the dynamic test sequence generation step specifically includes:

[0127] S331: Generate a list of key target points, adding all pre-calculated points and the initial target point list;

[0128] S332: Target point sorting and path planning. The target point list is sorted based on safety and efficiency principles. The sorting rules include:

[0129] (1) Prioritize testing low-risk points, and then test high-risk boundary points;

[0130] (2) On the flow axis, prioritize testing the middle flow point, and then extend to both ends of the large and small flow to reduce drastic load changes;

[0131] (3) Merge physically close points into the same test subsequence to reduce the time for switching between operating conditions.

[0132] S333: Generates a dynamic test sequence instruction set, which transforms the sorted list of target points into a series of control instructions that can be recognized by the test execution unit, including target flow rate, target pressure, target speed, allowable test time at that point, and safety monitoring threshold.

[0133] Step 3: Adaptive Basic Performance Sampling

[0134] First, a sparse scan test is performed on the key target points of the plan to obtain the skeleton of the performance curve. Then, based on the measured data, the local features of the curve are calculated, and the density and distribution of subsequent sampling points are adaptively adjusted to complete the fine scan, which specifically includes:

[0135] S41: Sparse scanning of key target points. The control drive unit and load adjustment unit adjust the pump to each key target operating point in sequence according to the generated dynamic test sequence and wait for it to stabilize before performing the first measurement to obtain a sparse initial sampling point dataset that covers the key area of ​​the performance curve.

[0136] S42: Curve local feature analysis. Based on the obtained initial sampling point data, the target performance curve is piecewise fitted, and the local mathematical features of each curve segment are calculated.

[0137] S43: Adaptive sampling decision-making, which dynamically determines the location and density of new sampling points to be added based on the calculated local features;

[0138] S44: Incremental sampling and model update. Perform incremental testing and sampling at the new determined position, add the new data points to the dataset, and update the performance curve model.

[0139] S45: Iteration and Termination. Repeat steps S41 to S44 until the overall fitting accuracy of the target performance curve meets the preset threshold, thus completing the fine scan.

[0140] Specifically, the steps for analyzing the local features of a curve include:

[0141] S421: Based on the initial sampling points, a piecewise polynomial fitting method is used to generate the initial head-flow performance curve function and efficiency-flow performance curve function.

[0142] S422: Within the flow domain, calculate the curvature of the initial head-flow performance curve function at each point, where the formula for calculating the curvature is:

[0143] ;

[0144] in, The first derivative of the head-flow performance curve function. The second derivative of the head-flow performance curve function. The initial head-flow performance curve function at point The curvature.

[0145] Preferably, in the adaptive sampling decision step, the adaptive sampling decision is based on curvature and a preset curvature threshold, with the following specific rules:

[0146] A11: For the smooth sections of a curve with curvature less than or equal to the minimum curvature threshold, set a low sampling density;

[0147] A12: For general variation segments of curves with curvature greater than the minimum curvature threshold and less than the maximum curvature threshold, set a medium sampling density;

[0148] A13: For high curvature segments of curves where the curvature is greater than or equal to the maximum curvature threshold, set a high sampling density;

[0149] A14: The specific location of the newly added sampling point is selected from the sub-interval with the largest prediction uncertainty in the corresponding interval.

[0150] In this embodiment, the prediction uncertainty is evaluated by a Bayesian optimization model. The Bayesian optimization model uses the data from the previous sampling points as the observation set and outputs the mean and variance of the predicted values ​​of the performance curve at each unsampled flow point Q. The variance is used to characterize the prediction uncertainty. New sampling points are preferentially selected within the interval that satisfies the density rule, so that the flow point with the largest variance is selected.

[0151] Specifically, after each incremental sampling, the curve is refitted using a new dataset that includes all historical sampling points, and the curvature distribution and parameters of the Bayesian optimization model are updated.

[0152] Specifically, the termination condition in the iteration and termination steps is:

[0153] Within the entire flow definition domain, the maximum value of the half-width of the confidence interval for curve prediction is less than the preset accuracy threshold.

[0154] Step 4: Boundary Performance Exploration

[0155] Based on the obtained basic performance data, a predictive AI agent is used to predict performance boundaries and drive the exploration AI agent to autonomously adjust operating parameters under safety constraints, actively exploring the pump's fault boundaries and performance boundaries, specifically including:

[0156] S51: Agent initialization and modeling. Based on the obtained basic performance curves and multi-source historical data, the predictive AI agent and the exploratory AI agent are initialized, and a reinforcement learning exploratory environment model including state space, action space, safety constraints and reward function is constructed.

[0157] S52: Boundary exploration loop. By exploring the AI ​​agent, it selects and executes control actions based on the current state and its policy network, driving the test bench to change the working conditions. It predicts the AI ​​agent's assessment of state risks and predicts the boundary. The system monitors safety constraints and collects multi-source response data.

[0158] S53: Boundary confirmation and marking. When the exploration process triggers the boundary determination condition, the system confirms and records the precise working condition and multi-source feature data of the boundary point.

[0159] S54: Model update and policy optimization. Using the state-action-reward data sequence generated by exploration, update the policy network of the exploration AI agent, and use the newly discovered boundary data to optimize the prediction AI agent model.

[0160] S55: Exploration terminated. When the preset exploration termination conditions are met, the test in this phase will be terminated.

[0161] In this embodiment, the agent initialization and modeling steps specifically include:

[0162] S511: Define the state space. The state of the state space includes the current operating parameters, key performance indicators, and distance from the predicted boundary.

[0163] S512: Define the action space. The actions in the action space are discrete or continuous control commands to the actuators of the test bench, including adjusting the opening change of the flow control valve, adjusting the opening change of the outlet pressure control valve, and adjusting the frequency change of the drive unit.

[0164] S513: Define a set of safety constraints. The constraints in the set of safety constraints are inviolable hard conditions, including maximum current, maximum vibration intensity, maximum temperature, minimum flow rate, and maximum pressure.

[0165] S514: Initialize the predictive AI agent. The predictive AI agent is a neural network model trained on historical data. The input is the state, and the output is the performance index prediction for the next state and the boundary proximity probability vector.

[0166] S515: Initialize the exploration AI agent. The exploration AI agent is a policy network based on reinforcement learning. The parameters of the exploration AI agent are randomly initialized to form the probability distribution of the output action in a given state.

[0167] Specifically, the reward function defined by the AI ​​agent is composed of the following weighted components:

[0168] A21: Boundary Discovery Reward: Triggered when exploration causes a sudden change in a key performance indicator, and the rate of change exceeds a threshold.

[0169]

[0170] in, Rewards for finding boundaries The weighting coefficients for the boundary discovery reward. For time step t, For time step t+1, The mutation threshold;

[0171] A22: Boundary Approach Reward: Encourages the agent to approach but not exceed the predicted boundary, and is positively correlated with the probability of approaching the predicted AI output boundary.

[0172] ;

[0173] in, Rewards for approaching the boundary; The weighting coefficients are those that are close to the boundary of the reward. For general safety boundary distance, This is the cavitation boundary distance. This is the distance to the overheating boundary.

[0174] A23: Safety Penalty: A penalty is imposed when a state approaches or violates a safety constraint. The penalty for violating a constraint is extremely high. Specifically:

[0175] ;

[0176] in, For safety reasons, To approximate the constraint penalty weight, To impose penalty weights for violating constraints, For a set of safety constraints, For a single security constraint, As a proximity constraint indicator function, when the state Approaching but not violating constraints The value is 1 when the time is right, and 0 otherwise. For a violation of the constraint indication function, when the state Actual violation of constraints The value is 1 when it is active, and 0 otherwise.

[0177] In this embodiment, the boundary exploration loop step specifically includes:

[0178] S521: In each exploration step, the exploration AI agent samples an action based on the current state through its policy network;

[0179] S522: Perform the action strategy network sampling action. The test bench smoothly adjusts to the new working condition within a preset short time. After the system stabilizes, the data of the new state is collected.

[0180] S523: The predictive AI agent predicts the next state based on the current state and the current action, and compares the predicted value with the actual value. If the deviation exceeds the threshold, the model uncertainty in that area is marked as high.

[0181] S524: Check new state: whether it violates the set of safety constraints. If it does, immediately perform a safety recovery action and terminate the current exploration sequence.

[0182] In this embodiment, the predictive AI agent is a time-series prediction model based on a long short-term memory network, deeply integrated with the field of pump performance testing. It is used to identify early patterns of performance mutations and quantify the risk of approaching boundaries from current and historical multi-source system data. Specifically, the agent takes a sequence of system states from the current and several past time steps as input. This sequence includes not only real-time operating parameters but also deeply integrates multi-source information characterizing operational health, such as flow rate, head, speed, efficiency, the dominant frequency amplitude of the vibration spectrum, the total harmonic distortion rate of the current, and the instantaneous bearing temperature. Its core output consists of two key vectors: one is a multi-step prediction of changes in key performance indicators in the near future, used to predict inflection point trends such as a sharp drop in efficiency or head failure; the other is a boundary approach probability vector, where each element represents the quantified risk probability of the system evolving towards a specific performance or fault boundary (such as cavitation boundary, overheating boundary, or vibration exceeding limit boundary) at the current position. This model learns the complex mapping relationship between multi-source heterogeneous features and boundary states by being trained on a large amount of historical test data, especially data containing various boundary cases. This enables it to serve as a prophet and risk assessor for exploring AI agents, providing crucial forward-looking information for their safe and efficient exploration decisions.

[0183] In this embodiment, the exploration AI agent is a reinforcement learning agent deeply coupled with the water pump boundary exploration test scenario and trained based on a near-end policy optimization algorithm. In a simulated-realistic environment comprised of a digital twin simulation and a physical test bench, it learns how to efficiently and safely manipulate the test system to discover unknown performance boundaries. The agent's input is a high-dimensional state space vector, which integrates real-time readings from multiple sensors such as flow rate, pressure, rotational speed, and vibration, as well as quantitative information on distances to various estimated boundaries provided by the predictive AI agent. Its output is specific control commands in the action space, calculated by its deep neural network policy model based on the current input state. The agent's unique feature lies in its carefully designed reward function, which encourages the discovery of performance mutations while severely penalizing any behavior that approaches or violates hard safety constraints. This ensures that it can autonomously learn a bold yet cautious boundary exploration strategy in real physical testing, rather than engaging in blind random attempts.

[0184] In this embodiment, the boundary determination condition in the boundary confirmation and marking step is that any of the following conditions are met:

[0185] A31: Performance mutation boundary: The absolute value of the derivative of a key performance indicator exceeds a threshold;

[0186] A32: Safety Hard Boundary: Any monitored parameter reaches its safety alarm threshold, but does not trigger an emergency shutdown;

[0187] A33: Loss of stability boundary: Uncontrollable periodic oscillations occur in flow or pressure, and the amplitude exceeds the set proportion of the stability threshold;

[0188] Once a boundary is confirmed, the system records complete multi-source data for that point and marks the boundary type.

[0189] In this embodiment, the preset conditions for terminating the exploration in the exploration termination step specifically include:

[0190] A41: Within the specified number of exploration steps, predefined target boundary types exceeding a preset proportion have been discovered and confirmed;

[0191] A42: When the cumulative reward obtained by the AI ​​agent in multiple consecutive exploration sequences is lower than a preset threshold, it indicates that the discovery of new boundaries tends to stagnate.

[0192] A43: The total system uptime or cumulative load has reached the preset safe exploration limit.

[0193] Step 5:

[0194] Real-time security monitoring and prediction involves running high-speed simulations of the digital twin in parallel during testing to predict the system's state over a future period. It also issues warnings and triggers proactive intervention before the predicted safety threshold is exceeded. Specifically, this includes:

[0195] S61: Real-time data synchronization and twin state initialization, synchronously injecting the real-time control commands and sensor data streams of the physical test system into the digital twin, so that the virtual state of the twin is aligned with the current state of the physical system;

[0196] S62: Future state multi-step prediction. Using the current alignment state as the initial condition, the digital twin runs a closed-loop simulation at a speed faster than real-time to predict the evolution trajectory of key system state parameters within a preset time window in the future.

[0197] S63: Predictive assessment of security risks, analyzing predicted trajectories to determine whether there is a risk of violating preset security constraints at any future time;

[0198] S64: Tiered early warning and proactive intervention: Based on the urgency and severity of the risk, different levels of early warning are triggered and predetermined safety control strategies are automatically executed.

[0199] In this embodiment, the real-time data synchronization and twin state initialization steps specifically include:

[0200] S611: Real-time acquisition of real-time data streams, including real-time control commands from the physical system and sensor feedback vectors;

[0201] S612: At each synchronization moment, the digital twin receives an estimate of the current state vector of the physical system and, using the received control commands, resets the virtual state vector to be aligned with the current state vector through its internal state observer, thus completing the state initialization.

[0202] Specifically, in the multi-step prediction of future states, the digital twin performs a prediction simulation at each synchronization moment to predict the system state within a future time period, where the future time period is set to 20 to 30 seconds. The prediction simulation is achieved by solving the state-space equations of the digital twin, which are expressed as follows:

[0203] ;

[0204] ;

[0205] in, For virtual state vectors, Projecting inputs for future control. For model parameters, To predict the output, This is the output function;

[0206] Among them, future control input projection The setting rules are as follows:

[0207] During the most recent control cycle, the control commands of the current physical system are used.

[0208] In the more distant future, if the test plan is known, the plan instructions will be used; otherwise, the control inputs will be assumed to remain unchanged.

[0209] In this embodiment, the predictive assessment of safety risks is achieved by analyzing the predicted output trajectory, specifically as follows:

[0210] S631: Define key safety parameters and their corresponding safety thresholds, including vibration intensity, motor current, bearing temperature, and outlet pressure;

[0211] S632: Calculate the time when the predicted trajectory of each safety parameter first exceeds its safety threshold in the future prediction time domain;

[0212] S633: Calculate the predicted safe time margin from the current time to the first time of exceeding the limit. Based on the shortest predicted safe time margin and the severity of the exceeding parameter, conduct a risk assessment. The formula for calculating the predicted safe time margin is as follows:

[0213] ;

[0214] in, This represents a key security parameter that is being monitored. Representing the current moment, The predicted time of exceeding the limit represents the parameter J. This represents the predicted safety margin for parameter J.

[0215] In this embodiment, tiered early warning and proactive intervention are triggered based on the predicted safety time margin and risk level, specifically tiered as follows:

[0216] A51: Level 1 warning. When the predicted safety time margin is less than the high threshold but greater than the low threshold, the system sends a visual and audible warning signal to the main control interface, but does not interrupt the current test.

[0217] A52: Level 2 warning. When the predicted safety time margin is less than or equal to the low threshold but greater than zero, the system issues a warning and sends an instruction to the test execution unit to prepare to execute the preset mitigation safety strategy, including reducing the test change rate or pausing the progression to more dangerous conditions.

[0218] A53: Level 3 intervention. When it is predicted that an over-limit will occur in a very short time or has already occurred, the system immediately overrides the current test command and automatically executes a mandatory safety protection strategy, including reducing the drive frequency at a preset smooth rate, gradually opening the safety relief valve, and executing a sequential shutdown.

[0219] Specifically, the mitigation and protection security strategies are optimized control sequences selected from the strategy library based on the effects of different intervention measures simulated by the digital twin in the predictive simulation. These sequences enable the system to escape the risk state most quickly and smoothly.

[0220] Step Six:

[0221] Synchronous acquisition of multi-source heterogeneous information: During the test, time-series and image data from vibration, current, thermal imaging, and noise sensors are simultaneously acquired and synchronized using a high-precision hardware clock. The multi-source heterogeneous information specifically includes:

[0222] Vibration signals containing spectral information are collected by a vibration acceleration sensor;

[0223] Motor current and voltage signals containing harmonic components are acquired by a power analyzer or a high-precision current probe;

[0224] Temperature field distribution images of key parts of the pump body and motor, periodically captured by an infrared thermal imager;

[0225] Operating noise signals containing spectral characteristics are collected by acoustic sensors.

[0226] Step 7: Digital Profile Generation and Model Training

[0227] By integrating all test data and multi-source information, and using an interpretable machine learning model for training and analysis, a multi-dimensional and quantifiable comprehensive performance and health profile of the tested pump is generated, and the historical database is updated. Specifically, this includes:

[0228] S71: Multi-source heterogeneous data fusion and feature engineering, which gathers all the obtained structured test data and unstructured multi-source sensing data, performs spatiotemporal alignment, cleaning and feature extraction, and constructs a feature matrix;

[0229] S72: Interpretable profile model training, using the feature matrix as input, trains an interpretable machine learning model, and the output of the machine learning model is a multi-dimensional quantitative indicator representing the overall state of the pump.

[0230] S73: Digital profile synthesis and report generation, which compares and combines the quantitative indicators output by the model with key features and thresholds to generate a structured digital profile report that includes multi-dimensional labels and scores.

[0231] S74: Knowledge base iterative update, archive the complete data chain, feature matrix, digital profile and corresponding interpretable model feature contribution of this test to the historical database.

[0232] In this embodiment, the multi-source heterogeneous data fusion and feature engineering steps specifically include:

[0233] S711: Data alignment, based on a unified timestamp, aligns time-series performance data, vibration spectrum data, current harmonic data, thermal imaging temperature field data, and noise spectrum data.

[0234] S712: Feature extraction and computation. This involves extracting and computing features from the aligned data to construct a feature vector. The computed features include:

[0235] A61: Performance characteristics, including flow rate, head, and efficiency at the optimal efficiency point; flow rate range in the high-efficiency zone; head curve flatness index;

[0236] A62: Vibration health characteristics, including the amplitude of bearing pass frequency and its harmonics, as well as the amplitude of rotational frequency under BEP conditions, and calculation of the comprehensive vibration health index. The formula for calculating the comprehensive vibration health index is as follows:

[0237]

[0238] in, To comprehensively assess the vibration health index, The vibration amplitude is the frequency of rotation. This represents the amplitude of the frequency at which the inner ring of the bearing passes. The amplitude of the frequency at which the bearing outer ring passes. The amplitude of the rotation frequency of the rolling element. , and These are the weighting coefficients. It is a very small positive number;

[0239] A63: Electrical health characteristics, including total harmonic distortion of current under rated load and amplitude ratio of specific order harmonics to the fundamental frequency, which characterizes the health of the rotor bars;

[0240] A64: Thermal characteristics, including the temperature rise and temperature difference of the bearing housings at the motor drive end and non-drive end;

[0241] A65: Acoustic characteristics, including noise sound pressure level in a specific frequency band under BEP conditions.

[0242] In this embodiment,

[0243] Preferably, the interpretable machine learning model in the training step of the interpretable profile model is a tree ensemble model. The input of the tree ensemble model is a feature vector, and the output is a series of quantifiable scores, including:

[0244] The overall performance score characterizes the degree of conformity with design standards for similar pumps.

[0245] Mechanical health rating indicates the condition of bearings and shaft systems;

[0246] Electrical health level, characterizing the electrical condition of a motor;

[0247] Operational stability score, comprehensive evaluation of vibration and noise;

[0248] Specifically, the interpretability of a machine learning model is achieved by calculating the contribution of each input feature to each output decision. The contribution is calculated using the SHAP value. Specifically, the contribution of feature i to the output of sample y is approximated by the following formula:

[0249] ;

[0250] in, Features SHAP value, For machine learning models, For the set of all features, Index the specific features for which the contribution is being calculated. For feature set A subset of, and Not containing features , Representative model Using only subsets When considering features in a sample The predicted value.

[0251] In this embodiment, the generated digital profile report is a structured data object, containing the following dimensions:

[0252] The identification and specifications include the pump model, unique test ID, and identified basic parameters.

[0253] The performance profile dimension includes the overall performance score and the percentage deviation of key performance characteristics from their design values ​​or the average value of similar products;

[0254] The health status profile includes mechanical health level, electrical health level, key features that lead to the level judgment and their SHAP contribution;

[0255] The operational quality profile includes operational stability score, vibration intensity level, and noise level;

[0256] The characteristics and boundary profile dimensions include the key boundary points discovered and their corresponding operating conditions, as well as a description of the dynamic response characteristics to changes in flow and pressure.

[0257] In this embodiment, the knowledge base iterative update step specifically includes:

[0258] S741: Store the complete data chain of this test, including raw sensor data, processed feature matrix, digital profile report, and feature contribution data provided by the interpretable model, as a data package;

[0259] S742: Expand the historical database with new data packets and retrain and fine-tune the startup identification model, predictive AI agent model and explore the value function of AI agent using the updated database;

[0260] S743: Use the newly acquired actual performance curves and boundary data to calibrate and optimize the parameters of the digital twin model.

[0261] Example 1:

[0262] This embodiment uses a submersible pump with a rated power of 15 kW and an unknown model as the test object. The intelligent control and analysis server first sends a command to the drive unit to start the pump under test at an extremely low frequency of 12 Hz, which is about 20% of its rated frequency. At the moment of startup, the sensor array simultaneously collects the starting current, voltage, bearing housing vibration, and controller error signals. Data analysis shows that the total harmonic distortion rate of the starting current is high, and the estimated system composite moment of inertia is large. These features are extracted as feature vectors and input into a pre-trained random forest model. The model quickly determines that the pump is a "submersible pump," its rated power level is "medium power," and its impeller type is likely "mixed flow."

[0263] Based on this identification result, the system matched a submersible pump template from the model library and, combined with user-input design parameters such as a rated flow rate of 80 cubic meters per hour and a rated head of 30 meters, constructed an initial digital twin of the pump. Simulation analysis automatically pre-calculated seven key target points on its performance curve, including the highest efficiency point, the starting point of the hump region, and the cavitation critical point, and generated a dynamic test sequence according to safety principles. Adaptive basic performance sampling then began. The system first controlled the load unit, sequentially adjusting the pump to these seven planned points for steady-state measurement, obtaining the skeleton of the head-flow rate curve. Based on these sparse points, the system calculated the local curvature of the curve, finding that the curve was flat in the high-efficiency region, while the curvature was larger in the low-flow region. Therefore, the Bayesian optimization model decided to add three fine sampling points each in the low-flow region and at the end of the high-flow region. After two rounds of iteration, a fine scan was completed with a total of thirteen test points, and the overall prediction confidence interval half-width was less than one percent, meeting the accuracy requirements.

[0264] Next, boundary performance exploration was initiated. Based on the obtained basic performance curves, the predictive AI agent (a long short-term memory network model) assessed the risk of the current state being close to various boundaries. The exploratory AI agent (a near-end policy optimization network) attempted to fine-tune the outlet valve opening and drive frequency within the safety constraint set to explore higher head conditions. During the exploration process, when the flow rate dropped to approximately 15 cubic meters per hour, the amplitude of the high-frequency component of the vibration sensor signal suddenly increased, and at the same time, the sound pressure level of a specific frequency band in the noise spectrum rose sharply. The probability of the cavitation boundary being approached by the predictive AI agent exceeded 90%. The system determined that the cavitation boundary had been triggered and recorded the precise operating conditions and multi-source feature data of this boundary point.

[0265] Throughout the testing process, the digital twin consistently ran parallel simulations at a speed thirty times faster than real-time. When the exploring AI agent attempted to approach the low-flow boundary, the twin predicted that the bearing temperature would exceed the alarm threshold of 85 degrees Celsius in the next twenty seconds. The system immediately triggered a level-two warning and automatically implemented a mitigation strategy, pausing the command to further reduce the flow rate, thereby avoiding the risk of overheating. Multi-source heterogeneous information, including vibration spectra, current harmonics, thermal imaging images, and noise signals, was all synchronously acquired using a high-precision clock.

[0266] Finally, all data was fused and analyzed. An interpretable tree ensemble model calculated the extracted features, outputting a comprehensive performance score of 88 points and a mechanical health level of B. The vibration health index was calculated to be 0.76, indicating slight, acceptable bearing wear. The digital profile report also noted that its high-efficiency range was narrower than similar pumps. The complete data chain and profile results of this test were archived to a historical database for updating the startup identification model, optimizing the AI ​​agent's strategy, and calibrating the digital twin model parameters, thereby enhancing the system's knowledge base.

[0267] Example 2:

[0268] This embodiment uses a circulating pump for a heating system as the test object. During the system startup identification phase, after starting at an 18 Hz frequency, its startup transient signal was analyzed. It was found that the starting current harmonic distortion rate was low, and the integral error of the control system reaching steady state was small. The pre-trained classification model determined that it was a "circulating pump", with a rated power level of "low power" and an impeller type tendency of "centrifugal".

[0269] The digital twin model is rapidly instantiated based on the identified type and input rated parameters. Through simulation pre-calculation, a slight hump feature is identified in the pump's head-flow curve, and nine key target test points, including points on both sides of the hump region, are planned accordingly. In the adaptive sampling phase, the system first completes sparse testing of these nine points. In the subsequent fine scanning, curvature analysis shows significant curve changes near the hump region. The system decides to add four additional high-density sampling points in this area, thereby accurately depicting the shape of the hump. Finally, basic performance testing is completed with fifteen test points, and the efficiency at the high-efficiency point is determined to be 78.5%.

[0270] During the boundary exploration phase, the AI ​​agent, guided by the predictive AI, proactively adjusts operating conditions to explore performance degradation trends and vibration boundaries under high flow rates. When the flow rate increases to 130% of the rated value, the vibration intensity continues to rise, increasing the "boundary approach reward" for the AI ​​agent and prompting it to continue exploring. The system determines that the "vibration exceeding the limit" has been reached and records it when the total vibration value approaches 95% of the upper limit of the safety threshold. Simultaneously, the digital twin's real-time advanced simulation predicts that if operation continues at the current rate, the motor current will exceed the limit after 15 seconds, triggering a level one warning to alert the operator.

[0271] Temperature field images under all operating conditions were simultaneously acquired from multiple sources, showing that the temperature of the bearing housing at the non-drive end of the motor remained normal. During the digital profile generation phase, feature fusion analysis confirmed the pump's hump characteristics, with an operational stability score of 85. The interpretable model, through SHAP value analysis, indicated that the main factor affecting its mechanical health level was the high vibration amplitude at twice the rotational frequency, contributing 40%, suggesting a possible slight misalignment issue. The final structured digital profile was stored in a database, and its actual performance curve data was used to back-calibrate the digital twin template of this pump model, making its fluid dynamics model parameters more accurate and providing a more accurate simulation benchmark for future testing of similar pumps.

[0272] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A performance testing system for both circulating pumps and submersible pumps, characterized in that: The system, comprising a test bench, sensor array, data acquisition and control unit, drive unit, load conditioning unit, and intelligent control and analysis server, is configured to perform the following steps: S11: Start-up and intelligent identification, start the pump under test at a preset extremely low frequency and power, and simultaneously collect multi-dimensional weak signals during the start-up transient. S12: Digital twin model construction and test planning. Based on the identified type and parameters, combined with the input design parameters, a parameterized initial digital twin is constructed, and a sequence of key physical test target points, including boundary points and performance inflection points, is automatically planned through simulation analysis. S13: Adaptive basic performance sampling. First, sparse scanning tests are performed on the key target points of the plan to obtain the skeleton of the performance curve. Then, the local features of the curve are calculated based on the measured data, and the density and distribution of subsequent sampling points are adaptively adjusted to complete the fine scanning. S14: Boundary performance exploration. Based on the obtained basic performance data, the predictive AI agent is used to predict the performance boundary and drive the exploration AI agent to autonomously adjust the operating parameters under safety constraints, actively exploring the pump's fault boundary and performance boundary. S15: Real-time security monitoring and prediction. During the test, the high-speed simulation of the digital twin is run in parallel to predict the system state in the future and issue an early warning and trigger active intervention before the system exceeds the safety threshold. S16: Synchronous acquisition of multi-source heterogeneous information. During the test, time-series and image data from vibration, current, thermal imaging and noise sensors are acquired synchronously and aligned in time. S17: Digital profile generation and model training. Integrate all test data and multi-source information, use interpretable machine learning models for training and analysis, generate a multi-dimensional, quantifiable comprehensive performance and health digital profile of the tested pump, and update the historical database.

2. The performance testing system for both circulating pumps and submersible pumps according to claim 1, characterized in that: The specific steps of startup and intelligent identification include: S21: Control command sent. The intelligent control and analysis server sends a control command to the drive unit, which instructs the drive unit to start the pump under test at an initial frequency that is 10% to 20% lower than the rated frequency of the pump under test, and a corresponding starting torque that is lower than the rated torque. S22: Multi-dimensional weak signal synchronous acquisition. During the startup process, electrical characteristic signals, mechanical characteristic signals and control response signals are synchronously acquired through a sensor array. S23: Feature extraction and fusion judgment. The collected signal is processed to extract feature vectors representing the pump type and basic parameters, and input into the pre-trained classification judgment model to output the judgment result of the type and key parameters of the pump under test.

3. The performance testing system for both circulating pumps and submersible pumps according to claim 2, characterized in that: The specific steps for building and testing a digital twin model include: S31: Model template matching and parameterized instantiation. Based on the identified pump type and parameters, the corresponding multi-domain physical model template is matched from the digital twin model library, and the identified parameters and user-input design parameters are used as initialization parameters to generate a parameterized initial digital twin. S32: Simulation analysis and feature point pre-calculation. Within the initial digital twin, the performance curve cluster of the pump is calculated through numerical simulation over the entire operating range. Based on the mathematical characteristics of the curves and preset rules, the performance inflection point and safety boundary point are pre-calculated. S33: Dynamic test sequence generation. All pre-calculated key target points are evaluated and sorted to generate a dynamic test sequence that guides the execution of physical tests.

4. The performance testing system for both circulating pumps and submersible pumps according to claim 3, characterized in that: The simulation analysis and feature point pre-calculation steps specifically include: S321: Pre-location of the highest efficiency point. The predicted flow rate corresponding to the highest efficiency point is obtained by solving for the point where the first derivative of the efficiency-flow curve is zero. S322: Head curve hump identification. By analyzing the head-flow curve, the point where the head change rate changes from negative to positive within a specific flow range is identified as the starting point of the hump. S323: Cavitation Boundary Point Prediction. Based on the required net positive suction head (NPSH) model, the required NPSH of the pump at different flow rates is calculated and compared with the effective NPSH of the system to predict the critical flow rate at which cavitation occurs. The judgment condition is as follows: ； in, For safety margin, This represents the system's effective net positive suction head (NPSH). This is the required net positive suction head (NPSH). S324: Minimum continuous thermally stable flow point prediction. Based on the pump power-flow curve and heat dissipation model, it calculates the pump temperature rise under different low flow conditions and predicts the critical flow rate at which the temperature rise exceeds the allowable limits of materials and seals. S325: Maximum allowable working pressure point prediction. Based on the pump body structural strength model, the stress of the pump pressure-bearing components is calculated at different head near the shut-off point, and the critical head at which the allowable stress is reached is predicted.

5. The performance testing system for both circulating pumps and submersible pumps according to claim 4, characterized in that: The adaptive basic performance sampling steps specifically include: S41: Sparse scanning of key target points. The control drive unit and load adjustment unit adjust the pump to each key target operating point in sequence according to the generated dynamic test sequence and wait for it to stabilize before performing the first measurement to obtain a sparse initial sampling point dataset that covers the key area of ​​the performance curve. S42: Curve local feature analysis. Based on the obtained initial sampling point data, the target performance curve is piecewise fitted, and the local mathematical features of each curve segment are calculated. S43: Adaptive sampling decision-making, which dynamically determines the location and density of new sampling points to be added based on the calculated local features; S44: Incremental sampling and model update. Perform incremental testing and sampling at the new determined position, add the new data points to the dataset, and update the performance curve model. S45: Iteration and Termination. Repeat steps S41 to S44 until the overall fitting accuracy of the target performance curve meets the preset threshold, thus completing the fine scan.

6. The performance testing system for both circulating pumps and submersible pumps according to claim 5, characterized in that: In the adaptive sampling decision-making step, the adaptive sampling decision is based on curvature and a preset curvature threshold, with the following specific rules: A11: For the smooth sections of a curve with curvature less than or equal to the minimum curvature threshold, set a low sampling density; A12: For general variation segments of curves with curvature greater than the minimum curvature threshold and less than the maximum curvature threshold, set a medium sampling density; A13: For high curvature segments of curves where the curvature is greater than or equal to the maximum curvature threshold, set a high sampling density; A14: The specific location of the newly added sampling point is selected from the sub-interval with the largest prediction uncertainty in the corresponding interval.

7. The performance testing system for both circulating pumps and submersible pumps according to claim 6, characterized in that: The specific steps for boundary performance exploration include: S51: Agent initialization and modeling. Based on the obtained basic performance curves and multi-source historical data, the predictive AI agent and the exploratory AI agent are initialized, and a reinforcement learning exploratory environment model including state space, action space, safety constraints and reward function is constructed. S52: Boundary exploration loop. By exploring the AI ​​agent, it selects and executes control actions based on the current state and its policy network, driving the test bench to change the working conditions. It predicts the AI ​​agent's assessment of state risks and predicts the boundary. The system monitors safety constraints and collects multi-source response data. S53: Boundary confirmation and marking. When the exploration process triggers the boundary determination condition, the system confirms and records the precise working condition and multi-source feature data of the boundary point. S54: Model update and policy optimization. Using the state-action-reward data sequence generated by exploration, update the policy network of the exploration AI agent, and use the newly discovered boundary data to optimize the prediction AI agent model. S55: Exploration terminated. When the preset exploration termination conditions are met, the test in this phase will be terminated.

8. The performance testing system for both circulating pumps and submersible pumps according to claim 7, characterized in that: The real-time security monitoring and prediction steps specifically include: S61: Real-time data synchronization and twin state initialization, synchronously injecting the real-time control commands and sensor data streams of the physical test system into the digital twin, so that the virtual state of the twin is aligned with the current state of the physical system; S62: Future state multi-step prediction. Using the current alignment state as the initial condition, the digital twin runs a closed-loop simulation at a speed faster than real-time to predict the evolution trajectory of key system state parameters within a preset time window in the future. S63: Predictive assessment of security risks, analyzing predicted trajectories to determine whether there is a risk of violating preset security constraints at any future time; S64: Tiered early warning and proactive intervention: Based on the urgency and severity of the risk, different levels of early warning are triggered and predetermined safety control strategies are automatically executed.

9. A performance testing system for both circulating pumps and submersible pumps according to claim 8, characterized in that: The specific steps of digital profile generation and model training include: S71: Multi-source heterogeneous data fusion and feature engineering, which gathers all the obtained structured test data and unstructured multi-source sensing data, performs spatiotemporal alignment, cleaning and feature extraction, and constructs a feature matrix; S72: Interpretable profile model training, using the feature matrix as input, trains an interpretable machine learning model, and the output of the machine learning model is a multi-dimensional quantitative indicator representing the overall state of the pump. S73: Digital profile synthesis and report generation, which compares and combines the quantitative indicators output by the model with key features and thresholds to generate a structured digital profile report that includes multi-dimensional labels and scores. S74: Knowledge base iterative update, archive the complete data chain, feature matrix, digital profile and corresponding interpretable model feature contribution of this test to the historical database.

10. A performance testing system for both circulating pumps and submersible pumps according to claim 9, characterized in that: In the training steps of the interpretable profiling model, the interpretable machine learning model is a tree ensemble model. The input of the tree ensemble model is a feature vector, and the output is a series of quantifiable scores, including: The overall performance score characterizes the degree of conformity with design standards for similar pumps. Mechanical health rating indicates the condition of bearings and shaft systems; Electrical health level, characterizing the electrical condition of a motor; Operational stability score, comprehensive evaluation of vibration and noise.

Images